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/CodeGen/PseudoSourceValue.h"
26 #include "llvm/Analysis/AliasAnalysis.h"
27 #include "llvm/Target/TargetData.h"
28 #include "llvm/Target/TargetFrameInfo.h"
29 #include "llvm/Target/TargetLowering.h"
30 #include "llvm/Target/TargetMachine.h"
31 #include "llvm/Target/TargetOptions.h"
32 #include "llvm/ADT/SmallPtrSet.h"
33 #include "llvm/ADT/Statistic.h"
34 #include "llvm/Support/CommandLine.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/ErrorHandling.h"
37 #include "llvm/Support/MathExtras.h"
38 #include "llvm/Support/raw_ostream.h"
42 STATISTIC(NodesCombined , "Number of dag nodes combined");
43 STATISTIC(PreIndexedNodes , "Number of pre-indexed nodes created");
44 STATISTIC(PostIndexedNodes, "Number of post-indexed nodes created");
45 STATISTIC(OpsNarrowed , "Number of load/op/store narrowed");
49 CombinerAA("combiner-alias-analysis", cl::Hidden,
50 cl::desc("Turn on alias analysis during testing"));
53 CombinerGlobalAA("combiner-global-alias-analysis", cl::Hidden,
54 cl::desc("Include global information in alias analysis"));
56 //------------------------------ DAGCombiner ---------------------------------//
60 const TargetLowering &TLI;
62 CodeGenOpt::Level OptLevel;
66 // Worklist of all of the nodes that need to be simplified.
67 std::vector<SDNode*> WorkList;
69 // AA - Used for DAG load/store alias analysis.
72 /// AddUsersToWorkList - When an instruction is simplified, add all users of
73 /// the instruction to the work lists because they might get more simplified
76 void AddUsersToWorkList(SDNode *N) {
77 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
82 /// visit - call the node-specific routine that knows how to fold each
83 /// particular type of node.
84 SDValue visit(SDNode *N);
87 /// AddToWorkList - Add to the work list making sure it's instance is at the
88 /// the back (next to be processed.)
89 void AddToWorkList(SDNode *N) {
90 removeFromWorkList(N);
91 WorkList.push_back(N);
94 /// removeFromWorkList - remove all instances of N from the worklist.
96 void removeFromWorkList(SDNode *N) {
97 WorkList.erase(std::remove(WorkList.begin(), WorkList.end(), N),
101 SDValue CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
104 SDValue CombineTo(SDNode *N, SDValue Res, bool AddTo = true) {
105 return CombineTo(N, &Res, 1, AddTo);
108 SDValue CombineTo(SDNode *N, SDValue Res0, SDValue Res1,
110 SDValue To[] = { Res0, Res1 };
111 return CombineTo(N, To, 2, AddTo);
114 void CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO);
118 /// SimplifyDemandedBits - Check the specified integer node value to see if
119 /// it can be simplified or if things it uses can be simplified by bit
120 /// propagation. If so, return true.
121 bool SimplifyDemandedBits(SDValue Op) {
122 unsigned BitWidth = Op.getValueType().getScalarType().getSizeInBits();
123 APInt Demanded = APInt::getAllOnesValue(BitWidth);
124 return SimplifyDemandedBits(Op, Demanded);
127 bool SimplifyDemandedBits(SDValue Op, const APInt &Demanded);
129 bool CombineToPreIndexedLoadStore(SDNode *N);
130 bool CombineToPostIndexedLoadStore(SDNode *N);
132 void ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad);
133 SDValue PromoteOperand(SDValue Op, EVT PVT, bool &Replace);
134 SDValue SExtPromoteOperand(SDValue Op, EVT PVT);
135 SDValue ZExtPromoteOperand(SDValue Op, EVT PVT);
136 SDValue PromoteIntBinOp(SDValue Op);
137 SDValue PromoteIntShiftOp(SDValue Op);
138 SDValue PromoteExtend(SDValue Op);
139 bool PromoteLoad(SDValue Op);
141 /// combine - call the node-specific routine that knows how to fold each
142 /// particular type of node. If that doesn't do anything, try the
143 /// target-specific DAG combines.
144 SDValue combine(SDNode *N);
146 // Visitation implementation - Implement dag node combining for different
147 // node types. The semantics are as follows:
149 // SDValue.getNode() == 0 - No change was made
150 // SDValue.getNode() == N - N was replaced, is dead and has been handled.
151 // otherwise - N should be replaced by the returned Operand.
153 SDValue visitTokenFactor(SDNode *N);
154 SDValue visitMERGE_VALUES(SDNode *N);
155 SDValue visitADD(SDNode *N);
156 SDValue visitSUB(SDNode *N);
157 SDValue visitADDC(SDNode *N);
158 SDValue visitADDE(SDNode *N);
159 SDValue visitMUL(SDNode *N);
160 SDValue visitSDIV(SDNode *N);
161 SDValue visitUDIV(SDNode *N);
162 SDValue visitSREM(SDNode *N);
163 SDValue visitUREM(SDNode *N);
164 SDValue visitMULHU(SDNode *N);
165 SDValue visitMULHS(SDNode *N);
166 SDValue visitSMUL_LOHI(SDNode *N);
167 SDValue visitUMUL_LOHI(SDNode *N);
168 SDValue visitSDIVREM(SDNode *N);
169 SDValue visitUDIVREM(SDNode *N);
170 SDValue visitAND(SDNode *N);
171 SDValue visitOR(SDNode *N);
172 SDValue visitXOR(SDNode *N);
173 SDValue SimplifyVBinOp(SDNode *N);
174 SDValue visitSHL(SDNode *N);
175 SDValue visitSRA(SDNode *N);
176 SDValue visitSRL(SDNode *N);
177 SDValue visitCTLZ(SDNode *N);
178 SDValue visitCTTZ(SDNode *N);
179 SDValue visitCTPOP(SDNode *N);
180 SDValue visitSELECT(SDNode *N);
181 SDValue visitSELECT_CC(SDNode *N);
182 SDValue visitSETCC(SDNode *N);
183 SDValue visitSIGN_EXTEND(SDNode *N);
184 SDValue visitZERO_EXTEND(SDNode *N);
185 SDValue visitANY_EXTEND(SDNode *N);
186 SDValue visitSIGN_EXTEND_INREG(SDNode *N);
187 SDValue visitTRUNCATE(SDNode *N);
188 SDValue visitBITCAST(SDNode *N);
189 SDValue visitBUILD_PAIR(SDNode *N);
190 SDValue visitFADD(SDNode *N);
191 SDValue visitFSUB(SDNode *N);
192 SDValue visitFMUL(SDNode *N);
193 SDValue visitFDIV(SDNode *N);
194 SDValue visitFREM(SDNode *N);
195 SDValue visitFCOPYSIGN(SDNode *N);
196 SDValue visitSINT_TO_FP(SDNode *N);
197 SDValue visitUINT_TO_FP(SDNode *N);
198 SDValue visitFP_TO_SINT(SDNode *N);
199 SDValue visitFP_TO_UINT(SDNode *N);
200 SDValue visitFP_ROUND(SDNode *N);
201 SDValue visitFP_ROUND_INREG(SDNode *N);
202 SDValue visitFP_EXTEND(SDNode *N);
203 SDValue visitFNEG(SDNode *N);
204 SDValue visitFABS(SDNode *N);
205 SDValue visitBRCOND(SDNode *N);
206 SDValue visitBR_CC(SDNode *N);
207 SDValue visitLOAD(SDNode *N);
208 SDValue visitSTORE(SDNode *N);
209 SDValue visitINSERT_VECTOR_ELT(SDNode *N);
210 SDValue visitEXTRACT_VECTOR_ELT(SDNode *N);
211 SDValue visitBUILD_VECTOR(SDNode *N);
212 SDValue visitCONCAT_VECTORS(SDNode *N);
213 SDValue visitVECTOR_SHUFFLE(SDNode *N);
214 SDValue visitMEMBARRIER(SDNode *N);
216 SDValue XformToShuffleWithZero(SDNode *N);
217 SDValue ReassociateOps(unsigned Opc, DebugLoc DL, SDValue LHS, SDValue RHS);
219 SDValue visitShiftByConstant(SDNode *N, unsigned Amt);
221 bool SimplifySelectOps(SDNode *SELECT, SDValue LHS, SDValue RHS);
222 SDValue SimplifyBinOpWithSameOpcodeHands(SDNode *N);
223 SDValue SimplifySelect(DebugLoc DL, SDValue N0, SDValue N1, SDValue N2);
224 SDValue SimplifySelectCC(DebugLoc DL, SDValue N0, SDValue N1, SDValue N2,
225 SDValue N3, ISD::CondCode CC,
226 bool NotExtCompare = false);
227 SDValue SimplifySetCC(EVT VT, SDValue N0, SDValue N1, ISD::CondCode Cond,
228 DebugLoc DL, bool foldBooleans = true);
229 SDValue SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
231 SDValue CombineConsecutiveLoads(SDNode *N, EVT VT);
232 SDValue ConstantFoldBITCASTofBUILD_VECTOR(SDNode *, EVT);
233 SDValue BuildSDIV(SDNode *N);
234 SDValue BuildUDIV(SDNode *N);
235 SDNode *MatchRotate(SDValue LHS, SDValue RHS, DebugLoc DL);
236 SDValue ReduceLoadWidth(SDNode *N);
237 SDValue ReduceLoadOpStoreWidth(SDNode *N);
239 SDValue GetDemandedBits(SDValue V, const APInt &Mask);
241 /// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes,
242 /// looking for aliasing nodes and adding them to the Aliases vector.
243 void GatherAllAliases(SDNode *N, SDValue OriginalChain,
244 SmallVector<SDValue, 8> &Aliases);
246 /// isAlias - Return true if there is any possibility that the two addresses
248 bool isAlias(SDValue Ptr1, int64_t Size1,
249 const Value *SrcValue1, int SrcValueOffset1,
250 unsigned SrcValueAlign1,
251 const MDNode *TBAAInfo1,
252 SDValue Ptr2, int64_t Size2,
253 const Value *SrcValue2, int SrcValueOffset2,
254 unsigned SrcValueAlign2,
255 const MDNode *TBAAInfo2) const;
257 /// FindAliasInfo - Extracts the relevant alias information from the memory
258 /// node. Returns true if the operand was a load.
259 bool FindAliasInfo(SDNode *N,
260 SDValue &Ptr, int64_t &Size,
261 const Value *&SrcValue, int &SrcValueOffset,
262 unsigned &SrcValueAlignment,
263 const MDNode *&TBAAInfo) const;
265 /// FindBetterChain - Walk up chain skipping non-aliasing memory nodes,
266 /// looking for a better chain (aliasing node.)
267 SDValue FindBetterChain(SDNode *N, SDValue Chain);
270 DAGCombiner(SelectionDAG &D, AliasAnalysis &A, CodeGenOpt::Level OL)
271 : DAG(D), TLI(D.getTargetLoweringInfo()), Level(Unrestricted),
272 OptLevel(OL), LegalOperations(false), LegalTypes(false), AA(A) {}
274 /// Run - runs the dag combiner on all nodes in the work list
275 void Run(CombineLevel AtLevel);
277 SelectionDAG &getDAG() const { return DAG; }
279 /// getShiftAmountTy - Returns a type large enough to hold any valid
280 /// shift amount - before type legalization these can be huge.
281 EVT getShiftAmountTy() {
282 return LegalTypes ? TLI.getShiftAmountTy() : TLI.getPointerTy();
285 /// isTypeLegal - This method returns true if we are running before type
286 /// legalization or if the specified VT is legal.
287 bool isTypeLegal(const EVT &VT) {
288 if (!LegalTypes) return true;
289 return TLI.isTypeLegal(VT);
296 /// WorkListRemover - This class is a DAGUpdateListener that removes any deleted
297 /// nodes from the worklist.
298 class WorkListRemover : public SelectionDAG::DAGUpdateListener {
301 explicit WorkListRemover(DAGCombiner &dc) : DC(dc) {}
303 virtual void NodeDeleted(SDNode *N, SDNode *E) {
304 DC.removeFromWorkList(N);
307 virtual void NodeUpdated(SDNode *N) {
313 //===----------------------------------------------------------------------===//
314 // TargetLowering::DAGCombinerInfo implementation
315 //===----------------------------------------------------------------------===//
317 void TargetLowering::DAGCombinerInfo::AddToWorklist(SDNode *N) {
318 ((DAGCombiner*)DC)->AddToWorkList(N);
321 SDValue TargetLowering::DAGCombinerInfo::
322 CombineTo(SDNode *N, const std::vector<SDValue> &To, bool AddTo) {
323 return ((DAGCombiner*)DC)->CombineTo(N, &To[0], To.size(), AddTo);
326 SDValue TargetLowering::DAGCombinerInfo::
327 CombineTo(SDNode *N, SDValue Res, bool AddTo) {
328 return ((DAGCombiner*)DC)->CombineTo(N, Res, AddTo);
332 SDValue TargetLowering::DAGCombinerInfo::
333 CombineTo(SDNode *N, SDValue Res0, SDValue Res1, bool AddTo) {
334 return ((DAGCombiner*)DC)->CombineTo(N, Res0, Res1, AddTo);
337 void TargetLowering::DAGCombinerInfo::
338 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
339 return ((DAGCombiner*)DC)->CommitTargetLoweringOpt(TLO);
342 //===----------------------------------------------------------------------===//
344 //===----------------------------------------------------------------------===//
346 /// isNegatibleForFree - Return 1 if we can compute the negated form of the
347 /// specified expression for the same cost as the expression itself, or 2 if we
348 /// can compute the negated form more cheaply than the expression itself.
349 static char isNegatibleForFree(SDValue Op, bool LegalOperations,
350 unsigned Depth = 0) {
351 // No compile time optimizations on this type.
352 if (Op.getValueType() == MVT::ppcf128)
355 // fneg is removable even if it has multiple uses.
356 if (Op.getOpcode() == ISD::FNEG) return 2;
358 // Don't allow anything with multiple uses.
359 if (!Op.hasOneUse()) return 0;
361 // Don't recurse exponentially.
362 if (Depth > 6) return 0;
364 switch (Op.getOpcode()) {
365 default: return false;
366 case ISD::ConstantFP:
367 // Don't invert constant FP values after legalize. The negated constant
368 // isn't necessarily legal.
369 return LegalOperations ? 0 : 1;
371 // FIXME: determine better conditions for this xform.
372 if (!UnsafeFPMath) return 0;
374 // fold (fsub (fadd A, B)) -> (fsub (fneg A), B)
375 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, Depth+1))
377 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
378 return isNegatibleForFree(Op.getOperand(1), LegalOperations, Depth+1);
380 // We can't turn -(A-B) into B-A when we honor signed zeros.
381 if (!UnsafeFPMath) return 0;
383 // fold (fneg (fsub A, B)) -> (fsub B, A)
388 if (HonorSignDependentRoundingFPMath()) return 0;
390 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y) or (fmul X, (fneg Y))
391 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, Depth+1))
394 return isNegatibleForFree(Op.getOperand(1), LegalOperations, Depth+1);
399 return isNegatibleForFree(Op.getOperand(0), LegalOperations, Depth+1);
403 /// GetNegatedExpression - If isNegatibleForFree returns true, this function
404 /// returns the newly negated expression.
405 static SDValue GetNegatedExpression(SDValue Op, SelectionDAG &DAG,
406 bool LegalOperations, unsigned Depth = 0) {
407 // fneg is removable even if it has multiple uses.
408 if (Op.getOpcode() == ISD::FNEG) return Op.getOperand(0);
410 // Don't allow anything with multiple uses.
411 assert(Op.hasOneUse() && "Unknown reuse!");
413 assert(Depth <= 6 && "GetNegatedExpression doesn't match isNegatibleForFree");
414 switch (Op.getOpcode()) {
415 default: llvm_unreachable("Unknown code");
416 case ISD::ConstantFP: {
417 APFloat V = cast<ConstantFPSDNode>(Op)->getValueAPF();
419 return DAG.getConstantFP(V, Op.getValueType());
422 // FIXME: determine better conditions for this xform.
423 assert(UnsafeFPMath);
425 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B)
426 if (isNegatibleForFree(Op.getOperand(0), LegalOperations, Depth+1))
427 return DAG.getNode(ISD::FSUB, Op.getDebugLoc(), Op.getValueType(),
428 GetNegatedExpression(Op.getOperand(0), DAG,
429 LegalOperations, Depth+1),
431 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
432 return DAG.getNode(ISD::FSUB, Op.getDebugLoc(), Op.getValueType(),
433 GetNegatedExpression(Op.getOperand(1), DAG,
434 LegalOperations, Depth+1),
437 // We can't turn -(A-B) into B-A when we honor signed zeros.
438 assert(UnsafeFPMath);
440 // fold (fneg (fsub 0, B)) -> B
441 if (ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(Op.getOperand(0)))
442 if (N0CFP->getValueAPF().isZero())
443 return Op.getOperand(1);
445 // fold (fneg (fsub A, B)) -> (fsub B, A)
446 return DAG.getNode(ISD::FSUB, Op.getDebugLoc(), Op.getValueType(),
447 Op.getOperand(1), Op.getOperand(0));
451 assert(!HonorSignDependentRoundingFPMath());
453 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y)
454 if (isNegatibleForFree(Op.getOperand(0), LegalOperations, Depth+1))
455 return DAG.getNode(Op.getOpcode(), Op.getDebugLoc(), Op.getValueType(),
456 GetNegatedExpression(Op.getOperand(0), DAG,
457 LegalOperations, Depth+1),
460 // fold (fneg (fmul X, Y)) -> (fmul X, (fneg Y))
461 return DAG.getNode(Op.getOpcode(), Op.getDebugLoc(), Op.getValueType(),
463 GetNegatedExpression(Op.getOperand(1), DAG,
464 LegalOperations, Depth+1));
468 return DAG.getNode(Op.getOpcode(), Op.getDebugLoc(), Op.getValueType(),
469 GetNegatedExpression(Op.getOperand(0), DAG,
470 LegalOperations, Depth+1));
472 return DAG.getNode(ISD::FP_ROUND, Op.getDebugLoc(), Op.getValueType(),
473 GetNegatedExpression(Op.getOperand(0), DAG,
474 LegalOperations, Depth+1),
480 // isSetCCEquivalent - Return true if this node is a setcc, or is a select_cc
481 // that selects between the values 1 and 0, making it equivalent to a setcc.
482 // Also, set the incoming LHS, RHS, and CC references to the appropriate
483 // nodes based on the type of node we are checking. This simplifies life a
484 // bit for the callers.
485 static bool isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS,
487 if (N.getOpcode() == ISD::SETCC) {
488 LHS = N.getOperand(0);
489 RHS = N.getOperand(1);
490 CC = N.getOperand(2);
493 if (N.getOpcode() == ISD::SELECT_CC &&
494 N.getOperand(2).getOpcode() == ISD::Constant &&
495 N.getOperand(3).getOpcode() == ISD::Constant &&
496 cast<ConstantSDNode>(N.getOperand(2))->getAPIntValue() == 1 &&
497 cast<ConstantSDNode>(N.getOperand(3))->isNullValue()) {
498 LHS = N.getOperand(0);
499 RHS = N.getOperand(1);
500 CC = N.getOperand(4);
506 // isOneUseSetCC - Return true if this is a SetCC-equivalent operation with only
507 // one use. If this is true, it allows the users to invert the operation for
508 // free when it is profitable to do so.
509 static bool isOneUseSetCC(SDValue N) {
511 if (isSetCCEquivalent(N, N0, N1, N2) && N.getNode()->hasOneUse())
516 SDValue DAGCombiner::ReassociateOps(unsigned Opc, DebugLoc DL,
517 SDValue N0, SDValue N1) {
518 EVT VT = N0.getValueType();
519 if (N0.getOpcode() == Opc && isa<ConstantSDNode>(N0.getOperand(1))) {
520 if (isa<ConstantSDNode>(N1)) {
521 // reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2))
523 DAG.FoldConstantArithmetic(Opc, VT,
524 cast<ConstantSDNode>(N0.getOperand(1)),
525 cast<ConstantSDNode>(N1));
526 return DAG.getNode(Opc, DL, VT, N0.getOperand(0), OpNode);
527 } else if (N0.hasOneUse()) {
528 // reassoc. (op (op x, c1), y) -> (op (op x, y), c1) iff x+c1 has one use
529 SDValue OpNode = DAG.getNode(Opc, N0.getDebugLoc(), VT,
530 N0.getOperand(0), N1);
531 AddToWorkList(OpNode.getNode());
532 return DAG.getNode(Opc, DL, VT, OpNode, N0.getOperand(1));
536 if (N1.getOpcode() == Opc && isa<ConstantSDNode>(N1.getOperand(1))) {
537 if (isa<ConstantSDNode>(N0)) {
538 // reassoc. (op c2, (op x, c1)) -> (op x, (op c1, c2))
540 DAG.FoldConstantArithmetic(Opc, VT,
541 cast<ConstantSDNode>(N1.getOperand(1)),
542 cast<ConstantSDNode>(N0));
543 return DAG.getNode(Opc, DL, VT, N1.getOperand(0), OpNode);
544 } else if (N1.hasOneUse()) {
545 // reassoc. (op y, (op x, c1)) -> (op (op x, y), c1) iff x+c1 has one use
546 SDValue OpNode = DAG.getNode(Opc, N0.getDebugLoc(), VT,
547 N1.getOperand(0), N0);
548 AddToWorkList(OpNode.getNode());
549 return DAG.getNode(Opc, DL, VT, OpNode, N1.getOperand(1));
556 SDValue DAGCombiner::CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
558 assert(N->getNumValues() == NumTo && "Broken CombineTo call!");
560 DEBUG(dbgs() << "\nReplacing.1 ";
562 dbgs() << "\nWith: ";
563 To[0].getNode()->dump(&DAG);
564 dbgs() << " and " << NumTo-1 << " other values\n";
565 for (unsigned i = 0, e = NumTo; i != e; ++i)
566 assert((!To[i].getNode() ||
567 N->getValueType(i) == To[i].getValueType()) &&
568 "Cannot combine value to value of different type!"));
569 WorkListRemover DeadNodes(*this);
570 DAG.ReplaceAllUsesWith(N, To, &DeadNodes);
573 // Push the new nodes and any users onto the worklist
574 for (unsigned i = 0, e = NumTo; i != e; ++i) {
575 if (To[i].getNode()) {
576 AddToWorkList(To[i].getNode());
577 AddUsersToWorkList(To[i].getNode());
582 // Finally, if the node is now dead, remove it from the graph. The node
583 // may not be dead if the replacement process recursively simplified to
584 // something else needing this node.
585 if (N->use_empty()) {
586 // Nodes can be reintroduced into the worklist. Make sure we do not
587 // process a node that has been replaced.
588 removeFromWorkList(N);
590 // Finally, since the node is now dead, remove it from the graph.
593 return SDValue(N, 0);
597 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
598 // Replace all uses. If any nodes become isomorphic to other nodes and
599 // are deleted, make sure to remove them from our worklist.
600 WorkListRemover DeadNodes(*this);
601 DAG.ReplaceAllUsesOfValueWith(TLO.Old, TLO.New, &DeadNodes);
603 // Push the new node and any (possibly new) users onto the worklist.
604 AddToWorkList(TLO.New.getNode());
605 AddUsersToWorkList(TLO.New.getNode());
607 // Finally, if the node is now dead, remove it from the graph. The node
608 // may not be dead if the replacement process recursively simplified to
609 // something else needing this node.
610 if (TLO.Old.getNode()->use_empty()) {
611 removeFromWorkList(TLO.Old.getNode());
613 // If the operands of this node are only used by the node, they will now
614 // be dead. Make sure to visit them first to delete dead nodes early.
615 for (unsigned i = 0, e = TLO.Old.getNode()->getNumOperands(); i != e; ++i)
616 if (TLO.Old.getNode()->getOperand(i).getNode()->hasOneUse())
617 AddToWorkList(TLO.Old.getNode()->getOperand(i).getNode());
619 DAG.DeleteNode(TLO.Old.getNode());
623 /// SimplifyDemandedBits - Check the specified integer node value to see if
624 /// it can be simplified or if things it uses can be simplified by bit
625 /// propagation. If so, return true.
626 bool DAGCombiner::SimplifyDemandedBits(SDValue Op, const APInt &Demanded) {
627 TargetLowering::TargetLoweringOpt TLO(DAG, LegalTypes, LegalOperations);
628 APInt KnownZero, KnownOne;
629 if (!TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
633 AddToWorkList(Op.getNode());
635 // Replace the old value with the new one.
637 DEBUG(dbgs() << "\nReplacing.2 ";
638 TLO.Old.getNode()->dump(&DAG);
639 dbgs() << "\nWith: ";
640 TLO.New.getNode()->dump(&DAG);
643 CommitTargetLoweringOpt(TLO);
647 void DAGCombiner::ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad) {
648 DebugLoc dl = Load->getDebugLoc();
649 EVT VT = Load->getValueType(0);
650 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, VT, SDValue(ExtLoad, 0));
652 DEBUG(dbgs() << "\nReplacing.9 ";
654 dbgs() << "\nWith: ";
655 Trunc.getNode()->dump(&DAG);
657 WorkListRemover DeadNodes(*this);
658 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 0), Trunc, &DeadNodes);
659 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), SDValue(ExtLoad, 1),
661 removeFromWorkList(Load);
662 DAG.DeleteNode(Load);
663 AddToWorkList(Trunc.getNode());
666 SDValue DAGCombiner::PromoteOperand(SDValue Op, EVT PVT, bool &Replace) {
668 DebugLoc dl = Op.getDebugLoc();
669 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Op)) {
670 EVT MemVT = LD->getMemoryVT();
671 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
672 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT) ? ISD::ZEXTLOAD
674 : LD->getExtensionType();
676 return DAG.getExtLoad(ExtType, PVT, dl,
677 LD->getChain(), LD->getBasePtr(),
678 LD->getPointerInfo(),
679 MemVT, LD->isVolatile(),
680 LD->isNonTemporal(), LD->getAlignment());
683 unsigned Opc = Op.getOpcode();
686 case ISD::AssertSext:
687 return DAG.getNode(ISD::AssertSext, dl, PVT,
688 SExtPromoteOperand(Op.getOperand(0), PVT),
690 case ISD::AssertZext:
691 return DAG.getNode(ISD::AssertZext, dl, PVT,
692 ZExtPromoteOperand(Op.getOperand(0), PVT),
694 case ISD::Constant: {
696 Op.getValueType().isByteSized() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
697 return DAG.getNode(ExtOpc, dl, PVT, Op);
701 if (!TLI.isOperationLegal(ISD::ANY_EXTEND, PVT))
703 return DAG.getNode(ISD::ANY_EXTEND, dl, PVT, Op);
706 SDValue DAGCombiner::SExtPromoteOperand(SDValue Op, EVT PVT) {
707 if (!TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, PVT))
709 EVT OldVT = Op.getValueType();
710 DebugLoc dl = Op.getDebugLoc();
711 bool Replace = false;
712 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
713 if (NewOp.getNode() == 0)
715 AddToWorkList(NewOp.getNode());
718 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
719 return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, NewOp.getValueType(), NewOp,
720 DAG.getValueType(OldVT));
723 SDValue DAGCombiner::ZExtPromoteOperand(SDValue Op, EVT PVT) {
724 EVT OldVT = Op.getValueType();
725 DebugLoc dl = Op.getDebugLoc();
726 bool Replace = false;
727 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
728 if (NewOp.getNode() == 0)
730 AddToWorkList(NewOp.getNode());
733 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
734 return DAG.getZeroExtendInReg(NewOp, dl, OldVT);
737 /// PromoteIntBinOp - Promote the specified integer binary operation if the
738 /// target indicates it is beneficial. e.g. On x86, it's usually better to
739 /// promote i16 operations to i32 since i16 instructions are longer.
740 SDValue DAGCombiner::PromoteIntBinOp(SDValue Op) {
741 if (!LegalOperations)
744 EVT VT = Op.getValueType();
745 if (VT.isVector() || !VT.isInteger())
748 // If operation type is 'undesirable', e.g. i16 on x86, consider
750 unsigned Opc = Op.getOpcode();
751 if (TLI.isTypeDesirableForOp(Opc, VT))
755 // Consult target whether it is a good idea to promote this operation and
756 // what's the right type to promote it to.
757 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
758 assert(PVT != VT && "Don't know what type to promote to!");
760 bool Replace0 = false;
761 SDValue N0 = Op.getOperand(0);
762 SDValue NN0 = PromoteOperand(N0, PVT, Replace0);
763 if (NN0.getNode() == 0)
766 bool Replace1 = false;
767 SDValue N1 = Op.getOperand(1);
772 NN1 = PromoteOperand(N1, PVT, Replace1);
773 if (NN1.getNode() == 0)
777 AddToWorkList(NN0.getNode());
779 AddToWorkList(NN1.getNode());
782 ReplaceLoadWithPromotedLoad(N0.getNode(), NN0.getNode());
784 ReplaceLoadWithPromotedLoad(N1.getNode(), NN1.getNode());
786 DEBUG(dbgs() << "\nPromoting ";
787 Op.getNode()->dump(&DAG));
788 DebugLoc dl = Op.getDebugLoc();
789 return DAG.getNode(ISD::TRUNCATE, dl, VT,
790 DAG.getNode(Opc, dl, PVT, NN0, NN1));
795 /// PromoteIntShiftOp - Promote the specified integer shift operation if the
796 /// target indicates it is beneficial. e.g. On x86, it's usually better to
797 /// promote i16 operations to i32 since i16 instructions are longer.
798 SDValue DAGCombiner::PromoteIntShiftOp(SDValue Op) {
799 if (!LegalOperations)
802 EVT VT = Op.getValueType();
803 if (VT.isVector() || !VT.isInteger())
806 // If operation type is 'undesirable', e.g. i16 on x86, consider
808 unsigned Opc = Op.getOpcode();
809 if (TLI.isTypeDesirableForOp(Opc, VT))
813 // Consult target whether it is a good idea to promote this operation and
814 // what's the right type to promote it to.
815 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
816 assert(PVT != VT && "Don't know what type to promote to!");
818 bool Replace = false;
819 SDValue N0 = Op.getOperand(0);
821 N0 = SExtPromoteOperand(Op.getOperand(0), PVT);
822 else if (Opc == ISD::SRL)
823 N0 = ZExtPromoteOperand(Op.getOperand(0), PVT);
825 N0 = PromoteOperand(N0, PVT, Replace);
826 if (N0.getNode() == 0)
829 AddToWorkList(N0.getNode());
831 ReplaceLoadWithPromotedLoad(Op.getOperand(0).getNode(), N0.getNode());
833 DEBUG(dbgs() << "\nPromoting ";
834 Op.getNode()->dump(&DAG));
835 DebugLoc dl = Op.getDebugLoc();
836 return DAG.getNode(ISD::TRUNCATE, dl, VT,
837 DAG.getNode(Opc, dl, PVT, N0, Op.getOperand(1)));
842 SDValue DAGCombiner::PromoteExtend(SDValue Op) {
843 if (!LegalOperations)
846 EVT VT = Op.getValueType();
847 if (VT.isVector() || !VT.isInteger())
850 // If operation type is 'undesirable', e.g. i16 on x86, consider
852 unsigned Opc = Op.getOpcode();
853 if (TLI.isTypeDesirableForOp(Opc, VT))
857 // Consult target whether it is a good idea to promote this operation and
858 // what's the right type to promote it to.
859 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
860 assert(PVT != VT && "Don't know what type to promote to!");
861 // fold (aext (aext x)) -> (aext x)
862 // fold (aext (zext x)) -> (zext x)
863 // fold (aext (sext x)) -> (sext x)
864 DEBUG(dbgs() << "\nPromoting ";
865 Op.getNode()->dump(&DAG));
866 return DAG.getNode(Op.getOpcode(), Op.getDebugLoc(), VT, Op.getOperand(0));
871 bool DAGCombiner::PromoteLoad(SDValue Op) {
872 if (!LegalOperations)
875 EVT VT = Op.getValueType();
876 if (VT.isVector() || !VT.isInteger())
879 // If operation type is 'undesirable', e.g. i16 on x86, consider
881 unsigned Opc = Op.getOpcode();
882 if (TLI.isTypeDesirableForOp(Opc, VT))
886 // Consult target whether it is a good idea to promote this operation and
887 // what's the right type to promote it to.
888 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
889 assert(PVT != VT && "Don't know what type to promote to!");
891 DebugLoc dl = Op.getDebugLoc();
892 SDNode *N = Op.getNode();
893 LoadSDNode *LD = cast<LoadSDNode>(N);
894 EVT MemVT = LD->getMemoryVT();
895 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
896 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT) ? ISD::ZEXTLOAD
898 : LD->getExtensionType();
899 SDValue NewLD = DAG.getExtLoad(ExtType, PVT, dl,
900 LD->getChain(), LD->getBasePtr(),
901 LD->getPointerInfo(),
902 MemVT, LD->isVolatile(),
903 LD->isNonTemporal(), LD->getAlignment());
904 SDValue Result = DAG.getNode(ISD::TRUNCATE, dl, VT, NewLD);
906 DEBUG(dbgs() << "\nPromoting ";
909 Result.getNode()->dump(&DAG);
911 WorkListRemover DeadNodes(*this);
912 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result, &DeadNodes);
913 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), NewLD.getValue(1), &DeadNodes);
914 removeFromWorkList(N);
916 AddToWorkList(Result.getNode());
923 //===----------------------------------------------------------------------===//
924 // Main DAG Combiner implementation
925 //===----------------------------------------------------------------------===//
927 void DAGCombiner::Run(CombineLevel AtLevel) {
928 // set the instance variables, so that the various visit routines may use it.
930 LegalOperations = Level >= NoIllegalOperations;
931 LegalTypes = Level >= NoIllegalTypes;
933 // Add all the dag nodes to the worklist.
934 WorkList.reserve(DAG.allnodes_size());
935 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
936 E = DAG.allnodes_end(); I != E; ++I)
937 WorkList.push_back(I);
939 // Create a dummy node (which is not added to allnodes), that adds a reference
940 // to the root node, preventing it from being deleted, and tracking any
941 // changes of the root.
942 HandleSDNode Dummy(DAG.getRoot());
944 // The root of the dag may dangle to deleted nodes until the dag combiner is
945 // done. Set it to null to avoid confusion.
946 DAG.setRoot(SDValue());
948 // while the worklist isn't empty, inspect the node on the end of it and
949 // try and combine it.
950 while (!WorkList.empty()) {
951 SDNode *N = WorkList.back();
954 // If N has no uses, it is dead. Make sure to revisit all N's operands once
955 // N is deleted from the DAG, since they too may now be dead or may have a
956 // reduced number of uses, allowing other xforms.
957 if (N->use_empty() && N != &Dummy) {
958 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
959 AddToWorkList(N->getOperand(i).getNode());
965 SDValue RV = combine(N);
967 if (RV.getNode() == 0)
972 // If we get back the same node we passed in, rather than a new node or
973 // zero, we know that the node must have defined multiple values and
974 // CombineTo was used. Since CombineTo takes care of the worklist
975 // mechanics for us, we have no work to do in this case.
976 if (RV.getNode() == N)
979 assert(N->getOpcode() != ISD::DELETED_NODE &&
980 RV.getNode()->getOpcode() != ISD::DELETED_NODE &&
981 "Node was deleted but visit returned new node!");
983 DEBUG(dbgs() << "\nReplacing.3 ";
985 dbgs() << "\nWith: ";
986 RV.getNode()->dump(&DAG);
988 WorkListRemover DeadNodes(*this);
989 if (N->getNumValues() == RV.getNode()->getNumValues())
990 DAG.ReplaceAllUsesWith(N, RV.getNode(), &DeadNodes);
992 assert(N->getValueType(0) == RV.getValueType() &&
993 N->getNumValues() == 1 && "Type mismatch");
995 DAG.ReplaceAllUsesWith(N, &OpV, &DeadNodes);
998 // Push the new node and any users onto the worklist
999 AddToWorkList(RV.getNode());
1000 AddUsersToWorkList(RV.getNode());
1002 // Add any uses of the old node to the worklist in case this node is the
1003 // last one that uses them. They may become dead after this node is
1005 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1006 AddToWorkList(N->getOperand(i).getNode());
1008 // Finally, if the node is now dead, remove it from the graph. The node
1009 // may not be dead if the replacement process recursively simplified to
1010 // something else needing this node.
1011 if (N->use_empty()) {
1012 // Nodes can be reintroduced into the worklist. Make sure we do not
1013 // process a node that has been replaced.
1014 removeFromWorkList(N);
1016 // Finally, since the node is now dead, remove it from the graph.
1021 // If the root changed (e.g. it was a dead load, update the root).
1022 DAG.setRoot(Dummy.getValue());
1025 SDValue DAGCombiner::visit(SDNode *N) {
1026 switch (N->getOpcode()) {
1028 case ISD::TokenFactor: return visitTokenFactor(N);
1029 case ISD::MERGE_VALUES: return visitMERGE_VALUES(N);
1030 case ISD::ADD: return visitADD(N);
1031 case ISD::SUB: return visitSUB(N);
1032 case ISD::ADDC: return visitADDC(N);
1033 case ISD::ADDE: return visitADDE(N);
1034 case ISD::MUL: return visitMUL(N);
1035 case ISD::SDIV: return visitSDIV(N);
1036 case ISD::UDIV: return visitUDIV(N);
1037 case ISD::SREM: return visitSREM(N);
1038 case ISD::UREM: return visitUREM(N);
1039 case ISD::MULHU: return visitMULHU(N);
1040 case ISD::MULHS: return visitMULHS(N);
1041 case ISD::SMUL_LOHI: return visitSMUL_LOHI(N);
1042 case ISD::UMUL_LOHI: return visitUMUL_LOHI(N);
1043 case ISD::SDIVREM: return visitSDIVREM(N);
1044 case ISD::UDIVREM: return visitUDIVREM(N);
1045 case ISD::AND: return visitAND(N);
1046 case ISD::OR: return visitOR(N);
1047 case ISD::XOR: return visitXOR(N);
1048 case ISD::SHL: return visitSHL(N);
1049 case ISD::SRA: return visitSRA(N);
1050 case ISD::SRL: return visitSRL(N);
1051 case ISD::CTLZ: return visitCTLZ(N);
1052 case ISD::CTTZ: return visitCTTZ(N);
1053 case ISD::CTPOP: return visitCTPOP(N);
1054 case ISD::SELECT: return visitSELECT(N);
1055 case ISD::SELECT_CC: return visitSELECT_CC(N);
1056 case ISD::SETCC: return visitSETCC(N);
1057 case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N);
1058 case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N);
1059 case ISD::ANY_EXTEND: return visitANY_EXTEND(N);
1060 case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N);
1061 case ISD::TRUNCATE: return visitTRUNCATE(N);
1062 case ISD::BITCAST: return visitBITCAST(N);
1063 case ISD::BUILD_PAIR: return visitBUILD_PAIR(N);
1064 case ISD::FADD: return visitFADD(N);
1065 case ISD::FSUB: return visitFSUB(N);
1066 case ISD::FMUL: return visitFMUL(N);
1067 case ISD::FDIV: return visitFDIV(N);
1068 case ISD::FREM: return visitFREM(N);
1069 case ISD::FCOPYSIGN: return visitFCOPYSIGN(N);
1070 case ISD::SINT_TO_FP: return visitSINT_TO_FP(N);
1071 case ISD::UINT_TO_FP: return visitUINT_TO_FP(N);
1072 case ISD::FP_TO_SINT: return visitFP_TO_SINT(N);
1073 case ISD::FP_TO_UINT: return visitFP_TO_UINT(N);
1074 case ISD::FP_ROUND: return visitFP_ROUND(N);
1075 case ISD::FP_ROUND_INREG: return visitFP_ROUND_INREG(N);
1076 case ISD::FP_EXTEND: return visitFP_EXTEND(N);
1077 case ISD::FNEG: return visitFNEG(N);
1078 case ISD::FABS: return visitFABS(N);
1079 case ISD::BRCOND: return visitBRCOND(N);
1080 case ISD::BR_CC: return visitBR_CC(N);
1081 case ISD::LOAD: return visitLOAD(N);
1082 case ISD::STORE: return visitSTORE(N);
1083 case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N);
1084 case ISD::EXTRACT_VECTOR_ELT: return visitEXTRACT_VECTOR_ELT(N);
1085 case ISD::BUILD_VECTOR: return visitBUILD_VECTOR(N);
1086 case ISD::CONCAT_VECTORS: return visitCONCAT_VECTORS(N);
1087 case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N);
1088 case ISD::MEMBARRIER: return visitMEMBARRIER(N);
1093 SDValue DAGCombiner::combine(SDNode *N) {
1094 SDValue RV = visit(N);
1096 // If nothing happened, try a target-specific DAG combine.
1097 if (RV.getNode() == 0) {
1098 assert(N->getOpcode() != ISD::DELETED_NODE &&
1099 "Node was deleted but visit returned NULL!");
1101 if (N->getOpcode() >= ISD::BUILTIN_OP_END ||
1102 TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode())) {
1104 // Expose the DAG combiner to the target combiner impls.
1105 TargetLowering::DAGCombinerInfo
1106 DagCombineInfo(DAG, !LegalTypes, !LegalOperations, false, this);
1108 RV = TLI.PerformDAGCombine(N, DagCombineInfo);
1112 // If nothing happened still, try promoting the operation.
1113 if (RV.getNode() == 0) {
1114 switch (N->getOpcode()) {
1122 RV = PromoteIntBinOp(SDValue(N, 0));
1127 RV = PromoteIntShiftOp(SDValue(N, 0));
1129 case ISD::SIGN_EXTEND:
1130 case ISD::ZERO_EXTEND:
1131 case ISD::ANY_EXTEND:
1132 RV = PromoteExtend(SDValue(N, 0));
1135 if (PromoteLoad(SDValue(N, 0)))
1141 // If N is a commutative binary node, try commuting it to enable more
1143 if (RV.getNode() == 0 &&
1144 SelectionDAG::isCommutativeBinOp(N->getOpcode()) &&
1145 N->getNumValues() == 1) {
1146 SDValue N0 = N->getOperand(0);
1147 SDValue N1 = N->getOperand(1);
1149 // Constant operands are canonicalized to RHS.
1150 if (isa<ConstantSDNode>(N0) || !isa<ConstantSDNode>(N1)) {
1151 SDValue Ops[] = { N1, N0 };
1152 SDNode *CSENode = DAG.getNodeIfExists(N->getOpcode(), N->getVTList(),
1155 return SDValue(CSENode, 0);
1162 /// getInputChainForNode - Given a node, return its input chain if it has one,
1163 /// otherwise return a null sd operand.
1164 static SDValue getInputChainForNode(SDNode *N) {
1165 if (unsigned NumOps = N->getNumOperands()) {
1166 if (N->getOperand(0).getValueType() == MVT::Other)
1167 return N->getOperand(0);
1168 else if (N->getOperand(NumOps-1).getValueType() == MVT::Other)
1169 return N->getOperand(NumOps-1);
1170 for (unsigned i = 1; i < NumOps-1; ++i)
1171 if (N->getOperand(i).getValueType() == MVT::Other)
1172 return N->getOperand(i);
1177 SDValue DAGCombiner::visitTokenFactor(SDNode *N) {
1178 // If N has two operands, where one has an input chain equal to the other,
1179 // the 'other' chain is redundant.
1180 if (N->getNumOperands() == 2) {
1181 if (getInputChainForNode(N->getOperand(0).getNode()) == N->getOperand(1))
1182 return N->getOperand(0);
1183 if (getInputChainForNode(N->getOperand(1).getNode()) == N->getOperand(0))
1184 return N->getOperand(1);
1187 SmallVector<SDNode *, 8> TFs; // List of token factors to visit.
1188 SmallVector<SDValue, 8> Ops; // Ops for replacing token factor.
1189 SmallPtrSet<SDNode*, 16> SeenOps;
1190 bool Changed = false; // If we should replace this token factor.
1192 // Start out with this token factor.
1195 // Iterate through token factors. The TFs grows when new token factors are
1197 for (unsigned i = 0; i < TFs.size(); ++i) {
1198 SDNode *TF = TFs[i];
1200 // Check each of the operands.
1201 for (unsigned i = 0, ie = TF->getNumOperands(); i != ie; ++i) {
1202 SDValue Op = TF->getOperand(i);
1204 switch (Op.getOpcode()) {
1205 case ISD::EntryToken:
1206 // Entry tokens don't need to be added to the list. They are
1211 case ISD::TokenFactor:
1212 if (Op.hasOneUse() &&
1213 std::find(TFs.begin(), TFs.end(), Op.getNode()) == TFs.end()) {
1214 // Queue up for processing.
1215 TFs.push_back(Op.getNode());
1216 // Clean up in case the token factor is removed.
1217 AddToWorkList(Op.getNode());
1224 // Only add if it isn't already in the list.
1225 if (SeenOps.insert(Op.getNode()))
1236 // If we've change things around then replace token factor.
1239 // The entry token is the only possible outcome.
1240 Result = DAG.getEntryNode();
1242 // New and improved token factor.
1243 Result = DAG.getNode(ISD::TokenFactor, N->getDebugLoc(),
1244 MVT::Other, &Ops[0], Ops.size());
1247 // Don't add users to work list.
1248 return CombineTo(N, Result, false);
1254 /// MERGE_VALUES can always be eliminated.
1255 SDValue DAGCombiner::visitMERGE_VALUES(SDNode *N) {
1256 WorkListRemover DeadNodes(*this);
1257 // Replacing results may cause a different MERGE_VALUES to suddenly
1258 // be CSE'd with N, and carry its uses with it. Iterate until no
1259 // uses remain, to ensure that the node can be safely deleted.
1261 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1262 DAG.ReplaceAllUsesOfValueWith(SDValue(N, i), N->getOperand(i),
1264 } while (!N->use_empty());
1265 removeFromWorkList(N);
1267 return SDValue(N, 0); // Return N so it doesn't get rechecked!
1271 SDValue combineShlAddConstant(DebugLoc DL, SDValue N0, SDValue N1,
1272 SelectionDAG &DAG) {
1273 EVT VT = N0.getValueType();
1274 SDValue N00 = N0.getOperand(0);
1275 SDValue N01 = N0.getOperand(1);
1276 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N01);
1278 if (N01C && N00.getOpcode() == ISD::ADD && N00.getNode()->hasOneUse() &&
1279 isa<ConstantSDNode>(N00.getOperand(1))) {
1280 // fold (add (shl (add x, c1), c2), ) -> (add (add (shl x, c2), c1<<c2), )
1281 N0 = DAG.getNode(ISD::ADD, N0.getDebugLoc(), VT,
1282 DAG.getNode(ISD::SHL, N00.getDebugLoc(), VT,
1283 N00.getOperand(0), N01),
1284 DAG.getNode(ISD::SHL, N01.getDebugLoc(), VT,
1285 N00.getOperand(1), N01));
1286 return DAG.getNode(ISD::ADD, DL, VT, N0, N1);
1292 SDValue DAGCombiner::visitADD(SDNode *N) {
1293 SDValue N0 = N->getOperand(0);
1294 SDValue N1 = N->getOperand(1);
1295 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1296 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1297 EVT VT = N0.getValueType();
1300 if (VT.isVector()) {
1301 SDValue FoldedVOp = SimplifyVBinOp(N);
1302 if (FoldedVOp.getNode()) return FoldedVOp;
1305 // fold (add x, undef) -> undef
1306 if (N0.getOpcode() == ISD::UNDEF)
1308 if (N1.getOpcode() == ISD::UNDEF)
1310 // fold (add c1, c2) -> c1+c2
1312 return DAG.FoldConstantArithmetic(ISD::ADD, VT, N0C, N1C);
1313 // canonicalize constant to RHS
1315 return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N1, N0);
1316 // fold (add x, 0) -> x
1317 if (N1C && N1C->isNullValue())
1319 // fold (add Sym, c) -> Sym+c
1320 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1321 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA) && N1C &&
1322 GA->getOpcode() == ISD::GlobalAddress)
1323 return DAG.getGlobalAddress(GA->getGlobal(), N1C->getDebugLoc(), VT,
1325 (uint64_t)N1C->getSExtValue());
1326 // fold ((c1-A)+c2) -> (c1+c2)-A
1327 if (N1C && N0.getOpcode() == ISD::SUB)
1328 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getOperand(0)))
1329 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1330 DAG.getConstant(N1C->getAPIntValue()+
1331 N0C->getAPIntValue(), VT),
1334 SDValue RADD = ReassociateOps(ISD::ADD, N->getDebugLoc(), N0, N1);
1335 if (RADD.getNode() != 0)
1337 // fold ((0-A) + B) -> B-A
1338 if (N0.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N0.getOperand(0)) &&
1339 cast<ConstantSDNode>(N0.getOperand(0))->isNullValue())
1340 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N1, N0.getOperand(1));
1341 // fold (A + (0-B)) -> A-B
1342 if (N1.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N1.getOperand(0)) &&
1343 cast<ConstantSDNode>(N1.getOperand(0))->isNullValue())
1344 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N0, N1.getOperand(1));
1345 // fold (A+(B-A)) -> B
1346 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1))
1347 return N1.getOperand(0);
1348 // fold ((B-A)+A) -> B
1349 if (N0.getOpcode() == ISD::SUB && N1 == N0.getOperand(1))
1350 return N0.getOperand(0);
1351 // fold (A+(B-(A+C))) to (B-C)
1352 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1353 N0 == N1.getOperand(1).getOperand(0))
1354 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N1.getOperand(0),
1355 N1.getOperand(1).getOperand(1));
1356 // fold (A+(B-(C+A))) to (B-C)
1357 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1358 N0 == N1.getOperand(1).getOperand(1))
1359 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N1.getOperand(0),
1360 N1.getOperand(1).getOperand(0));
1361 // fold (A+((B-A)+or-C)) to (B+or-C)
1362 if ((N1.getOpcode() == ISD::SUB || N1.getOpcode() == ISD::ADD) &&
1363 N1.getOperand(0).getOpcode() == ISD::SUB &&
1364 N0 == N1.getOperand(0).getOperand(1))
1365 return DAG.getNode(N1.getOpcode(), N->getDebugLoc(), VT,
1366 N1.getOperand(0).getOperand(0), N1.getOperand(1));
1368 // fold (A-B)+(C-D) to (A+C)-(B+D) when A or C is constant
1369 if (N0.getOpcode() == ISD::SUB && N1.getOpcode() == ISD::SUB) {
1370 SDValue N00 = N0.getOperand(0);
1371 SDValue N01 = N0.getOperand(1);
1372 SDValue N10 = N1.getOperand(0);
1373 SDValue N11 = N1.getOperand(1);
1375 if (isa<ConstantSDNode>(N00) || isa<ConstantSDNode>(N10))
1376 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1377 DAG.getNode(ISD::ADD, N0.getDebugLoc(), VT, N00, N10),
1378 DAG.getNode(ISD::ADD, N1.getDebugLoc(), VT, N01, N11));
1381 if (!VT.isVector() && SimplifyDemandedBits(SDValue(N, 0)))
1382 return SDValue(N, 0);
1384 // fold (a+b) -> (a|b) iff a and b share no bits.
1385 if (VT.isInteger() && !VT.isVector()) {
1386 APInt LHSZero, LHSOne;
1387 APInt RHSZero, RHSOne;
1388 APInt Mask = APInt::getAllOnesValue(VT.getScalarType().getSizeInBits());
1389 DAG.ComputeMaskedBits(N0, Mask, LHSZero, LHSOne);
1391 if (LHSZero.getBoolValue()) {
1392 DAG.ComputeMaskedBits(N1, Mask, RHSZero, RHSOne);
1394 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1395 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1396 if ((RHSZero & (~LHSZero & Mask)) == (~LHSZero & Mask) ||
1397 (LHSZero & (~RHSZero & Mask)) == (~RHSZero & Mask))
1398 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N0, N1);
1402 // fold (add (shl (add x, c1), c2), ) -> (add (add (shl x, c2), c1<<c2), )
1403 if (N0.getOpcode() == ISD::SHL && N0.getNode()->hasOneUse()) {
1404 SDValue Result = combineShlAddConstant(N->getDebugLoc(), N0, N1, DAG);
1405 if (Result.getNode()) return Result;
1407 if (N1.getOpcode() == ISD::SHL && N1.getNode()->hasOneUse()) {
1408 SDValue Result = combineShlAddConstant(N->getDebugLoc(), N1, N0, DAG);
1409 if (Result.getNode()) return Result;
1412 // fold (add x, shl(0 - y, n)) -> sub(x, shl(y, n))
1413 if (N1.getOpcode() == ISD::SHL &&
1414 N1.getOperand(0).getOpcode() == ISD::SUB)
1415 if (ConstantSDNode *C =
1416 dyn_cast<ConstantSDNode>(N1.getOperand(0).getOperand(0)))
1417 if (C->getAPIntValue() == 0)
1418 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N0,
1419 DAG.getNode(ISD::SHL, N->getDebugLoc(), VT,
1420 N1.getOperand(0).getOperand(1),
1422 if (N0.getOpcode() == ISD::SHL &&
1423 N0.getOperand(0).getOpcode() == ISD::SUB)
1424 if (ConstantSDNode *C =
1425 dyn_cast<ConstantSDNode>(N0.getOperand(0).getOperand(0)))
1426 if (C->getAPIntValue() == 0)
1427 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N1,
1428 DAG.getNode(ISD::SHL, N->getDebugLoc(), VT,
1429 N0.getOperand(0).getOperand(1),
1432 if (N1.getOpcode() == ISD::AND) {
1433 SDValue AndOp0 = N1.getOperand(0);
1434 ConstantSDNode *AndOp1 = dyn_cast<ConstantSDNode>(N1->getOperand(1));
1435 unsigned NumSignBits = DAG.ComputeNumSignBits(AndOp0);
1436 unsigned DestBits = VT.getScalarType().getSizeInBits();
1438 // (add z, (and (sbbl x, x), 1)) -> (sub z, (sbbl x, x))
1439 // and similar xforms where the inner op is either ~0 or 0.
1440 if (NumSignBits == DestBits && AndOp1 && AndOp1->isOne()) {
1441 DebugLoc DL = N->getDebugLoc();
1442 return DAG.getNode(ISD::SUB, DL, VT, N->getOperand(0), AndOp0);
1446 // add (sext i1), X -> sub X, (zext i1)
1447 if (N0.getOpcode() == ISD::SIGN_EXTEND &&
1448 N0.getOperand(0).getValueType() == MVT::i1 &&
1449 !TLI.isOperationLegal(ISD::SIGN_EXTEND, MVT::i1)) {
1450 DebugLoc DL = N->getDebugLoc();
1451 SDValue ZExt = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0));
1452 return DAG.getNode(ISD::SUB, DL, VT, N1, ZExt);
1458 SDValue DAGCombiner::visitADDC(SDNode *N) {
1459 SDValue N0 = N->getOperand(0);
1460 SDValue N1 = N->getOperand(1);
1461 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1462 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1463 EVT VT = N0.getValueType();
1465 // If the flag result is dead, turn this into an ADD.
1466 if (N->hasNUsesOfValue(0, 1))
1467 return CombineTo(N, DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N1, N0),
1468 DAG.getNode(ISD::CARRY_FALSE,
1469 N->getDebugLoc(), MVT::Glue));
1471 // canonicalize constant to RHS.
1473 return DAG.getNode(ISD::ADDC, N->getDebugLoc(), N->getVTList(), N1, N0);
1475 // fold (addc x, 0) -> x + no carry out
1476 if (N1C && N1C->isNullValue())
1477 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE,
1478 N->getDebugLoc(), MVT::Glue));
1480 // fold (addc a, b) -> (or a, b), CARRY_FALSE iff a and b share no bits.
1481 APInt LHSZero, LHSOne;
1482 APInt RHSZero, RHSOne;
1483 APInt Mask = APInt::getAllOnesValue(VT.getScalarType().getSizeInBits());
1484 DAG.ComputeMaskedBits(N0, Mask, LHSZero, LHSOne);
1486 if (LHSZero.getBoolValue()) {
1487 DAG.ComputeMaskedBits(N1, Mask, RHSZero, RHSOne);
1489 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1490 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1491 if ((RHSZero & (~LHSZero & Mask)) == (~LHSZero & Mask) ||
1492 (LHSZero & (~RHSZero & Mask)) == (~RHSZero & Mask))
1493 return CombineTo(N, DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N0, N1),
1494 DAG.getNode(ISD::CARRY_FALSE,
1495 N->getDebugLoc(), MVT::Glue));
1501 SDValue DAGCombiner::visitADDE(SDNode *N) {
1502 SDValue N0 = N->getOperand(0);
1503 SDValue N1 = N->getOperand(1);
1504 SDValue CarryIn = N->getOperand(2);
1505 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1506 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1508 // canonicalize constant to RHS
1510 return DAG.getNode(ISD::ADDE, N->getDebugLoc(), N->getVTList(),
1513 // fold (adde x, y, false) -> (addc x, y)
1514 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1515 return DAG.getNode(ISD::ADDC, N->getDebugLoc(), N->getVTList(), N1, N0);
1520 SDValue DAGCombiner::visitSUB(SDNode *N) {
1521 SDValue N0 = N->getOperand(0);
1522 SDValue N1 = N->getOperand(1);
1523 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1524 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1525 EVT VT = N0.getValueType();
1528 if (VT.isVector()) {
1529 SDValue FoldedVOp = SimplifyVBinOp(N);
1530 if (FoldedVOp.getNode()) return FoldedVOp;
1533 // fold (sub x, x) -> 0
1535 return DAG.getConstant(0, N->getValueType(0));
1536 // fold (sub c1, c2) -> c1-c2
1538 return DAG.FoldConstantArithmetic(ISD::SUB, VT, N0C, N1C);
1539 // fold (sub x, c) -> (add x, -c)
1541 return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N0,
1542 DAG.getConstant(-N1C->getAPIntValue(), VT));
1543 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1)
1544 if (N0C && N0C->isAllOnesValue())
1545 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N1, N0);
1546 // fold (A+B)-A -> B
1547 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1)
1548 return N0.getOperand(1);
1549 // fold (A+B)-B -> A
1550 if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1)
1551 return N0.getOperand(0);
1552 // fold ((A+(B+or-C))-B) -> A+or-C
1553 if (N0.getOpcode() == ISD::ADD &&
1554 (N0.getOperand(1).getOpcode() == ISD::SUB ||
1555 N0.getOperand(1).getOpcode() == ISD::ADD) &&
1556 N0.getOperand(1).getOperand(0) == N1)
1557 return DAG.getNode(N0.getOperand(1).getOpcode(), N->getDebugLoc(), VT,
1558 N0.getOperand(0), N0.getOperand(1).getOperand(1));
1559 // fold ((A+(C+B))-B) -> A+C
1560 if (N0.getOpcode() == ISD::ADD &&
1561 N0.getOperand(1).getOpcode() == ISD::ADD &&
1562 N0.getOperand(1).getOperand(1) == N1)
1563 return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT,
1564 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1565 // fold ((A-(B-C))-C) -> A-B
1566 if (N0.getOpcode() == ISD::SUB &&
1567 N0.getOperand(1).getOpcode() == ISD::SUB &&
1568 N0.getOperand(1).getOperand(1) == N1)
1569 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1570 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1572 // If either operand of a sub is undef, the result is undef
1573 if (N0.getOpcode() == ISD::UNDEF)
1575 if (N1.getOpcode() == ISD::UNDEF)
1578 // If the relocation model supports it, consider symbol offsets.
1579 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1580 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA)) {
1581 // fold (sub Sym, c) -> Sym-c
1582 if (N1C && GA->getOpcode() == ISD::GlobalAddress)
1583 return DAG.getGlobalAddress(GA->getGlobal(), N1C->getDebugLoc(), VT,
1585 (uint64_t)N1C->getSExtValue());
1586 // fold (sub Sym+c1, Sym+c2) -> c1-c2
1587 if (GlobalAddressSDNode *GB = dyn_cast<GlobalAddressSDNode>(N1))
1588 if (GA->getGlobal() == GB->getGlobal())
1589 return DAG.getConstant((uint64_t)GA->getOffset() - GB->getOffset(),
1596 SDValue DAGCombiner::visitMUL(SDNode *N) {
1597 SDValue N0 = N->getOperand(0);
1598 SDValue N1 = N->getOperand(1);
1599 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1600 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1601 EVT VT = N0.getValueType();
1604 if (VT.isVector()) {
1605 SDValue FoldedVOp = SimplifyVBinOp(N);
1606 if (FoldedVOp.getNode()) return FoldedVOp;
1609 // fold (mul x, undef) -> 0
1610 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
1611 return DAG.getConstant(0, VT);
1612 // fold (mul c1, c2) -> c1*c2
1614 return DAG.FoldConstantArithmetic(ISD::MUL, VT, N0C, N1C);
1615 // canonicalize constant to RHS
1617 return DAG.getNode(ISD::MUL, N->getDebugLoc(), VT, N1, N0);
1618 // fold (mul x, 0) -> 0
1619 if (N1C && N1C->isNullValue())
1621 // fold (mul x, -1) -> 0-x
1622 if (N1C && N1C->isAllOnesValue())
1623 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1624 DAG.getConstant(0, VT), N0);
1625 // fold (mul x, (1 << c)) -> x << c
1626 if (N1C && N1C->getAPIntValue().isPowerOf2())
1627 return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0,
1628 DAG.getConstant(N1C->getAPIntValue().logBase2(),
1629 getShiftAmountTy()));
1630 // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c
1631 if (N1C && (-N1C->getAPIntValue()).isPowerOf2()) {
1632 unsigned Log2Val = (-N1C->getAPIntValue()).logBase2();
1633 // FIXME: If the input is something that is easily negated (e.g. a
1634 // single-use add), we should put the negate there.
1635 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1636 DAG.getConstant(0, VT),
1637 DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0,
1638 DAG.getConstant(Log2Val, getShiftAmountTy())));
1640 // (mul (shl X, c1), c2) -> (mul X, c2 << c1)
1641 if (N1C && N0.getOpcode() == ISD::SHL &&
1642 isa<ConstantSDNode>(N0.getOperand(1))) {
1643 SDValue C3 = DAG.getNode(ISD::SHL, N->getDebugLoc(), VT,
1644 N1, N0.getOperand(1));
1645 AddToWorkList(C3.getNode());
1646 return DAG.getNode(ISD::MUL, N->getDebugLoc(), VT,
1647 N0.getOperand(0), C3);
1650 // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one
1653 SDValue Sh(0,0), Y(0,0);
1654 // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)).
1655 if (N0.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N0.getOperand(1)) &&
1656 N0.getNode()->hasOneUse()) {
1658 } else if (N1.getOpcode() == ISD::SHL &&
1659 isa<ConstantSDNode>(N1.getOperand(1)) &&
1660 N1.getNode()->hasOneUse()) {
1665 SDValue Mul = DAG.getNode(ISD::MUL, N->getDebugLoc(), VT,
1666 Sh.getOperand(0), Y);
1667 return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT,
1668 Mul, Sh.getOperand(1));
1672 // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2)
1673 if (N1C && N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() &&
1674 isa<ConstantSDNode>(N0.getOperand(1)))
1675 return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT,
1676 DAG.getNode(ISD::MUL, N0.getDebugLoc(), VT,
1677 N0.getOperand(0), N1),
1678 DAG.getNode(ISD::MUL, N1.getDebugLoc(), VT,
1679 N0.getOperand(1), N1));
1682 SDValue RMUL = ReassociateOps(ISD::MUL, N->getDebugLoc(), N0, N1);
1683 if (RMUL.getNode() != 0)
1689 SDValue DAGCombiner::visitSDIV(SDNode *N) {
1690 SDValue N0 = N->getOperand(0);
1691 SDValue N1 = N->getOperand(1);
1692 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1693 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1694 EVT VT = N->getValueType(0);
1697 if (VT.isVector()) {
1698 SDValue FoldedVOp = SimplifyVBinOp(N);
1699 if (FoldedVOp.getNode()) return FoldedVOp;
1702 // fold (sdiv c1, c2) -> c1/c2
1703 if (N0C && N1C && !N1C->isNullValue())
1704 return DAG.FoldConstantArithmetic(ISD::SDIV, VT, N0C, N1C);
1705 // fold (sdiv X, 1) -> X
1706 if (N1C && N1C->getSExtValue() == 1LL)
1708 // fold (sdiv X, -1) -> 0-X
1709 if (N1C && N1C->isAllOnesValue())
1710 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1711 DAG.getConstant(0, VT), N0);
1712 // If we know the sign bits of both operands are zero, strength reduce to a
1713 // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2
1714 if (!VT.isVector()) {
1715 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
1716 return DAG.getNode(ISD::UDIV, N->getDebugLoc(), N1.getValueType(),
1719 // fold (sdiv X, pow2) -> simple ops after legalize
1720 if (N1C && !N1C->isNullValue() && !TLI.isIntDivCheap() &&
1721 (isPowerOf2_64(N1C->getSExtValue()) ||
1722 isPowerOf2_64(-N1C->getSExtValue()))) {
1723 // If dividing by powers of two is cheap, then don't perform the following
1725 if (TLI.isPow2DivCheap())
1728 int64_t pow2 = N1C->getSExtValue();
1729 int64_t abs2 = pow2 > 0 ? pow2 : -pow2;
1730 unsigned lg2 = Log2_64(abs2);
1732 // Splat the sign bit into the register
1733 SDValue SGN = DAG.getNode(ISD::SRA, N->getDebugLoc(), VT, N0,
1734 DAG.getConstant(VT.getSizeInBits()-1,
1735 getShiftAmountTy()));
1736 AddToWorkList(SGN.getNode());
1738 // Add (N0 < 0) ? abs2 - 1 : 0;
1739 SDValue SRL = DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, SGN,
1740 DAG.getConstant(VT.getSizeInBits() - lg2,
1741 getShiftAmountTy()));
1742 SDValue ADD = DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N0, SRL);
1743 AddToWorkList(SRL.getNode());
1744 AddToWorkList(ADD.getNode()); // Divide by pow2
1745 SDValue SRA = DAG.getNode(ISD::SRA, N->getDebugLoc(), VT, ADD,
1746 DAG.getConstant(lg2, getShiftAmountTy()));
1748 // If we're dividing by a positive value, we're done. Otherwise, we must
1749 // negate the result.
1753 AddToWorkList(SRA.getNode());
1754 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1755 DAG.getConstant(0, VT), SRA);
1758 // if integer divide is expensive and we satisfy the requirements, emit an
1759 // alternate sequence.
1760 if (N1C && (N1C->getSExtValue() < -1 || N1C->getSExtValue() > 1) &&
1761 !TLI.isIntDivCheap()) {
1762 SDValue Op = BuildSDIV(N);
1763 if (Op.getNode()) return Op;
1767 if (N0.getOpcode() == ISD::UNDEF)
1768 return DAG.getConstant(0, VT);
1769 // X / undef -> undef
1770 if (N1.getOpcode() == ISD::UNDEF)
1776 SDValue DAGCombiner::visitUDIV(SDNode *N) {
1777 SDValue N0 = N->getOperand(0);
1778 SDValue N1 = N->getOperand(1);
1779 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1780 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1781 EVT VT = N->getValueType(0);
1784 if (VT.isVector()) {
1785 SDValue FoldedVOp = SimplifyVBinOp(N);
1786 if (FoldedVOp.getNode()) return FoldedVOp;
1789 // fold (udiv c1, c2) -> c1/c2
1790 if (N0C && N1C && !N1C->isNullValue())
1791 return DAG.FoldConstantArithmetic(ISD::UDIV, VT, N0C, N1C);
1792 // fold (udiv x, (1 << c)) -> x >>u c
1793 if (N1C && N1C->getAPIntValue().isPowerOf2())
1794 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0,
1795 DAG.getConstant(N1C->getAPIntValue().logBase2(),
1796 getShiftAmountTy()));
1797 // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2
1798 if (N1.getOpcode() == ISD::SHL) {
1799 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
1800 if (SHC->getAPIntValue().isPowerOf2()) {
1801 EVT ADDVT = N1.getOperand(1).getValueType();
1802 SDValue Add = DAG.getNode(ISD::ADD, N->getDebugLoc(), ADDVT,
1804 DAG.getConstant(SHC->getAPIntValue()
1807 AddToWorkList(Add.getNode());
1808 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0, Add);
1812 // fold (udiv x, c) -> alternate
1813 if (N1C && !N1C->isNullValue() && !TLI.isIntDivCheap()) {
1814 SDValue Op = BuildUDIV(N);
1815 if (Op.getNode()) return Op;
1819 if (N0.getOpcode() == ISD::UNDEF)
1820 return DAG.getConstant(0, VT);
1821 // X / undef -> undef
1822 if (N1.getOpcode() == ISD::UNDEF)
1828 SDValue DAGCombiner::visitSREM(SDNode *N) {
1829 SDValue N0 = N->getOperand(0);
1830 SDValue N1 = N->getOperand(1);
1831 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1832 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1833 EVT VT = N->getValueType(0);
1835 // fold (srem c1, c2) -> c1%c2
1836 if (N0C && N1C && !N1C->isNullValue())
1837 return DAG.FoldConstantArithmetic(ISD::SREM, VT, N0C, N1C);
1838 // If we know the sign bits of both operands are zero, strength reduce to a
1839 // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15
1840 if (!VT.isVector()) {
1841 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
1842 return DAG.getNode(ISD::UREM, N->getDebugLoc(), VT, N0, N1);
1845 // If X/C can be simplified by the division-by-constant logic, lower
1846 // X%C to the equivalent of X-X/C*C.
1847 if (N1C && !N1C->isNullValue()) {
1848 SDValue Div = DAG.getNode(ISD::SDIV, N->getDebugLoc(), VT, N0, N1);
1849 AddToWorkList(Div.getNode());
1850 SDValue OptimizedDiv = combine(Div.getNode());
1851 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
1852 SDValue Mul = DAG.getNode(ISD::MUL, N->getDebugLoc(), VT,
1854 SDValue Sub = DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N0, Mul);
1855 AddToWorkList(Mul.getNode());
1861 if (N0.getOpcode() == ISD::UNDEF)
1862 return DAG.getConstant(0, VT);
1863 // X % undef -> undef
1864 if (N1.getOpcode() == ISD::UNDEF)
1870 SDValue DAGCombiner::visitUREM(SDNode *N) {
1871 SDValue N0 = N->getOperand(0);
1872 SDValue N1 = N->getOperand(1);
1873 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1874 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1875 EVT VT = N->getValueType(0);
1877 // fold (urem c1, c2) -> c1%c2
1878 if (N0C && N1C && !N1C->isNullValue())
1879 return DAG.FoldConstantArithmetic(ISD::UREM, VT, N0C, N1C);
1880 // fold (urem x, pow2) -> (and x, pow2-1)
1881 if (N1C && !N1C->isNullValue() && N1C->getAPIntValue().isPowerOf2())
1882 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0,
1883 DAG.getConstant(N1C->getAPIntValue()-1,VT));
1884 // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1))
1885 if (N1.getOpcode() == ISD::SHL) {
1886 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
1887 if (SHC->getAPIntValue().isPowerOf2()) {
1889 DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N1,
1890 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()),
1892 AddToWorkList(Add.getNode());
1893 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0, Add);
1898 // If X/C can be simplified by the division-by-constant logic, lower
1899 // X%C to the equivalent of X-X/C*C.
1900 if (N1C && !N1C->isNullValue()) {
1901 SDValue Div = DAG.getNode(ISD::UDIV, N->getDebugLoc(), VT, N0, N1);
1902 AddToWorkList(Div.getNode());
1903 SDValue OptimizedDiv = combine(Div.getNode());
1904 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
1905 SDValue Mul = DAG.getNode(ISD::MUL, N->getDebugLoc(), VT,
1907 SDValue Sub = DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N0, Mul);
1908 AddToWorkList(Mul.getNode());
1914 if (N0.getOpcode() == ISD::UNDEF)
1915 return DAG.getConstant(0, VT);
1916 // X % undef -> undef
1917 if (N1.getOpcode() == ISD::UNDEF)
1923 SDValue DAGCombiner::visitMULHS(SDNode *N) {
1924 SDValue N0 = N->getOperand(0);
1925 SDValue N1 = N->getOperand(1);
1926 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1927 EVT VT = N->getValueType(0);
1928 DebugLoc DL = N->getDebugLoc();
1930 // fold (mulhs x, 0) -> 0
1931 if (N1C && N1C->isNullValue())
1933 // fold (mulhs x, 1) -> (sra x, size(x)-1)
1934 if (N1C && N1C->getAPIntValue() == 1)
1935 return DAG.getNode(ISD::SRA, N->getDebugLoc(), N0.getValueType(), N0,
1936 DAG.getConstant(N0.getValueType().getSizeInBits() - 1,
1937 getShiftAmountTy()));
1938 // fold (mulhs x, undef) -> 0
1939 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
1940 return DAG.getConstant(0, VT);
1942 // If the type twice as wide is legal, transform the mulhs to a wider multiply
1944 if (VT.isSimple() && !VT.isVector()) {
1945 MVT Simple = VT.getSimpleVT();
1946 unsigned SimpleSize = Simple.getSizeInBits();
1947 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
1948 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
1949 N0 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N0);
1950 N1 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N1);
1951 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
1952 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
1953 DAG.getConstant(SimpleSize, getShiftAmountTy()));
1954 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
1961 SDValue DAGCombiner::visitMULHU(SDNode *N) {
1962 SDValue N0 = N->getOperand(0);
1963 SDValue N1 = N->getOperand(1);
1964 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1965 EVT VT = N->getValueType(0);
1966 DebugLoc DL = N->getDebugLoc();
1968 // fold (mulhu x, 0) -> 0
1969 if (N1C && N1C->isNullValue())
1971 // fold (mulhu x, 1) -> 0
1972 if (N1C && N1C->getAPIntValue() == 1)
1973 return DAG.getConstant(0, N0.getValueType());
1974 // fold (mulhu x, undef) -> 0
1975 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
1976 return DAG.getConstant(0, VT);
1978 // If the type twice as wide is legal, transform the mulhu to a wider multiply
1980 if (VT.isSimple() && !VT.isVector()) {
1981 MVT Simple = VT.getSimpleVT();
1982 unsigned SimpleSize = Simple.getSizeInBits();
1983 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
1984 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
1985 N0 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N0);
1986 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N1);
1987 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
1988 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
1989 DAG.getConstant(SimpleSize, getShiftAmountTy()));
1990 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
1997 /// SimplifyNodeWithTwoResults - Perform optimizations common to nodes that
1998 /// compute two values. LoOp and HiOp give the opcodes for the two computations
1999 /// that are being performed. Return true if a simplification was made.
2001 SDValue DAGCombiner::SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
2003 // If the high half is not needed, just compute the low half.
2004 bool HiExists = N->hasAnyUseOfValue(1);
2006 (!LegalOperations ||
2007 TLI.isOperationLegal(LoOp, N->getValueType(0)))) {
2008 SDValue Res = DAG.getNode(LoOp, N->getDebugLoc(), N->getValueType(0),
2009 N->op_begin(), N->getNumOperands());
2010 return CombineTo(N, Res, Res);
2013 // If the low half is not needed, just compute the high half.
2014 bool LoExists = N->hasAnyUseOfValue(0);
2016 (!LegalOperations ||
2017 TLI.isOperationLegal(HiOp, N->getValueType(1)))) {
2018 SDValue Res = DAG.getNode(HiOp, N->getDebugLoc(), N->getValueType(1),
2019 N->op_begin(), N->getNumOperands());
2020 return CombineTo(N, Res, Res);
2023 // If both halves are used, return as it is.
2024 if (LoExists && HiExists)
2027 // If the two computed results can be simplified separately, separate them.
2029 SDValue Lo = DAG.getNode(LoOp, N->getDebugLoc(), N->getValueType(0),
2030 N->op_begin(), N->getNumOperands());
2031 AddToWorkList(Lo.getNode());
2032 SDValue LoOpt = combine(Lo.getNode());
2033 if (LoOpt.getNode() && LoOpt.getNode() != Lo.getNode() &&
2034 (!LegalOperations ||
2035 TLI.isOperationLegal(LoOpt.getOpcode(), LoOpt.getValueType())))
2036 return CombineTo(N, LoOpt, LoOpt);
2040 SDValue Hi = DAG.getNode(HiOp, N->getDebugLoc(), N->getValueType(1),
2041 N->op_begin(), N->getNumOperands());
2042 AddToWorkList(Hi.getNode());
2043 SDValue HiOpt = combine(Hi.getNode());
2044 if (HiOpt.getNode() && HiOpt != Hi &&
2045 (!LegalOperations ||
2046 TLI.isOperationLegal(HiOpt.getOpcode(), HiOpt.getValueType())))
2047 return CombineTo(N, HiOpt, HiOpt);
2053 SDValue DAGCombiner::visitSMUL_LOHI(SDNode *N) {
2054 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHS);
2055 if (Res.getNode()) return Res;
2057 EVT VT = N->getValueType(0);
2058 DebugLoc DL = N->getDebugLoc();
2060 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2062 if (VT.isSimple() && !VT.isVector()) {
2063 MVT Simple = VT.getSimpleVT();
2064 unsigned SimpleSize = Simple.getSizeInBits();
2065 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2066 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2067 SDValue Lo = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(0));
2068 SDValue Hi = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(1));
2069 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2070 // Compute the high part as N1.
2071 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2072 DAG.getConstant(SimpleSize, getShiftAmountTy()));
2073 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2074 // Compute the low part as N0.
2075 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2076 return CombineTo(N, Lo, Hi);
2083 SDValue DAGCombiner::visitUMUL_LOHI(SDNode *N) {
2084 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHU);
2085 if (Res.getNode()) return Res;
2087 EVT VT = N->getValueType(0);
2088 DebugLoc DL = N->getDebugLoc();
2090 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2092 if (VT.isSimple() && !VT.isVector()) {
2093 MVT Simple = VT.getSimpleVT();
2094 unsigned SimpleSize = Simple.getSizeInBits();
2095 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2096 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2097 SDValue Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(0));
2098 SDValue Hi = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(1));
2099 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2100 // Compute the high part as N1.
2101 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2102 DAG.getConstant(SimpleSize, getShiftAmountTy()));
2103 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2104 // Compute the low part as N0.
2105 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2106 return CombineTo(N, Lo, Hi);
2113 SDValue DAGCombiner::visitSDIVREM(SDNode *N) {
2114 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::SDIV, ISD::SREM);
2115 if (Res.getNode()) return Res;
2120 SDValue DAGCombiner::visitUDIVREM(SDNode *N) {
2121 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::UDIV, ISD::UREM);
2122 if (Res.getNode()) return Res;
2127 /// SimplifyBinOpWithSameOpcodeHands - If this is a binary operator with
2128 /// two operands of the same opcode, try to simplify it.
2129 SDValue DAGCombiner::SimplifyBinOpWithSameOpcodeHands(SDNode *N) {
2130 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
2131 EVT VT = N0.getValueType();
2132 assert(N0.getOpcode() == N1.getOpcode() && "Bad input!");
2134 // Bail early if none of these transforms apply.
2135 if (N0.getNode()->getNumOperands() == 0) return SDValue();
2137 // For each of OP in AND/OR/XOR:
2138 // fold (OP (zext x), (zext y)) -> (zext (OP x, y))
2139 // fold (OP (sext x), (sext y)) -> (sext (OP x, y))
2140 // fold (OP (aext x), (aext y)) -> (aext (OP x, y))
2141 // fold (OP (trunc x), (trunc y)) -> (trunc (OP x, y)) (if trunc isn't free)
2143 // do not sink logical op inside of a vector extend, since it may combine
2145 EVT Op0VT = N0.getOperand(0).getValueType();
2146 if ((N0.getOpcode() == ISD::ZERO_EXTEND ||
2147 N0.getOpcode() == ISD::SIGN_EXTEND ||
2148 // Avoid infinite looping with PromoteIntBinOp.
2149 (N0.getOpcode() == ISD::ANY_EXTEND &&
2150 (!LegalTypes || TLI.isTypeDesirableForOp(N->getOpcode(), Op0VT))) ||
2151 (N0.getOpcode() == ISD::TRUNCATE &&
2152 (!TLI.isZExtFree(VT, Op0VT) ||
2153 !TLI.isTruncateFree(Op0VT, VT)) &&
2154 TLI.isTypeLegal(Op0VT))) &&
2156 Op0VT == N1.getOperand(0).getValueType() &&
2157 (!LegalOperations || TLI.isOperationLegal(N->getOpcode(), Op0VT))) {
2158 SDValue ORNode = DAG.getNode(N->getOpcode(), N0.getDebugLoc(),
2159 N0.getOperand(0).getValueType(),
2160 N0.getOperand(0), N1.getOperand(0));
2161 AddToWorkList(ORNode.getNode());
2162 return DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT, ORNode);
2165 // For each of OP in SHL/SRL/SRA/AND...
2166 // fold (and (OP x, z), (OP y, z)) -> (OP (and x, y), z)
2167 // fold (or (OP x, z), (OP y, z)) -> (OP (or x, y), z)
2168 // fold (xor (OP x, z), (OP y, z)) -> (OP (xor x, y), z)
2169 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL ||
2170 N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::AND) &&
2171 N0.getOperand(1) == N1.getOperand(1)) {
2172 SDValue ORNode = DAG.getNode(N->getOpcode(), N0.getDebugLoc(),
2173 N0.getOperand(0).getValueType(),
2174 N0.getOperand(0), N1.getOperand(0));
2175 AddToWorkList(ORNode.getNode());
2176 return DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT,
2177 ORNode, N0.getOperand(1));
2183 SDValue DAGCombiner::visitAND(SDNode *N) {
2184 SDValue N0 = N->getOperand(0);
2185 SDValue N1 = N->getOperand(1);
2186 SDValue LL, LR, RL, RR, CC0, CC1;
2187 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2188 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2189 EVT VT = N1.getValueType();
2190 unsigned BitWidth = VT.getScalarType().getSizeInBits();
2193 if (VT.isVector()) {
2194 SDValue FoldedVOp = SimplifyVBinOp(N);
2195 if (FoldedVOp.getNode()) return FoldedVOp;
2198 // fold (and x, undef) -> 0
2199 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2200 return DAG.getConstant(0, VT);
2201 // fold (and c1, c2) -> c1&c2
2203 return DAG.FoldConstantArithmetic(ISD::AND, VT, N0C, N1C);
2204 // canonicalize constant to RHS
2206 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N1, N0);
2207 // fold (and x, -1) -> x
2208 if (N1C && N1C->isAllOnesValue())
2210 // if (and x, c) is known to be zero, return 0
2211 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
2212 APInt::getAllOnesValue(BitWidth)))
2213 return DAG.getConstant(0, VT);
2215 SDValue RAND = ReassociateOps(ISD::AND, N->getDebugLoc(), N0, N1);
2216 if (RAND.getNode() != 0)
2218 // fold (and (or x, C), D) -> D if (C & D) == D
2219 if (N1C && N0.getOpcode() == ISD::OR)
2220 if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
2221 if ((ORI->getAPIntValue() & N1C->getAPIntValue()) == N1C->getAPIntValue())
2223 // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits.
2224 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
2225 SDValue N0Op0 = N0.getOperand(0);
2226 APInt Mask = ~N1C->getAPIntValue();
2227 Mask = Mask.trunc(N0Op0.getValueSizeInBits());
2228 if (DAG.MaskedValueIsZero(N0Op0, Mask)) {
2229 SDValue Zext = DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(),
2230 N0.getValueType(), N0Op0);
2232 // Replace uses of the AND with uses of the Zero extend node.
2235 // We actually want to replace all uses of the any_extend with the
2236 // zero_extend, to avoid duplicating things. This will later cause this
2237 // AND to be folded.
2238 CombineTo(N0.getNode(), Zext);
2239 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2242 // fold (and (setcc x), (setcc y)) -> (setcc (and x, y))
2243 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
2244 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
2245 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
2247 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
2248 LL.getValueType().isInteger()) {
2249 // fold (and (seteq X, 0), (seteq Y, 0)) -> (seteq (or X, Y), 0)
2250 if (cast<ConstantSDNode>(LR)->isNullValue() && Op1 == ISD::SETEQ) {
2251 SDValue ORNode = DAG.getNode(ISD::OR, N0.getDebugLoc(),
2252 LR.getValueType(), LL, RL);
2253 AddToWorkList(ORNode.getNode());
2254 return DAG.getSetCC(N->getDebugLoc(), VT, ORNode, LR, Op1);
2256 // fold (and (seteq X, -1), (seteq Y, -1)) -> (seteq (and X, Y), -1)
2257 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETEQ) {
2258 SDValue ANDNode = DAG.getNode(ISD::AND, N0.getDebugLoc(),
2259 LR.getValueType(), LL, RL);
2260 AddToWorkList(ANDNode.getNode());
2261 return DAG.getSetCC(N->getDebugLoc(), VT, ANDNode, LR, Op1);
2263 // fold (and (setgt X, -1), (setgt Y, -1)) -> (setgt (or X, Y), -1)
2264 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETGT) {
2265 SDValue ORNode = DAG.getNode(ISD::OR, N0.getDebugLoc(),
2266 LR.getValueType(), LL, RL);
2267 AddToWorkList(ORNode.getNode());
2268 return DAG.getSetCC(N->getDebugLoc(), VT, ORNode, LR, Op1);
2271 // canonicalize equivalent to ll == rl
2272 if (LL == RR && LR == RL) {
2273 Op1 = ISD::getSetCCSwappedOperands(Op1);
2276 if (LL == RL && LR == RR) {
2277 bool isInteger = LL.getValueType().isInteger();
2278 ISD::CondCode Result = ISD::getSetCCAndOperation(Op0, Op1, isInteger);
2279 if (Result != ISD::SETCC_INVALID &&
2280 (!LegalOperations || TLI.isCondCodeLegal(Result, LL.getValueType())))
2281 return DAG.getSetCC(N->getDebugLoc(), N0.getValueType(),
2286 // Simplify: (and (op x...), (op y...)) -> (op (and x, y))
2287 if (N0.getOpcode() == N1.getOpcode()) {
2288 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
2289 if (Tmp.getNode()) return Tmp;
2292 // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1)
2293 // fold (and (sra)) -> (and (srl)) when possible.
2294 if (!VT.isVector() &&
2295 SimplifyDemandedBits(SDValue(N, 0)))
2296 return SDValue(N, 0);
2298 // fold (zext_inreg (extload x)) -> (zextload x)
2299 if (ISD::isEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode())) {
2300 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2301 EVT MemVT = LN0->getMemoryVT();
2302 // If we zero all the possible extended bits, then we can turn this into
2303 // a zextload if we are running before legalize or the operation is legal.
2304 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
2305 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
2306 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
2307 ((!LegalOperations && !LN0->isVolatile()) ||
2308 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
2309 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, N0.getDebugLoc(),
2310 LN0->getChain(), LN0->getBasePtr(),
2311 LN0->getPointerInfo(), MemVT,
2312 LN0->isVolatile(), LN0->isNonTemporal(),
2313 LN0->getAlignment());
2315 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
2316 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2319 // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use
2320 if (ISD::isSEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
2322 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2323 EVT MemVT = LN0->getMemoryVT();
2324 // If we zero all the possible extended bits, then we can turn this into
2325 // a zextload if we are running before legalize or the operation is legal.
2326 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
2327 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
2328 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
2329 ((!LegalOperations && !LN0->isVolatile()) ||
2330 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
2331 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, N0.getDebugLoc(),
2333 LN0->getBasePtr(), LN0->getPointerInfo(),
2335 LN0->isVolatile(), LN0->isNonTemporal(),
2336 LN0->getAlignment());
2338 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
2339 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2343 // fold (and (load x), 255) -> (zextload x, i8)
2344 // fold (and (extload x, i16), 255) -> (zextload x, i8)
2345 // fold (and (any_ext (extload x, i16)), 255) -> (zextload x, i8)
2346 if (N1C && (N0.getOpcode() == ISD::LOAD ||
2347 (N0.getOpcode() == ISD::ANY_EXTEND &&
2348 N0.getOperand(0).getOpcode() == ISD::LOAD))) {
2349 bool HasAnyExt = N0.getOpcode() == ISD::ANY_EXTEND;
2350 LoadSDNode *LN0 = HasAnyExt
2351 ? cast<LoadSDNode>(N0.getOperand(0))
2352 : cast<LoadSDNode>(N0);
2353 if (LN0->getExtensionType() != ISD::SEXTLOAD &&
2354 LN0->isUnindexed() && N0.hasOneUse() && LN0->hasOneUse()) {
2355 uint32_t ActiveBits = N1C->getAPIntValue().getActiveBits();
2356 if (ActiveBits > 0 && APIntOps::isMask(ActiveBits, N1C->getAPIntValue())){
2357 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits);
2358 EVT LoadedVT = LN0->getMemoryVT();
2360 if (ExtVT == LoadedVT &&
2361 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) {
2362 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
2365 DAG.getExtLoad(ISD::ZEXTLOAD, LoadResultTy, LN0->getDebugLoc(),
2366 LN0->getChain(), LN0->getBasePtr(),
2367 LN0->getPointerInfo(),
2368 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
2369 LN0->getAlignment());
2371 CombineTo(LN0, NewLoad, NewLoad.getValue(1));
2372 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2375 // Do not change the width of a volatile load.
2376 // Do not generate loads of non-round integer types since these can
2377 // be expensive (and would be wrong if the type is not byte sized).
2378 if (!LN0->isVolatile() && LoadedVT.bitsGT(ExtVT) && ExtVT.isRound() &&
2379 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) {
2380 EVT PtrType = LN0->getOperand(1).getValueType();
2382 unsigned Alignment = LN0->getAlignment();
2383 SDValue NewPtr = LN0->getBasePtr();
2385 // For big endian targets, we need to add an offset to the pointer
2386 // to load the correct bytes. For little endian systems, we merely
2387 // need to read fewer bytes from the same pointer.
2388 if (TLI.isBigEndian()) {
2389 unsigned LVTStoreBytes = LoadedVT.getStoreSize();
2390 unsigned EVTStoreBytes = ExtVT.getStoreSize();
2391 unsigned PtrOff = LVTStoreBytes - EVTStoreBytes;
2392 NewPtr = DAG.getNode(ISD::ADD, LN0->getDebugLoc(), PtrType,
2393 NewPtr, DAG.getConstant(PtrOff, PtrType));
2394 Alignment = MinAlign(Alignment, PtrOff);
2397 AddToWorkList(NewPtr.getNode());
2399 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
2401 DAG.getExtLoad(ISD::ZEXTLOAD, LoadResultTy, LN0->getDebugLoc(),
2402 LN0->getChain(), NewPtr,
2403 LN0->getPointerInfo(),
2404 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
2407 CombineTo(LN0, Load, Load.getValue(1));
2408 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2417 SDValue DAGCombiner::visitOR(SDNode *N) {
2418 SDValue N0 = N->getOperand(0);
2419 SDValue N1 = N->getOperand(1);
2420 SDValue LL, LR, RL, RR, CC0, CC1;
2421 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2422 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2423 EVT VT = N1.getValueType();
2426 if (VT.isVector()) {
2427 SDValue FoldedVOp = SimplifyVBinOp(N);
2428 if (FoldedVOp.getNode()) return FoldedVOp;
2431 // fold (or x, undef) -> -1
2432 if (!LegalOperations &&
2433 (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)) {
2434 EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT;
2435 return DAG.getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()), VT);
2437 // fold (or c1, c2) -> c1|c2
2439 return DAG.FoldConstantArithmetic(ISD::OR, VT, N0C, N1C);
2440 // canonicalize constant to RHS
2442 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N1, N0);
2443 // fold (or x, 0) -> x
2444 if (N1C && N1C->isNullValue())
2446 // fold (or x, -1) -> -1
2447 if (N1C && N1C->isAllOnesValue())
2449 // fold (or x, c) -> c iff (x & ~c) == 0
2450 if (N1C && DAG.MaskedValueIsZero(N0, ~N1C->getAPIntValue()))
2453 SDValue ROR = ReassociateOps(ISD::OR, N->getDebugLoc(), N0, N1);
2454 if (ROR.getNode() != 0)
2456 // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2)
2457 // iff (c1 & c2) == 0.
2458 if (N1C && N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
2459 isa<ConstantSDNode>(N0.getOperand(1))) {
2460 ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1));
2461 if ((C1->getAPIntValue() & N1C->getAPIntValue()) != 0)
2462 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
2463 DAG.getNode(ISD::OR, N0.getDebugLoc(), VT,
2464 N0.getOperand(0), N1),
2465 DAG.FoldConstantArithmetic(ISD::OR, VT, N1C, C1));
2467 // fold (or (setcc x), (setcc y)) -> (setcc (or x, y))
2468 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
2469 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
2470 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
2472 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
2473 LL.getValueType().isInteger()) {
2474 // fold (or (setne X, 0), (setne Y, 0)) -> (setne (or X, Y), 0)
2475 // fold (or (setlt X, 0), (setlt Y, 0)) -> (setne (or X, Y), 0)
2476 if (cast<ConstantSDNode>(LR)->isNullValue() &&
2477 (Op1 == ISD::SETNE || Op1 == ISD::SETLT)) {
2478 SDValue ORNode = DAG.getNode(ISD::OR, LR.getDebugLoc(),
2479 LR.getValueType(), LL, RL);
2480 AddToWorkList(ORNode.getNode());
2481 return DAG.getSetCC(N->getDebugLoc(), VT, ORNode, LR, Op1);
2483 // fold (or (setne X, -1), (setne Y, -1)) -> (setne (and X, Y), -1)
2484 // fold (or (setgt X, -1), (setgt Y -1)) -> (setgt (and X, Y), -1)
2485 if (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
2486 (Op1 == ISD::SETNE || Op1 == ISD::SETGT)) {
2487 SDValue ANDNode = DAG.getNode(ISD::AND, LR.getDebugLoc(),
2488 LR.getValueType(), LL, RL);
2489 AddToWorkList(ANDNode.getNode());
2490 return DAG.getSetCC(N->getDebugLoc(), VT, ANDNode, LR, Op1);
2493 // canonicalize equivalent to ll == rl
2494 if (LL == RR && LR == RL) {
2495 Op1 = ISD::getSetCCSwappedOperands(Op1);
2498 if (LL == RL && LR == RR) {
2499 bool isInteger = LL.getValueType().isInteger();
2500 ISD::CondCode Result = ISD::getSetCCOrOperation(Op0, Op1, isInteger);
2501 if (Result != ISD::SETCC_INVALID &&
2502 (!LegalOperations || TLI.isCondCodeLegal(Result, LL.getValueType())))
2503 return DAG.getSetCC(N->getDebugLoc(), N0.getValueType(),
2508 // Simplify: (or (op x...), (op y...)) -> (op (or x, y))
2509 if (N0.getOpcode() == N1.getOpcode()) {
2510 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
2511 if (Tmp.getNode()) return Tmp;
2514 // (or (and X, C1), (and Y, C2)) -> (and (or X, Y), C3) if possible.
2515 if (N0.getOpcode() == ISD::AND &&
2516 N1.getOpcode() == ISD::AND &&
2517 N0.getOperand(1).getOpcode() == ISD::Constant &&
2518 N1.getOperand(1).getOpcode() == ISD::Constant &&
2519 // Don't increase # computations.
2520 (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) {
2521 // We can only do this xform if we know that bits from X that are set in C2
2522 // but not in C1 are already zero. Likewise for Y.
2523 const APInt &LHSMask =
2524 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
2525 const APInt &RHSMask =
2526 cast<ConstantSDNode>(N1.getOperand(1))->getAPIntValue();
2528 if (DAG.MaskedValueIsZero(N0.getOperand(0), RHSMask&~LHSMask) &&
2529 DAG.MaskedValueIsZero(N1.getOperand(0), LHSMask&~RHSMask)) {
2530 SDValue X = DAG.getNode(ISD::OR, N0.getDebugLoc(), VT,
2531 N0.getOperand(0), N1.getOperand(0));
2532 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, X,
2533 DAG.getConstant(LHSMask | RHSMask, VT));
2537 // See if this is some rotate idiom.
2538 if (SDNode *Rot = MatchRotate(N0, N1, N->getDebugLoc()))
2539 return SDValue(Rot, 0);
2541 // Simplify the operands using demanded-bits information.
2542 if (!VT.isVector() &&
2543 SimplifyDemandedBits(SDValue(N, 0)))
2544 return SDValue(N, 0);
2549 /// MatchRotateHalf - Match "(X shl/srl V1) & V2" where V2 may not be present.
2550 static bool MatchRotateHalf(SDValue Op, SDValue &Shift, SDValue &Mask) {
2551 if (Op.getOpcode() == ISD::AND) {
2552 if (isa<ConstantSDNode>(Op.getOperand(1))) {
2553 Mask = Op.getOperand(1);
2554 Op = Op.getOperand(0);
2560 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SHL) {
2568 // MatchRotate - Handle an 'or' of two operands. If this is one of the many
2569 // idioms for rotate, and if the target supports rotation instructions, generate
2571 SDNode *DAGCombiner::MatchRotate(SDValue LHS, SDValue RHS, DebugLoc DL) {
2572 // Must be a legal type. Expanded 'n promoted things won't work with rotates.
2573 EVT VT = LHS.getValueType();
2574 if (!TLI.isTypeLegal(VT)) return 0;
2576 // The target must have at least one rotate flavor.
2577 bool HasROTL = TLI.isOperationLegalOrCustom(ISD::ROTL, VT);
2578 bool HasROTR = TLI.isOperationLegalOrCustom(ISD::ROTR, VT);
2579 if (!HasROTL && !HasROTR) return 0;
2581 // Match "(X shl/srl V1) & V2" where V2 may not be present.
2582 SDValue LHSShift; // The shift.
2583 SDValue LHSMask; // AND value if any.
2584 if (!MatchRotateHalf(LHS, LHSShift, LHSMask))
2585 return 0; // Not part of a rotate.
2587 SDValue RHSShift; // The shift.
2588 SDValue RHSMask; // AND value if any.
2589 if (!MatchRotateHalf(RHS, RHSShift, RHSMask))
2590 return 0; // Not part of a rotate.
2592 if (LHSShift.getOperand(0) != RHSShift.getOperand(0))
2593 return 0; // Not shifting the same value.
2595 if (LHSShift.getOpcode() == RHSShift.getOpcode())
2596 return 0; // Shifts must disagree.
2598 // Canonicalize shl to left side in a shl/srl pair.
2599 if (RHSShift.getOpcode() == ISD::SHL) {
2600 std::swap(LHS, RHS);
2601 std::swap(LHSShift, RHSShift);
2602 std::swap(LHSMask , RHSMask );
2605 unsigned OpSizeInBits = VT.getSizeInBits();
2606 SDValue LHSShiftArg = LHSShift.getOperand(0);
2607 SDValue LHSShiftAmt = LHSShift.getOperand(1);
2608 SDValue RHSShiftAmt = RHSShift.getOperand(1);
2610 // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1)
2611 // fold (or (shl x, C1), (srl x, C2)) -> (rotr x, C2)
2612 if (LHSShiftAmt.getOpcode() == ISD::Constant &&
2613 RHSShiftAmt.getOpcode() == ISD::Constant) {
2614 uint64_t LShVal = cast<ConstantSDNode>(LHSShiftAmt)->getZExtValue();
2615 uint64_t RShVal = cast<ConstantSDNode>(RHSShiftAmt)->getZExtValue();
2616 if ((LShVal + RShVal) != OpSizeInBits)
2621 Rot = DAG.getNode(ISD::ROTL, DL, VT, LHSShiftArg, LHSShiftAmt);
2623 Rot = DAG.getNode(ISD::ROTR, DL, VT, LHSShiftArg, RHSShiftAmt);
2625 // If there is an AND of either shifted operand, apply it to the result.
2626 if (LHSMask.getNode() || RHSMask.getNode()) {
2627 APInt Mask = APInt::getAllOnesValue(OpSizeInBits);
2629 if (LHSMask.getNode()) {
2630 APInt RHSBits = APInt::getLowBitsSet(OpSizeInBits, LShVal);
2631 Mask &= cast<ConstantSDNode>(LHSMask)->getAPIntValue() | RHSBits;
2633 if (RHSMask.getNode()) {
2634 APInt LHSBits = APInt::getHighBitsSet(OpSizeInBits, RShVal);
2635 Mask &= cast<ConstantSDNode>(RHSMask)->getAPIntValue() | LHSBits;
2638 Rot = DAG.getNode(ISD::AND, DL, VT, Rot, DAG.getConstant(Mask, VT));
2641 return Rot.getNode();
2644 // If there is a mask here, and we have a variable shift, we can't be sure
2645 // that we're masking out the right stuff.
2646 if (LHSMask.getNode() || RHSMask.getNode())
2649 // fold (or (shl x, y), (srl x, (sub 32, y))) -> (rotl x, y)
2650 // fold (or (shl x, y), (srl x, (sub 32, y))) -> (rotr x, (sub 32, y))
2651 if (RHSShiftAmt.getOpcode() == ISD::SUB &&
2652 LHSShiftAmt == RHSShiftAmt.getOperand(1)) {
2653 if (ConstantSDNode *SUBC =
2654 dyn_cast<ConstantSDNode>(RHSShiftAmt.getOperand(0))) {
2655 if (SUBC->getAPIntValue() == OpSizeInBits) {
2657 return DAG.getNode(ISD::ROTL, DL, VT,
2658 LHSShiftArg, LHSShiftAmt).getNode();
2660 return DAG.getNode(ISD::ROTR, DL, VT,
2661 LHSShiftArg, RHSShiftAmt).getNode();
2666 // fold (or (shl x, (sub 32, y)), (srl x, r)) -> (rotr x, y)
2667 // fold (or (shl x, (sub 32, y)), (srl x, r)) -> (rotl x, (sub 32, y))
2668 if (LHSShiftAmt.getOpcode() == ISD::SUB &&
2669 RHSShiftAmt == LHSShiftAmt.getOperand(1)) {
2670 if (ConstantSDNode *SUBC =
2671 dyn_cast<ConstantSDNode>(LHSShiftAmt.getOperand(0))) {
2672 if (SUBC->getAPIntValue() == OpSizeInBits) {
2674 return DAG.getNode(ISD::ROTR, DL, VT,
2675 LHSShiftArg, RHSShiftAmt).getNode();
2677 return DAG.getNode(ISD::ROTL, DL, VT,
2678 LHSShiftArg, LHSShiftAmt).getNode();
2683 // Look for sign/zext/any-extended or truncate cases:
2684 if ((LHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND
2685 || LHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND
2686 || LHSShiftAmt.getOpcode() == ISD::ANY_EXTEND
2687 || LHSShiftAmt.getOpcode() == ISD::TRUNCATE) &&
2688 (RHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND
2689 || RHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND
2690 || RHSShiftAmt.getOpcode() == ISD::ANY_EXTEND
2691 || RHSShiftAmt.getOpcode() == ISD::TRUNCATE)) {
2692 SDValue LExtOp0 = LHSShiftAmt.getOperand(0);
2693 SDValue RExtOp0 = RHSShiftAmt.getOperand(0);
2694 if (RExtOp0.getOpcode() == ISD::SUB &&
2695 RExtOp0.getOperand(1) == LExtOp0) {
2696 // fold (or (shl x, (*ext y)), (srl x, (*ext (sub 32, y)))) ->
2698 // fold (or (shl x, (*ext y)), (srl x, (*ext (sub 32, y)))) ->
2699 // (rotr x, (sub 32, y))
2700 if (ConstantSDNode *SUBC =
2701 dyn_cast<ConstantSDNode>(RExtOp0.getOperand(0))) {
2702 if (SUBC->getAPIntValue() == OpSizeInBits) {
2703 return DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT,
2705 HasROTL ? LHSShiftAmt : RHSShiftAmt).getNode();
2708 } else if (LExtOp0.getOpcode() == ISD::SUB &&
2709 RExtOp0 == LExtOp0.getOperand(1)) {
2710 // fold (or (shl x, (*ext (sub 32, y))), (srl x, (*ext y))) ->
2712 // fold (or (shl x, (*ext (sub 32, y))), (srl x, (*ext y))) ->
2713 // (rotl x, (sub 32, y))
2714 if (ConstantSDNode *SUBC =
2715 dyn_cast<ConstantSDNode>(LExtOp0.getOperand(0))) {
2716 if (SUBC->getAPIntValue() == OpSizeInBits) {
2717 return DAG.getNode(HasROTR ? ISD::ROTR : ISD::ROTL, DL, VT,
2719 HasROTR ? RHSShiftAmt : LHSShiftAmt).getNode();
2728 SDValue DAGCombiner::visitXOR(SDNode *N) {
2729 SDValue N0 = N->getOperand(0);
2730 SDValue N1 = N->getOperand(1);
2731 SDValue LHS, RHS, CC;
2732 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2733 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2734 EVT VT = N0.getValueType();
2737 if (VT.isVector()) {
2738 SDValue FoldedVOp = SimplifyVBinOp(N);
2739 if (FoldedVOp.getNode()) return FoldedVOp;
2742 // fold (xor undef, undef) -> 0. This is a common idiom (misuse).
2743 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
2744 return DAG.getConstant(0, VT);
2745 // fold (xor x, undef) -> undef
2746 if (N0.getOpcode() == ISD::UNDEF)
2748 if (N1.getOpcode() == ISD::UNDEF)
2750 // fold (xor c1, c2) -> c1^c2
2752 return DAG.FoldConstantArithmetic(ISD::XOR, VT, N0C, N1C);
2753 // canonicalize constant to RHS
2755 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N1, N0);
2756 // fold (xor x, 0) -> x
2757 if (N1C && N1C->isNullValue())
2760 SDValue RXOR = ReassociateOps(ISD::XOR, N->getDebugLoc(), N0, N1);
2761 if (RXOR.getNode() != 0)
2764 // fold !(x cc y) -> (x !cc y)
2765 if (N1C && N1C->getAPIntValue() == 1 && isSetCCEquivalent(N0, LHS, RHS, CC)) {
2766 bool isInt = LHS.getValueType().isInteger();
2767 ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
2770 if (!LegalOperations || TLI.isCondCodeLegal(NotCC, LHS.getValueType())) {
2771 switch (N0.getOpcode()) {
2773 llvm_unreachable("Unhandled SetCC Equivalent!");
2775 return DAG.getSetCC(N->getDebugLoc(), VT, LHS, RHS, NotCC);
2776 case ISD::SELECT_CC:
2777 return DAG.getSelectCC(N->getDebugLoc(), LHS, RHS, N0.getOperand(2),
2778 N0.getOperand(3), NotCC);
2783 // fold (not (zext (setcc x, y))) -> (zext (not (setcc x, y)))
2784 if (N1C && N1C->getAPIntValue() == 1 && N0.getOpcode() == ISD::ZERO_EXTEND &&
2785 N0.getNode()->hasOneUse() &&
2786 isSetCCEquivalent(N0.getOperand(0), LHS, RHS, CC)){
2787 SDValue V = N0.getOperand(0);
2788 V = DAG.getNode(ISD::XOR, N0.getDebugLoc(), V.getValueType(), V,
2789 DAG.getConstant(1, V.getValueType()));
2790 AddToWorkList(V.getNode());
2791 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, V);
2794 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are setcc
2795 if (N1C && N1C->getAPIntValue() == 1 && VT == MVT::i1 &&
2796 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
2797 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
2798 if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) {
2799 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
2800 LHS = DAG.getNode(ISD::XOR, LHS.getDebugLoc(), VT, LHS, N1); // LHS = ~LHS
2801 RHS = DAG.getNode(ISD::XOR, RHS.getDebugLoc(), VT, RHS, N1); // RHS = ~RHS
2802 AddToWorkList(LHS.getNode()); AddToWorkList(RHS.getNode());
2803 return DAG.getNode(NewOpcode, N->getDebugLoc(), VT, LHS, RHS);
2806 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are constants
2807 if (N1C && N1C->isAllOnesValue() &&
2808 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
2809 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
2810 if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) {
2811 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
2812 LHS = DAG.getNode(ISD::XOR, LHS.getDebugLoc(), VT, LHS, N1); // LHS = ~LHS
2813 RHS = DAG.getNode(ISD::XOR, RHS.getDebugLoc(), VT, RHS, N1); // RHS = ~RHS
2814 AddToWorkList(LHS.getNode()); AddToWorkList(RHS.getNode());
2815 return DAG.getNode(NewOpcode, N->getDebugLoc(), VT, LHS, RHS);
2818 // fold (xor (xor x, c1), c2) -> (xor x, (xor c1, c2))
2819 if (N1C && N0.getOpcode() == ISD::XOR) {
2820 ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0));
2821 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
2823 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N0.getOperand(1),
2824 DAG.getConstant(N1C->getAPIntValue() ^
2825 N00C->getAPIntValue(), VT));
2827 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N0.getOperand(0),
2828 DAG.getConstant(N1C->getAPIntValue() ^
2829 N01C->getAPIntValue(), VT));
2831 // fold (xor x, x) -> 0
2833 if (!VT.isVector()) {
2834 return DAG.getConstant(0, VT);
2835 } else if (!LegalOperations || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT)){
2836 // Produce a vector of zeros.
2837 SDValue El = DAG.getConstant(0, VT.getVectorElementType());
2838 std::vector<SDValue> Ops(VT.getVectorNumElements(), El);
2839 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), VT,
2840 &Ops[0], Ops.size());
2844 // Simplify: xor (op x...), (op y...) -> (op (xor x, y))
2845 if (N0.getOpcode() == N1.getOpcode()) {
2846 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
2847 if (Tmp.getNode()) return Tmp;
2850 // Simplify the expression using non-local knowledge.
2851 if (!VT.isVector() &&
2852 SimplifyDemandedBits(SDValue(N, 0)))
2853 return SDValue(N, 0);
2858 /// visitShiftByConstant - Handle transforms common to the three shifts, when
2859 /// the shift amount is a constant.
2860 SDValue DAGCombiner::visitShiftByConstant(SDNode *N, unsigned Amt) {
2861 SDNode *LHS = N->getOperand(0).getNode();
2862 if (!LHS->hasOneUse()) return SDValue();
2864 // We want to pull some binops through shifts, so that we have (and (shift))
2865 // instead of (shift (and)), likewise for add, or, xor, etc. This sort of
2866 // thing happens with address calculations, so it's important to canonicalize
2868 bool HighBitSet = false; // Can we transform this if the high bit is set?
2870 switch (LHS->getOpcode()) {
2871 default: return SDValue();
2874 HighBitSet = false; // We can only transform sra if the high bit is clear.
2877 HighBitSet = true; // We can only transform sra if the high bit is set.
2880 if (N->getOpcode() != ISD::SHL)
2881 return SDValue(); // only shl(add) not sr[al](add).
2882 HighBitSet = false; // We can only transform sra if the high bit is clear.
2886 // We require the RHS of the binop to be a constant as well.
2887 ConstantSDNode *BinOpCst = dyn_cast<ConstantSDNode>(LHS->getOperand(1));
2888 if (!BinOpCst) return SDValue();
2890 // FIXME: disable this unless the input to the binop is a shift by a constant.
2891 // If it is not a shift, it pessimizes some common cases like:
2893 // void foo(int *X, int i) { X[i & 1235] = 1; }
2894 // int bar(int *X, int i) { return X[i & 255]; }
2895 SDNode *BinOpLHSVal = LHS->getOperand(0).getNode();
2896 if ((BinOpLHSVal->getOpcode() != ISD::SHL &&
2897 BinOpLHSVal->getOpcode() != ISD::SRA &&
2898 BinOpLHSVal->getOpcode() != ISD::SRL) ||
2899 !isa<ConstantSDNode>(BinOpLHSVal->getOperand(1)))
2902 EVT VT = N->getValueType(0);
2904 // If this is a signed shift right, and the high bit is modified by the
2905 // logical operation, do not perform the transformation. The highBitSet
2906 // boolean indicates the value of the high bit of the constant which would
2907 // cause it to be modified for this operation.
2908 if (N->getOpcode() == ISD::SRA) {
2909 bool BinOpRHSSignSet = BinOpCst->getAPIntValue().isNegative();
2910 if (BinOpRHSSignSet != HighBitSet)
2914 // Fold the constants, shifting the binop RHS by the shift amount.
2915 SDValue NewRHS = DAG.getNode(N->getOpcode(), LHS->getOperand(1).getDebugLoc(),
2917 LHS->getOperand(1), N->getOperand(1));
2919 // Create the new shift.
2920 SDValue NewShift = DAG.getNode(N->getOpcode(),
2921 LHS->getOperand(0).getDebugLoc(),
2922 VT, LHS->getOperand(0), N->getOperand(1));
2924 // Create the new binop.
2925 return DAG.getNode(LHS->getOpcode(), N->getDebugLoc(), VT, NewShift, NewRHS);
2928 SDValue DAGCombiner::visitSHL(SDNode *N) {
2929 SDValue N0 = N->getOperand(0);
2930 SDValue N1 = N->getOperand(1);
2931 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2932 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2933 EVT VT = N0.getValueType();
2934 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
2936 // fold (shl c1, c2) -> c1<<c2
2938 return DAG.FoldConstantArithmetic(ISD::SHL, VT, N0C, N1C);
2939 // fold (shl 0, x) -> 0
2940 if (N0C && N0C->isNullValue())
2942 // fold (shl x, c >= size(x)) -> undef
2943 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
2944 return DAG.getUNDEF(VT);
2945 // fold (shl x, 0) -> x
2946 if (N1C && N1C->isNullValue())
2948 // if (shl x, c) is known to be zero, return 0
2949 if (DAG.MaskedValueIsZero(SDValue(N, 0),
2950 APInt::getAllOnesValue(OpSizeInBits)))
2951 return DAG.getConstant(0, VT);
2952 // fold (shl x, (trunc (and y, c))) -> (shl x, (and (trunc y), (trunc c))).
2953 if (N1.getOpcode() == ISD::TRUNCATE &&
2954 N1.getOperand(0).getOpcode() == ISD::AND &&
2955 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
2956 SDValue N101 = N1.getOperand(0).getOperand(1);
2957 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
2958 EVT TruncVT = N1.getValueType();
2959 SDValue N100 = N1.getOperand(0).getOperand(0);
2960 APInt TruncC = N101C->getAPIntValue();
2961 TruncC = TruncC.trunc(TruncVT.getSizeInBits());
2962 return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0,
2963 DAG.getNode(ISD::AND, N->getDebugLoc(), TruncVT,
2964 DAG.getNode(ISD::TRUNCATE,
2967 DAG.getConstant(TruncC, TruncVT)));
2971 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
2972 return SDValue(N, 0);
2974 // fold (shl (shl x, c1), c2) -> 0 or (shl x, (add c1, c2))
2975 if (N1C && N0.getOpcode() == ISD::SHL &&
2976 N0.getOperand(1).getOpcode() == ISD::Constant) {
2977 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
2978 uint64_t c2 = N1C->getZExtValue();
2979 if (c1 + c2 >= OpSizeInBits)
2980 return DAG.getConstant(0, VT);
2981 return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0.getOperand(0),
2982 DAG.getConstant(c1 + c2, N1.getValueType()));
2985 // fold (shl (ext (shl x, c1)), c2) -> (ext (shl x, (add c1, c2)))
2986 // For this to be valid, the second form must not preserve any of the bits
2987 // that are shifted out by the inner shift in the first form. This means
2988 // the outer shift size must be >= the number of bits added by the ext.
2989 // As a corollary, we don't care what kind of ext it is.
2990 if (N1C && (N0.getOpcode() == ISD::ZERO_EXTEND ||
2991 N0.getOpcode() == ISD::ANY_EXTEND ||
2992 N0.getOpcode() == ISD::SIGN_EXTEND) &&
2993 N0.getOperand(0).getOpcode() == ISD::SHL &&
2994 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
2996 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
2997 uint64_t c2 = N1C->getZExtValue();
2998 EVT InnerShiftVT = N0.getOperand(0).getValueType();
2999 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
3000 if (c2 >= OpSizeInBits - InnerShiftSize) {
3001 if (c1 + c2 >= OpSizeInBits)
3002 return DAG.getConstant(0, VT);
3003 return DAG.getNode(ISD::SHL, N0->getDebugLoc(), VT,
3004 DAG.getNode(N0.getOpcode(), N0->getDebugLoc(), VT,
3005 N0.getOperand(0)->getOperand(0)),
3006 DAG.getConstant(c1 + c2, N1.getValueType()));
3010 // fold (shl (srl x, c1), c2) -> (shl (and x, (shl -1, c1)), (sub c2, c1)) or
3011 // (srl (and x, (shl -1, c1)), (sub c1, c2))
3012 if (N1C && N0.getOpcode() == ISD::SRL &&
3013 N0.getOperand(1).getOpcode() == ISD::Constant) {
3014 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
3015 if (c1 < VT.getSizeInBits()) {
3016 uint64_t c2 = N1C->getZExtValue();
3017 SDValue HiBitsMask =
3018 DAG.getConstant(APInt::getHighBitsSet(VT.getSizeInBits(),
3019 VT.getSizeInBits() - c1),
3021 SDValue Mask = DAG.getNode(ISD::AND, N0.getDebugLoc(), VT,
3025 return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, Mask,
3026 DAG.getConstant(c2-c1, N1.getValueType()));
3028 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, Mask,
3029 DAG.getConstant(c1-c2, N1.getValueType()));
3032 // fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1))
3033 if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1)) {
3034 SDValue HiBitsMask =
3035 DAG.getConstant(APInt::getHighBitsSet(VT.getSizeInBits(),
3036 VT.getSizeInBits() -
3037 N1C->getZExtValue()),
3039 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0.getOperand(0),
3044 SDValue NewSHL = visitShiftByConstant(N, N1C->getZExtValue());
3045 if (NewSHL.getNode())
3052 SDValue DAGCombiner::visitSRA(SDNode *N) {
3053 SDValue N0 = N->getOperand(0);
3054 SDValue N1 = N->getOperand(1);
3055 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3056 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3057 EVT VT = N0.getValueType();
3058 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
3060 // fold (sra c1, c2) -> (sra c1, c2)
3062 return DAG.FoldConstantArithmetic(ISD::SRA, VT, N0C, N1C);
3063 // fold (sra 0, x) -> 0
3064 if (N0C && N0C->isNullValue())
3066 // fold (sra -1, x) -> -1
3067 if (N0C && N0C->isAllOnesValue())
3069 // fold (sra x, (setge c, size(x))) -> undef
3070 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
3071 return DAG.getUNDEF(VT);
3072 // fold (sra x, 0) -> x
3073 if (N1C && N1C->isNullValue())
3075 // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports
3077 if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) {
3078 unsigned LowBits = OpSizeInBits - (unsigned)N1C->getZExtValue();
3079 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), LowBits);
3081 ExtVT = EVT::getVectorVT(*DAG.getContext(),
3082 ExtVT, VT.getVectorNumElements());
3083 if ((!LegalOperations ||
3084 TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, ExtVT)))
3085 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT,
3086 N0.getOperand(0), DAG.getValueType(ExtVT));
3089 // fold (sra (sra x, c1), c2) -> (sra x, (add c1, c2))
3090 if (N1C && N0.getOpcode() == ISD::SRA) {
3091 if (ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
3092 unsigned Sum = N1C->getZExtValue() + C1->getZExtValue();
3093 if (Sum >= OpSizeInBits) Sum = OpSizeInBits-1;
3094 return DAG.getNode(ISD::SRA, N->getDebugLoc(), VT, N0.getOperand(0),
3095 DAG.getConstant(Sum, N1C->getValueType(0)));
3099 // fold (sra (shl X, m), (sub result_size, n))
3100 // -> (sign_extend (trunc (shl X, (sub (sub result_size, n), m)))) for
3101 // result_size - n != m.
3102 // If truncate is free for the target sext(shl) is likely to result in better
3104 if (N0.getOpcode() == ISD::SHL) {
3105 // Get the two constanst of the shifts, CN0 = m, CN = n.
3106 const ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3108 // Determine what the truncate's result bitsize and type would be.
3110 EVT::getIntegerVT(*DAG.getContext(),
3111 OpSizeInBits - N1C->getZExtValue());
3112 // Determine the residual right-shift amount.
3113 signed ShiftAmt = N1C->getZExtValue() - N01C->getZExtValue();
3115 // If the shift is not a no-op (in which case this should be just a sign
3116 // extend already), the truncated to type is legal, sign_extend is legal
3117 // on that type, and the truncate to that type is both legal and free,
3118 // perform the transform.
3119 if ((ShiftAmt > 0) &&
3120 TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND, TruncVT) &&
3121 TLI.isOperationLegalOrCustom(ISD::TRUNCATE, VT) &&
3122 TLI.isTruncateFree(VT, TruncVT)) {
3124 SDValue Amt = DAG.getConstant(ShiftAmt, getShiftAmountTy());
3125 SDValue Shift = DAG.getNode(ISD::SRL, N0.getDebugLoc(), VT,
3126 N0.getOperand(0), Amt);
3127 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(), TruncVT,
3129 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(),
3130 N->getValueType(0), Trunc);
3135 // fold (sra x, (trunc (and y, c))) -> (sra x, (and (trunc y), (trunc c))).
3136 if (N1.getOpcode() == ISD::TRUNCATE &&
3137 N1.getOperand(0).getOpcode() == ISD::AND &&
3138 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
3139 SDValue N101 = N1.getOperand(0).getOperand(1);
3140 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
3141 EVT TruncVT = N1.getValueType();
3142 SDValue N100 = N1.getOperand(0).getOperand(0);
3143 APInt TruncC = N101C->getAPIntValue();
3144 TruncC = TruncC.trunc(TruncVT.getScalarType().getSizeInBits());
3145 return DAG.getNode(ISD::SRA, N->getDebugLoc(), VT, N0,
3146 DAG.getNode(ISD::AND, N->getDebugLoc(),
3148 DAG.getNode(ISD::TRUNCATE,
3151 DAG.getConstant(TruncC, TruncVT)));
3155 // Simplify, based on bits shifted out of the LHS.
3156 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
3157 return SDValue(N, 0);
3160 // If the sign bit is known to be zero, switch this to a SRL.
3161 if (DAG.SignBitIsZero(N0))
3162 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0, N1);
3165 SDValue NewSRA = visitShiftByConstant(N, N1C->getZExtValue());
3166 if (NewSRA.getNode())
3173 SDValue DAGCombiner::visitSRL(SDNode *N) {
3174 SDValue N0 = N->getOperand(0);
3175 SDValue N1 = N->getOperand(1);
3176 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3177 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3178 EVT VT = N0.getValueType();
3179 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
3181 // fold (srl c1, c2) -> c1 >>u c2
3183 return DAG.FoldConstantArithmetic(ISD::SRL, VT, N0C, N1C);
3184 // fold (srl 0, x) -> 0
3185 if (N0C && N0C->isNullValue())
3187 // fold (srl x, c >= size(x)) -> undef
3188 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
3189 return DAG.getUNDEF(VT);
3190 // fold (srl x, 0) -> x
3191 if (N1C && N1C->isNullValue())
3193 // if (srl x, c) is known to be zero, return 0
3194 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
3195 APInt::getAllOnesValue(OpSizeInBits)))
3196 return DAG.getConstant(0, VT);
3198 // fold (srl (srl x, c1), c2) -> 0 or (srl x, (add c1, c2))
3199 if (N1C && N0.getOpcode() == ISD::SRL &&
3200 N0.getOperand(1).getOpcode() == ISD::Constant) {
3201 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
3202 uint64_t c2 = N1C->getZExtValue();
3203 if (c1 + c2 >= OpSizeInBits)
3204 return DAG.getConstant(0, VT);
3205 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0.getOperand(0),
3206 DAG.getConstant(c1 + c2, N1.getValueType()));
3209 // fold (srl (trunc (srl x, c1)), c2) -> 0 or (trunc (srl x, (add c1, c2)))
3210 if (N1C && N0.getOpcode() == ISD::TRUNCATE &&
3211 N0.getOperand(0).getOpcode() == ISD::SRL &&
3212 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
3214 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
3215 uint64_t c2 = N1C->getZExtValue();
3216 EVT InnerShiftVT = N0.getOperand(0).getValueType();
3217 EVT ShiftCountVT = N0.getOperand(0)->getOperand(1).getValueType();
3218 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
3219 // This is only valid if the OpSizeInBits + c1 = size of inner shift.
3220 if (c1 + OpSizeInBits == InnerShiftSize) {
3221 if (c1 + c2 >= InnerShiftSize)
3222 return DAG.getConstant(0, VT);
3223 return DAG.getNode(ISD::TRUNCATE, N0->getDebugLoc(), VT,
3224 DAG.getNode(ISD::SRL, N0->getDebugLoc(), InnerShiftVT,
3225 N0.getOperand(0)->getOperand(0),
3226 DAG.getConstant(c1 + c2, ShiftCountVT)));
3230 // fold (srl (shl x, c), c) -> (and x, cst2)
3231 if (N1C && N0.getOpcode() == ISD::SHL && N0.getOperand(1) == N1 &&
3232 N0.getValueSizeInBits() <= 64) {
3233 uint64_t ShAmt = N1C->getZExtValue()+64-N0.getValueSizeInBits();
3234 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0.getOperand(0),
3235 DAG.getConstant(~0ULL >> ShAmt, VT));
3239 // fold (srl (anyextend x), c) -> (anyextend (srl x, c))
3240 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
3241 // Shifting in all undef bits?
3242 EVT SmallVT = N0.getOperand(0).getValueType();
3243 if (N1C->getZExtValue() >= SmallVT.getSizeInBits())
3244 return DAG.getUNDEF(VT);
3246 if (!LegalTypes || TLI.isTypeDesirableForOp(ISD::SRL, SmallVT)) {
3247 SDValue SmallShift = DAG.getNode(ISD::SRL, N0.getDebugLoc(), SmallVT,
3248 N0.getOperand(0), N1);
3249 AddToWorkList(SmallShift.getNode());
3250 return DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, SmallShift);
3254 // fold (srl (sra X, Y), 31) -> (srl X, 31). This srl only looks at the sign
3255 // bit, which is unmodified by sra.
3256 if (N1C && N1C->getZExtValue() + 1 == VT.getSizeInBits()) {
3257 if (N0.getOpcode() == ISD::SRA)
3258 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0.getOperand(0), N1);
3261 // fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit).
3262 if (N1C && N0.getOpcode() == ISD::CTLZ &&
3263 N1C->getAPIntValue() == Log2_32(VT.getSizeInBits())) {
3264 APInt KnownZero, KnownOne;
3265 APInt Mask = APInt::getAllOnesValue(VT.getScalarType().getSizeInBits());
3266 DAG.ComputeMaskedBits(N0.getOperand(0), Mask, KnownZero, KnownOne);
3268 // If any of the input bits are KnownOne, then the input couldn't be all
3269 // zeros, thus the result of the srl will always be zero.
3270 if (KnownOne.getBoolValue()) return DAG.getConstant(0, VT);
3272 // If all of the bits input the to ctlz node are known to be zero, then
3273 // the result of the ctlz is "32" and the result of the shift is one.
3274 APInt UnknownBits = ~KnownZero & Mask;
3275 if (UnknownBits == 0) return DAG.getConstant(1, VT);
3277 // Otherwise, check to see if there is exactly one bit input to the ctlz.
3278 if ((UnknownBits & (UnknownBits - 1)) == 0) {
3279 // Okay, we know that only that the single bit specified by UnknownBits
3280 // could be set on input to the CTLZ node. If this bit is set, the SRL
3281 // will return 0, if it is clear, it returns 1. Change the CTLZ/SRL pair
3282 // to an SRL/XOR pair, which is likely to simplify more.
3283 unsigned ShAmt = UnknownBits.countTrailingZeros();
3284 SDValue Op = N0.getOperand(0);
3287 Op = DAG.getNode(ISD::SRL, N0.getDebugLoc(), VT, Op,
3288 DAG.getConstant(ShAmt, getShiftAmountTy()));
3289 AddToWorkList(Op.getNode());
3292 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT,
3293 Op, DAG.getConstant(1, VT));
3297 // fold (srl x, (trunc (and y, c))) -> (srl x, (and (trunc y), (trunc c))).
3298 if (N1.getOpcode() == ISD::TRUNCATE &&
3299 N1.getOperand(0).getOpcode() == ISD::AND &&
3300 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
3301 SDValue N101 = N1.getOperand(0).getOperand(1);
3302 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
3303 EVT TruncVT = N1.getValueType();
3304 SDValue N100 = N1.getOperand(0).getOperand(0);
3305 APInt TruncC = N101C->getAPIntValue();
3306 TruncC = TruncC.trunc(TruncVT.getSizeInBits());
3307 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0,
3308 DAG.getNode(ISD::AND, N->getDebugLoc(),
3310 DAG.getNode(ISD::TRUNCATE,
3313 DAG.getConstant(TruncC, TruncVT)));
3317 // fold operands of srl based on knowledge that the low bits are not
3319 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
3320 return SDValue(N, 0);
3323 SDValue NewSRL = visitShiftByConstant(N, N1C->getZExtValue());
3324 if (NewSRL.getNode())
3328 // Attempt to convert a srl of a load into a narrower zero-extending load.
3329 SDValue NarrowLoad = ReduceLoadWidth(N);
3330 if (NarrowLoad.getNode())
3333 // Here is a common situation. We want to optimize:
3336 // %b = and i32 %a, 2
3337 // %c = srl i32 %b, 1
3338 // brcond i32 %c ...
3344 // %c = setcc eq %b, 0
3347 // However when after the source operand of SRL is optimized into AND, the SRL
3348 // itself may not be optimized further. Look for it and add the BRCOND into
3350 if (N->hasOneUse()) {
3351 SDNode *Use = *N->use_begin();
3352 if (Use->getOpcode() == ISD::BRCOND)
3354 else if (Use->getOpcode() == ISD::TRUNCATE && Use->hasOneUse()) {
3355 // Also look pass the truncate.
3356 Use = *Use->use_begin();
3357 if (Use->getOpcode() == ISD::BRCOND)
3365 SDValue DAGCombiner::visitCTLZ(SDNode *N) {
3366 SDValue N0 = N->getOperand(0);
3367 EVT VT = N->getValueType(0);
3369 // fold (ctlz c1) -> c2
3370 if (isa<ConstantSDNode>(N0))
3371 return DAG.getNode(ISD::CTLZ, N->getDebugLoc(), VT, N0);
3375 SDValue DAGCombiner::visitCTTZ(SDNode *N) {
3376 SDValue N0 = N->getOperand(0);
3377 EVT VT = N->getValueType(0);
3379 // fold (cttz c1) -> c2
3380 if (isa<ConstantSDNode>(N0))
3381 return DAG.getNode(ISD::CTTZ, N->getDebugLoc(), VT, N0);
3385 SDValue DAGCombiner::visitCTPOP(SDNode *N) {
3386 SDValue N0 = N->getOperand(0);
3387 EVT VT = N->getValueType(0);
3389 // fold (ctpop c1) -> c2
3390 if (isa<ConstantSDNode>(N0))
3391 return DAG.getNode(ISD::CTPOP, N->getDebugLoc(), VT, N0);
3395 SDValue DAGCombiner::visitSELECT(SDNode *N) {
3396 SDValue N0 = N->getOperand(0);
3397 SDValue N1 = N->getOperand(1);
3398 SDValue N2 = N->getOperand(2);
3399 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3400 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3401 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
3402 EVT VT = N->getValueType(0);
3403 EVT VT0 = N0.getValueType();
3405 // fold (select C, X, X) -> X
3408 // fold (select true, X, Y) -> X
3409 if (N0C && !N0C->isNullValue())
3411 // fold (select false, X, Y) -> Y
3412 if (N0C && N0C->isNullValue())
3414 // fold (select C, 1, X) -> (or C, X)
3415 if (VT == MVT::i1 && N1C && N1C->getAPIntValue() == 1)
3416 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N0, N2);
3417 // fold (select C, 0, 1) -> (xor C, 1)
3418 if (VT.isInteger() &&
3421 TLI.getBooleanContents() == TargetLowering::ZeroOrOneBooleanContent)) &&
3422 N1C && N2C && N1C->isNullValue() && N2C->getAPIntValue() == 1) {
3425 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT0,
3426 N0, DAG.getConstant(1, VT0));
3427 XORNode = DAG.getNode(ISD::XOR, N0.getDebugLoc(), VT0,
3428 N0, DAG.getConstant(1, VT0));
3429 AddToWorkList(XORNode.getNode());
3431 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, XORNode);
3432 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, XORNode);
3434 // fold (select C, 0, X) -> (and (not C), X)
3435 if (VT == VT0 && VT == MVT::i1 && N1C && N1C->isNullValue()) {
3436 SDValue NOTNode = DAG.getNOT(N0.getDebugLoc(), N0, VT);
3437 AddToWorkList(NOTNode.getNode());
3438 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, NOTNode, N2);
3440 // fold (select C, X, 1) -> (or (not C), X)
3441 if (VT == VT0 && VT == MVT::i1 && N2C && N2C->getAPIntValue() == 1) {
3442 SDValue NOTNode = DAG.getNOT(N0.getDebugLoc(), N0, VT);
3443 AddToWorkList(NOTNode.getNode());
3444 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, NOTNode, N1);
3446 // fold (select C, X, 0) -> (and C, X)
3447 if (VT == MVT::i1 && N2C && N2C->isNullValue())
3448 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0, N1);
3449 // fold (select X, X, Y) -> (or X, Y)
3450 // fold (select X, 1, Y) -> (or X, Y)
3451 if (VT == MVT::i1 && (N0 == N1 || (N1C && N1C->getAPIntValue() == 1)))
3452 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N0, N2);
3453 // fold (select X, Y, X) -> (and X, Y)
3454 // fold (select X, Y, 0) -> (and X, Y)
3455 if (VT == MVT::i1 && (N0 == N2 || (N2C && N2C->getAPIntValue() == 0)))
3456 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0, N1);
3458 // If we can fold this based on the true/false value, do so.
3459 if (SimplifySelectOps(N, N1, N2))
3460 return SDValue(N, 0); // Don't revisit N.
3462 // fold selects based on a setcc into other things, such as min/max/abs
3463 if (N0.getOpcode() == ISD::SETCC) {
3465 // Check against MVT::Other for SELECT_CC, which is a workaround for targets
3466 // having to say they don't support SELECT_CC on every type the DAG knows
3467 // about, since there is no way to mark an opcode illegal at all value types
3468 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, MVT::Other) &&
3469 TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT))
3470 return DAG.getNode(ISD::SELECT_CC, N->getDebugLoc(), VT,
3471 N0.getOperand(0), N0.getOperand(1),
3472 N1, N2, N0.getOperand(2));
3473 return SimplifySelect(N->getDebugLoc(), N0, N1, N2);
3479 SDValue DAGCombiner::visitSELECT_CC(SDNode *N) {
3480 SDValue N0 = N->getOperand(0);
3481 SDValue N1 = N->getOperand(1);
3482 SDValue N2 = N->getOperand(2);
3483 SDValue N3 = N->getOperand(3);
3484 SDValue N4 = N->getOperand(4);
3485 ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get();
3487 // fold select_cc lhs, rhs, x, x, cc -> x
3491 // Determine if the condition we're dealing with is constant
3492 SDValue SCC = SimplifySetCC(TLI.getSetCCResultType(N0.getValueType()),
3493 N0, N1, CC, N->getDebugLoc(), false);
3494 if (SCC.getNode()) AddToWorkList(SCC.getNode());
3496 if (ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode())) {
3497 if (!SCCC->isNullValue())
3498 return N2; // cond always true -> true val
3500 return N3; // cond always false -> false val
3503 // Fold to a simpler select_cc
3504 if (SCC.getNode() && SCC.getOpcode() == ISD::SETCC)
3505 return DAG.getNode(ISD::SELECT_CC, N->getDebugLoc(), N2.getValueType(),
3506 SCC.getOperand(0), SCC.getOperand(1), N2, N3,
3509 // If we can fold this based on the true/false value, do so.
3510 if (SimplifySelectOps(N, N2, N3))
3511 return SDValue(N, 0); // Don't revisit N.
3513 // fold select_cc into other things, such as min/max/abs
3514 return SimplifySelectCC(N->getDebugLoc(), N0, N1, N2, N3, CC);
3517 SDValue DAGCombiner::visitSETCC(SDNode *N) {
3518 return SimplifySetCC(N->getValueType(0), N->getOperand(0), N->getOperand(1),
3519 cast<CondCodeSDNode>(N->getOperand(2))->get(),
3523 // ExtendUsesToFormExtLoad - Trying to extend uses of a load to enable this:
3524 // "fold ({s|z|a}ext (load x)) -> ({s|z|a}ext (truncate ({s|z|a}extload x)))"
3525 // transformation. Returns true if extension are possible and the above
3526 // mentioned transformation is profitable.
3527 static bool ExtendUsesToFormExtLoad(SDNode *N, SDValue N0,
3529 SmallVector<SDNode*, 4> &ExtendNodes,
3530 const TargetLowering &TLI) {
3531 bool HasCopyToRegUses = false;
3532 bool isTruncFree = TLI.isTruncateFree(N->getValueType(0), N0.getValueType());
3533 for (SDNode::use_iterator UI = N0.getNode()->use_begin(),
3534 UE = N0.getNode()->use_end();
3539 if (UI.getUse().getResNo() != N0.getResNo())
3541 // FIXME: Only extend SETCC N, N and SETCC N, c for now.
3542 if (ExtOpc != ISD::ANY_EXTEND && User->getOpcode() == ISD::SETCC) {
3543 ISD::CondCode CC = cast<CondCodeSDNode>(User->getOperand(2))->get();
3544 if (ExtOpc == ISD::ZERO_EXTEND && ISD::isSignedIntSetCC(CC))
3545 // Sign bits will be lost after a zext.
3548 for (unsigned i = 0; i != 2; ++i) {
3549 SDValue UseOp = User->getOperand(i);
3552 if (!isa<ConstantSDNode>(UseOp))
3557 ExtendNodes.push_back(User);
3560 // If truncates aren't free and there are users we can't
3561 // extend, it isn't worthwhile.
3564 // Remember if this value is live-out.
3565 if (User->getOpcode() == ISD::CopyToReg)
3566 HasCopyToRegUses = true;
3569 if (HasCopyToRegUses) {
3570 bool BothLiveOut = false;
3571 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
3573 SDUse &Use = UI.getUse();
3574 if (Use.getResNo() == 0 && Use.getUser()->getOpcode() == ISD::CopyToReg) {
3580 // Both unextended and extended values are live out. There had better be
3581 // a good reason for the transformation.
3582 return ExtendNodes.size();
3587 SDValue DAGCombiner::visitSIGN_EXTEND(SDNode *N) {
3588 SDValue N0 = N->getOperand(0);
3589 EVT VT = N->getValueType(0);
3591 // fold (sext c1) -> c1
3592 if (isa<ConstantSDNode>(N0))
3593 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, N0);
3595 // fold (sext (sext x)) -> (sext x)
3596 // fold (sext (aext x)) -> (sext x)
3597 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
3598 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT,
3601 if (N0.getOpcode() == ISD::TRUNCATE) {
3602 // fold (sext (truncate (load x))) -> (sext (smaller load x))
3603 // fold (sext (truncate (srl (load x), c))) -> (sext (smaller load (x+c/n)))
3604 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
3605 if (NarrowLoad.getNode()) {
3606 SDNode* oye = N0.getNode()->getOperand(0).getNode();
3607 if (NarrowLoad.getNode() != N0.getNode()) {
3608 CombineTo(N0.getNode(), NarrowLoad);
3609 // CombineTo deleted the truncate, if needed, but not what's under it.
3612 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3615 // See if the value being truncated is already sign extended. If so, just
3616 // eliminate the trunc/sext pair.
3617 SDValue Op = N0.getOperand(0);
3618 unsigned OpBits = Op.getValueType().getScalarType().getSizeInBits();
3619 unsigned MidBits = N0.getValueType().getScalarType().getSizeInBits();
3620 unsigned DestBits = VT.getScalarType().getSizeInBits();
3621 unsigned NumSignBits = DAG.ComputeNumSignBits(Op);
3623 if (OpBits == DestBits) {
3624 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
3625 // bits, it is already ready.
3626 if (NumSignBits > DestBits-MidBits)
3628 } else if (OpBits < DestBits) {
3629 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
3630 // bits, just sext from i32.
3631 if (NumSignBits > OpBits-MidBits)
3632 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, Op);
3634 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
3635 // bits, just truncate to i32.
3636 if (NumSignBits > OpBits-MidBits)
3637 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, Op);
3640 // fold (sext (truncate x)) -> (sextinreg x).
3641 if (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG,
3642 N0.getValueType())) {
3643 if (OpBits < DestBits)
3644 Op = DAG.getNode(ISD::ANY_EXTEND, N0.getDebugLoc(), VT, Op);
3645 else if (OpBits > DestBits)
3646 Op = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(), VT, Op);
3647 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT, Op,
3648 DAG.getValueType(N0.getValueType()));
3652 // fold (sext (load x)) -> (sext (truncate (sextload x)))
3653 if (ISD::isNON_EXTLoad(N0.getNode()) &&
3654 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
3655 TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()))) {
3656 bool DoXform = true;
3657 SmallVector<SDNode*, 4> SetCCs;
3658 if (!N0.hasOneUse())
3659 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::SIGN_EXTEND, SetCCs, TLI);
3661 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
3662 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, N->getDebugLoc(),
3664 LN0->getBasePtr(), LN0->getPointerInfo(),
3666 LN0->isVolatile(), LN0->isNonTemporal(),
3667 LN0->getAlignment());
3668 CombineTo(N, ExtLoad);
3669 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
3670 N0.getValueType(), ExtLoad);
3671 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
3673 // Extend SetCC uses if necessary.
3674 for (unsigned i = 0, e = SetCCs.size(); i != e; ++i) {
3675 SDNode *SetCC = SetCCs[i];
3676 SmallVector<SDValue, 4> Ops;
3678 for (unsigned j = 0; j != 2; ++j) {
3679 SDValue SOp = SetCC->getOperand(j);
3681 Ops.push_back(ExtLoad);
3683 Ops.push_back(DAG.getNode(ISD::SIGN_EXTEND,
3684 N->getDebugLoc(), VT, SOp));
3687 Ops.push_back(SetCC->getOperand(2));
3688 CombineTo(SetCC, DAG.getNode(ISD::SETCC, N->getDebugLoc(),
3689 SetCC->getValueType(0),
3690 &Ops[0], Ops.size()));
3693 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3697 // fold (sext (sextload x)) -> (sext (truncate (sextload x)))
3698 // fold (sext ( extload x)) -> (sext (truncate (sextload x)))
3699 if ((ISD::isSEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
3700 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
3701 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
3702 EVT MemVT = LN0->getMemoryVT();
3703 if ((!LegalOperations && !LN0->isVolatile()) ||
3704 TLI.isLoadExtLegal(ISD::SEXTLOAD, MemVT)) {
3705 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, N->getDebugLoc(),
3707 LN0->getBasePtr(), LN0->getPointerInfo(),
3709 LN0->isVolatile(), LN0->isNonTemporal(),
3710 LN0->getAlignment());
3711 CombineTo(N, ExtLoad);
3712 CombineTo(N0.getNode(),
3713 DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
3714 N0.getValueType(), ExtLoad),
3715 ExtLoad.getValue(1));
3716 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3720 if (N0.getOpcode() == ISD::SETCC) {
3721 // sext(setcc) -> sext_in_reg(vsetcc) for vectors.
3722 // Only do this before legalize for now.
3723 if (VT.isVector() && !LegalOperations) {
3724 EVT N0VT = N0.getOperand(0).getValueType();
3725 // We know that the # elements of the results is the same as the
3726 // # elements of the compare (and the # elements of the compare result
3727 // for that matter). Check to see that they are the same size. If so,
3728 // we know that the element size of the sext'd result matches the
3729 // element size of the compare operands.
3730 if (VT.getSizeInBits() == N0VT.getSizeInBits())
3731 return DAG.getVSetCC(N->getDebugLoc(), VT, N0.getOperand(0),
3733 cast<CondCodeSDNode>(N0.getOperand(2))->get());
3734 // If the desired elements are smaller or larger than the source
3735 // elements we can use a matching integer vector type and then
3736 // truncate/sign extend
3738 EVT MatchingElementType =
3739 EVT::getIntegerVT(*DAG.getContext(),
3740 N0VT.getScalarType().getSizeInBits());
3741 EVT MatchingVectorType =
3742 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
3743 N0VT.getVectorNumElements());
3745 DAG.getVSetCC(N->getDebugLoc(), MatchingVectorType, N0.getOperand(0),
3747 cast<CondCodeSDNode>(N0.getOperand(2))->get());
3748 return DAG.getSExtOrTrunc(VsetCC, N->getDebugLoc(), VT);
3752 // sext(setcc x, y, cc) -> (select_cc x, y, -1, 0, cc)
3753 unsigned ElementWidth = VT.getScalarType().getSizeInBits();
3755 DAG.getConstant(APInt::getAllOnesValue(ElementWidth), VT);
3757 SimplifySelectCC(N->getDebugLoc(), N0.getOperand(0), N0.getOperand(1),
3758 NegOne, DAG.getConstant(0, VT),
3759 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
3760 if (SCC.getNode()) return SCC;
3761 if (!LegalOperations ||
3762 TLI.isOperationLegal(ISD::SETCC, TLI.getSetCCResultType(VT)))
3763 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
3764 DAG.getSetCC(N->getDebugLoc(),
3765 TLI.getSetCCResultType(VT),
3766 N0.getOperand(0), N0.getOperand(1),
3767 cast<CondCodeSDNode>(N0.getOperand(2))->get()),
3768 NegOne, DAG.getConstant(0, VT));
3771 // fold (sext x) -> (zext x) if the sign bit is known zero.
3772 if ((!LegalOperations || TLI.isOperationLegal(ISD::ZERO_EXTEND, VT)) &&
3773 DAG.SignBitIsZero(N0))
3774 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, N0);
3779 SDValue DAGCombiner::visitZERO_EXTEND(SDNode *N) {
3780 SDValue N0 = N->getOperand(0);
3781 EVT VT = N->getValueType(0);
3783 // fold (zext c1) -> c1
3784 if (isa<ConstantSDNode>(N0))
3785 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, N0);
3786 // fold (zext (zext x)) -> (zext x)
3787 // fold (zext (aext x)) -> (zext x)
3788 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
3789 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT,
3792 // fold (zext (truncate (load x))) -> (zext (smaller load x))
3793 // fold (zext (truncate (srl (load x), c))) -> (zext (small load (x+c/n)))
3794 if (N0.getOpcode() == ISD::TRUNCATE) {
3795 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
3796 if (NarrowLoad.getNode()) {
3797 SDNode* oye = N0.getNode()->getOperand(0).getNode();
3798 if (NarrowLoad.getNode() != N0.getNode()) {
3799 CombineTo(N0.getNode(), NarrowLoad);
3800 // CombineTo deleted the truncate, if needed, but not what's under it.
3803 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, NarrowLoad);
3807 // fold (zext (truncate x)) -> (and x, mask)
3808 if (N0.getOpcode() == ISD::TRUNCATE &&
3809 (!LegalOperations || TLI.isOperationLegal(ISD::AND, VT))) {
3811 // fold (zext (truncate (load x))) -> (zext (smaller load x))
3812 // fold (zext (truncate (srl (load x), c))) -> (zext (smaller load (x+c/n)))
3813 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
3814 if (NarrowLoad.getNode()) {
3815 SDNode* oye = N0.getNode()->getOperand(0).getNode();
3816 if (NarrowLoad.getNode() != N0.getNode()) {
3817 CombineTo(N0.getNode(), NarrowLoad);
3818 // CombineTo deleted the truncate, if needed, but not what's under it.
3821 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3824 SDValue Op = N0.getOperand(0);
3825 if (Op.getValueType().bitsLT(VT)) {
3826 Op = DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, Op);
3827 } else if (Op.getValueType().bitsGT(VT)) {
3828 Op = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, Op);
3830 return DAG.getZeroExtendInReg(Op, N->getDebugLoc(),
3831 N0.getValueType().getScalarType());
3834 // Fold (zext (and (trunc x), cst)) -> (and x, cst),
3835 // if either of the casts is not free.
3836 if (N0.getOpcode() == ISD::AND &&
3837 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
3838 N0.getOperand(1).getOpcode() == ISD::Constant &&
3839 (!TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
3840 N0.getValueType()) ||
3841 !TLI.isZExtFree(N0.getValueType(), VT))) {
3842 SDValue X = N0.getOperand(0).getOperand(0);
3843 if (X.getValueType().bitsLT(VT)) {
3844 X = DAG.getNode(ISD::ANY_EXTEND, X.getDebugLoc(), VT, X);
3845 } else if (X.getValueType().bitsGT(VT)) {
3846 X = DAG.getNode(ISD::TRUNCATE, X.getDebugLoc(), VT, X);
3848 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
3849 Mask = Mask.zext(VT.getSizeInBits());
3850 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
3851 X, DAG.getConstant(Mask, VT));
3854 // fold (zext (load x)) -> (zext (truncate (zextload x)))
3855 if (ISD::isNON_EXTLoad(N0.getNode()) &&
3856 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
3857 TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()))) {
3858 bool DoXform = true;
3859 SmallVector<SDNode*, 4> SetCCs;
3860 if (!N0.hasOneUse())
3861 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ZERO_EXTEND, SetCCs, TLI);
3863 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
3864 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, N->getDebugLoc(),
3866 LN0->getBasePtr(), LN0->getPointerInfo(),
3868 LN0->isVolatile(), LN0->isNonTemporal(),
3869 LN0->getAlignment());
3870 CombineTo(N, ExtLoad);
3871 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
3872 N0.getValueType(), ExtLoad);
3873 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
3875 // Extend SetCC uses if necessary.
3876 for (unsigned i = 0, e = SetCCs.size(); i != e; ++i) {
3877 SDNode *SetCC = SetCCs[i];
3878 SmallVector<SDValue, 4> Ops;
3880 for (unsigned j = 0; j != 2; ++j) {
3881 SDValue SOp = SetCC->getOperand(j);
3883 Ops.push_back(ExtLoad);
3885 Ops.push_back(DAG.getNode(ISD::ZERO_EXTEND,
3886 N->getDebugLoc(), VT, SOp));
3889 Ops.push_back(SetCC->getOperand(2));
3890 CombineTo(SetCC, DAG.getNode(ISD::SETCC, N->getDebugLoc(),
3891 SetCC->getValueType(0),
3892 &Ops[0], Ops.size()));
3895 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3899 // fold (zext (zextload x)) -> (zext (truncate (zextload x)))
3900 // fold (zext ( extload x)) -> (zext (truncate (zextload x)))
3901 if ((ISD::isZEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
3902 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
3903 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
3904 EVT MemVT = LN0->getMemoryVT();
3905 if ((!LegalOperations && !LN0->isVolatile()) ||
3906 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT)) {
3907 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, N->getDebugLoc(),
3909 LN0->getBasePtr(), LN0->getPointerInfo(),
3911 LN0->isVolatile(), LN0->isNonTemporal(),
3912 LN0->getAlignment());
3913 CombineTo(N, ExtLoad);
3914 CombineTo(N0.getNode(),
3915 DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(), N0.getValueType(),
3917 ExtLoad.getValue(1));
3918 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3922 if (N0.getOpcode() == ISD::SETCC) {
3923 if (!LegalOperations && VT.isVector()) {
3924 // zext(setcc) -> (and (vsetcc), (1, 1, ...) for vectors.
3925 // Only do this before legalize for now.
3926 EVT N0VT = N0.getOperand(0).getValueType();
3927 EVT EltVT = VT.getVectorElementType();
3928 SmallVector<SDValue,8> OneOps(VT.getVectorNumElements(),
3929 DAG.getConstant(1, EltVT));
3930 if (VT.getSizeInBits() == N0VT.getSizeInBits()) {
3931 // We know that the # elements of the results is the same as the
3932 // # elements of the compare (and the # elements of the compare result
3933 // for that matter). Check to see that they are the same size. If so,
3934 // we know that the element size of the sext'd result matches the
3935 // element size of the compare operands.
3936 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
3937 DAG.getVSetCC(N->getDebugLoc(), VT, N0.getOperand(0),
3939 cast<CondCodeSDNode>(N0.getOperand(2))->get()),
3940 DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), VT,
3941 &OneOps[0], OneOps.size()));
3943 // If the desired elements are smaller or larger than the source
3944 // elements we can use a matching integer vector type and then
3945 // truncate/sign extend
3946 EVT MatchingElementType =
3947 EVT::getIntegerVT(*DAG.getContext(),
3948 N0VT.getScalarType().getSizeInBits());
3949 EVT MatchingVectorType =
3950 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
3951 N0VT.getVectorNumElements());
3953 DAG.getVSetCC(N->getDebugLoc(), MatchingVectorType, N0.getOperand(0),
3955 cast<CondCodeSDNode>(N0.getOperand(2))->get());
3956 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
3957 DAG.getSExtOrTrunc(VsetCC, N->getDebugLoc(), VT),
3958 DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), VT,
3959 &OneOps[0], OneOps.size()));
3963 // zext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
3965 SimplifySelectCC(N->getDebugLoc(), N0.getOperand(0), N0.getOperand(1),
3966 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
3967 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
3968 if (SCC.getNode()) return SCC;
3971 // (zext (shl (zext x), cst)) -> (shl (zext x), cst)
3972 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL) &&
3973 isa<ConstantSDNode>(N0.getOperand(1)) &&
3974 N0.getOperand(0).getOpcode() == ISD::ZERO_EXTEND &&
3976 if (N0.getOpcode() == ISD::SHL) {
3977 // If the original shl may be shifting out bits, do not perform this
3979 unsigned ShAmt = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
3980 unsigned KnownZeroBits = N0.getOperand(0).getValueType().getSizeInBits() -
3981 N0.getOperand(0).getOperand(0).getValueType().getSizeInBits();
3982 if (ShAmt > KnownZeroBits)
3985 DebugLoc dl = N->getDebugLoc();
3986 return DAG.getNode(N0.getOpcode(), dl, VT,
3987 DAG.getNode(ISD::ZERO_EXTEND, dl, VT, N0.getOperand(0)),
3988 DAG.getNode(ISD::ZERO_EXTEND, dl,
3989 N0.getOperand(1).getValueType(),
3996 SDValue DAGCombiner::visitANY_EXTEND(SDNode *N) {
3997 SDValue N0 = N->getOperand(0);
3998 EVT VT = N->getValueType(0);
4000 // fold (aext c1) -> c1
4001 if (isa<ConstantSDNode>(N0))
4002 return DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, N0);
4003 // fold (aext (aext x)) -> (aext x)
4004 // fold (aext (zext x)) -> (zext x)
4005 // fold (aext (sext x)) -> (sext x)
4006 if (N0.getOpcode() == ISD::ANY_EXTEND ||
4007 N0.getOpcode() == ISD::ZERO_EXTEND ||
4008 N0.getOpcode() == ISD::SIGN_EXTEND)
4009 return DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT, N0.getOperand(0));
4011 // fold (aext (truncate (load x))) -> (aext (smaller load x))
4012 // fold (aext (truncate (srl (load x), c))) -> (aext (small load (x+c/n)))
4013 if (N0.getOpcode() == ISD::TRUNCATE) {
4014 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
4015 if (NarrowLoad.getNode()) {
4016 SDNode* oye = N0.getNode()->getOperand(0).getNode();
4017 if (NarrowLoad.getNode() != N0.getNode()) {
4018 CombineTo(N0.getNode(), NarrowLoad);
4019 // CombineTo deleted the truncate, if needed, but not what's under it.
4022 return DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, NarrowLoad);
4026 // fold (aext (truncate x))
4027 if (N0.getOpcode() == ISD::TRUNCATE) {
4028 SDValue TruncOp = N0.getOperand(0);
4029 if (TruncOp.getValueType() == VT)
4030 return TruncOp; // x iff x size == zext size.
4031 if (TruncOp.getValueType().bitsGT(VT))
4032 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, TruncOp);
4033 return DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, TruncOp);
4036 // Fold (aext (and (trunc x), cst)) -> (and x, cst)
4037 // if the trunc is not free.
4038 if (N0.getOpcode() == ISD::AND &&
4039 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
4040 N0.getOperand(1).getOpcode() == ISD::Constant &&
4041 !TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
4042 N0.getValueType())) {
4043 SDValue X = N0.getOperand(0).getOperand(0);
4044 if (X.getValueType().bitsLT(VT)) {
4045 X = DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, X);
4046 } else if (X.getValueType().bitsGT(VT)) {
4047 X = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, X);
4049 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
4050 Mask = Mask.zext(VT.getSizeInBits());
4051 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
4052 X, DAG.getConstant(Mask, VT));
4055 // fold (aext (load x)) -> (aext (truncate (extload x)))
4056 if (ISD::isNON_EXTLoad(N0.getNode()) &&
4057 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
4058 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType()))) {
4059 bool DoXform = true;
4060 SmallVector<SDNode*, 4> SetCCs;
4061 if (!N0.hasOneUse())
4062 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ANY_EXTEND, SetCCs, TLI);
4064 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4065 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, VT, N->getDebugLoc(),
4067 LN0->getBasePtr(), LN0->getPointerInfo(),
4069 LN0->isVolatile(), LN0->isNonTemporal(),
4070 LN0->getAlignment());
4071 CombineTo(N, ExtLoad);
4072 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4073 N0.getValueType(), ExtLoad);
4074 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
4076 // Extend SetCC uses if necessary.
4077 for (unsigned i = 0, e = SetCCs.size(); i != e; ++i) {
4078 SDNode *SetCC = SetCCs[i];
4079 SmallVector<SDValue, 4> Ops;
4081 for (unsigned j = 0; j != 2; ++j) {
4082 SDValue SOp = SetCC->getOperand(j);
4084 Ops.push_back(ExtLoad);
4086 Ops.push_back(DAG.getNode(ISD::ANY_EXTEND,
4087 N->getDebugLoc(), VT, SOp));
4090 Ops.push_back(SetCC->getOperand(2));
4091 CombineTo(SetCC, DAG.getNode(ISD::SETCC, N->getDebugLoc(),
4092 SetCC->getValueType(0),
4093 &Ops[0], Ops.size()));
4096 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4100 // fold (aext (zextload x)) -> (aext (truncate (zextload x)))
4101 // fold (aext (sextload x)) -> (aext (truncate (sextload x)))
4102 // fold (aext ( extload x)) -> (aext (truncate (extload x)))
4103 if (N0.getOpcode() == ISD::LOAD &&
4104 !ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
4106 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4107 EVT MemVT = LN0->getMemoryVT();
4108 SDValue ExtLoad = DAG.getExtLoad(LN0->getExtensionType(), VT,
4110 LN0->getChain(), LN0->getBasePtr(),
4111 LN0->getPointerInfo(), MemVT,
4112 LN0->isVolatile(), LN0->isNonTemporal(),
4113 LN0->getAlignment());
4114 CombineTo(N, ExtLoad);
4115 CombineTo(N0.getNode(),
4116 DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4117 N0.getValueType(), ExtLoad),
4118 ExtLoad.getValue(1));
4119 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4122 if (N0.getOpcode() == ISD::SETCC) {
4123 // aext(setcc) -> sext_in_reg(vsetcc) for vectors.
4124 // Only do this before legalize for now.
4125 if (VT.isVector() && !LegalOperations) {
4126 EVT N0VT = N0.getOperand(0).getValueType();
4127 // We know that the # elements of the results is the same as the
4128 // # elements of the compare (and the # elements of the compare result
4129 // for that matter). Check to see that they are the same size. If so,
4130 // we know that the element size of the sext'd result matches the
4131 // element size of the compare operands.
4132 if (VT.getSizeInBits() == N0VT.getSizeInBits())
4133 return DAG.getVSetCC(N->getDebugLoc(), VT, N0.getOperand(0),
4135 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4136 // If the desired elements are smaller or larger than the source
4137 // elements we can use a matching integer vector type and then
4138 // truncate/sign extend
4140 EVT MatchingElementType =
4141 EVT::getIntegerVT(*DAG.getContext(),
4142 N0VT.getScalarType().getSizeInBits());
4143 EVT MatchingVectorType =
4144 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
4145 N0VT.getVectorNumElements());
4147 DAG.getVSetCC(N->getDebugLoc(), MatchingVectorType, N0.getOperand(0),
4149 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4150 return DAG.getSExtOrTrunc(VsetCC, N->getDebugLoc(), VT);
4154 // aext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
4156 SimplifySelectCC(N->getDebugLoc(), N0.getOperand(0), N0.getOperand(1),
4157 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
4158 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
4166 /// GetDemandedBits - See if the specified operand can be simplified with the
4167 /// knowledge that only the bits specified by Mask are used. If so, return the
4168 /// simpler operand, otherwise return a null SDValue.
4169 SDValue DAGCombiner::GetDemandedBits(SDValue V, const APInt &Mask) {
4170 switch (V.getOpcode()) {
4174 // If the LHS or RHS don't contribute bits to the or, drop them.
4175 if (DAG.MaskedValueIsZero(V.getOperand(0), Mask))
4176 return V.getOperand(1);
4177 if (DAG.MaskedValueIsZero(V.getOperand(1), Mask))
4178 return V.getOperand(0);
4181 // Only look at single-use SRLs.
4182 if (!V.getNode()->hasOneUse())
4184 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(V.getOperand(1))) {
4185 // See if we can recursively simplify the LHS.
4186 unsigned Amt = RHSC->getZExtValue();
4188 // Watch out for shift count overflow though.
4189 if (Amt >= Mask.getBitWidth()) break;
4190 APInt NewMask = Mask << Amt;
4191 SDValue SimplifyLHS = GetDemandedBits(V.getOperand(0), NewMask);
4192 if (SimplifyLHS.getNode())
4193 return DAG.getNode(ISD::SRL, V.getDebugLoc(), V.getValueType(),
4194 SimplifyLHS, V.getOperand(1));
4200 /// ReduceLoadWidth - If the result of a wider load is shifted to right of N
4201 /// bits and then truncated to a narrower type and where N is a multiple
4202 /// of number of bits of the narrower type, transform it to a narrower load
4203 /// from address + N / num of bits of new type. If the result is to be
4204 /// extended, also fold the extension to form a extending load.
4205 SDValue DAGCombiner::ReduceLoadWidth(SDNode *N) {
4206 unsigned Opc = N->getOpcode();
4208 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
4209 SDValue N0 = N->getOperand(0);
4210 EVT VT = N->getValueType(0);
4213 // This transformation isn't valid for vector loads.
4217 // Special case: SIGN_EXTEND_INREG is basically truncating to ExtVT then
4219 if (Opc == ISD::SIGN_EXTEND_INREG) {
4220 ExtType = ISD::SEXTLOAD;
4221 ExtVT = cast<VTSDNode>(N->getOperand(1))->getVT();
4222 if (LegalOperations && !TLI.isLoadExtLegal(ISD::SEXTLOAD, ExtVT))
4224 } else if (Opc == ISD::SRL) {
4225 // Another special-case: SRL is basically zero-extending a narrower value.
4226 ExtType = ISD::ZEXTLOAD;
4228 ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1));
4229 if (!N01) return SDValue();
4230 ExtVT = EVT::getIntegerVT(*DAG.getContext(),
4231 VT.getSizeInBits() - N01->getZExtValue());
4234 unsigned EVTBits = ExtVT.getSizeInBits();
4236 // Do not generate loads of non-round integer types since these can
4237 // be expensive (and would be wrong if the type is not byte sized).
4238 if (!ExtVT.isRound())
4242 if (N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
4243 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
4244 ShAmt = N01->getZExtValue();
4245 // Is the shift amount a multiple of size of VT?
4246 if ((ShAmt & (EVTBits-1)) == 0) {
4247 N0 = N0.getOperand(0);
4248 // Is the load width a multiple of size of VT?
4249 if ((N0.getValueType().getSizeInBits() & (EVTBits-1)) != 0)
4253 // At this point, we must have a load or else we can't do the transform.
4254 if (!isa<LoadSDNode>(N0)) return SDValue();
4256 // If the shift amount is larger than the input type then we're not
4257 // accessing any of the loaded bytes. If the load was a zextload/extload
4258 // then the result of the shift+trunc is zero/undef (handled elsewhere).
4259 // If the load was a sextload then the result is a splat of the sign bit
4260 // of the extended byte. This is not worth optimizing for.
4261 if (ShAmt >= cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits())
4266 // If the load is shifted left (and the result isn't shifted back right),
4267 // we can fold the truncate through the shift.
4268 unsigned ShLeftAmt = 0;
4269 if (ShAmt == 0 && N0.getOpcode() == ISD::SHL && N0.hasOneUse() &&
4270 ExtVT == VT && TLI.isNarrowingProfitable(N0.getValueType(), VT)) {
4271 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
4272 ShLeftAmt = N01->getZExtValue();
4273 N0 = N0.getOperand(0);
4277 // If we haven't found a load, we can't narrow it. Don't transform one with
4278 // multiple uses, this would require adding a new load.
4279 if (!isa<LoadSDNode>(N0) || !N0.hasOneUse() ||
4280 // Don't change the width of a volatile load.
4281 cast<LoadSDNode>(N0)->isVolatile())
4284 // Verify that we are actually reducing a load width here.
4285 if (cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits() < EVTBits)
4288 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4289 EVT PtrType = N0.getOperand(1).getValueType();
4291 // For big endian targets, we need to adjust the offset to the pointer to
4292 // load the correct bytes.
4293 if (TLI.isBigEndian()) {
4294 unsigned LVTStoreBits = LN0->getMemoryVT().getStoreSizeInBits();
4295 unsigned EVTStoreBits = ExtVT.getStoreSizeInBits();
4296 ShAmt = LVTStoreBits - EVTStoreBits - ShAmt;
4299 uint64_t PtrOff = ShAmt / 8;
4300 unsigned NewAlign = MinAlign(LN0->getAlignment(), PtrOff);
4301 SDValue NewPtr = DAG.getNode(ISD::ADD, LN0->getDebugLoc(),
4302 PtrType, LN0->getBasePtr(),
4303 DAG.getConstant(PtrOff, PtrType));
4304 AddToWorkList(NewPtr.getNode());
4307 if (ExtType == ISD::NON_EXTLOAD)
4308 Load = DAG.getLoad(VT, N0.getDebugLoc(), LN0->getChain(), NewPtr,
4309 LN0->getPointerInfo().getWithOffset(PtrOff),
4310 LN0->isVolatile(), LN0->isNonTemporal(), NewAlign);
4312 Load = DAG.getExtLoad(ExtType, VT, N0.getDebugLoc(), LN0->getChain(),NewPtr,
4313 LN0->getPointerInfo().getWithOffset(PtrOff),
4314 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
4317 // Replace the old load's chain with the new load's chain.
4318 WorkListRemover DeadNodes(*this);
4319 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1),
4322 // Shift the result left, if we've swallowed a left shift.
4323 SDValue Result = Load;
4324 if (ShLeftAmt != 0) {
4325 EVT ShImmTy = getShiftAmountTy();
4326 if (!isUIntN(ShImmTy.getSizeInBits(), ShLeftAmt))
4328 Result = DAG.getNode(ISD::SHL, N0.getDebugLoc(), VT,
4329 Result, DAG.getConstant(ShLeftAmt, ShImmTy));
4332 // Return the new loaded value.
4336 SDValue DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
4337 SDValue N0 = N->getOperand(0);
4338 SDValue N1 = N->getOperand(1);
4339 EVT VT = N->getValueType(0);
4340 EVT EVT = cast<VTSDNode>(N1)->getVT();
4341 unsigned VTBits = VT.getScalarType().getSizeInBits();
4342 unsigned EVTBits = EVT.getScalarType().getSizeInBits();
4344 // fold (sext_in_reg c1) -> c1
4345 if (isa<ConstantSDNode>(N0) || N0.getOpcode() == ISD::UNDEF)
4346 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT, N0, N1);
4348 // If the input is already sign extended, just drop the extension.
4349 if (DAG.ComputeNumSignBits(N0) >= VTBits-EVTBits+1)
4352 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2
4353 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
4354 EVT.bitsLT(cast<VTSDNode>(N0.getOperand(1))->getVT())) {
4355 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT,
4356 N0.getOperand(0), N1);
4359 // fold (sext_in_reg (sext x)) -> (sext x)
4360 // fold (sext_in_reg (aext x)) -> (sext x)
4361 // if x is small enough.
4362 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) {
4363 SDValue N00 = N0.getOperand(0);
4364 if (N00.getValueType().getScalarType().getSizeInBits() <= EVTBits &&
4365 (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND, VT)))
4366 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, N00, N1);
4369 // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is known zero.
4370 if (DAG.MaskedValueIsZero(N0, APInt::getBitsSet(VTBits, EVTBits-1, EVTBits)))
4371 return DAG.getZeroExtendInReg(N0, N->getDebugLoc(), EVT);
4373 // fold operands of sext_in_reg based on knowledge that the top bits are not
4375 if (SimplifyDemandedBits(SDValue(N, 0)))
4376 return SDValue(N, 0);
4378 // fold (sext_in_reg (load x)) -> (smaller sextload x)
4379 // fold (sext_in_reg (srl (load x), c)) -> (smaller sextload (x+c/evtbits))
4380 SDValue NarrowLoad = ReduceLoadWidth(N);
4381 if (NarrowLoad.getNode())
4384 // fold (sext_in_reg (srl X, 24), i8) -> (sra X, 24)
4385 // fold (sext_in_reg (srl X, 23), i8) -> (sra X, 23) iff possible.
4386 // We already fold "(sext_in_reg (srl X, 25), i8) -> srl X, 25" above.
4387 if (N0.getOpcode() == ISD::SRL) {
4388 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
4389 if (ShAmt->getZExtValue()+EVTBits <= VTBits) {
4390 // We can turn this into an SRA iff the input to the SRL is already sign
4392 unsigned InSignBits = DAG.ComputeNumSignBits(N0.getOperand(0));
4393 if (VTBits-(ShAmt->getZExtValue()+EVTBits) < InSignBits)
4394 return DAG.getNode(ISD::SRA, N->getDebugLoc(), VT,
4395 N0.getOperand(0), N0.getOperand(1));
4399 // fold (sext_inreg (extload x)) -> (sextload x)
4400 if (ISD::isEXTLoad(N0.getNode()) &&
4401 ISD::isUNINDEXEDLoad(N0.getNode()) &&
4402 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
4403 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
4404 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) {
4405 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4406 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, N->getDebugLoc(),
4408 LN0->getBasePtr(), LN0->getPointerInfo(),
4410 LN0->isVolatile(), LN0->isNonTemporal(),
4411 LN0->getAlignment());
4412 CombineTo(N, ExtLoad);
4413 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
4414 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4416 // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use
4417 if (ISD::isZEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
4419 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
4420 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
4421 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) {
4422 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4423 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, N->getDebugLoc(),
4425 LN0->getBasePtr(), LN0->getPointerInfo(),
4427 LN0->isVolatile(), LN0->isNonTemporal(),
4428 LN0->getAlignment());
4429 CombineTo(N, ExtLoad);
4430 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
4431 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4436 SDValue DAGCombiner::visitTRUNCATE(SDNode *N) {
4437 SDValue N0 = N->getOperand(0);
4438 EVT VT = N->getValueType(0);
4441 if (N0.getValueType() == N->getValueType(0))
4443 // fold (truncate c1) -> c1
4444 if (isa<ConstantSDNode>(N0))
4445 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, N0);
4446 // fold (truncate (truncate x)) -> (truncate x)
4447 if (N0.getOpcode() == ISD::TRUNCATE)
4448 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, N0.getOperand(0));
4449 // fold (truncate (ext x)) -> (ext x) or (truncate x) or x
4450 if (N0.getOpcode() == ISD::ZERO_EXTEND ||
4451 N0.getOpcode() == ISD::SIGN_EXTEND ||
4452 N0.getOpcode() == ISD::ANY_EXTEND) {
4453 if (N0.getOperand(0).getValueType().bitsLT(VT))
4454 // if the source is smaller than the dest, we still need an extend
4455 return DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT,
4457 else if (N0.getOperand(0).getValueType().bitsGT(VT))
4458 // if the source is larger than the dest, than we just need the truncate
4459 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, N0.getOperand(0));
4461 // if the source and dest are the same type, we can drop both the extend
4462 // and the truncate.
4463 return N0.getOperand(0);
4466 // See if we can simplify the input to this truncate through knowledge that
4467 // only the low bits are being used. For example "trunc (or (shl x, 8), y)"
4470 GetDemandedBits(N0, APInt::getLowBitsSet(N0.getValueSizeInBits(),
4471 VT.getSizeInBits()));
4472 if (Shorter.getNode())
4473 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, Shorter);
4475 // fold (truncate (load x)) -> (smaller load x)
4476 // fold (truncate (srl (load x), c)) -> (smaller load (x+c/evtbits))
4477 if (!LegalTypes || TLI.isTypeDesirableForOp(N0.getOpcode(), VT)) {
4478 SDValue Reduced = ReduceLoadWidth(N);
4479 if (Reduced.getNode())
4483 // Simplify the operands using demanded-bits information.
4484 if (!VT.isVector() &&
4485 SimplifyDemandedBits(SDValue(N, 0)))
4486 return SDValue(N, 0);
4491 static SDNode *getBuildPairElt(SDNode *N, unsigned i) {
4492 SDValue Elt = N->getOperand(i);
4493 if (Elt.getOpcode() != ISD::MERGE_VALUES)
4494 return Elt.getNode();
4495 return Elt.getOperand(Elt.getResNo()).getNode();
4498 /// CombineConsecutiveLoads - build_pair (load, load) -> load
4499 /// if load locations are consecutive.
4500 SDValue DAGCombiner::CombineConsecutiveLoads(SDNode *N, EVT VT) {
4501 assert(N->getOpcode() == ISD::BUILD_PAIR);
4503 LoadSDNode *LD1 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 0));
4504 LoadSDNode *LD2 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 1));
4505 if (!LD1 || !LD2 || !ISD::isNON_EXTLoad(LD1) || !LD1->hasOneUse() ||
4506 LD1->getPointerInfo().getAddrSpace() !=
4507 LD2->getPointerInfo().getAddrSpace())
4509 EVT LD1VT = LD1->getValueType(0);
4511 if (ISD::isNON_EXTLoad(LD2) &&
4513 // If both are volatile this would reduce the number of volatile loads.
4514 // If one is volatile it might be ok, but play conservative and bail out.
4515 !LD1->isVolatile() &&
4516 !LD2->isVolatile() &&
4517 DAG.isConsecutiveLoad(LD2, LD1, LD1VT.getSizeInBits()/8, 1)) {
4518 unsigned Align = LD1->getAlignment();
4519 unsigned NewAlign = TLI.getTargetData()->
4520 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
4522 if (NewAlign <= Align &&
4523 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)))
4524 return DAG.getLoad(VT, N->getDebugLoc(), LD1->getChain(),
4525 LD1->getBasePtr(), LD1->getPointerInfo(),
4526 false, false, Align);
4532 SDValue DAGCombiner::visitBITCAST(SDNode *N) {
4533 SDValue N0 = N->getOperand(0);
4534 EVT VT = N->getValueType(0);
4536 // If the input is a BUILD_VECTOR with all constant elements, fold this now.
4537 // Only do this before legalize, since afterward the target may be depending
4538 // on the bitconvert.
4539 // First check to see if this is all constant.
4541 N0.getOpcode() == ISD::BUILD_VECTOR && N0.getNode()->hasOneUse() &&
4543 bool isSimple = true;
4544 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i)
4545 if (N0.getOperand(i).getOpcode() != ISD::UNDEF &&
4546 N0.getOperand(i).getOpcode() != ISD::Constant &&
4547 N0.getOperand(i).getOpcode() != ISD::ConstantFP) {
4552 EVT DestEltVT = N->getValueType(0).getVectorElementType();
4553 assert(!DestEltVT.isVector() &&
4554 "Element type of vector ValueType must not be vector!");
4556 return ConstantFoldBITCASTofBUILD_VECTOR(N0.getNode(), DestEltVT);
4559 // If the input is a constant, let getNode fold it.
4560 if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) {
4561 SDValue Res = DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, N0);
4562 if (Res.getNode() != N) {
4563 if (!LegalOperations ||
4564 TLI.isOperationLegal(Res.getNode()->getOpcode(), VT))
4567 // Folding it resulted in an illegal node, and it's too late to
4568 // do that. Clean up the old node and forego the transformation.
4569 // Ideally this won't happen very often, because instcombine
4570 // and the earlier dagcombine runs (where illegal nodes are
4571 // permitted) should have folded most of them already.
4572 DAG.DeleteNode(Res.getNode());
4576 // (conv (conv x, t1), t2) -> (conv x, t2)
4577 if (N0.getOpcode() == ISD::BITCAST)
4578 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT,
4581 // fold (conv (load x)) -> (load (conv*)x)
4582 // If the resultant load doesn't need a higher alignment than the original!
4583 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
4584 // Do not change the width of a volatile load.
4585 !cast<LoadSDNode>(N0)->isVolatile() &&
4586 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT))) {
4587 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4588 unsigned Align = TLI.getTargetData()->
4589 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
4590 unsigned OrigAlign = LN0->getAlignment();
4592 if (Align <= OrigAlign) {
4593 SDValue Load = DAG.getLoad(VT, N->getDebugLoc(), LN0->getChain(),
4594 LN0->getBasePtr(), LN0->getPointerInfo(),
4595 LN0->isVolatile(), LN0->isNonTemporal(),
4598 CombineTo(N0.getNode(),
4599 DAG.getNode(ISD::BITCAST, N0.getDebugLoc(),
4600 N0.getValueType(), Load),
4606 // fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit)
4607 // fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit))
4608 // This often reduces constant pool loads.
4609 if ((N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FABS) &&
4610 N0.getNode()->hasOneUse() && VT.isInteger() && !VT.isVector()) {
4611 SDValue NewConv = DAG.getNode(ISD::BITCAST, N0.getDebugLoc(), VT,
4613 AddToWorkList(NewConv.getNode());
4615 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
4616 if (N0.getOpcode() == ISD::FNEG)
4617 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT,
4618 NewConv, DAG.getConstant(SignBit, VT));
4619 assert(N0.getOpcode() == ISD::FABS);
4620 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
4621 NewConv, DAG.getConstant(~SignBit, VT));
4624 // fold (bitconvert (fcopysign cst, x)) ->
4625 // (or (and (bitconvert x), sign), (and cst, (not sign)))
4626 // Note that we don't handle (copysign x, cst) because this can always be
4627 // folded to an fneg or fabs.
4628 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse() &&
4629 isa<ConstantFPSDNode>(N0.getOperand(0)) &&
4630 VT.isInteger() && !VT.isVector()) {
4631 unsigned OrigXWidth = N0.getOperand(1).getValueType().getSizeInBits();
4632 EVT IntXVT = EVT::getIntegerVT(*DAG.getContext(), OrigXWidth);
4633 if (isTypeLegal(IntXVT)) {
4634 SDValue X = DAG.getNode(ISD::BITCAST, N0.getDebugLoc(),
4635 IntXVT, N0.getOperand(1));
4636 AddToWorkList(X.getNode());
4638 // If X has a different width than the result/lhs, sext it or truncate it.
4639 unsigned VTWidth = VT.getSizeInBits();
4640 if (OrigXWidth < VTWidth) {
4641 X = DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, X);
4642 AddToWorkList(X.getNode());
4643 } else if (OrigXWidth > VTWidth) {
4644 // To get the sign bit in the right place, we have to shift it right
4645 // before truncating.
4646 X = DAG.getNode(ISD::SRL, X.getDebugLoc(),
4647 X.getValueType(), X,
4648 DAG.getConstant(OrigXWidth-VTWidth, X.getValueType()));
4649 AddToWorkList(X.getNode());
4650 X = DAG.getNode(ISD::TRUNCATE, X.getDebugLoc(), VT, X);
4651 AddToWorkList(X.getNode());
4654 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
4655 X = DAG.getNode(ISD::AND, X.getDebugLoc(), VT,
4656 X, DAG.getConstant(SignBit, VT));
4657 AddToWorkList(X.getNode());
4659 SDValue Cst = DAG.getNode(ISD::BITCAST, N0.getDebugLoc(),
4660 VT, N0.getOperand(0));
4661 Cst = DAG.getNode(ISD::AND, Cst.getDebugLoc(), VT,
4662 Cst, DAG.getConstant(~SignBit, VT));
4663 AddToWorkList(Cst.getNode());
4665 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, X, Cst);
4669 // bitconvert(build_pair(ld, ld)) -> ld iff load locations are consecutive.
4670 if (N0.getOpcode() == ISD::BUILD_PAIR) {
4671 SDValue CombineLD = CombineConsecutiveLoads(N0.getNode(), VT);
4672 if (CombineLD.getNode())
4679 SDValue DAGCombiner::visitBUILD_PAIR(SDNode *N) {
4680 EVT VT = N->getValueType(0);
4681 return CombineConsecutiveLoads(N, VT);
4684 /// ConstantFoldBITCASTofBUILD_VECTOR - We know that BV is a build_vector
4685 /// node with Constant, ConstantFP or Undef operands. DstEltVT indicates the
4686 /// destination element value type.
4687 SDValue DAGCombiner::
4688 ConstantFoldBITCASTofBUILD_VECTOR(SDNode *BV, EVT DstEltVT) {
4689 EVT SrcEltVT = BV->getValueType(0).getVectorElementType();
4691 // If this is already the right type, we're done.
4692 if (SrcEltVT == DstEltVT) return SDValue(BV, 0);
4694 unsigned SrcBitSize = SrcEltVT.getSizeInBits();
4695 unsigned DstBitSize = DstEltVT.getSizeInBits();
4697 // If this is a conversion of N elements of one type to N elements of another
4698 // type, convert each element. This handles FP<->INT cases.
4699 if (SrcBitSize == DstBitSize) {
4700 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
4701 BV->getValueType(0).getVectorNumElements());
4703 // Due to the FP element handling below calling this routine recursively,
4704 // we can end up with a scalar-to-vector node here.
4705 if (BV->getOpcode() == ISD::SCALAR_TO_VECTOR)
4706 return DAG.getNode(ISD::SCALAR_TO_VECTOR, BV->getDebugLoc(), VT,
4707 DAG.getNode(ISD::BITCAST, BV->getDebugLoc(),
4708 DstEltVT, BV->getOperand(0)));
4710 SmallVector<SDValue, 8> Ops;
4711 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
4712 SDValue Op = BV->getOperand(i);
4713 // If the vector element type is not legal, the BUILD_VECTOR operands
4714 // are promoted and implicitly truncated. Make that explicit here.
4715 if (Op.getValueType() != SrcEltVT)
4716 Op = DAG.getNode(ISD::TRUNCATE, BV->getDebugLoc(), SrcEltVT, Op);
4717 Ops.push_back(DAG.getNode(ISD::BITCAST, BV->getDebugLoc(),
4719 AddToWorkList(Ops.back().getNode());
4721 return DAG.getNode(ISD::BUILD_VECTOR, BV->getDebugLoc(), VT,
4722 &Ops[0], Ops.size());
4725 // Otherwise, we're growing or shrinking the elements. To avoid having to
4726 // handle annoying details of growing/shrinking FP values, we convert them to
4728 if (SrcEltVT.isFloatingPoint()) {
4729 // Convert the input float vector to a int vector where the elements are the
4731 assert((SrcEltVT == MVT::f32 || SrcEltVT == MVT::f64) && "Unknown FP VT!");
4732 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), SrcEltVT.getSizeInBits());
4733 BV = ConstantFoldBITCASTofBUILD_VECTOR(BV, IntVT).getNode();
4737 // Now we know the input is an integer vector. If the output is a FP type,
4738 // convert to integer first, then to FP of the right size.
4739 if (DstEltVT.isFloatingPoint()) {
4740 assert((DstEltVT == MVT::f32 || DstEltVT == MVT::f64) && "Unknown FP VT!");
4741 EVT TmpVT = EVT::getIntegerVT(*DAG.getContext(), DstEltVT.getSizeInBits());
4742 SDNode *Tmp = ConstantFoldBITCASTofBUILD_VECTOR(BV, TmpVT).getNode();
4744 // Next, convert to FP elements of the same size.
4745 return ConstantFoldBITCASTofBUILD_VECTOR(Tmp, DstEltVT);
4748 // Okay, we know the src/dst types are both integers of differing types.
4749 // Handling growing first.
4750 assert(SrcEltVT.isInteger() && DstEltVT.isInteger());
4751 if (SrcBitSize < DstBitSize) {
4752 unsigned NumInputsPerOutput = DstBitSize/SrcBitSize;
4754 SmallVector<SDValue, 8> Ops;
4755 for (unsigned i = 0, e = BV->getNumOperands(); i != e;
4756 i += NumInputsPerOutput) {
4757 bool isLE = TLI.isLittleEndian();
4758 APInt NewBits = APInt(DstBitSize, 0);
4759 bool EltIsUndef = true;
4760 for (unsigned j = 0; j != NumInputsPerOutput; ++j) {
4761 // Shift the previously computed bits over.
4762 NewBits <<= SrcBitSize;
4763 SDValue Op = BV->getOperand(i+ (isLE ? (NumInputsPerOutput-j-1) : j));
4764 if (Op.getOpcode() == ISD::UNDEF) continue;
4767 NewBits |= cast<ConstantSDNode>(Op)->getAPIntValue().
4768 zextOrTrunc(SrcBitSize).zext(DstBitSize);
4772 Ops.push_back(DAG.getUNDEF(DstEltVT));
4774 Ops.push_back(DAG.getConstant(NewBits, DstEltVT));
4777 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, Ops.size());
4778 return DAG.getNode(ISD::BUILD_VECTOR, BV->getDebugLoc(), VT,
4779 &Ops[0], Ops.size());
4782 // Finally, this must be the case where we are shrinking elements: each input
4783 // turns into multiple outputs.
4784 bool isS2V = ISD::isScalarToVector(BV);
4785 unsigned NumOutputsPerInput = SrcBitSize/DstBitSize;
4786 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
4787 NumOutputsPerInput*BV->getNumOperands());
4788 SmallVector<SDValue, 8> Ops;
4790 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
4791 if (BV->getOperand(i).getOpcode() == ISD::UNDEF) {
4792 for (unsigned j = 0; j != NumOutputsPerInput; ++j)
4793 Ops.push_back(DAG.getUNDEF(DstEltVT));
4797 APInt OpVal = cast<ConstantSDNode>(BV->getOperand(i))->
4798 getAPIntValue().zextOrTrunc(SrcBitSize);
4800 for (unsigned j = 0; j != NumOutputsPerInput; ++j) {
4801 APInt ThisVal = OpVal.trunc(DstBitSize);
4802 Ops.push_back(DAG.getConstant(ThisVal, DstEltVT));
4803 if (isS2V && i == 0 && j == 0 && ThisVal.zext(SrcBitSize) == OpVal)
4804 // Simply turn this into a SCALAR_TO_VECTOR of the new type.
4805 return DAG.getNode(ISD::SCALAR_TO_VECTOR, BV->getDebugLoc(), VT,
4807 OpVal = OpVal.lshr(DstBitSize);
4810 // For big endian targets, swap the order of the pieces of each element.
4811 if (TLI.isBigEndian())
4812 std::reverse(Ops.end()-NumOutputsPerInput, Ops.end());
4815 return DAG.getNode(ISD::BUILD_VECTOR, BV->getDebugLoc(), VT,
4816 &Ops[0], Ops.size());
4819 SDValue DAGCombiner::visitFADD(SDNode *N) {
4820 SDValue N0 = N->getOperand(0);
4821 SDValue N1 = N->getOperand(1);
4822 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
4823 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
4824 EVT VT = N->getValueType(0);
4827 if (VT.isVector()) {
4828 SDValue FoldedVOp = SimplifyVBinOp(N);
4829 if (FoldedVOp.getNode()) return FoldedVOp;
4832 // fold (fadd c1, c2) -> (fadd c1, c2)
4833 if (N0CFP && N1CFP && VT != MVT::ppcf128)
4834 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0, N1);
4835 // canonicalize constant to RHS
4836 if (N0CFP && !N1CFP)
4837 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N1, N0);
4838 // fold (fadd A, 0) -> A
4839 if (UnsafeFPMath && N1CFP && N1CFP->getValueAPF().isZero())
4841 // fold (fadd A, (fneg B)) -> (fsub A, B)
4842 if (isNegatibleForFree(N1, LegalOperations) == 2)
4843 return DAG.getNode(ISD::FSUB, N->getDebugLoc(), VT, N0,
4844 GetNegatedExpression(N1, DAG, LegalOperations));
4845 // fold (fadd (fneg A), B) -> (fsub B, A)
4846 if (isNegatibleForFree(N0, LegalOperations) == 2)
4847 return DAG.getNode(ISD::FSUB, N->getDebugLoc(), VT, N1,
4848 GetNegatedExpression(N0, DAG, LegalOperations));
4850 // If allowed, fold (fadd (fadd x, c1), c2) -> (fadd x, (fadd c1, c2))
4851 if (UnsafeFPMath && N1CFP && N0.getOpcode() == ISD::FADD &&
4852 N0.getNode()->hasOneUse() && isa<ConstantFPSDNode>(N0.getOperand(1)))
4853 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0.getOperand(0),
4854 DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
4855 N0.getOperand(1), N1));
4860 SDValue DAGCombiner::visitFSUB(SDNode *N) {
4861 SDValue N0 = N->getOperand(0);
4862 SDValue N1 = N->getOperand(1);
4863 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
4864 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
4865 EVT VT = N->getValueType(0);
4868 if (VT.isVector()) {
4869 SDValue FoldedVOp = SimplifyVBinOp(N);
4870 if (FoldedVOp.getNode()) return FoldedVOp;
4873 // fold (fsub c1, c2) -> c1-c2
4874 if (N0CFP && N1CFP && VT != MVT::ppcf128)
4875 return DAG.getNode(ISD::FSUB, N->getDebugLoc(), VT, N0, N1);
4876 // fold (fsub A, 0) -> A
4877 if (UnsafeFPMath && N1CFP && N1CFP->getValueAPF().isZero())
4879 // fold (fsub 0, B) -> -B
4880 if (UnsafeFPMath && N0CFP && N0CFP->getValueAPF().isZero()) {
4881 if (isNegatibleForFree(N1, LegalOperations))
4882 return GetNegatedExpression(N1, DAG, LegalOperations);
4883 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
4884 return DAG.getNode(ISD::FNEG, N->getDebugLoc(), VT, N1);
4886 // fold (fsub A, (fneg B)) -> (fadd A, B)
4887 if (isNegatibleForFree(N1, LegalOperations))
4888 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0,
4889 GetNegatedExpression(N1, DAG, LegalOperations));
4894 SDValue DAGCombiner::visitFMUL(SDNode *N) {
4895 SDValue N0 = N->getOperand(0);
4896 SDValue N1 = N->getOperand(1);
4897 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
4898 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
4899 EVT VT = N->getValueType(0);
4902 if (VT.isVector()) {
4903 SDValue FoldedVOp = SimplifyVBinOp(N);
4904 if (FoldedVOp.getNode()) return FoldedVOp;
4907 // fold (fmul c1, c2) -> c1*c2
4908 if (N0CFP && N1CFP && VT != MVT::ppcf128)
4909 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT, N0, N1);
4910 // canonicalize constant to RHS
4911 if (N0CFP && !N1CFP)
4912 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT, N1, N0);
4913 // fold (fmul A, 0) -> 0
4914 if (UnsafeFPMath && N1CFP && N1CFP->getValueAPF().isZero())
4916 // fold (fmul A, 0) -> 0, vector edition.
4917 if (UnsafeFPMath && ISD::isBuildVectorAllZeros(N1.getNode()))
4919 // fold (fmul X, 2.0) -> (fadd X, X)
4920 if (N1CFP && N1CFP->isExactlyValue(+2.0))
4921 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0, N0);
4922 // fold (fmul X, -1.0) -> (fneg X)
4923 if (N1CFP && N1CFP->isExactlyValue(-1.0))
4924 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
4925 return DAG.getNode(ISD::FNEG, N->getDebugLoc(), VT, N0);
4927 // fold (fmul (fneg X), (fneg Y)) -> (fmul X, Y)
4928 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations)) {
4929 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations)) {
4930 // Both can be negated for free, check to see if at least one is cheaper
4932 if (LHSNeg == 2 || RHSNeg == 2)
4933 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
4934 GetNegatedExpression(N0, DAG, LegalOperations),
4935 GetNegatedExpression(N1, DAG, LegalOperations));
4939 // If allowed, fold (fmul (fmul x, c1), c2) -> (fmul x, (fmul c1, c2))
4940 if (UnsafeFPMath && N1CFP && N0.getOpcode() == ISD::FMUL &&
4941 N0.getNode()->hasOneUse() && isa<ConstantFPSDNode>(N0.getOperand(1)))
4942 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT, N0.getOperand(0),
4943 DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
4944 N0.getOperand(1), N1));
4949 SDValue DAGCombiner::visitFDIV(SDNode *N) {
4950 SDValue N0 = N->getOperand(0);
4951 SDValue N1 = N->getOperand(1);
4952 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
4953 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
4954 EVT VT = N->getValueType(0);
4957 if (VT.isVector()) {
4958 SDValue FoldedVOp = SimplifyVBinOp(N);
4959 if (FoldedVOp.getNode()) return FoldedVOp;
4962 // fold (fdiv c1, c2) -> c1/c2
4963 if (N0CFP && N1CFP && VT != MVT::ppcf128)
4964 return DAG.getNode(ISD::FDIV, N->getDebugLoc(), VT, N0, N1);
4967 // (fdiv (fneg X), (fneg Y)) -> (fdiv X, Y)
4968 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations)) {
4969 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations)) {
4970 // Both can be negated for free, check to see if at least one is cheaper
4972 if (LHSNeg == 2 || RHSNeg == 2)
4973 return DAG.getNode(ISD::FDIV, N->getDebugLoc(), VT,
4974 GetNegatedExpression(N0, DAG, LegalOperations),
4975 GetNegatedExpression(N1, DAG, LegalOperations));
4982 SDValue DAGCombiner::visitFREM(SDNode *N) {
4983 SDValue N0 = N->getOperand(0);
4984 SDValue N1 = N->getOperand(1);
4985 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
4986 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
4987 EVT VT = N->getValueType(0);
4989 // fold (frem c1, c2) -> fmod(c1,c2)
4990 if (N0CFP && N1CFP && VT != MVT::ppcf128)
4991 return DAG.getNode(ISD::FREM, N->getDebugLoc(), VT, N0, N1);
4996 SDValue DAGCombiner::visitFCOPYSIGN(SDNode *N) {
4997 SDValue N0 = N->getOperand(0);
4998 SDValue N1 = N->getOperand(1);
4999 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5000 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
5001 EVT VT = N->getValueType(0);
5003 if (N0CFP && N1CFP && VT != MVT::ppcf128) // Constant fold
5004 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT, N0, N1);
5007 const APFloat& V = N1CFP->getValueAPF();
5008 // copysign(x, c1) -> fabs(x) iff ispos(c1)
5009 // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1)
5010 if (!V.isNegative()) {
5011 if (!LegalOperations || TLI.isOperationLegal(ISD::FABS, VT))
5012 return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0);
5014 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
5015 return DAG.getNode(ISD::FNEG, N->getDebugLoc(), VT,
5016 DAG.getNode(ISD::FABS, N0.getDebugLoc(), VT, N0));
5020 // copysign(fabs(x), y) -> copysign(x, y)
5021 // copysign(fneg(x), y) -> copysign(x, y)
5022 // copysign(copysign(x,z), y) -> copysign(x, y)
5023 if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG ||
5024 N0.getOpcode() == ISD::FCOPYSIGN)
5025 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
5026 N0.getOperand(0), N1);
5028 // copysign(x, abs(y)) -> abs(x)
5029 if (N1.getOpcode() == ISD::FABS)
5030 return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0);
5032 // copysign(x, copysign(y,z)) -> copysign(x, z)
5033 if (N1.getOpcode() == ISD::FCOPYSIGN)
5034 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
5035 N0, N1.getOperand(1));
5037 // copysign(x, fp_extend(y)) -> copysign(x, y)
5038 // copysign(x, fp_round(y)) -> copysign(x, y)
5039 if (N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND)
5040 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
5041 N0, N1.getOperand(0));
5046 SDValue DAGCombiner::visitSINT_TO_FP(SDNode *N) {
5047 SDValue N0 = N->getOperand(0);
5048 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
5049 EVT VT = N->getValueType(0);
5050 EVT OpVT = N0.getValueType();
5052 // fold (sint_to_fp c1) -> c1fp
5053 if (N0C && OpVT != MVT::ppcf128)
5054 return DAG.getNode(ISD::SINT_TO_FP, N->getDebugLoc(), VT, N0);
5056 // If the input is a legal type, and SINT_TO_FP is not legal on this target,
5057 // but UINT_TO_FP is legal on this target, try to convert.
5058 if (!TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT) &&
5059 TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT)) {
5060 // If the sign bit is known to be zero, we can change this to UINT_TO_FP.
5061 if (DAG.SignBitIsZero(N0))
5062 return DAG.getNode(ISD::UINT_TO_FP, N->getDebugLoc(), VT, N0);
5068 SDValue DAGCombiner::visitUINT_TO_FP(SDNode *N) {
5069 SDValue N0 = N->getOperand(0);
5070 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
5071 EVT VT = N->getValueType(0);
5072 EVT OpVT = N0.getValueType();
5074 // fold (uint_to_fp c1) -> c1fp
5075 if (N0C && OpVT != MVT::ppcf128)
5076 return DAG.getNode(ISD::UINT_TO_FP, N->getDebugLoc(), VT, N0);
5078 // If the input is a legal type, and UINT_TO_FP is not legal on this target,
5079 // but SINT_TO_FP is legal on this target, try to convert.
5080 if (!TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT) &&
5081 TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT)) {
5082 // If the sign bit is known to be zero, we can change this to SINT_TO_FP.
5083 if (DAG.SignBitIsZero(N0))
5084 return DAG.getNode(ISD::SINT_TO_FP, N->getDebugLoc(), VT, N0);
5090 SDValue DAGCombiner::visitFP_TO_SINT(SDNode *N) {
5091 SDValue N0 = N->getOperand(0);
5092 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5093 EVT VT = N->getValueType(0);
5095 // fold (fp_to_sint c1fp) -> c1
5097 return DAG.getNode(ISD::FP_TO_SINT, N->getDebugLoc(), VT, N0);
5102 SDValue DAGCombiner::visitFP_TO_UINT(SDNode *N) {
5103 SDValue N0 = N->getOperand(0);
5104 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5105 EVT VT = N->getValueType(0);
5107 // fold (fp_to_uint c1fp) -> c1
5108 if (N0CFP && VT != MVT::ppcf128)
5109 return DAG.getNode(ISD::FP_TO_UINT, N->getDebugLoc(), VT, N0);
5114 SDValue DAGCombiner::visitFP_ROUND(SDNode *N) {
5115 SDValue N0 = N->getOperand(0);
5116 SDValue N1 = N->getOperand(1);
5117 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5118 EVT VT = N->getValueType(0);
5120 // fold (fp_round c1fp) -> c1fp
5121 if (N0CFP && N0.getValueType() != MVT::ppcf128)
5122 return DAG.getNode(ISD::FP_ROUND, N->getDebugLoc(), VT, N0, N1);
5124 // fold (fp_round (fp_extend x)) -> x
5125 if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType())
5126 return N0.getOperand(0);
5128 // fold (fp_round (fp_round x)) -> (fp_round x)
5129 if (N0.getOpcode() == ISD::FP_ROUND) {
5130 // This is a value preserving truncation if both round's are.
5131 bool IsTrunc = N->getConstantOperandVal(1) == 1 &&
5132 N0.getNode()->getConstantOperandVal(1) == 1;
5133 return DAG.getNode(ISD::FP_ROUND, N->getDebugLoc(), VT, N0.getOperand(0),
5134 DAG.getIntPtrConstant(IsTrunc));
5137 // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y)
5138 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse()) {
5139 SDValue Tmp = DAG.getNode(ISD::FP_ROUND, N0.getDebugLoc(), VT,
5140 N0.getOperand(0), N1);
5141 AddToWorkList(Tmp.getNode());
5142 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
5143 Tmp, N0.getOperand(1));
5149 SDValue DAGCombiner::visitFP_ROUND_INREG(SDNode *N) {
5150 SDValue N0 = N->getOperand(0);
5151 EVT VT = N->getValueType(0);
5152 EVT EVT = cast<VTSDNode>(N->getOperand(1))->getVT();
5153 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5155 // fold (fp_round_inreg c1fp) -> c1fp
5156 if (N0CFP && isTypeLegal(EVT)) {
5157 SDValue Round = DAG.getConstantFP(*N0CFP->getConstantFPValue(), EVT);
5158 return DAG.getNode(ISD::FP_EXTEND, N->getDebugLoc(), VT, Round);
5164 SDValue DAGCombiner::visitFP_EXTEND(SDNode *N) {
5165 SDValue N0 = N->getOperand(0);
5166 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5167 EVT VT = N->getValueType(0);
5169 // If this is fp_round(fpextend), don't fold it, allow ourselves to be folded.
5170 if (N->hasOneUse() &&
5171 N->use_begin()->getOpcode() == ISD::FP_ROUND)
5174 // fold (fp_extend c1fp) -> c1fp
5175 if (N0CFP && VT != MVT::ppcf128)
5176 return DAG.getNode(ISD::FP_EXTEND, N->getDebugLoc(), VT, N0);
5178 // Turn fp_extend(fp_round(X, 1)) -> x since the fp_round doesn't affect the
5180 if (N0.getOpcode() == ISD::FP_ROUND
5181 && N0.getNode()->getConstantOperandVal(1) == 1) {
5182 SDValue In = N0.getOperand(0);
5183 if (In.getValueType() == VT) return In;
5184 if (VT.bitsLT(In.getValueType()))
5185 return DAG.getNode(ISD::FP_ROUND, N->getDebugLoc(), VT,
5186 In, N0.getOperand(1));
5187 return DAG.getNode(ISD::FP_EXTEND, N->getDebugLoc(), VT, In);
5190 // fold (fpext (load x)) -> (fpext (fptrunc (extload x)))
5191 if (ISD::isNON_EXTLoad(N0.getNode()) && N0.hasOneUse() &&
5192 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5193 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType()))) {
5194 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5195 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, VT, N->getDebugLoc(),
5197 LN0->getBasePtr(), LN0->getPointerInfo(),
5199 LN0->isVolatile(), LN0->isNonTemporal(),
5200 LN0->getAlignment());
5201 CombineTo(N, ExtLoad);
5202 CombineTo(N0.getNode(),
5203 DAG.getNode(ISD::FP_ROUND, N0.getDebugLoc(),
5204 N0.getValueType(), ExtLoad, DAG.getIntPtrConstant(1)),
5205 ExtLoad.getValue(1));
5206 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5212 SDValue DAGCombiner::visitFNEG(SDNode *N) {
5213 SDValue N0 = N->getOperand(0);
5214 EVT VT = N->getValueType(0);
5216 if (isNegatibleForFree(N0, LegalOperations))
5217 return GetNegatedExpression(N0, DAG, LegalOperations);
5219 // Transform fneg(bitconvert(x)) -> bitconvert(x^sign) to avoid loading
5220 // constant pool values.
5221 if (N0.getOpcode() == ISD::BITCAST &&
5223 N0.getNode()->hasOneUse() &&
5224 N0.getOperand(0).getValueType().isInteger()) {
5225 SDValue Int = N0.getOperand(0);
5226 EVT IntVT = Int.getValueType();
5227 if (IntVT.isInteger() && !IntVT.isVector()) {
5228 Int = DAG.getNode(ISD::XOR, N0.getDebugLoc(), IntVT, Int,
5229 DAG.getConstant(APInt::getSignBit(IntVT.getSizeInBits()), IntVT));
5230 AddToWorkList(Int.getNode());
5231 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
5239 SDValue DAGCombiner::visitFABS(SDNode *N) {
5240 SDValue N0 = N->getOperand(0);
5241 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5242 EVT VT = N->getValueType(0);
5244 // fold (fabs c1) -> fabs(c1)
5245 if (N0CFP && VT != MVT::ppcf128)
5246 return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0);
5247 // fold (fabs (fabs x)) -> (fabs x)
5248 if (N0.getOpcode() == ISD::FABS)
5249 return N->getOperand(0);
5250 // fold (fabs (fneg x)) -> (fabs x)
5251 // fold (fabs (fcopysign x, y)) -> (fabs x)
5252 if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN)
5253 return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0.getOperand(0));
5255 // Transform fabs(bitconvert(x)) -> bitconvert(x&~sign) to avoid loading
5256 // constant pool values.
5257 if (N0.getOpcode() == ISD::BITCAST && N0.getNode()->hasOneUse() &&
5258 N0.getOperand(0).getValueType().isInteger() &&
5259 !N0.getOperand(0).getValueType().isVector()) {
5260 SDValue Int = N0.getOperand(0);
5261 EVT IntVT = Int.getValueType();
5262 if (IntVT.isInteger() && !IntVT.isVector()) {
5263 Int = DAG.getNode(ISD::AND, N0.getDebugLoc(), IntVT, Int,
5264 DAG.getConstant(~APInt::getSignBit(IntVT.getSizeInBits()), IntVT));
5265 AddToWorkList(Int.getNode());
5266 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
5267 N->getValueType(0), Int);
5274 SDValue DAGCombiner::visitBRCOND(SDNode *N) {
5275 SDValue Chain = N->getOperand(0);
5276 SDValue N1 = N->getOperand(1);
5277 SDValue N2 = N->getOperand(2);
5279 // If N is a constant we could fold this into a fallthrough or unconditional
5280 // branch. However that doesn't happen very often in normal code, because
5281 // Instcombine/SimplifyCFG should have handled the available opportunities.
5282 // If we did this folding here, it would be necessary to update the
5283 // MachineBasicBlock CFG, which is awkward.
5285 // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal
5287 if (N1.getOpcode() == ISD::SETCC &&
5288 TLI.isOperationLegalOrCustom(ISD::BR_CC, MVT::Other)) {
5289 return DAG.getNode(ISD::BR_CC, N->getDebugLoc(), MVT::Other,
5290 Chain, N1.getOperand(2),
5291 N1.getOperand(0), N1.getOperand(1), N2);
5294 if ((N1.hasOneUse() && N1.getOpcode() == ISD::SRL) ||
5295 ((N1.getOpcode() == ISD::TRUNCATE && N1.hasOneUse()) &&
5296 (N1.getOperand(0).hasOneUse() &&
5297 N1.getOperand(0).getOpcode() == ISD::SRL))) {
5299 if (N1.getOpcode() == ISD::TRUNCATE) {
5300 // Look pass the truncate.
5301 Trunc = N1.getNode();
5302 N1 = N1.getOperand(0);
5305 // Match this pattern so that we can generate simpler code:
5308 // %b = and i32 %a, 2
5309 // %c = srl i32 %b, 1
5310 // brcond i32 %c ...
5315 // %b = and i32 %a, 2
5316 // %c = setcc eq %b, 0
5319 // This applies only when the AND constant value has one bit set and the
5320 // SRL constant is equal to the log2 of the AND constant. The back-end is
5321 // smart enough to convert the result into a TEST/JMP sequence.
5322 SDValue Op0 = N1.getOperand(0);
5323 SDValue Op1 = N1.getOperand(1);
5325 if (Op0.getOpcode() == ISD::AND &&
5326 Op1.getOpcode() == ISD::Constant) {
5327 SDValue AndOp1 = Op0.getOperand(1);
5329 if (AndOp1.getOpcode() == ISD::Constant) {
5330 const APInt &AndConst = cast<ConstantSDNode>(AndOp1)->getAPIntValue();
5332 if (AndConst.isPowerOf2() &&
5333 cast<ConstantSDNode>(Op1)->getAPIntValue()==AndConst.logBase2()) {
5335 DAG.getSetCC(N->getDebugLoc(),
5336 TLI.getSetCCResultType(Op0.getValueType()),
5337 Op0, DAG.getConstant(0, Op0.getValueType()),
5340 SDValue NewBRCond = DAG.getNode(ISD::BRCOND, N->getDebugLoc(),
5341 MVT::Other, Chain, SetCC, N2);
5342 // Don't add the new BRCond into the worklist or else SimplifySelectCC
5343 // will convert it back to (X & C1) >> C2.
5344 CombineTo(N, NewBRCond, false);
5345 // Truncate is dead.
5347 removeFromWorkList(Trunc);
5348 DAG.DeleteNode(Trunc);
5350 // Replace the uses of SRL with SETCC
5351 WorkListRemover DeadNodes(*this);
5352 DAG.ReplaceAllUsesOfValueWith(N1, SetCC, &DeadNodes);
5353 removeFromWorkList(N1.getNode());
5354 DAG.DeleteNode(N1.getNode());
5355 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5361 // Restore N1 if the above transformation doesn't match.
5362 N1 = N->getOperand(1);
5365 // Transform br(xor(x, y)) -> br(x != y)
5366 // Transform br(xor(xor(x,y), 1)) -> br (x == y)
5367 if (N1.hasOneUse() && N1.getOpcode() == ISD::XOR) {
5368 SDNode *TheXor = N1.getNode();
5369 SDValue Op0 = TheXor->getOperand(0);
5370 SDValue Op1 = TheXor->getOperand(1);
5371 if (Op0.getOpcode() == Op1.getOpcode()) {
5372 // Avoid missing important xor optimizations.
5373 SDValue Tmp = visitXOR(TheXor);
5374 if (Tmp.getNode() && Tmp.getNode() != TheXor) {
5375 DEBUG(dbgs() << "\nReplacing.8 ";
5377 dbgs() << "\nWith: ";
5378 Tmp.getNode()->dump(&DAG);
5380 WorkListRemover DeadNodes(*this);
5381 DAG.ReplaceAllUsesOfValueWith(N1, Tmp, &DeadNodes);
5382 removeFromWorkList(TheXor);
5383 DAG.DeleteNode(TheXor);
5384 return DAG.getNode(ISD::BRCOND, N->getDebugLoc(),
5385 MVT::Other, Chain, Tmp, N2);
5389 if (Op0.getOpcode() != ISD::SETCC && Op1.getOpcode() != ISD::SETCC) {
5391 if (ConstantSDNode *RHSCI = dyn_cast<ConstantSDNode>(Op0))
5392 if (RHSCI->getAPIntValue() == 1 && Op0.hasOneUse() &&
5393 Op0.getOpcode() == ISD::XOR) {
5394 TheXor = Op0.getNode();
5398 EVT SetCCVT = N1.getValueType();
5400 SetCCVT = TLI.getSetCCResultType(SetCCVT);
5401 SDValue SetCC = DAG.getSetCC(TheXor->getDebugLoc(),
5404 Equal ? ISD::SETEQ : ISD::SETNE);
5405 // Replace the uses of XOR with SETCC
5406 WorkListRemover DeadNodes(*this);
5407 DAG.ReplaceAllUsesOfValueWith(N1, SetCC, &DeadNodes);
5408 removeFromWorkList(N1.getNode());
5409 DAG.DeleteNode(N1.getNode());
5410 return DAG.getNode(ISD::BRCOND, N->getDebugLoc(),
5411 MVT::Other, Chain, SetCC, N2);
5418 // Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB.
5420 SDValue DAGCombiner::visitBR_CC(SDNode *N) {
5421 CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1));
5422 SDValue CondLHS = N->getOperand(2), CondRHS = N->getOperand(3);
5424 // If N is a constant we could fold this into a fallthrough or unconditional
5425 // branch. However that doesn't happen very often in normal code, because
5426 // Instcombine/SimplifyCFG should have handled the available opportunities.
5427 // If we did this folding here, it would be necessary to update the
5428 // MachineBasicBlock CFG, which is awkward.
5430 // Use SimplifySetCC to simplify SETCC's.
5431 SDValue Simp = SimplifySetCC(TLI.getSetCCResultType(CondLHS.getValueType()),
5432 CondLHS, CondRHS, CC->get(), N->getDebugLoc(),
5434 if (Simp.getNode()) AddToWorkList(Simp.getNode());
5436 // fold to a simpler setcc
5437 if (Simp.getNode() && Simp.getOpcode() == ISD::SETCC)
5438 return DAG.getNode(ISD::BR_CC, N->getDebugLoc(), MVT::Other,
5439 N->getOperand(0), Simp.getOperand(2),
5440 Simp.getOperand(0), Simp.getOperand(1),
5446 /// CombineToPreIndexedLoadStore - Try turning a load / store into a
5447 /// pre-indexed load / store when the base pointer is an add or subtract
5448 /// and it has other uses besides the load / store. After the
5449 /// transformation, the new indexed load / store has effectively folded
5450 /// the add / subtract in and all of its other uses are redirected to the
5451 /// new load / store.
5452 bool DAGCombiner::CombineToPreIndexedLoadStore(SDNode *N) {
5453 if (!LegalOperations)
5459 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
5460 if (LD->isIndexed())
5462 VT = LD->getMemoryVT();
5463 if (!TLI.isIndexedLoadLegal(ISD::PRE_INC, VT) &&
5464 !TLI.isIndexedLoadLegal(ISD::PRE_DEC, VT))
5466 Ptr = LD->getBasePtr();
5467 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
5468 if (ST->isIndexed())
5470 VT = ST->getMemoryVT();
5471 if (!TLI.isIndexedStoreLegal(ISD::PRE_INC, VT) &&
5472 !TLI.isIndexedStoreLegal(ISD::PRE_DEC, VT))
5474 Ptr = ST->getBasePtr();
5480 // If the pointer is not an add/sub, or if it doesn't have multiple uses, bail
5481 // out. There is no reason to make this a preinc/predec.
5482 if ((Ptr.getOpcode() != ISD::ADD && Ptr.getOpcode() != ISD::SUB) ||
5483 Ptr.getNode()->hasOneUse())
5486 // Ask the target to do addressing mode selection.
5489 ISD::MemIndexedMode AM = ISD::UNINDEXED;
5490 if (!TLI.getPreIndexedAddressParts(N, BasePtr, Offset, AM, DAG))
5492 // Don't create a indexed load / store with zero offset.
5493 if (isa<ConstantSDNode>(Offset) &&
5494 cast<ConstantSDNode>(Offset)->isNullValue())
5497 // Try turning it into a pre-indexed load / store except when:
5498 // 1) The new base ptr is a frame index.
5499 // 2) If N is a store and the new base ptr is either the same as or is a
5500 // predecessor of the value being stored.
5501 // 3) Another use of old base ptr is a predecessor of N. If ptr is folded
5502 // that would create a cycle.
5503 // 4) All uses are load / store ops that use it as old base ptr.
5505 // Check #1. Preinc'ing a frame index would require copying the stack pointer
5506 // (plus the implicit offset) to a register to preinc anyway.
5507 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
5512 SDValue Val = cast<StoreSDNode>(N)->getValue();
5513 if (Val == BasePtr || BasePtr.getNode()->isPredecessorOf(Val.getNode()))
5517 // Now check for #3 and #4.
5518 bool RealUse = false;
5519 for (SDNode::use_iterator I = Ptr.getNode()->use_begin(),
5520 E = Ptr.getNode()->use_end(); I != E; ++I) {
5524 if (Use->isPredecessorOf(N))
5527 if (!((Use->getOpcode() == ISD::LOAD &&
5528 cast<LoadSDNode>(Use)->getBasePtr() == Ptr) ||
5529 (Use->getOpcode() == ISD::STORE &&
5530 cast<StoreSDNode>(Use)->getBasePtr() == Ptr)))
5539 Result = DAG.getIndexedLoad(SDValue(N,0), N->getDebugLoc(),
5540 BasePtr, Offset, AM);
5542 Result = DAG.getIndexedStore(SDValue(N,0), N->getDebugLoc(),
5543 BasePtr, Offset, AM);
5546 DEBUG(dbgs() << "\nReplacing.4 ";
5548 dbgs() << "\nWith: ";
5549 Result.getNode()->dump(&DAG);
5551 WorkListRemover DeadNodes(*this);
5553 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0),
5555 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2),
5558 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1),
5562 // Finally, since the node is now dead, remove it from the graph.
5565 // Replace the uses of Ptr with uses of the updated base value.
5566 DAG.ReplaceAllUsesOfValueWith(Ptr, Result.getValue(isLoad ? 1 : 0),
5568 removeFromWorkList(Ptr.getNode());
5569 DAG.DeleteNode(Ptr.getNode());
5574 /// CombineToPostIndexedLoadStore - Try to combine a load / store with a
5575 /// add / sub of the base pointer node into a post-indexed load / store.
5576 /// The transformation folded the add / subtract into the new indexed
5577 /// load / store effectively and all of its uses are redirected to the
5578 /// new load / store.
5579 bool DAGCombiner::CombineToPostIndexedLoadStore(SDNode *N) {
5580 if (!LegalOperations)
5586 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
5587 if (LD->isIndexed())
5589 VT = LD->getMemoryVT();
5590 if (!TLI.isIndexedLoadLegal(ISD::POST_INC, VT) &&
5591 !TLI.isIndexedLoadLegal(ISD::POST_DEC, VT))
5593 Ptr = LD->getBasePtr();
5594 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
5595 if (ST->isIndexed())
5597 VT = ST->getMemoryVT();
5598 if (!TLI.isIndexedStoreLegal(ISD::POST_INC, VT) &&
5599 !TLI.isIndexedStoreLegal(ISD::POST_DEC, VT))
5601 Ptr = ST->getBasePtr();
5607 if (Ptr.getNode()->hasOneUse())
5610 for (SDNode::use_iterator I = Ptr.getNode()->use_begin(),
5611 E = Ptr.getNode()->use_end(); I != E; ++I) {
5614 (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB))
5619 ISD::MemIndexedMode AM = ISD::UNINDEXED;
5620 if (TLI.getPostIndexedAddressParts(N, Op, BasePtr, Offset, AM, DAG)) {
5621 // Don't create a indexed load / store with zero offset.
5622 if (isa<ConstantSDNode>(Offset) &&
5623 cast<ConstantSDNode>(Offset)->isNullValue())
5626 // Try turning it into a post-indexed load / store except when
5627 // 1) All uses are load / store ops that use it as base ptr.
5628 // 2) Op must be independent of N, i.e. Op is neither a predecessor
5629 // nor a successor of N. Otherwise, if Op is folded that would
5632 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
5636 bool TryNext = false;
5637 for (SDNode::use_iterator II = BasePtr.getNode()->use_begin(),
5638 EE = BasePtr.getNode()->use_end(); II != EE; ++II) {
5640 if (Use == Ptr.getNode())
5643 // If all the uses are load / store addresses, then don't do the
5645 if (Use->getOpcode() == ISD::ADD || Use->getOpcode() == ISD::SUB){
5646 bool RealUse = false;
5647 for (SDNode::use_iterator III = Use->use_begin(),
5648 EEE = Use->use_end(); III != EEE; ++III) {
5649 SDNode *UseUse = *III;
5650 if (!((UseUse->getOpcode() == ISD::LOAD &&
5651 cast<LoadSDNode>(UseUse)->getBasePtr().getNode() == Use) ||
5652 (UseUse->getOpcode() == ISD::STORE &&
5653 cast<StoreSDNode>(UseUse)->getBasePtr().getNode() == Use)))
5668 if (!Op->isPredecessorOf(N) && !N->isPredecessorOf(Op)) {
5669 SDValue Result = isLoad
5670 ? DAG.getIndexedLoad(SDValue(N,0), N->getDebugLoc(),
5671 BasePtr, Offset, AM)
5672 : DAG.getIndexedStore(SDValue(N,0), N->getDebugLoc(),
5673 BasePtr, Offset, AM);
5676 DEBUG(dbgs() << "\nReplacing.5 ";
5678 dbgs() << "\nWith: ";
5679 Result.getNode()->dump(&DAG);
5681 WorkListRemover DeadNodes(*this);
5683 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0),
5685 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2),
5688 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1),
5692 // Finally, since the node is now dead, remove it from the graph.
5695 // Replace the uses of Use with uses of the updated base value.
5696 DAG.ReplaceAllUsesOfValueWith(SDValue(Op, 0),
5697 Result.getValue(isLoad ? 1 : 0),
5699 removeFromWorkList(Op);
5709 SDValue DAGCombiner::visitLOAD(SDNode *N) {
5710 LoadSDNode *LD = cast<LoadSDNode>(N);
5711 SDValue Chain = LD->getChain();
5712 SDValue Ptr = LD->getBasePtr();
5714 // If load is not volatile and there are no uses of the loaded value (and
5715 // the updated indexed value in case of indexed loads), change uses of the
5716 // chain value into uses of the chain input (i.e. delete the dead load).
5717 if (!LD->isVolatile()) {
5718 if (N->getValueType(1) == MVT::Other) {
5720 if (N->hasNUsesOfValue(0, 0)) {
5721 // It's not safe to use the two value CombineTo variant here. e.g.
5722 // v1, chain2 = load chain1, loc
5723 // v2, chain3 = load chain2, loc
5725 // Now we replace use of chain2 with chain1. This makes the second load
5726 // isomorphic to the one we are deleting, and thus makes this load live.
5727 DEBUG(dbgs() << "\nReplacing.6 ";
5729 dbgs() << "\nWith chain: ";
5730 Chain.getNode()->dump(&DAG);
5732 WorkListRemover DeadNodes(*this);
5733 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain, &DeadNodes);
5735 if (N->use_empty()) {
5736 removeFromWorkList(N);
5740 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5744 assert(N->getValueType(2) == MVT::Other && "Malformed indexed loads?");
5745 if (N->hasNUsesOfValue(0, 0) && N->hasNUsesOfValue(0, 1)) {
5746 SDValue Undef = DAG.getUNDEF(N->getValueType(0));
5747 DEBUG(dbgs() << "\nReplacing.7 ";
5749 dbgs() << "\nWith: ";
5750 Undef.getNode()->dump(&DAG);
5751 dbgs() << " and 2 other values\n");
5752 WorkListRemover DeadNodes(*this);
5753 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Undef, &DeadNodes);
5754 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1),
5755 DAG.getUNDEF(N->getValueType(1)),
5757 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 2), Chain, &DeadNodes);
5758 removeFromWorkList(N);
5760 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5765 // If this load is directly stored, replace the load value with the stored
5767 // TODO: Handle store large -> read small portion.
5768 // TODO: Handle TRUNCSTORE/LOADEXT
5769 if (LD->getExtensionType() == ISD::NON_EXTLOAD &&
5770 !LD->isVolatile()) {
5771 if (ISD::isNON_TRUNCStore(Chain.getNode())) {
5772 StoreSDNode *PrevST = cast<StoreSDNode>(Chain);
5773 if (PrevST->getBasePtr() == Ptr &&
5774 PrevST->getValue().getValueType() == N->getValueType(0))
5775 return CombineTo(N, Chain.getOperand(1), Chain);
5779 // Try to infer better alignment information than the load already has.
5780 if (OptLevel != CodeGenOpt::None && LD->isUnindexed()) {
5781 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
5782 if (Align > LD->getAlignment())
5783 return DAG.getExtLoad(LD->getExtensionType(), LD->getValueType(0),
5785 Chain, Ptr, LD->getPointerInfo(),
5787 LD->isVolatile(), LD->isNonTemporal(), Align);
5792 // Walk up chain skipping non-aliasing memory nodes.
5793 SDValue BetterChain = FindBetterChain(N, Chain);
5795 // If there is a better chain.
5796 if (Chain != BetterChain) {
5799 // Replace the chain to void dependency.
5800 if (LD->getExtensionType() == ISD::NON_EXTLOAD) {
5801 ReplLoad = DAG.getLoad(N->getValueType(0), LD->getDebugLoc(),
5802 BetterChain, Ptr, LD->getPointerInfo(),
5803 LD->isVolatile(), LD->isNonTemporal(),
5804 LD->getAlignment());
5806 ReplLoad = DAG.getExtLoad(LD->getExtensionType(), LD->getValueType(0),
5808 BetterChain, Ptr, LD->getPointerInfo(),
5811 LD->isNonTemporal(),
5812 LD->getAlignment());
5815 // Create token factor to keep old chain connected.
5816 SDValue Token = DAG.getNode(ISD::TokenFactor, N->getDebugLoc(),
5817 MVT::Other, Chain, ReplLoad.getValue(1));
5819 // Make sure the new and old chains are cleaned up.
5820 AddToWorkList(Token.getNode());
5822 // Replace uses with load result and token factor. Don't add users
5824 return CombineTo(N, ReplLoad.getValue(0), Token, false);
5828 // Try transforming N to an indexed load.
5829 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
5830 return SDValue(N, 0);
5835 /// CheckForMaskedLoad - Check to see if V is (and load (ptr), imm), where the
5836 /// load is having specific bytes cleared out. If so, return the byte size
5837 /// being masked out and the shift amount.
5838 static std::pair<unsigned, unsigned>
5839 CheckForMaskedLoad(SDValue V, SDValue Ptr, SDValue Chain) {
5840 std::pair<unsigned, unsigned> Result(0, 0);
5842 // Check for the structure we're looking for.
5843 if (V->getOpcode() != ISD::AND ||
5844 !isa<ConstantSDNode>(V->getOperand(1)) ||
5845 !ISD::isNormalLoad(V->getOperand(0).getNode()))
5848 // Check the chain and pointer.
5849 LoadSDNode *LD = cast<LoadSDNode>(V->getOperand(0));
5850 if (LD->getBasePtr() != Ptr) return Result; // Not from same pointer.
5852 // The store should be chained directly to the load or be an operand of a
5854 if (LD == Chain.getNode())
5856 else if (Chain->getOpcode() != ISD::TokenFactor)
5857 return Result; // Fail.
5860 for (unsigned i = 0, e = Chain->getNumOperands(); i != e; ++i)
5861 if (Chain->getOperand(i).getNode() == LD) {
5865 if (!isOk) return Result;
5868 // This only handles simple types.
5869 if (V.getValueType() != MVT::i16 &&
5870 V.getValueType() != MVT::i32 &&
5871 V.getValueType() != MVT::i64)
5874 // Check the constant mask. Invert it so that the bits being masked out are
5875 // 0 and the bits being kept are 1. Use getSExtValue so that leading bits
5876 // follow the sign bit for uniformity.
5877 uint64_t NotMask = ~cast<ConstantSDNode>(V->getOperand(1))->getSExtValue();
5878 unsigned NotMaskLZ = CountLeadingZeros_64(NotMask);
5879 if (NotMaskLZ & 7) return Result; // Must be multiple of a byte.
5880 unsigned NotMaskTZ = CountTrailingZeros_64(NotMask);
5881 if (NotMaskTZ & 7) return Result; // Must be multiple of a byte.
5882 if (NotMaskLZ == 64) return Result; // All zero mask.
5884 // See if we have a continuous run of bits. If so, we have 0*1+0*
5885 if (CountTrailingOnes_64(NotMask >> NotMaskTZ)+NotMaskTZ+NotMaskLZ != 64)
5888 // Adjust NotMaskLZ down to be from the actual size of the int instead of i64.
5889 if (V.getValueType() != MVT::i64 && NotMaskLZ)
5890 NotMaskLZ -= 64-V.getValueSizeInBits();
5892 unsigned MaskedBytes = (V.getValueSizeInBits()-NotMaskLZ-NotMaskTZ)/8;
5893 switch (MaskedBytes) {
5897 default: return Result; // All one mask, or 5-byte mask.
5900 // Verify that the first bit starts at a multiple of mask so that the access
5901 // is aligned the same as the access width.
5902 if (NotMaskTZ && NotMaskTZ/8 % MaskedBytes) return Result;
5904 Result.first = MaskedBytes;
5905 Result.second = NotMaskTZ/8;
5910 /// ShrinkLoadReplaceStoreWithStore - Check to see if IVal is something that
5911 /// provides a value as specified by MaskInfo. If so, replace the specified
5912 /// store with a narrower store of truncated IVal.
5914 ShrinkLoadReplaceStoreWithStore(const std::pair<unsigned, unsigned> &MaskInfo,
5915 SDValue IVal, StoreSDNode *St,
5917 unsigned NumBytes = MaskInfo.first;
5918 unsigned ByteShift = MaskInfo.second;
5919 SelectionDAG &DAG = DC->getDAG();
5921 // Check to see if IVal is all zeros in the part being masked in by the 'or'
5922 // that uses this. If not, this is not a replacement.
5923 APInt Mask = ~APInt::getBitsSet(IVal.getValueSizeInBits(),
5924 ByteShift*8, (ByteShift+NumBytes)*8);
5925 if (!DAG.MaskedValueIsZero(IVal, Mask)) return 0;
5927 // Check that it is legal on the target to do this. It is legal if the new
5928 // VT we're shrinking to (i8/i16/i32) is legal or we're still before type
5930 MVT VT = MVT::getIntegerVT(NumBytes*8);
5931 if (!DC->isTypeLegal(VT))
5934 // Okay, we can do this! Replace the 'St' store with a store of IVal that is
5935 // shifted by ByteShift and truncated down to NumBytes.
5937 IVal = DAG.getNode(ISD::SRL, IVal->getDebugLoc(), IVal.getValueType(), IVal,
5938 DAG.getConstant(ByteShift*8, DC->getShiftAmountTy()));
5940 // Figure out the offset for the store and the alignment of the access.
5942 unsigned NewAlign = St->getAlignment();
5944 if (DAG.getTargetLoweringInfo().isLittleEndian())
5945 StOffset = ByteShift;
5947 StOffset = IVal.getValueType().getStoreSize() - ByteShift - NumBytes;
5949 SDValue Ptr = St->getBasePtr();
5951 Ptr = DAG.getNode(ISD::ADD, IVal->getDebugLoc(), Ptr.getValueType(),
5952 Ptr, DAG.getConstant(StOffset, Ptr.getValueType()));
5953 NewAlign = MinAlign(NewAlign, StOffset);
5956 // Truncate down to the new size.
5957 IVal = DAG.getNode(ISD::TRUNCATE, IVal->getDebugLoc(), VT, IVal);
5960 return DAG.getStore(St->getChain(), St->getDebugLoc(), IVal, Ptr,
5961 St->getPointerInfo().getWithOffset(StOffset),
5962 false, false, NewAlign).getNode();
5966 /// ReduceLoadOpStoreWidth - Look for sequence of load / op / store where op is
5967 /// one of 'or', 'xor', and 'and' of immediates. If 'op' is only touching some
5968 /// of the loaded bits, try narrowing the load and store if it would end up
5969 /// being a win for performance or code size.
5970 SDValue DAGCombiner::ReduceLoadOpStoreWidth(SDNode *N) {
5971 StoreSDNode *ST = cast<StoreSDNode>(N);
5972 if (ST->isVolatile())
5975 SDValue Chain = ST->getChain();
5976 SDValue Value = ST->getValue();
5977 SDValue Ptr = ST->getBasePtr();
5978 EVT VT = Value.getValueType();
5980 if (ST->isTruncatingStore() || VT.isVector() || !Value.hasOneUse())
5983 unsigned Opc = Value.getOpcode();
5985 // If this is "store (or X, Y), P" and X is "(and (load P), cst)", where cst
5986 // is a byte mask indicating a consecutive number of bytes, check to see if
5987 // Y is known to provide just those bytes. If so, we try to replace the
5988 // load + replace + store sequence with a single (narrower) store, which makes
5990 if (Opc == ISD::OR) {
5991 std::pair<unsigned, unsigned> MaskedLoad;
5992 MaskedLoad = CheckForMaskedLoad(Value.getOperand(0), Ptr, Chain);
5993 if (MaskedLoad.first)
5994 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
5995 Value.getOperand(1), ST,this))
5996 return SDValue(NewST, 0);
5998 // Or is commutative, so try swapping X and Y.
5999 MaskedLoad = CheckForMaskedLoad(Value.getOperand(1), Ptr, Chain);
6000 if (MaskedLoad.first)
6001 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
6002 Value.getOperand(0), ST,this))
6003 return SDValue(NewST, 0);
6006 if ((Opc != ISD::OR && Opc != ISD::XOR && Opc != ISD::AND) ||
6007 Value.getOperand(1).getOpcode() != ISD::Constant)
6010 SDValue N0 = Value.getOperand(0);
6011 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
6012 Chain == SDValue(N0.getNode(), 1)) {
6013 LoadSDNode *LD = cast<LoadSDNode>(N0);
6014 if (LD->getBasePtr() != Ptr ||
6015 LD->getPointerInfo().getAddrSpace() !=
6016 ST->getPointerInfo().getAddrSpace())
6019 // Find the type to narrow it the load / op / store to.
6020 SDValue N1 = Value.getOperand(1);
6021 unsigned BitWidth = N1.getValueSizeInBits();
6022 APInt Imm = cast<ConstantSDNode>(N1)->getAPIntValue();
6023 if (Opc == ISD::AND)
6024 Imm ^= APInt::getAllOnesValue(BitWidth);
6025 if (Imm == 0 || Imm.isAllOnesValue())
6027 unsigned ShAmt = Imm.countTrailingZeros();
6028 unsigned MSB = BitWidth - Imm.countLeadingZeros() - 1;
6029 unsigned NewBW = NextPowerOf2(MSB - ShAmt);
6030 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
6031 while (NewBW < BitWidth &&
6032 !(TLI.isOperationLegalOrCustom(Opc, NewVT) &&
6033 TLI.isNarrowingProfitable(VT, NewVT))) {
6034 NewBW = NextPowerOf2(NewBW);
6035 NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
6037 if (NewBW >= BitWidth)
6040 // If the lsb changed does not start at the type bitwidth boundary,
6041 // start at the previous one.
6043 ShAmt = (((ShAmt + NewBW - 1) / NewBW) * NewBW) - NewBW;
6044 APInt Mask = APInt::getBitsSet(BitWidth, ShAmt, ShAmt + NewBW);
6045 if ((Imm & Mask) == Imm) {
6046 APInt NewImm = (Imm & Mask).lshr(ShAmt).trunc(NewBW);
6047 if (Opc == ISD::AND)
6048 NewImm ^= APInt::getAllOnesValue(NewBW);
6049 uint64_t PtrOff = ShAmt / 8;
6050 // For big endian targets, we need to adjust the offset to the pointer to
6051 // load the correct bytes.
6052 if (TLI.isBigEndian())
6053 PtrOff = (BitWidth + 7 - NewBW) / 8 - PtrOff;
6055 unsigned NewAlign = MinAlign(LD->getAlignment(), PtrOff);
6056 const Type *NewVTTy = NewVT.getTypeForEVT(*DAG.getContext());
6057 if (NewAlign < TLI.getTargetData()->getABITypeAlignment(NewVTTy))
6060 SDValue NewPtr = DAG.getNode(ISD::ADD, LD->getDebugLoc(),
6061 Ptr.getValueType(), Ptr,
6062 DAG.getConstant(PtrOff, Ptr.getValueType()));
6063 SDValue NewLD = DAG.getLoad(NewVT, N0.getDebugLoc(),
6064 LD->getChain(), NewPtr,
6065 LD->getPointerInfo().getWithOffset(PtrOff),
6066 LD->isVolatile(), LD->isNonTemporal(),
6068 SDValue NewVal = DAG.getNode(Opc, Value.getDebugLoc(), NewVT, NewLD,
6069 DAG.getConstant(NewImm, NewVT));
6070 SDValue NewST = DAG.getStore(Chain, N->getDebugLoc(),
6072 ST->getPointerInfo().getWithOffset(PtrOff),
6073 false, false, NewAlign);
6075 AddToWorkList(NewPtr.getNode());
6076 AddToWorkList(NewLD.getNode());
6077 AddToWorkList(NewVal.getNode());
6078 WorkListRemover DeadNodes(*this);
6079 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLD.getValue(1),
6089 SDValue DAGCombiner::visitSTORE(SDNode *N) {
6090 StoreSDNode *ST = cast<StoreSDNode>(N);
6091 SDValue Chain = ST->getChain();
6092 SDValue Value = ST->getValue();
6093 SDValue Ptr = ST->getBasePtr();
6095 // If this is a store of a bit convert, store the input value if the
6096 // resultant store does not need a higher alignment than the original.
6097 if (Value.getOpcode() == ISD::BITCAST && !ST->isTruncatingStore() &&
6098 ST->isUnindexed()) {
6099 unsigned OrigAlign = ST->getAlignment();
6100 EVT SVT = Value.getOperand(0).getValueType();
6101 unsigned Align = TLI.getTargetData()->
6102 getABITypeAlignment(SVT.getTypeForEVT(*DAG.getContext()));
6103 if (Align <= OrigAlign &&
6104 ((!LegalOperations && !ST->isVolatile()) ||
6105 TLI.isOperationLegalOrCustom(ISD::STORE, SVT)))
6106 return DAG.getStore(Chain, N->getDebugLoc(), Value.getOperand(0),
6107 Ptr, ST->getPointerInfo(), ST->isVolatile(),
6108 ST->isNonTemporal(), OrigAlign);
6111 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
6112 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Value)) {
6113 // NOTE: If the original store is volatile, this transform must not increase
6114 // the number of stores. For example, on x86-32 an f64 can be stored in one
6115 // processor operation but an i64 (which is not legal) requires two. So the
6116 // transform should not be done in this case.
6117 if (Value.getOpcode() != ISD::TargetConstantFP) {
6119 switch (CFP->getValueType(0).getSimpleVT().SimpleTy) {
6120 default: llvm_unreachable("Unknown FP type");
6121 case MVT::f80: // We don't do this for these yet.
6126 if ((isTypeLegal(MVT::i32) && !LegalOperations && !ST->isVolatile()) ||
6127 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
6128 Tmp = DAG.getConstant((uint32_t)CFP->getValueAPF().
6129 bitcastToAPInt().getZExtValue(), MVT::i32);
6130 return DAG.getStore(Chain, N->getDebugLoc(), Tmp,
6131 Ptr, ST->getPointerInfo(), ST->isVolatile(),
6132 ST->isNonTemporal(), ST->getAlignment());
6136 if ((TLI.isTypeLegal(MVT::i64) && !LegalOperations &&
6137 !ST->isVolatile()) ||
6138 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i64)) {
6139 Tmp = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
6140 getZExtValue(), MVT::i64);
6141 return DAG.getStore(Chain, N->getDebugLoc(), Tmp,
6142 Ptr, ST->getPointerInfo(), ST->isVolatile(),
6143 ST->isNonTemporal(), ST->getAlignment());
6144 } else if (!ST->isVolatile() &&
6145 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
6146 // Many FP stores are not made apparent until after legalize, e.g. for
6147 // argument passing. Since this is so common, custom legalize the
6148 // 64-bit integer store into two 32-bit stores.
6149 uint64_t Val = CFP->getValueAPF().bitcastToAPInt().getZExtValue();
6150 SDValue Lo = DAG.getConstant(Val & 0xFFFFFFFF, MVT::i32);
6151 SDValue Hi = DAG.getConstant(Val >> 32, MVT::i32);
6152 if (TLI.isBigEndian()) std::swap(Lo, Hi);
6154 unsigned Alignment = ST->getAlignment();
6155 bool isVolatile = ST->isVolatile();
6156 bool isNonTemporal = ST->isNonTemporal();
6158 SDValue St0 = DAG.getStore(Chain, ST->getDebugLoc(), Lo,
6159 Ptr, ST->getPointerInfo(),
6160 isVolatile, isNonTemporal,
6161 ST->getAlignment());
6162 Ptr = DAG.getNode(ISD::ADD, N->getDebugLoc(), Ptr.getValueType(), Ptr,
6163 DAG.getConstant(4, Ptr.getValueType()));
6164 Alignment = MinAlign(Alignment, 4U);
6165 SDValue St1 = DAG.getStore(Chain, ST->getDebugLoc(), Hi,
6166 Ptr, ST->getPointerInfo().getWithOffset(4),
6167 isVolatile, isNonTemporal,
6169 return DAG.getNode(ISD::TokenFactor, N->getDebugLoc(), MVT::Other,
6178 // Try to infer better alignment information than the store already has.
6179 if (OptLevel != CodeGenOpt::None && ST->isUnindexed()) {
6180 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
6181 if (Align > ST->getAlignment())
6182 return DAG.getTruncStore(Chain, N->getDebugLoc(), Value,
6183 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
6184 ST->isVolatile(), ST->isNonTemporal(), Align);
6189 // Walk up chain skipping non-aliasing memory nodes.
6190 SDValue BetterChain = FindBetterChain(N, Chain);
6192 // If there is a better chain.
6193 if (Chain != BetterChain) {
6196 // Replace the chain to avoid dependency.
6197 if (ST->isTruncatingStore()) {
6198 ReplStore = DAG.getTruncStore(BetterChain, N->getDebugLoc(), Value, Ptr,
6199 ST->getPointerInfo(),
6200 ST->getMemoryVT(), ST->isVolatile(),
6201 ST->isNonTemporal(), ST->getAlignment());
6203 ReplStore = DAG.getStore(BetterChain, N->getDebugLoc(), Value, Ptr,
6204 ST->getPointerInfo(),
6205 ST->isVolatile(), ST->isNonTemporal(),
6206 ST->getAlignment());
6209 // Create token to keep both nodes around.
6210 SDValue Token = DAG.getNode(ISD::TokenFactor, N->getDebugLoc(),
6211 MVT::Other, Chain, ReplStore);
6213 // Make sure the new and old chains are cleaned up.
6214 AddToWorkList(Token.getNode());
6216 // Don't add users to work list.
6217 return CombineTo(N, Token, false);
6221 // Try transforming N to an indexed store.
6222 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
6223 return SDValue(N, 0);
6225 // FIXME: is there such a thing as a truncating indexed store?
6226 if (ST->isTruncatingStore() && ST->isUnindexed() &&
6227 Value.getValueType().isInteger()) {
6228 // See if we can simplify the input to this truncstore with knowledge that
6229 // only the low bits are being used. For example:
6230 // "truncstore (or (shl x, 8), y), i8" -> "truncstore y, i8"
6232 GetDemandedBits(Value,
6233 APInt::getLowBitsSet(Value.getValueSizeInBits(),
6234 ST->getMemoryVT().getSizeInBits()));
6235 AddToWorkList(Value.getNode());
6236 if (Shorter.getNode())
6237 return DAG.getTruncStore(Chain, N->getDebugLoc(), Shorter,
6238 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
6239 ST->isVolatile(), ST->isNonTemporal(),
6240 ST->getAlignment());
6242 // Otherwise, see if we can simplify the operation with
6243 // SimplifyDemandedBits, which only works if the value has a single use.
6244 if (SimplifyDemandedBits(Value,
6245 APInt::getLowBitsSet(
6246 Value.getValueType().getScalarType().getSizeInBits(),
6247 ST->getMemoryVT().getScalarType().getSizeInBits())))
6248 return SDValue(N, 0);
6251 // If this is a load followed by a store to the same location, then the store
6253 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Value)) {
6254 if (Ld->getBasePtr() == Ptr && ST->getMemoryVT() == Ld->getMemoryVT() &&
6255 ST->isUnindexed() && !ST->isVolatile() &&
6256 // There can't be any side effects between the load and store, such as
6258 Chain.reachesChainWithoutSideEffects(SDValue(Ld, 1))) {
6259 // The store is dead, remove it.
6264 // If this is an FP_ROUND or TRUNC followed by a store, fold this into a
6265 // truncating store. We can do this even if this is already a truncstore.
6266 if ((Value.getOpcode() == ISD::FP_ROUND || Value.getOpcode() == ISD::TRUNCATE)
6267 && Value.getNode()->hasOneUse() && ST->isUnindexed() &&
6268 TLI.isTruncStoreLegal(Value.getOperand(0).getValueType(),
6269 ST->getMemoryVT())) {
6270 return DAG.getTruncStore(Chain, N->getDebugLoc(), Value.getOperand(0),
6271 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
6272 ST->isVolatile(), ST->isNonTemporal(),
6273 ST->getAlignment());
6276 return ReduceLoadOpStoreWidth(N);
6279 SDValue DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) {
6280 SDValue InVec = N->getOperand(0);
6281 SDValue InVal = N->getOperand(1);
6282 SDValue EltNo = N->getOperand(2);
6284 // If the inserted element is an UNDEF, just use the input vector.
6285 if (InVal.getOpcode() == ISD::UNDEF)
6288 // If the invec is a BUILD_VECTOR and if EltNo is a constant, build a new
6289 // vector with the inserted element.
6290 if (InVec.getOpcode() == ISD::BUILD_VECTOR && isa<ConstantSDNode>(EltNo)) {
6291 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
6292 SmallVector<SDValue, 8> Ops(InVec.getNode()->op_begin(),
6293 InVec.getNode()->op_end());
6294 if (Elt < Ops.size())
6296 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
6297 InVec.getValueType(), &Ops[0], Ops.size());
6299 // If the invec is an UNDEF and if EltNo is a constant, create a new
6300 // BUILD_VECTOR with undef elements and the inserted element.
6301 if (!LegalOperations && InVec.getOpcode() == ISD::UNDEF &&
6302 isa<ConstantSDNode>(EltNo)) {
6303 EVT VT = InVec.getValueType();
6304 EVT EltVT = VT.getVectorElementType();
6305 unsigned NElts = VT.getVectorNumElements();
6306 SmallVector<SDValue, 8> Ops(NElts, DAG.getUNDEF(EltVT));
6308 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
6309 if (Elt < Ops.size())
6311 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
6312 InVec.getValueType(), &Ops[0], Ops.size());
6317 SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) {
6318 // (vextract (scalar_to_vector val, 0) -> val
6319 SDValue InVec = N->getOperand(0);
6321 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
6322 // Check if the result type doesn't match the inserted element type. A
6323 // SCALAR_TO_VECTOR may truncate the inserted element and the
6324 // EXTRACT_VECTOR_ELT may widen the extracted vector.
6325 SDValue InOp = InVec.getOperand(0);
6326 EVT NVT = N->getValueType(0);
6327 if (InOp.getValueType() != NVT) {
6328 assert(InOp.getValueType().isInteger() && NVT.isInteger());
6329 return DAG.getSExtOrTrunc(InOp, InVec.getDebugLoc(), NVT);
6334 // Perform only after legalization to ensure build_vector / vector_shuffle
6335 // optimizations have already been done.
6336 if (!LegalOperations) return SDValue();
6338 // (vextract (v4f32 load $addr), c) -> (f32 load $addr+c*size)
6339 // (vextract (v4f32 s2v (f32 load $addr)), c) -> (f32 load $addr+c*size)
6340 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), 0) -> (f32 load $addr)
6341 SDValue EltNo = N->getOperand(1);
6343 if (isa<ConstantSDNode>(EltNo)) {
6344 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
6345 bool NewLoad = false;
6346 bool BCNumEltsChanged = false;
6347 EVT VT = InVec.getValueType();
6348 EVT ExtVT = VT.getVectorElementType();
6351 if (InVec.getOpcode() == ISD::BITCAST) {
6352 EVT BCVT = InVec.getOperand(0).getValueType();
6353 if (!BCVT.isVector() || ExtVT.bitsGT(BCVT.getVectorElementType()))
6355 if (VT.getVectorNumElements() != BCVT.getVectorNumElements())
6356 BCNumEltsChanged = true;
6357 InVec = InVec.getOperand(0);
6358 ExtVT = BCVT.getVectorElementType();
6362 LoadSDNode *LN0 = NULL;
6363 const ShuffleVectorSDNode *SVN = NULL;
6364 if (ISD::isNormalLoad(InVec.getNode())) {
6365 LN0 = cast<LoadSDNode>(InVec);
6366 } else if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR &&
6367 InVec.getOperand(0).getValueType() == ExtVT &&
6368 ISD::isNormalLoad(InVec.getOperand(0).getNode())) {
6369 LN0 = cast<LoadSDNode>(InVec.getOperand(0));
6370 } else if ((SVN = dyn_cast<ShuffleVectorSDNode>(InVec))) {
6371 // (vextract (vector_shuffle (load $addr), v2, <1, u, u, u>), 1)
6373 // (load $addr+1*size)
6375 // If the bit convert changed the number of elements, it is unsafe
6376 // to examine the mask.
6377 if (BCNumEltsChanged)
6380 // Select the input vector, guarding against out of range extract vector.
6381 unsigned NumElems = VT.getVectorNumElements();
6382 int Idx = (Elt > (int)NumElems) ? -1 : SVN->getMaskElt(Elt);
6383 InVec = (Idx < (int)NumElems) ? InVec.getOperand(0) : InVec.getOperand(1);
6385 if (InVec.getOpcode() == ISD::BITCAST)
6386 InVec = InVec.getOperand(0);
6387 if (ISD::isNormalLoad(InVec.getNode())) {
6388 LN0 = cast<LoadSDNode>(InVec);
6389 Elt = (Idx < (int)NumElems) ? Idx : Idx - (int)NumElems;
6393 if (!LN0 || !LN0->hasOneUse() || LN0->isVolatile())
6396 // If Idx was -1 above, Elt is going to be -1, so just return undef.
6398 return DAG.getUNDEF(LN0->getBasePtr().getValueType());
6400 unsigned Align = LN0->getAlignment();
6402 // Check the resultant load doesn't need a higher alignment than the
6406 ->getABITypeAlignment(LVT.getTypeForEVT(*DAG.getContext()));
6408 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, LVT))
6414 SDValue NewPtr = LN0->getBasePtr();
6415 unsigned PtrOff = 0;
6418 PtrOff = LVT.getSizeInBits() * Elt / 8;
6419 EVT PtrType = NewPtr.getValueType();
6420 if (TLI.isBigEndian())
6421 PtrOff = VT.getSizeInBits() / 8 - PtrOff;
6422 NewPtr = DAG.getNode(ISD::ADD, N->getDebugLoc(), PtrType, NewPtr,
6423 DAG.getConstant(PtrOff, PtrType));
6426 return DAG.getLoad(LVT, N->getDebugLoc(), LN0->getChain(), NewPtr,
6427 LN0->getPointerInfo().getWithOffset(PtrOff),
6428 LN0->isVolatile(), LN0->isNonTemporal(), Align);
6434 SDValue DAGCombiner::visitBUILD_VECTOR(SDNode *N) {
6435 unsigned NumInScalars = N->getNumOperands();
6436 EVT VT = N->getValueType(0);
6438 // Check to see if this is a BUILD_VECTOR of a bunch of EXTRACT_VECTOR_ELT
6439 // operations. If so, and if the EXTRACT_VECTOR_ELT vector inputs come from
6440 // at most two distinct vectors, turn this into a shuffle node.
6441 SDValue VecIn1, VecIn2;
6442 for (unsigned i = 0; i != NumInScalars; ++i) {
6443 // Ignore undef inputs.
6444 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
6446 // If this input is something other than a EXTRACT_VECTOR_ELT with a
6447 // constant index, bail out.
6448 if (N->getOperand(i).getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
6449 !isa<ConstantSDNode>(N->getOperand(i).getOperand(1))) {
6450 VecIn1 = VecIn2 = SDValue(0, 0);
6454 // If the input vector type disagrees with the result of the build_vector,
6455 // we can't make a shuffle.
6456 SDValue ExtractedFromVec = N->getOperand(i).getOperand(0);
6457 if (ExtractedFromVec.getValueType() != VT) {
6458 VecIn1 = VecIn2 = SDValue(0, 0);
6462 // Otherwise, remember this. We allow up to two distinct input vectors.
6463 if (ExtractedFromVec == VecIn1 || ExtractedFromVec == VecIn2)
6466 if (VecIn1.getNode() == 0) {
6467 VecIn1 = ExtractedFromVec;
6468 } else if (VecIn2.getNode() == 0) {
6469 VecIn2 = ExtractedFromVec;
6472 VecIn1 = VecIn2 = SDValue(0, 0);
6477 // If everything is good, we can make a shuffle operation.
6478 if (VecIn1.getNode()) {
6479 SmallVector<int, 8> Mask;
6480 for (unsigned i = 0; i != NumInScalars; ++i) {
6481 if (N->getOperand(i).getOpcode() == ISD::UNDEF) {
6486 // If extracting from the first vector, just use the index directly.
6487 SDValue Extract = N->getOperand(i);
6488 SDValue ExtVal = Extract.getOperand(1);
6489 if (Extract.getOperand(0) == VecIn1) {
6490 unsigned ExtIndex = cast<ConstantSDNode>(ExtVal)->getZExtValue();
6491 if (ExtIndex > VT.getVectorNumElements())
6494 Mask.push_back(ExtIndex);
6498 // Otherwise, use InIdx + VecSize
6499 unsigned Idx = cast<ConstantSDNode>(ExtVal)->getZExtValue();
6500 Mask.push_back(Idx+NumInScalars);
6503 // Add count and size info.
6504 if (!isTypeLegal(VT))
6507 // Return the new VECTOR_SHUFFLE node.
6510 Ops[1] = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
6511 return DAG.getVectorShuffle(VT, N->getDebugLoc(), Ops[0], Ops[1], &Mask[0]);
6517 SDValue DAGCombiner::visitCONCAT_VECTORS(SDNode *N) {
6518 // TODO: Check to see if this is a CONCAT_VECTORS of a bunch of
6519 // EXTRACT_SUBVECTOR operations. If so, and if the EXTRACT_SUBVECTOR vector
6520 // inputs come from at most two distinct vectors, turn this into a shuffle
6523 // If we only have one input vector, we don't need to do any concatenation.
6524 if (N->getNumOperands() == 1)
6525 return N->getOperand(0);
6530 SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
6531 EVT VT = N->getValueType(0);
6532 unsigned NumElts = VT.getVectorNumElements();
6534 SDValue N0 = N->getOperand(0);
6536 assert(N0.getValueType().getVectorNumElements() == NumElts &&
6537 "Vector shuffle must be normalized in DAG");
6539 // FIXME: implement canonicalizations from DAG.getVectorShuffle()
6541 // If it is a splat, check if the argument vector is another splat or a
6542 // build_vector with all scalar elements the same.
6543 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
6544 if (SVN->isSplat() && SVN->getSplatIndex() < (int)NumElts) {
6545 SDNode *V = N0.getNode();
6547 // If this is a bit convert that changes the element type of the vector but
6548 // not the number of vector elements, look through it. Be careful not to
6549 // look though conversions that change things like v4f32 to v2f64.
6550 if (V->getOpcode() == ISD::BITCAST) {
6551 SDValue ConvInput = V->getOperand(0);
6552 if (ConvInput.getValueType().isVector() &&
6553 ConvInput.getValueType().getVectorNumElements() == NumElts)
6554 V = ConvInput.getNode();
6557 if (V->getOpcode() == ISD::BUILD_VECTOR) {
6558 assert(V->getNumOperands() == NumElts &&
6559 "BUILD_VECTOR has wrong number of operands");
6561 bool AllSame = true;
6562 for (unsigned i = 0; i != NumElts; ++i) {
6563 if (V->getOperand(i).getOpcode() != ISD::UNDEF) {
6564 Base = V->getOperand(i);
6568 // Splat of <u, u, u, u>, return <u, u, u, u>
6569 if (!Base.getNode())
6571 for (unsigned i = 0; i != NumElts; ++i) {
6572 if (V->getOperand(i) != Base) {
6577 // Splat of <x, x, x, x>, return <x, x, x, x>
6585 SDValue DAGCombiner::visitMEMBARRIER(SDNode* N) {
6586 if (!TLI.getShouldFoldAtomicFences())
6589 SDValue atomic = N->getOperand(0);
6590 switch (atomic.getOpcode()) {
6591 case ISD::ATOMIC_CMP_SWAP:
6592 case ISD::ATOMIC_SWAP:
6593 case ISD::ATOMIC_LOAD_ADD:
6594 case ISD::ATOMIC_LOAD_SUB:
6595 case ISD::ATOMIC_LOAD_AND:
6596 case ISD::ATOMIC_LOAD_OR:
6597 case ISD::ATOMIC_LOAD_XOR:
6598 case ISD::ATOMIC_LOAD_NAND:
6599 case ISD::ATOMIC_LOAD_MIN:
6600 case ISD::ATOMIC_LOAD_MAX:
6601 case ISD::ATOMIC_LOAD_UMIN:
6602 case ISD::ATOMIC_LOAD_UMAX:
6608 SDValue fence = atomic.getOperand(0);
6609 if (fence.getOpcode() != ISD::MEMBARRIER)
6612 switch (atomic.getOpcode()) {
6613 case ISD::ATOMIC_CMP_SWAP:
6614 return SDValue(DAG.UpdateNodeOperands(atomic.getNode(),
6615 fence.getOperand(0),
6616 atomic.getOperand(1), atomic.getOperand(2),
6617 atomic.getOperand(3)), atomic.getResNo());
6618 case ISD::ATOMIC_SWAP:
6619 case ISD::ATOMIC_LOAD_ADD:
6620 case ISD::ATOMIC_LOAD_SUB:
6621 case ISD::ATOMIC_LOAD_AND:
6622 case ISD::ATOMIC_LOAD_OR:
6623 case ISD::ATOMIC_LOAD_XOR:
6624 case ISD::ATOMIC_LOAD_NAND:
6625 case ISD::ATOMIC_LOAD_MIN:
6626 case ISD::ATOMIC_LOAD_MAX:
6627 case ISD::ATOMIC_LOAD_UMIN:
6628 case ISD::ATOMIC_LOAD_UMAX:
6629 return SDValue(DAG.UpdateNodeOperands(atomic.getNode(),
6630 fence.getOperand(0),
6631 atomic.getOperand(1), atomic.getOperand(2)),
6638 /// XformToShuffleWithZero - Returns a vector_shuffle if it able to transform
6639 /// an AND to a vector_shuffle with the destination vector and a zero vector.
6640 /// e.g. AND V, <0xffffffff, 0, 0xffffffff, 0>. ==>
6641 /// vector_shuffle V, Zero, <0, 4, 2, 4>
6642 SDValue DAGCombiner::XformToShuffleWithZero(SDNode *N) {
6643 EVT VT = N->getValueType(0);
6644 DebugLoc dl = N->getDebugLoc();
6645 SDValue LHS = N->getOperand(0);
6646 SDValue RHS = N->getOperand(1);
6647 if (N->getOpcode() == ISD::AND) {
6648 if (RHS.getOpcode() == ISD::BITCAST)
6649 RHS = RHS.getOperand(0);
6650 if (RHS.getOpcode() == ISD::BUILD_VECTOR) {
6651 SmallVector<int, 8> Indices;
6652 unsigned NumElts = RHS.getNumOperands();
6653 for (unsigned i = 0; i != NumElts; ++i) {
6654 SDValue Elt = RHS.getOperand(i);
6655 if (!isa<ConstantSDNode>(Elt))
6657 else if (cast<ConstantSDNode>(Elt)->isAllOnesValue())
6658 Indices.push_back(i);
6659 else if (cast<ConstantSDNode>(Elt)->isNullValue())
6660 Indices.push_back(NumElts);
6665 // Let's see if the target supports this vector_shuffle.
6666 EVT RVT = RHS.getValueType();
6667 if (!TLI.isVectorClearMaskLegal(Indices, RVT))
6670 // Return the new VECTOR_SHUFFLE node.
6671 EVT EltVT = RVT.getVectorElementType();
6672 SmallVector<SDValue,8> ZeroOps(RVT.getVectorNumElements(),
6673 DAG.getConstant(0, EltVT));
6674 SDValue Zero = DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
6675 RVT, &ZeroOps[0], ZeroOps.size());
6676 LHS = DAG.getNode(ISD::BITCAST, dl, RVT, LHS);
6677 SDValue Shuf = DAG.getVectorShuffle(RVT, dl, LHS, Zero, &Indices[0]);
6678 return DAG.getNode(ISD::BITCAST, dl, VT, Shuf);
6685 /// SimplifyVBinOp - Visit a binary vector operation, like ADD.
6686 SDValue DAGCombiner::SimplifyVBinOp(SDNode *N) {
6687 // After legalize, the target may be depending on adds and other
6688 // binary ops to provide legal ways to construct constants or other
6689 // things. Simplifying them may result in a loss of legality.
6690 if (LegalOperations) return SDValue();
6692 assert(N->getValueType(0).isVector() &&
6693 "SimplifyVBinOp only works on vectors!");
6695 SDValue LHS = N->getOperand(0);
6696 SDValue RHS = N->getOperand(1);
6697 SDValue Shuffle = XformToShuffleWithZero(N);
6698 if (Shuffle.getNode()) return Shuffle;
6700 // If the LHS and RHS are BUILD_VECTOR nodes, see if we can constant fold
6702 if (LHS.getOpcode() == ISD::BUILD_VECTOR &&
6703 RHS.getOpcode() == ISD::BUILD_VECTOR) {
6704 SmallVector<SDValue, 8> Ops;
6705 for (unsigned i = 0, e = LHS.getNumOperands(); i != e; ++i) {
6706 SDValue LHSOp = LHS.getOperand(i);
6707 SDValue RHSOp = RHS.getOperand(i);
6708 // If these two elements can't be folded, bail out.
6709 if ((LHSOp.getOpcode() != ISD::UNDEF &&
6710 LHSOp.getOpcode() != ISD::Constant &&
6711 LHSOp.getOpcode() != ISD::ConstantFP) ||
6712 (RHSOp.getOpcode() != ISD::UNDEF &&
6713 RHSOp.getOpcode() != ISD::Constant &&
6714 RHSOp.getOpcode() != ISD::ConstantFP))
6717 // Can't fold divide by zero.
6718 if (N->getOpcode() == ISD::SDIV || N->getOpcode() == ISD::UDIV ||
6719 N->getOpcode() == ISD::FDIV) {
6720 if ((RHSOp.getOpcode() == ISD::Constant &&
6721 cast<ConstantSDNode>(RHSOp.getNode())->isNullValue()) ||
6722 (RHSOp.getOpcode() == ISD::ConstantFP &&
6723 cast<ConstantFPSDNode>(RHSOp.getNode())->getValueAPF().isZero()))
6727 EVT VT = LHSOp.getValueType();
6728 assert(RHSOp.getValueType() == VT &&
6729 "SimplifyVBinOp with different BUILD_VECTOR element types");
6730 SDValue FoldOp = DAG.getNode(N->getOpcode(), LHS.getDebugLoc(), VT,
6732 if (FoldOp.getOpcode() != ISD::UNDEF &&
6733 FoldOp.getOpcode() != ISD::Constant &&
6734 FoldOp.getOpcode() != ISD::ConstantFP)
6736 Ops.push_back(FoldOp);
6737 AddToWorkList(FoldOp.getNode());
6740 if (Ops.size() == LHS.getNumOperands())
6741 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
6742 LHS.getValueType(), &Ops[0], Ops.size());
6748 SDValue DAGCombiner::SimplifySelect(DebugLoc DL, SDValue N0,
6749 SDValue N1, SDValue N2){
6750 assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!");
6752 SDValue SCC = SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1), N1, N2,
6753 cast<CondCodeSDNode>(N0.getOperand(2))->get());
6755 // If we got a simplified select_cc node back from SimplifySelectCC, then
6756 // break it down into a new SETCC node, and a new SELECT node, and then return
6757 // the SELECT node, since we were called with a SELECT node.
6758 if (SCC.getNode()) {
6759 // Check to see if we got a select_cc back (to turn into setcc/select).
6760 // Otherwise, just return whatever node we got back, like fabs.
6761 if (SCC.getOpcode() == ISD::SELECT_CC) {
6762 SDValue SETCC = DAG.getNode(ISD::SETCC, N0.getDebugLoc(),
6764 SCC.getOperand(0), SCC.getOperand(1),
6766 AddToWorkList(SETCC.getNode());
6767 return DAG.getNode(ISD::SELECT, SCC.getDebugLoc(), SCC.getValueType(),
6768 SCC.getOperand(2), SCC.getOperand(3), SETCC);
6776 /// SimplifySelectOps - Given a SELECT or a SELECT_CC node, where LHS and RHS
6777 /// are the two values being selected between, see if we can simplify the
6778 /// select. Callers of this should assume that TheSelect is deleted if this
6779 /// returns true. As such, they should return the appropriate thing (e.g. the
6780 /// node) back to the top-level of the DAG combiner loop to avoid it being
6782 bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDValue LHS,
6785 // If this is a select from two identical things, try to pull the operation
6786 // through the select.
6787 if (LHS.getOpcode() != RHS.getOpcode() ||
6788 !LHS.hasOneUse() || !RHS.hasOneUse())
6791 // If this is a load and the token chain is identical, replace the select
6792 // of two loads with a load through a select of the address to load from.
6793 // This triggers in things like "select bool X, 10.0, 123.0" after the FP
6794 // constants have been dropped into the constant pool.
6795 if (LHS.getOpcode() == ISD::LOAD) {
6796 LoadSDNode *LLD = cast<LoadSDNode>(LHS);
6797 LoadSDNode *RLD = cast<LoadSDNode>(RHS);
6799 // Token chains must be identical.
6800 if (LHS.getOperand(0) != RHS.getOperand(0) ||
6801 // Do not let this transformation reduce the number of volatile loads.
6802 LLD->isVolatile() || RLD->isVolatile() ||
6803 // If this is an EXTLOAD, the VT's must match.
6804 LLD->getMemoryVT() != RLD->getMemoryVT() ||
6805 // If this is an EXTLOAD, the kind of extension must match.
6806 (LLD->getExtensionType() != RLD->getExtensionType() &&
6807 // The only exception is if one of the extensions is anyext.
6808 LLD->getExtensionType() != ISD::EXTLOAD &&
6809 RLD->getExtensionType() != ISD::EXTLOAD) ||
6810 // FIXME: this discards src value information. This is
6811 // over-conservative. It would be beneficial to be able to remember
6812 // both potential memory locations. Since we are discarding
6813 // src value info, don't do the transformation if the memory
6814 // locations are not in the default address space.
6815 LLD->getPointerInfo().getAddrSpace() != 0 ||
6816 RLD->getPointerInfo().getAddrSpace() != 0)
6819 // Check that the select condition doesn't reach either load. If so,
6820 // folding this will induce a cycle into the DAG. If not, this is safe to
6821 // xform, so create a select of the addresses.
6823 if (TheSelect->getOpcode() == ISD::SELECT) {
6824 SDNode *CondNode = TheSelect->getOperand(0).getNode();
6825 if ((LLD->hasAnyUseOfValue(1) && LLD->isPredecessorOf(CondNode)) ||
6826 (RLD->hasAnyUseOfValue(1) && RLD->isPredecessorOf(CondNode)))
6828 Addr = DAG.getNode(ISD::SELECT, TheSelect->getDebugLoc(),
6829 LLD->getBasePtr().getValueType(),
6830 TheSelect->getOperand(0), LLD->getBasePtr(),
6832 } else { // Otherwise SELECT_CC
6833 SDNode *CondLHS = TheSelect->getOperand(0).getNode();
6834 SDNode *CondRHS = TheSelect->getOperand(1).getNode();
6836 if ((LLD->hasAnyUseOfValue(1) &&
6837 (LLD->isPredecessorOf(CondLHS) || LLD->isPredecessorOf(CondRHS))) ||
6838 (LLD->hasAnyUseOfValue(1) &&
6839 (LLD->isPredecessorOf(CondLHS) || LLD->isPredecessorOf(CondRHS))))
6842 Addr = DAG.getNode(ISD::SELECT_CC, TheSelect->getDebugLoc(),
6843 LLD->getBasePtr().getValueType(),
6844 TheSelect->getOperand(0),
6845 TheSelect->getOperand(1),
6846 LLD->getBasePtr(), RLD->getBasePtr(),
6847 TheSelect->getOperand(4));
6851 if (LLD->getExtensionType() == ISD::NON_EXTLOAD) {
6852 Load = DAG.getLoad(TheSelect->getValueType(0),
6853 TheSelect->getDebugLoc(),
6854 // FIXME: Discards pointer info.
6855 LLD->getChain(), Addr, MachinePointerInfo(),
6856 LLD->isVolatile(), LLD->isNonTemporal(),
6857 LLD->getAlignment());
6859 Load = DAG.getExtLoad(LLD->getExtensionType() == ISD::EXTLOAD ?
6860 RLD->getExtensionType() : LLD->getExtensionType(),
6861 TheSelect->getValueType(0),
6862 TheSelect->getDebugLoc(),
6863 // FIXME: Discards pointer info.
6864 LLD->getChain(), Addr, MachinePointerInfo(),
6865 LLD->getMemoryVT(), LLD->isVolatile(),
6866 LLD->isNonTemporal(), LLD->getAlignment());
6869 // Users of the select now use the result of the load.
6870 CombineTo(TheSelect, Load);
6872 // Users of the old loads now use the new load's chain. We know the
6873 // old-load value is dead now.
6874 CombineTo(LHS.getNode(), Load.getValue(0), Load.getValue(1));
6875 CombineTo(RHS.getNode(), Load.getValue(0), Load.getValue(1));
6882 /// SimplifySelectCC - Simplify an expression of the form (N0 cond N1) ? N2 : N3
6883 /// where 'cond' is the comparison specified by CC.
6884 SDValue DAGCombiner::SimplifySelectCC(DebugLoc DL, SDValue N0, SDValue N1,
6885 SDValue N2, SDValue N3,
6886 ISD::CondCode CC, bool NotExtCompare) {
6887 // (x ? y : y) -> y.
6888 if (N2 == N3) return N2;
6890 EVT VT = N2.getValueType();
6891 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
6892 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
6893 ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N3.getNode());
6895 // Determine if the condition we're dealing with is constant
6896 SDValue SCC = SimplifySetCC(TLI.getSetCCResultType(N0.getValueType()),
6897 N0, N1, CC, DL, false);
6898 if (SCC.getNode()) AddToWorkList(SCC.getNode());
6899 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode());
6901 // fold select_cc true, x, y -> x
6902 if (SCCC && !SCCC->isNullValue())
6904 // fold select_cc false, x, y -> y
6905 if (SCCC && SCCC->isNullValue())
6908 // Check to see if we can simplify the select into an fabs node
6909 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) {
6910 // Allow either -0.0 or 0.0
6911 if (CFP->getValueAPF().isZero()) {
6912 // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs
6913 if ((CC == ISD::SETGE || CC == ISD::SETGT) &&
6914 N0 == N2 && N3.getOpcode() == ISD::FNEG &&
6915 N2 == N3.getOperand(0))
6916 return DAG.getNode(ISD::FABS, DL, VT, N0);
6918 // select (setl[te] X, +/-0.0), fneg(X), X -> fabs
6919 if ((CC == ISD::SETLT || CC == ISD::SETLE) &&
6920 N0 == N3 && N2.getOpcode() == ISD::FNEG &&
6921 N2.getOperand(0) == N3)
6922 return DAG.getNode(ISD::FABS, DL, VT, N3);
6926 // Turn "(a cond b) ? 1.0f : 2.0f" into "load (tmp + ((a cond b) ? 0 : 4)"
6927 // where "tmp" is a constant pool entry containing an array with 1.0 and 2.0
6928 // in it. This is a win when the constant is not otherwise available because
6929 // it replaces two constant pool loads with one. We only do this if the FP
6930 // type is known to be legal, because if it isn't, then we are before legalize
6931 // types an we want the other legalization to happen first (e.g. to avoid
6932 // messing with soft float) and if the ConstantFP is not legal, because if
6933 // it is legal, we may not need to store the FP constant in a constant pool.
6934 if (ConstantFPSDNode *TV = dyn_cast<ConstantFPSDNode>(N2))
6935 if (ConstantFPSDNode *FV = dyn_cast<ConstantFPSDNode>(N3)) {
6936 if (TLI.isTypeLegal(N2.getValueType()) &&
6937 (TLI.getOperationAction(ISD::ConstantFP, N2.getValueType()) !=
6938 TargetLowering::Legal) &&
6939 // If both constants have multiple uses, then we won't need to do an
6940 // extra load, they are likely around in registers for other users.
6941 (TV->hasOneUse() || FV->hasOneUse())) {
6942 Constant *Elts[] = {
6943 const_cast<ConstantFP*>(FV->getConstantFPValue()),
6944 const_cast<ConstantFP*>(TV->getConstantFPValue())
6946 const Type *FPTy = Elts[0]->getType();
6947 const TargetData &TD = *TLI.getTargetData();
6949 // Create a ConstantArray of the two constants.
6950 Constant *CA = ConstantArray::get(ArrayType::get(FPTy, 2), Elts, 2);
6951 SDValue CPIdx = DAG.getConstantPool(CA, TLI.getPointerTy(),
6952 TD.getPrefTypeAlignment(FPTy));
6953 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
6955 // Get the offsets to the 0 and 1 element of the array so that we can
6956 // select between them.
6957 SDValue Zero = DAG.getIntPtrConstant(0);
6958 unsigned EltSize = (unsigned)TD.getTypeAllocSize(Elts[0]->getType());
6959 SDValue One = DAG.getIntPtrConstant(EltSize);
6961 SDValue Cond = DAG.getSetCC(DL,
6962 TLI.getSetCCResultType(N0.getValueType()),
6964 SDValue CstOffset = DAG.getNode(ISD::SELECT, DL, Zero.getValueType(),
6966 CPIdx = DAG.getNode(ISD::ADD, DL, TLI.getPointerTy(), CPIdx,
6968 return DAG.getLoad(TV->getValueType(0), DL, DAG.getEntryNode(), CPIdx,
6969 MachinePointerInfo::getConstantPool(), false,
6975 // Check to see if we can perform the "gzip trick", transforming
6976 // (select_cc setlt X, 0, A, 0) -> (and (sra X, (sub size(X), 1), A)
6977 if (N1C && N3C && N3C->isNullValue() && CC == ISD::SETLT &&
6978 N0.getValueType().isInteger() &&
6979 N2.getValueType().isInteger() &&
6980 (N1C->isNullValue() || // (a < 0) ? b : 0
6981 (N1C->getAPIntValue() == 1 && N0 == N2))) { // (a < 1) ? a : 0
6982 EVT XType = N0.getValueType();
6983 EVT AType = N2.getValueType();
6984 if (XType.bitsGE(AType)) {
6985 // and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a
6986 // single-bit constant.
6987 if (N2C && ((N2C->getAPIntValue() & (N2C->getAPIntValue()-1)) == 0)) {
6988 unsigned ShCtV = N2C->getAPIntValue().logBase2();
6989 ShCtV = XType.getSizeInBits()-ShCtV-1;
6990 SDValue ShCt = DAG.getConstant(ShCtV, getShiftAmountTy());
6991 SDValue Shift = DAG.getNode(ISD::SRL, N0.getDebugLoc(),
6993 AddToWorkList(Shift.getNode());
6995 if (XType.bitsGT(AType)) {
6996 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
6997 AddToWorkList(Shift.getNode());
7000 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
7003 SDValue Shift = DAG.getNode(ISD::SRA, N0.getDebugLoc(),
7005 DAG.getConstant(XType.getSizeInBits()-1,
7006 getShiftAmountTy()));
7007 AddToWorkList(Shift.getNode());
7009 if (XType.bitsGT(AType)) {
7010 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
7011 AddToWorkList(Shift.getNode());
7014 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
7018 // fold (select_cc seteq (and x, y), 0, 0, A) -> (and (shr (shl x)) A)
7019 // where y is has a single bit set.
7020 // A plaintext description would be, we can turn the SELECT_CC into an AND
7021 // when the condition can be materialized as an all-ones register. Any
7022 // single bit-test can be materialized as an all-ones register with
7023 // shift-left and shift-right-arith.
7024 if (CC == ISD::SETEQ && N0->getOpcode() == ISD::AND &&
7025 N0->getValueType(0) == VT &&
7026 N1C && N1C->isNullValue() &&
7027 N2C && N2C->isNullValue()) {
7028 SDValue AndLHS = N0->getOperand(0);
7029 ConstantSDNode *ConstAndRHS = dyn_cast<ConstantSDNode>(N0->getOperand(1));
7030 if (ConstAndRHS && ConstAndRHS->getAPIntValue().countPopulation() == 1) {
7031 // Shift the tested bit over the sign bit.
7032 APInt AndMask = ConstAndRHS->getAPIntValue();
7034 DAG.getConstant(AndMask.countLeadingZeros(), getShiftAmountTy());
7035 SDValue Shl = DAG.getNode(ISD::SHL, N0.getDebugLoc(), VT, AndLHS, ShlAmt);
7037 // Now arithmetic right shift it all the way over, so the result is either
7038 // all-ones, or zero.
7040 DAG.getConstant(AndMask.getBitWidth()-1, getShiftAmountTy());
7041 SDValue Shr = DAG.getNode(ISD::SRA, N0.getDebugLoc(), VT, Shl, ShrAmt);
7043 return DAG.getNode(ISD::AND, DL, VT, Shr, N3);
7047 // fold select C, 16, 0 -> shl C, 4
7048 if (N2C && N3C && N3C->isNullValue() && N2C->getAPIntValue().isPowerOf2() &&
7049 TLI.getBooleanContents() == TargetLowering::ZeroOrOneBooleanContent) {
7051 // If the caller doesn't want us to simplify this into a zext of a compare,
7053 if (NotExtCompare && N2C->getAPIntValue() == 1)
7056 // Get a SetCC of the condition
7057 // FIXME: Should probably make sure that setcc is legal if we ever have a
7058 // target where it isn't.
7060 // cast from setcc result type to select result type
7062 SCC = DAG.getSetCC(DL, TLI.getSetCCResultType(N0.getValueType()),
7064 if (N2.getValueType().bitsLT(SCC.getValueType()))
7065 Temp = DAG.getZeroExtendInReg(SCC, N2.getDebugLoc(), N2.getValueType());
7067 Temp = DAG.getNode(ISD::ZERO_EXTEND, N2.getDebugLoc(),
7068 N2.getValueType(), SCC);
7070 SCC = DAG.getSetCC(N0.getDebugLoc(), MVT::i1, N0, N1, CC);
7071 Temp = DAG.getNode(ISD::ZERO_EXTEND, N2.getDebugLoc(),
7072 N2.getValueType(), SCC);
7075 AddToWorkList(SCC.getNode());
7076 AddToWorkList(Temp.getNode());
7078 if (N2C->getAPIntValue() == 1)
7081 // shl setcc result by log2 n2c
7082 return DAG.getNode(ISD::SHL, DL, N2.getValueType(), Temp,
7083 DAG.getConstant(N2C->getAPIntValue().logBase2(),
7084 getShiftAmountTy()));
7087 // Check to see if this is the equivalent of setcc
7088 // FIXME: Turn all of these into setcc if setcc if setcc is legal
7089 // otherwise, go ahead with the folds.
7090 if (0 && N3C && N3C->isNullValue() && N2C && (N2C->getAPIntValue() == 1ULL)) {
7091 EVT XType = N0.getValueType();
7092 if (!LegalOperations ||
7093 TLI.isOperationLegal(ISD::SETCC, TLI.getSetCCResultType(XType))) {
7094 SDValue Res = DAG.getSetCC(DL, TLI.getSetCCResultType(XType), N0, N1, CC);
7095 if (Res.getValueType() != VT)
7096 Res = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Res);
7100 // fold (seteq X, 0) -> (srl (ctlz X, log2(size(X))))
7101 if (N1C && N1C->isNullValue() && CC == ISD::SETEQ &&
7102 (!LegalOperations ||
7103 TLI.isOperationLegal(ISD::CTLZ, XType))) {
7104 SDValue Ctlz = DAG.getNode(ISD::CTLZ, N0.getDebugLoc(), XType, N0);
7105 return DAG.getNode(ISD::SRL, DL, XType, Ctlz,
7106 DAG.getConstant(Log2_32(XType.getSizeInBits()),
7107 getShiftAmountTy()));
7109 // fold (setgt X, 0) -> (srl (and (-X, ~X), size(X)-1))
7110 if (N1C && N1C->isNullValue() && CC == ISD::SETGT) {
7111 SDValue NegN0 = DAG.getNode(ISD::SUB, N0.getDebugLoc(),
7112 XType, DAG.getConstant(0, XType), N0);
7113 SDValue NotN0 = DAG.getNOT(N0.getDebugLoc(), N0, XType);
7114 return DAG.getNode(ISD::SRL, DL, XType,
7115 DAG.getNode(ISD::AND, DL, XType, NegN0, NotN0),
7116 DAG.getConstant(XType.getSizeInBits()-1,
7117 getShiftAmountTy()));
7119 // fold (setgt X, -1) -> (xor (srl (X, size(X)-1), 1))
7120 if (N1C && N1C->isAllOnesValue() && CC == ISD::SETGT) {
7121 SDValue Sign = DAG.getNode(ISD::SRL, N0.getDebugLoc(), XType, N0,
7122 DAG.getConstant(XType.getSizeInBits()-1,
7123 getShiftAmountTy()));
7124 return DAG.getNode(ISD::XOR, DL, XType, Sign, DAG.getConstant(1, XType));
7128 // Check to see if this is an integer abs.
7129 // select_cc setg[te] X, 0, X, -X ->
7130 // select_cc setgt X, -1, X, -X ->
7131 // select_cc setl[te] X, 0, -X, X ->
7132 // select_cc setlt X, 1, -X, X ->
7133 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
7135 ConstantSDNode *SubC = NULL;
7136 if (((N1C->isNullValue() && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
7137 (N1C->isAllOnesValue() && CC == ISD::SETGT)) &&
7138 N0 == N2 && N3.getOpcode() == ISD::SUB && N0 == N3.getOperand(1))
7139 SubC = dyn_cast<ConstantSDNode>(N3.getOperand(0));
7140 else if (((N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE)) ||
7141 (N1C->isOne() && CC == ISD::SETLT)) &&
7142 N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1))
7143 SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0));
7145 EVT XType = N0.getValueType();
7146 if (SubC && SubC->isNullValue() && XType.isInteger()) {
7147 SDValue Shift = DAG.getNode(ISD::SRA, N0.getDebugLoc(), XType,
7149 DAG.getConstant(XType.getSizeInBits()-1,
7150 getShiftAmountTy()));
7151 SDValue Add = DAG.getNode(ISD::ADD, N0.getDebugLoc(),
7153 AddToWorkList(Shift.getNode());
7154 AddToWorkList(Add.getNode());
7155 return DAG.getNode(ISD::XOR, DL, XType, Add, Shift);
7162 /// SimplifySetCC - This is a stub for TargetLowering::SimplifySetCC.
7163 SDValue DAGCombiner::SimplifySetCC(EVT VT, SDValue N0,
7164 SDValue N1, ISD::CondCode Cond,
7165 DebugLoc DL, bool foldBooleans) {
7166 TargetLowering::DAGCombinerInfo
7167 DagCombineInfo(DAG, !LegalTypes, !LegalOperations, false, this);
7168 return TLI.SimplifySetCC(VT, N0, N1, Cond, foldBooleans, DagCombineInfo, DL);
7171 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
7172 /// return a DAG expression to select that will generate the same value by
7173 /// multiplying by a magic number. See:
7174 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
7175 SDValue DAGCombiner::BuildSDIV(SDNode *N) {
7176 std::vector<SDNode*> Built;
7177 SDValue S = TLI.BuildSDIV(N, DAG, &Built);
7179 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end();
7185 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
7186 /// return a DAG expression to select that will generate the same value by
7187 /// multiplying by a magic number. See:
7188 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
7189 SDValue DAGCombiner::BuildUDIV(SDNode *N) {
7190 std::vector<SDNode*> Built;
7191 SDValue S = TLI.BuildUDIV(N, DAG, &Built);
7193 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end();
7199 /// FindBaseOffset - Return true if base is a frame index, which is known not
7200 // to alias with anything but itself. Provides base object and offset as
7202 static bool FindBaseOffset(SDValue Ptr, SDValue &Base, int64_t &Offset,
7203 const GlobalValue *&GV, void *&CV) {
7204 // Assume it is a primitive operation.
7205 Base = Ptr; Offset = 0; GV = 0; CV = 0;
7207 // If it's an adding a simple constant then integrate the offset.
7208 if (Base.getOpcode() == ISD::ADD) {
7209 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Base.getOperand(1))) {
7210 Base = Base.getOperand(0);
7211 Offset += C->getZExtValue();
7215 // Return the underlying GlobalValue, and update the Offset. Return false
7216 // for GlobalAddressSDNode since the same GlobalAddress may be represented
7217 // by multiple nodes with different offsets.
7218 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Base)) {
7219 GV = G->getGlobal();
7220 Offset += G->getOffset();
7224 // Return the underlying Constant value, and update the Offset. Return false
7225 // for ConstantSDNodes since the same constant pool entry may be represented
7226 // by multiple nodes with different offsets.
7227 if (ConstantPoolSDNode *C = dyn_cast<ConstantPoolSDNode>(Base)) {
7228 CV = C->isMachineConstantPoolEntry() ? (void *)C->getMachineCPVal()
7229 : (void *)C->getConstVal();
7230 Offset += C->getOffset();
7233 // If it's any of the following then it can't alias with anything but itself.
7234 return isa<FrameIndexSDNode>(Base);
7237 /// isAlias - Return true if there is any possibility that the two addresses
7239 bool DAGCombiner::isAlias(SDValue Ptr1, int64_t Size1,
7240 const Value *SrcValue1, int SrcValueOffset1,
7241 unsigned SrcValueAlign1,
7242 const MDNode *TBAAInfo1,
7243 SDValue Ptr2, int64_t Size2,
7244 const Value *SrcValue2, int SrcValueOffset2,
7245 unsigned SrcValueAlign2,
7246 const MDNode *TBAAInfo2) const {
7247 // If they are the same then they must be aliases.
7248 if (Ptr1 == Ptr2) return true;
7250 // Gather base node and offset information.
7251 SDValue Base1, Base2;
7252 int64_t Offset1, Offset2;
7253 const GlobalValue *GV1, *GV2;
7255 bool isFrameIndex1 = FindBaseOffset(Ptr1, Base1, Offset1, GV1, CV1);
7256 bool isFrameIndex2 = FindBaseOffset(Ptr2, Base2, Offset2, GV2, CV2);
7258 // If they have a same base address then check to see if they overlap.
7259 if (Base1 == Base2 || (GV1 && (GV1 == GV2)) || (CV1 && (CV1 == CV2)))
7260 return !((Offset1 + Size1) <= Offset2 || (Offset2 + Size2) <= Offset1);
7262 // It is possible for different frame indices to alias each other, mostly
7263 // when tail call optimization reuses return address slots for arguments.
7264 // To catch this case, look up the actual index of frame indices to compute
7265 // the real alias relationship.
7266 if (isFrameIndex1 && isFrameIndex2) {
7267 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7268 Offset1 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base1)->getIndex());
7269 Offset2 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base2)->getIndex());
7270 return !((Offset1 + Size1) <= Offset2 || (Offset2 + Size2) <= Offset1);
7273 // Otherwise, if we know what the bases are, and they aren't identical, then
7274 // we know they cannot alias.
7275 if ((isFrameIndex1 || CV1 || GV1) && (isFrameIndex2 || CV2 || GV2))
7278 // If we know required SrcValue1 and SrcValue2 have relatively large alignment
7279 // compared to the size and offset of the access, we may be able to prove they
7280 // do not alias. This check is conservative for now to catch cases created by
7281 // splitting vector types.
7282 if ((SrcValueAlign1 == SrcValueAlign2) &&
7283 (SrcValueOffset1 != SrcValueOffset2) &&
7284 (Size1 == Size2) && (SrcValueAlign1 > Size1)) {
7285 int64_t OffAlign1 = SrcValueOffset1 % SrcValueAlign1;
7286 int64_t OffAlign2 = SrcValueOffset2 % SrcValueAlign1;
7288 // There is no overlap between these relatively aligned accesses of similar
7289 // size, return no alias.
7290 if ((OffAlign1 + Size1) <= OffAlign2 || (OffAlign2 + Size2) <= OffAlign1)
7294 if (CombinerGlobalAA) {
7295 // Use alias analysis information.
7296 int64_t MinOffset = std::min(SrcValueOffset1, SrcValueOffset2);
7297 int64_t Overlap1 = Size1 + SrcValueOffset1 - MinOffset;
7298 int64_t Overlap2 = Size2 + SrcValueOffset2 - MinOffset;
7299 AliasAnalysis::AliasResult AAResult =
7300 AA.alias(AliasAnalysis::Location(SrcValue1, Overlap1, TBAAInfo1),
7301 AliasAnalysis::Location(SrcValue2, Overlap2, TBAAInfo2));
7302 if (AAResult == AliasAnalysis::NoAlias)
7306 // Otherwise we have to assume they alias.
7310 /// FindAliasInfo - Extracts the relevant alias information from the memory
7311 /// node. Returns true if the operand was a load.
7312 bool DAGCombiner::FindAliasInfo(SDNode *N,
7313 SDValue &Ptr, int64_t &Size,
7314 const Value *&SrcValue,
7315 int &SrcValueOffset,
7316 unsigned &SrcValueAlign,
7317 const MDNode *&TBAAInfo) const {
7318 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7319 Ptr = LD->getBasePtr();
7320 Size = LD->getMemoryVT().getSizeInBits() >> 3;
7321 SrcValue = LD->getSrcValue();
7322 SrcValueOffset = LD->getSrcValueOffset();
7323 SrcValueAlign = LD->getOriginalAlignment();
7324 TBAAInfo = LD->getTBAAInfo();
7326 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7327 Ptr = ST->getBasePtr();
7328 Size = ST->getMemoryVT().getSizeInBits() >> 3;
7329 SrcValue = ST->getSrcValue();
7330 SrcValueOffset = ST->getSrcValueOffset();
7331 SrcValueAlign = ST->getOriginalAlignment();
7332 TBAAInfo = ST->getTBAAInfo();
7334 llvm_unreachable("FindAliasInfo expected a memory operand");
7340 /// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes,
7341 /// looking for aliasing nodes and adding them to the Aliases vector.
7342 void DAGCombiner::GatherAllAliases(SDNode *N, SDValue OriginalChain,
7343 SmallVector<SDValue, 8> &Aliases) {
7344 SmallVector<SDValue, 8> Chains; // List of chains to visit.
7345 SmallPtrSet<SDNode *, 16> Visited; // Visited node set.
7347 // Get alias information for node.
7350 const Value *SrcValue;
7352 unsigned SrcValueAlign;
7353 const MDNode *SrcTBAAInfo;
7354 bool IsLoad = FindAliasInfo(N, Ptr, Size, SrcValue, SrcValueOffset,
7355 SrcValueAlign, SrcTBAAInfo);
7358 Chains.push_back(OriginalChain);
7361 // Look at each chain and determine if it is an alias. If so, add it to the
7362 // aliases list. If not, then continue up the chain looking for the next
7364 while (!Chains.empty()) {
7365 SDValue Chain = Chains.back();
7368 // For TokenFactor nodes, look at each operand and only continue up the
7369 // chain until we find two aliases. If we've seen two aliases, assume we'll
7370 // find more and revert to original chain since the xform is unlikely to be
7373 // FIXME: The depth check could be made to return the last non-aliasing
7374 // chain we found before we hit a tokenfactor rather than the original
7376 if (Depth > 6 || Aliases.size() == 2) {
7378 Aliases.push_back(OriginalChain);
7382 // Don't bother if we've been before.
7383 if (!Visited.insert(Chain.getNode()))
7386 switch (Chain.getOpcode()) {
7387 case ISD::EntryToken:
7388 // Entry token is ideal chain operand, but handled in FindBetterChain.
7393 // Get alias information for Chain.
7396 const Value *OpSrcValue;
7397 int OpSrcValueOffset;
7398 unsigned OpSrcValueAlign;
7399 const MDNode *OpSrcTBAAInfo;
7400 bool IsOpLoad = FindAliasInfo(Chain.getNode(), OpPtr, OpSize,
7401 OpSrcValue, OpSrcValueOffset,
7405 // If chain is alias then stop here.
7406 if (!(IsLoad && IsOpLoad) &&
7407 isAlias(Ptr, Size, SrcValue, SrcValueOffset, SrcValueAlign,
7409 OpPtr, OpSize, OpSrcValue, OpSrcValueOffset,
7410 OpSrcValueAlign, OpSrcTBAAInfo)) {
7411 Aliases.push_back(Chain);
7413 // Look further up the chain.
7414 Chains.push_back(Chain.getOperand(0));
7420 case ISD::TokenFactor:
7421 // We have to check each of the operands of the token factor for "small"
7422 // token factors, so we queue them up. Adding the operands to the queue
7423 // (stack) in reverse order maintains the original order and increases the
7424 // likelihood that getNode will find a matching token factor (CSE.)
7425 if (Chain.getNumOperands() > 16) {
7426 Aliases.push_back(Chain);
7429 for (unsigned n = Chain.getNumOperands(); n;)
7430 Chains.push_back(Chain.getOperand(--n));
7435 // For all other instructions we will just have to take what we can get.
7436 Aliases.push_back(Chain);
7442 /// FindBetterChain - Walk up chain skipping non-aliasing memory nodes, looking
7443 /// for a better chain (aliasing node.)
7444 SDValue DAGCombiner::FindBetterChain(SDNode *N, SDValue OldChain) {
7445 SmallVector<SDValue, 8> Aliases; // Ops for replacing token factor.
7447 // Accumulate all the aliases to this node.
7448 GatherAllAliases(N, OldChain, Aliases);
7450 if (Aliases.size() == 0) {
7451 // If no operands then chain to entry token.
7452 return DAG.getEntryNode();
7453 } else if (Aliases.size() == 1) {
7454 // If a single operand then chain to it. We don't need to revisit it.
7458 // Construct a custom tailored token factor.
7459 return DAG.getNode(ISD::TokenFactor, N->getDebugLoc(), MVT::Other,
7460 &Aliases[0], Aliases.size());
7463 // SelectionDAG::Combine - This is the entry point for the file.
7465 void SelectionDAG::Combine(CombineLevel Level, AliasAnalysis &AA,
7466 CodeGenOpt::Level OptLevel) {
7467 /// run - This is the main entry point to this class.
7469 DAGCombiner(*this, AA, OptLevel).Run(Level);