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.
66 std::vector<SDNode*> WorkList;
68 // AA - Used for DAG load/store alias analysis.
71 /// AddUsersToWorkList - When an instruction is simplified, add all users of
72 /// the instruction to the work lists because they might get more simplified
75 void AddUsersToWorkList(SDNode *N) {
76 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
81 /// visit - call the node-specific routine that knows how to fold each
82 /// particular type of node.
83 SDValue visit(SDNode *N);
86 /// AddToWorkList - Add to the work list making sure it's instance is at the
87 /// the back (next to be processed.)
88 void AddToWorkList(SDNode *N) {
89 removeFromWorkList(N);
90 WorkList.push_back(N);
93 /// removeFromWorkList - remove all instances of N from the worklist.
95 void removeFromWorkList(SDNode *N) {
96 WorkList.erase(std::remove(WorkList.begin(), WorkList.end(), N),
100 SDValue CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
103 SDValue CombineTo(SDNode *N, SDValue Res, bool AddTo = true) {
104 return CombineTo(N, &Res, 1, AddTo);
107 SDValue CombineTo(SDNode *N, SDValue Res0, SDValue Res1,
109 SDValue To[] = { Res0, Res1 };
110 return CombineTo(N, To, 2, AddTo);
113 void CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO);
117 /// SimplifyDemandedBits - Check the specified integer node value to see if
118 /// it can be simplified or if things it uses can be simplified by bit
119 /// propagation. If so, return true.
120 bool SimplifyDemandedBits(SDValue Op) {
121 unsigned BitWidth = Op.getValueType().getScalarType().getSizeInBits();
122 APInt Demanded = APInt::getAllOnesValue(BitWidth);
123 return SimplifyDemandedBits(Op, Demanded);
126 bool SimplifyDemandedBits(SDValue Op, const APInt &Demanded);
128 bool CombineToPreIndexedLoadStore(SDNode *N);
129 bool CombineToPostIndexedLoadStore(SDNode *N);
131 void ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad);
132 SDValue PromoteOperand(SDValue Op, EVT PVT, bool &Replace);
133 SDValue SExtPromoteOperand(SDValue Op, EVT PVT);
134 SDValue ZExtPromoteOperand(SDValue Op, EVT PVT);
135 SDValue PromoteIntBinOp(SDValue Op);
136 SDValue PromoteIntShiftOp(SDValue Op);
137 SDValue PromoteExtend(SDValue Op);
138 bool PromoteLoad(SDValue Op);
140 void ExtendSetCCUses(SmallVector<SDNode*, 4> SetCCs,
141 SDValue Trunc, SDValue ExtLoad, DebugLoc DL,
142 ISD::NodeType ExtType);
144 /// combine - call the node-specific routine that knows how to fold each
145 /// particular type of node. If that doesn't do anything, try the
146 /// target-specific DAG combines.
147 SDValue combine(SDNode *N);
149 // Visitation implementation - Implement dag node combining for different
150 // node types. The semantics are as follows:
152 // SDValue.getNode() == 0 - No change was made
153 // SDValue.getNode() == N - N was replaced, is dead and has been handled.
154 // otherwise - N should be replaced by the returned Operand.
156 SDValue visitTokenFactor(SDNode *N);
157 SDValue visitMERGE_VALUES(SDNode *N);
158 SDValue visitADD(SDNode *N);
159 SDValue visitSUB(SDNode *N);
160 SDValue visitADDC(SDNode *N);
161 SDValue visitADDE(SDNode *N);
162 SDValue visitMUL(SDNode *N);
163 SDValue visitSDIV(SDNode *N);
164 SDValue visitUDIV(SDNode *N);
165 SDValue visitSREM(SDNode *N);
166 SDValue visitUREM(SDNode *N);
167 SDValue visitMULHU(SDNode *N);
168 SDValue visitMULHS(SDNode *N);
169 SDValue visitSMUL_LOHI(SDNode *N);
170 SDValue visitUMUL_LOHI(SDNode *N);
171 SDValue visitSMULO(SDNode *N);
172 SDValue visitUMULO(SDNode *N);
173 SDValue visitSDIVREM(SDNode *N);
174 SDValue visitUDIVREM(SDNode *N);
175 SDValue visitAND(SDNode *N);
176 SDValue visitOR(SDNode *N);
177 SDValue visitXOR(SDNode *N);
178 SDValue SimplifyVBinOp(SDNode *N);
179 SDValue visitSHL(SDNode *N);
180 SDValue visitSRA(SDNode *N);
181 SDValue visitSRL(SDNode *N);
182 SDValue visitCTLZ(SDNode *N);
183 SDValue visitCTTZ(SDNode *N);
184 SDValue visitCTPOP(SDNode *N);
185 SDValue visitSELECT(SDNode *N);
186 SDValue visitSELECT_CC(SDNode *N);
187 SDValue visitSETCC(SDNode *N);
188 SDValue visitSIGN_EXTEND(SDNode *N);
189 SDValue visitZERO_EXTEND(SDNode *N);
190 SDValue visitANY_EXTEND(SDNode *N);
191 SDValue visitSIGN_EXTEND_INREG(SDNode *N);
192 SDValue visitTRUNCATE(SDNode *N);
193 SDValue visitBITCAST(SDNode *N);
194 SDValue visitBUILD_PAIR(SDNode *N);
195 SDValue visitFADD(SDNode *N);
196 SDValue visitFSUB(SDNode *N);
197 SDValue visitFMUL(SDNode *N);
198 SDValue visitFDIV(SDNode *N);
199 SDValue visitFREM(SDNode *N);
200 SDValue visitFCOPYSIGN(SDNode *N);
201 SDValue visitSINT_TO_FP(SDNode *N);
202 SDValue visitUINT_TO_FP(SDNode *N);
203 SDValue visitFP_TO_SINT(SDNode *N);
204 SDValue visitFP_TO_UINT(SDNode *N);
205 SDValue visitFP_ROUND(SDNode *N);
206 SDValue visitFP_ROUND_INREG(SDNode *N);
207 SDValue visitFP_EXTEND(SDNode *N);
208 SDValue visitFNEG(SDNode *N);
209 SDValue visitFABS(SDNode *N);
210 SDValue visitBRCOND(SDNode *N);
211 SDValue visitBR_CC(SDNode *N);
212 SDValue visitLOAD(SDNode *N);
213 SDValue visitSTORE(SDNode *N);
214 SDValue visitINSERT_VECTOR_ELT(SDNode *N);
215 SDValue visitEXTRACT_VECTOR_ELT(SDNode *N);
216 SDValue visitBUILD_VECTOR(SDNode *N);
217 SDValue visitCONCAT_VECTORS(SDNode *N);
218 SDValue visitEXTRACT_SUBVECTOR(SDNode *N);
219 SDValue visitVECTOR_SHUFFLE(SDNode *N);
220 SDValue visitMEMBARRIER(SDNode *N);
222 SDValue XformToShuffleWithZero(SDNode *N);
223 SDValue ReassociateOps(unsigned Opc, DebugLoc DL, SDValue LHS, SDValue RHS);
225 SDValue visitShiftByConstant(SDNode *N, unsigned Amt);
227 bool SimplifySelectOps(SDNode *SELECT, SDValue LHS, SDValue RHS);
228 SDValue SimplifyBinOpWithSameOpcodeHands(SDNode *N);
229 SDValue SimplifySelect(DebugLoc DL, SDValue N0, SDValue N1, SDValue N2);
230 SDValue SimplifySelectCC(DebugLoc DL, SDValue N0, SDValue N1, SDValue N2,
231 SDValue N3, ISD::CondCode CC,
232 bool NotExtCompare = false);
233 SDValue SimplifySetCC(EVT VT, SDValue N0, SDValue N1, ISD::CondCode Cond,
234 DebugLoc DL, bool foldBooleans = true);
235 SDValue SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
237 SDValue CombineConsecutiveLoads(SDNode *N, EVT VT);
238 SDValue ConstantFoldBITCASTofBUILD_VECTOR(SDNode *, EVT);
239 SDValue BuildSDIV(SDNode *N);
240 SDValue BuildUDIV(SDNode *N);
241 SDValue MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
242 bool DemandHighBits = true);
243 SDValue MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1);
244 SDNode *MatchRotate(SDValue LHS, SDValue RHS, DebugLoc DL);
245 SDValue ReduceLoadWidth(SDNode *N);
246 SDValue ReduceLoadOpStoreWidth(SDNode *N);
247 SDValue TransformFPLoadStorePair(SDNode *N);
249 SDValue GetDemandedBits(SDValue V, const APInt &Mask);
251 /// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes,
252 /// looking for aliasing nodes and adding them to the Aliases vector.
253 void GatherAllAliases(SDNode *N, SDValue OriginalChain,
254 SmallVector<SDValue, 8> &Aliases);
256 /// isAlias - Return true if there is any possibility that the two addresses
258 bool isAlias(SDValue Ptr1, int64_t Size1,
259 const Value *SrcValue1, int SrcValueOffset1,
260 unsigned SrcValueAlign1,
261 const MDNode *TBAAInfo1,
262 SDValue Ptr2, int64_t Size2,
263 const Value *SrcValue2, int SrcValueOffset2,
264 unsigned SrcValueAlign2,
265 const MDNode *TBAAInfo2) const;
267 /// FindAliasInfo - Extracts the relevant alias information from the memory
268 /// node. Returns true if the operand was a load.
269 bool FindAliasInfo(SDNode *N,
270 SDValue &Ptr, int64_t &Size,
271 const Value *&SrcValue, int &SrcValueOffset,
272 unsigned &SrcValueAlignment,
273 const MDNode *&TBAAInfo) const;
275 /// FindBetterChain - Walk up chain skipping non-aliasing memory nodes,
276 /// looking for a better chain (aliasing node.)
277 SDValue FindBetterChain(SDNode *N, SDValue Chain);
280 DAGCombiner(SelectionDAG &D, AliasAnalysis &A, CodeGenOpt::Level OL)
281 : DAG(D), TLI(D.getTargetLoweringInfo()), Level(BeforeLegalizeTypes),
282 OptLevel(OL), LegalOperations(false), LegalTypes(false), AA(A) {}
284 /// Run - runs the dag combiner on all nodes in the work list
285 void Run(CombineLevel AtLevel);
287 SelectionDAG &getDAG() const { return DAG; }
289 /// getShiftAmountTy - Returns a type large enough to hold any valid
290 /// shift amount - before type legalization these can be huge.
291 EVT getShiftAmountTy(EVT LHSTy) {
292 return LegalTypes ? TLI.getShiftAmountTy(LHSTy) : TLI.getPointerTy();
295 /// isTypeLegal - This method returns true if we are running before type
296 /// legalization or if the specified VT is legal.
297 bool isTypeLegal(const EVT &VT) {
298 if (!LegalTypes) return true;
299 return TLI.isTypeLegal(VT);
306 /// WorkListRemover - This class is a DAGUpdateListener that removes any deleted
307 /// nodes from the worklist.
308 class WorkListRemover : public SelectionDAG::DAGUpdateListener {
311 explicit WorkListRemover(DAGCombiner &dc) : DC(dc) {}
313 virtual void NodeDeleted(SDNode *N, SDNode *E) {
314 DC.removeFromWorkList(N);
317 virtual void NodeUpdated(SDNode *N) {
323 //===----------------------------------------------------------------------===//
324 // TargetLowering::DAGCombinerInfo implementation
325 //===----------------------------------------------------------------------===//
327 void TargetLowering::DAGCombinerInfo::AddToWorklist(SDNode *N) {
328 ((DAGCombiner*)DC)->AddToWorkList(N);
331 void TargetLowering::DAGCombinerInfo::RemoveFromWorklist(SDNode *N) {
332 ((DAGCombiner*)DC)->removeFromWorkList(N);
335 SDValue TargetLowering::DAGCombinerInfo::
336 CombineTo(SDNode *N, const std::vector<SDValue> &To, bool AddTo) {
337 return ((DAGCombiner*)DC)->CombineTo(N, &To[0], To.size(), AddTo);
340 SDValue TargetLowering::DAGCombinerInfo::
341 CombineTo(SDNode *N, SDValue Res, bool AddTo) {
342 return ((DAGCombiner*)DC)->CombineTo(N, Res, AddTo);
346 SDValue TargetLowering::DAGCombinerInfo::
347 CombineTo(SDNode *N, SDValue Res0, SDValue Res1, bool AddTo) {
348 return ((DAGCombiner*)DC)->CombineTo(N, Res0, Res1, AddTo);
351 void TargetLowering::DAGCombinerInfo::
352 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
353 return ((DAGCombiner*)DC)->CommitTargetLoweringOpt(TLO);
356 //===----------------------------------------------------------------------===//
358 //===----------------------------------------------------------------------===//
360 /// isNegatibleForFree - Return 1 if we can compute the negated form of the
361 /// specified expression for the same cost as the expression itself, or 2 if we
362 /// can compute the negated form more cheaply than the expression itself.
363 static char isNegatibleForFree(SDValue Op, bool LegalOperations,
364 const TargetOptions *Options,
365 unsigned Depth = 0) {
366 // No compile time optimizations on this type.
367 if (Op.getValueType() == MVT::ppcf128)
370 // fneg is removable even if it has multiple uses.
371 if (Op.getOpcode() == ISD::FNEG) return 2;
373 // Don't allow anything with multiple uses.
374 if (!Op.hasOneUse()) return 0;
376 // Don't recurse exponentially.
377 if (Depth > 6) return 0;
379 switch (Op.getOpcode()) {
380 default: return false;
381 case ISD::ConstantFP:
382 // Don't invert constant FP values after legalize. The negated constant
383 // isn't necessarily legal.
384 return LegalOperations ? 0 : 1;
386 // FIXME: determine better conditions for this xform.
387 if (!Options->UnsafeFPMath) return 0;
389 // fold (fsub (fadd A, B)) -> (fsub (fneg A), B)
390 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, Options,
393 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
394 return isNegatibleForFree(Op.getOperand(1), LegalOperations, Options,
397 // We can't turn -(A-B) into B-A when we honor signed zeros.
398 if (!Options->UnsafeFPMath) return 0;
400 // fold (fneg (fsub A, B)) -> (fsub B, A)
405 if (Options->HonorSignDependentRoundingFPMath()) return 0;
407 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y) or (fmul X, (fneg Y))
408 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, Options,
412 return isNegatibleForFree(Op.getOperand(1), LegalOperations, Options,
418 return isNegatibleForFree(Op.getOperand(0), LegalOperations, Options,
423 /// GetNegatedExpression - If isNegatibleForFree returns true, this function
424 /// returns the newly negated expression.
425 static SDValue GetNegatedExpression(SDValue Op, SelectionDAG &DAG,
426 bool LegalOperations, unsigned Depth = 0) {
427 // fneg is removable even if it has multiple uses.
428 if (Op.getOpcode() == ISD::FNEG) return Op.getOperand(0);
430 // Don't allow anything with multiple uses.
431 assert(Op.hasOneUse() && "Unknown reuse!");
433 assert(Depth <= 6 && "GetNegatedExpression doesn't match isNegatibleForFree");
434 switch (Op.getOpcode()) {
435 default: llvm_unreachable("Unknown code");
436 case ISD::ConstantFP: {
437 APFloat V = cast<ConstantFPSDNode>(Op)->getValueAPF();
439 return DAG.getConstantFP(V, Op.getValueType());
442 // FIXME: determine better conditions for this xform.
443 assert(DAG.getTarget().Options.UnsafeFPMath);
445 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B)
446 if (isNegatibleForFree(Op.getOperand(0), LegalOperations,
447 &DAG.getTarget().Options, Depth+1))
448 return DAG.getNode(ISD::FSUB, Op.getDebugLoc(), Op.getValueType(),
449 GetNegatedExpression(Op.getOperand(0), DAG,
450 LegalOperations, Depth+1),
452 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
453 return DAG.getNode(ISD::FSUB, Op.getDebugLoc(), Op.getValueType(),
454 GetNegatedExpression(Op.getOperand(1), DAG,
455 LegalOperations, Depth+1),
458 // We can't turn -(A-B) into B-A when we honor signed zeros.
459 assert(DAG.getTarget().Options.UnsafeFPMath);
461 // fold (fneg (fsub 0, B)) -> B
462 if (ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(Op.getOperand(0)))
463 if (N0CFP->getValueAPF().isZero())
464 return Op.getOperand(1);
466 // fold (fneg (fsub A, B)) -> (fsub B, A)
467 return DAG.getNode(ISD::FSUB, Op.getDebugLoc(), Op.getValueType(),
468 Op.getOperand(1), Op.getOperand(0));
472 assert(!DAG.getTarget().Options.HonorSignDependentRoundingFPMath());
474 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y)
475 if (isNegatibleForFree(Op.getOperand(0), LegalOperations,
476 &DAG.getTarget().Options, Depth+1))
477 return DAG.getNode(Op.getOpcode(), Op.getDebugLoc(), Op.getValueType(),
478 GetNegatedExpression(Op.getOperand(0), DAG,
479 LegalOperations, Depth+1),
482 // fold (fneg (fmul X, Y)) -> (fmul X, (fneg Y))
483 return DAG.getNode(Op.getOpcode(), Op.getDebugLoc(), Op.getValueType(),
485 GetNegatedExpression(Op.getOperand(1), DAG,
486 LegalOperations, Depth+1));
490 return DAG.getNode(Op.getOpcode(), Op.getDebugLoc(), Op.getValueType(),
491 GetNegatedExpression(Op.getOperand(0), DAG,
492 LegalOperations, Depth+1));
494 return DAG.getNode(ISD::FP_ROUND, Op.getDebugLoc(), Op.getValueType(),
495 GetNegatedExpression(Op.getOperand(0), DAG,
496 LegalOperations, Depth+1),
502 // isSetCCEquivalent - Return true if this node is a setcc, or is a select_cc
503 // that selects between the values 1 and 0, making it equivalent to a setcc.
504 // Also, set the incoming LHS, RHS, and CC references to the appropriate
505 // nodes based on the type of node we are checking. This simplifies life a
506 // bit for the callers.
507 static bool isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS,
509 if (N.getOpcode() == ISD::SETCC) {
510 LHS = N.getOperand(0);
511 RHS = N.getOperand(1);
512 CC = N.getOperand(2);
515 if (N.getOpcode() == ISD::SELECT_CC &&
516 N.getOperand(2).getOpcode() == ISD::Constant &&
517 N.getOperand(3).getOpcode() == ISD::Constant &&
518 cast<ConstantSDNode>(N.getOperand(2))->getAPIntValue() == 1 &&
519 cast<ConstantSDNode>(N.getOperand(3))->isNullValue()) {
520 LHS = N.getOperand(0);
521 RHS = N.getOperand(1);
522 CC = N.getOperand(4);
528 // isOneUseSetCC - Return true if this is a SetCC-equivalent operation with only
529 // one use. If this is true, it allows the users to invert the operation for
530 // free when it is profitable to do so.
531 static bool isOneUseSetCC(SDValue N) {
533 if (isSetCCEquivalent(N, N0, N1, N2) && N.getNode()->hasOneUse())
538 SDValue DAGCombiner::ReassociateOps(unsigned Opc, DebugLoc DL,
539 SDValue N0, SDValue N1) {
540 EVT VT = N0.getValueType();
541 if (N0.getOpcode() == Opc && isa<ConstantSDNode>(N0.getOperand(1))) {
542 if (isa<ConstantSDNode>(N1)) {
543 // reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2))
545 DAG.FoldConstantArithmetic(Opc, VT,
546 cast<ConstantSDNode>(N0.getOperand(1)),
547 cast<ConstantSDNode>(N1));
548 return DAG.getNode(Opc, DL, VT, N0.getOperand(0), OpNode);
550 if (N0.hasOneUse()) {
551 // reassoc. (op (op x, c1), y) -> (op (op x, y), c1) iff x+c1 has one use
552 SDValue OpNode = DAG.getNode(Opc, N0.getDebugLoc(), VT,
553 N0.getOperand(0), N1);
554 AddToWorkList(OpNode.getNode());
555 return DAG.getNode(Opc, DL, VT, OpNode, N0.getOperand(1));
559 if (N1.getOpcode() == Opc && isa<ConstantSDNode>(N1.getOperand(1))) {
560 if (isa<ConstantSDNode>(N0)) {
561 // reassoc. (op c2, (op x, c1)) -> (op x, (op c1, c2))
563 DAG.FoldConstantArithmetic(Opc, VT,
564 cast<ConstantSDNode>(N1.getOperand(1)),
565 cast<ConstantSDNode>(N0));
566 return DAG.getNode(Opc, DL, VT, N1.getOperand(0), OpNode);
568 if (N1.hasOneUse()) {
569 // reassoc. (op y, (op x, c1)) -> (op (op x, y), c1) iff x+c1 has one use
570 SDValue OpNode = DAG.getNode(Opc, N0.getDebugLoc(), VT,
571 N1.getOperand(0), N0);
572 AddToWorkList(OpNode.getNode());
573 return DAG.getNode(Opc, DL, VT, OpNode, N1.getOperand(1));
580 SDValue DAGCombiner::CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
582 assert(N->getNumValues() == NumTo && "Broken CombineTo call!");
584 DEBUG(dbgs() << "\nReplacing.1 ";
586 dbgs() << "\nWith: ";
587 To[0].getNode()->dump(&DAG);
588 dbgs() << " and " << NumTo-1 << " other values\n";
589 for (unsigned i = 0, e = NumTo; i != e; ++i)
590 assert((!To[i].getNode() ||
591 N->getValueType(i) == To[i].getValueType()) &&
592 "Cannot combine value to value of different type!"));
593 WorkListRemover DeadNodes(*this);
594 DAG.ReplaceAllUsesWith(N, To, &DeadNodes);
597 // Push the new nodes and any users onto the worklist
598 for (unsigned i = 0, e = NumTo; i != e; ++i) {
599 if (To[i].getNode()) {
600 AddToWorkList(To[i].getNode());
601 AddUsersToWorkList(To[i].getNode());
606 // Finally, if the node is now dead, remove it from the graph. The node
607 // may not be dead if the replacement process recursively simplified to
608 // something else needing this node.
609 if (N->use_empty()) {
610 // Nodes can be reintroduced into the worklist. Make sure we do not
611 // process a node that has been replaced.
612 removeFromWorkList(N);
614 // Finally, since the node is now dead, remove it from the graph.
617 return SDValue(N, 0);
621 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
622 // Replace all uses. If any nodes become isomorphic to other nodes and
623 // are deleted, make sure to remove them from our worklist.
624 WorkListRemover DeadNodes(*this);
625 DAG.ReplaceAllUsesOfValueWith(TLO.Old, TLO.New, &DeadNodes);
627 // Push the new node and any (possibly new) users onto the worklist.
628 AddToWorkList(TLO.New.getNode());
629 AddUsersToWorkList(TLO.New.getNode());
631 // Finally, if the node is now dead, remove it from the graph. The node
632 // may not be dead if the replacement process recursively simplified to
633 // something else needing this node.
634 if (TLO.Old.getNode()->use_empty()) {
635 removeFromWorkList(TLO.Old.getNode());
637 // If the operands of this node are only used by the node, they will now
638 // be dead. Make sure to visit them first to delete dead nodes early.
639 for (unsigned i = 0, e = TLO.Old.getNode()->getNumOperands(); i != e; ++i)
640 if (TLO.Old.getNode()->getOperand(i).getNode()->hasOneUse())
641 AddToWorkList(TLO.Old.getNode()->getOperand(i).getNode());
643 DAG.DeleteNode(TLO.Old.getNode());
647 /// SimplifyDemandedBits - Check the specified integer node value to see if
648 /// it can be simplified or if things it uses can be simplified by bit
649 /// propagation. If so, return true.
650 bool DAGCombiner::SimplifyDemandedBits(SDValue Op, const APInt &Demanded) {
651 TargetLowering::TargetLoweringOpt TLO(DAG, LegalTypes, LegalOperations);
652 APInt KnownZero, KnownOne;
653 if (!TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
657 AddToWorkList(Op.getNode());
659 // Replace the old value with the new one.
661 DEBUG(dbgs() << "\nReplacing.2 ";
662 TLO.Old.getNode()->dump(&DAG);
663 dbgs() << "\nWith: ";
664 TLO.New.getNode()->dump(&DAG);
667 CommitTargetLoweringOpt(TLO);
671 void DAGCombiner::ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad) {
672 DebugLoc dl = Load->getDebugLoc();
673 EVT VT = Load->getValueType(0);
674 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, VT, SDValue(ExtLoad, 0));
676 DEBUG(dbgs() << "\nReplacing.9 ";
678 dbgs() << "\nWith: ";
679 Trunc.getNode()->dump(&DAG);
681 WorkListRemover DeadNodes(*this);
682 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 0), Trunc, &DeadNodes);
683 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), SDValue(ExtLoad, 1),
685 removeFromWorkList(Load);
686 DAG.DeleteNode(Load);
687 AddToWorkList(Trunc.getNode());
690 SDValue DAGCombiner::PromoteOperand(SDValue Op, EVT PVT, bool &Replace) {
692 DebugLoc dl = Op.getDebugLoc();
693 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Op)) {
694 EVT MemVT = LD->getMemoryVT();
695 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
696 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT) ? ISD::ZEXTLOAD
698 : LD->getExtensionType();
700 return DAG.getExtLoad(ExtType, dl, PVT,
701 LD->getChain(), LD->getBasePtr(),
702 LD->getPointerInfo(),
703 MemVT, LD->isVolatile(),
704 LD->isNonTemporal(), LD->getAlignment());
707 unsigned Opc = Op.getOpcode();
710 case ISD::AssertSext:
711 return DAG.getNode(ISD::AssertSext, dl, PVT,
712 SExtPromoteOperand(Op.getOperand(0), PVT),
714 case ISD::AssertZext:
715 return DAG.getNode(ISD::AssertZext, dl, PVT,
716 ZExtPromoteOperand(Op.getOperand(0), PVT),
718 case ISD::Constant: {
720 Op.getValueType().isByteSized() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
721 return DAG.getNode(ExtOpc, dl, PVT, Op);
725 if (!TLI.isOperationLegal(ISD::ANY_EXTEND, PVT))
727 return DAG.getNode(ISD::ANY_EXTEND, dl, PVT, Op);
730 SDValue DAGCombiner::SExtPromoteOperand(SDValue Op, EVT PVT) {
731 if (!TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, PVT))
733 EVT OldVT = Op.getValueType();
734 DebugLoc dl = Op.getDebugLoc();
735 bool Replace = false;
736 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
737 if (NewOp.getNode() == 0)
739 AddToWorkList(NewOp.getNode());
742 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
743 return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, NewOp.getValueType(), NewOp,
744 DAG.getValueType(OldVT));
747 SDValue DAGCombiner::ZExtPromoteOperand(SDValue Op, EVT PVT) {
748 EVT OldVT = Op.getValueType();
749 DebugLoc dl = Op.getDebugLoc();
750 bool Replace = false;
751 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
752 if (NewOp.getNode() == 0)
754 AddToWorkList(NewOp.getNode());
757 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
758 return DAG.getZeroExtendInReg(NewOp, dl, OldVT);
761 /// PromoteIntBinOp - Promote the specified integer binary operation if the
762 /// target indicates it is beneficial. e.g. On x86, it's usually better to
763 /// promote i16 operations to i32 since i16 instructions are longer.
764 SDValue DAGCombiner::PromoteIntBinOp(SDValue Op) {
765 if (!LegalOperations)
768 EVT VT = Op.getValueType();
769 if (VT.isVector() || !VT.isInteger())
772 // If operation type is 'undesirable', e.g. i16 on x86, consider
774 unsigned Opc = Op.getOpcode();
775 if (TLI.isTypeDesirableForOp(Opc, VT))
779 // Consult target whether it is a good idea to promote this operation and
780 // what's the right type to promote it to.
781 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
782 assert(PVT != VT && "Don't know what type to promote to!");
784 bool Replace0 = false;
785 SDValue N0 = Op.getOperand(0);
786 SDValue NN0 = PromoteOperand(N0, PVT, Replace0);
787 if (NN0.getNode() == 0)
790 bool Replace1 = false;
791 SDValue N1 = Op.getOperand(1);
796 NN1 = PromoteOperand(N1, PVT, Replace1);
797 if (NN1.getNode() == 0)
801 AddToWorkList(NN0.getNode());
803 AddToWorkList(NN1.getNode());
806 ReplaceLoadWithPromotedLoad(N0.getNode(), NN0.getNode());
808 ReplaceLoadWithPromotedLoad(N1.getNode(), NN1.getNode());
810 DEBUG(dbgs() << "\nPromoting ";
811 Op.getNode()->dump(&DAG));
812 DebugLoc dl = Op.getDebugLoc();
813 return DAG.getNode(ISD::TRUNCATE, dl, VT,
814 DAG.getNode(Opc, dl, PVT, NN0, NN1));
819 /// PromoteIntShiftOp - Promote the specified integer shift operation if the
820 /// target indicates it is beneficial. e.g. On x86, it's usually better to
821 /// promote i16 operations to i32 since i16 instructions are longer.
822 SDValue DAGCombiner::PromoteIntShiftOp(SDValue Op) {
823 if (!LegalOperations)
826 EVT VT = Op.getValueType();
827 if (VT.isVector() || !VT.isInteger())
830 // If operation type is 'undesirable', e.g. i16 on x86, consider
832 unsigned Opc = Op.getOpcode();
833 if (TLI.isTypeDesirableForOp(Opc, VT))
837 // Consult target whether it is a good idea to promote this operation and
838 // what's the right type to promote it to.
839 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
840 assert(PVT != VT && "Don't know what type to promote to!");
842 bool Replace = false;
843 SDValue N0 = Op.getOperand(0);
845 N0 = SExtPromoteOperand(Op.getOperand(0), PVT);
846 else if (Opc == ISD::SRL)
847 N0 = ZExtPromoteOperand(Op.getOperand(0), PVT);
849 N0 = PromoteOperand(N0, PVT, Replace);
850 if (N0.getNode() == 0)
853 AddToWorkList(N0.getNode());
855 ReplaceLoadWithPromotedLoad(Op.getOperand(0).getNode(), N0.getNode());
857 DEBUG(dbgs() << "\nPromoting ";
858 Op.getNode()->dump(&DAG));
859 DebugLoc dl = Op.getDebugLoc();
860 return DAG.getNode(ISD::TRUNCATE, dl, VT,
861 DAG.getNode(Opc, dl, PVT, N0, Op.getOperand(1)));
866 SDValue DAGCombiner::PromoteExtend(SDValue Op) {
867 if (!LegalOperations)
870 EVT VT = Op.getValueType();
871 if (VT.isVector() || !VT.isInteger())
874 // If operation type is 'undesirable', e.g. i16 on x86, consider
876 unsigned Opc = Op.getOpcode();
877 if (TLI.isTypeDesirableForOp(Opc, VT))
881 // Consult target whether it is a good idea to promote this operation and
882 // what's the right type to promote it to.
883 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
884 assert(PVT != VT && "Don't know what type to promote to!");
885 // fold (aext (aext x)) -> (aext x)
886 // fold (aext (zext x)) -> (zext x)
887 // fold (aext (sext x)) -> (sext x)
888 DEBUG(dbgs() << "\nPromoting ";
889 Op.getNode()->dump(&DAG));
890 return DAG.getNode(Op.getOpcode(), Op.getDebugLoc(), VT, Op.getOperand(0));
895 bool DAGCombiner::PromoteLoad(SDValue Op) {
896 if (!LegalOperations)
899 EVT VT = Op.getValueType();
900 if (VT.isVector() || !VT.isInteger())
903 // If operation type is 'undesirable', e.g. i16 on x86, consider
905 unsigned Opc = Op.getOpcode();
906 if (TLI.isTypeDesirableForOp(Opc, VT))
910 // Consult target whether it is a good idea to promote this operation and
911 // what's the right type to promote it to.
912 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
913 assert(PVT != VT && "Don't know what type to promote to!");
915 DebugLoc dl = Op.getDebugLoc();
916 SDNode *N = Op.getNode();
917 LoadSDNode *LD = cast<LoadSDNode>(N);
918 EVT MemVT = LD->getMemoryVT();
919 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
920 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT) ? ISD::ZEXTLOAD
922 : LD->getExtensionType();
923 SDValue NewLD = DAG.getExtLoad(ExtType, dl, PVT,
924 LD->getChain(), LD->getBasePtr(),
925 LD->getPointerInfo(),
926 MemVT, LD->isVolatile(),
927 LD->isNonTemporal(), LD->getAlignment());
928 SDValue Result = DAG.getNode(ISD::TRUNCATE, dl, VT, NewLD);
930 DEBUG(dbgs() << "\nPromoting ";
933 Result.getNode()->dump(&DAG);
935 WorkListRemover DeadNodes(*this);
936 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result, &DeadNodes);
937 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), NewLD.getValue(1), &DeadNodes);
938 removeFromWorkList(N);
940 AddToWorkList(Result.getNode());
947 //===----------------------------------------------------------------------===//
948 // Main DAG Combiner implementation
949 //===----------------------------------------------------------------------===//
951 void DAGCombiner::Run(CombineLevel AtLevel) {
952 // set the instance variables, so that the various visit routines may use it.
954 LegalOperations = Level >= AfterLegalizeVectorOps;
955 LegalTypes = Level >= AfterLegalizeTypes;
957 // Add all the dag nodes to the worklist.
958 WorkList.reserve(DAG.allnodes_size());
959 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
960 E = DAG.allnodes_end(); I != E; ++I)
961 WorkList.push_back(I);
963 // Create a dummy node (which is not added to allnodes), that adds a reference
964 // to the root node, preventing it from being deleted, and tracking any
965 // changes of the root.
966 HandleSDNode Dummy(DAG.getRoot());
968 // The root of the dag may dangle to deleted nodes until the dag combiner is
969 // done. Set it to null to avoid confusion.
970 DAG.setRoot(SDValue());
972 // while the worklist isn't empty, inspect the node on the end of it and
973 // try and combine it.
974 while (!WorkList.empty()) {
975 SDNode *N = WorkList.back();
978 // If N has no uses, it is dead. Make sure to revisit all N's operands once
979 // N is deleted from the DAG, since they too may now be dead or may have a
980 // reduced number of uses, allowing other xforms.
981 if (N->use_empty() && N != &Dummy) {
982 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
983 AddToWorkList(N->getOperand(i).getNode());
989 SDValue RV = combine(N);
991 if (RV.getNode() == 0)
996 // If we get back the same node we passed in, rather than a new node or
997 // zero, we know that the node must have defined multiple values and
998 // CombineTo was used. Since CombineTo takes care of the worklist
999 // mechanics for us, we have no work to do in this case.
1000 if (RV.getNode() == N)
1003 assert(N->getOpcode() != ISD::DELETED_NODE &&
1004 RV.getNode()->getOpcode() != ISD::DELETED_NODE &&
1005 "Node was deleted but visit returned new node!");
1007 DEBUG(dbgs() << "\nReplacing.3 ";
1009 dbgs() << "\nWith: ";
1010 RV.getNode()->dump(&DAG);
1013 // Transfer debug value.
1014 DAG.TransferDbgValues(SDValue(N, 0), RV);
1015 WorkListRemover DeadNodes(*this);
1016 if (N->getNumValues() == RV.getNode()->getNumValues())
1017 DAG.ReplaceAllUsesWith(N, RV.getNode(), &DeadNodes);
1019 assert(N->getValueType(0) == RV.getValueType() &&
1020 N->getNumValues() == 1 && "Type mismatch");
1022 DAG.ReplaceAllUsesWith(N, &OpV, &DeadNodes);
1025 // Push the new node and any users onto the worklist
1026 AddToWorkList(RV.getNode());
1027 AddUsersToWorkList(RV.getNode());
1029 // Add any uses of the old node to the worklist in case this node is the
1030 // last one that uses them. They may become dead after this node is
1032 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1033 AddToWorkList(N->getOperand(i).getNode());
1035 // Finally, if the node is now dead, remove it from the graph. The node
1036 // may not be dead if the replacement process recursively simplified to
1037 // something else needing this node.
1038 if (N->use_empty()) {
1039 // Nodes can be reintroduced into the worklist. Make sure we do not
1040 // process a node that has been replaced.
1041 removeFromWorkList(N);
1043 // Finally, since the node is now dead, remove it from the graph.
1048 // If the root changed (e.g. it was a dead load, update the root).
1049 DAG.setRoot(Dummy.getValue());
1052 SDValue DAGCombiner::visit(SDNode *N) {
1053 switch (N->getOpcode()) {
1055 case ISD::TokenFactor: return visitTokenFactor(N);
1056 case ISD::MERGE_VALUES: return visitMERGE_VALUES(N);
1057 case ISD::ADD: return visitADD(N);
1058 case ISD::SUB: return visitSUB(N);
1059 case ISD::ADDC: return visitADDC(N);
1060 case ISD::ADDE: return visitADDE(N);
1061 case ISD::MUL: return visitMUL(N);
1062 case ISD::SDIV: return visitSDIV(N);
1063 case ISD::UDIV: return visitUDIV(N);
1064 case ISD::SREM: return visitSREM(N);
1065 case ISD::UREM: return visitUREM(N);
1066 case ISD::MULHU: return visitMULHU(N);
1067 case ISD::MULHS: return visitMULHS(N);
1068 case ISD::SMUL_LOHI: return visitSMUL_LOHI(N);
1069 case ISD::UMUL_LOHI: return visitUMUL_LOHI(N);
1070 case ISD::SMULO: return visitSMULO(N);
1071 case ISD::UMULO: return visitUMULO(N);
1072 case ISD::SDIVREM: return visitSDIVREM(N);
1073 case ISD::UDIVREM: return visitUDIVREM(N);
1074 case ISD::AND: return visitAND(N);
1075 case ISD::OR: return visitOR(N);
1076 case ISD::XOR: return visitXOR(N);
1077 case ISD::SHL: return visitSHL(N);
1078 case ISD::SRA: return visitSRA(N);
1079 case ISD::SRL: return visitSRL(N);
1080 case ISD::CTLZ: return visitCTLZ(N);
1081 case ISD::CTTZ: return visitCTTZ(N);
1082 case ISD::CTPOP: return visitCTPOP(N);
1083 case ISD::SELECT: return visitSELECT(N);
1084 case ISD::SELECT_CC: return visitSELECT_CC(N);
1085 case ISD::SETCC: return visitSETCC(N);
1086 case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N);
1087 case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N);
1088 case ISD::ANY_EXTEND: return visitANY_EXTEND(N);
1089 case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N);
1090 case ISD::TRUNCATE: return visitTRUNCATE(N);
1091 case ISD::BITCAST: return visitBITCAST(N);
1092 case ISD::BUILD_PAIR: return visitBUILD_PAIR(N);
1093 case ISD::FADD: return visitFADD(N);
1094 case ISD::FSUB: return visitFSUB(N);
1095 case ISD::FMUL: return visitFMUL(N);
1096 case ISD::FDIV: return visitFDIV(N);
1097 case ISD::FREM: return visitFREM(N);
1098 case ISD::FCOPYSIGN: return visitFCOPYSIGN(N);
1099 case ISD::SINT_TO_FP: return visitSINT_TO_FP(N);
1100 case ISD::UINT_TO_FP: return visitUINT_TO_FP(N);
1101 case ISD::FP_TO_SINT: return visitFP_TO_SINT(N);
1102 case ISD::FP_TO_UINT: return visitFP_TO_UINT(N);
1103 case ISD::FP_ROUND: return visitFP_ROUND(N);
1104 case ISD::FP_ROUND_INREG: return visitFP_ROUND_INREG(N);
1105 case ISD::FP_EXTEND: return visitFP_EXTEND(N);
1106 case ISD::FNEG: return visitFNEG(N);
1107 case ISD::FABS: return visitFABS(N);
1108 case ISD::BRCOND: return visitBRCOND(N);
1109 case ISD::BR_CC: return visitBR_CC(N);
1110 case ISD::LOAD: return visitLOAD(N);
1111 case ISD::STORE: return visitSTORE(N);
1112 case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N);
1113 case ISD::EXTRACT_VECTOR_ELT: return visitEXTRACT_VECTOR_ELT(N);
1114 case ISD::BUILD_VECTOR: return visitBUILD_VECTOR(N);
1115 case ISD::CONCAT_VECTORS: return visitCONCAT_VECTORS(N);
1116 case ISD::EXTRACT_SUBVECTOR: return visitEXTRACT_SUBVECTOR(N);
1117 case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N);
1118 case ISD::MEMBARRIER: return visitMEMBARRIER(N);
1123 SDValue DAGCombiner::combine(SDNode *N) {
1124 SDValue RV = visit(N);
1126 // If nothing happened, try a target-specific DAG combine.
1127 if (RV.getNode() == 0) {
1128 assert(N->getOpcode() != ISD::DELETED_NODE &&
1129 "Node was deleted but visit returned NULL!");
1131 if (N->getOpcode() >= ISD::BUILTIN_OP_END ||
1132 TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode())) {
1134 // Expose the DAG combiner to the target combiner impls.
1135 TargetLowering::DAGCombinerInfo
1136 DagCombineInfo(DAG, !LegalTypes, !LegalOperations, false, this);
1138 RV = TLI.PerformDAGCombine(N, DagCombineInfo);
1142 // If nothing happened still, try promoting the operation.
1143 if (RV.getNode() == 0) {
1144 switch (N->getOpcode()) {
1152 RV = PromoteIntBinOp(SDValue(N, 0));
1157 RV = PromoteIntShiftOp(SDValue(N, 0));
1159 case ISD::SIGN_EXTEND:
1160 case ISD::ZERO_EXTEND:
1161 case ISD::ANY_EXTEND:
1162 RV = PromoteExtend(SDValue(N, 0));
1165 if (PromoteLoad(SDValue(N, 0)))
1171 // If N is a commutative binary node, try commuting it to enable more
1173 if (RV.getNode() == 0 &&
1174 SelectionDAG::isCommutativeBinOp(N->getOpcode()) &&
1175 N->getNumValues() == 1) {
1176 SDValue N0 = N->getOperand(0);
1177 SDValue N1 = N->getOperand(1);
1179 // Constant operands are canonicalized to RHS.
1180 if (isa<ConstantSDNode>(N0) || !isa<ConstantSDNode>(N1)) {
1181 SDValue Ops[] = { N1, N0 };
1182 SDNode *CSENode = DAG.getNodeIfExists(N->getOpcode(), N->getVTList(),
1185 return SDValue(CSENode, 0);
1192 /// getInputChainForNode - Given a node, return its input chain if it has one,
1193 /// otherwise return a null sd operand.
1194 static SDValue getInputChainForNode(SDNode *N) {
1195 if (unsigned NumOps = N->getNumOperands()) {
1196 if (N->getOperand(0).getValueType() == MVT::Other)
1197 return N->getOperand(0);
1198 else if (N->getOperand(NumOps-1).getValueType() == MVT::Other)
1199 return N->getOperand(NumOps-1);
1200 for (unsigned i = 1; i < NumOps-1; ++i)
1201 if (N->getOperand(i).getValueType() == MVT::Other)
1202 return N->getOperand(i);
1207 SDValue DAGCombiner::visitTokenFactor(SDNode *N) {
1208 // If N has two operands, where one has an input chain equal to the other,
1209 // the 'other' chain is redundant.
1210 if (N->getNumOperands() == 2) {
1211 if (getInputChainForNode(N->getOperand(0).getNode()) == N->getOperand(1))
1212 return N->getOperand(0);
1213 if (getInputChainForNode(N->getOperand(1).getNode()) == N->getOperand(0))
1214 return N->getOperand(1);
1217 SmallVector<SDNode *, 8> TFs; // List of token factors to visit.
1218 SmallVector<SDValue, 8> Ops; // Ops for replacing token factor.
1219 SmallPtrSet<SDNode*, 16> SeenOps;
1220 bool Changed = false; // If we should replace this token factor.
1222 // Start out with this token factor.
1225 // Iterate through token factors. The TFs grows when new token factors are
1227 for (unsigned i = 0; i < TFs.size(); ++i) {
1228 SDNode *TF = TFs[i];
1230 // Check each of the operands.
1231 for (unsigned i = 0, ie = TF->getNumOperands(); i != ie; ++i) {
1232 SDValue Op = TF->getOperand(i);
1234 switch (Op.getOpcode()) {
1235 case ISD::EntryToken:
1236 // Entry tokens don't need to be added to the list. They are
1241 case ISD::TokenFactor:
1242 if (Op.hasOneUse() &&
1243 std::find(TFs.begin(), TFs.end(), Op.getNode()) == TFs.end()) {
1244 // Queue up for processing.
1245 TFs.push_back(Op.getNode());
1246 // Clean up in case the token factor is removed.
1247 AddToWorkList(Op.getNode());
1254 // Only add if it isn't already in the list.
1255 if (SeenOps.insert(Op.getNode()))
1266 // If we've change things around then replace token factor.
1269 // The entry token is the only possible outcome.
1270 Result = DAG.getEntryNode();
1272 // New and improved token factor.
1273 Result = DAG.getNode(ISD::TokenFactor, N->getDebugLoc(),
1274 MVT::Other, &Ops[0], Ops.size());
1277 // Don't add users to work list.
1278 return CombineTo(N, Result, false);
1284 /// MERGE_VALUES can always be eliminated.
1285 SDValue DAGCombiner::visitMERGE_VALUES(SDNode *N) {
1286 WorkListRemover DeadNodes(*this);
1287 // Replacing results may cause a different MERGE_VALUES to suddenly
1288 // be CSE'd with N, and carry its uses with it. Iterate until no
1289 // uses remain, to ensure that the node can be safely deleted.
1291 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1292 DAG.ReplaceAllUsesOfValueWith(SDValue(N, i), N->getOperand(i),
1294 } while (!N->use_empty());
1295 removeFromWorkList(N);
1297 return SDValue(N, 0); // Return N so it doesn't get rechecked!
1301 SDValue combineShlAddConstant(DebugLoc DL, SDValue N0, SDValue N1,
1302 SelectionDAG &DAG) {
1303 EVT VT = N0.getValueType();
1304 SDValue N00 = N0.getOperand(0);
1305 SDValue N01 = N0.getOperand(1);
1306 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N01);
1308 if (N01C && N00.getOpcode() == ISD::ADD && N00.getNode()->hasOneUse() &&
1309 isa<ConstantSDNode>(N00.getOperand(1))) {
1310 // fold (add (shl (add x, c1), c2), ) -> (add (add (shl x, c2), c1<<c2), )
1311 N0 = DAG.getNode(ISD::ADD, N0.getDebugLoc(), VT,
1312 DAG.getNode(ISD::SHL, N00.getDebugLoc(), VT,
1313 N00.getOperand(0), N01),
1314 DAG.getNode(ISD::SHL, N01.getDebugLoc(), VT,
1315 N00.getOperand(1), N01));
1316 return DAG.getNode(ISD::ADD, DL, VT, N0, N1);
1322 SDValue DAGCombiner::visitADD(SDNode *N) {
1323 SDValue N0 = N->getOperand(0);
1324 SDValue N1 = N->getOperand(1);
1325 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1326 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1327 EVT VT = N0.getValueType();
1330 if (VT.isVector()) {
1331 SDValue FoldedVOp = SimplifyVBinOp(N);
1332 if (FoldedVOp.getNode()) return FoldedVOp;
1335 // fold (add x, undef) -> undef
1336 if (N0.getOpcode() == ISD::UNDEF)
1338 if (N1.getOpcode() == ISD::UNDEF)
1340 // fold (add c1, c2) -> c1+c2
1342 return DAG.FoldConstantArithmetic(ISD::ADD, VT, N0C, N1C);
1343 // canonicalize constant to RHS
1345 return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N1, N0);
1346 // fold (add x, 0) -> x
1347 if (N1C && N1C->isNullValue())
1349 // fold (add Sym, c) -> Sym+c
1350 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1351 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA) && N1C &&
1352 GA->getOpcode() == ISD::GlobalAddress)
1353 return DAG.getGlobalAddress(GA->getGlobal(), N1C->getDebugLoc(), VT,
1355 (uint64_t)N1C->getSExtValue());
1356 // fold ((c1-A)+c2) -> (c1+c2)-A
1357 if (N1C && N0.getOpcode() == ISD::SUB)
1358 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getOperand(0)))
1359 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1360 DAG.getConstant(N1C->getAPIntValue()+
1361 N0C->getAPIntValue(), VT),
1364 SDValue RADD = ReassociateOps(ISD::ADD, N->getDebugLoc(), N0, N1);
1365 if (RADD.getNode() != 0)
1367 // fold ((0-A) + B) -> B-A
1368 if (N0.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N0.getOperand(0)) &&
1369 cast<ConstantSDNode>(N0.getOperand(0))->isNullValue())
1370 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N1, N0.getOperand(1));
1371 // fold (A + (0-B)) -> A-B
1372 if (N1.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N1.getOperand(0)) &&
1373 cast<ConstantSDNode>(N1.getOperand(0))->isNullValue())
1374 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N0, N1.getOperand(1));
1375 // fold (A+(B-A)) -> B
1376 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1))
1377 return N1.getOperand(0);
1378 // fold ((B-A)+A) -> B
1379 if (N0.getOpcode() == ISD::SUB && N1 == N0.getOperand(1))
1380 return N0.getOperand(0);
1381 // fold (A+(B-(A+C))) to (B-C)
1382 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1383 N0 == N1.getOperand(1).getOperand(0))
1384 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N1.getOperand(0),
1385 N1.getOperand(1).getOperand(1));
1386 // fold (A+(B-(C+A))) to (B-C)
1387 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1388 N0 == N1.getOperand(1).getOperand(1))
1389 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N1.getOperand(0),
1390 N1.getOperand(1).getOperand(0));
1391 // fold (A+((B-A)+or-C)) to (B+or-C)
1392 if ((N1.getOpcode() == ISD::SUB || N1.getOpcode() == ISD::ADD) &&
1393 N1.getOperand(0).getOpcode() == ISD::SUB &&
1394 N0 == N1.getOperand(0).getOperand(1))
1395 return DAG.getNode(N1.getOpcode(), N->getDebugLoc(), VT,
1396 N1.getOperand(0).getOperand(0), N1.getOperand(1));
1398 // fold (A-B)+(C-D) to (A+C)-(B+D) when A or C is constant
1399 if (N0.getOpcode() == ISD::SUB && N1.getOpcode() == ISD::SUB) {
1400 SDValue N00 = N0.getOperand(0);
1401 SDValue N01 = N0.getOperand(1);
1402 SDValue N10 = N1.getOperand(0);
1403 SDValue N11 = N1.getOperand(1);
1405 if (isa<ConstantSDNode>(N00) || isa<ConstantSDNode>(N10))
1406 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1407 DAG.getNode(ISD::ADD, N0.getDebugLoc(), VT, N00, N10),
1408 DAG.getNode(ISD::ADD, N1.getDebugLoc(), VT, N01, N11));
1411 if (!VT.isVector() && SimplifyDemandedBits(SDValue(N, 0)))
1412 return SDValue(N, 0);
1414 // fold (a+b) -> (a|b) iff a and b share no bits.
1415 if (VT.isInteger() && !VT.isVector()) {
1416 APInt LHSZero, LHSOne;
1417 APInt RHSZero, RHSOne;
1418 APInt Mask = APInt::getAllOnesValue(VT.getScalarType().getSizeInBits());
1419 DAG.ComputeMaskedBits(N0, Mask, LHSZero, LHSOne);
1421 if (LHSZero.getBoolValue()) {
1422 DAG.ComputeMaskedBits(N1, Mask, RHSZero, RHSOne);
1424 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1425 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1426 if ((RHSZero & (~LHSZero & Mask)) == (~LHSZero & Mask) ||
1427 (LHSZero & (~RHSZero & Mask)) == (~RHSZero & Mask))
1428 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N0, N1);
1432 // fold (add (shl (add x, c1), c2), ) -> (add (add (shl x, c2), c1<<c2), )
1433 if (N0.getOpcode() == ISD::SHL && N0.getNode()->hasOneUse()) {
1434 SDValue Result = combineShlAddConstant(N->getDebugLoc(), N0, N1, DAG);
1435 if (Result.getNode()) return Result;
1437 if (N1.getOpcode() == ISD::SHL && N1.getNode()->hasOneUse()) {
1438 SDValue Result = combineShlAddConstant(N->getDebugLoc(), N1, N0, DAG);
1439 if (Result.getNode()) return Result;
1442 // fold (add x, shl(0 - y, n)) -> sub(x, shl(y, n))
1443 if (N1.getOpcode() == ISD::SHL &&
1444 N1.getOperand(0).getOpcode() == ISD::SUB)
1445 if (ConstantSDNode *C =
1446 dyn_cast<ConstantSDNode>(N1.getOperand(0).getOperand(0)))
1447 if (C->getAPIntValue() == 0)
1448 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N0,
1449 DAG.getNode(ISD::SHL, N->getDebugLoc(), VT,
1450 N1.getOperand(0).getOperand(1),
1452 if (N0.getOpcode() == ISD::SHL &&
1453 N0.getOperand(0).getOpcode() == ISD::SUB)
1454 if (ConstantSDNode *C =
1455 dyn_cast<ConstantSDNode>(N0.getOperand(0).getOperand(0)))
1456 if (C->getAPIntValue() == 0)
1457 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N1,
1458 DAG.getNode(ISD::SHL, N->getDebugLoc(), VT,
1459 N0.getOperand(0).getOperand(1),
1462 if (N1.getOpcode() == ISD::AND) {
1463 SDValue AndOp0 = N1.getOperand(0);
1464 ConstantSDNode *AndOp1 = dyn_cast<ConstantSDNode>(N1->getOperand(1));
1465 unsigned NumSignBits = DAG.ComputeNumSignBits(AndOp0);
1466 unsigned DestBits = VT.getScalarType().getSizeInBits();
1468 // (add z, (and (sbbl x, x), 1)) -> (sub z, (sbbl x, x))
1469 // and similar xforms where the inner op is either ~0 or 0.
1470 if (NumSignBits == DestBits && AndOp1 && AndOp1->isOne()) {
1471 DebugLoc DL = N->getDebugLoc();
1472 return DAG.getNode(ISD::SUB, DL, VT, N->getOperand(0), AndOp0);
1476 // add (sext i1), X -> sub X, (zext i1)
1477 if (N0.getOpcode() == ISD::SIGN_EXTEND &&
1478 N0.getOperand(0).getValueType() == MVT::i1 &&
1479 !TLI.isOperationLegal(ISD::SIGN_EXTEND, MVT::i1)) {
1480 DebugLoc DL = N->getDebugLoc();
1481 SDValue ZExt = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0));
1482 return DAG.getNode(ISD::SUB, DL, VT, N1, ZExt);
1488 SDValue DAGCombiner::visitADDC(SDNode *N) {
1489 SDValue N0 = N->getOperand(0);
1490 SDValue N1 = N->getOperand(1);
1491 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1492 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1493 EVT VT = N0.getValueType();
1495 // If the flag result is dead, turn this into an ADD.
1496 if (N->hasNUsesOfValue(0, 1))
1497 return CombineTo(N, DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N1, N0),
1498 DAG.getNode(ISD::CARRY_FALSE,
1499 N->getDebugLoc(), MVT::Glue));
1501 // canonicalize constant to RHS.
1503 return DAG.getNode(ISD::ADDC, N->getDebugLoc(), N->getVTList(), N1, N0);
1505 // fold (addc x, 0) -> x + no carry out
1506 if (N1C && N1C->isNullValue())
1507 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE,
1508 N->getDebugLoc(), MVT::Glue));
1510 // fold (addc a, b) -> (or a, b), CARRY_FALSE iff a and b share no bits.
1511 APInt LHSZero, LHSOne;
1512 APInt RHSZero, RHSOne;
1513 APInt Mask = APInt::getAllOnesValue(VT.getScalarType().getSizeInBits());
1514 DAG.ComputeMaskedBits(N0, Mask, LHSZero, LHSOne);
1516 if (LHSZero.getBoolValue()) {
1517 DAG.ComputeMaskedBits(N1, Mask, RHSZero, RHSOne);
1519 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1520 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1521 if ((RHSZero & (~LHSZero & Mask)) == (~LHSZero & Mask) ||
1522 (LHSZero & (~RHSZero & Mask)) == (~RHSZero & Mask))
1523 return CombineTo(N, DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N0, N1),
1524 DAG.getNode(ISD::CARRY_FALSE,
1525 N->getDebugLoc(), MVT::Glue));
1531 SDValue DAGCombiner::visitADDE(SDNode *N) {
1532 SDValue N0 = N->getOperand(0);
1533 SDValue N1 = N->getOperand(1);
1534 SDValue CarryIn = N->getOperand(2);
1535 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1536 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1538 // canonicalize constant to RHS
1540 return DAG.getNode(ISD::ADDE, N->getDebugLoc(), N->getVTList(),
1543 // fold (adde x, y, false) -> (addc x, y)
1544 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1545 return DAG.getNode(ISD::ADDC, N->getDebugLoc(), N->getVTList(), N1, N0);
1550 // Since it may not be valid to emit a fold to zero for vector initializers
1551 // check if we can before folding.
1552 static SDValue tryFoldToZero(DebugLoc DL, const TargetLowering &TLI, EVT VT,
1553 SelectionDAG &DAG, bool LegalOperations) {
1554 if (!VT.isVector()) {
1555 return DAG.getConstant(0, VT);
1557 if (!LegalOperations || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT)) {
1558 // Produce a vector of zeros.
1559 SDValue El = DAG.getConstant(0, VT.getVectorElementType());
1560 std::vector<SDValue> Ops(VT.getVectorNumElements(), El);
1561 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT,
1562 &Ops[0], Ops.size());
1567 SDValue DAGCombiner::visitSUB(SDNode *N) {
1568 SDValue N0 = N->getOperand(0);
1569 SDValue N1 = N->getOperand(1);
1570 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1571 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1572 ConstantSDNode *N1C1 = N1.getOpcode() != ISD::ADD ? 0 :
1573 dyn_cast<ConstantSDNode>(N1.getOperand(1).getNode());
1574 EVT VT = N0.getValueType();
1577 if (VT.isVector()) {
1578 SDValue FoldedVOp = SimplifyVBinOp(N);
1579 if (FoldedVOp.getNode()) return FoldedVOp;
1582 // fold (sub x, x) -> 0
1583 // FIXME: Refactor this and xor and other similar operations together.
1585 return tryFoldToZero(N->getDebugLoc(), TLI, VT, DAG, LegalOperations);
1586 // fold (sub c1, c2) -> c1-c2
1588 return DAG.FoldConstantArithmetic(ISD::SUB, VT, N0C, N1C);
1589 // fold (sub x, c) -> (add x, -c)
1591 return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N0,
1592 DAG.getConstant(-N1C->getAPIntValue(), VT));
1593 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1)
1594 if (N0C && N0C->isAllOnesValue())
1595 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N1, N0);
1596 // fold A-(A-B) -> B
1597 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(0))
1598 return N1.getOperand(1);
1599 // fold (A+B)-A -> B
1600 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1)
1601 return N0.getOperand(1);
1602 // fold (A+B)-B -> A
1603 if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1)
1604 return N0.getOperand(0);
1605 // fold C2-(A+C1) -> (C2-C1)-A
1606 if (N1.getOpcode() == ISD::ADD && N0C && N1C1) {
1607 SDValue NewC = DAG.getConstant((N0C->getAPIntValue() - N1C1->getAPIntValue()), VT);
1608 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, NewC,
1611 // fold ((A+(B+or-C))-B) -> A+or-C
1612 if (N0.getOpcode() == ISD::ADD &&
1613 (N0.getOperand(1).getOpcode() == ISD::SUB ||
1614 N0.getOperand(1).getOpcode() == ISD::ADD) &&
1615 N0.getOperand(1).getOperand(0) == N1)
1616 return DAG.getNode(N0.getOperand(1).getOpcode(), N->getDebugLoc(), VT,
1617 N0.getOperand(0), N0.getOperand(1).getOperand(1));
1618 // fold ((A+(C+B))-B) -> A+C
1619 if (N0.getOpcode() == ISD::ADD &&
1620 N0.getOperand(1).getOpcode() == ISD::ADD &&
1621 N0.getOperand(1).getOperand(1) == N1)
1622 return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT,
1623 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1624 // fold ((A-(B-C))-C) -> A-B
1625 if (N0.getOpcode() == ISD::SUB &&
1626 N0.getOperand(1).getOpcode() == ISD::SUB &&
1627 N0.getOperand(1).getOperand(1) == N1)
1628 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1629 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1631 // If either operand of a sub is undef, the result is undef
1632 if (N0.getOpcode() == ISD::UNDEF)
1634 if (N1.getOpcode() == ISD::UNDEF)
1637 // If the relocation model supports it, consider symbol offsets.
1638 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1639 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA)) {
1640 // fold (sub Sym, c) -> Sym-c
1641 if (N1C && GA->getOpcode() == ISD::GlobalAddress)
1642 return DAG.getGlobalAddress(GA->getGlobal(), N1C->getDebugLoc(), VT,
1644 (uint64_t)N1C->getSExtValue());
1645 // fold (sub Sym+c1, Sym+c2) -> c1-c2
1646 if (GlobalAddressSDNode *GB = dyn_cast<GlobalAddressSDNode>(N1))
1647 if (GA->getGlobal() == GB->getGlobal())
1648 return DAG.getConstant((uint64_t)GA->getOffset() - GB->getOffset(),
1655 SDValue DAGCombiner::visitMUL(SDNode *N) {
1656 SDValue N0 = N->getOperand(0);
1657 SDValue N1 = N->getOperand(1);
1658 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1659 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1660 EVT VT = N0.getValueType();
1663 if (VT.isVector()) {
1664 SDValue FoldedVOp = SimplifyVBinOp(N);
1665 if (FoldedVOp.getNode()) return FoldedVOp;
1668 // fold (mul x, undef) -> 0
1669 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
1670 return DAG.getConstant(0, VT);
1671 // fold (mul c1, c2) -> c1*c2
1673 return DAG.FoldConstantArithmetic(ISD::MUL, VT, N0C, N1C);
1674 // canonicalize constant to RHS
1676 return DAG.getNode(ISD::MUL, N->getDebugLoc(), VT, N1, N0);
1677 // fold (mul x, 0) -> 0
1678 if (N1C && N1C->isNullValue())
1680 // fold (mul x, -1) -> 0-x
1681 if (N1C && N1C->isAllOnesValue())
1682 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1683 DAG.getConstant(0, VT), N0);
1684 // fold (mul x, (1 << c)) -> x << c
1685 if (N1C && N1C->getAPIntValue().isPowerOf2())
1686 return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0,
1687 DAG.getConstant(N1C->getAPIntValue().logBase2(),
1688 getShiftAmountTy(N0.getValueType())));
1689 // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c
1690 if (N1C && (-N1C->getAPIntValue()).isPowerOf2()) {
1691 unsigned Log2Val = (-N1C->getAPIntValue()).logBase2();
1692 // FIXME: If the input is something that is easily negated (e.g. a
1693 // single-use add), we should put the negate there.
1694 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1695 DAG.getConstant(0, VT),
1696 DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0,
1697 DAG.getConstant(Log2Val,
1698 getShiftAmountTy(N0.getValueType()))));
1700 // (mul (shl X, c1), c2) -> (mul X, c2 << c1)
1701 if (N1C && N0.getOpcode() == ISD::SHL &&
1702 isa<ConstantSDNode>(N0.getOperand(1))) {
1703 SDValue C3 = DAG.getNode(ISD::SHL, N->getDebugLoc(), VT,
1704 N1, N0.getOperand(1));
1705 AddToWorkList(C3.getNode());
1706 return DAG.getNode(ISD::MUL, N->getDebugLoc(), VT,
1707 N0.getOperand(0), C3);
1710 // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one
1713 SDValue Sh(0,0), Y(0,0);
1714 // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)).
1715 if (N0.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N0.getOperand(1)) &&
1716 N0.getNode()->hasOneUse()) {
1718 } else if (N1.getOpcode() == ISD::SHL &&
1719 isa<ConstantSDNode>(N1.getOperand(1)) &&
1720 N1.getNode()->hasOneUse()) {
1725 SDValue Mul = DAG.getNode(ISD::MUL, N->getDebugLoc(), VT,
1726 Sh.getOperand(0), Y);
1727 return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT,
1728 Mul, Sh.getOperand(1));
1732 // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2)
1733 if (N1C && N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() &&
1734 isa<ConstantSDNode>(N0.getOperand(1)))
1735 return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT,
1736 DAG.getNode(ISD::MUL, N0.getDebugLoc(), VT,
1737 N0.getOperand(0), N1),
1738 DAG.getNode(ISD::MUL, N1.getDebugLoc(), VT,
1739 N0.getOperand(1), N1));
1742 SDValue RMUL = ReassociateOps(ISD::MUL, N->getDebugLoc(), N0, N1);
1743 if (RMUL.getNode() != 0)
1749 SDValue DAGCombiner::visitSDIV(SDNode *N) {
1750 SDValue N0 = N->getOperand(0);
1751 SDValue N1 = N->getOperand(1);
1752 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1753 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1754 EVT VT = N->getValueType(0);
1757 if (VT.isVector()) {
1758 SDValue FoldedVOp = SimplifyVBinOp(N);
1759 if (FoldedVOp.getNode()) return FoldedVOp;
1762 // fold (sdiv c1, c2) -> c1/c2
1763 if (N0C && N1C && !N1C->isNullValue())
1764 return DAG.FoldConstantArithmetic(ISD::SDIV, VT, N0C, N1C);
1765 // fold (sdiv X, 1) -> X
1766 if (N1C && N1C->getAPIntValue() == 1LL)
1768 // fold (sdiv X, -1) -> 0-X
1769 if (N1C && N1C->isAllOnesValue())
1770 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1771 DAG.getConstant(0, VT), N0);
1772 // If we know the sign bits of both operands are zero, strength reduce to a
1773 // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2
1774 if (!VT.isVector()) {
1775 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
1776 return DAG.getNode(ISD::UDIV, N->getDebugLoc(), N1.getValueType(),
1779 // fold (sdiv X, pow2) -> simple ops after legalize
1780 if (N1C && !N1C->isNullValue() &&
1781 (N1C->getAPIntValue().isPowerOf2() ||
1782 (-N1C->getAPIntValue()).isPowerOf2())) {
1783 // If dividing by powers of two is cheap, then don't perform the following
1785 if (TLI.isPow2DivCheap())
1788 unsigned lg2 = N1C->getAPIntValue().countTrailingZeros();
1790 // Splat the sign bit into the register
1791 SDValue SGN = DAG.getNode(ISD::SRA, N->getDebugLoc(), VT, N0,
1792 DAG.getConstant(VT.getSizeInBits()-1,
1793 getShiftAmountTy(N0.getValueType())));
1794 AddToWorkList(SGN.getNode());
1796 // Add (N0 < 0) ? abs2 - 1 : 0;
1797 SDValue SRL = DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, SGN,
1798 DAG.getConstant(VT.getSizeInBits() - lg2,
1799 getShiftAmountTy(SGN.getValueType())));
1800 SDValue ADD = DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N0, SRL);
1801 AddToWorkList(SRL.getNode());
1802 AddToWorkList(ADD.getNode()); // Divide by pow2
1803 SDValue SRA = DAG.getNode(ISD::SRA, N->getDebugLoc(), VT, ADD,
1804 DAG.getConstant(lg2, getShiftAmountTy(ADD.getValueType())));
1806 // If we're dividing by a positive value, we're done. Otherwise, we must
1807 // negate the result.
1808 if (N1C->getAPIntValue().isNonNegative())
1811 AddToWorkList(SRA.getNode());
1812 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1813 DAG.getConstant(0, VT), SRA);
1816 // if integer divide is expensive and we satisfy the requirements, emit an
1817 // alternate sequence.
1818 if (N1C && !N1C->isNullValue() && !TLI.isIntDivCheap()) {
1819 SDValue Op = BuildSDIV(N);
1820 if (Op.getNode()) return Op;
1824 if (N0.getOpcode() == ISD::UNDEF)
1825 return DAG.getConstant(0, VT);
1826 // X / undef -> undef
1827 if (N1.getOpcode() == ISD::UNDEF)
1833 SDValue DAGCombiner::visitUDIV(SDNode *N) {
1834 SDValue N0 = N->getOperand(0);
1835 SDValue N1 = N->getOperand(1);
1836 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1837 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1838 EVT VT = N->getValueType(0);
1841 if (VT.isVector()) {
1842 SDValue FoldedVOp = SimplifyVBinOp(N);
1843 if (FoldedVOp.getNode()) return FoldedVOp;
1846 // fold (udiv c1, c2) -> c1/c2
1847 if (N0C && N1C && !N1C->isNullValue())
1848 return DAG.FoldConstantArithmetic(ISD::UDIV, VT, N0C, N1C);
1849 // fold (udiv x, (1 << c)) -> x >>u c
1850 if (N1C && N1C->getAPIntValue().isPowerOf2())
1851 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0,
1852 DAG.getConstant(N1C->getAPIntValue().logBase2(),
1853 getShiftAmountTy(N0.getValueType())));
1854 // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2
1855 if (N1.getOpcode() == ISD::SHL) {
1856 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
1857 if (SHC->getAPIntValue().isPowerOf2()) {
1858 EVT ADDVT = N1.getOperand(1).getValueType();
1859 SDValue Add = DAG.getNode(ISD::ADD, N->getDebugLoc(), ADDVT,
1861 DAG.getConstant(SHC->getAPIntValue()
1864 AddToWorkList(Add.getNode());
1865 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0, Add);
1869 // fold (udiv x, c) -> alternate
1870 if (N1C && !N1C->isNullValue() && !TLI.isIntDivCheap()) {
1871 SDValue Op = BuildUDIV(N);
1872 if (Op.getNode()) return Op;
1876 if (N0.getOpcode() == ISD::UNDEF)
1877 return DAG.getConstant(0, VT);
1878 // X / undef -> undef
1879 if (N1.getOpcode() == ISD::UNDEF)
1885 SDValue DAGCombiner::visitSREM(SDNode *N) {
1886 SDValue N0 = N->getOperand(0);
1887 SDValue N1 = N->getOperand(1);
1888 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1889 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1890 EVT VT = N->getValueType(0);
1892 // fold (srem c1, c2) -> c1%c2
1893 if (N0C && N1C && !N1C->isNullValue())
1894 return DAG.FoldConstantArithmetic(ISD::SREM, VT, N0C, N1C);
1895 // If we know the sign bits of both operands are zero, strength reduce to a
1896 // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15
1897 if (!VT.isVector()) {
1898 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
1899 return DAG.getNode(ISD::UREM, N->getDebugLoc(), VT, N0, N1);
1902 // If X/C can be simplified by the division-by-constant logic, lower
1903 // X%C to the equivalent of X-X/C*C.
1904 if (N1C && !N1C->isNullValue()) {
1905 SDValue Div = DAG.getNode(ISD::SDIV, N->getDebugLoc(), VT, N0, N1);
1906 AddToWorkList(Div.getNode());
1907 SDValue OptimizedDiv = combine(Div.getNode());
1908 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
1909 SDValue Mul = DAG.getNode(ISD::MUL, N->getDebugLoc(), VT,
1911 SDValue Sub = DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N0, Mul);
1912 AddToWorkList(Mul.getNode());
1918 if (N0.getOpcode() == ISD::UNDEF)
1919 return DAG.getConstant(0, VT);
1920 // X % undef -> undef
1921 if (N1.getOpcode() == ISD::UNDEF)
1927 SDValue DAGCombiner::visitUREM(SDNode *N) {
1928 SDValue N0 = N->getOperand(0);
1929 SDValue N1 = N->getOperand(1);
1930 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1931 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1932 EVT VT = N->getValueType(0);
1934 // fold (urem c1, c2) -> c1%c2
1935 if (N0C && N1C && !N1C->isNullValue())
1936 return DAG.FoldConstantArithmetic(ISD::UREM, VT, N0C, N1C);
1937 // fold (urem x, pow2) -> (and x, pow2-1)
1938 if (N1C && !N1C->isNullValue() && N1C->getAPIntValue().isPowerOf2())
1939 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0,
1940 DAG.getConstant(N1C->getAPIntValue()-1,VT));
1941 // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1))
1942 if (N1.getOpcode() == ISD::SHL) {
1943 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
1944 if (SHC->getAPIntValue().isPowerOf2()) {
1946 DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N1,
1947 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()),
1949 AddToWorkList(Add.getNode());
1950 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0, Add);
1955 // If X/C can be simplified by the division-by-constant logic, lower
1956 // X%C to the equivalent of X-X/C*C.
1957 if (N1C && !N1C->isNullValue()) {
1958 SDValue Div = DAG.getNode(ISD::UDIV, N->getDebugLoc(), VT, N0, N1);
1959 AddToWorkList(Div.getNode());
1960 SDValue OptimizedDiv = combine(Div.getNode());
1961 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
1962 SDValue Mul = DAG.getNode(ISD::MUL, N->getDebugLoc(), VT,
1964 SDValue Sub = DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N0, Mul);
1965 AddToWorkList(Mul.getNode());
1971 if (N0.getOpcode() == ISD::UNDEF)
1972 return DAG.getConstant(0, VT);
1973 // X % undef -> undef
1974 if (N1.getOpcode() == ISD::UNDEF)
1980 SDValue DAGCombiner::visitMULHS(SDNode *N) {
1981 SDValue N0 = N->getOperand(0);
1982 SDValue N1 = N->getOperand(1);
1983 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1984 EVT VT = N->getValueType(0);
1985 DebugLoc DL = N->getDebugLoc();
1987 // fold (mulhs x, 0) -> 0
1988 if (N1C && N1C->isNullValue())
1990 // fold (mulhs x, 1) -> (sra x, size(x)-1)
1991 if (N1C && N1C->getAPIntValue() == 1)
1992 return DAG.getNode(ISD::SRA, N->getDebugLoc(), N0.getValueType(), N0,
1993 DAG.getConstant(N0.getValueType().getSizeInBits() - 1,
1994 getShiftAmountTy(N0.getValueType())));
1995 // fold (mulhs x, undef) -> 0
1996 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
1997 return DAG.getConstant(0, VT);
1999 // If the type twice as wide is legal, transform the mulhs to a wider multiply
2001 if (VT.isSimple() && !VT.isVector()) {
2002 MVT Simple = VT.getSimpleVT();
2003 unsigned SimpleSize = Simple.getSizeInBits();
2004 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2005 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2006 N0 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N0);
2007 N1 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N1);
2008 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2009 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2010 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2011 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2018 SDValue DAGCombiner::visitMULHU(SDNode *N) {
2019 SDValue N0 = N->getOperand(0);
2020 SDValue N1 = N->getOperand(1);
2021 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2022 EVT VT = N->getValueType(0);
2023 DebugLoc DL = N->getDebugLoc();
2025 // fold (mulhu x, 0) -> 0
2026 if (N1C && N1C->isNullValue())
2028 // fold (mulhu x, 1) -> 0
2029 if (N1C && N1C->getAPIntValue() == 1)
2030 return DAG.getConstant(0, N0.getValueType());
2031 // fold (mulhu x, undef) -> 0
2032 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2033 return DAG.getConstant(0, VT);
2035 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2037 if (VT.isSimple() && !VT.isVector()) {
2038 MVT Simple = VT.getSimpleVT();
2039 unsigned SimpleSize = Simple.getSizeInBits();
2040 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2041 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2042 N0 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N0);
2043 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N1);
2044 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2045 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2046 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2047 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2054 /// SimplifyNodeWithTwoResults - Perform optimizations common to nodes that
2055 /// compute two values. LoOp and HiOp give the opcodes for the two computations
2056 /// that are being performed. Return true if a simplification was made.
2058 SDValue DAGCombiner::SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
2060 // If the high half is not needed, just compute the low half.
2061 bool HiExists = N->hasAnyUseOfValue(1);
2063 (!LegalOperations ||
2064 TLI.isOperationLegal(LoOp, N->getValueType(0)))) {
2065 SDValue Res = DAG.getNode(LoOp, N->getDebugLoc(), N->getValueType(0),
2066 N->op_begin(), N->getNumOperands());
2067 return CombineTo(N, Res, Res);
2070 // If the low half is not needed, just compute the high half.
2071 bool LoExists = N->hasAnyUseOfValue(0);
2073 (!LegalOperations ||
2074 TLI.isOperationLegal(HiOp, N->getValueType(1)))) {
2075 SDValue Res = DAG.getNode(HiOp, N->getDebugLoc(), N->getValueType(1),
2076 N->op_begin(), N->getNumOperands());
2077 return CombineTo(N, Res, Res);
2080 // If both halves are used, return as it is.
2081 if (LoExists && HiExists)
2084 // If the two computed results can be simplified separately, separate them.
2086 SDValue Lo = DAG.getNode(LoOp, N->getDebugLoc(), N->getValueType(0),
2087 N->op_begin(), N->getNumOperands());
2088 AddToWorkList(Lo.getNode());
2089 SDValue LoOpt = combine(Lo.getNode());
2090 if (LoOpt.getNode() && LoOpt.getNode() != Lo.getNode() &&
2091 (!LegalOperations ||
2092 TLI.isOperationLegal(LoOpt.getOpcode(), LoOpt.getValueType())))
2093 return CombineTo(N, LoOpt, LoOpt);
2097 SDValue Hi = DAG.getNode(HiOp, N->getDebugLoc(), N->getValueType(1),
2098 N->op_begin(), N->getNumOperands());
2099 AddToWorkList(Hi.getNode());
2100 SDValue HiOpt = combine(Hi.getNode());
2101 if (HiOpt.getNode() && HiOpt != Hi &&
2102 (!LegalOperations ||
2103 TLI.isOperationLegal(HiOpt.getOpcode(), HiOpt.getValueType())))
2104 return CombineTo(N, HiOpt, HiOpt);
2110 SDValue DAGCombiner::visitSMUL_LOHI(SDNode *N) {
2111 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHS);
2112 if (Res.getNode()) return Res;
2114 EVT VT = N->getValueType(0);
2115 DebugLoc DL = N->getDebugLoc();
2117 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2119 if (VT.isSimple() && !VT.isVector()) {
2120 MVT Simple = VT.getSimpleVT();
2121 unsigned SimpleSize = Simple.getSizeInBits();
2122 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2123 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2124 SDValue Lo = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(0));
2125 SDValue Hi = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(1));
2126 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2127 // Compute the high part as N1.
2128 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2129 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2130 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2131 // Compute the low part as N0.
2132 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2133 return CombineTo(N, Lo, Hi);
2140 SDValue DAGCombiner::visitUMUL_LOHI(SDNode *N) {
2141 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHU);
2142 if (Res.getNode()) return Res;
2144 EVT VT = N->getValueType(0);
2145 DebugLoc DL = N->getDebugLoc();
2147 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2149 if (VT.isSimple() && !VT.isVector()) {
2150 MVT Simple = VT.getSimpleVT();
2151 unsigned SimpleSize = Simple.getSizeInBits();
2152 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2153 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2154 SDValue Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(0));
2155 SDValue Hi = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(1));
2156 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2157 // Compute the high part as N1.
2158 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2159 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2160 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2161 // Compute the low part as N0.
2162 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2163 return CombineTo(N, Lo, Hi);
2170 SDValue DAGCombiner::visitSMULO(SDNode *N) {
2171 // (smulo x, 2) -> (saddo x, x)
2172 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2173 if (C2->getAPIntValue() == 2)
2174 return DAG.getNode(ISD::SADDO, N->getDebugLoc(), N->getVTList(),
2175 N->getOperand(0), N->getOperand(0));
2180 SDValue DAGCombiner::visitUMULO(SDNode *N) {
2181 // (umulo x, 2) -> (uaddo x, x)
2182 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2183 if (C2->getAPIntValue() == 2)
2184 return DAG.getNode(ISD::UADDO, N->getDebugLoc(), N->getVTList(),
2185 N->getOperand(0), N->getOperand(0));
2190 SDValue DAGCombiner::visitSDIVREM(SDNode *N) {
2191 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::SDIV, ISD::SREM);
2192 if (Res.getNode()) return Res;
2197 SDValue DAGCombiner::visitUDIVREM(SDNode *N) {
2198 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::UDIV, ISD::UREM);
2199 if (Res.getNode()) return Res;
2204 /// SimplifyBinOpWithSameOpcodeHands - If this is a binary operator with
2205 /// two operands of the same opcode, try to simplify it.
2206 SDValue DAGCombiner::SimplifyBinOpWithSameOpcodeHands(SDNode *N) {
2207 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
2208 EVT VT = N0.getValueType();
2209 assert(N0.getOpcode() == N1.getOpcode() && "Bad input!");
2211 // Bail early if none of these transforms apply.
2212 if (N0.getNode()->getNumOperands() == 0) return SDValue();
2214 // For each of OP in AND/OR/XOR:
2215 // fold (OP (zext x), (zext y)) -> (zext (OP x, y))
2216 // fold (OP (sext x), (sext y)) -> (sext (OP x, y))
2217 // fold (OP (aext x), (aext y)) -> (aext (OP x, y))
2218 // fold (OP (trunc x), (trunc y)) -> (trunc (OP x, y)) (if trunc isn't free)
2220 // do not sink logical op inside of a vector extend, since it may combine
2222 EVT Op0VT = N0.getOperand(0).getValueType();
2223 if ((N0.getOpcode() == ISD::ZERO_EXTEND ||
2224 N0.getOpcode() == ISD::SIGN_EXTEND ||
2225 // Avoid infinite looping with PromoteIntBinOp.
2226 (N0.getOpcode() == ISD::ANY_EXTEND &&
2227 (!LegalTypes || TLI.isTypeDesirableForOp(N->getOpcode(), Op0VT))) ||
2228 (N0.getOpcode() == ISD::TRUNCATE &&
2229 (!TLI.isZExtFree(VT, Op0VT) ||
2230 !TLI.isTruncateFree(Op0VT, VT)) &&
2231 TLI.isTypeLegal(Op0VT))) &&
2233 Op0VT == N1.getOperand(0).getValueType() &&
2234 (!LegalOperations || TLI.isOperationLegal(N->getOpcode(), Op0VT))) {
2235 SDValue ORNode = DAG.getNode(N->getOpcode(), N0.getDebugLoc(),
2236 N0.getOperand(0).getValueType(),
2237 N0.getOperand(0), N1.getOperand(0));
2238 AddToWorkList(ORNode.getNode());
2239 return DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT, ORNode);
2242 // For each of OP in SHL/SRL/SRA/AND...
2243 // fold (and (OP x, z), (OP y, z)) -> (OP (and x, y), z)
2244 // fold (or (OP x, z), (OP y, z)) -> (OP (or x, y), z)
2245 // fold (xor (OP x, z), (OP y, z)) -> (OP (xor x, y), z)
2246 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL ||
2247 N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::AND) &&
2248 N0.getOperand(1) == N1.getOperand(1)) {
2249 SDValue ORNode = DAG.getNode(N->getOpcode(), N0.getDebugLoc(),
2250 N0.getOperand(0).getValueType(),
2251 N0.getOperand(0), N1.getOperand(0));
2252 AddToWorkList(ORNode.getNode());
2253 return DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT,
2254 ORNode, N0.getOperand(1));
2260 SDValue DAGCombiner::visitAND(SDNode *N) {
2261 SDValue N0 = N->getOperand(0);
2262 SDValue N1 = N->getOperand(1);
2263 SDValue LL, LR, RL, RR, CC0, CC1;
2264 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2265 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2266 EVT VT = N1.getValueType();
2267 unsigned BitWidth = VT.getScalarType().getSizeInBits();
2270 if (VT.isVector()) {
2271 SDValue FoldedVOp = SimplifyVBinOp(N);
2272 if (FoldedVOp.getNode()) return FoldedVOp;
2275 // fold (and x, undef) -> 0
2276 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2277 return DAG.getConstant(0, VT);
2278 // fold (and c1, c2) -> c1&c2
2280 return DAG.FoldConstantArithmetic(ISD::AND, VT, N0C, N1C);
2281 // canonicalize constant to RHS
2283 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N1, N0);
2284 // fold (and x, -1) -> x
2285 if (N1C && N1C->isAllOnesValue())
2287 // if (and x, c) is known to be zero, return 0
2288 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
2289 APInt::getAllOnesValue(BitWidth)))
2290 return DAG.getConstant(0, VT);
2292 SDValue RAND = ReassociateOps(ISD::AND, N->getDebugLoc(), N0, N1);
2293 if (RAND.getNode() != 0)
2295 // fold (and (or x, C), D) -> D if (C & D) == D
2296 if (N1C && N0.getOpcode() == ISD::OR)
2297 if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
2298 if ((ORI->getAPIntValue() & N1C->getAPIntValue()) == N1C->getAPIntValue())
2300 // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits.
2301 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
2302 SDValue N0Op0 = N0.getOperand(0);
2303 APInt Mask = ~N1C->getAPIntValue();
2304 Mask = Mask.trunc(N0Op0.getValueSizeInBits());
2305 if (DAG.MaskedValueIsZero(N0Op0, Mask)) {
2306 SDValue Zext = DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(),
2307 N0.getValueType(), N0Op0);
2309 // Replace uses of the AND with uses of the Zero extend node.
2312 // We actually want to replace all uses of the any_extend with the
2313 // zero_extend, to avoid duplicating things. This will later cause this
2314 // AND to be folded.
2315 CombineTo(N0.getNode(), Zext);
2316 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2319 // fold (and (setcc x), (setcc y)) -> (setcc (and x, y))
2320 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
2321 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
2322 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
2324 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
2325 LL.getValueType().isInteger()) {
2326 // fold (and (seteq X, 0), (seteq Y, 0)) -> (seteq (or X, Y), 0)
2327 if (cast<ConstantSDNode>(LR)->isNullValue() && Op1 == ISD::SETEQ) {
2328 SDValue ORNode = DAG.getNode(ISD::OR, N0.getDebugLoc(),
2329 LR.getValueType(), LL, RL);
2330 AddToWorkList(ORNode.getNode());
2331 return DAG.getSetCC(N->getDebugLoc(), VT, ORNode, LR, Op1);
2333 // fold (and (seteq X, -1), (seteq Y, -1)) -> (seteq (and X, Y), -1)
2334 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETEQ) {
2335 SDValue ANDNode = DAG.getNode(ISD::AND, N0.getDebugLoc(),
2336 LR.getValueType(), LL, RL);
2337 AddToWorkList(ANDNode.getNode());
2338 return DAG.getSetCC(N->getDebugLoc(), VT, ANDNode, LR, Op1);
2340 // fold (and (setgt X, -1), (setgt Y, -1)) -> (setgt (or X, Y), -1)
2341 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETGT) {
2342 SDValue ORNode = DAG.getNode(ISD::OR, N0.getDebugLoc(),
2343 LR.getValueType(), LL, RL);
2344 AddToWorkList(ORNode.getNode());
2345 return DAG.getSetCC(N->getDebugLoc(), VT, ORNode, LR, Op1);
2348 // canonicalize equivalent to ll == rl
2349 if (LL == RR && LR == RL) {
2350 Op1 = ISD::getSetCCSwappedOperands(Op1);
2353 if (LL == RL && LR == RR) {
2354 bool isInteger = LL.getValueType().isInteger();
2355 ISD::CondCode Result = ISD::getSetCCAndOperation(Op0, Op1, isInteger);
2356 if (Result != ISD::SETCC_INVALID &&
2357 (!LegalOperations || TLI.isCondCodeLegal(Result, LL.getValueType())))
2358 return DAG.getSetCC(N->getDebugLoc(), N0.getValueType(),
2363 // Simplify: (and (op x...), (op y...)) -> (op (and x, y))
2364 if (N0.getOpcode() == N1.getOpcode()) {
2365 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
2366 if (Tmp.getNode()) return Tmp;
2369 // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1)
2370 // fold (and (sra)) -> (and (srl)) when possible.
2371 if (!VT.isVector() &&
2372 SimplifyDemandedBits(SDValue(N, 0)))
2373 return SDValue(N, 0);
2375 // fold (zext_inreg (extload x)) -> (zextload x)
2376 if (ISD::isEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode())) {
2377 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2378 EVT MemVT = LN0->getMemoryVT();
2379 // If we zero all the possible extended bits, then we can turn this into
2380 // a zextload if we are running before legalize or the operation is legal.
2381 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
2382 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
2383 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
2384 ((!LegalOperations && !LN0->isVolatile()) ||
2385 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
2386 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N0.getDebugLoc(), VT,
2387 LN0->getChain(), LN0->getBasePtr(),
2388 LN0->getPointerInfo(), MemVT,
2389 LN0->isVolatile(), LN0->isNonTemporal(),
2390 LN0->getAlignment());
2392 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
2393 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2396 // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use
2397 if (ISD::isSEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
2399 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2400 EVT MemVT = LN0->getMemoryVT();
2401 // If we zero all the possible extended bits, then we can turn this into
2402 // a zextload if we are running before legalize or the operation is legal.
2403 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
2404 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
2405 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
2406 ((!LegalOperations && !LN0->isVolatile()) ||
2407 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
2408 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N0.getDebugLoc(), VT,
2410 LN0->getBasePtr(), LN0->getPointerInfo(),
2412 LN0->isVolatile(), LN0->isNonTemporal(),
2413 LN0->getAlignment());
2415 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
2416 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2420 // fold (and (load x), 255) -> (zextload x, i8)
2421 // fold (and (extload x, i16), 255) -> (zextload x, i8)
2422 // fold (and (any_ext (extload x, i16)), 255) -> (zextload x, i8)
2423 if (N1C && (N0.getOpcode() == ISD::LOAD ||
2424 (N0.getOpcode() == ISD::ANY_EXTEND &&
2425 N0.getOperand(0).getOpcode() == ISD::LOAD))) {
2426 bool HasAnyExt = N0.getOpcode() == ISD::ANY_EXTEND;
2427 LoadSDNode *LN0 = HasAnyExt
2428 ? cast<LoadSDNode>(N0.getOperand(0))
2429 : cast<LoadSDNode>(N0);
2430 if (LN0->getExtensionType() != ISD::SEXTLOAD &&
2431 LN0->isUnindexed() && N0.hasOneUse() && LN0->hasOneUse()) {
2432 uint32_t ActiveBits = N1C->getAPIntValue().getActiveBits();
2433 if (ActiveBits > 0 && APIntOps::isMask(ActiveBits, N1C->getAPIntValue())){
2434 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits);
2435 EVT LoadedVT = LN0->getMemoryVT();
2437 if (ExtVT == LoadedVT &&
2438 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) {
2439 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
2442 DAG.getExtLoad(ISD::ZEXTLOAD, LN0->getDebugLoc(), LoadResultTy,
2443 LN0->getChain(), LN0->getBasePtr(),
2444 LN0->getPointerInfo(),
2445 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
2446 LN0->getAlignment());
2448 CombineTo(LN0, NewLoad, NewLoad.getValue(1));
2449 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2452 // Do not change the width of a volatile load.
2453 // Do not generate loads of non-round integer types since these can
2454 // be expensive (and would be wrong if the type is not byte sized).
2455 if (!LN0->isVolatile() && LoadedVT.bitsGT(ExtVT) && ExtVT.isRound() &&
2456 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) {
2457 EVT PtrType = LN0->getOperand(1).getValueType();
2459 unsigned Alignment = LN0->getAlignment();
2460 SDValue NewPtr = LN0->getBasePtr();
2462 // For big endian targets, we need to add an offset to the pointer
2463 // to load the correct bytes. For little endian systems, we merely
2464 // need to read fewer bytes from the same pointer.
2465 if (TLI.isBigEndian()) {
2466 unsigned LVTStoreBytes = LoadedVT.getStoreSize();
2467 unsigned EVTStoreBytes = ExtVT.getStoreSize();
2468 unsigned PtrOff = LVTStoreBytes - EVTStoreBytes;
2469 NewPtr = DAG.getNode(ISD::ADD, LN0->getDebugLoc(), PtrType,
2470 NewPtr, DAG.getConstant(PtrOff, PtrType));
2471 Alignment = MinAlign(Alignment, PtrOff);
2474 AddToWorkList(NewPtr.getNode());
2476 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
2478 DAG.getExtLoad(ISD::ZEXTLOAD, LN0->getDebugLoc(), LoadResultTy,
2479 LN0->getChain(), NewPtr,
2480 LN0->getPointerInfo(),
2481 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
2484 CombineTo(LN0, Load, Load.getValue(1));
2485 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2494 /// MatchBSwapHWord - Match (a >> 8) | (a << 8) as (bswap a) >> 16
2496 SDValue DAGCombiner::MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
2497 bool DemandHighBits) {
2498 if (!LegalOperations)
2501 EVT VT = N->getValueType(0);
2502 if (VT != MVT::i64 && VT != MVT::i32 && VT != MVT::i16)
2504 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
2507 // Recognize (and (shl a, 8), 0xff), (and (srl a, 8), 0xff00)
2508 bool LookPassAnd0 = false;
2509 bool LookPassAnd1 = false;
2510 if (N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode() == ISD::SRL)
2512 if (N1.getOpcode() == ISD::AND && N1.getOperand(0).getOpcode() == ISD::SHL)
2514 if (N0.getOpcode() == ISD::AND) {
2515 if (!N0.getNode()->hasOneUse())
2517 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
2518 if (!N01C || N01C->getZExtValue() != 0xFF00)
2520 N0 = N0.getOperand(0);
2521 LookPassAnd0 = true;
2524 if (N1.getOpcode() == ISD::AND) {
2525 if (!N1.getNode()->hasOneUse())
2527 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
2528 if (!N11C || N11C->getZExtValue() != 0xFF)
2530 N1 = N1.getOperand(0);
2531 LookPassAnd1 = true;
2534 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
2536 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
2538 if (!N0.getNode()->hasOneUse() ||
2539 !N1.getNode()->hasOneUse())
2542 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
2543 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
2546 if (N01C->getZExtValue() != 8 || N11C->getZExtValue() != 8)
2549 // Look for (shl (and a, 0xff), 8), (srl (and a, 0xff00), 8)
2550 SDValue N00 = N0->getOperand(0);
2551 if (!LookPassAnd0 && N00.getOpcode() == ISD::AND) {
2552 if (!N00.getNode()->hasOneUse())
2554 ConstantSDNode *N001C = dyn_cast<ConstantSDNode>(N00.getOperand(1));
2555 if (!N001C || N001C->getZExtValue() != 0xFF)
2557 N00 = N00.getOperand(0);
2558 LookPassAnd0 = true;
2561 SDValue N10 = N1->getOperand(0);
2562 if (!LookPassAnd1 && N10.getOpcode() == ISD::AND) {
2563 if (!N10.getNode()->hasOneUse())
2565 ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N10.getOperand(1));
2566 if (!N101C || N101C->getZExtValue() != 0xFF00)
2568 N10 = N10.getOperand(0);
2569 LookPassAnd1 = true;
2575 // Make sure everything beyond the low halfword is zero since the SRL 16
2576 // will clear the top bits.
2577 unsigned OpSizeInBits = VT.getSizeInBits();
2578 if (DemandHighBits && OpSizeInBits > 16 &&
2579 (!LookPassAnd0 || !LookPassAnd1) &&
2580 !DAG.MaskedValueIsZero(N10, APInt::getHighBitsSet(OpSizeInBits, 16)))
2583 SDValue Res = DAG.getNode(ISD::BSWAP, N->getDebugLoc(), VT, N00);
2584 if (OpSizeInBits > 16)
2585 Res = DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, Res,
2586 DAG.getConstant(OpSizeInBits-16, getShiftAmountTy(VT)));
2590 /// isBSwapHWordElement - Return true if the specified node is an element
2591 /// that makes up a 32-bit packed halfword byteswap. i.e.
2592 /// ((x&0xff)<<8)|((x&0xff00)>>8)|((x&0x00ff0000)<<8)|((x&0xff000000)>>8)
2593 static bool isBSwapHWordElement(SDValue N, SmallVector<SDNode*,4> &Parts) {
2594 if (!N.getNode()->hasOneUse())
2597 unsigned Opc = N.getOpcode();
2598 if (Opc != ISD::AND && Opc != ISD::SHL && Opc != ISD::SRL)
2601 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N.getOperand(1));
2606 switch (N1C->getZExtValue()) {
2609 case 0xFF: Num = 0; break;
2610 case 0xFF00: Num = 1; break;
2611 case 0xFF0000: Num = 2; break;
2612 case 0xFF000000: Num = 3; break;
2615 // Look for (x & 0xff) << 8 as well as ((x << 8) & 0xff00).
2616 SDValue N0 = N.getOperand(0);
2617 if (Opc == ISD::AND) {
2618 if (Num == 0 || Num == 2) {
2620 // (x >> 8) & 0xff0000
2621 if (N0.getOpcode() != ISD::SRL)
2623 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
2624 if (!C || C->getZExtValue() != 8)
2627 // (x << 8) & 0xff00
2628 // (x << 8) & 0xff000000
2629 if (N0.getOpcode() != ISD::SHL)
2631 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
2632 if (!C || C->getZExtValue() != 8)
2635 } else if (Opc == ISD::SHL) {
2637 // (x & 0xff0000) << 8
2638 if (Num != 0 && Num != 2)
2640 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
2641 if (!C || C->getZExtValue() != 8)
2643 } else { // Opc == ISD::SRL
2644 // (x & 0xff00) >> 8
2645 // (x & 0xff000000) >> 8
2646 if (Num != 1 && Num != 3)
2648 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
2649 if (!C || C->getZExtValue() != 8)
2656 Parts[Num] = N0.getOperand(0).getNode();
2660 /// MatchBSwapHWord - Match a 32-bit packed halfword bswap. That is
2661 /// ((x&0xff)<<8)|((x&0xff00)>>8)|((x&0x00ff0000)<<8)|((x&0xff000000)>>8)
2662 /// => (rotl (bswap x), 16)
2663 SDValue DAGCombiner::MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1) {
2664 if (!LegalOperations)
2667 EVT VT = N->getValueType(0);
2670 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
2673 SmallVector<SDNode*,4> Parts(4, (SDNode*)0);
2675 // (or (or (and), (and)), (or (and), (and)))
2676 // (or (or (or (and), (and)), (and)), (and))
2677 if (N0.getOpcode() != ISD::OR)
2679 SDValue N00 = N0.getOperand(0);
2680 SDValue N01 = N0.getOperand(1);
2682 if (N1.getOpcode() == ISD::OR) {
2683 // (or (or (and), (and)), (or (and), (and)))
2684 SDValue N000 = N00.getOperand(0);
2685 if (!isBSwapHWordElement(N000, Parts))
2688 SDValue N001 = N00.getOperand(1);
2689 if (!isBSwapHWordElement(N001, Parts))
2691 SDValue N010 = N01.getOperand(0);
2692 if (!isBSwapHWordElement(N010, Parts))
2694 SDValue N011 = N01.getOperand(1);
2695 if (!isBSwapHWordElement(N011, Parts))
2698 // (or (or (or (and), (and)), (and)), (and))
2699 if (!isBSwapHWordElement(N1, Parts))
2701 if (!isBSwapHWordElement(N01, Parts))
2703 if (N00.getOpcode() != ISD::OR)
2705 SDValue N000 = N00.getOperand(0);
2706 if (!isBSwapHWordElement(N000, Parts))
2708 SDValue N001 = N00.getOperand(1);
2709 if (!isBSwapHWordElement(N001, Parts))
2713 // Make sure the parts are all coming from the same node.
2714 if (Parts[0] != Parts[1] || Parts[0] != Parts[2] || Parts[0] != Parts[3])
2717 SDValue BSwap = DAG.getNode(ISD::BSWAP, N->getDebugLoc(), VT,
2718 SDValue(Parts[0],0));
2720 // Result of the bswap should be rotated by 16. If it's not legal, than
2721 // do (x << 16) | (x >> 16).
2722 SDValue ShAmt = DAG.getConstant(16, getShiftAmountTy(VT));
2723 if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT))
2724 return DAG.getNode(ISD::ROTL, N->getDebugLoc(), VT, BSwap, ShAmt);
2725 else if (TLI.isOperationLegalOrCustom(ISD::ROTR, VT))
2726 return DAG.getNode(ISD::ROTR, N->getDebugLoc(), VT, BSwap, ShAmt);
2727 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT,
2728 DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, BSwap, ShAmt),
2729 DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, BSwap, ShAmt));
2732 SDValue DAGCombiner::visitOR(SDNode *N) {
2733 SDValue N0 = N->getOperand(0);
2734 SDValue N1 = N->getOperand(1);
2735 SDValue LL, LR, RL, RR, CC0, CC1;
2736 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2737 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2738 EVT VT = N1.getValueType();
2741 if (VT.isVector()) {
2742 SDValue FoldedVOp = SimplifyVBinOp(N);
2743 if (FoldedVOp.getNode()) return FoldedVOp;
2746 // fold (or x, undef) -> -1
2747 if (!LegalOperations &&
2748 (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)) {
2749 EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT;
2750 return DAG.getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()), VT);
2752 // fold (or c1, c2) -> c1|c2
2754 return DAG.FoldConstantArithmetic(ISD::OR, VT, N0C, N1C);
2755 // canonicalize constant to RHS
2757 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N1, N0);
2758 // fold (or x, 0) -> x
2759 if (N1C && N1C->isNullValue())
2761 // fold (or x, -1) -> -1
2762 if (N1C && N1C->isAllOnesValue())
2764 // fold (or x, c) -> c iff (x & ~c) == 0
2765 if (N1C && DAG.MaskedValueIsZero(N0, ~N1C->getAPIntValue()))
2768 // Recognize halfword bswaps as (bswap + rotl 16) or (bswap + shl 16)
2769 SDValue BSwap = MatchBSwapHWord(N, N0, N1);
2770 if (BSwap.getNode() != 0)
2772 BSwap = MatchBSwapHWordLow(N, N0, N1);
2773 if (BSwap.getNode() != 0)
2777 SDValue ROR = ReassociateOps(ISD::OR, N->getDebugLoc(), N0, N1);
2778 if (ROR.getNode() != 0)
2780 // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2)
2781 // iff (c1 & c2) == 0.
2782 if (N1C && N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
2783 isa<ConstantSDNode>(N0.getOperand(1))) {
2784 ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1));
2785 if ((C1->getAPIntValue() & N1C->getAPIntValue()) != 0)
2786 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
2787 DAG.getNode(ISD::OR, N0.getDebugLoc(), VT,
2788 N0.getOperand(0), N1),
2789 DAG.FoldConstantArithmetic(ISD::OR, VT, N1C, C1));
2791 // fold (or (setcc x), (setcc y)) -> (setcc (or x, y))
2792 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
2793 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
2794 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
2796 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
2797 LL.getValueType().isInteger()) {
2798 // fold (or (setne X, 0), (setne Y, 0)) -> (setne (or X, Y), 0)
2799 // fold (or (setlt X, 0), (setlt Y, 0)) -> (setne (or X, Y), 0)
2800 if (cast<ConstantSDNode>(LR)->isNullValue() &&
2801 (Op1 == ISD::SETNE || Op1 == ISD::SETLT)) {
2802 SDValue ORNode = DAG.getNode(ISD::OR, LR.getDebugLoc(),
2803 LR.getValueType(), LL, RL);
2804 AddToWorkList(ORNode.getNode());
2805 return DAG.getSetCC(N->getDebugLoc(), VT, ORNode, LR, Op1);
2807 // fold (or (setne X, -1), (setne Y, -1)) -> (setne (and X, Y), -1)
2808 // fold (or (setgt X, -1), (setgt Y -1)) -> (setgt (and X, Y), -1)
2809 if (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
2810 (Op1 == ISD::SETNE || Op1 == ISD::SETGT)) {
2811 SDValue ANDNode = DAG.getNode(ISD::AND, LR.getDebugLoc(),
2812 LR.getValueType(), LL, RL);
2813 AddToWorkList(ANDNode.getNode());
2814 return DAG.getSetCC(N->getDebugLoc(), VT, ANDNode, LR, Op1);
2817 // canonicalize equivalent to ll == rl
2818 if (LL == RR && LR == RL) {
2819 Op1 = ISD::getSetCCSwappedOperands(Op1);
2822 if (LL == RL && LR == RR) {
2823 bool isInteger = LL.getValueType().isInteger();
2824 ISD::CondCode Result = ISD::getSetCCOrOperation(Op0, Op1, isInteger);
2825 if (Result != ISD::SETCC_INVALID &&
2826 (!LegalOperations || TLI.isCondCodeLegal(Result, LL.getValueType())))
2827 return DAG.getSetCC(N->getDebugLoc(), N0.getValueType(),
2832 // Simplify: (or (op x...), (op y...)) -> (op (or x, y))
2833 if (N0.getOpcode() == N1.getOpcode()) {
2834 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
2835 if (Tmp.getNode()) return Tmp;
2838 // (or (and X, C1), (and Y, C2)) -> (and (or X, Y), C3) if possible.
2839 if (N0.getOpcode() == ISD::AND &&
2840 N1.getOpcode() == ISD::AND &&
2841 N0.getOperand(1).getOpcode() == ISD::Constant &&
2842 N1.getOperand(1).getOpcode() == ISD::Constant &&
2843 // Don't increase # computations.
2844 (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) {
2845 // We can only do this xform if we know that bits from X that are set in C2
2846 // but not in C1 are already zero. Likewise for Y.
2847 const APInt &LHSMask =
2848 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
2849 const APInt &RHSMask =
2850 cast<ConstantSDNode>(N1.getOperand(1))->getAPIntValue();
2852 if (DAG.MaskedValueIsZero(N0.getOperand(0), RHSMask&~LHSMask) &&
2853 DAG.MaskedValueIsZero(N1.getOperand(0), LHSMask&~RHSMask)) {
2854 SDValue X = DAG.getNode(ISD::OR, N0.getDebugLoc(), VT,
2855 N0.getOperand(0), N1.getOperand(0));
2856 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, X,
2857 DAG.getConstant(LHSMask | RHSMask, VT));
2861 // See if this is some rotate idiom.
2862 if (SDNode *Rot = MatchRotate(N0, N1, N->getDebugLoc()))
2863 return SDValue(Rot, 0);
2865 // Simplify the operands using demanded-bits information.
2866 if (!VT.isVector() &&
2867 SimplifyDemandedBits(SDValue(N, 0)))
2868 return SDValue(N, 0);
2873 /// MatchRotateHalf - Match "(X shl/srl V1) & V2" where V2 may not be present.
2874 static bool MatchRotateHalf(SDValue Op, SDValue &Shift, SDValue &Mask) {
2875 if (Op.getOpcode() == ISD::AND) {
2876 if (isa<ConstantSDNode>(Op.getOperand(1))) {
2877 Mask = Op.getOperand(1);
2878 Op = Op.getOperand(0);
2884 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SHL) {
2892 // MatchRotate - Handle an 'or' of two operands. If this is one of the many
2893 // idioms for rotate, and if the target supports rotation instructions, generate
2895 SDNode *DAGCombiner::MatchRotate(SDValue LHS, SDValue RHS, DebugLoc DL) {
2896 // Must be a legal type. Expanded 'n promoted things won't work with rotates.
2897 EVT VT = LHS.getValueType();
2898 if (!TLI.isTypeLegal(VT)) return 0;
2900 // The target must have at least one rotate flavor.
2901 bool HasROTL = TLI.isOperationLegalOrCustom(ISD::ROTL, VT);
2902 bool HasROTR = TLI.isOperationLegalOrCustom(ISD::ROTR, VT);
2903 if (!HasROTL && !HasROTR) return 0;
2905 // Match "(X shl/srl V1) & V2" where V2 may not be present.
2906 SDValue LHSShift; // The shift.
2907 SDValue LHSMask; // AND value if any.
2908 if (!MatchRotateHalf(LHS, LHSShift, LHSMask))
2909 return 0; // Not part of a rotate.
2911 SDValue RHSShift; // The shift.
2912 SDValue RHSMask; // AND value if any.
2913 if (!MatchRotateHalf(RHS, RHSShift, RHSMask))
2914 return 0; // Not part of a rotate.
2916 if (LHSShift.getOperand(0) != RHSShift.getOperand(0))
2917 return 0; // Not shifting the same value.
2919 if (LHSShift.getOpcode() == RHSShift.getOpcode())
2920 return 0; // Shifts must disagree.
2922 // Canonicalize shl to left side in a shl/srl pair.
2923 if (RHSShift.getOpcode() == ISD::SHL) {
2924 std::swap(LHS, RHS);
2925 std::swap(LHSShift, RHSShift);
2926 std::swap(LHSMask , RHSMask );
2929 unsigned OpSizeInBits = VT.getSizeInBits();
2930 SDValue LHSShiftArg = LHSShift.getOperand(0);
2931 SDValue LHSShiftAmt = LHSShift.getOperand(1);
2932 SDValue RHSShiftAmt = RHSShift.getOperand(1);
2934 // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1)
2935 // fold (or (shl x, C1), (srl x, C2)) -> (rotr x, C2)
2936 if (LHSShiftAmt.getOpcode() == ISD::Constant &&
2937 RHSShiftAmt.getOpcode() == ISD::Constant) {
2938 uint64_t LShVal = cast<ConstantSDNode>(LHSShiftAmt)->getZExtValue();
2939 uint64_t RShVal = cast<ConstantSDNode>(RHSShiftAmt)->getZExtValue();
2940 if ((LShVal + RShVal) != OpSizeInBits)
2945 Rot = DAG.getNode(ISD::ROTL, DL, VT, LHSShiftArg, LHSShiftAmt);
2947 Rot = DAG.getNode(ISD::ROTR, DL, VT, LHSShiftArg, RHSShiftAmt);
2949 // If there is an AND of either shifted operand, apply it to the result.
2950 if (LHSMask.getNode() || RHSMask.getNode()) {
2951 APInt Mask = APInt::getAllOnesValue(OpSizeInBits);
2953 if (LHSMask.getNode()) {
2954 APInt RHSBits = APInt::getLowBitsSet(OpSizeInBits, LShVal);
2955 Mask &= cast<ConstantSDNode>(LHSMask)->getAPIntValue() | RHSBits;
2957 if (RHSMask.getNode()) {
2958 APInt LHSBits = APInt::getHighBitsSet(OpSizeInBits, RShVal);
2959 Mask &= cast<ConstantSDNode>(RHSMask)->getAPIntValue() | LHSBits;
2962 Rot = DAG.getNode(ISD::AND, DL, VT, Rot, DAG.getConstant(Mask, VT));
2965 return Rot.getNode();
2968 // If there is a mask here, and we have a variable shift, we can't be sure
2969 // that we're masking out the right stuff.
2970 if (LHSMask.getNode() || RHSMask.getNode())
2973 // fold (or (shl x, y), (srl x, (sub 32, y))) -> (rotl x, y)
2974 // fold (or (shl x, y), (srl x, (sub 32, y))) -> (rotr x, (sub 32, y))
2975 if (RHSShiftAmt.getOpcode() == ISD::SUB &&
2976 LHSShiftAmt == RHSShiftAmt.getOperand(1)) {
2977 if (ConstantSDNode *SUBC =
2978 dyn_cast<ConstantSDNode>(RHSShiftAmt.getOperand(0))) {
2979 if (SUBC->getAPIntValue() == OpSizeInBits) {
2981 return DAG.getNode(ISD::ROTL, DL, VT,
2982 LHSShiftArg, LHSShiftAmt).getNode();
2984 return DAG.getNode(ISD::ROTR, DL, VT,
2985 LHSShiftArg, RHSShiftAmt).getNode();
2990 // fold (or (shl x, (sub 32, y)), (srl x, r)) -> (rotr x, y)
2991 // fold (or (shl x, (sub 32, y)), (srl x, r)) -> (rotl x, (sub 32, y))
2992 if (LHSShiftAmt.getOpcode() == ISD::SUB &&
2993 RHSShiftAmt == LHSShiftAmt.getOperand(1)) {
2994 if (ConstantSDNode *SUBC =
2995 dyn_cast<ConstantSDNode>(LHSShiftAmt.getOperand(0))) {
2996 if (SUBC->getAPIntValue() == OpSizeInBits) {
2998 return DAG.getNode(ISD::ROTR, DL, VT,
2999 LHSShiftArg, RHSShiftAmt).getNode();
3001 return DAG.getNode(ISD::ROTL, DL, VT,
3002 LHSShiftArg, LHSShiftAmt).getNode();
3007 // Look for sign/zext/any-extended or truncate cases:
3008 if ((LHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND
3009 || LHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND
3010 || LHSShiftAmt.getOpcode() == ISD::ANY_EXTEND
3011 || LHSShiftAmt.getOpcode() == ISD::TRUNCATE) &&
3012 (RHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND
3013 || RHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND
3014 || RHSShiftAmt.getOpcode() == ISD::ANY_EXTEND
3015 || RHSShiftAmt.getOpcode() == ISD::TRUNCATE)) {
3016 SDValue LExtOp0 = LHSShiftAmt.getOperand(0);
3017 SDValue RExtOp0 = RHSShiftAmt.getOperand(0);
3018 if (RExtOp0.getOpcode() == ISD::SUB &&
3019 RExtOp0.getOperand(1) == LExtOp0) {
3020 // fold (or (shl x, (*ext y)), (srl x, (*ext (sub 32, y)))) ->
3022 // fold (or (shl x, (*ext y)), (srl x, (*ext (sub 32, y)))) ->
3023 // (rotr x, (sub 32, y))
3024 if (ConstantSDNode *SUBC =
3025 dyn_cast<ConstantSDNode>(RExtOp0.getOperand(0))) {
3026 if (SUBC->getAPIntValue() == OpSizeInBits) {
3027 return DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT,
3029 HasROTL ? LHSShiftAmt : RHSShiftAmt).getNode();
3032 } else if (LExtOp0.getOpcode() == ISD::SUB &&
3033 RExtOp0 == LExtOp0.getOperand(1)) {
3034 // fold (or (shl x, (*ext (sub 32, y))), (srl x, (*ext y))) ->
3036 // fold (or (shl x, (*ext (sub 32, y))), (srl x, (*ext y))) ->
3037 // (rotl x, (sub 32, y))
3038 if (ConstantSDNode *SUBC =
3039 dyn_cast<ConstantSDNode>(LExtOp0.getOperand(0))) {
3040 if (SUBC->getAPIntValue() == OpSizeInBits) {
3041 return DAG.getNode(HasROTR ? ISD::ROTR : ISD::ROTL, DL, VT,
3043 HasROTR ? RHSShiftAmt : LHSShiftAmt).getNode();
3052 SDValue DAGCombiner::visitXOR(SDNode *N) {
3053 SDValue N0 = N->getOperand(0);
3054 SDValue N1 = N->getOperand(1);
3055 SDValue LHS, RHS, CC;
3056 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3057 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3058 EVT VT = N0.getValueType();
3061 if (VT.isVector()) {
3062 SDValue FoldedVOp = SimplifyVBinOp(N);
3063 if (FoldedVOp.getNode()) return FoldedVOp;
3066 // fold (xor undef, undef) -> 0. This is a common idiom (misuse).
3067 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
3068 return DAG.getConstant(0, VT);
3069 // fold (xor x, undef) -> undef
3070 if (N0.getOpcode() == ISD::UNDEF)
3072 if (N1.getOpcode() == ISD::UNDEF)
3074 // fold (xor c1, c2) -> c1^c2
3076 return DAG.FoldConstantArithmetic(ISD::XOR, VT, N0C, N1C);
3077 // canonicalize constant to RHS
3079 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N1, N0);
3080 // fold (xor x, 0) -> x
3081 if (N1C && N1C->isNullValue())
3084 SDValue RXOR = ReassociateOps(ISD::XOR, N->getDebugLoc(), N0, N1);
3085 if (RXOR.getNode() != 0)
3088 // fold !(x cc y) -> (x !cc y)
3089 if (N1C && N1C->getAPIntValue() == 1 && isSetCCEquivalent(N0, LHS, RHS, CC)) {
3090 bool isInt = LHS.getValueType().isInteger();
3091 ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
3094 if (!LegalOperations || TLI.isCondCodeLegal(NotCC, LHS.getValueType())) {
3095 switch (N0.getOpcode()) {
3097 llvm_unreachable("Unhandled SetCC Equivalent!");
3099 return DAG.getSetCC(N->getDebugLoc(), VT, LHS, RHS, NotCC);
3100 case ISD::SELECT_CC:
3101 return DAG.getSelectCC(N->getDebugLoc(), LHS, RHS, N0.getOperand(2),
3102 N0.getOperand(3), NotCC);
3107 // fold (not (zext (setcc x, y))) -> (zext (not (setcc x, y)))
3108 if (N1C && N1C->getAPIntValue() == 1 && N0.getOpcode() == ISD::ZERO_EXTEND &&
3109 N0.getNode()->hasOneUse() &&
3110 isSetCCEquivalent(N0.getOperand(0), LHS, RHS, CC)){
3111 SDValue V = N0.getOperand(0);
3112 V = DAG.getNode(ISD::XOR, N0.getDebugLoc(), V.getValueType(), V,
3113 DAG.getConstant(1, V.getValueType()));
3114 AddToWorkList(V.getNode());
3115 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, V);
3118 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are setcc
3119 if (N1C && N1C->getAPIntValue() == 1 && VT == MVT::i1 &&
3120 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3121 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3122 if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) {
3123 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3124 LHS = DAG.getNode(ISD::XOR, LHS.getDebugLoc(), VT, LHS, N1); // LHS = ~LHS
3125 RHS = DAG.getNode(ISD::XOR, RHS.getDebugLoc(), VT, RHS, N1); // RHS = ~RHS
3126 AddToWorkList(LHS.getNode()); AddToWorkList(RHS.getNode());
3127 return DAG.getNode(NewOpcode, N->getDebugLoc(), VT, LHS, RHS);
3130 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are constants
3131 if (N1C && N1C->isAllOnesValue() &&
3132 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3133 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3134 if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) {
3135 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3136 LHS = DAG.getNode(ISD::XOR, LHS.getDebugLoc(), VT, LHS, N1); // LHS = ~LHS
3137 RHS = DAG.getNode(ISD::XOR, RHS.getDebugLoc(), VT, RHS, N1); // RHS = ~RHS
3138 AddToWorkList(LHS.getNode()); AddToWorkList(RHS.getNode());
3139 return DAG.getNode(NewOpcode, N->getDebugLoc(), VT, LHS, RHS);
3142 // fold (xor (xor x, c1), c2) -> (xor x, (xor c1, c2))
3143 if (N1C && N0.getOpcode() == ISD::XOR) {
3144 ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0));
3145 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3147 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N0.getOperand(1),
3148 DAG.getConstant(N1C->getAPIntValue() ^
3149 N00C->getAPIntValue(), VT));
3151 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N0.getOperand(0),
3152 DAG.getConstant(N1C->getAPIntValue() ^
3153 N01C->getAPIntValue(), VT));
3155 // fold (xor x, x) -> 0
3157 return tryFoldToZero(N->getDebugLoc(), TLI, VT, DAG, LegalOperations);
3159 // Simplify: xor (op x...), (op y...) -> (op (xor x, y))
3160 if (N0.getOpcode() == N1.getOpcode()) {
3161 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3162 if (Tmp.getNode()) return Tmp;
3165 // Simplify the expression using non-local knowledge.
3166 if (!VT.isVector() &&
3167 SimplifyDemandedBits(SDValue(N, 0)))
3168 return SDValue(N, 0);
3173 /// visitShiftByConstant - Handle transforms common to the three shifts, when
3174 /// the shift amount is a constant.
3175 SDValue DAGCombiner::visitShiftByConstant(SDNode *N, unsigned Amt) {
3176 SDNode *LHS = N->getOperand(0).getNode();
3177 if (!LHS->hasOneUse()) return SDValue();
3179 // We want to pull some binops through shifts, so that we have (and (shift))
3180 // instead of (shift (and)), likewise for add, or, xor, etc. This sort of
3181 // thing happens with address calculations, so it's important to canonicalize
3183 bool HighBitSet = false; // Can we transform this if the high bit is set?
3185 switch (LHS->getOpcode()) {
3186 default: return SDValue();
3189 HighBitSet = false; // We can only transform sra if the high bit is clear.
3192 HighBitSet = true; // We can only transform sra if the high bit is set.
3195 if (N->getOpcode() != ISD::SHL)
3196 return SDValue(); // only shl(add) not sr[al](add).
3197 HighBitSet = false; // We can only transform sra if the high bit is clear.
3201 // We require the RHS of the binop to be a constant as well.
3202 ConstantSDNode *BinOpCst = dyn_cast<ConstantSDNode>(LHS->getOperand(1));
3203 if (!BinOpCst) return SDValue();
3205 // FIXME: disable this unless the input to the binop is a shift by a constant.
3206 // If it is not a shift, it pessimizes some common cases like:
3208 // void foo(int *X, int i) { X[i & 1235] = 1; }
3209 // int bar(int *X, int i) { return X[i & 255]; }
3210 SDNode *BinOpLHSVal = LHS->getOperand(0).getNode();
3211 if ((BinOpLHSVal->getOpcode() != ISD::SHL &&
3212 BinOpLHSVal->getOpcode() != ISD::SRA &&
3213 BinOpLHSVal->getOpcode() != ISD::SRL) ||
3214 !isa<ConstantSDNode>(BinOpLHSVal->getOperand(1)))
3217 EVT VT = N->getValueType(0);
3219 // If this is a signed shift right, and the high bit is modified by the
3220 // logical operation, do not perform the transformation. The highBitSet
3221 // boolean indicates the value of the high bit of the constant which would
3222 // cause it to be modified for this operation.
3223 if (N->getOpcode() == ISD::SRA) {
3224 bool BinOpRHSSignSet = BinOpCst->getAPIntValue().isNegative();
3225 if (BinOpRHSSignSet != HighBitSet)
3229 // Fold the constants, shifting the binop RHS by the shift amount.
3230 SDValue NewRHS = DAG.getNode(N->getOpcode(), LHS->getOperand(1).getDebugLoc(),
3232 LHS->getOperand(1), N->getOperand(1));
3234 // Create the new shift.
3235 SDValue NewShift = DAG.getNode(N->getOpcode(),
3236 LHS->getOperand(0).getDebugLoc(),
3237 VT, LHS->getOperand(0), N->getOperand(1));
3239 // Create the new binop.
3240 return DAG.getNode(LHS->getOpcode(), N->getDebugLoc(), VT, NewShift, NewRHS);
3243 SDValue DAGCombiner::visitSHL(SDNode *N) {
3244 SDValue N0 = N->getOperand(0);
3245 SDValue N1 = N->getOperand(1);
3246 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3247 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3248 EVT VT = N0.getValueType();
3249 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
3251 // fold (shl c1, c2) -> c1<<c2
3253 return DAG.FoldConstantArithmetic(ISD::SHL, VT, N0C, N1C);
3254 // fold (shl 0, x) -> 0
3255 if (N0C && N0C->isNullValue())
3257 // fold (shl x, c >= size(x)) -> undef
3258 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
3259 return DAG.getUNDEF(VT);
3260 // fold (shl x, 0) -> x
3261 if (N1C && N1C->isNullValue())
3263 // fold (shl undef, x) -> 0
3264 if (N0.getOpcode() == ISD::UNDEF)
3265 return DAG.getConstant(0, VT);
3266 // if (shl x, c) is known to be zero, return 0
3267 if (DAG.MaskedValueIsZero(SDValue(N, 0),
3268 APInt::getAllOnesValue(OpSizeInBits)))
3269 return DAG.getConstant(0, VT);
3270 // fold (shl x, (trunc (and y, c))) -> (shl x, (and (trunc y), (trunc c))).
3271 if (N1.getOpcode() == ISD::TRUNCATE &&
3272 N1.getOperand(0).getOpcode() == ISD::AND &&
3273 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
3274 SDValue N101 = N1.getOperand(0).getOperand(1);
3275 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
3276 EVT TruncVT = N1.getValueType();
3277 SDValue N100 = N1.getOperand(0).getOperand(0);
3278 APInt TruncC = N101C->getAPIntValue();
3279 TruncC = TruncC.trunc(TruncVT.getSizeInBits());
3280 return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0,
3281 DAG.getNode(ISD::AND, N->getDebugLoc(), TruncVT,
3282 DAG.getNode(ISD::TRUNCATE,
3285 DAG.getConstant(TruncC, TruncVT)));
3289 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
3290 return SDValue(N, 0);
3292 // fold (shl (shl x, c1), c2) -> 0 or (shl x, (add c1, c2))
3293 if (N1C && N0.getOpcode() == ISD::SHL &&
3294 N0.getOperand(1).getOpcode() == ISD::Constant) {
3295 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
3296 uint64_t c2 = N1C->getZExtValue();
3297 if (c1 + c2 >= OpSizeInBits)
3298 return DAG.getConstant(0, VT);
3299 return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0.getOperand(0),
3300 DAG.getConstant(c1 + c2, N1.getValueType()));
3303 // fold (shl (ext (shl x, c1)), c2) -> (ext (shl x, (add c1, c2)))
3304 // For this to be valid, the second form must not preserve any of the bits
3305 // that are shifted out by the inner shift in the first form. This means
3306 // the outer shift size must be >= the number of bits added by the ext.
3307 // As a corollary, we don't care what kind of ext it is.
3308 if (N1C && (N0.getOpcode() == ISD::ZERO_EXTEND ||
3309 N0.getOpcode() == ISD::ANY_EXTEND ||
3310 N0.getOpcode() == ISD::SIGN_EXTEND) &&
3311 N0.getOperand(0).getOpcode() == ISD::SHL &&
3312 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
3314 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
3315 uint64_t c2 = N1C->getZExtValue();
3316 EVT InnerShiftVT = N0.getOperand(0).getValueType();
3317 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
3318 if (c2 >= OpSizeInBits - InnerShiftSize) {
3319 if (c1 + c2 >= OpSizeInBits)
3320 return DAG.getConstant(0, VT);
3321 return DAG.getNode(ISD::SHL, N0->getDebugLoc(), VT,
3322 DAG.getNode(N0.getOpcode(), N0->getDebugLoc(), VT,
3323 N0.getOperand(0)->getOperand(0)),
3324 DAG.getConstant(c1 + c2, N1.getValueType()));
3328 // fold (shl (srl x, c1), c2) -> (and (shl x, (sub c2, c1), MASK) or
3329 // (and (srl x, (sub c1, c2), MASK)
3330 if (N1C && N0.getOpcode() == ISD::SRL &&
3331 N0.getOperand(1).getOpcode() == ISD::Constant) {
3332 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
3333 if (c1 < VT.getSizeInBits()) {
3334 uint64_t c2 = N1C->getZExtValue();
3335 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(),
3336 VT.getSizeInBits() - c1);
3339 Mask = Mask.shl(c2-c1);
3340 Shift = DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0.getOperand(0),
3341 DAG.getConstant(c2-c1, N1.getValueType()));
3343 Mask = Mask.lshr(c1-c2);
3344 Shift = DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0.getOperand(0),
3345 DAG.getConstant(c1-c2, N1.getValueType()));
3347 return DAG.getNode(ISD::AND, N0.getDebugLoc(), VT, Shift,
3348 DAG.getConstant(Mask, VT));
3351 // fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1))
3352 if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1)) {
3353 SDValue HiBitsMask =
3354 DAG.getConstant(APInt::getHighBitsSet(VT.getSizeInBits(),
3355 VT.getSizeInBits() -
3356 N1C->getZExtValue()),
3358 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0.getOperand(0),
3363 SDValue NewSHL = visitShiftByConstant(N, N1C->getZExtValue());
3364 if (NewSHL.getNode())
3371 SDValue DAGCombiner::visitSRA(SDNode *N) {
3372 SDValue N0 = N->getOperand(0);
3373 SDValue N1 = N->getOperand(1);
3374 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3375 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3376 EVT VT = N0.getValueType();
3377 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
3379 // fold (sra c1, c2) -> (sra c1, c2)
3381 return DAG.FoldConstantArithmetic(ISD::SRA, VT, N0C, N1C);
3382 // fold (sra 0, x) -> 0
3383 if (N0C && N0C->isNullValue())
3385 // fold (sra -1, x) -> -1
3386 if (N0C && N0C->isAllOnesValue())
3388 // fold (sra x, (setge c, size(x))) -> undef
3389 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
3390 return DAG.getUNDEF(VT);
3391 // fold (sra x, 0) -> x
3392 if (N1C && N1C->isNullValue())
3394 // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports
3396 if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) {
3397 unsigned LowBits = OpSizeInBits - (unsigned)N1C->getZExtValue();
3398 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), LowBits);
3400 ExtVT = EVT::getVectorVT(*DAG.getContext(),
3401 ExtVT, VT.getVectorNumElements());
3402 if ((!LegalOperations ||
3403 TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, ExtVT)))
3404 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT,
3405 N0.getOperand(0), DAG.getValueType(ExtVT));
3408 // fold (sra (sra x, c1), c2) -> (sra x, (add c1, c2))
3409 if (N1C && N0.getOpcode() == ISD::SRA) {
3410 if (ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
3411 unsigned Sum = N1C->getZExtValue() + C1->getZExtValue();
3412 if (Sum >= OpSizeInBits) Sum = OpSizeInBits-1;
3413 return DAG.getNode(ISD::SRA, N->getDebugLoc(), VT, N0.getOperand(0),
3414 DAG.getConstant(Sum, N1C->getValueType(0)));
3418 // fold (sra (shl X, m), (sub result_size, n))
3419 // -> (sign_extend (trunc (shl X, (sub (sub result_size, n), m)))) for
3420 // result_size - n != m.
3421 // If truncate is free for the target sext(shl) is likely to result in better
3423 if (N0.getOpcode() == ISD::SHL) {
3424 // Get the two constanst of the shifts, CN0 = m, CN = n.
3425 const ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3427 // Determine what the truncate's result bitsize and type would be.
3429 EVT::getIntegerVT(*DAG.getContext(),
3430 OpSizeInBits - N1C->getZExtValue());
3431 // Determine the residual right-shift amount.
3432 signed ShiftAmt = N1C->getZExtValue() - N01C->getZExtValue();
3434 // If the shift is not a no-op (in which case this should be just a sign
3435 // extend already), the truncated to type is legal, sign_extend is legal
3436 // on that type, and the truncate to that type is both legal and free,
3437 // perform the transform.
3438 if ((ShiftAmt > 0) &&
3439 TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND, TruncVT) &&
3440 TLI.isOperationLegalOrCustom(ISD::TRUNCATE, VT) &&
3441 TLI.isTruncateFree(VT, TruncVT)) {
3443 SDValue Amt = DAG.getConstant(ShiftAmt,
3444 getShiftAmountTy(N0.getOperand(0).getValueType()));
3445 SDValue Shift = DAG.getNode(ISD::SRL, N0.getDebugLoc(), VT,
3446 N0.getOperand(0), Amt);
3447 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(), TruncVT,
3449 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(),
3450 N->getValueType(0), Trunc);
3455 // fold (sra x, (trunc (and y, c))) -> (sra x, (and (trunc y), (trunc c))).
3456 if (N1.getOpcode() == ISD::TRUNCATE &&
3457 N1.getOperand(0).getOpcode() == ISD::AND &&
3458 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
3459 SDValue N101 = N1.getOperand(0).getOperand(1);
3460 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
3461 EVT TruncVT = N1.getValueType();
3462 SDValue N100 = N1.getOperand(0).getOperand(0);
3463 APInt TruncC = N101C->getAPIntValue();
3464 TruncC = TruncC.trunc(TruncVT.getScalarType().getSizeInBits());
3465 return DAG.getNode(ISD::SRA, N->getDebugLoc(), VT, N0,
3466 DAG.getNode(ISD::AND, N->getDebugLoc(),
3468 DAG.getNode(ISD::TRUNCATE,
3471 DAG.getConstant(TruncC, TruncVT)));
3475 // fold (sra (trunc (sr x, c1)), c2) -> (trunc (sra x, c1+c2))
3476 // if c1 is equal to the number of bits the trunc removes
3477 if (N0.getOpcode() == ISD::TRUNCATE &&
3478 (N0.getOperand(0).getOpcode() == ISD::SRL ||
3479 N0.getOperand(0).getOpcode() == ISD::SRA) &&
3480 N0.getOperand(0).hasOneUse() &&
3481 N0.getOperand(0).getOperand(1).hasOneUse() &&
3482 N1C && isa<ConstantSDNode>(N0.getOperand(0).getOperand(1))) {
3483 EVT LargeVT = N0.getOperand(0).getValueType();
3484 ConstantSDNode *LargeShiftAmt =
3485 cast<ConstantSDNode>(N0.getOperand(0).getOperand(1));
3487 if (LargeVT.getScalarType().getSizeInBits() - OpSizeInBits ==
3488 LargeShiftAmt->getZExtValue()) {
3490 DAG.getConstant(LargeShiftAmt->getZExtValue() + N1C->getZExtValue(),
3491 getShiftAmountTy(N0.getOperand(0).getOperand(0).getValueType()));
3492 SDValue SRA = DAG.getNode(ISD::SRA, N->getDebugLoc(), LargeVT,
3493 N0.getOperand(0).getOperand(0), Amt);
3494 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, SRA);
3498 // Simplify, based on bits shifted out of the LHS.
3499 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
3500 return SDValue(N, 0);
3503 // If the sign bit is known to be zero, switch this to a SRL.
3504 if (DAG.SignBitIsZero(N0))
3505 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0, N1);
3508 SDValue NewSRA = visitShiftByConstant(N, N1C->getZExtValue());
3509 if (NewSRA.getNode())
3516 SDValue DAGCombiner::visitSRL(SDNode *N) {
3517 SDValue N0 = N->getOperand(0);
3518 SDValue N1 = N->getOperand(1);
3519 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3520 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3521 EVT VT = N0.getValueType();
3522 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
3524 // fold (srl c1, c2) -> c1 >>u c2
3526 return DAG.FoldConstantArithmetic(ISD::SRL, VT, N0C, N1C);
3527 // fold (srl 0, x) -> 0
3528 if (N0C && N0C->isNullValue())
3530 // fold (srl x, c >= size(x)) -> undef
3531 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
3532 return DAG.getUNDEF(VT);
3533 // fold (srl x, 0) -> x
3534 if (N1C && N1C->isNullValue())
3536 // if (srl x, c) is known to be zero, return 0
3537 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
3538 APInt::getAllOnesValue(OpSizeInBits)))
3539 return DAG.getConstant(0, VT);
3541 // fold (srl (srl x, c1), c2) -> 0 or (srl x, (add c1, c2))
3542 if (N1C && N0.getOpcode() == ISD::SRL &&
3543 N0.getOperand(1).getOpcode() == ISD::Constant) {
3544 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
3545 uint64_t c2 = N1C->getZExtValue();
3546 if (c1 + c2 >= OpSizeInBits)
3547 return DAG.getConstant(0, VT);
3548 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0.getOperand(0),
3549 DAG.getConstant(c1 + c2, N1.getValueType()));
3552 // fold (srl (trunc (srl x, c1)), c2) -> 0 or (trunc (srl x, (add c1, c2)))
3553 if (N1C && N0.getOpcode() == ISD::TRUNCATE &&
3554 N0.getOperand(0).getOpcode() == ISD::SRL &&
3555 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
3557 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
3558 uint64_t c2 = N1C->getZExtValue();
3559 EVT InnerShiftVT = N0.getOperand(0).getValueType();
3560 EVT ShiftCountVT = N0.getOperand(0)->getOperand(1).getValueType();
3561 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
3562 // This is only valid if the OpSizeInBits + c1 = size of inner shift.
3563 if (c1 + OpSizeInBits == InnerShiftSize) {
3564 if (c1 + c2 >= InnerShiftSize)
3565 return DAG.getConstant(0, VT);
3566 return DAG.getNode(ISD::TRUNCATE, N0->getDebugLoc(), VT,
3567 DAG.getNode(ISD::SRL, N0->getDebugLoc(), InnerShiftVT,
3568 N0.getOperand(0)->getOperand(0),
3569 DAG.getConstant(c1 + c2, ShiftCountVT)));
3573 // fold (srl (shl x, c), c) -> (and x, cst2)
3574 if (N1C && N0.getOpcode() == ISD::SHL && N0.getOperand(1) == N1 &&
3575 N0.getValueSizeInBits() <= 64) {
3576 uint64_t ShAmt = N1C->getZExtValue()+64-N0.getValueSizeInBits();
3577 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0.getOperand(0),
3578 DAG.getConstant(~0ULL >> ShAmt, VT));
3582 // fold (srl (anyextend x), c) -> (anyextend (srl x, c))
3583 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
3584 // Shifting in all undef bits?
3585 EVT SmallVT = N0.getOperand(0).getValueType();
3586 if (N1C->getZExtValue() >= SmallVT.getSizeInBits())
3587 return DAG.getUNDEF(VT);
3589 if (!LegalTypes || TLI.isTypeDesirableForOp(ISD::SRL, SmallVT)) {
3590 uint64_t ShiftAmt = N1C->getZExtValue();
3591 SDValue SmallShift = DAG.getNode(ISD::SRL, N0.getDebugLoc(), SmallVT,
3593 DAG.getConstant(ShiftAmt, getShiftAmountTy(SmallVT)));
3594 AddToWorkList(SmallShift.getNode());
3595 return DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, SmallShift);
3599 // fold (srl (sra X, Y), 31) -> (srl X, 31). This srl only looks at the sign
3600 // bit, which is unmodified by sra.
3601 if (N1C && N1C->getZExtValue() + 1 == VT.getSizeInBits()) {
3602 if (N0.getOpcode() == ISD::SRA)
3603 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0.getOperand(0), N1);
3606 // fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit).
3607 if (N1C && N0.getOpcode() == ISD::CTLZ &&
3608 N1C->getAPIntValue() == Log2_32(VT.getSizeInBits())) {
3609 APInt KnownZero, KnownOne;
3610 APInt Mask = APInt::getAllOnesValue(VT.getScalarType().getSizeInBits());
3611 DAG.ComputeMaskedBits(N0.getOperand(0), Mask, KnownZero, KnownOne);
3613 // If any of the input bits are KnownOne, then the input couldn't be all
3614 // zeros, thus the result of the srl will always be zero.
3615 if (KnownOne.getBoolValue()) return DAG.getConstant(0, VT);
3617 // If all of the bits input the to ctlz node are known to be zero, then
3618 // the result of the ctlz is "32" and the result of the shift is one.
3619 APInt UnknownBits = ~KnownZero & Mask;
3620 if (UnknownBits == 0) return DAG.getConstant(1, VT);
3622 // Otherwise, check to see if there is exactly one bit input to the ctlz.
3623 if ((UnknownBits & (UnknownBits - 1)) == 0) {
3624 // Okay, we know that only that the single bit specified by UnknownBits
3625 // could be set on input to the CTLZ node. If this bit is set, the SRL
3626 // will return 0, if it is clear, it returns 1. Change the CTLZ/SRL pair
3627 // to an SRL/XOR pair, which is likely to simplify more.
3628 unsigned ShAmt = UnknownBits.countTrailingZeros();
3629 SDValue Op = N0.getOperand(0);
3632 Op = DAG.getNode(ISD::SRL, N0.getDebugLoc(), VT, Op,
3633 DAG.getConstant(ShAmt, getShiftAmountTy(Op.getValueType())));
3634 AddToWorkList(Op.getNode());
3637 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT,
3638 Op, DAG.getConstant(1, VT));
3642 // fold (srl x, (trunc (and y, c))) -> (srl x, (and (trunc y), (trunc c))).
3643 if (N1.getOpcode() == ISD::TRUNCATE &&
3644 N1.getOperand(0).getOpcode() == ISD::AND &&
3645 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
3646 SDValue N101 = N1.getOperand(0).getOperand(1);
3647 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
3648 EVT TruncVT = N1.getValueType();
3649 SDValue N100 = N1.getOperand(0).getOperand(0);
3650 APInt TruncC = N101C->getAPIntValue();
3651 TruncC = TruncC.trunc(TruncVT.getSizeInBits());
3652 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0,
3653 DAG.getNode(ISD::AND, N->getDebugLoc(),
3655 DAG.getNode(ISD::TRUNCATE,
3658 DAG.getConstant(TruncC, TruncVT)));
3662 // fold operands of srl based on knowledge that the low bits are not
3664 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
3665 return SDValue(N, 0);
3668 SDValue NewSRL = visitShiftByConstant(N, N1C->getZExtValue());
3669 if (NewSRL.getNode())
3673 // Attempt to convert a srl of a load into a narrower zero-extending load.
3674 SDValue NarrowLoad = ReduceLoadWidth(N);
3675 if (NarrowLoad.getNode())
3678 // Here is a common situation. We want to optimize:
3681 // %b = and i32 %a, 2
3682 // %c = srl i32 %b, 1
3683 // brcond i32 %c ...
3689 // %c = setcc eq %b, 0
3692 // However when after the source operand of SRL is optimized into AND, the SRL
3693 // itself may not be optimized further. Look for it and add the BRCOND into
3695 if (N->hasOneUse()) {
3696 SDNode *Use = *N->use_begin();
3697 if (Use->getOpcode() == ISD::BRCOND)
3699 else if (Use->getOpcode() == ISD::TRUNCATE && Use->hasOneUse()) {
3700 // Also look pass the truncate.
3701 Use = *Use->use_begin();
3702 if (Use->getOpcode() == ISD::BRCOND)
3710 SDValue DAGCombiner::visitCTLZ(SDNode *N) {
3711 SDValue N0 = N->getOperand(0);
3712 EVT VT = N->getValueType(0);
3714 // fold (ctlz c1) -> c2
3715 if (isa<ConstantSDNode>(N0))
3716 return DAG.getNode(ISD::CTLZ, N->getDebugLoc(), VT, N0);
3720 SDValue DAGCombiner::visitCTTZ(SDNode *N) {
3721 SDValue N0 = N->getOperand(0);
3722 EVT VT = N->getValueType(0);
3724 // fold (cttz c1) -> c2
3725 if (isa<ConstantSDNode>(N0))
3726 return DAG.getNode(ISD::CTTZ, N->getDebugLoc(), VT, N0);
3730 SDValue DAGCombiner::visitCTPOP(SDNode *N) {
3731 SDValue N0 = N->getOperand(0);
3732 EVT VT = N->getValueType(0);
3734 // fold (ctpop c1) -> c2
3735 if (isa<ConstantSDNode>(N0))
3736 return DAG.getNode(ISD::CTPOP, N->getDebugLoc(), VT, N0);
3740 SDValue DAGCombiner::visitSELECT(SDNode *N) {
3741 SDValue N0 = N->getOperand(0);
3742 SDValue N1 = N->getOperand(1);
3743 SDValue N2 = N->getOperand(2);
3744 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3745 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3746 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
3747 EVT VT = N->getValueType(0);
3748 EVT VT0 = N0.getValueType();
3750 // fold (select C, X, X) -> X
3753 // fold (select true, X, Y) -> X
3754 if (N0C && !N0C->isNullValue())
3756 // fold (select false, X, Y) -> Y
3757 if (N0C && N0C->isNullValue())
3759 // fold (select C, 1, X) -> (or C, X)
3760 if (VT == MVT::i1 && N1C && N1C->getAPIntValue() == 1)
3761 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N0, N2);
3762 // fold (select C, 0, 1) -> (xor C, 1)
3763 if (VT.isInteger() &&
3766 TLI.getBooleanContents(false) == TargetLowering::ZeroOrOneBooleanContent)) &&
3767 N1C && N2C && N1C->isNullValue() && N2C->getAPIntValue() == 1) {
3770 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT0,
3771 N0, DAG.getConstant(1, VT0));
3772 XORNode = DAG.getNode(ISD::XOR, N0.getDebugLoc(), VT0,
3773 N0, DAG.getConstant(1, VT0));
3774 AddToWorkList(XORNode.getNode());
3776 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, XORNode);
3777 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, XORNode);
3779 // fold (select C, 0, X) -> (and (not C), X)
3780 if (VT == VT0 && VT == MVT::i1 && N1C && N1C->isNullValue()) {
3781 SDValue NOTNode = DAG.getNOT(N0.getDebugLoc(), N0, VT);
3782 AddToWorkList(NOTNode.getNode());
3783 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, NOTNode, N2);
3785 // fold (select C, X, 1) -> (or (not C), X)
3786 if (VT == VT0 && VT == MVT::i1 && N2C && N2C->getAPIntValue() == 1) {
3787 SDValue NOTNode = DAG.getNOT(N0.getDebugLoc(), N0, VT);
3788 AddToWorkList(NOTNode.getNode());
3789 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, NOTNode, N1);
3791 // fold (select C, X, 0) -> (and C, X)
3792 if (VT == MVT::i1 && N2C && N2C->isNullValue())
3793 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0, N1);
3794 // fold (select X, X, Y) -> (or X, Y)
3795 // fold (select X, 1, Y) -> (or X, Y)
3796 if (VT == MVT::i1 && (N0 == N1 || (N1C && N1C->getAPIntValue() == 1)))
3797 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N0, N2);
3798 // fold (select X, Y, X) -> (and X, Y)
3799 // fold (select X, Y, 0) -> (and X, Y)
3800 if (VT == MVT::i1 && (N0 == N2 || (N2C && N2C->getAPIntValue() == 0)))
3801 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0, N1);
3803 // If we can fold this based on the true/false value, do so.
3804 if (SimplifySelectOps(N, N1, N2))
3805 return SDValue(N, 0); // Don't revisit N.
3807 // fold selects based on a setcc into other things, such as min/max/abs
3808 if (N0.getOpcode() == ISD::SETCC) {
3810 // Check against MVT::Other for SELECT_CC, which is a workaround for targets
3811 // having to say they don't support SELECT_CC on every type the DAG knows
3812 // about, since there is no way to mark an opcode illegal at all value types
3813 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, MVT::Other) &&
3814 TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT))
3815 return DAG.getNode(ISD::SELECT_CC, N->getDebugLoc(), VT,
3816 N0.getOperand(0), N0.getOperand(1),
3817 N1, N2, N0.getOperand(2));
3818 return SimplifySelect(N->getDebugLoc(), N0, N1, N2);
3824 SDValue DAGCombiner::visitSELECT_CC(SDNode *N) {
3825 SDValue N0 = N->getOperand(0);
3826 SDValue N1 = N->getOperand(1);
3827 SDValue N2 = N->getOperand(2);
3828 SDValue N3 = N->getOperand(3);
3829 SDValue N4 = N->getOperand(4);
3830 ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get();
3832 // fold select_cc lhs, rhs, x, x, cc -> x
3836 // Determine if the condition we're dealing with is constant
3837 SDValue SCC = SimplifySetCC(TLI.getSetCCResultType(N0.getValueType()),
3838 N0, N1, CC, N->getDebugLoc(), false);
3839 if (SCC.getNode()) AddToWorkList(SCC.getNode());
3841 if (ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode())) {
3842 if (!SCCC->isNullValue())
3843 return N2; // cond always true -> true val
3845 return N3; // cond always false -> false val
3848 // Fold to a simpler select_cc
3849 if (SCC.getNode() && SCC.getOpcode() == ISD::SETCC)
3850 return DAG.getNode(ISD::SELECT_CC, N->getDebugLoc(), N2.getValueType(),
3851 SCC.getOperand(0), SCC.getOperand(1), N2, N3,
3854 // If we can fold this based on the true/false value, do so.
3855 if (SimplifySelectOps(N, N2, N3))
3856 return SDValue(N, 0); // Don't revisit N.
3858 // fold select_cc into other things, such as min/max/abs
3859 return SimplifySelectCC(N->getDebugLoc(), N0, N1, N2, N3, CC);
3862 SDValue DAGCombiner::visitSETCC(SDNode *N) {
3863 return SimplifySetCC(N->getValueType(0), N->getOperand(0), N->getOperand(1),
3864 cast<CondCodeSDNode>(N->getOperand(2))->get(),
3868 // ExtendUsesToFormExtLoad - Trying to extend uses of a load to enable this:
3869 // "fold ({s|z|a}ext (load x)) -> ({s|z|a}ext (truncate ({s|z|a}extload x)))"
3870 // transformation. Returns true if extension are possible and the above
3871 // mentioned transformation is profitable.
3872 static bool ExtendUsesToFormExtLoad(SDNode *N, SDValue N0,
3874 SmallVector<SDNode*, 4> &ExtendNodes,
3875 const TargetLowering &TLI) {
3876 bool HasCopyToRegUses = false;
3877 bool isTruncFree = TLI.isTruncateFree(N->getValueType(0), N0.getValueType());
3878 for (SDNode::use_iterator UI = N0.getNode()->use_begin(),
3879 UE = N0.getNode()->use_end();
3884 if (UI.getUse().getResNo() != N0.getResNo())
3886 // FIXME: Only extend SETCC N, N and SETCC N, c for now.
3887 if (ExtOpc != ISD::ANY_EXTEND && User->getOpcode() == ISD::SETCC) {
3888 ISD::CondCode CC = cast<CondCodeSDNode>(User->getOperand(2))->get();
3889 if (ExtOpc == ISD::ZERO_EXTEND && ISD::isSignedIntSetCC(CC))
3890 // Sign bits will be lost after a zext.
3893 for (unsigned i = 0; i != 2; ++i) {
3894 SDValue UseOp = User->getOperand(i);
3897 if (!isa<ConstantSDNode>(UseOp))
3902 ExtendNodes.push_back(User);
3905 // If truncates aren't free and there are users we can't
3906 // extend, it isn't worthwhile.
3909 // Remember if this value is live-out.
3910 if (User->getOpcode() == ISD::CopyToReg)
3911 HasCopyToRegUses = true;
3914 if (HasCopyToRegUses) {
3915 bool BothLiveOut = false;
3916 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
3918 SDUse &Use = UI.getUse();
3919 if (Use.getResNo() == 0 && Use.getUser()->getOpcode() == ISD::CopyToReg) {
3925 // Both unextended and extended values are live out. There had better be
3926 // a good reason for the transformation.
3927 return ExtendNodes.size();
3932 void DAGCombiner::ExtendSetCCUses(SmallVector<SDNode*, 4> SetCCs,
3933 SDValue Trunc, SDValue ExtLoad, DebugLoc DL,
3934 ISD::NodeType ExtType) {
3935 // Extend SetCC uses if necessary.
3936 for (unsigned i = 0, e = SetCCs.size(); i != e; ++i) {
3937 SDNode *SetCC = SetCCs[i];
3938 SmallVector<SDValue, 4> Ops;
3940 for (unsigned j = 0; j != 2; ++j) {
3941 SDValue SOp = SetCC->getOperand(j);
3943 Ops.push_back(ExtLoad);
3945 Ops.push_back(DAG.getNode(ExtType, DL, ExtLoad->getValueType(0), SOp));
3948 Ops.push_back(SetCC->getOperand(2));
3949 CombineTo(SetCC, DAG.getNode(ISD::SETCC, DL, SetCC->getValueType(0),
3950 &Ops[0], Ops.size()));
3954 SDValue DAGCombiner::visitSIGN_EXTEND(SDNode *N) {
3955 SDValue N0 = N->getOperand(0);
3956 EVT VT = N->getValueType(0);
3958 // fold (sext c1) -> c1
3959 if (isa<ConstantSDNode>(N0))
3960 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, N0);
3962 // fold (sext (sext x)) -> (sext x)
3963 // fold (sext (aext x)) -> (sext x)
3964 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
3965 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT,
3968 if (N0.getOpcode() == ISD::TRUNCATE) {
3969 // fold (sext (truncate (load x))) -> (sext (smaller load x))
3970 // fold (sext (truncate (srl (load x), c))) -> (sext (smaller load (x+c/n)))
3971 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
3972 if (NarrowLoad.getNode()) {
3973 SDNode* oye = N0.getNode()->getOperand(0).getNode();
3974 if (NarrowLoad.getNode() != N0.getNode()) {
3975 CombineTo(N0.getNode(), NarrowLoad);
3976 // CombineTo deleted the truncate, if needed, but not what's under it.
3979 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3982 // See if the value being truncated is already sign extended. If so, just
3983 // eliminate the trunc/sext pair.
3984 SDValue Op = N0.getOperand(0);
3985 unsigned OpBits = Op.getValueType().getScalarType().getSizeInBits();
3986 unsigned MidBits = N0.getValueType().getScalarType().getSizeInBits();
3987 unsigned DestBits = VT.getScalarType().getSizeInBits();
3988 unsigned NumSignBits = DAG.ComputeNumSignBits(Op);
3990 if (OpBits == DestBits) {
3991 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
3992 // bits, it is already ready.
3993 if (NumSignBits > DestBits-MidBits)
3995 } else if (OpBits < DestBits) {
3996 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
3997 // bits, just sext from i32.
3998 if (NumSignBits > OpBits-MidBits)
3999 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, Op);
4001 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
4002 // bits, just truncate to i32.
4003 if (NumSignBits > OpBits-MidBits)
4004 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, Op);
4007 // fold (sext (truncate x)) -> (sextinreg x).
4008 if (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG,
4009 N0.getValueType())) {
4010 if (OpBits < DestBits)
4011 Op = DAG.getNode(ISD::ANY_EXTEND, N0.getDebugLoc(), VT, Op);
4012 else if (OpBits > DestBits)
4013 Op = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(), VT, Op);
4014 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT, Op,
4015 DAG.getValueType(N0.getValueType()));
4019 // fold (sext (load x)) -> (sext (truncate (sextload x)))
4020 // None of the supported targets knows how to perform load and sign extend
4021 // on vectors in one instruction. We only perform this transformation on
4023 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
4024 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
4025 TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()))) {
4026 bool DoXform = true;
4027 SmallVector<SDNode*, 4> SetCCs;
4028 if (!N0.hasOneUse())
4029 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::SIGN_EXTEND, SetCCs, TLI);
4031 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4032 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
4034 LN0->getBasePtr(), LN0->getPointerInfo(),
4036 LN0->isVolatile(), LN0->isNonTemporal(),
4037 LN0->getAlignment());
4038 CombineTo(N, ExtLoad);
4039 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4040 N0.getValueType(), ExtLoad);
4041 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
4042 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, N->getDebugLoc(),
4044 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4048 // fold (sext (sextload x)) -> (sext (truncate (sextload x)))
4049 // fold (sext ( extload x)) -> (sext (truncate (sextload x)))
4050 if ((ISD::isSEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
4051 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
4052 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4053 EVT MemVT = LN0->getMemoryVT();
4054 if ((!LegalOperations && !LN0->isVolatile()) ||
4055 TLI.isLoadExtLegal(ISD::SEXTLOAD, MemVT)) {
4056 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
4058 LN0->getBasePtr(), LN0->getPointerInfo(),
4060 LN0->isVolatile(), LN0->isNonTemporal(),
4061 LN0->getAlignment());
4062 CombineTo(N, ExtLoad);
4063 CombineTo(N0.getNode(),
4064 DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4065 N0.getValueType(), ExtLoad),
4066 ExtLoad.getValue(1));
4067 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4071 // fold (sext (and/or/xor (load x), cst)) ->
4072 // (and/or/xor (sextload x), (sext cst))
4073 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
4074 N0.getOpcode() == ISD::XOR) &&
4075 isa<LoadSDNode>(N0.getOperand(0)) &&
4076 N0.getOperand(1).getOpcode() == ISD::Constant &&
4077 TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()) &&
4078 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
4079 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
4080 if (LN0->getExtensionType() != ISD::ZEXTLOAD) {
4081 bool DoXform = true;
4082 SmallVector<SDNode*, 4> SetCCs;
4083 if (!N0.hasOneUse())
4084 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::SIGN_EXTEND,
4087 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, LN0->getDebugLoc(), VT,
4088 LN0->getChain(), LN0->getBasePtr(),
4089 LN0->getPointerInfo(),
4092 LN0->isNonTemporal(),
4093 LN0->getAlignment());
4094 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
4095 Mask = Mask.sext(VT.getSizeInBits());
4096 SDValue And = DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT,
4097 ExtLoad, DAG.getConstant(Mask, VT));
4098 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
4099 N0.getOperand(0).getDebugLoc(),
4100 N0.getOperand(0).getValueType(), ExtLoad);
4102 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
4103 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, N->getDebugLoc(),
4105 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4110 if (N0.getOpcode() == ISD::SETCC) {
4111 // sext(setcc) -> sext_in_reg(vsetcc) for vectors.
4112 // Only do this before legalize for now.
4113 if (VT.isVector() && !LegalOperations) {
4114 EVT N0VT = N0.getOperand(0).getValueType();
4115 // We know that the # elements of the results is the same as the
4116 // # elements of the compare (and the # elements of the compare result
4117 // for that matter). Check to see that they are the same size. If so,
4118 // we know that the element size of the sext'd result matches the
4119 // element size of the compare operands.
4120 if (VT.getSizeInBits() == N0VT.getSizeInBits())
4121 return DAG.getSetCC(N->getDebugLoc(), VT, N0.getOperand(0),
4123 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4124 // If the desired elements are smaller or larger than the source
4125 // elements we can use a matching integer vector type and then
4126 // truncate/sign extend
4128 EVT MatchingElementType =
4129 EVT::getIntegerVT(*DAG.getContext(),
4130 N0VT.getScalarType().getSizeInBits());
4131 EVT MatchingVectorType =
4132 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
4133 N0VT.getVectorNumElements());
4135 DAG.getSetCC(N->getDebugLoc(), MatchingVectorType, N0.getOperand(0),
4137 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4138 return DAG.getSExtOrTrunc(VsetCC, N->getDebugLoc(), VT);
4142 // sext(setcc x, y, cc) -> (select_cc x, y, -1, 0, cc)
4143 unsigned ElementWidth = VT.getScalarType().getSizeInBits();
4145 DAG.getConstant(APInt::getAllOnesValue(ElementWidth), VT);
4147 SimplifySelectCC(N->getDebugLoc(), N0.getOperand(0), N0.getOperand(1),
4148 NegOne, DAG.getConstant(0, VT),
4149 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
4150 if (SCC.getNode()) return SCC;
4151 if (!LegalOperations ||
4152 TLI.isOperationLegal(ISD::SETCC, TLI.getSetCCResultType(VT)))
4153 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
4154 DAG.getSetCC(N->getDebugLoc(),
4155 TLI.getSetCCResultType(VT),
4156 N0.getOperand(0), N0.getOperand(1),
4157 cast<CondCodeSDNode>(N0.getOperand(2))->get()),
4158 NegOne, DAG.getConstant(0, VT));
4161 // fold (sext x) -> (zext x) if the sign bit is known zero.
4162 if ((!LegalOperations || TLI.isOperationLegal(ISD::ZERO_EXTEND, VT)) &&
4163 DAG.SignBitIsZero(N0))
4164 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, N0);
4169 SDValue DAGCombiner::visitZERO_EXTEND(SDNode *N) {
4170 SDValue N0 = N->getOperand(0);
4171 EVT VT = N->getValueType(0);
4173 // fold (zext c1) -> c1
4174 if (isa<ConstantSDNode>(N0))
4175 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, N0);
4176 // fold (zext (zext x)) -> (zext x)
4177 // fold (zext (aext x)) -> (zext x)
4178 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
4179 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT,
4182 // fold (zext (truncate (load x))) -> (zext (smaller load x))
4183 // fold (zext (truncate (srl (load x), c))) -> (zext (small load (x+c/n)))
4184 if (N0.getOpcode() == ISD::TRUNCATE) {
4185 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
4186 if (NarrowLoad.getNode()) {
4187 SDNode* oye = N0.getNode()->getOperand(0).getNode();
4188 if (NarrowLoad.getNode() != N0.getNode()) {
4189 CombineTo(N0.getNode(), NarrowLoad);
4190 // CombineTo deleted the truncate, if needed, but not what's under it.
4193 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4197 // fold (zext (truncate x)) -> (and x, mask)
4198 if (N0.getOpcode() == ISD::TRUNCATE &&
4199 (!LegalOperations || TLI.isOperationLegal(ISD::AND, VT))) {
4201 // fold (zext (truncate (load x))) -> (zext (smaller load x))
4202 // fold (zext (truncate (srl (load x), c))) -> (zext (smaller load (x+c/n)))
4203 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
4204 if (NarrowLoad.getNode()) {
4205 SDNode* oye = N0.getNode()->getOperand(0).getNode();
4206 if (NarrowLoad.getNode() != N0.getNode()) {
4207 CombineTo(N0.getNode(), NarrowLoad);
4208 // CombineTo deleted the truncate, if needed, but not what's under it.
4211 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4214 SDValue Op = N0.getOperand(0);
4215 if (Op.getValueType().bitsLT(VT)) {
4216 Op = DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, Op);
4217 } else if (Op.getValueType().bitsGT(VT)) {
4218 Op = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, Op);
4220 return DAG.getZeroExtendInReg(Op, N->getDebugLoc(),
4221 N0.getValueType().getScalarType());
4224 // Fold (zext (and (trunc x), cst)) -> (and x, cst),
4225 // if either of the casts is not free.
4226 if (N0.getOpcode() == ISD::AND &&
4227 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
4228 N0.getOperand(1).getOpcode() == ISD::Constant &&
4229 (!TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
4230 N0.getValueType()) ||
4231 !TLI.isZExtFree(N0.getValueType(), VT))) {
4232 SDValue X = N0.getOperand(0).getOperand(0);
4233 if (X.getValueType().bitsLT(VT)) {
4234 X = DAG.getNode(ISD::ANY_EXTEND, X.getDebugLoc(), VT, X);
4235 } else if (X.getValueType().bitsGT(VT)) {
4236 X = DAG.getNode(ISD::TRUNCATE, X.getDebugLoc(), VT, X);
4238 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
4239 Mask = Mask.zext(VT.getSizeInBits());
4240 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
4241 X, DAG.getConstant(Mask, VT));
4244 // fold (zext (load x)) -> (zext (truncate (zextload x)))
4245 // None of the supported targets knows how to perform load and vector_zext
4246 // on vectors in one instruction. We only perform this transformation on
4248 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
4249 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
4250 TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()))) {
4251 bool DoXform = true;
4252 SmallVector<SDNode*, 4> SetCCs;
4253 if (!N0.hasOneUse())
4254 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ZERO_EXTEND, SetCCs, TLI);
4256 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4257 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N->getDebugLoc(), VT,
4259 LN0->getBasePtr(), LN0->getPointerInfo(),
4261 LN0->isVolatile(), LN0->isNonTemporal(),
4262 LN0->getAlignment());
4263 CombineTo(N, ExtLoad);
4264 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4265 N0.getValueType(), ExtLoad);
4266 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
4268 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, N->getDebugLoc(),
4270 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4274 // fold (zext (and/or/xor (load x), cst)) ->
4275 // (and/or/xor (zextload x), (zext cst))
4276 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
4277 N0.getOpcode() == ISD::XOR) &&
4278 isa<LoadSDNode>(N0.getOperand(0)) &&
4279 N0.getOperand(1).getOpcode() == ISD::Constant &&
4280 TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()) &&
4281 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
4282 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
4283 if (LN0->getExtensionType() != ISD::SEXTLOAD) {
4284 bool DoXform = true;
4285 SmallVector<SDNode*, 4> SetCCs;
4286 if (!N0.hasOneUse())
4287 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::ZERO_EXTEND,
4290 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, LN0->getDebugLoc(), VT,
4291 LN0->getChain(), LN0->getBasePtr(),
4292 LN0->getPointerInfo(),
4295 LN0->isNonTemporal(),
4296 LN0->getAlignment());
4297 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
4298 Mask = Mask.zext(VT.getSizeInBits());
4299 SDValue And = DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT,
4300 ExtLoad, DAG.getConstant(Mask, VT));
4301 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
4302 N0.getOperand(0).getDebugLoc(),
4303 N0.getOperand(0).getValueType(), ExtLoad);
4305 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
4306 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, N->getDebugLoc(),
4308 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4313 // fold (zext (zextload x)) -> (zext (truncate (zextload x)))
4314 // fold (zext ( extload x)) -> (zext (truncate (zextload x)))
4315 if ((ISD::isZEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
4316 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
4317 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4318 EVT MemVT = LN0->getMemoryVT();
4319 if ((!LegalOperations && !LN0->isVolatile()) ||
4320 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT)) {
4321 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N->getDebugLoc(), VT,
4323 LN0->getBasePtr(), LN0->getPointerInfo(),
4325 LN0->isVolatile(), LN0->isNonTemporal(),
4326 LN0->getAlignment());
4327 CombineTo(N, ExtLoad);
4328 CombineTo(N0.getNode(),
4329 DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(), N0.getValueType(),
4331 ExtLoad.getValue(1));
4332 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4336 if (N0.getOpcode() == ISD::SETCC) {
4337 if (!LegalOperations && VT.isVector()) {
4338 // zext(setcc) -> (and (vsetcc), (1, 1, ...) for vectors.
4339 // Only do this before legalize for now.
4340 EVT N0VT = N0.getOperand(0).getValueType();
4341 EVT EltVT = VT.getVectorElementType();
4342 SmallVector<SDValue,8> OneOps(VT.getVectorNumElements(),
4343 DAG.getConstant(1, EltVT));
4344 if (VT.getSizeInBits() == N0VT.getSizeInBits())
4345 // We know that the # elements of the results is the same as the
4346 // # elements of the compare (and the # elements of the compare result
4347 // for that matter). Check to see that they are the same size. If so,
4348 // we know that the element size of the sext'd result matches the
4349 // element size of the compare operands.
4350 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
4351 DAG.getSetCC(N->getDebugLoc(), VT, N0.getOperand(0),
4353 cast<CondCodeSDNode>(N0.getOperand(2))->get()),
4354 DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), VT,
4355 &OneOps[0], OneOps.size()));
4357 // If the desired elements are smaller or larger than the source
4358 // elements we can use a matching integer vector type and then
4359 // truncate/sign extend
4360 EVT MatchingElementType =
4361 EVT::getIntegerVT(*DAG.getContext(),
4362 N0VT.getScalarType().getSizeInBits());
4363 EVT MatchingVectorType =
4364 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
4365 N0VT.getVectorNumElements());
4367 DAG.getSetCC(N->getDebugLoc(), MatchingVectorType, N0.getOperand(0),
4369 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4370 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
4371 DAG.getSExtOrTrunc(VsetCC, N->getDebugLoc(), VT),
4372 DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), VT,
4373 &OneOps[0], OneOps.size()));
4376 // zext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
4378 SimplifySelectCC(N->getDebugLoc(), N0.getOperand(0), N0.getOperand(1),
4379 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
4380 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
4381 if (SCC.getNode()) return SCC;
4384 // (zext (shl (zext x), cst)) -> (shl (zext x), cst)
4385 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL) &&
4386 isa<ConstantSDNode>(N0.getOperand(1)) &&
4387 N0.getOperand(0).getOpcode() == ISD::ZERO_EXTEND &&
4389 SDValue ShAmt = N0.getOperand(1);
4390 unsigned ShAmtVal = cast<ConstantSDNode>(ShAmt)->getZExtValue();
4391 if (N0.getOpcode() == ISD::SHL) {
4392 SDValue InnerZExt = N0.getOperand(0);
4393 // If the original shl may be shifting out bits, do not perform this
4395 unsigned KnownZeroBits = InnerZExt.getValueType().getSizeInBits() -
4396 InnerZExt.getOperand(0).getValueType().getSizeInBits();
4397 if (ShAmtVal > KnownZeroBits)
4401 DebugLoc DL = N->getDebugLoc();
4403 // Ensure that the shift amount is wide enough for the shifted value.
4404 if (VT.getSizeInBits() >= 256)
4405 ShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, ShAmt);
4407 return DAG.getNode(N0.getOpcode(), DL, VT,
4408 DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0)),
4415 SDValue DAGCombiner::visitANY_EXTEND(SDNode *N) {
4416 SDValue N0 = N->getOperand(0);
4417 EVT VT = N->getValueType(0);
4419 // fold (aext c1) -> c1
4420 if (isa<ConstantSDNode>(N0))
4421 return DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, N0);
4422 // fold (aext (aext x)) -> (aext x)
4423 // fold (aext (zext x)) -> (zext x)
4424 // fold (aext (sext x)) -> (sext x)
4425 if (N0.getOpcode() == ISD::ANY_EXTEND ||
4426 N0.getOpcode() == ISD::ZERO_EXTEND ||
4427 N0.getOpcode() == ISD::SIGN_EXTEND)
4428 return DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT, N0.getOperand(0));
4430 // fold (aext (truncate (load x))) -> (aext (smaller load x))
4431 // fold (aext (truncate (srl (load x), c))) -> (aext (small load (x+c/n)))
4432 if (N0.getOpcode() == ISD::TRUNCATE) {
4433 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
4434 if (NarrowLoad.getNode()) {
4435 SDNode* oye = N0.getNode()->getOperand(0).getNode();
4436 if (NarrowLoad.getNode() != N0.getNode()) {
4437 CombineTo(N0.getNode(), NarrowLoad);
4438 // CombineTo deleted the truncate, if needed, but not what's under it.
4441 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4445 // fold (aext (truncate x))
4446 if (N0.getOpcode() == ISD::TRUNCATE) {
4447 SDValue TruncOp = N0.getOperand(0);
4448 if (TruncOp.getValueType() == VT)
4449 return TruncOp; // x iff x size == zext size.
4450 if (TruncOp.getValueType().bitsGT(VT))
4451 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, TruncOp);
4452 return DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, TruncOp);
4455 // Fold (aext (and (trunc x), cst)) -> (and x, cst)
4456 // if the trunc is not free.
4457 if (N0.getOpcode() == ISD::AND &&
4458 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
4459 N0.getOperand(1).getOpcode() == ISD::Constant &&
4460 !TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
4461 N0.getValueType())) {
4462 SDValue X = N0.getOperand(0).getOperand(0);
4463 if (X.getValueType().bitsLT(VT)) {
4464 X = DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, X);
4465 } else if (X.getValueType().bitsGT(VT)) {
4466 X = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, X);
4468 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
4469 Mask = Mask.zext(VT.getSizeInBits());
4470 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
4471 X, DAG.getConstant(Mask, VT));
4474 // fold (aext (load x)) -> (aext (truncate (extload x)))
4475 // None of the supported targets knows how to perform load and any_ext
4476 // on vectors in one instruction. We only perform this transformation on
4478 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
4479 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
4480 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType()))) {
4481 bool DoXform = true;
4482 SmallVector<SDNode*, 4> SetCCs;
4483 if (!N0.hasOneUse())
4484 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ANY_EXTEND, SetCCs, TLI);
4486 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4487 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, N->getDebugLoc(), VT,
4489 LN0->getBasePtr(), LN0->getPointerInfo(),
4491 LN0->isVolatile(), LN0->isNonTemporal(),
4492 LN0->getAlignment());
4493 CombineTo(N, ExtLoad);
4494 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4495 N0.getValueType(), ExtLoad);
4496 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
4497 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, N->getDebugLoc(),
4499 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4503 // fold (aext (zextload x)) -> (aext (truncate (zextload x)))
4504 // fold (aext (sextload x)) -> (aext (truncate (sextload x)))
4505 // fold (aext ( extload x)) -> (aext (truncate (extload x)))
4506 if (N0.getOpcode() == ISD::LOAD &&
4507 !ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
4509 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4510 EVT MemVT = LN0->getMemoryVT();
4511 SDValue ExtLoad = DAG.getExtLoad(LN0->getExtensionType(), N->getDebugLoc(),
4512 VT, LN0->getChain(), LN0->getBasePtr(),
4513 LN0->getPointerInfo(), MemVT,
4514 LN0->isVolatile(), LN0->isNonTemporal(),
4515 LN0->getAlignment());
4516 CombineTo(N, ExtLoad);
4517 CombineTo(N0.getNode(),
4518 DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4519 N0.getValueType(), ExtLoad),
4520 ExtLoad.getValue(1));
4521 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4524 if (N0.getOpcode() == ISD::SETCC) {
4525 // aext(setcc) -> sext_in_reg(vsetcc) for vectors.
4526 // Only do this before legalize for now.
4527 if (VT.isVector() && !LegalOperations) {
4528 EVT N0VT = N0.getOperand(0).getValueType();
4529 // We know that the # elements of the results is the same as the
4530 // # elements of the compare (and the # elements of the compare result
4531 // for that matter). Check to see that they are the same size. If so,
4532 // we know that the element size of the sext'd result matches the
4533 // element size of the compare operands.
4534 if (VT.getSizeInBits() == N0VT.getSizeInBits())
4535 return DAG.getSetCC(N->getDebugLoc(), VT, N0.getOperand(0),
4537 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4538 // If the desired elements are smaller or larger than the source
4539 // elements we can use a matching integer vector type and then
4540 // truncate/sign extend
4542 EVT MatchingElementType =
4543 EVT::getIntegerVT(*DAG.getContext(),
4544 N0VT.getScalarType().getSizeInBits());
4545 EVT MatchingVectorType =
4546 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
4547 N0VT.getVectorNumElements());
4549 DAG.getSetCC(N->getDebugLoc(), MatchingVectorType, N0.getOperand(0),
4551 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4552 return DAG.getSExtOrTrunc(VsetCC, N->getDebugLoc(), VT);
4556 // aext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
4558 SimplifySelectCC(N->getDebugLoc(), N0.getOperand(0), N0.getOperand(1),
4559 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
4560 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
4568 /// GetDemandedBits - See if the specified operand can be simplified with the
4569 /// knowledge that only the bits specified by Mask are used. If so, return the
4570 /// simpler operand, otherwise return a null SDValue.
4571 SDValue DAGCombiner::GetDemandedBits(SDValue V, const APInt &Mask) {
4572 switch (V.getOpcode()) {
4574 case ISD::Constant: {
4575 const ConstantSDNode *CV = cast<ConstantSDNode>(V.getNode());
4576 assert(CV != 0 && "Const value should be ConstSDNode.");
4577 const APInt &CVal = CV->getAPIntValue();
4578 APInt NewVal = CVal & Mask;
4579 if (NewVal != CVal) {
4580 return DAG.getConstant(NewVal, V.getValueType());
4586 // If the LHS or RHS don't contribute bits to the or, drop them.
4587 if (DAG.MaskedValueIsZero(V.getOperand(0), Mask))
4588 return V.getOperand(1);
4589 if (DAG.MaskedValueIsZero(V.getOperand(1), Mask))
4590 return V.getOperand(0);
4593 // Only look at single-use SRLs.
4594 if (!V.getNode()->hasOneUse())
4596 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(V.getOperand(1))) {
4597 // See if we can recursively simplify the LHS.
4598 unsigned Amt = RHSC->getZExtValue();
4600 // Watch out for shift count overflow though.
4601 if (Amt >= Mask.getBitWidth()) break;
4602 APInt NewMask = Mask << Amt;
4603 SDValue SimplifyLHS = GetDemandedBits(V.getOperand(0), NewMask);
4604 if (SimplifyLHS.getNode())
4605 return DAG.getNode(ISD::SRL, V.getDebugLoc(), V.getValueType(),
4606 SimplifyLHS, V.getOperand(1));
4612 /// ReduceLoadWidth - If the result of a wider load is shifted to right of N
4613 /// bits and then truncated to a narrower type and where N is a multiple
4614 /// of number of bits of the narrower type, transform it to a narrower load
4615 /// from address + N / num of bits of new type. If the result is to be
4616 /// extended, also fold the extension to form a extending load.
4617 SDValue DAGCombiner::ReduceLoadWidth(SDNode *N) {
4618 unsigned Opc = N->getOpcode();
4620 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
4621 SDValue N0 = N->getOperand(0);
4622 EVT VT = N->getValueType(0);
4625 // This transformation isn't valid for vector loads.
4629 // Special case: SIGN_EXTEND_INREG is basically truncating to ExtVT then
4631 if (Opc == ISD::SIGN_EXTEND_INREG) {
4632 ExtType = ISD::SEXTLOAD;
4633 ExtVT = cast<VTSDNode>(N->getOperand(1))->getVT();
4634 } else if (Opc == ISD::SRL) {
4635 // Another special-case: SRL is basically zero-extending a narrower value.
4636 ExtType = ISD::ZEXTLOAD;
4638 ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1));
4639 if (!N01) return SDValue();
4640 ExtVT = EVT::getIntegerVT(*DAG.getContext(),
4641 VT.getSizeInBits() - N01->getZExtValue());
4643 if (LegalOperations && !TLI.isLoadExtLegal(ExtType, ExtVT))
4646 unsigned EVTBits = ExtVT.getSizeInBits();
4648 // Do not generate loads of non-round integer types since these can
4649 // be expensive (and would be wrong if the type is not byte sized).
4650 if (!ExtVT.isRound())
4654 if (N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
4655 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
4656 ShAmt = N01->getZExtValue();
4657 // Is the shift amount a multiple of size of VT?
4658 if ((ShAmt & (EVTBits-1)) == 0) {
4659 N0 = N0.getOperand(0);
4660 // Is the load width a multiple of size of VT?
4661 if ((N0.getValueType().getSizeInBits() & (EVTBits-1)) != 0)
4665 // At this point, we must have a load or else we can't do the transform.
4666 if (!isa<LoadSDNode>(N0)) return SDValue();
4668 // If the shift amount is larger than the input type then we're not
4669 // accessing any of the loaded bytes. If the load was a zextload/extload
4670 // then the result of the shift+trunc is zero/undef (handled elsewhere).
4671 // If the load was a sextload then the result is a splat of the sign bit
4672 // of the extended byte. This is not worth optimizing for.
4673 if (ShAmt >= cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits())
4678 // If the load is shifted left (and the result isn't shifted back right),
4679 // we can fold the truncate through the shift.
4680 unsigned ShLeftAmt = 0;
4681 if (ShAmt == 0 && N0.getOpcode() == ISD::SHL && N0.hasOneUse() &&
4682 ExtVT == VT && TLI.isNarrowingProfitable(N0.getValueType(), VT)) {
4683 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
4684 ShLeftAmt = N01->getZExtValue();
4685 N0 = N0.getOperand(0);
4689 // If we haven't found a load, we can't narrow it. Don't transform one with
4690 // multiple uses, this would require adding a new load.
4691 if (!isa<LoadSDNode>(N0) || !N0.hasOneUse() ||
4692 // Don't change the width of a volatile load.
4693 cast<LoadSDNode>(N0)->isVolatile())
4696 // Verify that we are actually reducing a load width here.
4697 if (cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits() < EVTBits)
4700 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4701 EVT PtrType = N0.getOperand(1).getValueType();
4703 // For big endian targets, we need to adjust the offset to the pointer to
4704 // load the correct bytes.
4705 if (TLI.isBigEndian()) {
4706 unsigned LVTStoreBits = LN0->getMemoryVT().getStoreSizeInBits();
4707 unsigned EVTStoreBits = ExtVT.getStoreSizeInBits();
4708 ShAmt = LVTStoreBits - EVTStoreBits - ShAmt;
4711 uint64_t PtrOff = ShAmt / 8;
4712 unsigned NewAlign = MinAlign(LN0->getAlignment(), PtrOff);
4713 SDValue NewPtr = DAG.getNode(ISD::ADD, LN0->getDebugLoc(),
4714 PtrType, LN0->getBasePtr(),
4715 DAG.getConstant(PtrOff, PtrType));
4716 AddToWorkList(NewPtr.getNode());
4719 if (ExtType == ISD::NON_EXTLOAD)
4720 Load = DAG.getLoad(VT, N0.getDebugLoc(), LN0->getChain(), NewPtr,
4721 LN0->getPointerInfo().getWithOffset(PtrOff),
4722 LN0->isVolatile(), LN0->isNonTemporal(),
4723 LN0->isInvariant(), NewAlign);
4725 Load = DAG.getExtLoad(ExtType, N0.getDebugLoc(), VT, LN0->getChain(),NewPtr,
4726 LN0->getPointerInfo().getWithOffset(PtrOff),
4727 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
4730 // Replace the old load's chain with the new load's chain.
4731 WorkListRemover DeadNodes(*this);
4732 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1),
4735 // Shift the result left, if we've swallowed a left shift.
4736 SDValue Result = Load;
4737 if (ShLeftAmt != 0) {
4738 EVT ShImmTy = getShiftAmountTy(Result.getValueType());
4739 if (!isUIntN(ShImmTy.getSizeInBits(), ShLeftAmt))
4741 Result = DAG.getNode(ISD::SHL, N0.getDebugLoc(), VT,
4742 Result, DAG.getConstant(ShLeftAmt, ShImmTy));
4745 // Return the new loaded value.
4749 SDValue DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
4750 SDValue N0 = N->getOperand(0);
4751 SDValue N1 = N->getOperand(1);
4752 EVT VT = N->getValueType(0);
4753 EVT EVT = cast<VTSDNode>(N1)->getVT();
4754 unsigned VTBits = VT.getScalarType().getSizeInBits();
4755 unsigned EVTBits = EVT.getScalarType().getSizeInBits();
4757 // fold (sext_in_reg c1) -> c1
4758 if (isa<ConstantSDNode>(N0) || N0.getOpcode() == ISD::UNDEF)
4759 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT, N0, N1);
4761 // If the input is already sign extended, just drop the extension.
4762 if (DAG.ComputeNumSignBits(N0) >= VTBits-EVTBits+1)
4765 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2
4766 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
4767 EVT.bitsLT(cast<VTSDNode>(N0.getOperand(1))->getVT())) {
4768 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT,
4769 N0.getOperand(0), N1);
4772 // fold (sext_in_reg (sext x)) -> (sext x)
4773 // fold (sext_in_reg (aext x)) -> (sext x)
4774 // if x is small enough.
4775 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) {
4776 SDValue N00 = N0.getOperand(0);
4777 if (N00.getValueType().getScalarType().getSizeInBits() <= EVTBits &&
4778 (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND, VT)))
4779 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, N00, N1);
4782 // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is known zero.
4783 if (DAG.MaskedValueIsZero(N0, APInt::getBitsSet(VTBits, EVTBits-1, EVTBits)))
4784 return DAG.getZeroExtendInReg(N0, N->getDebugLoc(), EVT);
4786 // fold operands of sext_in_reg based on knowledge that the top bits are not
4788 if (SimplifyDemandedBits(SDValue(N, 0)))
4789 return SDValue(N, 0);
4791 // fold (sext_in_reg (load x)) -> (smaller sextload x)
4792 // fold (sext_in_reg (srl (load x), c)) -> (smaller sextload (x+c/evtbits))
4793 SDValue NarrowLoad = ReduceLoadWidth(N);
4794 if (NarrowLoad.getNode())
4797 // fold (sext_in_reg (srl X, 24), i8) -> (sra X, 24)
4798 // fold (sext_in_reg (srl X, 23), i8) -> (sra X, 23) iff possible.
4799 // We already fold "(sext_in_reg (srl X, 25), i8) -> srl X, 25" above.
4800 if (N0.getOpcode() == ISD::SRL) {
4801 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
4802 if (ShAmt->getZExtValue()+EVTBits <= VTBits) {
4803 // We can turn this into an SRA iff the input to the SRL is already sign
4805 unsigned InSignBits = DAG.ComputeNumSignBits(N0.getOperand(0));
4806 if (VTBits-(ShAmt->getZExtValue()+EVTBits) < InSignBits)
4807 return DAG.getNode(ISD::SRA, N->getDebugLoc(), VT,
4808 N0.getOperand(0), N0.getOperand(1));
4812 // fold (sext_inreg (extload x)) -> (sextload x)
4813 if (ISD::isEXTLoad(N0.getNode()) &&
4814 ISD::isUNINDEXEDLoad(N0.getNode()) &&
4815 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
4816 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
4817 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) {
4818 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4819 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
4821 LN0->getBasePtr(), LN0->getPointerInfo(),
4823 LN0->isVolatile(), LN0->isNonTemporal(),
4824 LN0->getAlignment());
4825 CombineTo(N, ExtLoad);
4826 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
4827 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4829 // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use
4830 if (ISD::isZEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
4832 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
4833 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
4834 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) {
4835 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4836 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
4838 LN0->getBasePtr(), LN0->getPointerInfo(),
4840 LN0->isVolatile(), LN0->isNonTemporal(),
4841 LN0->getAlignment());
4842 CombineTo(N, ExtLoad);
4843 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
4844 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4847 // Form (sext_inreg (bswap >> 16)) or (sext_inreg (rotl (bswap) 16))
4848 if (EVTBits <= 16 && N0.getOpcode() == ISD::OR) {
4849 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
4850 N0.getOperand(1), false);
4851 if (BSwap.getNode() != 0)
4852 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT,
4859 SDValue DAGCombiner::visitTRUNCATE(SDNode *N) {
4860 SDValue N0 = N->getOperand(0);
4861 EVT VT = N->getValueType(0);
4864 if (N0.getValueType() == N->getValueType(0))
4866 // fold (truncate c1) -> c1
4867 if (isa<ConstantSDNode>(N0))
4868 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, N0);
4869 // fold (truncate (truncate x)) -> (truncate x)
4870 if (N0.getOpcode() == ISD::TRUNCATE)
4871 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, N0.getOperand(0));
4872 // fold (truncate (ext x)) -> (ext x) or (truncate x) or x
4873 if (N0.getOpcode() == ISD::ZERO_EXTEND ||
4874 N0.getOpcode() == ISD::SIGN_EXTEND ||
4875 N0.getOpcode() == ISD::ANY_EXTEND) {
4876 if (N0.getOperand(0).getValueType().bitsLT(VT))
4877 // if the source is smaller than the dest, we still need an extend
4878 return DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT,
4880 else if (N0.getOperand(0).getValueType().bitsGT(VT))
4881 // if the source is larger than the dest, than we just need the truncate
4882 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, N0.getOperand(0));
4884 // if the source and dest are the same type, we can drop both the extend
4885 // and the truncate.
4886 return N0.getOperand(0);
4889 // See if we can simplify the input to this truncate through knowledge that
4890 // only the low bits are being used.
4891 // For example "trunc (or (shl x, 8), y)" // -> trunc y
4892 // Currently we only perform this optimization on scalars because vectors
4893 // may have different active low bits.
4894 if (!VT.isVector()) {
4896 GetDemandedBits(N0, APInt::getLowBitsSet(N0.getValueSizeInBits(),
4897 VT.getSizeInBits()));
4898 if (Shorter.getNode())
4899 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, Shorter);
4901 // fold (truncate (load x)) -> (smaller load x)
4902 // fold (truncate (srl (load x), c)) -> (smaller load (x+c/evtbits))
4903 if (!LegalTypes || TLI.isTypeDesirableForOp(N0.getOpcode(), VT)) {
4904 SDValue Reduced = ReduceLoadWidth(N);
4905 if (Reduced.getNode())
4909 // Simplify the operands using demanded-bits information.
4910 if (!VT.isVector() &&
4911 SimplifyDemandedBits(SDValue(N, 0)))
4912 return SDValue(N, 0);
4917 static SDNode *getBuildPairElt(SDNode *N, unsigned i) {
4918 SDValue Elt = N->getOperand(i);
4919 if (Elt.getOpcode() != ISD::MERGE_VALUES)
4920 return Elt.getNode();
4921 return Elt.getOperand(Elt.getResNo()).getNode();
4924 /// CombineConsecutiveLoads - build_pair (load, load) -> load
4925 /// if load locations are consecutive.
4926 SDValue DAGCombiner::CombineConsecutiveLoads(SDNode *N, EVT VT) {
4927 assert(N->getOpcode() == ISD::BUILD_PAIR);
4929 LoadSDNode *LD1 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 0));
4930 LoadSDNode *LD2 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 1));
4931 if (!LD1 || !LD2 || !ISD::isNON_EXTLoad(LD1) || !LD1->hasOneUse() ||
4932 LD1->getPointerInfo().getAddrSpace() !=
4933 LD2->getPointerInfo().getAddrSpace())
4935 EVT LD1VT = LD1->getValueType(0);
4937 if (ISD::isNON_EXTLoad(LD2) &&
4939 // If both are volatile this would reduce the number of volatile loads.
4940 // If one is volatile it might be ok, but play conservative and bail out.
4941 !LD1->isVolatile() &&
4942 !LD2->isVolatile() &&
4943 DAG.isConsecutiveLoad(LD2, LD1, LD1VT.getSizeInBits()/8, 1)) {
4944 unsigned Align = LD1->getAlignment();
4945 unsigned NewAlign = TLI.getTargetData()->
4946 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
4948 if (NewAlign <= Align &&
4949 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)))
4950 return DAG.getLoad(VT, N->getDebugLoc(), LD1->getChain(),
4951 LD1->getBasePtr(), LD1->getPointerInfo(),
4952 false, false, false, Align);
4958 SDValue DAGCombiner::visitBITCAST(SDNode *N) {
4959 SDValue N0 = N->getOperand(0);
4960 EVT VT = N->getValueType(0);
4962 // If the input is a BUILD_VECTOR with all constant elements, fold this now.
4963 // Only do this before legalize, since afterward the target may be depending
4964 // on the bitconvert.
4965 // First check to see if this is all constant.
4967 N0.getOpcode() == ISD::BUILD_VECTOR && N0.getNode()->hasOneUse() &&
4969 bool isSimple = true;
4970 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i)
4971 if (N0.getOperand(i).getOpcode() != ISD::UNDEF &&
4972 N0.getOperand(i).getOpcode() != ISD::Constant &&
4973 N0.getOperand(i).getOpcode() != ISD::ConstantFP) {
4978 EVT DestEltVT = N->getValueType(0).getVectorElementType();
4979 assert(!DestEltVT.isVector() &&
4980 "Element type of vector ValueType must not be vector!");
4982 return ConstantFoldBITCASTofBUILD_VECTOR(N0.getNode(), DestEltVT);
4985 // If the input is a constant, let getNode fold it.
4986 if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) {
4987 SDValue Res = DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, N0);
4988 if (Res.getNode() != N) {
4989 if (!LegalOperations ||
4990 TLI.isOperationLegal(Res.getNode()->getOpcode(), VT))
4993 // Folding it resulted in an illegal node, and it's too late to
4994 // do that. Clean up the old node and forego the transformation.
4995 // Ideally this won't happen very often, because instcombine
4996 // and the earlier dagcombine runs (where illegal nodes are
4997 // permitted) should have folded most of them already.
4998 DAG.DeleteNode(Res.getNode());
5002 // (conv (conv x, t1), t2) -> (conv x, t2)
5003 if (N0.getOpcode() == ISD::BITCAST)
5004 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT,
5007 // fold (conv (load x)) -> (load (conv*)x)
5008 // If the resultant load doesn't need a higher alignment than the original!
5009 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
5010 // Do not change the width of a volatile load.
5011 !cast<LoadSDNode>(N0)->isVolatile() &&
5012 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT))) {
5013 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5014 unsigned Align = TLI.getTargetData()->
5015 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
5016 unsigned OrigAlign = LN0->getAlignment();
5018 if (Align <= OrigAlign) {
5019 SDValue Load = DAG.getLoad(VT, N->getDebugLoc(), LN0->getChain(),
5020 LN0->getBasePtr(), LN0->getPointerInfo(),
5021 LN0->isVolatile(), LN0->isNonTemporal(),
5022 LN0->isInvariant(), OrigAlign);
5024 CombineTo(N0.getNode(),
5025 DAG.getNode(ISD::BITCAST, N0.getDebugLoc(),
5026 N0.getValueType(), Load),
5032 // fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit)
5033 // fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit))
5034 // This often reduces constant pool loads.
5035 if ((N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FABS) &&
5036 N0.getNode()->hasOneUse() && VT.isInteger() && !VT.isVector()) {
5037 SDValue NewConv = DAG.getNode(ISD::BITCAST, N0.getDebugLoc(), VT,
5039 AddToWorkList(NewConv.getNode());
5041 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
5042 if (N0.getOpcode() == ISD::FNEG)
5043 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT,
5044 NewConv, DAG.getConstant(SignBit, VT));
5045 assert(N0.getOpcode() == ISD::FABS);
5046 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
5047 NewConv, DAG.getConstant(~SignBit, VT));
5050 // fold (bitconvert (fcopysign cst, x)) ->
5051 // (or (and (bitconvert x), sign), (and cst, (not sign)))
5052 // Note that we don't handle (copysign x, cst) because this can always be
5053 // folded to an fneg or fabs.
5054 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse() &&
5055 isa<ConstantFPSDNode>(N0.getOperand(0)) &&
5056 VT.isInteger() && !VT.isVector()) {
5057 unsigned OrigXWidth = N0.getOperand(1).getValueType().getSizeInBits();
5058 EVT IntXVT = EVT::getIntegerVT(*DAG.getContext(), OrigXWidth);
5059 if (isTypeLegal(IntXVT)) {
5060 SDValue X = DAG.getNode(ISD::BITCAST, N0.getDebugLoc(),
5061 IntXVT, N0.getOperand(1));
5062 AddToWorkList(X.getNode());
5064 // If X has a different width than the result/lhs, sext it or truncate it.
5065 unsigned VTWidth = VT.getSizeInBits();
5066 if (OrigXWidth < VTWidth) {
5067 X = DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, X);
5068 AddToWorkList(X.getNode());
5069 } else if (OrigXWidth > VTWidth) {
5070 // To get the sign bit in the right place, we have to shift it right
5071 // before truncating.
5072 X = DAG.getNode(ISD::SRL, X.getDebugLoc(),
5073 X.getValueType(), X,
5074 DAG.getConstant(OrigXWidth-VTWidth, X.getValueType()));
5075 AddToWorkList(X.getNode());
5076 X = DAG.getNode(ISD::TRUNCATE, X.getDebugLoc(), VT, X);
5077 AddToWorkList(X.getNode());
5080 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
5081 X = DAG.getNode(ISD::AND, X.getDebugLoc(), VT,
5082 X, DAG.getConstant(SignBit, VT));
5083 AddToWorkList(X.getNode());
5085 SDValue Cst = DAG.getNode(ISD::BITCAST, N0.getDebugLoc(),
5086 VT, N0.getOperand(0));
5087 Cst = DAG.getNode(ISD::AND, Cst.getDebugLoc(), VT,
5088 Cst, DAG.getConstant(~SignBit, VT));
5089 AddToWorkList(Cst.getNode());
5091 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, X, Cst);
5095 // bitconvert(build_pair(ld, ld)) -> ld iff load locations are consecutive.
5096 if (N0.getOpcode() == ISD::BUILD_PAIR) {
5097 SDValue CombineLD = CombineConsecutiveLoads(N0.getNode(), VT);
5098 if (CombineLD.getNode())
5105 SDValue DAGCombiner::visitBUILD_PAIR(SDNode *N) {
5106 EVT VT = N->getValueType(0);
5107 return CombineConsecutiveLoads(N, VT);
5110 /// ConstantFoldBITCASTofBUILD_VECTOR - We know that BV is a build_vector
5111 /// node with Constant, ConstantFP or Undef operands. DstEltVT indicates the
5112 /// destination element value type.
5113 SDValue DAGCombiner::
5114 ConstantFoldBITCASTofBUILD_VECTOR(SDNode *BV, EVT DstEltVT) {
5115 EVT SrcEltVT = BV->getValueType(0).getVectorElementType();
5117 // If this is already the right type, we're done.
5118 if (SrcEltVT == DstEltVT) return SDValue(BV, 0);
5120 unsigned SrcBitSize = SrcEltVT.getSizeInBits();
5121 unsigned DstBitSize = DstEltVT.getSizeInBits();
5123 // If this is a conversion of N elements of one type to N elements of another
5124 // type, convert each element. This handles FP<->INT cases.
5125 if (SrcBitSize == DstBitSize) {
5126 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
5127 BV->getValueType(0).getVectorNumElements());
5129 // Due to the FP element handling below calling this routine recursively,
5130 // we can end up with a scalar-to-vector node here.
5131 if (BV->getOpcode() == ISD::SCALAR_TO_VECTOR)
5132 return DAG.getNode(ISD::SCALAR_TO_VECTOR, BV->getDebugLoc(), VT,
5133 DAG.getNode(ISD::BITCAST, BV->getDebugLoc(),
5134 DstEltVT, BV->getOperand(0)));
5136 SmallVector<SDValue, 8> Ops;
5137 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
5138 SDValue Op = BV->getOperand(i);
5139 // If the vector element type is not legal, the BUILD_VECTOR operands
5140 // are promoted and implicitly truncated. Make that explicit here.
5141 if (Op.getValueType() != SrcEltVT)
5142 Op = DAG.getNode(ISD::TRUNCATE, BV->getDebugLoc(), SrcEltVT, Op);
5143 Ops.push_back(DAG.getNode(ISD::BITCAST, BV->getDebugLoc(),
5145 AddToWorkList(Ops.back().getNode());
5147 return DAG.getNode(ISD::BUILD_VECTOR, BV->getDebugLoc(), VT,
5148 &Ops[0], Ops.size());
5151 // Otherwise, we're growing or shrinking the elements. To avoid having to
5152 // handle annoying details of growing/shrinking FP values, we convert them to
5154 if (SrcEltVT.isFloatingPoint()) {
5155 // Convert the input float vector to a int vector where the elements are the
5157 assert((SrcEltVT == MVT::f32 || SrcEltVT == MVT::f64) && "Unknown FP VT!");
5158 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), SrcEltVT.getSizeInBits());
5159 BV = ConstantFoldBITCASTofBUILD_VECTOR(BV, IntVT).getNode();
5163 // Now we know the input is an integer vector. If the output is a FP type,
5164 // convert to integer first, then to FP of the right size.
5165 if (DstEltVT.isFloatingPoint()) {
5166 assert((DstEltVT == MVT::f32 || DstEltVT == MVT::f64) && "Unknown FP VT!");
5167 EVT TmpVT = EVT::getIntegerVT(*DAG.getContext(), DstEltVT.getSizeInBits());
5168 SDNode *Tmp = ConstantFoldBITCASTofBUILD_VECTOR(BV, TmpVT).getNode();
5170 // Next, convert to FP elements of the same size.
5171 return ConstantFoldBITCASTofBUILD_VECTOR(Tmp, DstEltVT);
5174 // Okay, we know the src/dst types are both integers of differing types.
5175 // Handling growing first.
5176 assert(SrcEltVT.isInteger() && DstEltVT.isInteger());
5177 if (SrcBitSize < DstBitSize) {
5178 unsigned NumInputsPerOutput = DstBitSize/SrcBitSize;
5180 SmallVector<SDValue, 8> Ops;
5181 for (unsigned i = 0, e = BV->getNumOperands(); i != e;
5182 i += NumInputsPerOutput) {
5183 bool isLE = TLI.isLittleEndian();
5184 APInt NewBits = APInt(DstBitSize, 0);
5185 bool EltIsUndef = true;
5186 for (unsigned j = 0; j != NumInputsPerOutput; ++j) {
5187 // Shift the previously computed bits over.
5188 NewBits <<= SrcBitSize;
5189 SDValue Op = BV->getOperand(i+ (isLE ? (NumInputsPerOutput-j-1) : j));
5190 if (Op.getOpcode() == ISD::UNDEF) continue;
5193 NewBits |= cast<ConstantSDNode>(Op)->getAPIntValue().
5194 zextOrTrunc(SrcBitSize).zext(DstBitSize);
5198 Ops.push_back(DAG.getUNDEF(DstEltVT));
5200 Ops.push_back(DAG.getConstant(NewBits, DstEltVT));
5203 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, Ops.size());
5204 return DAG.getNode(ISD::BUILD_VECTOR, BV->getDebugLoc(), VT,
5205 &Ops[0], Ops.size());
5208 // Finally, this must be the case where we are shrinking elements: each input
5209 // turns into multiple outputs.
5210 bool isS2V = ISD::isScalarToVector(BV);
5211 unsigned NumOutputsPerInput = SrcBitSize/DstBitSize;
5212 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
5213 NumOutputsPerInput*BV->getNumOperands());
5214 SmallVector<SDValue, 8> Ops;
5216 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
5217 if (BV->getOperand(i).getOpcode() == ISD::UNDEF) {
5218 for (unsigned j = 0; j != NumOutputsPerInput; ++j)
5219 Ops.push_back(DAG.getUNDEF(DstEltVT));
5223 APInt OpVal = cast<ConstantSDNode>(BV->getOperand(i))->
5224 getAPIntValue().zextOrTrunc(SrcBitSize);
5226 for (unsigned j = 0; j != NumOutputsPerInput; ++j) {
5227 APInt ThisVal = OpVal.trunc(DstBitSize);
5228 Ops.push_back(DAG.getConstant(ThisVal, DstEltVT));
5229 if (isS2V && i == 0 && j == 0 && ThisVal.zext(SrcBitSize) == OpVal)
5230 // Simply turn this into a SCALAR_TO_VECTOR of the new type.
5231 return DAG.getNode(ISD::SCALAR_TO_VECTOR, BV->getDebugLoc(), VT,
5233 OpVal = OpVal.lshr(DstBitSize);
5236 // For big endian targets, swap the order of the pieces of each element.
5237 if (TLI.isBigEndian())
5238 std::reverse(Ops.end()-NumOutputsPerInput, Ops.end());
5241 return DAG.getNode(ISD::BUILD_VECTOR, BV->getDebugLoc(), VT,
5242 &Ops[0], Ops.size());
5245 SDValue DAGCombiner::visitFADD(SDNode *N) {
5246 SDValue N0 = N->getOperand(0);
5247 SDValue N1 = N->getOperand(1);
5248 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5249 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
5250 EVT VT = N->getValueType(0);
5253 if (VT.isVector()) {
5254 SDValue FoldedVOp = SimplifyVBinOp(N);
5255 if (FoldedVOp.getNode()) return FoldedVOp;
5258 // fold (fadd c1, c2) -> (fadd c1, c2)
5259 if (N0CFP && N1CFP && VT != MVT::ppcf128)
5260 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0, N1);
5261 // canonicalize constant to RHS
5262 if (N0CFP && !N1CFP)
5263 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N1, N0);
5264 // fold (fadd A, 0) -> A
5265 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
5266 N1CFP->getValueAPF().isZero())
5268 // fold (fadd A, (fneg B)) -> (fsub A, B)
5269 if (isNegatibleForFree(N1, LegalOperations, &DAG.getTarget().Options) == 2)
5270 return DAG.getNode(ISD::FSUB, N->getDebugLoc(), VT, N0,
5271 GetNegatedExpression(N1, DAG, LegalOperations));
5272 // fold (fadd (fneg A), B) -> (fsub B, A)
5273 if (isNegatibleForFree(N0, LegalOperations, &DAG.getTarget().Options) == 2)
5274 return DAG.getNode(ISD::FSUB, N->getDebugLoc(), VT, N1,
5275 GetNegatedExpression(N0, DAG, LegalOperations));
5277 // If allowed, fold (fadd (fadd x, c1), c2) -> (fadd x, (fadd c1, c2))
5278 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
5279 N0.getOpcode() == ISD::FADD && N0.getNode()->hasOneUse() &&
5280 isa<ConstantFPSDNode>(N0.getOperand(1)))
5281 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0.getOperand(0),
5282 DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
5283 N0.getOperand(1), N1));
5288 SDValue DAGCombiner::visitFSUB(SDNode *N) {
5289 SDValue N0 = N->getOperand(0);
5290 SDValue N1 = N->getOperand(1);
5291 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5292 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
5293 EVT VT = N->getValueType(0);
5296 if (VT.isVector()) {
5297 SDValue FoldedVOp = SimplifyVBinOp(N);
5298 if (FoldedVOp.getNode()) return FoldedVOp;
5301 // fold (fsub c1, c2) -> c1-c2
5302 if (N0CFP && N1CFP && VT != MVT::ppcf128)
5303 return DAG.getNode(ISD::FSUB, N->getDebugLoc(), VT, N0, N1);
5304 // fold (fsub A, 0) -> A
5305 if (DAG.getTarget().Options.UnsafeFPMath &&
5306 N1CFP && N1CFP->getValueAPF().isZero())
5308 // fold (fsub 0, B) -> -B
5309 if (DAG.getTarget().Options.UnsafeFPMath &&
5310 N0CFP && N0CFP->getValueAPF().isZero()) {
5311 if (isNegatibleForFree(N1, LegalOperations, &DAG.getTarget().Options))
5312 return GetNegatedExpression(N1, DAG, LegalOperations);
5313 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
5314 return DAG.getNode(ISD::FNEG, N->getDebugLoc(), VT, N1);
5316 // fold (fsub A, (fneg B)) -> (fadd A, B)
5317 if (isNegatibleForFree(N1, LegalOperations, &DAG.getTarget().Options))
5318 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0,
5319 GetNegatedExpression(N1, DAG, LegalOperations));
5324 SDValue DAGCombiner::visitFMUL(SDNode *N) {
5325 SDValue N0 = N->getOperand(0);
5326 SDValue N1 = N->getOperand(1);
5327 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5328 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
5329 EVT VT = N->getValueType(0);
5332 if (VT.isVector()) {
5333 SDValue FoldedVOp = SimplifyVBinOp(N);
5334 if (FoldedVOp.getNode()) return FoldedVOp;
5337 // fold (fmul c1, c2) -> c1*c2
5338 if (N0CFP && N1CFP && VT != MVT::ppcf128)
5339 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT, N0, N1);
5340 // canonicalize constant to RHS
5341 if (N0CFP && !N1CFP)
5342 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT, N1, N0);
5343 // fold (fmul A, 0) -> 0
5344 if (DAG.getTarget().Options.UnsafeFPMath &&
5345 N1CFP && N1CFP->getValueAPF().isZero())
5347 // fold (fmul A, 0) -> 0, vector edition.
5348 if (DAG.getTarget().Options.UnsafeFPMath &&
5349 ISD::isBuildVectorAllZeros(N1.getNode()))
5351 // fold (fmul X, 2.0) -> (fadd X, X)
5352 if (N1CFP && N1CFP->isExactlyValue(+2.0))
5353 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0, N0);
5354 // fold (fmul X, -1.0) -> (fneg X)
5355 if (N1CFP && N1CFP->isExactlyValue(-1.0))
5356 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
5357 return DAG.getNode(ISD::FNEG, N->getDebugLoc(), VT, N0);
5359 // fold (fmul (fneg X), (fneg Y)) -> (fmul X, Y)
5360 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations,
5361 &DAG.getTarget().Options)) {
5362 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations,
5363 &DAG.getTarget().Options)) {
5364 // Both can be negated for free, check to see if at least one is cheaper
5366 if (LHSNeg == 2 || RHSNeg == 2)
5367 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5368 GetNegatedExpression(N0, DAG, LegalOperations),
5369 GetNegatedExpression(N1, DAG, LegalOperations));
5373 // If allowed, fold (fmul (fmul x, c1), c2) -> (fmul x, (fmul c1, c2))
5374 if (DAG.getTarget().Options.UnsafeFPMath &&
5375 N1CFP && N0.getOpcode() == ISD::FMUL &&
5376 N0.getNode()->hasOneUse() && isa<ConstantFPSDNode>(N0.getOperand(1)))
5377 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT, N0.getOperand(0),
5378 DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5379 N0.getOperand(1), N1));
5384 SDValue DAGCombiner::visitFDIV(SDNode *N) {
5385 SDValue N0 = N->getOperand(0);
5386 SDValue N1 = N->getOperand(1);
5387 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5388 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
5389 EVT VT = N->getValueType(0);
5392 if (VT.isVector()) {
5393 SDValue FoldedVOp = SimplifyVBinOp(N);
5394 if (FoldedVOp.getNode()) return FoldedVOp;
5397 // fold (fdiv c1, c2) -> c1/c2
5398 if (N0CFP && N1CFP && VT != MVT::ppcf128)
5399 return DAG.getNode(ISD::FDIV, N->getDebugLoc(), VT, N0, N1);
5402 // (fdiv (fneg X), (fneg Y)) -> (fdiv X, Y)
5403 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations,
5404 &DAG.getTarget().Options)) {
5405 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations,
5406 &DAG.getTarget().Options)) {
5407 // Both can be negated for free, check to see if at least one is cheaper
5409 if (LHSNeg == 2 || RHSNeg == 2)
5410 return DAG.getNode(ISD::FDIV, N->getDebugLoc(), VT,
5411 GetNegatedExpression(N0, DAG, LegalOperations),
5412 GetNegatedExpression(N1, DAG, LegalOperations));
5419 SDValue DAGCombiner::visitFREM(SDNode *N) {
5420 SDValue N0 = N->getOperand(0);
5421 SDValue N1 = N->getOperand(1);
5422 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5423 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
5424 EVT VT = N->getValueType(0);
5426 // fold (frem c1, c2) -> fmod(c1,c2)
5427 if (N0CFP && N1CFP && VT != MVT::ppcf128)
5428 return DAG.getNode(ISD::FREM, N->getDebugLoc(), VT, N0, N1);
5433 SDValue DAGCombiner::visitFCOPYSIGN(SDNode *N) {
5434 SDValue N0 = N->getOperand(0);
5435 SDValue N1 = N->getOperand(1);
5436 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5437 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
5438 EVT VT = N->getValueType(0);
5440 if (N0CFP && N1CFP && VT != MVT::ppcf128) // Constant fold
5441 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT, N0, N1);
5444 const APFloat& V = N1CFP->getValueAPF();
5445 // copysign(x, c1) -> fabs(x) iff ispos(c1)
5446 // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1)
5447 if (!V.isNegative()) {
5448 if (!LegalOperations || TLI.isOperationLegal(ISD::FABS, VT))
5449 return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0);
5451 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
5452 return DAG.getNode(ISD::FNEG, N->getDebugLoc(), VT,
5453 DAG.getNode(ISD::FABS, N0.getDebugLoc(), VT, N0));
5457 // copysign(fabs(x), y) -> copysign(x, y)
5458 // copysign(fneg(x), y) -> copysign(x, y)
5459 // copysign(copysign(x,z), y) -> copysign(x, y)
5460 if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG ||
5461 N0.getOpcode() == ISD::FCOPYSIGN)
5462 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
5463 N0.getOperand(0), N1);
5465 // copysign(x, abs(y)) -> abs(x)
5466 if (N1.getOpcode() == ISD::FABS)
5467 return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0);
5469 // copysign(x, copysign(y,z)) -> copysign(x, z)
5470 if (N1.getOpcode() == ISD::FCOPYSIGN)
5471 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
5472 N0, N1.getOperand(1));
5474 // copysign(x, fp_extend(y)) -> copysign(x, y)
5475 // copysign(x, fp_round(y)) -> copysign(x, y)
5476 if (N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND)
5477 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
5478 N0, N1.getOperand(0));
5483 SDValue DAGCombiner::visitSINT_TO_FP(SDNode *N) {
5484 SDValue N0 = N->getOperand(0);
5485 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
5486 EVT VT = N->getValueType(0);
5487 EVT OpVT = N0.getValueType();
5489 // fold (sint_to_fp c1) -> c1fp
5490 if (N0C && OpVT != MVT::ppcf128 &&
5491 // ...but only if the target supports immediate floating-point values
5492 (!LegalOperations ||
5493 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
5494 return DAG.getNode(ISD::SINT_TO_FP, N->getDebugLoc(), VT, N0);
5496 // If the input is a legal type, and SINT_TO_FP is not legal on this target,
5497 // but UINT_TO_FP is legal on this target, try to convert.
5498 if (!TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT) &&
5499 TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT)) {
5500 // If the sign bit is known to be zero, we can change this to UINT_TO_FP.
5501 if (DAG.SignBitIsZero(N0))
5502 return DAG.getNode(ISD::UINT_TO_FP, N->getDebugLoc(), VT, N0);
5508 SDValue DAGCombiner::visitUINT_TO_FP(SDNode *N) {
5509 SDValue N0 = N->getOperand(0);
5510 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
5511 EVT VT = N->getValueType(0);
5512 EVT OpVT = N0.getValueType();
5514 // fold (uint_to_fp c1) -> c1fp
5515 if (N0C && OpVT != MVT::ppcf128 &&
5516 // ...but only if the target supports immediate floating-point values
5517 (!LegalOperations ||
5518 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
5519 return DAG.getNode(ISD::UINT_TO_FP, N->getDebugLoc(), VT, N0);
5521 // If the input is a legal type, and UINT_TO_FP is not legal on this target,
5522 // but SINT_TO_FP is legal on this target, try to convert.
5523 if (!TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT) &&
5524 TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT)) {
5525 // If the sign bit is known to be zero, we can change this to SINT_TO_FP.
5526 if (DAG.SignBitIsZero(N0))
5527 return DAG.getNode(ISD::SINT_TO_FP, N->getDebugLoc(), VT, N0);
5533 SDValue DAGCombiner::visitFP_TO_SINT(SDNode *N) {
5534 SDValue N0 = N->getOperand(0);
5535 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5536 EVT VT = N->getValueType(0);
5538 // fold (fp_to_sint c1fp) -> c1
5540 return DAG.getNode(ISD::FP_TO_SINT, N->getDebugLoc(), VT, N0);
5545 SDValue DAGCombiner::visitFP_TO_UINT(SDNode *N) {
5546 SDValue N0 = N->getOperand(0);
5547 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5548 EVT VT = N->getValueType(0);
5550 // fold (fp_to_uint c1fp) -> c1
5551 if (N0CFP && VT != MVT::ppcf128)
5552 return DAG.getNode(ISD::FP_TO_UINT, N->getDebugLoc(), VT, N0);
5557 SDValue DAGCombiner::visitFP_ROUND(SDNode *N) {
5558 SDValue N0 = N->getOperand(0);
5559 SDValue N1 = N->getOperand(1);
5560 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5561 EVT VT = N->getValueType(0);
5563 // fold (fp_round c1fp) -> c1fp
5564 if (N0CFP && N0.getValueType() != MVT::ppcf128)
5565 return DAG.getNode(ISD::FP_ROUND, N->getDebugLoc(), VT, N0, N1);
5567 // fold (fp_round (fp_extend x)) -> x
5568 if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType())
5569 return N0.getOperand(0);
5571 // fold (fp_round (fp_round x)) -> (fp_round x)
5572 if (N0.getOpcode() == ISD::FP_ROUND) {
5573 // This is a value preserving truncation if both round's are.
5574 bool IsTrunc = N->getConstantOperandVal(1) == 1 &&
5575 N0.getNode()->getConstantOperandVal(1) == 1;
5576 return DAG.getNode(ISD::FP_ROUND, N->getDebugLoc(), VT, N0.getOperand(0),
5577 DAG.getIntPtrConstant(IsTrunc));
5580 // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y)
5581 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse()) {
5582 SDValue Tmp = DAG.getNode(ISD::FP_ROUND, N0.getDebugLoc(), VT,
5583 N0.getOperand(0), N1);
5584 AddToWorkList(Tmp.getNode());
5585 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
5586 Tmp, N0.getOperand(1));
5592 SDValue DAGCombiner::visitFP_ROUND_INREG(SDNode *N) {
5593 SDValue N0 = N->getOperand(0);
5594 EVT VT = N->getValueType(0);
5595 EVT EVT = cast<VTSDNode>(N->getOperand(1))->getVT();
5596 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5598 // fold (fp_round_inreg c1fp) -> c1fp
5599 if (N0CFP && isTypeLegal(EVT)) {
5600 SDValue Round = DAG.getConstantFP(*N0CFP->getConstantFPValue(), EVT);
5601 return DAG.getNode(ISD::FP_EXTEND, N->getDebugLoc(), VT, Round);
5607 SDValue DAGCombiner::visitFP_EXTEND(SDNode *N) {
5608 SDValue N0 = N->getOperand(0);
5609 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5610 EVT VT = N->getValueType(0);
5612 // If this is fp_round(fpextend), don't fold it, allow ourselves to be folded.
5613 if (N->hasOneUse() &&
5614 N->use_begin()->getOpcode() == ISD::FP_ROUND)
5617 // fold (fp_extend c1fp) -> c1fp
5618 if (N0CFP && VT != MVT::ppcf128)
5619 return DAG.getNode(ISD::FP_EXTEND, N->getDebugLoc(), VT, N0);
5621 // Turn fp_extend(fp_round(X, 1)) -> x since the fp_round doesn't affect the
5623 if (N0.getOpcode() == ISD::FP_ROUND
5624 && N0.getNode()->getConstantOperandVal(1) == 1) {
5625 SDValue In = N0.getOperand(0);
5626 if (In.getValueType() == VT) return In;
5627 if (VT.bitsLT(In.getValueType()))
5628 return DAG.getNode(ISD::FP_ROUND, N->getDebugLoc(), VT,
5629 In, N0.getOperand(1));
5630 return DAG.getNode(ISD::FP_EXTEND, N->getDebugLoc(), VT, In);
5633 // fold (fpext (load x)) -> (fpext (fptrunc (extload x)))
5634 if (ISD::isNON_EXTLoad(N0.getNode()) && N0.hasOneUse() &&
5635 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5636 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType()))) {
5637 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5638 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, N->getDebugLoc(), VT,
5640 LN0->getBasePtr(), LN0->getPointerInfo(),
5642 LN0->isVolatile(), LN0->isNonTemporal(),
5643 LN0->getAlignment());
5644 CombineTo(N, ExtLoad);
5645 CombineTo(N0.getNode(),
5646 DAG.getNode(ISD::FP_ROUND, N0.getDebugLoc(),
5647 N0.getValueType(), ExtLoad, DAG.getIntPtrConstant(1)),
5648 ExtLoad.getValue(1));
5649 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5655 SDValue DAGCombiner::visitFNEG(SDNode *N) {
5656 SDValue N0 = N->getOperand(0);
5657 EVT VT = N->getValueType(0);
5659 if (isNegatibleForFree(N0, LegalOperations, &DAG.getTarget().Options))
5660 return GetNegatedExpression(N0, DAG, LegalOperations);
5662 // Transform fneg(bitconvert(x)) -> bitconvert(x^sign) to avoid loading
5663 // constant pool values.
5664 if (N0.getOpcode() == ISD::BITCAST &&
5666 N0.getNode()->hasOneUse() &&
5667 N0.getOperand(0).getValueType().isInteger()) {
5668 SDValue Int = N0.getOperand(0);
5669 EVT IntVT = Int.getValueType();
5670 if (IntVT.isInteger() && !IntVT.isVector()) {
5671 Int = DAG.getNode(ISD::XOR, N0.getDebugLoc(), IntVT, Int,
5672 DAG.getConstant(APInt::getSignBit(IntVT.getSizeInBits()), IntVT));
5673 AddToWorkList(Int.getNode());
5674 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
5682 SDValue DAGCombiner::visitFABS(SDNode *N) {
5683 SDValue N0 = N->getOperand(0);
5684 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5685 EVT VT = N->getValueType(0);
5687 // fold (fabs c1) -> fabs(c1)
5688 if (N0CFP && VT != MVT::ppcf128)
5689 return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0);
5690 // fold (fabs (fabs x)) -> (fabs x)
5691 if (N0.getOpcode() == ISD::FABS)
5692 return N->getOperand(0);
5693 // fold (fabs (fneg x)) -> (fabs x)
5694 // fold (fabs (fcopysign x, y)) -> (fabs x)
5695 if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN)
5696 return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0.getOperand(0));
5698 // Transform fabs(bitconvert(x)) -> bitconvert(x&~sign) to avoid loading
5699 // constant pool values.
5700 if (N0.getOpcode() == ISD::BITCAST && N0.getNode()->hasOneUse() &&
5701 N0.getOperand(0).getValueType().isInteger() &&
5702 !N0.getOperand(0).getValueType().isVector()) {
5703 SDValue Int = N0.getOperand(0);
5704 EVT IntVT = Int.getValueType();
5705 if (IntVT.isInteger() && !IntVT.isVector()) {
5706 Int = DAG.getNode(ISD::AND, N0.getDebugLoc(), IntVT, Int,
5707 DAG.getConstant(~APInt::getSignBit(IntVT.getSizeInBits()), IntVT));
5708 AddToWorkList(Int.getNode());
5709 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
5710 N->getValueType(0), Int);
5717 SDValue DAGCombiner::visitBRCOND(SDNode *N) {
5718 SDValue Chain = N->getOperand(0);
5719 SDValue N1 = N->getOperand(1);
5720 SDValue N2 = N->getOperand(2);
5722 // If N is a constant we could fold this into a fallthrough or unconditional
5723 // branch. However that doesn't happen very often in normal code, because
5724 // Instcombine/SimplifyCFG should have handled the available opportunities.
5725 // If we did this folding here, it would be necessary to update the
5726 // MachineBasicBlock CFG, which is awkward.
5728 // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal
5730 if (N1.getOpcode() == ISD::SETCC &&
5731 TLI.isOperationLegalOrCustom(ISD::BR_CC, MVT::Other)) {
5732 return DAG.getNode(ISD::BR_CC, N->getDebugLoc(), MVT::Other,
5733 Chain, N1.getOperand(2),
5734 N1.getOperand(0), N1.getOperand(1), N2);
5737 if ((N1.hasOneUse() && N1.getOpcode() == ISD::SRL) ||
5738 ((N1.getOpcode() == ISD::TRUNCATE && N1.hasOneUse()) &&
5739 (N1.getOperand(0).hasOneUse() &&
5740 N1.getOperand(0).getOpcode() == ISD::SRL))) {
5742 if (N1.getOpcode() == ISD::TRUNCATE) {
5743 // Look pass the truncate.
5744 Trunc = N1.getNode();
5745 N1 = N1.getOperand(0);
5748 // Match this pattern so that we can generate simpler code:
5751 // %b = and i32 %a, 2
5752 // %c = srl i32 %b, 1
5753 // brcond i32 %c ...
5758 // %b = and i32 %a, 2
5759 // %c = setcc eq %b, 0
5762 // This applies only when the AND constant value has one bit set and the
5763 // SRL constant is equal to the log2 of the AND constant. The back-end is
5764 // smart enough to convert the result into a TEST/JMP sequence.
5765 SDValue Op0 = N1.getOperand(0);
5766 SDValue Op1 = N1.getOperand(1);
5768 if (Op0.getOpcode() == ISD::AND &&
5769 Op1.getOpcode() == ISD::Constant) {
5770 SDValue AndOp1 = Op0.getOperand(1);
5772 if (AndOp1.getOpcode() == ISD::Constant) {
5773 const APInt &AndConst = cast<ConstantSDNode>(AndOp1)->getAPIntValue();
5775 if (AndConst.isPowerOf2() &&
5776 cast<ConstantSDNode>(Op1)->getAPIntValue()==AndConst.logBase2()) {
5778 DAG.getSetCC(N->getDebugLoc(),
5779 TLI.getSetCCResultType(Op0.getValueType()),
5780 Op0, DAG.getConstant(0, Op0.getValueType()),
5783 SDValue NewBRCond = DAG.getNode(ISD::BRCOND, N->getDebugLoc(),
5784 MVT::Other, Chain, SetCC, N2);
5785 // Don't add the new BRCond into the worklist or else SimplifySelectCC
5786 // will convert it back to (X & C1) >> C2.
5787 CombineTo(N, NewBRCond, false);
5788 // Truncate is dead.
5790 removeFromWorkList(Trunc);
5791 DAG.DeleteNode(Trunc);
5793 // Replace the uses of SRL with SETCC
5794 WorkListRemover DeadNodes(*this);
5795 DAG.ReplaceAllUsesOfValueWith(N1, SetCC, &DeadNodes);
5796 removeFromWorkList(N1.getNode());
5797 DAG.DeleteNode(N1.getNode());
5798 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5804 // Restore N1 if the above transformation doesn't match.
5805 N1 = N->getOperand(1);
5808 // Transform br(xor(x, y)) -> br(x != y)
5809 // Transform br(xor(xor(x,y), 1)) -> br (x == y)
5810 if (N1.hasOneUse() && N1.getOpcode() == ISD::XOR) {
5811 SDNode *TheXor = N1.getNode();
5812 SDValue Op0 = TheXor->getOperand(0);
5813 SDValue Op1 = TheXor->getOperand(1);
5814 if (Op0.getOpcode() == Op1.getOpcode()) {
5815 // Avoid missing important xor optimizations.
5816 SDValue Tmp = visitXOR(TheXor);
5817 if (Tmp.getNode() && Tmp.getNode() != TheXor) {
5818 DEBUG(dbgs() << "\nReplacing.8 ";
5820 dbgs() << "\nWith: ";
5821 Tmp.getNode()->dump(&DAG);
5823 WorkListRemover DeadNodes(*this);
5824 DAG.ReplaceAllUsesOfValueWith(N1, Tmp, &DeadNodes);
5825 removeFromWorkList(TheXor);
5826 DAG.DeleteNode(TheXor);
5827 return DAG.getNode(ISD::BRCOND, N->getDebugLoc(),
5828 MVT::Other, Chain, Tmp, N2);
5832 if (Op0.getOpcode() != ISD::SETCC && Op1.getOpcode() != ISD::SETCC) {
5834 if (ConstantSDNode *RHSCI = dyn_cast<ConstantSDNode>(Op0))
5835 if (RHSCI->getAPIntValue() == 1 && Op0.hasOneUse() &&
5836 Op0.getOpcode() == ISD::XOR) {
5837 TheXor = Op0.getNode();
5841 EVT SetCCVT = N1.getValueType();
5843 SetCCVT = TLI.getSetCCResultType(SetCCVT);
5844 SDValue SetCC = DAG.getSetCC(TheXor->getDebugLoc(),
5847 Equal ? ISD::SETEQ : ISD::SETNE);
5848 // Replace the uses of XOR with SETCC
5849 WorkListRemover DeadNodes(*this);
5850 DAG.ReplaceAllUsesOfValueWith(N1, SetCC, &DeadNodes);
5851 removeFromWorkList(N1.getNode());
5852 DAG.DeleteNode(N1.getNode());
5853 return DAG.getNode(ISD::BRCOND, N->getDebugLoc(),
5854 MVT::Other, Chain, SetCC, N2);
5861 // Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB.
5863 SDValue DAGCombiner::visitBR_CC(SDNode *N) {
5864 CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1));
5865 SDValue CondLHS = N->getOperand(2), CondRHS = N->getOperand(3);
5867 // If N is a constant we could fold this into a fallthrough or unconditional
5868 // branch. However that doesn't happen very often in normal code, because
5869 // Instcombine/SimplifyCFG should have handled the available opportunities.
5870 // If we did this folding here, it would be necessary to update the
5871 // MachineBasicBlock CFG, which is awkward.
5873 // Use SimplifySetCC to simplify SETCC's.
5874 SDValue Simp = SimplifySetCC(TLI.getSetCCResultType(CondLHS.getValueType()),
5875 CondLHS, CondRHS, CC->get(), N->getDebugLoc(),
5877 if (Simp.getNode()) AddToWorkList(Simp.getNode());
5879 // fold to a simpler setcc
5880 if (Simp.getNode() && Simp.getOpcode() == ISD::SETCC)
5881 return DAG.getNode(ISD::BR_CC, N->getDebugLoc(), MVT::Other,
5882 N->getOperand(0), Simp.getOperand(2),
5883 Simp.getOperand(0), Simp.getOperand(1),
5889 /// CombineToPreIndexedLoadStore - Try turning a load / store into a
5890 /// pre-indexed load / store when the base pointer is an add or subtract
5891 /// and it has other uses besides the load / store. After the
5892 /// transformation, the new indexed load / store has effectively folded
5893 /// the add / subtract in and all of its other uses are redirected to the
5894 /// new load / store.
5895 bool DAGCombiner::CombineToPreIndexedLoadStore(SDNode *N) {
5896 if (Level < AfterLegalizeDAG)
5902 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
5903 if (LD->isIndexed())
5905 VT = LD->getMemoryVT();
5906 if (!TLI.isIndexedLoadLegal(ISD::PRE_INC, VT) &&
5907 !TLI.isIndexedLoadLegal(ISD::PRE_DEC, VT))
5909 Ptr = LD->getBasePtr();
5910 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
5911 if (ST->isIndexed())
5913 VT = ST->getMemoryVT();
5914 if (!TLI.isIndexedStoreLegal(ISD::PRE_INC, VT) &&
5915 !TLI.isIndexedStoreLegal(ISD::PRE_DEC, VT))
5917 Ptr = ST->getBasePtr();
5923 // If the pointer is not an add/sub, or if it doesn't have multiple uses, bail
5924 // out. There is no reason to make this a preinc/predec.
5925 if ((Ptr.getOpcode() != ISD::ADD && Ptr.getOpcode() != ISD::SUB) ||
5926 Ptr.getNode()->hasOneUse())
5929 // Ask the target to do addressing mode selection.
5932 ISD::MemIndexedMode AM = ISD::UNINDEXED;
5933 if (!TLI.getPreIndexedAddressParts(N, BasePtr, Offset, AM, DAG))
5935 // Don't create a indexed load / store with zero offset.
5936 if (isa<ConstantSDNode>(Offset) &&
5937 cast<ConstantSDNode>(Offset)->isNullValue())
5940 // Try turning it into a pre-indexed load / store except when:
5941 // 1) The new base ptr is a frame index.
5942 // 2) If N is a store and the new base ptr is either the same as or is a
5943 // predecessor of the value being stored.
5944 // 3) Another use of old base ptr is a predecessor of N. If ptr is folded
5945 // that would create a cycle.
5946 // 4) All uses are load / store ops that use it as old base ptr.
5948 // Check #1. Preinc'ing a frame index would require copying the stack pointer
5949 // (plus the implicit offset) to a register to preinc anyway.
5950 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
5955 SDValue Val = cast<StoreSDNode>(N)->getValue();
5956 if (Val == BasePtr || BasePtr.getNode()->isPredecessorOf(Val.getNode()))
5960 // Now check for #3 and #4.
5961 bool RealUse = false;
5963 // Caches for hasPredecessorHelper
5964 SmallPtrSet<const SDNode *, 32> Visited;
5965 SmallVector<const SDNode *, 16> Worklist;
5967 for (SDNode::use_iterator I = Ptr.getNode()->use_begin(),
5968 E = Ptr.getNode()->use_end(); I != E; ++I) {
5972 if (N->hasPredecessorHelper(Use, Visited, Worklist))
5975 if (!((Use->getOpcode() == ISD::LOAD &&
5976 cast<LoadSDNode>(Use)->getBasePtr() == Ptr) ||
5977 (Use->getOpcode() == ISD::STORE &&
5978 cast<StoreSDNode>(Use)->getBasePtr() == Ptr)))
5987 Result = DAG.getIndexedLoad(SDValue(N,0), N->getDebugLoc(),
5988 BasePtr, Offset, AM);
5990 Result = DAG.getIndexedStore(SDValue(N,0), N->getDebugLoc(),
5991 BasePtr, Offset, AM);
5994 DEBUG(dbgs() << "\nReplacing.4 ";
5996 dbgs() << "\nWith: ";
5997 Result.getNode()->dump(&DAG);
5999 WorkListRemover DeadNodes(*this);
6001 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0),
6003 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2),
6006 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1),
6010 // Finally, since the node is now dead, remove it from the graph.
6013 // Replace the uses of Ptr with uses of the updated base value.
6014 DAG.ReplaceAllUsesOfValueWith(Ptr, Result.getValue(isLoad ? 1 : 0),
6016 removeFromWorkList(Ptr.getNode());
6017 DAG.DeleteNode(Ptr.getNode());
6022 /// CombineToPostIndexedLoadStore - Try to combine a load / store with a
6023 /// add / sub of the base pointer node into a post-indexed load / store.
6024 /// The transformation folded the add / subtract into the new indexed
6025 /// load / store effectively and all of its uses are redirected to the
6026 /// new load / store.
6027 bool DAGCombiner::CombineToPostIndexedLoadStore(SDNode *N) {
6028 if (Level < AfterLegalizeDAG)
6034 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
6035 if (LD->isIndexed())
6037 VT = LD->getMemoryVT();
6038 if (!TLI.isIndexedLoadLegal(ISD::POST_INC, VT) &&
6039 !TLI.isIndexedLoadLegal(ISD::POST_DEC, VT))
6041 Ptr = LD->getBasePtr();
6042 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
6043 if (ST->isIndexed())
6045 VT = ST->getMemoryVT();
6046 if (!TLI.isIndexedStoreLegal(ISD::POST_INC, VT) &&
6047 !TLI.isIndexedStoreLegal(ISD::POST_DEC, VT))
6049 Ptr = ST->getBasePtr();
6055 if (Ptr.getNode()->hasOneUse())
6058 for (SDNode::use_iterator I = Ptr.getNode()->use_begin(),
6059 E = Ptr.getNode()->use_end(); I != E; ++I) {
6062 (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB))
6067 ISD::MemIndexedMode AM = ISD::UNINDEXED;
6068 if (TLI.getPostIndexedAddressParts(N, Op, BasePtr, Offset, AM, DAG)) {
6069 // Don't create a indexed load / store with zero offset.
6070 if (isa<ConstantSDNode>(Offset) &&
6071 cast<ConstantSDNode>(Offset)->isNullValue())
6074 // Try turning it into a post-indexed load / store except when
6075 // 1) All uses are load / store ops that use it as base ptr.
6076 // 2) Op must be independent of N, i.e. Op is neither a predecessor
6077 // nor a successor of N. Otherwise, if Op is folded that would
6080 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
6084 bool TryNext = false;
6085 for (SDNode::use_iterator II = BasePtr.getNode()->use_begin(),
6086 EE = BasePtr.getNode()->use_end(); II != EE; ++II) {
6088 if (Use == Ptr.getNode())
6091 // If all the uses are load / store addresses, then don't do the
6093 if (Use->getOpcode() == ISD::ADD || Use->getOpcode() == ISD::SUB){
6094 bool RealUse = false;
6095 for (SDNode::use_iterator III = Use->use_begin(),
6096 EEE = Use->use_end(); III != EEE; ++III) {
6097 SDNode *UseUse = *III;
6098 if (!((UseUse->getOpcode() == ISD::LOAD &&
6099 cast<LoadSDNode>(UseUse)->getBasePtr().getNode() == Use) ||
6100 (UseUse->getOpcode() == ISD::STORE &&
6101 cast<StoreSDNode>(UseUse)->getBasePtr().getNode() == Use)))
6116 if (!Op->isPredecessorOf(N) && !N->isPredecessorOf(Op)) {
6117 SDValue Result = isLoad
6118 ? DAG.getIndexedLoad(SDValue(N,0), N->getDebugLoc(),
6119 BasePtr, Offset, AM)
6120 : DAG.getIndexedStore(SDValue(N,0), N->getDebugLoc(),
6121 BasePtr, Offset, AM);
6124 DEBUG(dbgs() << "\nReplacing.5 ";
6126 dbgs() << "\nWith: ";
6127 Result.getNode()->dump(&DAG);
6129 WorkListRemover DeadNodes(*this);
6131 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0),
6133 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2),
6136 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1),
6140 // Finally, since the node is now dead, remove it from the graph.
6143 // Replace the uses of Use with uses of the updated base value.
6144 DAG.ReplaceAllUsesOfValueWith(SDValue(Op, 0),
6145 Result.getValue(isLoad ? 1 : 0),
6147 removeFromWorkList(Op);
6157 SDValue DAGCombiner::visitLOAD(SDNode *N) {
6158 LoadSDNode *LD = cast<LoadSDNode>(N);
6159 SDValue Chain = LD->getChain();
6160 SDValue Ptr = LD->getBasePtr();
6162 // If load is not volatile and there are no uses of the loaded value (and
6163 // the updated indexed value in case of indexed loads), change uses of the
6164 // chain value into uses of the chain input (i.e. delete the dead load).
6165 if (!LD->isVolatile()) {
6166 if (N->getValueType(1) == MVT::Other) {
6168 if (N->hasNUsesOfValue(0, 0)) {
6169 // It's not safe to use the two value CombineTo variant here. e.g.
6170 // v1, chain2 = load chain1, loc
6171 // v2, chain3 = load chain2, loc
6173 // Now we replace use of chain2 with chain1. This makes the second load
6174 // isomorphic to the one we are deleting, and thus makes this load live.
6175 DEBUG(dbgs() << "\nReplacing.6 ";
6177 dbgs() << "\nWith chain: ";
6178 Chain.getNode()->dump(&DAG);
6180 WorkListRemover DeadNodes(*this);
6181 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain, &DeadNodes);
6183 if (N->use_empty()) {
6184 removeFromWorkList(N);
6188 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6192 assert(N->getValueType(2) == MVT::Other && "Malformed indexed loads?");
6193 if (N->hasNUsesOfValue(0, 0) && N->hasNUsesOfValue(0, 1)) {
6194 SDValue Undef = DAG.getUNDEF(N->getValueType(0));
6195 DEBUG(dbgs() << "\nReplacing.7 ";
6197 dbgs() << "\nWith: ";
6198 Undef.getNode()->dump(&DAG);
6199 dbgs() << " and 2 other values\n");
6200 WorkListRemover DeadNodes(*this);
6201 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Undef, &DeadNodes);
6202 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1),
6203 DAG.getUNDEF(N->getValueType(1)),
6205 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 2), Chain, &DeadNodes);
6206 removeFromWorkList(N);
6208 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6213 // If this load is directly stored, replace the load value with the stored
6215 // TODO: Handle store large -> read small portion.
6216 // TODO: Handle TRUNCSTORE/LOADEXT
6217 if (ISD::isNormalLoad(N) && !LD->isVolatile()) {
6218 if (ISD::isNON_TRUNCStore(Chain.getNode())) {
6219 StoreSDNode *PrevST = cast<StoreSDNode>(Chain);
6220 if (PrevST->getBasePtr() == Ptr &&
6221 PrevST->getValue().getValueType() == N->getValueType(0))
6222 return CombineTo(N, Chain.getOperand(1), Chain);
6226 // Try to infer better alignment information than the load already has.
6227 if (OptLevel != CodeGenOpt::None && LD->isUnindexed()) {
6228 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
6229 if (Align > LD->getAlignment())
6230 return DAG.getExtLoad(LD->getExtensionType(), N->getDebugLoc(),
6231 LD->getValueType(0),
6232 Chain, Ptr, LD->getPointerInfo(),
6234 LD->isVolatile(), LD->isNonTemporal(), Align);
6239 // Walk up chain skipping non-aliasing memory nodes.
6240 SDValue BetterChain = FindBetterChain(N, Chain);
6242 // If there is a better chain.
6243 if (Chain != BetterChain) {
6246 // Replace the chain to void dependency.
6247 if (LD->getExtensionType() == ISD::NON_EXTLOAD) {
6248 ReplLoad = DAG.getLoad(N->getValueType(0), LD->getDebugLoc(),
6249 BetterChain, Ptr, LD->getPointerInfo(),
6250 LD->isVolatile(), LD->isNonTemporal(),
6251 LD->isInvariant(), LD->getAlignment());
6253 ReplLoad = DAG.getExtLoad(LD->getExtensionType(), LD->getDebugLoc(),
6254 LD->getValueType(0),
6255 BetterChain, Ptr, LD->getPointerInfo(),
6258 LD->isNonTemporal(),
6259 LD->getAlignment());
6262 // Create token factor to keep old chain connected.
6263 SDValue Token = DAG.getNode(ISD::TokenFactor, N->getDebugLoc(),
6264 MVT::Other, Chain, ReplLoad.getValue(1));
6266 // Make sure the new and old chains are cleaned up.
6267 AddToWorkList(Token.getNode());
6269 // Replace uses with load result and token factor. Don't add users
6271 return CombineTo(N, ReplLoad.getValue(0), Token, false);
6275 // Try transforming N to an indexed load.
6276 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
6277 return SDValue(N, 0);
6282 /// CheckForMaskedLoad - Check to see if V is (and load (ptr), imm), where the
6283 /// load is having specific bytes cleared out. If so, return the byte size
6284 /// being masked out and the shift amount.
6285 static std::pair<unsigned, unsigned>
6286 CheckForMaskedLoad(SDValue V, SDValue Ptr, SDValue Chain) {
6287 std::pair<unsigned, unsigned> Result(0, 0);
6289 // Check for the structure we're looking for.
6290 if (V->getOpcode() != ISD::AND ||
6291 !isa<ConstantSDNode>(V->getOperand(1)) ||
6292 !ISD::isNormalLoad(V->getOperand(0).getNode()))
6295 // Check the chain and pointer.
6296 LoadSDNode *LD = cast<LoadSDNode>(V->getOperand(0));
6297 if (LD->getBasePtr() != Ptr) return Result; // Not from same pointer.
6299 // The store should be chained directly to the load or be an operand of a
6301 if (LD == Chain.getNode())
6303 else if (Chain->getOpcode() != ISD::TokenFactor)
6304 return Result; // Fail.
6307 for (unsigned i = 0, e = Chain->getNumOperands(); i != e; ++i)
6308 if (Chain->getOperand(i).getNode() == LD) {
6312 if (!isOk) return Result;
6315 // This only handles simple types.
6316 if (V.getValueType() != MVT::i16 &&
6317 V.getValueType() != MVT::i32 &&
6318 V.getValueType() != MVT::i64)
6321 // Check the constant mask. Invert it so that the bits being masked out are
6322 // 0 and the bits being kept are 1. Use getSExtValue so that leading bits
6323 // follow the sign bit for uniformity.
6324 uint64_t NotMask = ~cast<ConstantSDNode>(V->getOperand(1))->getSExtValue();
6325 unsigned NotMaskLZ = CountLeadingZeros_64(NotMask);
6326 if (NotMaskLZ & 7) return Result; // Must be multiple of a byte.
6327 unsigned NotMaskTZ = CountTrailingZeros_64(NotMask);
6328 if (NotMaskTZ & 7) return Result; // Must be multiple of a byte.
6329 if (NotMaskLZ == 64) return Result; // All zero mask.
6331 // See if we have a continuous run of bits. If so, we have 0*1+0*
6332 if (CountTrailingOnes_64(NotMask >> NotMaskTZ)+NotMaskTZ+NotMaskLZ != 64)
6335 // Adjust NotMaskLZ down to be from the actual size of the int instead of i64.
6336 if (V.getValueType() != MVT::i64 && NotMaskLZ)
6337 NotMaskLZ -= 64-V.getValueSizeInBits();
6339 unsigned MaskedBytes = (V.getValueSizeInBits()-NotMaskLZ-NotMaskTZ)/8;
6340 switch (MaskedBytes) {
6344 default: return Result; // All one mask, or 5-byte mask.
6347 // Verify that the first bit starts at a multiple of mask so that the access
6348 // is aligned the same as the access width.
6349 if (NotMaskTZ && NotMaskTZ/8 % MaskedBytes) return Result;
6351 Result.first = MaskedBytes;
6352 Result.second = NotMaskTZ/8;
6357 /// ShrinkLoadReplaceStoreWithStore - Check to see if IVal is something that
6358 /// provides a value as specified by MaskInfo. If so, replace the specified
6359 /// store with a narrower store of truncated IVal.
6361 ShrinkLoadReplaceStoreWithStore(const std::pair<unsigned, unsigned> &MaskInfo,
6362 SDValue IVal, StoreSDNode *St,
6364 unsigned NumBytes = MaskInfo.first;
6365 unsigned ByteShift = MaskInfo.second;
6366 SelectionDAG &DAG = DC->getDAG();
6368 // Check to see if IVal is all zeros in the part being masked in by the 'or'
6369 // that uses this. If not, this is not a replacement.
6370 APInt Mask = ~APInt::getBitsSet(IVal.getValueSizeInBits(),
6371 ByteShift*8, (ByteShift+NumBytes)*8);
6372 if (!DAG.MaskedValueIsZero(IVal, Mask)) return 0;
6374 // Check that it is legal on the target to do this. It is legal if the new
6375 // VT we're shrinking to (i8/i16/i32) is legal or we're still before type
6377 MVT VT = MVT::getIntegerVT(NumBytes*8);
6378 if (!DC->isTypeLegal(VT))
6381 // Okay, we can do this! Replace the 'St' store with a store of IVal that is
6382 // shifted by ByteShift and truncated down to NumBytes.
6384 IVal = DAG.getNode(ISD::SRL, IVal->getDebugLoc(), IVal.getValueType(), IVal,
6385 DAG.getConstant(ByteShift*8,
6386 DC->getShiftAmountTy(IVal.getValueType())));
6388 // Figure out the offset for the store and the alignment of the access.
6390 unsigned NewAlign = St->getAlignment();
6392 if (DAG.getTargetLoweringInfo().isLittleEndian())
6393 StOffset = ByteShift;
6395 StOffset = IVal.getValueType().getStoreSize() - ByteShift - NumBytes;
6397 SDValue Ptr = St->getBasePtr();
6399 Ptr = DAG.getNode(ISD::ADD, IVal->getDebugLoc(), Ptr.getValueType(),
6400 Ptr, DAG.getConstant(StOffset, Ptr.getValueType()));
6401 NewAlign = MinAlign(NewAlign, StOffset);
6404 // Truncate down to the new size.
6405 IVal = DAG.getNode(ISD::TRUNCATE, IVal->getDebugLoc(), VT, IVal);
6408 return DAG.getStore(St->getChain(), St->getDebugLoc(), IVal, Ptr,
6409 St->getPointerInfo().getWithOffset(StOffset),
6410 false, false, NewAlign).getNode();
6414 /// ReduceLoadOpStoreWidth - Look for sequence of load / op / store where op is
6415 /// one of 'or', 'xor', and 'and' of immediates. If 'op' is only touching some
6416 /// of the loaded bits, try narrowing the load and store if it would end up
6417 /// being a win for performance or code size.
6418 SDValue DAGCombiner::ReduceLoadOpStoreWidth(SDNode *N) {
6419 StoreSDNode *ST = cast<StoreSDNode>(N);
6420 if (ST->isVolatile())
6423 SDValue Chain = ST->getChain();
6424 SDValue Value = ST->getValue();
6425 SDValue Ptr = ST->getBasePtr();
6426 EVT VT = Value.getValueType();
6428 if (ST->isTruncatingStore() || VT.isVector() || !Value.hasOneUse())
6431 unsigned Opc = Value.getOpcode();
6433 // If this is "store (or X, Y), P" and X is "(and (load P), cst)", where cst
6434 // is a byte mask indicating a consecutive number of bytes, check to see if
6435 // Y is known to provide just those bytes. If so, we try to replace the
6436 // load + replace + store sequence with a single (narrower) store, which makes
6438 if (Opc == ISD::OR) {
6439 std::pair<unsigned, unsigned> MaskedLoad;
6440 MaskedLoad = CheckForMaskedLoad(Value.getOperand(0), Ptr, Chain);
6441 if (MaskedLoad.first)
6442 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
6443 Value.getOperand(1), ST,this))
6444 return SDValue(NewST, 0);
6446 // Or is commutative, so try swapping X and Y.
6447 MaskedLoad = CheckForMaskedLoad(Value.getOperand(1), Ptr, Chain);
6448 if (MaskedLoad.first)
6449 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
6450 Value.getOperand(0), ST,this))
6451 return SDValue(NewST, 0);
6454 if ((Opc != ISD::OR && Opc != ISD::XOR && Opc != ISD::AND) ||
6455 Value.getOperand(1).getOpcode() != ISD::Constant)
6458 SDValue N0 = Value.getOperand(0);
6459 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
6460 Chain == SDValue(N0.getNode(), 1)) {
6461 LoadSDNode *LD = cast<LoadSDNode>(N0);
6462 if (LD->getBasePtr() != Ptr ||
6463 LD->getPointerInfo().getAddrSpace() !=
6464 ST->getPointerInfo().getAddrSpace())
6467 // Find the type to narrow it the load / op / store to.
6468 SDValue N1 = Value.getOperand(1);
6469 unsigned BitWidth = N1.getValueSizeInBits();
6470 APInt Imm = cast<ConstantSDNode>(N1)->getAPIntValue();
6471 if (Opc == ISD::AND)
6472 Imm ^= APInt::getAllOnesValue(BitWidth);
6473 if (Imm == 0 || Imm.isAllOnesValue())
6475 unsigned ShAmt = Imm.countTrailingZeros();
6476 unsigned MSB = BitWidth - Imm.countLeadingZeros() - 1;
6477 unsigned NewBW = NextPowerOf2(MSB - ShAmt);
6478 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
6479 while (NewBW < BitWidth &&
6480 !(TLI.isOperationLegalOrCustom(Opc, NewVT) &&
6481 TLI.isNarrowingProfitable(VT, NewVT))) {
6482 NewBW = NextPowerOf2(NewBW);
6483 NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
6485 if (NewBW >= BitWidth)
6488 // If the lsb changed does not start at the type bitwidth boundary,
6489 // start at the previous one.
6491 ShAmt = (((ShAmt + NewBW - 1) / NewBW) * NewBW) - NewBW;
6492 APInt Mask = APInt::getBitsSet(BitWidth, ShAmt, ShAmt + NewBW);
6493 if ((Imm & Mask) == Imm) {
6494 APInt NewImm = (Imm & Mask).lshr(ShAmt).trunc(NewBW);
6495 if (Opc == ISD::AND)
6496 NewImm ^= APInt::getAllOnesValue(NewBW);
6497 uint64_t PtrOff = ShAmt / 8;
6498 // For big endian targets, we need to adjust the offset to the pointer to
6499 // load the correct bytes.
6500 if (TLI.isBigEndian())
6501 PtrOff = (BitWidth + 7 - NewBW) / 8 - PtrOff;
6503 unsigned NewAlign = MinAlign(LD->getAlignment(), PtrOff);
6504 Type *NewVTTy = NewVT.getTypeForEVT(*DAG.getContext());
6505 if (NewAlign < TLI.getTargetData()->getABITypeAlignment(NewVTTy))
6508 SDValue NewPtr = DAG.getNode(ISD::ADD, LD->getDebugLoc(),
6509 Ptr.getValueType(), Ptr,
6510 DAG.getConstant(PtrOff, Ptr.getValueType()));
6511 SDValue NewLD = DAG.getLoad(NewVT, N0.getDebugLoc(),
6512 LD->getChain(), NewPtr,
6513 LD->getPointerInfo().getWithOffset(PtrOff),
6514 LD->isVolatile(), LD->isNonTemporal(),
6515 LD->isInvariant(), NewAlign);
6516 SDValue NewVal = DAG.getNode(Opc, Value.getDebugLoc(), NewVT, NewLD,
6517 DAG.getConstant(NewImm, NewVT));
6518 SDValue NewST = DAG.getStore(Chain, N->getDebugLoc(),
6520 ST->getPointerInfo().getWithOffset(PtrOff),
6521 false, false, NewAlign);
6523 AddToWorkList(NewPtr.getNode());
6524 AddToWorkList(NewLD.getNode());
6525 AddToWorkList(NewVal.getNode());
6526 WorkListRemover DeadNodes(*this);
6527 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLD.getValue(1),
6537 /// TransformFPLoadStorePair - For a given floating point load / store pair,
6538 /// if the load value isn't used by any other operations, then consider
6539 /// transforming the pair to integer load / store operations if the target
6540 /// deems the transformation profitable.
6541 SDValue DAGCombiner::TransformFPLoadStorePair(SDNode *N) {
6542 StoreSDNode *ST = cast<StoreSDNode>(N);
6543 SDValue Chain = ST->getChain();
6544 SDValue Value = ST->getValue();
6545 if (ISD::isNormalStore(ST) && ISD::isNormalLoad(Value.getNode()) &&
6546 Value.hasOneUse() &&
6547 Chain == SDValue(Value.getNode(), 1)) {
6548 LoadSDNode *LD = cast<LoadSDNode>(Value);
6549 EVT VT = LD->getMemoryVT();
6550 if (!VT.isFloatingPoint() ||
6551 VT != ST->getMemoryVT() ||
6552 LD->isNonTemporal() ||
6553 ST->isNonTemporal() ||
6554 LD->getPointerInfo().getAddrSpace() != 0 ||
6555 ST->getPointerInfo().getAddrSpace() != 0)
6558 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
6559 if (!TLI.isOperationLegal(ISD::LOAD, IntVT) ||
6560 !TLI.isOperationLegal(ISD::STORE, IntVT) ||
6561 !TLI.isDesirableToTransformToIntegerOp(ISD::LOAD, VT) ||
6562 !TLI.isDesirableToTransformToIntegerOp(ISD::STORE, VT))
6565 unsigned LDAlign = LD->getAlignment();
6566 unsigned STAlign = ST->getAlignment();
6567 Type *IntVTTy = IntVT.getTypeForEVT(*DAG.getContext());
6568 unsigned ABIAlign = TLI.getTargetData()->getABITypeAlignment(IntVTTy);
6569 if (LDAlign < ABIAlign || STAlign < ABIAlign)
6572 SDValue NewLD = DAG.getLoad(IntVT, Value.getDebugLoc(),
6573 LD->getChain(), LD->getBasePtr(),
6574 LD->getPointerInfo(),
6575 false, false, false, LDAlign);
6577 SDValue NewST = DAG.getStore(NewLD.getValue(1), N->getDebugLoc(),
6578 NewLD, ST->getBasePtr(),
6579 ST->getPointerInfo(),
6580 false, false, STAlign);
6582 AddToWorkList(NewLD.getNode());
6583 AddToWorkList(NewST.getNode());
6584 WorkListRemover DeadNodes(*this);
6585 DAG.ReplaceAllUsesOfValueWith(Value.getValue(1), NewLD.getValue(1),
6594 SDValue DAGCombiner::visitSTORE(SDNode *N) {
6595 StoreSDNode *ST = cast<StoreSDNode>(N);
6596 SDValue Chain = ST->getChain();
6597 SDValue Value = ST->getValue();
6598 SDValue Ptr = ST->getBasePtr();
6600 // If this is a store of a bit convert, store the input value if the
6601 // resultant store does not need a higher alignment than the original.
6602 if (Value.getOpcode() == ISD::BITCAST && !ST->isTruncatingStore() &&
6603 ST->isUnindexed()) {
6604 unsigned OrigAlign = ST->getAlignment();
6605 EVT SVT = Value.getOperand(0).getValueType();
6606 unsigned Align = TLI.getTargetData()->
6607 getABITypeAlignment(SVT.getTypeForEVT(*DAG.getContext()));
6608 if (Align <= OrigAlign &&
6609 ((!LegalOperations && !ST->isVolatile()) ||
6610 TLI.isOperationLegalOrCustom(ISD::STORE, SVT)))
6611 return DAG.getStore(Chain, N->getDebugLoc(), Value.getOperand(0),
6612 Ptr, ST->getPointerInfo(), ST->isVolatile(),
6613 ST->isNonTemporal(), OrigAlign);
6616 // Turn 'store undef, Ptr' -> nothing.
6617 if (Value.getOpcode() == ISD::UNDEF && ST->isUnindexed())
6620 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
6621 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Value)) {
6622 // NOTE: If the original store is volatile, this transform must not increase
6623 // the number of stores. For example, on x86-32 an f64 can be stored in one
6624 // processor operation but an i64 (which is not legal) requires two. So the
6625 // transform should not be done in this case.
6626 if (Value.getOpcode() != ISD::TargetConstantFP) {
6628 switch (CFP->getValueType(0).getSimpleVT().SimpleTy) {
6629 default: llvm_unreachable("Unknown FP type");
6630 case MVT::f80: // We don't do this for these yet.
6635 if ((isTypeLegal(MVT::i32) && !LegalOperations && !ST->isVolatile()) ||
6636 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
6637 Tmp = DAG.getConstant((uint32_t)CFP->getValueAPF().
6638 bitcastToAPInt().getZExtValue(), MVT::i32);
6639 return DAG.getStore(Chain, N->getDebugLoc(), Tmp,
6640 Ptr, ST->getPointerInfo(), ST->isVolatile(),
6641 ST->isNonTemporal(), ST->getAlignment());
6645 if ((TLI.isTypeLegal(MVT::i64) && !LegalOperations &&
6646 !ST->isVolatile()) ||
6647 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i64)) {
6648 Tmp = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
6649 getZExtValue(), MVT::i64);
6650 return DAG.getStore(Chain, N->getDebugLoc(), Tmp,
6651 Ptr, ST->getPointerInfo(), ST->isVolatile(),
6652 ST->isNonTemporal(), ST->getAlignment());
6655 if (!ST->isVolatile() &&
6656 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
6657 // Many FP stores are not made apparent until after legalize, e.g. for
6658 // argument passing. Since this is so common, custom legalize the
6659 // 64-bit integer store into two 32-bit stores.
6660 uint64_t Val = CFP->getValueAPF().bitcastToAPInt().getZExtValue();
6661 SDValue Lo = DAG.getConstant(Val & 0xFFFFFFFF, MVT::i32);
6662 SDValue Hi = DAG.getConstant(Val >> 32, MVT::i32);
6663 if (TLI.isBigEndian()) std::swap(Lo, Hi);
6665 unsigned Alignment = ST->getAlignment();
6666 bool isVolatile = ST->isVolatile();
6667 bool isNonTemporal = ST->isNonTemporal();
6669 SDValue St0 = DAG.getStore(Chain, ST->getDebugLoc(), Lo,
6670 Ptr, ST->getPointerInfo(),
6671 isVolatile, isNonTemporal,
6672 ST->getAlignment());
6673 Ptr = DAG.getNode(ISD::ADD, N->getDebugLoc(), Ptr.getValueType(), Ptr,
6674 DAG.getConstant(4, Ptr.getValueType()));
6675 Alignment = MinAlign(Alignment, 4U);
6676 SDValue St1 = DAG.getStore(Chain, ST->getDebugLoc(), Hi,
6677 Ptr, ST->getPointerInfo().getWithOffset(4),
6678 isVolatile, isNonTemporal,
6680 return DAG.getNode(ISD::TokenFactor, N->getDebugLoc(), MVT::Other,
6689 // Try to infer better alignment information than the store already has.
6690 if (OptLevel != CodeGenOpt::None && ST->isUnindexed()) {
6691 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
6692 if (Align > ST->getAlignment())
6693 return DAG.getTruncStore(Chain, N->getDebugLoc(), Value,
6694 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
6695 ST->isVolatile(), ST->isNonTemporal(), Align);
6699 // Try transforming a pair floating point load / store ops to integer
6700 // load / store ops.
6701 SDValue NewST = TransformFPLoadStorePair(N);
6702 if (NewST.getNode())
6706 // Walk up chain skipping non-aliasing memory nodes.
6707 SDValue BetterChain = FindBetterChain(N, Chain);
6709 // If there is a better chain.
6710 if (Chain != BetterChain) {
6713 // Replace the chain to avoid dependency.
6714 if (ST->isTruncatingStore()) {
6715 ReplStore = DAG.getTruncStore(BetterChain, N->getDebugLoc(), Value, Ptr,
6716 ST->getPointerInfo(),
6717 ST->getMemoryVT(), ST->isVolatile(),
6718 ST->isNonTemporal(), ST->getAlignment());
6720 ReplStore = DAG.getStore(BetterChain, N->getDebugLoc(), Value, Ptr,
6721 ST->getPointerInfo(),
6722 ST->isVolatile(), ST->isNonTemporal(),
6723 ST->getAlignment());
6726 // Create token to keep both nodes around.
6727 SDValue Token = DAG.getNode(ISD::TokenFactor, N->getDebugLoc(),
6728 MVT::Other, Chain, ReplStore);
6730 // Make sure the new and old chains are cleaned up.
6731 AddToWorkList(Token.getNode());
6733 // Don't add users to work list.
6734 return CombineTo(N, Token, false);
6738 // Try transforming N to an indexed store.
6739 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
6740 return SDValue(N, 0);
6742 // FIXME: is there such a thing as a truncating indexed store?
6743 if (ST->isTruncatingStore() && ST->isUnindexed() &&
6744 Value.getValueType().isInteger()) {
6745 // See if we can simplify the input to this truncstore with knowledge that
6746 // only the low bits are being used. For example:
6747 // "truncstore (or (shl x, 8), y), i8" -> "truncstore y, i8"
6749 GetDemandedBits(Value,
6750 APInt::getLowBitsSet(
6751 Value.getValueType().getScalarType().getSizeInBits(),
6752 ST->getMemoryVT().getScalarType().getSizeInBits()));
6753 AddToWorkList(Value.getNode());
6754 if (Shorter.getNode())
6755 return DAG.getTruncStore(Chain, N->getDebugLoc(), Shorter,
6756 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
6757 ST->isVolatile(), ST->isNonTemporal(),
6758 ST->getAlignment());
6760 // Otherwise, see if we can simplify the operation with
6761 // SimplifyDemandedBits, which only works if the value has a single use.
6762 if (SimplifyDemandedBits(Value,
6763 APInt::getLowBitsSet(
6764 Value.getValueType().getScalarType().getSizeInBits(),
6765 ST->getMemoryVT().getScalarType().getSizeInBits())))
6766 return SDValue(N, 0);
6769 // If this is a load followed by a store to the same location, then the store
6771 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Value)) {
6772 if (Ld->getBasePtr() == Ptr && ST->getMemoryVT() == Ld->getMemoryVT() &&
6773 ST->isUnindexed() && !ST->isVolatile() &&
6774 // There can't be any side effects between the load and store, such as
6776 Chain.reachesChainWithoutSideEffects(SDValue(Ld, 1))) {
6777 // The store is dead, remove it.
6782 // If this is an FP_ROUND or TRUNC followed by a store, fold this into a
6783 // truncating store. We can do this even if this is already a truncstore.
6784 if ((Value.getOpcode() == ISD::FP_ROUND || Value.getOpcode() == ISD::TRUNCATE)
6785 && Value.getNode()->hasOneUse() && ST->isUnindexed() &&
6786 TLI.isTruncStoreLegal(Value.getOperand(0).getValueType(),
6787 ST->getMemoryVT())) {
6788 return DAG.getTruncStore(Chain, N->getDebugLoc(), Value.getOperand(0),
6789 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
6790 ST->isVolatile(), ST->isNonTemporal(),
6791 ST->getAlignment());
6794 return ReduceLoadOpStoreWidth(N);
6797 SDValue DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) {
6798 SDValue InVec = N->getOperand(0);
6799 SDValue InVal = N->getOperand(1);
6800 SDValue EltNo = N->getOperand(2);
6801 DebugLoc dl = N->getDebugLoc();
6803 // If the inserted element is an UNDEF, just use the input vector.
6804 if (InVal.getOpcode() == ISD::UNDEF)
6807 EVT VT = InVec.getValueType();
6809 // If we can't generate a legal BUILD_VECTOR, exit
6810 if (LegalOperations && !TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
6813 // Check that we know which element is being inserted
6814 if (!isa<ConstantSDNode>(EltNo))
6816 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
6818 // Check that the operand is a BUILD_VECTOR (or UNDEF, which can essentially
6819 // be converted to a BUILD_VECTOR). Fill in the Ops vector with the
6821 SmallVector<SDValue, 8> Ops;
6822 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
6823 Ops.append(InVec.getNode()->op_begin(),
6824 InVec.getNode()->op_end());
6825 } else if (InVec.getOpcode() == ISD::UNDEF) {
6826 unsigned NElts = VT.getVectorNumElements();
6827 Ops.append(NElts, DAG.getUNDEF(InVal.getValueType()));
6832 // Insert the element
6833 if (Elt < Ops.size()) {
6834 // All the operands of BUILD_VECTOR must have the same type;
6835 // we enforce that here.
6836 EVT OpVT = Ops[0].getValueType();
6837 if (InVal.getValueType() != OpVT)
6838 InVal = OpVT.bitsGT(InVal.getValueType()) ?
6839 DAG.getNode(ISD::ANY_EXTEND, dl, OpVT, InVal) :
6840 DAG.getNode(ISD::TRUNCATE, dl, OpVT, InVal);
6844 // Return the new vector
6845 return DAG.getNode(ISD::BUILD_VECTOR, dl,
6846 VT, &Ops[0], Ops.size());
6849 SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) {
6850 // (vextract (scalar_to_vector val, 0) -> val
6851 SDValue InVec = N->getOperand(0);
6853 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
6854 // Check if the result type doesn't match the inserted element type. A
6855 // SCALAR_TO_VECTOR may truncate the inserted element and the
6856 // EXTRACT_VECTOR_ELT may widen the extracted vector.
6857 SDValue InOp = InVec.getOperand(0);
6858 EVT NVT = N->getValueType(0);
6859 if (InOp.getValueType() != NVT) {
6860 assert(InOp.getValueType().isInteger() && NVT.isInteger());
6861 return DAG.getSExtOrTrunc(InOp, InVec.getDebugLoc(), NVT);
6866 // Perform only after legalization to ensure build_vector / vector_shuffle
6867 // optimizations have already been done.
6868 if (!LegalOperations) return SDValue();
6870 // (vextract (v4f32 load $addr), c) -> (f32 load $addr+c*size)
6871 // (vextract (v4f32 s2v (f32 load $addr)), c) -> (f32 load $addr+c*size)
6872 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), 0) -> (f32 load $addr)
6873 SDValue EltNo = N->getOperand(1);
6875 if (isa<ConstantSDNode>(EltNo)) {
6876 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
6877 bool NewLoad = false;
6878 bool BCNumEltsChanged = false;
6879 EVT VT = InVec.getValueType();
6880 EVT ExtVT = VT.getVectorElementType();
6883 if (InVec.getOpcode() == ISD::BITCAST) {
6884 EVT BCVT = InVec.getOperand(0).getValueType();
6885 if (!BCVT.isVector() || ExtVT.bitsGT(BCVT.getVectorElementType()))
6887 if (VT.getVectorNumElements() != BCVT.getVectorNumElements())
6888 BCNumEltsChanged = true;
6889 InVec = InVec.getOperand(0);
6890 ExtVT = BCVT.getVectorElementType();
6894 LoadSDNode *LN0 = NULL;
6895 const ShuffleVectorSDNode *SVN = NULL;
6896 if (ISD::isNormalLoad(InVec.getNode())) {
6897 LN0 = cast<LoadSDNode>(InVec);
6898 } else if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR &&
6899 InVec.getOperand(0).getValueType() == ExtVT &&
6900 ISD::isNormalLoad(InVec.getOperand(0).getNode())) {
6901 LN0 = cast<LoadSDNode>(InVec.getOperand(0));
6902 } else if ((SVN = dyn_cast<ShuffleVectorSDNode>(InVec))) {
6903 // (vextract (vector_shuffle (load $addr), v2, <1, u, u, u>), 1)
6905 // (load $addr+1*size)
6907 // If the bit convert changed the number of elements, it is unsafe
6908 // to examine the mask.
6909 if (BCNumEltsChanged)
6912 // Select the input vector, guarding against out of range extract vector.
6913 unsigned NumElems = VT.getVectorNumElements();
6914 int Idx = (Elt > (int)NumElems) ? -1 : SVN->getMaskElt(Elt);
6915 InVec = (Idx < (int)NumElems) ? InVec.getOperand(0) : InVec.getOperand(1);
6917 if (InVec.getOpcode() == ISD::BITCAST)
6918 InVec = InVec.getOperand(0);
6919 if (ISD::isNormalLoad(InVec.getNode())) {
6920 LN0 = cast<LoadSDNode>(InVec);
6921 Elt = (Idx < (int)NumElems) ? Idx : Idx - (int)NumElems;
6925 if (!LN0 || !LN0->hasNUsesOfValue(1,0) || LN0->isVolatile())
6928 // If Idx was -1 above, Elt is going to be -1, so just return undef.
6930 return DAG.getUNDEF(LVT);
6932 unsigned Align = LN0->getAlignment();
6934 // Check the resultant load doesn't need a higher alignment than the
6938 ->getABITypeAlignment(LVT.getTypeForEVT(*DAG.getContext()));
6940 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, LVT))
6946 SDValue NewPtr = LN0->getBasePtr();
6947 unsigned PtrOff = 0;
6950 PtrOff = LVT.getSizeInBits() * Elt / 8;
6951 EVT PtrType = NewPtr.getValueType();
6952 if (TLI.isBigEndian())
6953 PtrOff = VT.getSizeInBits() / 8 - PtrOff;
6954 NewPtr = DAG.getNode(ISD::ADD, N->getDebugLoc(), PtrType, NewPtr,
6955 DAG.getConstant(PtrOff, PtrType));
6958 // The replacement we need to do here is a little tricky: we need to
6959 // replace an extractelement of a load with a load.
6960 // Use ReplaceAllUsesOfValuesWith to do the replacement.
6961 SDValue Load = DAG.getLoad(LVT, N->getDebugLoc(), LN0->getChain(), NewPtr,
6962 LN0->getPointerInfo().getWithOffset(PtrOff),
6963 LN0->isVolatile(), LN0->isNonTemporal(),
6964 LN0->isInvariant(), Align);
6965 WorkListRemover DeadNodes(*this);
6966 SDValue From[] = { SDValue(N, 0), SDValue(LN0,1) };
6967 SDValue To[] = { Load.getValue(0), Load.getValue(1) };
6968 DAG.ReplaceAllUsesOfValuesWith(From, To, 2, &DeadNodes);
6969 // Since we're explcitly calling ReplaceAllUses, add the new node to the
6970 // worklist explicitly as well.
6971 AddToWorkList(Load.getNode());
6972 // Make sure to revisit this node to clean it up; it will usually be dead.
6974 return SDValue(N, 0);
6980 SDValue DAGCombiner::visitBUILD_VECTOR(SDNode *N) {
6981 unsigned NumInScalars = N->getNumOperands();
6982 DebugLoc dl = N->getDebugLoc();
6983 EVT VT = N->getValueType(0);
6984 // Check to see if this is a BUILD_VECTOR of a bunch of values
6985 // which come from any_extend or zero_extend nodes. If so, we can create
6986 // a new BUILD_VECTOR using bit-casts which may enable other BUILD_VECTOR
6987 // optimizations. We do not handle sign-extend because we can't fill the sign
6989 EVT SourceType = MVT::Other;
6990 bool allAnyExt = true;
6991 for (unsigned i = 0; i < NumInScalars; ++i) {
6992 SDValue In = N->getOperand(i);
6993 // Ignore undef inputs.
6994 if (In.getOpcode() == ISD::UNDEF) continue;
6996 bool AnyExt = In.getOpcode() == ISD::ANY_EXTEND;
6997 bool ZeroExt = In.getOpcode() == ISD::ZERO_EXTEND;
6999 // Abort if the element is not an extension.
7000 if (!ZeroExt && !AnyExt) {
7001 SourceType = MVT::Other;
7005 // The input is a ZeroExt or AnyExt. Check the original type.
7006 EVT InTy = In.getOperand(0).getValueType();
7008 // Check that all of the widened source types are the same.
7009 if (SourceType == MVT::Other)
7012 else if (InTy != SourceType) {
7013 // Multiple income types. Abort.
7014 SourceType = MVT::Other;
7018 // Check if all of the extends are ANY_EXTENDs.
7019 allAnyExt &= AnyExt;
7023 // In order to have valid types, all of the inputs must be extended from the
7024 // same source type and all of the inputs must be any or zero extend.
7025 // Scalar sizes must be a power of two.
7026 EVT OutScalarTy = N->getValueType(0).getScalarType();
7027 bool validTypes = SourceType != MVT::Other &&
7028 isPowerOf2_32(OutScalarTy.getSizeInBits()) &&
7029 isPowerOf2_32(SourceType.getSizeInBits());
7031 // We perform this optimization post type-legalization because
7032 // the type-legalizer often scalarizes integer-promoted vectors.
7033 // Performing this optimization before may create bit-casts which
7034 // will be type-legalized to complex code sequences.
7035 // We perform this optimization only before the operation legalizer because we
7036 // may introduce illegal operations.
7037 if (LegalTypes && !LegalOperations && validTypes) {
7038 bool isLE = TLI.isLittleEndian();
7039 unsigned ElemRatio = OutScalarTy.getSizeInBits()/SourceType.getSizeInBits();
7040 assert(ElemRatio > 1 && "Invalid element size ratio");
7041 SDValue Filler = allAnyExt ? DAG.getUNDEF(SourceType):
7042 DAG.getConstant(0, SourceType);
7044 unsigned NewBVElems = ElemRatio * N->getValueType(0).getVectorNumElements();
7045 SmallVector<SDValue, 8> Ops(NewBVElems, Filler);
7047 // Populate the new build_vector
7048 for (unsigned i=0; i < N->getNumOperands(); ++i) {
7049 SDValue Cast = N->getOperand(i);
7050 assert((Cast.getOpcode() == ISD::ANY_EXTEND ||
7051 Cast.getOpcode() == ISD::ZERO_EXTEND ||
7052 Cast.getOpcode() == ISD::UNDEF) && "Invalid cast opcode");
7054 if (Cast.getOpcode() == ISD::UNDEF)
7055 In = DAG.getUNDEF(SourceType);
7057 In = Cast->getOperand(0);
7058 unsigned Index = isLE ? (i * ElemRatio) :
7059 (i * ElemRatio + (ElemRatio - 1));
7061 assert(Index < Ops.size() && "Invalid index");
7065 // The type of the new BUILD_VECTOR node.
7066 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), SourceType, NewBVElems);
7067 assert(VecVT.getSizeInBits() == N->getValueType(0).getSizeInBits() &&
7068 "Invalid vector size");
7069 // Check if the new vector type is legal.
7070 if (!isTypeLegal(VecVT)) return SDValue();
7072 // Make the new BUILD_VECTOR.
7073 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
7074 VecVT, &Ops[0], Ops.size());
7076 // Bitcast to the desired type.
7077 return DAG.getNode(ISD::BITCAST, dl, N->getValueType(0), BV);
7080 // Check to see if this is a BUILD_VECTOR of a bunch of EXTRACT_VECTOR_ELT
7081 // operations. If so, and if the EXTRACT_VECTOR_ELT vector inputs come from
7082 // at most two distinct vectors, turn this into a shuffle node.
7083 SDValue VecIn1, VecIn2;
7084 for (unsigned i = 0; i != NumInScalars; ++i) {
7085 // Ignore undef inputs.
7086 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
7088 // If this input is something other than a EXTRACT_VECTOR_ELT with a
7089 // constant index, bail out.
7090 if (N->getOperand(i).getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
7091 !isa<ConstantSDNode>(N->getOperand(i).getOperand(1))) {
7092 VecIn1 = VecIn2 = SDValue(0, 0);
7096 // If the input vector type disagrees with the result of the build_vector,
7097 // we can't make a shuffle.
7098 SDValue ExtractedFromVec = N->getOperand(i).getOperand(0);
7099 if (ExtractedFromVec.getValueType() != VT) {
7100 VecIn1 = VecIn2 = SDValue(0, 0);
7104 // Otherwise, remember this. We allow up to two distinct input vectors.
7105 if (ExtractedFromVec == VecIn1 || ExtractedFromVec == VecIn2)
7108 if (VecIn1.getNode() == 0) {
7109 VecIn1 = ExtractedFromVec;
7110 } else if (VecIn2.getNode() == 0) {
7111 VecIn2 = ExtractedFromVec;
7114 VecIn1 = VecIn2 = SDValue(0, 0);
7119 // If everything is good, we can make a shuffle operation.
7120 if (VecIn1.getNode()) {
7121 SmallVector<int, 8> Mask;
7122 for (unsigned i = 0; i != NumInScalars; ++i) {
7123 if (N->getOperand(i).getOpcode() == ISD::UNDEF) {
7128 // If extracting from the first vector, just use the index directly.
7129 SDValue Extract = N->getOperand(i);
7130 SDValue ExtVal = Extract.getOperand(1);
7131 if (Extract.getOperand(0) == VecIn1) {
7132 unsigned ExtIndex = cast<ConstantSDNode>(ExtVal)->getZExtValue();
7133 if (ExtIndex > VT.getVectorNumElements())
7136 Mask.push_back(ExtIndex);
7140 // Otherwise, use InIdx + VecSize
7141 unsigned Idx = cast<ConstantSDNode>(ExtVal)->getZExtValue();
7142 Mask.push_back(Idx+NumInScalars);
7145 // Add count and size info.
7146 if (!isTypeLegal(VT))
7149 // Return the new VECTOR_SHUFFLE node.
7152 Ops[1] = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
7153 return DAG.getVectorShuffle(VT, N->getDebugLoc(), Ops[0], Ops[1], &Mask[0]);
7159 SDValue DAGCombiner::visitCONCAT_VECTORS(SDNode *N) {
7160 // TODO: Check to see if this is a CONCAT_VECTORS of a bunch of
7161 // EXTRACT_SUBVECTOR operations. If so, and if the EXTRACT_SUBVECTOR vector
7162 // inputs come from at most two distinct vectors, turn this into a shuffle
7165 // If we only have one input vector, we don't need to do any concatenation.
7166 if (N->getNumOperands() == 1)
7167 return N->getOperand(0);
7172 SDValue DAGCombiner::visitEXTRACT_SUBVECTOR(SDNode* N) {
7173 EVT NVT = N->getValueType(0);
7174 SDValue V = N->getOperand(0);
7176 if (V->getOpcode() == ISD::INSERT_SUBVECTOR) {
7177 // Handle only simple case where vector being inserted and vector
7178 // being extracted are of same type, and are half size of larger vectors.
7179 EVT BigVT = V->getOperand(0).getValueType();
7180 EVT SmallVT = V->getOperand(1).getValueType();
7181 if (NVT != SmallVT || NVT.getSizeInBits()*2 != BigVT.getSizeInBits())
7184 // Only handle cases where both indexes are constants with the same type.
7185 ConstantSDNode *InsIdx = dyn_cast<ConstantSDNode>(N->getOperand(1));
7186 ConstantSDNode *ExtIdx = dyn_cast<ConstantSDNode>(V->getOperand(2));
7188 if (InsIdx && ExtIdx &&
7189 InsIdx->getValueType(0).getSizeInBits() <= 64 &&
7190 ExtIdx->getValueType(0).getSizeInBits() <= 64) {
7192 // (extract_subvec (insert_subvec V1, V2, InsIdx), ExtIdx)
7194 // indices are equal => V1
7195 // otherwise => (extract_subvec V1, ExtIdx)
7196 if (InsIdx->getZExtValue() == ExtIdx->getZExtValue())
7197 return V->getOperand(1);
7198 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, N->getDebugLoc(), NVT,
7199 V->getOperand(0), N->getOperand(1));
7206 SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
7207 EVT VT = N->getValueType(0);
7208 unsigned NumElts = VT.getVectorNumElements();
7210 SDValue N0 = N->getOperand(0);
7212 assert(N0.getValueType().getVectorNumElements() == NumElts &&
7213 "Vector shuffle must be normalized in DAG");
7215 // FIXME: implement canonicalizations from DAG.getVectorShuffle()
7217 // If it is a splat, check if the argument vector is another splat or a
7218 // build_vector with all scalar elements the same.
7219 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
7220 if (SVN->isSplat() && SVN->getSplatIndex() < (int)NumElts) {
7221 SDNode *V = N0.getNode();
7223 // If this is a bit convert that changes the element type of the vector but
7224 // not the number of vector elements, look through it. Be careful not to
7225 // look though conversions that change things like v4f32 to v2f64.
7226 if (V->getOpcode() == ISD::BITCAST) {
7227 SDValue ConvInput = V->getOperand(0);
7228 if (ConvInput.getValueType().isVector() &&
7229 ConvInput.getValueType().getVectorNumElements() == NumElts)
7230 V = ConvInput.getNode();
7233 if (V->getOpcode() == ISD::BUILD_VECTOR) {
7234 assert(V->getNumOperands() == NumElts &&
7235 "BUILD_VECTOR has wrong number of operands");
7237 bool AllSame = true;
7238 for (unsigned i = 0; i != NumElts; ++i) {
7239 if (V->getOperand(i).getOpcode() != ISD::UNDEF) {
7240 Base = V->getOperand(i);
7244 // Splat of <u, u, u, u>, return <u, u, u, u>
7245 if (!Base.getNode())
7247 for (unsigned i = 0; i != NumElts; ++i) {
7248 if (V->getOperand(i) != Base) {
7253 // Splat of <x, x, x, x>, return <x, x, x, x>
7261 SDValue DAGCombiner::visitMEMBARRIER(SDNode* N) {
7262 if (!TLI.getShouldFoldAtomicFences())
7265 SDValue atomic = N->getOperand(0);
7266 switch (atomic.getOpcode()) {
7267 case ISD::ATOMIC_CMP_SWAP:
7268 case ISD::ATOMIC_SWAP:
7269 case ISD::ATOMIC_LOAD_ADD:
7270 case ISD::ATOMIC_LOAD_SUB:
7271 case ISD::ATOMIC_LOAD_AND:
7272 case ISD::ATOMIC_LOAD_OR:
7273 case ISD::ATOMIC_LOAD_XOR:
7274 case ISD::ATOMIC_LOAD_NAND:
7275 case ISD::ATOMIC_LOAD_MIN:
7276 case ISD::ATOMIC_LOAD_MAX:
7277 case ISD::ATOMIC_LOAD_UMIN:
7278 case ISD::ATOMIC_LOAD_UMAX:
7284 SDValue fence = atomic.getOperand(0);
7285 if (fence.getOpcode() != ISD::MEMBARRIER)
7288 switch (atomic.getOpcode()) {
7289 case ISD::ATOMIC_CMP_SWAP:
7290 return SDValue(DAG.UpdateNodeOperands(atomic.getNode(),
7291 fence.getOperand(0),
7292 atomic.getOperand(1), atomic.getOperand(2),
7293 atomic.getOperand(3)), atomic.getResNo());
7294 case ISD::ATOMIC_SWAP:
7295 case ISD::ATOMIC_LOAD_ADD:
7296 case ISD::ATOMIC_LOAD_SUB:
7297 case ISD::ATOMIC_LOAD_AND:
7298 case ISD::ATOMIC_LOAD_OR:
7299 case ISD::ATOMIC_LOAD_XOR:
7300 case ISD::ATOMIC_LOAD_NAND:
7301 case ISD::ATOMIC_LOAD_MIN:
7302 case ISD::ATOMIC_LOAD_MAX:
7303 case ISD::ATOMIC_LOAD_UMIN:
7304 case ISD::ATOMIC_LOAD_UMAX:
7305 return SDValue(DAG.UpdateNodeOperands(atomic.getNode(),
7306 fence.getOperand(0),
7307 atomic.getOperand(1), atomic.getOperand(2)),
7314 /// XformToShuffleWithZero - Returns a vector_shuffle if it able to transform
7315 /// an AND to a vector_shuffle with the destination vector and a zero vector.
7316 /// e.g. AND V, <0xffffffff, 0, 0xffffffff, 0>. ==>
7317 /// vector_shuffle V, Zero, <0, 4, 2, 4>
7318 SDValue DAGCombiner::XformToShuffleWithZero(SDNode *N) {
7319 EVT VT = N->getValueType(0);
7320 DebugLoc dl = N->getDebugLoc();
7321 SDValue LHS = N->getOperand(0);
7322 SDValue RHS = N->getOperand(1);
7323 if (N->getOpcode() == ISD::AND) {
7324 if (RHS.getOpcode() == ISD::BITCAST)
7325 RHS = RHS.getOperand(0);
7326 if (RHS.getOpcode() == ISD::BUILD_VECTOR) {
7327 SmallVector<int, 8> Indices;
7328 unsigned NumElts = RHS.getNumOperands();
7329 for (unsigned i = 0; i != NumElts; ++i) {
7330 SDValue Elt = RHS.getOperand(i);
7331 if (!isa<ConstantSDNode>(Elt))
7333 else if (cast<ConstantSDNode>(Elt)->isAllOnesValue())
7334 Indices.push_back(i);
7335 else if (cast<ConstantSDNode>(Elt)->isNullValue())
7336 Indices.push_back(NumElts);
7341 // Let's see if the target supports this vector_shuffle.
7342 EVT RVT = RHS.getValueType();
7343 if (!TLI.isVectorClearMaskLegal(Indices, RVT))
7346 // Return the new VECTOR_SHUFFLE node.
7347 EVT EltVT = RVT.getVectorElementType();
7348 SmallVector<SDValue,8> ZeroOps(RVT.getVectorNumElements(),
7349 DAG.getConstant(0, EltVT));
7350 SDValue Zero = DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
7351 RVT, &ZeroOps[0], ZeroOps.size());
7352 LHS = DAG.getNode(ISD::BITCAST, dl, RVT, LHS);
7353 SDValue Shuf = DAG.getVectorShuffle(RVT, dl, LHS, Zero, &Indices[0]);
7354 return DAG.getNode(ISD::BITCAST, dl, VT, Shuf);
7361 /// SimplifyVBinOp - Visit a binary vector operation, like ADD.
7362 SDValue DAGCombiner::SimplifyVBinOp(SDNode *N) {
7363 // After legalize, the target may be depending on adds and other
7364 // binary ops to provide legal ways to construct constants or other
7365 // things. Simplifying them may result in a loss of legality.
7366 if (LegalOperations) return SDValue();
7368 assert(N->getValueType(0).isVector() &&
7369 "SimplifyVBinOp only works on vectors!");
7371 SDValue LHS = N->getOperand(0);
7372 SDValue RHS = N->getOperand(1);
7373 SDValue Shuffle = XformToShuffleWithZero(N);
7374 if (Shuffle.getNode()) return Shuffle;
7376 // If the LHS and RHS are BUILD_VECTOR nodes, see if we can constant fold
7378 if (LHS.getOpcode() == ISD::BUILD_VECTOR &&
7379 RHS.getOpcode() == ISD::BUILD_VECTOR) {
7380 SmallVector<SDValue, 8> Ops;
7381 for (unsigned i = 0, e = LHS.getNumOperands(); i != e; ++i) {
7382 SDValue LHSOp = LHS.getOperand(i);
7383 SDValue RHSOp = RHS.getOperand(i);
7384 // If these two elements can't be folded, bail out.
7385 if ((LHSOp.getOpcode() != ISD::UNDEF &&
7386 LHSOp.getOpcode() != ISD::Constant &&
7387 LHSOp.getOpcode() != ISD::ConstantFP) ||
7388 (RHSOp.getOpcode() != ISD::UNDEF &&
7389 RHSOp.getOpcode() != ISD::Constant &&
7390 RHSOp.getOpcode() != ISD::ConstantFP))
7393 // Can't fold divide by zero.
7394 if (N->getOpcode() == ISD::SDIV || N->getOpcode() == ISD::UDIV ||
7395 N->getOpcode() == ISD::FDIV) {
7396 if ((RHSOp.getOpcode() == ISD::Constant &&
7397 cast<ConstantSDNode>(RHSOp.getNode())->isNullValue()) ||
7398 (RHSOp.getOpcode() == ISD::ConstantFP &&
7399 cast<ConstantFPSDNode>(RHSOp.getNode())->getValueAPF().isZero()))
7403 EVT VT = LHSOp.getValueType();
7404 EVT RVT = RHSOp.getValueType();
7406 // Integer BUILD_VECTOR operands may have types larger than the element
7407 // size (e.g., when the element type is not legal). Prior to type
7408 // legalization, the types may not match between the two BUILD_VECTORS.
7409 // Truncate one of the operands to make them match.
7410 if (RVT.getSizeInBits() > VT.getSizeInBits()) {
7411 RHSOp = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, RHSOp);
7413 LHSOp = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), RVT, LHSOp);
7417 SDValue FoldOp = DAG.getNode(N->getOpcode(), LHS.getDebugLoc(), VT,
7419 if (FoldOp.getOpcode() != ISD::UNDEF &&
7420 FoldOp.getOpcode() != ISD::Constant &&
7421 FoldOp.getOpcode() != ISD::ConstantFP)
7423 Ops.push_back(FoldOp);
7424 AddToWorkList(FoldOp.getNode());
7427 if (Ops.size() == LHS.getNumOperands())
7428 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
7429 LHS.getValueType(), &Ops[0], Ops.size());
7435 SDValue DAGCombiner::SimplifySelect(DebugLoc DL, SDValue N0,
7436 SDValue N1, SDValue N2){
7437 assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!");
7439 SDValue SCC = SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1), N1, N2,
7440 cast<CondCodeSDNode>(N0.getOperand(2))->get());
7442 // If we got a simplified select_cc node back from SimplifySelectCC, then
7443 // break it down into a new SETCC node, and a new SELECT node, and then return
7444 // the SELECT node, since we were called with a SELECT node.
7445 if (SCC.getNode()) {
7446 // Check to see if we got a select_cc back (to turn into setcc/select).
7447 // Otherwise, just return whatever node we got back, like fabs.
7448 if (SCC.getOpcode() == ISD::SELECT_CC) {
7449 SDValue SETCC = DAG.getNode(ISD::SETCC, N0.getDebugLoc(),
7451 SCC.getOperand(0), SCC.getOperand(1),
7453 AddToWorkList(SETCC.getNode());
7454 return DAG.getNode(ISD::SELECT, SCC.getDebugLoc(), SCC.getValueType(),
7455 SCC.getOperand(2), SCC.getOperand(3), SETCC);
7463 /// SimplifySelectOps - Given a SELECT or a SELECT_CC node, where LHS and RHS
7464 /// are the two values being selected between, see if we can simplify the
7465 /// select. Callers of this should assume that TheSelect is deleted if this
7466 /// returns true. As such, they should return the appropriate thing (e.g. the
7467 /// node) back to the top-level of the DAG combiner loop to avoid it being
7469 bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDValue LHS,
7472 // Cannot simplify select with vector condition
7473 if (TheSelect->getOperand(0).getValueType().isVector()) return false;
7475 // If this is a select from two identical things, try to pull the operation
7476 // through the select.
7477 if (LHS.getOpcode() != RHS.getOpcode() ||
7478 !LHS.hasOneUse() || !RHS.hasOneUse())
7481 // If this is a load and the token chain is identical, replace the select
7482 // of two loads with a load through a select of the address to load from.
7483 // This triggers in things like "select bool X, 10.0, 123.0" after the FP
7484 // constants have been dropped into the constant pool.
7485 if (LHS.getOpcode() == ISD::LOAD) {
7486 LoadSDNode *LLD = cast<LoadSDNode>(LHS);
7487 LoadSDNode *RLD = cast<LoadSDNode>(RHS);
7489 // Token chains must be identical.
7490 if (LHS.getOperand(0) != RHS.getOperand(0) ||
7491 // Do not let this transformation reduce the number of volatile loads.
7492 LLD->isVolatile() || RLD->isVolatile() ||
7493 // If this is an EXTLOAD, the VT's must match.
7494 LLD->getMemoryVT() != RLD->getMemoryVT() ||
7495 // If this is an EXTLOAD, the kind of extension must match.
7496 (LLD->getExtensionType() != RLD->getExtensionType() &&
7497 // The only exception is if one of the extensions is anyext.
7498 LLD->getExtensionType() != ISD::EXTLOAD &&
7499 RLD->getExtensionType() != ISD::EXTLOAD) ||
7500 // FIXME: this discards src value information. This is
7501 // over-conservative. It would be beneficial to be able to remember
7502 // both potential memory locations. Since we are discarding
7503 // src value info, don't do the transformation if the memory
7504 // locations are not in the default address space.
7505 LLD->getPointerInfo().getAddrSpace() != 0 ||
7506 RLD->getPointerInfo().getAddrSpace() != 0)
7509 // Check that the select condition doesn't reach either load. If so,
7510 // folding this will induce a cycle into the DAG. If not, this is safe to
7511 // xform, so create a select of the addresses.
7513 if (TheSelect->getOpcode() == ISD::SELECT) {
7514 SDNode *CondNode = TheSelect->getOperand(0).getNode();
7515 if ((LLD->hasAnyUseOfValue(1) && LLD->isPredecessorOf(CondNode)) ||
7516 (RLD->hasAnyUseOfValue(1) && RLD->isPredecessorOf(CondNode)))
7518 Addr = DAG.getNode(ISD::SELECT, TheSelect->getDebugLoc(),
7519 LLD->getBasePtr().getValueType(),
7520 TheSelect->getOperand(0), LLD->getBasePtr(),
7522 } else { // Otherwise SELECT_CC
7523 SDNode *CondLHS = TheSelect->getOperand(0).getNode();
7524 SDNode *CondRHS = TheSelect->getOperand(1).getNode();
7526 if ((LLD->hasAnyUseOfValue(1) &&
7527 (LLD->isPredecessorOf(CondLHS) || LLD->isPredecessorOf(CondRHS))) ||
7528 (LLD->hasAnyUseOfValue(1) &&
7529 (LLD->isPredecessorOf(CondLHS) || LLD->isPredecessorOf(CondRHS))))
7532 Addr = DAG.getNode(ISD::SELECT_CC, TheSelect->getDebugLoc(),
7533 LLD->getBasePtr().getValueType(),
7534 TheSelect->getOperand(0),
7535 TheSelect->getOperand(1),
7536 LLD->getBasePtr(), RLD->getBasePtr(),
7537 TheSelect->getOperand(4));
7541 if (LLD->getExtensionType() == ISD::NON_EXTLOAD) {
7542 Load = DAG.getLoad(TheSelect->getValueType(0),
7543 TheSelect->getDebugLoc(),
7544 // FIXME: Discards pointer info.
7545 LLD->getChain(), Addr, MachinePointerInfo(),
7546 LLD->isVolatile(), LLD->isNonTemporal(),
7547 LLD->isInvariant(), LLD->getAlignment());
7549 Load = DAG.getExtLoad(LLD->getExtensionType() == ISD::EXTLOAD ?
7550 RLD->getExtensionType() : LLD->getExtensionType(),
7551 TheSelect->getDebugLoc(),
7552 TheSelect->getValueType(0),
7553 // FIXME: Discards pointer info.
7554 LLD->getChain(), Addr, MachinePointerInfo(),
7555 LLD->getMemoryVT(), LLD->isVolatile(),
7556 LLD->isNonTemporal(), LLD->getAlignment());
7559 // Users of the select now use the result of the load.
7560 CombineTo(TheSelect, Load);
7562 // Users of the old loads now use the new load's chain. We know the
7563 // old-load value is dead now.
7564 CombineTo(LHS.getNode(), Load.getValue(0), Load.getValue(1));
7565 CombineTo(RHS.getNode(), Load.getValue(0), Load.getValue(1));
7572 /// SimplifySelectCC - Simplify an expression of the form (N0 cond N1) ? N2 : N3
7573 /// where 'cond' is the comparison specified by CC.
7574 SDValue DAGCombiner::SimplifySelectCC(DebugLoc DL, SDValue N0, SDValue N1,
7575 SDValue N2, SDValue N3,
7576 ISD::CondCode CC, bool NotExtCompare) {
7577 // (x ? y : y) -> y.
7578 if (N2 == N3) return N2;
7580 EVT VT = N2.getValueType();
7581 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
7582 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
7583 ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N3.getNode());
7585 // Determine if the condition we're dealing with is constant
7586 SDValue SCC = SimplifySetCC(TLI.getSetCCResultType(N0.getValueType()),
7587 N0, N1, CC, DL, false);
7588 if (SCC.getNode()) AddToWorkList(SCC.getNode());
7589 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode());
7591 // fold select_cc true, x, y -> x
7592 if (SCCC && !SCCC->isNullValue())
7594 // fold select_cc false, x, y -> y
7595 if (SCCC && SCCC->isNullValue())
7598 // Check to see if we can simplify the select into an fabs node
7599 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) {
7600 // Allow either -0.0 or 0.0
7601 if (CFP->getValueAPF().isZero()) {
7602 // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs
7603 if ((CC == ISD::SETGE || CC == ISD::SETGT) &&
7604 N0 == N2 && N3.getOpcode() == ISD::FNEG &&
7605 N2 == N3.getOperand(0))
7606 return DAG.getNode(ISD::FABS, DL, VT, N0);
7608 // select (setl[te] X, +/-0.0), fneg(X), X -> fabs
7609 if ((CC == ISD::SETLT || CC == ISD::SETLE) &&
7610 N0 == N3 && N2.getOpcode() == ISD::FNEG &&
7611 N2.getOperand(0) == N3)
7612 return DAG.getNode(ISD::FABS, DL, VT, N3);
7616 // Turn "(a cond b) ? 1.0f : 2.0f" into "load (tmp + ((a cond b) ? 0 : 4)"
7617 // where "tmp" is a constant pool entry containing an array with 1.0 and 2.0
7618 // in it. This is a win when the constant is not otherwise available because
7619 // it replaces two constant pool loads with one. We only do this if the FP
7620 // type is known to be legal, because if it isn't, then we are before legalize
7621 // types an we want the other legalization to happen first (e.g. to avoid
7622 // messing with soft float) and if the ConstantFP is not legal, because if
7623 // it is legal, we may not need to store the FP constant in a constant pool.
7624 if (ConstantFPSDNode *TV = dyn_cast<ConstantFPSDNode>(N2))
7625 if (ConstantFPSDNode *FV = dyn_cast<ConstantFPSDNode>(N3)) {
7626 if (TLI.isTypeLegal(N2.getValueType()) &&
7627 (TLI.getOperationAction(ISD::ConstantFP, N2.getValueType()) !=
7628 TargetLowering::Legal) &&
7629 // If both constants have multiple uses, then we won't need to do an
7630 // extra load, they are likely around in registers for other users.
7631 (TV->hasOneUse() || FV->hasOneUse())) {
7632 Constant *Elts[] = {
7633 const_cast<ConstantFP*>(FV->getConstantFPValue()),
7634 const_cast<ConstantFP*>(TV->getConstantFPValue())
7636 Type *FPTy = Elts[0]->getType();
7637 const TargetData &TD = *TLI.getTargetData();
7639 // Create a ConstantArray of the two constants.
7640 Constant *CA = ConstantArray::get(ArrayType::get(FPTy, 2), Elts);
7641 SDValue CPIdx = DAG.getConstantPool(CA, TLI.getPointerTy(),
7642 TD.getPrefTypeAlignment(FPTy));
7643 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
7645 // Get the offsets to the 0 and 1 element of the array so that we can
7646 // select between them.
7647 SDValue Zero = DAG.getIntPtrConstant(0);
7648 unsigned EltSize = (unsigned)TD.getTypeAllocSize(Elts[0]->getType());
7649 SDValue One = DAG.getIntPtrConstant(EltSize);
7651 SDValue Cond = DAG.getSetCC(DL,
7652 TLI.getSetCCResultType(N0.getValueType()),
7654 AddToWorkList(Cond.getNode());
7655 SDValue CstOffset = DAG.getNode(ISD::SELECT, DL, Zero.getValueType(),
7657 AddToWorkList(CstOffset.getNode());
7658 CPIdx = DAG.getNode(ISD::ADD, DL, TLI.getPointerTy(), CPIdx,
7660 AddToWorkList(CPIdx.getNode());
7661 return DAG.getLoad(TV->getValueType(0), DL, DAG.getEntryNode(), CPIdx,
7662 MachinePointerInfo::getConstantPool(), false,
7663 false, false, Alignment);
7668 // Check to see if we can perform the "gzip trick", transforming
7669 // (select_cc setlt X, 0, A, 0) -> (and (sra X, (sub size(X), 1), A)
7670 if (N1C && N3C && N3C->isNullValue() && CC == ISD::SETLT &&
7671 (N1C->isNullValue() || // (a < 0) ? b : 0
7672 (N1C->getAPIntValue() == 1 && N0 == N2))) { // (a < 1) ? a : 0
7673 EVT XType = N0.getValueType();
7674 EVT AType = N2.getValueType();
7675 if (XType.bitsGE(AType)) {
7676 // and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a
7677 // single-bit constant.
7678 if (N2C && ((N2C->getAPIntValue() & (N2C->getAPIntValue()-1)) == 0)) {
7679 unsigned ShCtV = N2C->getAPIntValue().logBase2();
7680 ShCtV = XType.getSizeInBits()-ShCtV-1;
7681 SDValue ShCt = DAG.getConstant(ShCtV,
7682 getShiftAmountTy(N0.getValueType()));
7683 SDValue Shift = DAG.getNode(ISD::SRL, N0.getDebugLoc(),
7685 AddToWorkList(Shift.getNode());
7687 if (XType.bitsGT(AType)) {
7688 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
7689 AddToWorkList(Shift.getNode());
7692 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
7695 SDValue Shift = DAG.getNode(ISD::SRA, N0.getDebugLoc(),
7697 DAG.getConstant(XType.getSizeInBits()-1,
7698 getShiftAmountTy(N0.getValueType())));
7699 AddToWorkList(Shift.getNode());
7701 if (XType.bitsGT(AType)) {
7702 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
7703 AddToWorkList(Shift.getNode());
7706 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
7710 // fold (select_cc seteq (and x, y), 0, 0, A) -> (and (shr (shl x)) A)
7711 // where y is has a single bit set.
7712 // A plaintext description would be, we can turn the SELECT_CC into an AND
7713 // when the condition can be materialized as an all-ones register. Any
7714 // single bit-test can be materialized as an all-ones register with
7715 // shift-left and shift-right-arith.
7716 if (CC == ISD::SETEQ && N0->getOpcode() == ISD::AND &&
7717 N0->getValueType(0) == VT &&
7718 N1C && N1C->isNullValue() &&
7719 N2C && N2C->isNullValue()) {
7720 SDValue AndLHS = N0->getOperand(0);
7721 ConstantSDNode *ConstAndRHS = dyn_cast<ConstantSDNode>(N0->getOperand(1));
7722 if (ConstAndRHS && ConstAndRHS->getAPIntValue().countPopulation() == 1) {
7723 // Shift the tested bit over the sign bit.
7724 APInt AndMask = ConstAndRHS->getAPIntValue();
7726 DAG.getConstant(AndMask.countLeadingZeros(),
7727 getShiftAmountTy(AndLHS.getValueType()));
7728 SDValue Shl = DAG.getNode(ISD::SHL, N0.getDebugLoc(), VT, AndLHS, ShlAmt);
7730 // Now arithmetic right shift it all the way over, so the result is either
7731 // all-ones, or zero.
7733 DAG.getConstant(AndMask.getBitWidth()-1,
7734 getShiftAmountTy(Shl.getValueType()));
7735 SDValue Shr = DAG.getNode(ISD::SRA, N0.getDebugLoc(), VT, Shl, ShrAmt);
7737 return DAG.getNode(ISD::AND, DL, VT, Shr, N3);
7741 // fold select C, 16, 0 -> shl C, 4
7742 if (N2C && N3C && N3C->isNullValue() && N2C->getAPIntValue().isPowerOf2() &&
7743 TLI.getBooleanContents(N0.getValueType().isVector()) ==
7744 TargetLowering::ZeroOrOneBooleanContent) {
7746 // If the caller doesn't want us to simplify this into a zext of a compare,
7748 if (NotExtCompare && N2C->getAPIntValue() == 1)
7751 // Get a SetCC of the condition
7752 // FIXME: Should probably make sure that setcc is legal if we ever have a
7753 // target where it isn't.
7755 // cast from setcc result type to select result type
7757 SCC = DAG.getSetCC(DL, TLI.getSetCCResultType(N0.getValueType()),
7759 if (N2.getValueType().bitsLT(SCC.getValueType()))
7760 Temp = DAG.getZeroExtendInReg(SCC, N2.getDebugLoc(), N2.getValueType());
7762 Temp = DAG.getNode(ISD::ZERO_EXTEND, N2.getDebugLoc(),
7763 N2.getValueType(), SCC);
7765 SCC = DAG.getSetCC(N0.getDebugLoc(), MVT::i1, N0, N1, CC);
7766 Temp = DAG.getNode(ISD::ZERO_EXTEND, N2.getDebugLoc(),
7767 N2.getValueType(), SCC);
7770 AddToWorkList(SCC.getNode());
7771 AddToWorkList(Temp.getNode());
7773 if (N2C->getAPIntValue() == 1)
7776 // shl setcc result by log2 n2c
7777 return DAG.getNode(ISD::SHL, DL, N2.getValueType(), Temp,
7778 DAG.getConstant(N2C->getAPIntValue().logBase2(),
7779 getShiftAmountTy(Temp.getValueType())));
7782 // Check to see if this is the equivalent of setcc
7783 // FIXME: Turn all of these into setcc if setcc if setcc is legal
7784 // otherwise, go ahead with the folds.
7785 if (0 && N3C && N3C->isNullValue() && N2C && (N2C->getAPIntValue() == 1ULL)) {
7786 EVT XType = N0.getValueType();
7787 if (!LegalOperations ||
7788 TLI.isOperationLegal(ISD::SETCC, TLI.getSetCCResultType(XType))) {
7789 SDValue Res = DAG.getSetCC(DL, TLI.getSetCCResultType(XType), N0, N1, CC);
7790 if (Res.getValueType() != VT)
7791 Res = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Res);
7795 // fold (seteq X, 0) -> (srl (ctlz X, log2(size(X))))
7796 if (N1C && N1C->isNullValue() && CC == ISD::SETEQ &&
7797 (!LegalOperations ||
7798 TLI.isOperationLegal(ISD::CTLZ, XType))) {
7799 SDValue Ctlz = DAG.getNode(ISD::CTLZ, N0.getDebugLoc(), XType, N0);
7800 return DAG.getNode(ISD::SRL, DL, XType, Ctlz,
7801 DAG.getConstant(Log2_32(XType.getSizeInBits()),
7802 getShiftAmountTy(Ctlz.getValueType())));
7804 // fold (setgt X, 0) -> (srl (and (-X, ~X), size(X)-1))
7805 if (N1C && N1C->isNullValue() && CC == ISD::SETGT) {
7806 SDValue NegN0 = DAG.getNode(ISD::SUB, N0.getDebugLoc(),
7807 XType, DAG.getConstant(0, XType), N0);
7808 SDValue NotN0 = DAG.getNOT(N0.getDebugLoc(), N0, XType);
7809 return DAG.getNode(ISD::SRL, DL, XType,
7810 DAG.getNode(ISD::AND, DL, XType, NegN0, NotN0),
7811 DAG.getConstant(XType.getSizeInBits()-1,
7812 getShiftAmountTy(XType)));
7814 // fold (setgt X, -1) -> (xor (srl (X, size(X)-1), 1))
7815 if (N1C && N1C->isAllOnesValue() && CC == ISD::SETGT) {
7816 SDValue Sign = DAG.getNode(ISD::SRL, N0.getDebugLoc(), XType, N0,
7817 DAG.getConstant(XType.getSizeInBits()-1,
7818 getShiftAmountTy(N0.getValueType())));
7819 return DAG.getNode(ISD::XOR, DL, XType, Sign, DAG.getConstant(1, XType));
7823 // Check to see if this is an integer abs.
7824 // select_cc setg[te] X, 0, X, -X ->
7825 // select_cc setgt X, -1, X, -X ->
7826 // select_cc setl[te] X, 0, -X, X ->
7827 // select_cc setlt X, 1, -X, X ->
7828 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
7830 ConstantSDNode *SubC = NULL;
7831 if (((N1C->isNullValue() && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
7832 (N1C->isAllOnesValue() && CC == ISD::SETGT)) &&
7833 N0 == N2 && N3.getOpcode() == ISD::SUB && N0 == N3.getOperand(1))
7834 SubC = dyn_cast<ConstantSDNode>(N3.getOperand(0));
7835 else if (((N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE)) ||
7836 (N1C->isOne() && CC == ISD::SETLT)) &&
7837 N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1))
7838 SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0));
7840 EVT XType = N0.getValueType();
7841 if (SubC && SubC->isNullValue() && XType.isInteger()) {
7842 SDValue Shift = DAG.getNode(ISD::SRA, N0.getDebugLoc(), XType,
7844 DAG.getConstant(XType.getSizeInBits()-1,
7845 getShiftAmountTy(N0.getValueType())));
7846 SDValue Add = DAG.getNode(ISD::ADD, N0.getDebugLoc(),
7848 AddToWorkList(Shift.getNode());
7849 AddToWorkList(Add.getNode());
7850 return DAG.getNode(ISD::XOR, DL, XType, Add, Shift);
7857 /// SimplifySetCC - This is a stub for TargetLowering::SimplifySetCC.
7858 SDValue DAGCombiner::SimplifySetCC(EVT VT, SDValue N0,
7859 SDValue N1, ISD::CondCode Cond,
7860 DebugLoc DL, bool foldBooleans) {
7861 TargetLowering::DAGCombinerInfo
7862 DagCombineInfo(DAG, !LegalTypes, !LegalOperations, false, this);
7863 return TLI.SimplifySetCC(VT, N0, N1, Cond, foldBooleans, DagCombineInfo, DL);
7866 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
7867 /// return a DAG expression to select that will generate the same value by
7868 /// multiplying by a magic number. See:
7869 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
7870 SDValue DAGCombiner::BuildSDIV(SDNode *N) {
7871 std::vector<SDNode*> Built;
7872 SDValue S = TLI.BuildSDIV(N, DAG, LegalOperations, &Built);
7874 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end();
7880 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
7881 /// return a DAG expression to select that will generate the same value by
7882 /// multiplying by a magic number. See:
7883 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
7884 SDValue DAGCombiner::BuildUDIV(SDNode *N) {
7885 std::vector<SDNode*> Built;
7886 SDValue S = TLI.BuildUDIV(N, DAG, LegalOperations, &Built);
7888 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end();
7894 /// FindBaseOffset - Return true if base is a frame index, which is known not
7895 // to alias with anything but itself. Provides base object and offset as
7897 static bool FindBaseOffset(SDValue Ptr, SDValue &Base, int64_t &Offset,
7898 const GlobalValue *&GV, void *&CV) {
7899 // Assume it is a primitive operation.
7900 Base = Ptr; Offset = 0; GV = 0; CV = 0;
7902 // If it's an adding a simple constant then integrate the offset.
7903 if (Base.getOpcode() == ISD::ADD) {
7904 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Base.getOperand(1))) {
7905 Base = Base.getOperand(0);
7906 Offset += C->getZExtValue();
7910 // Return the underlying GlobalValue, and update the Offset. Return false
7911 // for GlobalAddressSDNode since the same GlobalAddress may be represented
7912 // by multiple nodes with different offsets.
7913 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Base)) {
7914 GV = G->getGlobal();
7915 Offset += G->getOffset();
7919 // Return the underlying Constant value, and update the Offset. Return false
7920 // for ConstantSDNodes since the same constant pool entry may be represented
7921 // by multiple nodes with different offsets.
7922 if (ConstantPoolSDNode *C = dyn_cast<ConstantPoolSDNode>(Base)) {
7923 CV = C->isMachineConstantPoolEntry() ? (void *)C->getMachineCPVal()
7924 : (void *)C->getConstVal();
7925 Offset += C->getOffset();
7928 // If it's any of the following then it can't alias with anything but itself.
7929 return isa<FrameIndexSDNode>(Base);
7932 /// isAlias - Return true if there is any possibility that the two addresses
7934 bool DAGCombiner::isAlias(SDValue Ptr1, int64_t Size1,
7935 const Value *SrcValue1, int SrcValueOffset1,
7936 unsigned SrcValueAlign1,
7937 const MDNode *TBAAInfo1,
7938 SDValue Ptr2, int64_t Size2,
7939 const Value *SrcValue2, int SrcValueOffset2,
7940 unsigned SrcValueAlign2,
7941 const MDNode *TBAAInfo2) const {
7942 // If they are the same then they must be aliases.
7943 if (Ptr1 == Ptr2) return true;
7945 // Gather base node and offset information.
7946 SDValue Base1, Base2;
7947 int64_t Offset1, Offset2;
7948 const GlobalValue *GV1, *GV2;
7950 bool isFrameIndex1 = FindBaseOffset(Ptr1, Base1, Offset1, GV1, CV1);
7951 bool isFrameIndex2 = FindBaseOffset(Ptr2, Base2, Offset2, GV2, CV2);
7953 // If they have a same base address then check to see if they overlap.
7954 if (Base1 == Base2 || (GV1 && (GV1 == GV2)) || (CV1 && (CV1 == CV2)))
7955 return !((Offset1 + Size1) <= Offset2 || (Offset2 + Size2) <= Offset1);
7957 // It is possible for different frame indices to alias each other, mostly
7958 // when tail call optimization reuses return address slots for arguments.
7959 // To catch this case, look up the actual index of frame indices to compute
7960 // the real alias relationship.
7961 if (isFrameIndex1 && isFrameIndex2) {
7962 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7963 Offset1 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base1)->getIndex());
7964 Offset2 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base2)->getIndex());
7965 return !((Offset1 + Size1) <= Offset2 || (Offset2 + Size2) <= Offset1);
7968 // Otherwise, if we know what the bases are, and they aren't identical, then
7969 // we know they cannot alias.
7970 if ((isFrameIndex1 || CV1 || GV1) && (isFrameIndex2 || CV2 || GV2))
7973 // If we know required SrcValue1 and SrcValue2 have relatively large alignment
7974 // compared to the size and offset of the access, we may be able to prove they
7975 // do not alias. This check is conservative for now to catch cases created by
7976 // splitting vector types.
7977 if ((SrcValueAlign1 == SrcValueAlign2) &&
7978 (SrcValueOffset1 != SrcValueOffset2) &&
7979 (Size1 == Size2) && (SrcValueAlign1 > Size1)) {
7980 int64_t OffAlign1 = SrcValueOffset1 % SrcValueAlign1;
7981 int64_t OffAlign2 = SrcValueOffset2 % SrcValueAlign1;
7983 // There is no overlap between these relatively aligned accesses of similar
7984 // size, return no alias.
7985 if ((OffAlign1 + Size1) <= OffAlign2 || (OffAlign2 + Size2) <= OffAlign1)
7989 if (CombinerGlobalAA) {
7990 // Use alias analysis information.
7991 int64_t MinOffset = std::min(SrcValueOffset1, SrcValueOffset2);
7992 int64_t Overlap1 = Size1 + SrcValueOffset1 - MinOffset;
7993 int64_t Overlap2 = Size2 + SrcValueOffset2 - MinOffset;
7994 AliasAnalysis::AliasResult AAResult =
7995 AA.alias(AliasAnalysis::Location(SrcValue1, Overlap1, TBAAInfo1),
7996 AliasAnalysis::Location(SrcValue2, Overlap2, TBAAInfo2));
7997 if (AAResult == AliasAnalysis::NoAlias)
8001 // Otherwise we have to assume they alias.
8005 /// FindAliasInfo - Extracts the relevant alias information from the memory
8006 /// node. Returns true if the operand was a load.
8007 bool DAGCombiner::FindAliasInfo(SDNode *N,
8008 SDValue &Ptr, int64_t &Size,
8009 const Value *&SrcValue,
8010 int &SrcValueOffset,
8011 unsigned &SrcValueAlign,
8012 const MDNode *&TBAAInfo) const {
8013 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8014 Ptr = LD->getBasePtr();
8015 Size = LD->getMemoryVT().getSizeInBits() >> 3;
8016 SrcValue = LD->getSrcValue();
8017 SrcValueOffset = LD->getSrcValueOffset();
8018 SrcValueAlign = LD->getOriginalAlignment();
8019 TBAAInfo = LD->getTBAAInfo();
8022 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8023 Ptr = ST->getBasePtr();
8024 Size = ST->getMemoryVT().getSizeInBits() >> 3;
8025 SrcValue = ST->getSrcValue();
8026 SrcValueOffset = ST->getSrcValueOffset();
8027 SrcValueAlign = ST->getOriginalAlignment();
8028 TBAAInfo = ST->getTBAAInfo();
8031 llvm_unreachable("FindAliasInfo expected a memory operand");
8034 /// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes,
8035 /// looking for aliasing nodes and adding them to the Aliases vector.
8036 void DAGCombiner::GatherAllAliases(SDNode *N, SDValue OriginalChain,
8037 SmallVector<SDValue, 8> &Aliases) {
8038 SmallVector<SDValue, 8> Chains; // List of chains to visit.
8039 SmallPtrSet<SDNode *, 16> Visited; // Visited node set.
8041 // Get alias information for node.
8044 const Value *SrcValue;
8046 unsigned SrcValueAlign;
8047 const MDNode *SrcTBAAInfo;
8048 bool IsLoad = FindAliasInfo(N, Ptr, Size, SrcValue, SrcValueOffset,
8049 SrcValueAlign, SrcTBAAInfo);
8052 Chains.push_back(OriginalChain);
8055 // Look at each chain and determine if it is an alias. If so, add it to the
8056 // aliases list. If not, then continue up the chain looking for the next
8058 while (!Chains.empty()) {
8059 SDValue Chain = Chains.back();
8062 // For TokenFactor nodes, look at each operand and only continue up the
8063 // chain until we find two aliases. If we've seen two aliases, assume we'll
8064 // find more and revert to original chain since the xform is unlikely to be
8067 // FIXME: The depth check could be made to return the last non-aliasing
8068 // chain we found before we hit a tokenfactor rather than the original
8070 if (Depth > 6 || Aliases.size() == 2) {
8072 Aliases.push_back(OriginalChain);
8076 // Don't bother if we've been before.
8077 if (!Visited.insert(Chain.getNode()))
8080 switch (Chain.getOpcode()) {
8081 case ISD::EntryToken:
8082 // Entry token is ideal chain operand, but handled in FindBetterChain.
8087 // Get alias information for Chain.
8090 const Value *OpSrcValue;
8091 int OpSrcValueOffset;
8092 unsigned OpSrcValueAlign;
8093 const MDNode *OpSrcTBAAInfo;
8094 bool IsOpLoad = FindAliasInfo(Chain.getNode(), OpPtr, OpSize,
8095 OpSrcValue, OpSrcValueOffset,
8099 // If chain is alias then stop here.
8100 if (!(IsLoad && IsOpLoad) &&
8101 isAlias(Ptr, Size, SrcValue, SrcValueOffset, SrcValueAlign,
8103 OpPtr, OpSize, OpSrcValue, OpSrcValueOffset,
8104 OpSrcValueAlign, OpSrcTBAAInfo)) {
8105 Aliases.push_back(Chain);
8107 // Look further up the chain.
8108 Chains.push_back(Chain.getOperand(0));
8114 case ISD::TokenFactor:
8115 // We have to check each of the operands of the token factor for "small"
8116 // token factors, so we queue them up. Adding the operands to the queue
8117 // (stack) in reverse order maintains the original order and increases the
8118 // likelihood that getNode will find a matching token factor (CSE.)
8119 if (Chain.getNumOperands() > 16) {
8120 Aliases.push_back(Chain);
8123 for (unsigned n = Chain.getNumOperands(); n;)
8124 Chains.push_back(Chain.getOperand(--n));
8129 // For all other instructions we will just have to take what we can get.
8130 Aliases.push_back(Chain);
8136 /// FindBetterChain - Walk up chain skipping non-aliasing memory nodes, looking
8137 /// for a better chain (aliasing node.)
8138 SDValue DAGCombiner::FindBetterChain(SDNode *N, SDValue OldChain) {
8139 SmallVector<SDValue, 8> Aliases; // Ops for replacing token factor.
8141 // Accumulate all the aliases to this node.
8142 GatherAllAliases(N, OldChain, Aliases);
8144 // If no operands then chain to entry token.
8145 if (Aliases.size() == 0)
8146 return DAG.getEntryNode();
8148 // If a single operand then chain to it. We don't need to revisit it.
8149 if (Aliases.size() == 1)
8152 // Construct a custom tailored token factor.
8153 return DAG.getNode(ISD::TokenFactor, N->getDebugLoc(), MVT::Other,
8154 &Aliases[0], Aliases.size());
8157 // SelectionDAG::Combine - This is the entry point for the file.
8159 void SelectionDAG::Combine(CombineLevel Level, AliasAnalysis &AA,
8160 CodeGenOpt::Level OptLevel) {
8161 /// run - This is the main entry point to this class.
8163 DAGCombiner(*this, AA, OptLevel).Run(Level);