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);
133 /// combine - call the node-specific routine that knows how to fold each
134 /// particular type of node. If that doesn't do anything, try the
135 /// target-specific DAG combines.
136 SDValue combine(SDNode *N);
138 // Visitation implementation - Implement dag node combining for different
139 // node types. The semantics are as follows:
141 // SDValue.getNode() == 0 - No change was made
142 // SDValue.getNode() == N - N was replaced, is dead and has been handled.
143 // otherwise - N should be replaced by the returned Operand.
145 SDValue visitTokenFactor(SDNode *N);
146 SDValue visitMERGE_VALUES(SDNode *N);
147 SDValue visitADD(SDNode *N);
148 SDValue visitSUB(SDNode *N);
149 SDValue visitADDC(SDNode *N);
150 SDValue visitADDE(SDNode *N);
151 SDValue visitMUL(SDNode *N);
152 SDValue visitSDIV(SDNode *N);
153 SDValue visitUDIV(SDNode *N);
154 SDValue visitSREM(SDNode *N);
155 SDValue visitUREM(SDNode *N);
156 SDValue visitMULHU(SDNode *N);
157 SDValue visitMULHS(SDNode *N);
158 SDValue visitSMUL_LOHI(SDNode *N);
159 SDValue visitUMUL_LOHI(SDNode *N);
160 SDValue visitSDIVREM(SDNode *N);
161 SDValue visitUDIVREM(SDNode *N);
162 SDValue visitAND(SDNode *N);
163 SDValue visitOR(SDNode *N);
164 SDValue visitXOR(SDNode *N);
165 SDValue SimplifyVBinOp(SDNode *N);
166 SDValue visitSHL(SDNode *N);
167 SDValue visitSRA(SDNode *N);
168 SDValue visitSRL(SDNode *N);
169 SDValue visitCTLZ(SDNode *N);
170 SDValue visitCTTZ(SDNode *N);
171 SDValue visitCTPOP(SDNode *N);
172 SDValue visitSELECT(SDNode *N);
173 SDValue visitSELECT_CC(SDNode *N);
174 SDValue visitSETCC(SDNode *N);
175 SDValue visitSIGN_EXTEND(SDNode *N);
176 SDValue visitZERO_EXTEND(SDNode *N);
177 SDValue visitANY_EXTEND(SDNode *N);
178 SDValue visitSIGN_EXTEND_INREG(SDNode *N);
179 SDValue visitTRUNCATE(SDNode *N);
180 SDValue visitBIT_CONVERT(SDNode *N);
181 SDValue visitBUILD_PAIR(SDNode *N);
182 SDValue visitFADD(SDNode *N);
183 SDValue visitFSUB(SDNode *N);
184 SDValue visitFMUL(SDNode *N);
185 SDValue visitFDIV(SDNode *N);
186 SDValue visitFREM(SDNode *N);
187 SDValue visitFCOPYSIGN(SDNode *N);
188 SDValue visitSINT_TO_FP(SDNode *N);
189 SDValue visitUINT_TO_FP(SDNode *N);
190 SDValue visitFP_TO_SINT(SDNode *N);
191 SDValue visitFP_TO_UINT(SDNode *N);
192 SDValue visitFP_ROUND(SDNode *N);
193 SDValue visitFP_ROUND_INREG(SDNode *N);
194 SDValue visitFP_EXTEND(SDNode *N);
195 SDValue visitFNEG(SDNode *N);
196 SDValue visitFABS(SDNode *N);
197 SDValue visitBRCOND(SDNode *N);
198 SDValue visitBR_CC(SDNode *N);
199 SDValue visitLOAD(SDNode *N);
200 SDValue visitSTORE(SDNode *N);
201 SDValue visitINSERT_VECTOR_ELT(SDNode *N);
202 SDValue visitEXTRACT_VECTOR_ELT(SDNode *N);
203 SDValue visitBUILD_VECTOR(SDNode *N);
204 SDValue visitCONCAT_VECTORS(SDNode *N);
205 SDValue visitVECTOR_SHUFFLE(SDNode *N);
207 SDValue XformToShuffleWithZero(SDNode *N);
208 SDValue ReassociateOps(unsigned Opc, DebugLoc DL, SDValue LHS, SDValue RHS);
210 SDValue visitShiftByConstant(SDNode *N, unsigned Amt);
212 bool SimplifySelectOps(SDNode *SELECT, SDValue LHS, SDValue RHS);
213 SDValue SimplifyBinOpWithSameOpcodeHands(SDNode *N);
214 SDValue SimplifySelect(DebugLoc DL, SDValue N0, SDValue N1, SDValue N2);
215 SDValue SimplifySelectCC(DebugLoc DL, SDValue N0, SDValue N1, SDValue N2,
216 SDValue N3, ISD::CondCode CC,
217 bool NotExtCompare = false);
218 SDValue SimplifySetCC(EVT VT, SDValue N0, SDValue N1, ISD::CondCode Cond,
219 DebugLoc DL, bool foldBooleans = true);
220 SDValue SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
222 SDValue CombineConsecutiveLoads(SDNode *N, EVT VT);
223 SDValue ConstantFoldBIT_CONVERTofBUILD_VECTOR(SDNode *, EVT);
224 SDValue BuildSDIV(SDNode *N);
225 SDValue BuildUDIV(SDNode *N);
226 SDNode *MatchRotate(SDValue LHS, SDValue RHS, DebugLoc DL);
227 SDValue ReduceLoadWidth(SDNode *N);
228 SDValue ReduceLoadOpStoreWidth(SDNode *N);
230 SDValue GetDemandedBits(SDValue V, const APInt &Mask);
232 /// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes,
233 /// looking for aliasing nodes and adding them to the Aliases vector.
234 void GatherAllAliases(SDNode *N, SDValue OriginalChain,
235 SmallVector<SDValue, 8> &Aliases);
237 /// isAlias - Return true if there is any possibility that the two addresses
239 bool isAlias(SDValue Ptr1, int64_t Size1,
240 const Value *SrcValue1, int SrcValueOffset1,
241 unsigned SrcValueAlign1,
242 SDValue Ptr2, int64_t Size2,
243 const Value *SrcValue2, int SrcValueOffset2,
244 unsigned SrcValueAlign2) const;
246 /// FindAliasInfo - Extracts the relevant alias information from the memory
247 /// node. Returns true if the operand was a load.
248 bool FindAliasInfo(SDNode *N,
249 SDValue &Ptr, int64_t &Size,
250 const Value *&SrcValue, int &SrcValueOffset,
251 unsigned &SrcValueAlignment) const;
253 /// FindBetterChain - Walk up chain skipping non-aliasing memory nodes,
254 /// looking for a better chain (aliasing node.)
255 SDValue FindBetterChain(SDNode *N, SDValue Chain);
257 /// getShiftAmountTy - Returns a type large enough to hold any valid
258 /// shift amount - before type legalization these can be huge.
259 EVT getShiftAmountTy() {
260 return LegalTypes ? TLI.getShiftAmountTy() : TLI.getPointerTy();
264 DAGCombiner(SelectionDAG &D, AliasAnalysis &A, CodeGenOpt::Level OL)
266 TLI(D.getTargetLoweringInfo()),
269 LegalOperations(false),
273 /// Run - runs the dag combiner on all nodes in the work list
274 void Run(CombineLevel AtLevel);
280 /// WorkListRemover - This class is a DAGUpdateListener that removes any deleted
281 /// nodes from the worklist.
282 class WorkListRemover : public SelectionDAG::DAGUpdateListener {
285 explicit WorkListRemover(DAGCombiner &dc) : DC(dc) {}
287 virtual void NodeDeleted(SDNode *N, SDNode *E) {
288 DC.removeFromWorkList(N);
291 virtual void NodeUpdated(SDNode *N) {
297 //===----------------------------------------------------------------------===//
298 // TargetLowering::DAGCombinerInfo implementation
299 //===----------------------------------------------------------------------===//
301 void TargetLowering::DAGCombinerInfo::AddToWorklist(SDNode *N) {
302 ((DAGCombiner*)DC)->AddToWorkList(N);
305 SDValue TargetLowering::DAGCombinerInfo::
306 CombineTo(SDNode *N, const std::vector<SDValue> &To, bool AddTo) {
307 return ((DAGCombiner*)DC)->CombineTo(N, &To[0], To.size(), AddTo);
310 SDValue TargetLowering::DAGCombinerInfo::
311 CombineTo(SDNode *N, SDValue Res, bool AddTo) {
312 return ((DAGCombiner*)DC)->CombineTo(N, Res, AddTo);
316 SDValue TargetLowering::DAGCombinerInfo::
317 CombineTo(SDNode *N, SDValue Res0, SDValue Res1, bool AddTo) {
318 return ((DAGCombiner*)DC)->CombineTo(N, Res0, Res1, AddTo);
321 void TargetLowering::DAGCombinerInfo::
322 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
323 return ((DAGCombiner*)DC)->CommitTargetLoweringOpt(TLO);
326 //===----------------------------------------------------------------------===//
328 //===----------------------------------------------------------------------===//
330 /// isNegatibleForFree - Return 1 if we can compute the negated form of the
331 /// specified expression for the same cost as the expression itself, or 2 if we
332 /// can compute the negated form more cheaply than the expression itself.
333 static char isNegatibleForFree(SDValue Op, bool LegalOperations,
334 unsigned Depth = 0) {
335 // No compile time optimizations on this type.
336 if (Op.getValueType() == MVT::ppcf128)
339 // fneg is removable even if it has multiple uses.
340 if (Op.getOpcode() == ISD::FNEG) return 2;
342 // Don't allow anything with multiple uses.
343 if (!Op.hasOneUse()) return 0;
345 // Don't recurse exponentially.
346 if (Depth > 6) return 0;
348 switch (Op.getOpcode()) {
349 default: return false;
350 case ISD::ConstantFP:
351 // Don't invert constant FP values after legalize. The negated constant
352 // isn't necessarily legal.
353 return LegalOperations ? 0 : 1;
355 // FIXME: determine better conditions for this xform.
356 if (!UnsafeFPMath) return 0;
358 // fold (fsub (fadd A, B)) -> (fsub (fneg A), B)
359 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, Depth+1))
361 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
362 return isNegatibleForFree(Op.getOperand(1), LegalOperations, Depth+1);
364 // We can't turn -(A-B) into B-A when we honor signed zeros.
365 if (!UnsafeFPMath) return 0;
367 // fold (fneg (fsub A, B)) -> (fsub B, A)
372 if (HonorSignDependentRoundingFPMath()) return 0;
374 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y) or (fmul X, (fneg Y))
375 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, Depth+1))
378 return isNegatibleForFree(Op.getOperand(1), LegalOperations, Depth+1);
383 return isNegatibleForFree(Op.getOperand(0), LegalOperations, Depth+1);
387 /// GetNegatedExpression - If isNegatibleForFree returns true, this function
388 /// returns the newly negated expression.
389 static SDValue GetNegatedExpression(SDValue Op, SelectionDAG &DAG,
390 bool LegalOperations, unsigned Depth = 0) {
391 // fneg is removable even if it has multiple uses.
392 if (Op.getOpcode() == ISD::FNEG) return Op.getOperand(0);
394 // Don't allow anything with multiple uses.
395 assert(Op.hasOneUse() && "Unknown reuse!");
397 assert(Depth <= 6 && "GetNegatedExpression doesn't match isNegatibleForFree");
398 switch (Op.getOpcode()) {
399 default: llvm_unreachable("Unknown code");
400 case ISD::ConstantFP: {
401 APFloat V = cast<ConstantFPSDNode>(Op)->getValueAPF();
403 return DAG.getConstantFP(V, Op.getValueType());
406 // FIXME: determine better conditions for this xform.
407 assert(UnsafeFPMath);
409 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B)
410 if (isNegatibleForFree(Op.getOperand(0), LegalOperations, Depth+1))
411 return DAG.getNode(ISD::FSUB, Op.getDebugLoc(), Op.getValueType(),
412 GetNegatedExpression(Op.getOperand(0), DAG,
413 LegalOperations, Depth+1),
415 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
416 return DAG.getNode(ISD::FSUB, Op.getDebugLoc(), Op.getValueType(),
417 GetNegatedExpression(Op.getOperand(1), DAG,
418 LegalOperations, Depth+1),
421 // We can't turn -(A-B) into B-A when we honor signed zeros.
422 assert(UnsafeFPMath);
424 // fold (fneg (fsub 0, B)) -> B
425 if (ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(Op.getOperand(0)))
426 if (N0CFP->getValueAPF().isZero())
427 return Op.getOperand(1);
429 // fold (fneg (fsub A, B)) -> (fsub B, A)
430 return DAG.getNode(ISD::FSUB, Op.getDebugLoc(), Op.getValueType(),
431 Op.getOperand(1), Op.getOperand(0));
435 assert(!HonorSignDependentRoundingFPMath());
437 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y)
438 if (isNegatibleForFree(Op.getOperand(0), LegalOperations, Depth+1))
439 return DAG.getNode(Op.getOpcode(), Op.getDebugLoc(), Op.getValueType(),
440 GetNegatedExpression(Op.getOperand(0), DAG,
441 LegalOperations, Depth+1),
444 // fold (fneg (fmul X, Y)) -> (fmul X, (fneg Y))
445 return DAG.getNode(Op.getOpcode(), Op.getDebugLoc(), Op.getValueType(),
447 GetNegatedExpression(Op.getOperand(1), DAG,
448 LegalOperations, Depth+1));
452 return DAG.getNode(Op.getOpcode(), Op.getDebugLoc(), Op.getValueType(),
453 GetNegatedExpression(Op.getOperand(0), DAG,
454 LegalOperations, Depth+1));
456 return DAG.getNode(ISD::FP_ROUND, Op.getDebugLoc(), Op.getValueType(),
457 GetNegatedExpression(Op.getOperand(0), DAG,
458 LegalOperations, Depth+1),
464 // isSetCCEquivalent - Return true if this node is a setcc, or is a select_cc
465 // that selects between the values 1 and 0, making it equivalent to a setcc.
466 // Also, set the incoming LHS, RHS, and CC references to the appropriate
467 // nodes based on the type of node we are checking. This simplifies life a
468 // bit for the callers.
469 static bool isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS,
471 if (N.getOpcode() == ISD::SETCC) {
472 LHS = N.getOperand(0);
473 RHS = N.getOperand(1);
474 CC = N.getOperand(2);
477 if (N.getOpcode() == ISD::SELECT_CC &&
478 N.getOperand(2).getOpcode() == ISD::Constant &&
479 N.getOperand(3).getOpcode() == ISD::Constant &&
480 cast<ConstantSDNode>(N.getOperand(2))->getAPIntValue() == 1 &&
481 cast<ConstantSDNode>(N.getOperand(3))->isNullValue()) {
482 LHS = N.getOperand(0);
483 RHS = N.getOperand(1);
484 CC = N.getOperand(4);
490 // isOneUseSetCC - Return true if this is a SetCC-equivalent operation with only
491 // one use. If this is true, it allows the users to invert the operation for
492 // free when it is profitable to do so.
493 static bool isOneUseSetCC(SDValue N) {
495 if (isSetCCEquivalent(N, N0, N1, N2) && N.getNode()->hasOneUse())
500 SDValue DAGCombiner::ReassociateOps(unsigned Opc, DebugLoc DL,
501 SDValue N0, SDValue N1) {
502 EVT VT = N0.getValueType();
503 if (N0.getOpcode() == Opc && isa<ConstantSDNode>(N0.getOperand(1))) {
504 if (isa<ConstantSDNode>(N1)) {
505 // reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2))
507 DAG.FoldConstantArithmetic(Opc, VT,
508 cast<ConstantSDNode>(N0.getOperand(1)),
509 cast<ConstantSDNode>(N1));
510 return DAG.getNode(Opc, DL, VT, N0.getOperand(0), OpNode);
511 } else if (N0.hasOneUse()) {
512 // reassoc. (op (op x, c1), y) -> (op (op x, y), c1) iff x+c1 has one use
513 SDValue OpNode = DAG.getNode(Opc, N0.getDebugLoc(), VT,
514 N0.getOperand(0), N1);
515 AddToWorkList(OpNode.getNode());
516 return DAG.getNode(Opc, DL, VT, OpNode, N0.getOperand(1));
520 if (N1.getOpcode() == Opc && isa<ConstantSDNode>(N1.getOperand(1))) {
521 if (isa<ConstantSDNode>(N0)) {
522 // reassoc. (op c2, (op x, c1)) -> (op x, (op c1, c2))
524 DAG.FoldConstantArithmetic(Opc, VT,
525 cast<ConstantSDNode>(N1.getOperand(1)),
526 cast<ConstantSDNode>(N0));
527 return DAG.getNode(Opc, DL, VT, N1.getOperand(0), OpNode);
528 } else if (N1.hasOneUse()) {
529 // reassoc. (op y, (op x, c1)) -> (op (op x, y), c1) iff x+c1 has one use
530 SDValue OpNode = DAG.getNode(Opc, N0.getDebugLoc(), VT,
531 N1.getOperand(0), N0);
532 AddToWorkList(OpNode.getNode());
533 return DAG.getNode(Opc, DL, VT, OpNode, N1.getOperand(1));
540 SDValue DAGCombiner::CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
542 assert(N->getNumValues() == NumTo && "Broken CombineTo call!");
544 DEBUG(dbgs() << "\nReplacing.1 ";
546 dbgs() << "\nWith: ";
547 To[0].getNode()->dump(&DAG);
548 dbgs() << " and " << NumTo-1 << " other values\n";
549 for (unsigned i = 0, e = NumTo; i != e; ++i)
550 assert((!To[i].getNode() ||
551 N->getValueType(i) == To[i].getValueType()) &&
552 "Cannot combine value to value of different type!"));
553 WorkListRemover DeadNodes(*this);
554 DAG.ReplaceAllUsesWith(N, To, &DeadNodes);
557 // Push the new nodes and any users onto the worklist
558 for (unsigned i = 0, e = NumTo; i != e; ++i) {
559 if (To[i].getNode()) {
560 AddToWorkList(To[i].getNode());
561 AddUsersToWorkList(To[i].getNode());
566 // Finally, if the node is now dead, remove it from the graph. The node
567 // may not be dead if the replacement process recursively simplified to
568 // something else needing this node.
569 if (N->use_empty()) {
570 // Nodes can be reintroduced into the worklist. Make sure we do not
571 // process a node that has been replaced.
572 removeFromWorkList(N);
574 // Finally, since the node is now dead, remove it from the graph.
577 return SDValue(N, 0);
581 DAGCombiner::CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &
583 // Replace all uses. If any nodes become isomorphic to other nodes and
584 // are deleted, make sure to remove them from our worklist.
585 WorkListRemover DeadNodes(*this);
586 DAG.ReplaceAllUsesOfValueWith(TLO.Old, TLO.New, &DeadNodes);
588 // Push the new node and any (possibly new) users onto the worklist.
589 AddToWorkList(TLO.New.getNode());
590 AddUsersToWorkList(TLO.New.getNode());
592 // Finally, if the node is now dead, remove it from the graph. The node
593 // may not be dead if the replacement process recursively simplified to
594 // something else needing this node.
595 if (TLO.Old.getNode()->use_empty()) {
596 removeFromWorkList(TLO.Old.getNode());
598 // If the operands of this node are only used by the node, they will now
599 // be dead. Make sure to visit them first to delete dead nodes early.
600 for (unsigned i = 0, e = TLO.Old.getNode()->getNumOperands(); i != e; ++i)
601 if (TLO.Old.getNode()->getOperand(i).getNode()->hasOneUse())
602 AddToWorkList(TLO.Old.getNode()->getOperand(i).getNode());
604 DAG.DeleteNode(TLO.Old.getNode());
608 /// SimplifyDemandedBits - Check the specified integer node value to see if
609 /// it can be simplified or if things it uses can be simplified by bit
610 /// propagation. If so, return true.
611 bool DAGCombiner::SimplifyDemandedBits(SDValue Op, const APInt &Demanded) {
612 TargetLowering::TargetLoweringOpt TLO(DAG);
613 APInt KnownZero, KnownOne;
614 if (!TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
618 AddToWorkList(Op.getNode());
620 // Replace the old value with the new one.
622 DEBUG(dbgs() << "\nReplacing.2 ";
623 TLO.Old.getNode()->dump(&DAG);
624 dbgs() << "\nWith: ";
625 TLO.New.getNode()->dump(&DAG);
628 CommitTargetLoweringOpt(TLO);
632 //===----------------------------------------------------------------------===//
633 // Main DAG Combiner implementation
634 //===----------------------------------------------------------------------===//
636 void DAGCombiner::Run(CombineLevel AtLevel) {
637 // set the instance variables, so that the various visit routines may use it.
639 LegalOperations = Level >= NoIllegalOperations;
640 LegalTypes = Level >= NoIllegalTypes;
642 // Add all the dag nodes to the worklist.
643 WorkList.reserve(DAG.allnodes_size());
644 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
645 E = DAG.allnodes_end(); I != E; ++I)
646 WorkList.push_back(I);
648 // Create a dummy node (which is not added to allnodes), that adds a reference
649 // to the root node, preventing it from being deleted, and tracking any
650 // changes of the root.
651 HandleSDNode Dummy(DAG.getRoot());
653 // The root of the dag may dangle to deleted nodes until the dag combiner is
654 // done. Set it to null to avoid confusion.
655 DAG.setRoot(SDValue());
657 // while the worklist isn't empty, inspect the node on the end of it and
658 // try and combine it.
659 while (!WorkList.empty()) {
660 SDNode *N = WorkList.back();
663 // If N has no uses, it is dead. Make sure to revisit all N's operands once
664 // N is deleted from the DAG, since they too may now be dead or may have a
665 // reduced number of uses, allowing other xforms.
666 if (N->use_empty() && N != &Dummy) {
667 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
668 AddToWorkList(N->getOperand(i).getNode());
674 SDValue RV = combine(N);
676 if (RV.getNode() == 0)
681 // If we get back the same node we passed in, rather than a new node or
682 // zero, we know that the node must have defined multiple values and
683 // CombineTo was used. Since CombineTo takes care of the worklist
684 // mechanics for us, we have no work to do in this case.
685 if (RV.getNode() == N)
688 assert(N->getOpcode() != ISD::DELETED_NODE &&
689 RV.getNode()->getOpcode() != ISD::DELETED_NODE &&
690 "Node was deleted but visit returned new node!");
692 DEBUG(dbgs() << "\nReplacing.3 ";
694 dbgs() << "\nWith: ";
695 RV.getNode()->dump(&DAG);
697 WorkListRemover DeadNodes(*this);
698 if (N->getNumValues() == RV.getNode()->getNumValues())
699 DAG.ReplaceAllUsesWith(N, RV.getNode(), &DeadNodes);
701 assert(N->getValueType(0) == RV.getValueType() &&
702 N->getNumValues() == 1 && "Type mismatch");
704 DAG.ReplaceAllUsesWith(N, &OpV, &DeadNodes);
707 // Push the new node and any users onto the worklist
708 AddToWorkList(RV.getNode());
709 AddUsersToWorkList(RV.getNode());
711 // Add any uses of the old node to the worklist in case this node is the
712 // last one that uses them. They may become dead after this node is
714 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
715 AddToWorkList(N->getOperand(i).getNode());
717 // Finally, if the node is now dead, remove it from the graph. The node
718 // may not be dead if the replacement process recursively simplified to
719 // something else needing this node.
720 if (N->use_empty()) {
721 // Nodes can be reintroduced into the worklist. Make sure we do not
722 // process a node that has been replaced.
723 removeFromWorkList(N);
725 // Finally, since the node is now dead, remove it from the graph.
730 // If the root changed (e.g. it was a dead load, update the root).
731 DAG.setRoot(Dummy.getValue());
734 SDValue DAGCombiner::visit(SDNode *N) {
735 switch(N->getOpcode()) {
737 case ISD::TokenFactor: return visitTokenFactor(N);
738 case ISD::MERGE_VALUES: return visitMERGE_VALUES(N);
739 case ISD::ADD: return visitADD(N);
740 case ISD::SUB: return visitSUB(N);
741 case ISD::ADDC: return visitADDC(N);
742 case ISD::ADDE: return visitADDE(N);
743 case ISD::MUL: return visitMUL(N);
744 case ISD::SDIV: return visitSDIV(N);
745 case ISD::UDIV: return visitUDIV(N);
746 case ISD::SREM: return visitSREM(N);
747 case ISD::UREM: return visitUREM(N);
748 case ISD::MULHU: return visitMULHU(N);
749 case ISD::MULHS: return visitMULHS(N);
750 case ISD::SMUL_LOHI: return visitSMUL_LOHI(N);
751 case ISD::UMUL_LOHI: return visitUMUL_LOHI(N);
752 case ISD::SDIVREM: return visitSDIVREM(N);
753 case ISD::UDIVREM: return visitUDIVREM(N);
754 case ISD::AND: return visitAND(N);
755 case ISD::OR: return visitOR(N);
756 case ISD::XOR: return visitXOR(N);
757 case ISD::SHL: return visitSHL(N);
758 case ISD::SRA: return visitSRA(N);
759 case ISD::SRL: return visitSRL(N);
760 case ISD::CTLZ: return visitCTLZ(N);
761 case ISD::CTTZ: return visitCTTZ(N);
762 case ISD::CTPOP: return visitCTPOP(N);
763 case ISD::SELECT: return visitSELECT(N);
764 case ISD::SELECT_CC: return visitSELECT_CC(N);
765 case ISD::SETCC: return visitSETCC(N);
766 case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N);
767 case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N);
768 case ISD::ANY_EXTEND: return visitANY_EXTEND(N);
769 case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N);
770 case ISD::TRUNCATE: return visitTRUNCATE(N);
771 case ISD::BIT_CONVERT: return visitBIT_CONVERT(N);
772 case ISD::BUILD_PAIR: return visitBUILD_PAIR(N);
773 case ISD::FADD: return visitFADD(N);
774 case ISD::FSUB: return visitFSUB(N);
775 case ISD::FMUL: return visitFMUL(N);
776 case ISD::FDIV: return visitFDIV(N);
777 case ISD::FREM: return visitFREM(N);
778 case ISD::FCOPYSIGN: return visitFCOPYSIGN(N);
779 case ISD::SINT_TO_FP: return visitSINT_TO_FP(N);
780 case ISD::UINT_TO_FP: return visitUINT_TO_FP(N);
781 case ISD::FP_TO_SINT: return visitFP_TO_SINT(N);
782 case ISD::FP_TO_UINT: return visitFP_TO_UINT(N);
783 case ISD::FP_ROUND: return visitFP_ROUND(N);
784 case ISD::FP_ROUND_INREG: return visitFP_ROUND_INREG(N);
785 case ISD::FP_EXTEND: return visitFP_EXTEND(N);
786 case ISD::FNEG: return visitFNEG(N);
787 case ISD::FABS: return visitFABS(N);
788 case ISD::BRCOND: return visitBRCOND(N);
789 case ISD::BR_CC: return visitBR_CC(N);
790 case ISD::LOAD: return visitLOAD(N);
791 case ISD::STORE: return visitSTORE(N);
792 case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N);
793 case ISD::EXTRACT_VECTOR_ELT: return visitEXTRACT_VECTOR_ELT(N);
794 case ISD::BUILD_VECTOR: return visitBUILD_VECTOR(N);
795 case ISD::CONCAT_VECTORS: return visitCONCAT_VECTORS(N);
796 case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N);
801 SDValue DAGCombiner::combine(SDNode *N) {
802 SDValue RV = visit(N);
804 // If nothing happened, try a target-specific DAG combine.
805 if (RV.getNode() == 0) {
806 assert(N->getOpcode() != ISD::DELETED_NODE &&
807 "Node was deleted but visit returned NULL!");
809 if (N->getOpcode() >= ISD::BUILTIN_OP_END ||
810 TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode())) {
812 // Expose the DAG combiner to the target combiner impls.
813 TargetLowering::DAGCombinerInfo
814 DagCombineInfo(DAG, !LegalTypes, !LegalOperations, false, this);
816 RV = TLI.PerformDAGCombine(N, DagCombineInfo);
820 // If N is a commutative binary node, try commuting it to enable more
822 if (RV.getNode() == 0 &&
823 SelectionDAG::isCommutativeBinOp(N->getOpcode()) &&
824 N->getNumValues() == 1) {
825 SDValue N0 = N->getOperand(0);
826 SDValue N1 = N->getOperand(1);
828 // Constant operands are canonicalized to RHS.
829 if (isa<ConstantSDNode>(N0) || !isa<ConstantSDNode>(N1)) {
830 SDValue Ops[] = { N1, N0 };
831 SDNode *CSENode = DAG.getNodeIfExists(N->getOpcode(), N->getVTList(),
834 return SDValue(CSENode, 0);
841 /// getInputChainForNode - Given a node, return its input chain if it has one,
842 /// otherwise return a null sd operand.
843 static SDValue getInputChainForNode(SDNode *N) {
844 if (unsigned NumOps = N->getNumOperands()) {
845 if (N->getOperand(0).getValueType() == MVT::Other)
846 return N->getOperand(0);
847 else if (N->getOperand(NumOps-1).getValueType() == MVT::Other)
848 return N->getOperand(NumOps-1);
849 for (unsigned i = 1; i < NumOps-1; ++i)
850 if (N->getOperand(i).getValueType() == MVT::Other)
851 return N->getOperand(i);
856 SDValue DAGCombiner::visitTokenFactor(SDNode *N) {
857 // If N has two operands, where one has an input chain equal to the other,
858 // the 'other' chain is redundant.
859 if (N->getNumOperands() == 2) {
860 if (getInputChainForNode(N->getOperand(0).getNode()) == N->getOperand(1))
861 return N->getOperand(0);
862 if (getInputChainForNode(N->getOperand(1).getNode()) == N->getOperand(0))
863 return N->getOperand(1);
866 SmallVector<SDNode *, 8> TFs; // List of token factors to visit.
867 SmallVector<SDValue, 8> Ops; // Ops for replacing token factor.
868 SmallPtrSet<SDNode*, 16> SeenOps;
869 bool Changed = false; // If we should replace this token factor.
871 // Start out with this token factor.
874 // Iterate through token factors. The TFs grows when new token factors are
876 for (unsigned i = 0; i < TFs.size(); ++i) {
879 // Check each of the operands.
880 for (unsigned i = 0, ie = TF->getNumOperands(); i != ie; ++i) {
881 SDValue Op = TF->getOperand(i);
883 switch (Op.getOpcode()) {
884 case ISD::EntryToken:
885 // Entry tokens don't need to be added to the list. They are
890 case ISD::TokenFactor:
891 if (Op.hasOneUse() &&
892 std::find(TFs.begin(), TFs.end(), Op.getNode()) == TFs.end()) {
893 // Queue up for processing.
894 TFs.push_back(Op.getNode());
895 // Clean up in case the token factor is removed.
896 AddToWorkList(Op.getNode());
903 // Only add if it isn't already in the list.
904 if (SeenOps.insert(Op.getNode()))
915 // If we've change things around then replace token factor.
918 // The entry token is the only possible outcome.
919 Result = DAG.getEntryNode();
921 // New and improved token factor.
922 Result = DAG.getNode(ISD::TokenFactor, N->getDebugLoc(),
923 MVT::Other, &Ops[0], Ops.size());
926 // Don't add users to work list.
927 return CombineTo(N, Result, false);
933 /// MERGE_VALUES can always be eliminated.
934 SDValue DAGCombiner::visitMERGE_VALUES(SDNode *N) {
935 WorkListRemover DeadNodes(*this);
936 // Replacing results may cause a different MERGE_VALUES to suddenly
937 // be CSE'd with N, and carry its uses with it. Iterate until no
938 // uses remain, to ensure that the node can be safely deleted.
940 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
941 DAG.ReplaceAllUsesOfValueWith(SDValue(N, i), N->getOperand(i),
943 } while (!N->use_empty());
944 removeFromWorkList(N);
946 return SDValue(N, 0); // Return N so it doesn't get rechecked!
950 SDValue combineShlAddConstant(DebugLoc DL, SDValue N0, SDValue N1,
952 EVT VT = N0.getValueType();
953 SDValue N00 = N0.getOperand(0);
954 SDValue N01 = N0.getOperand(1);
955 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N01);
957 if (N01C && N00.getOpcode() == ISD::ADD && N00.getNode()->hasOneUse() &&
958 isa<ConstantSDNode>(N00.getOperand(1))) {
959 // fold (add (shl (add x, c1), c2), ) -> (add (add (shl x, c2), c1<<c2), )
960 N0 = DAG.getNode(ISD::ADD, N0.getDebugLoc(), VT,
961 DAG.getNode(ISD::SHL, N00.getDebugLoc(), VT,
962 N00.getOperand(0), N01),
963 DAG.getNode(ISD::SHL, N01.getDebugLoc(), VT,
964 N00.getOperand(1), N01));
965 return DAG.getNode(ISD::ADD, DL, VT, N0, N1);
971 SDValue DAGCombiner::visitADD(SDNode *N) {
972 SDValue N0 = N->getOperand(0);
973 SDValue N1 = N->getOperand(1);
974 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
975 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
976 EVT VT = N0.getValueType();
980 SDValue FoldedVOp = SimplifyVBinOp(N);
981 if (FoldedVOp.getNode()) return FoldedVOp;
984 // fold (add x, undef) -> undef
985 if (N0.getOpcode() == ISD::UNDEF)
987 if (N1.getOpcode() == ISD::UNDEF)
989 // fold (add c1, c2) -> c1+c2
991 return DAG.FoldConstantArithmetic(ISD::ADD, VT, N0C, N1C);
992 // canonicalize constant to RHS
994 return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N1, N0);
995 // fold (add x, 0) -> x
996 if (N1C && N1C->isNullValue())
998 // fold (add Sym, c) -> Sym+c
999 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1000 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA) && N1C &&
1001 GA->getOpcode() == ISD::GlobalAddress)
1002 return DAG.getGlobalAddress(GA->getGlobal(), VT,
1004 (uint64_t)N1C->getSExtValue());
1005 // fold ((c1-A)+c2) -> (c1+c2)-A
1006 if (N1C && N0.getOpcode() == ISD::SUB)
1007 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getOperand(0)))
1008 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1009 DAG.getConstant(N1C->getAPIntValue()+
1010 N0C->getAPIntValue(), VT),
1013 SDValue RADD = ReassociateOps(ISD::ADD, N->getDebugLoc(), N0, N1);
1014 if (RADD.getNode() != 0)
1016 // fold ((0-A) + B) -> B-A
1017 if (N0.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N0.getOperand(0)) &&
1018 cast<ConstantSDNode>(N0.getOperand(0))->isNullValue())
1019 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N1, N0.getOperand(1));
1020 // fold (A + (0-B)) -> A-B
1021 if (N1.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N1.getOperand(0)) &&
1022 cast<ConstantSDNode>(N1.getOperand(0))->isNullValue())
1023 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N0, N1.getOperand(1));
1024 // fold (A+(B-A)) -> B
1025 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1))
1026 return N1.getOperand(0);
1027 // fold ((B-A)+A) -> B
1028 if (N0.getOpcode() == ISD::SUB && N1 == N0.getOperand(1))
1029 return N0.getOperand(0);
1030 // fold (A+(B-(A+C))) to (B-C)
1031 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1032 N0 == N1.getOperand(1).getOperand(0))
1033 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N1.getOperand(0),
1034 N1.getOperand(1).getOperand(1));
1035 // fold (A+(B-(C+A))) to (B-C)
1036 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1037 N0 == N1.getOperand(1).getOperand(1))
1038 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N1.getOperand(0),
1039 N1.getOperand(1).getOperand(0));
1040 // fold (A+((B-A)+or-C)) to (B+or-C)
1041 if ((N1.getOpcode() == ISD::SUB || N1.getOpcode() == ISD::ADD) &&
1042 N1.getOperand(0).getOpcode() == ISD::SUB &&
1043 N0 == N1.getOperand(0).getOperand(1))
1044 return DAG.getNode(N1.getOpcode(), N->getDebugLoc(), VT,
1045 N1.getOperand(0).getOperand(0), N1.getOperand(1));
1047 // fold (A-B)+(C-D) to (A+C)-(B+D) when A or C is constant
1048 if (N0.getOpcode() == ISD::SUB && N1.getOpcode() == ISD::SUB) {
1049 SDValue N00 = N0.getOperand(0);
1050 SDValue N01 = N0.getOperand(1);
1051 SDValue N10 = N1.getOperand(0);
1052 SDValue N11 = N1.getOperand(1);
1054 if (isa<ConstantSDNode>(N00) || isa<ConstantSDNode>(N10))
1055 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1056 DAG.getNode(ISD::ADD, N0.getDebugLoc(), VT, N00, N10),
1057 DAG.getNode(ISD::ADD, N1.getDebugLoc(), VT, N01, N11));
1060 if (!VT.isVector() && SimplifyDemandedBits(SDValue(N, 0)))
1061 return SDValue(N, 0);
1063 // fold (a+b) -> (a|b) iff a and b share no bits.
1064 if (VT.isInteger() && !VT.isVector()) {
1065 APInt LHSZero, LHSOne;
1066 APInt RHSZero, RHSOne;
1067 APInt Mask = APInt::getAllOnesValue(VT.getSizeInBits());
1068 DAG.ComputeMaskedBits(N0, Mask, LHSZero, LHSOne);
1070 if (LHSZero.getBoolValue()) {
1071 DAG.ComputeMaskedBits(N1, Mask, RHSZero, RHSOne);
1073 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1074 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1075 if ((RHSZero & (~LHSZero & Mask)) == (~LHSZero & Mask) ||
1076 (LHSZero & (~RHSZero & Mask)) == (~RHSZero & Mask))
1077 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N0, N1);
1081 // fold (add (shl (add x, c1), c2), ) -> (add (add (shl x, c2), c1<<c2), )
1082 if (N0.getOpcode() == ISD::SHL && N0.getNode()->hasOneUse()) {
1083 SDValue Result = combineShlAddConstant(N->getDebugLoc(), N0, N1, DAG);
1084 if (Result.getNode()) return Result;
1086 if (N1.getOpcode() == ISD::SHL && N1.getNode()->hasOneUse()) {
1087 SDValue Result = combineShlAddConstant(N->getDebugLoc(), N1, N0, DAG);
1088 if (Result.getNode()) return Result;
1094 SDValue DAGCombiner::visitADDC(SDNode *N) {
1095 SDValue N0 = N->getOperand(0);
1096 SDValue N1 = N->getOperand(1);
1097 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1098 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1099 EVT VT = N0.getValueType();
1101 // If the flag result is dead, turn this into an ADD.
1102 if (N->hasNUsesOfValue(0, 1))
1103 return CombineTo(N, DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N1, N0),
1104 DAG.getNode(ISD::CARRY_FALSE,
1105 N->getDebugLoc(), MVT::Flag));
1107 // canonicalize constant to RHS.
1109 return DAG.getNode(ISD::ADDC, N->getDebugLoc(), N->getVTList(), N1, N0);
1111 // fold (addc x, 0) -> x + no carry out
1112 if (N1C && N1C->isNullValue())
1113 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE,
1114 N->getDebugLoc(), MVT::Flag));
1116 // fold (addc a, b) -> (or a, b), CARRY_FALSE iff a and b share no bits.
1117 APInt LHSZero, LHSOne;
1118 APInt RHSZero, RHSOne;
1119 APInt Mask = APInt::getAllOnesValue(VT.getSizeInBits());
1120 DAG.ComputeMaskedBits(N0, Mask, LHSZero, LHSOne);
1122 if (LHSZero.getBoolValue()) {
1123 DAG.ComputeMaskedBits(N1, Mask, RHSZero, RHSOne);
1125 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1126 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1127 if ((RHSZero & (~LHSZero & Mask)) == (~LHSZero & Mask) ||
1128 (LHSZero & (~RHSZero & Mask)) == (~RHSZero & Mask))
1129 return CombineTo(N, DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N0, N1),
1130 DAG.getNode(ISD::CARRY_FALSE,
1131 N->getDebugLoc(), MVT::Flag));
1137 SDValue DAGCombiner::visitADDE(SDNode *N) {
1138 SDValue N0 = N->getOperand(0);
1139 SDValue N1 = N->getOperand(1);
1140 SDValue CarryIn = N->getOperand(2);
1141 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1142 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1144 // canonicalize constant to RHS
1146 return DAG.getNode(ISD::ADDE, N->getDebugLoc(), N->getVTList(),
1149 // fold (adde x, y, false) -> (addc x, y)
1150 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1151 return DAG.getNode(ISD::ADDC, N->getDebugLoc(), N->getVTList(), N1, N0);
1156 SDValue DAGCombiner::visitSUB(SDNode *N) {
1157 SDValue N0 = N->getOperand(0);
1158 SDValue N1 = N->getOperand(1);
1159 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1160 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1161 EVT VT = N0.getValueType();
1164 if (VT.isVector()) {
1165 SDValue FoldedVOp = SimplifyVBinOp(N);
1166 if (FoldedVOp.getNode()) return FoldedVOp;
1169 // fold (sub x, x) -> 0
1171 return DAG.getConstant(0, N->getValueType(0));
1172 // fold (sub c1, c2) -> c1-c2
1174 return DAG.FoldConstantArithmetic(ISD::SUB, VT, N0C, N1C);
1175 // fold (sub x, c) -> (add x, -c)
1177 return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N0,
1178 DAG.getConstant(-N1C->getAPIntValue(), VT));
1179 // fold (A+B)-A -> B
1180 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1)
1181 return N0.getOperand(1);
1182 // fold (A+B)-B -> A
1183 if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1)
1184 return N0.getOperand(0);
1185 // fold ((A+(B+or-C))-B) -> A+or-C
1186 if (N0.getOpcode() == ISD::ADD &&
1187 (N0.getOperand(1).getOpcode() == ISD::SUB ||
1188 N0.getOperand(1).getOpcode() == ISD::ADD) &&
1189 N0.getOperand(1).getOperand(0) == N1)
1190 return DAG.getNode(N0.getOperand(1).getOpcode(), N->getDebugLoc(), VT,
1191 N0.getOperand(0), N0.getOperand(1).getOperand(1));
1192 // fold ((A+(C+B))-B) -> A+C
1193 if (N0.getOpcode() == ISD::ADD &&
1194 N0.getOperand(1).getOpcode() == ISD::ADD &&
1195 N0.getOperand(1).getOperand(1) == N1)
1196 return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT,
1197 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1198 // fold ((A-(B-C))-C) -> A-B
1199 if (N0.getOpcode() == ISD::SUB &&
1200 N0.getOperand(1).getOpcode() == ISD::SUB &&
1201 N0.getOperand(1).getOperand(1) == N1)
1202 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1203 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1205 // If either operand of a sub is undef, the result is undef
1206 if (N0.getOpcode() == ISD::UNDEF)
1208 if (N1.getOpcode() == ISD::UNDEF)
1211 // If the relocation model supports it, consider symbol offsets.
1212 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1213 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA)) {
1214 // fold (sub Sym, c) -> Sym-c
1215 if (N1C && GA->getOpcode() == ISD::GlobalAddress)
1216 return DAG.getGlobalAddress(GA->getGlobal(), VT,
1218 (uint64_t)N1C->getSExtValue());
1219 // fold (sub Sym+c1, Sym+c2) -> c1-c2
1220 if (GlobalAddressSDNode *GB = dyn_cast<GlobalAddressSDNode>(N1))
1221 if (GA->getGlobal() == GB->getGlobal())
1222 return DAG.getConstant((uint64_t)GA->getOffset() - GB->getOffset(),
1229 SDValue DAGCombiner::visitMUL(SDNode *N) {
1230 SDValue N0 = N->getOperand(0);
1231 SDValue N1 = N->getOperand(1);
1232 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1233 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1234 EVT VT = N0.getValueType();
1237 if (VT.isVector()) {
1238 SDValue FoldedVOp = SimplifyVBinOp(N);
1239 if (FoldedVOp.getNode()) return FoldedVOp;
1242 // fold (mul x, undef) -> 0
1243 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
1244 return DAG.getConstant(0, VT);
1245 // fold (mul c1, c2) -> c1*c2
1247 return DAG.FoldConstantArithmetic(ISD::MUL, VT, N0C, N1C);
1248 // canonicalize constant to RHS
1250 return DAG.getNode(ISD::MUL, N->getDebugLoc(), VT, N1, N0);
1251 // fold (mul x, 0) -> 0
1252 if (N1C && N1C->isNullValue())
1254 // fold (mul x, -1) -> 0-x
1255 if (N1C && N1C->isAllOnesValue())
1256 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1257 DAG.getConstant(0, VT), N0);
1258 // fold (mul x, (1 << c)) -> x << c
1259 if (N1C && N1C->getAPIntValue().isPowerOf2())
1260 return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0,
1261 DAG.getConstant(N1C->getAPIntValue().logBase2(),
1262 getShiftAmountTy()));
1263 // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c
1264 if (N1C && (-N1C->getAPIntValue()).isPowerOf2()) {
1265 unsigned Log2Val = (-N1C->getAPIntValue()).logBase2();
1266 // FIXME: If the input is something that is easily negated (e.g. a
1267 // single-use add), we should put the negate there.
1268 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1269 DAG.getConstant(0, VT),
1270 DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0,
1271 DAG.getConstant(Log2Val, getShiftAmountTy())));
1273 // (mul (shl X, c1), c2) -> (mul X, c2 << c1)
1274 if (N1C && N0.getOpcode() == ISD::SHL &&
1275 isa<ConstantSDNode>(N0.getOperand(1))) {
1276 SDValue C3 = DAG.getNode(ISD::SHL, N->getDebugLoc(), VT,
1277 N1, N0.getOperand(1));
1278 AddToWorkList(C3.getNode());
1279 return DAG.getNode(ISD::MUL, N->getDebugLoc(), VT,
1280 N0.getOperand(0), C3);
1283 // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one
1286 SDValue Sh(0,0), Y(0,0);
1287 // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)).
1288 if (N0.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N0.getOperand(1)) &&
1289 N0.getNode()->hasOneUse()) {
1291 } else if (N1.getOpcode() == ISD::SHL &&
1292 isa<ConstantSDNode>(N1.getOperand(1)) &&
1293 N1.getNode()->hasOneUse()) {
1298 SDValue Mul = DAG.getNode(ISD::MUL, N->getDebugLoc(), VT,
1299 Sh.getOperand(0), Y);
1300 return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT,
1301 Mul, Sh.getOperand(1));
1305 // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2)
1306 if (N1C && N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() &&
1307 isa<ConstantSDNode>(N0.getOperand(1)))
1308 return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT,
1309 DAG.getNode(ISD::MUL, N0.getDebugLoc(), VT,
1310 N0.getOperand(0), N1),
1311 DAG.getNode(ISD::MUL, N1.getDebugLoc(), VT,
1312 N0.getOperand(1), N1));
1315 SDValue RMUL = ReassociateOps(ISD::MUL, N->getDebugLoc(), N0, N1);
1316 if (RMUL.getNode() != 0)
1322 SDValue DAGCombiner::visitSDIV(SDNode *N) {
1323 SDValue N0 = N->getOperand(0);
1324 SDValue N1 = N->getOperand(1);
1325 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1326 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1327 EVT VT = N->getValueType(0);
1330 if (VT.isVector()) {
1331 SDValue FoldedVOp = SimplifyVBinOp(N);
1332 if (FoldedVOp.getNode()) return FoldedVOp;
1335 // fold (sdiv c1, c2) -> c1/c2
1336 if (N0C && N1C && !N1C->isNullValue())
1337 return DAG.FoldConstantArithmetic(ISD::SDIV, VT, N0C, N1C);
1338 // fold (sdiv X, 1) -> X
1339 if (N1C && N1C->getSExtValue() == 1LL)
1341 // fold (sdiv X, -1) -> 0-X
1342 if (N1C && N1C->isAllOnesValue())
1343 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1344 DAG.getConstant(0, VT), N0);
1345 // If we know the sign bits of both operands are zero, strength reduce to a
1346 // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2
1347 if (!VT.isVector()) {
1348 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
1349 return DAG.getNode(ISD::UDIV, N->getDebugLoc(), N1.getValueType(),
1352 // fold (sdiv X, pow2) -> simple ops after legalize
1353 if (N1C && !N1C->isNullValue() && !TLI.isIntDivCheap() &&
1354 (isPowerOf2_64(N1C->getSExtValue()) ||
1355 isPowerOf2_64(-N1C->getSExtValue()))) {
1356 // If dividing by powers of two is cheap, then don't perform the following
1358 if (TLI.isPow2DivCheap())
1361 int64_t pow2 = N1C->getSExtValue();
1362 int64_t abs2 = pow2 > 0 ? pow2 : -pow2;
1363 unsigned lg2 = Log2_64(abs2);
1365 // Splat the sign bit into the register
1366 SDValue SGN = DAG.getNode(ISD::SRA, N->getDebugLoc(), VT, N0,
1367 DAG.getConstant(VT.getSizeInBits()-1,
1368 getShiftAmountTy()));
1369 AddToWorkList(SGN.getNode());
1371 // Add (N0 < 0) ? abs2 - 1 : 0;
1372 SDValue SRL = DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, SGN,
1373 DAG.getConstant(VT.getSizeInBits() - lg2,
1374 getShiftAmountTy()));
1375 SDValue ADD = DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N0, SRL);
1376 AddToWorkList(SRL.getNode());
1377 AddToWorkList(ADD.getNode()); // Divide by pow2
1378 SDValue SRA = DAG.getNode(ISD::SRA, N->getDebugLoc(), VT, ADD,
1379 DAG.getConstant(lg2, getShiftAmountTy()));
1381 // If we're dividing by a positive value, we're done. Otherwise, we must
1382 // negate the result.
1386 AddToWorkList(SRA.getNode());
1387 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1388 DAG.getConstant(0, VT), SRA);
1391 // if integer divide is expensive and we satisfy the requirements, emit an
1392 // alternate sequence.
1393 if (N1C && (N1C->getSExtValue() < -1 || N1C->getSExtValue() > 1) &&
1394 !TLI.isIntDivCheap()) {
1395 SDValue Op = BuildSDIV(N);
1396 if (Op.getNode()) return Op;
1400 if (N0.getOpcode() == ISD::UNDEF)
1401 return DAG.getConstant(0, VT);
1402 // X / undef -> undef
1403 if (N1.getOpcode() == ISD::UNDEF)
1409 SDValue DAGCombiner::visitUDIV(SDNode *N) {
1410 SDValue N0 = N->getOperand(0);
1411 SDValue N1 = N->getOperand(1);
1412 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1413 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1414 EVT VT = N->getValueType(0);
1417 if (VT.isVector()) {
1418 SDValue FoldedVOp = SimplifyVBinOp(N);
1419 if (FoldedVOp.getNode()) return FoldedVOp;
1422 // fold (udiv c1, c2) -> c1/c2
1423 if (N0C && N1C && !N1C->isNullValue())
1424 return DAG.FoldConstantArithmetic(ISD::UDIV, VT, N0C, N1C);
1425 // fold (udiv x, (1 << c)) -> x >>u c
1426 if (N1C && N1C->getAPIntValue().isPowerOf2())
1427 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0,
1428 DAG.getConstant(N1C->getAPIntValue().logBase2(),
1429 getShiftAmountTy()));
1430 // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2
1431 if (N1.getOpcode() == ISD::SHL) {
1432 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
1433 if (SHC->getAPIntValue().isPowerOf2()) {
1434 EVT ADDVT = N1.getOperand(1).getValueType();
1435 SDValue Add = DAG.getNode(ISD::ADD, N->getDebugLoc(), ADDVT,
1437 DAG.getConstant(SHC->getAPIntValue()
1440 AddToWorkList(Add.getNode());
1441 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0, Add);
1445 // fold (udiv x, c) -> alternate
1446 if (N1C && !N1C->isNullValue() && !TLI.isIntDivCheap()) {
1447 SDValue Op = BuildUDIV(N);
1448 if (Op.getNode()) return Op;
1452 if (N0.getOpcode() == ISD::UNDEF)
1453 return DAG.getConstant(0, VT);
1454 // X / undef -> undef
1455 if (N1.getOpcode() == ISD::UNDEF)
1461 SDValue DAGCombiner::visitSREM(SDNode *N) {
1462 SDValue N0 = N->getOperand(0);
1463 SDValue N1 = N->getOperand(1);
1464 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1465 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1466 EVT VT = N->getValueType(0);
1468 // fold (srem c1, c2) -> c1%c2
1469 if (N0C && N1C && !N1C->isNullValue())
1470 return DAG.FoldConstantArithmetic(ISD::SREM, VT, N0C, N1C);
1471 // If we know the sign bits of both operands are zero, strength reduce to a
1472 // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15
1473 if (!VT.isVector()) {
1474 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
1475 return DAG.getNode(ISD::UREM, N->getDebugLoc(), VT, N0, N1);
1478 // If X/C can be simplified by the division-by-constant logic, lower
1479 // X%C to the equivalent of X-X/C*C.
1480 if (N1C && !N1C->isNullValue()) {
1481 SDValue Div = DAG.getNode(ISD::SDIV, N->getDebugLoc(), VT, N0, N1);
1482 AddToWorkList(Div.getNode());
1483 SDValue OptimizedDiv = combine(Div.getNode());
1484 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
1485 SDValue Mul = DAG.getNode(ISD::MUL, N->getDebugLoc(), VT,
1487 SDValue Sub = DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N0, Mul);
1488 AddToWorkList(Mul.getNode());
1494 if (N0.getOpcode() == ISD::UNDEF)
1495 return DAG.getConstant(0, VT);
1496 // X % undef -> undef
1497 if (N1.getOpcode() == ISD::UNDEF)
1503 SDValue DAGCombiner::visitUREM(SDNode *N) {
1504 SDValue N0 = N->getOperand(0);
1505 SDValue N1 = N->getOperand(1);
1506 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1507 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1508 EVT VT = N->getValueType(0);
1510 // fold (urem c1, c2) -> c1%c2
1511 if (N0C && N1C && !N1C->isNullValue())
1512 return DAG.FoldConstantArithmetic(ISD::UREM, VT, N0C, N1C);
1513 // fold (urem x, pow2) -> (and x, pow2-1)
1514 if (N1C && !N1C->isNullValue() && N1C->getAPIntValue().isPowerOf2())
1515 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0,
1516 DAG.getConstant(N1C->getAPIntValue()-1,VT));
1517 // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1))
1518 if (N1.getOpcode() == ISD::SHL) {
1519 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
1520 if (SHC->getAPIntValue().isPowerOf2()) {
1522 DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N1,
1523 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()),
1525 AddToWorkList(Add.getNode());
1526 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0, Add);
1531 // If X/C can be simplified by the division-by-constant logic, lower
1532 // X%C to the equivalent of X-X/C*C.
1533 if (N1C && !N1C->isNullValue()) {
1534 SDValue Div = DAG.getNode(ISD::UDIV, N->getDebugLoc(), VT, N0, N1);
1535 AddToWorkList(Div.getNode());
1536 SDValue OptimizedDiv = combine(Div.getNode());
1537 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
1538 SDValue Mul = DAG.getNode(ISD::MUL, N->getDebugLoc(), VT,
1540 SDValue Sub = DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N0, Mul);
1541 AddToWorkList(Mul.getNode());
1547 if (N0.getOpcode() == ISD::UNDEF)
1548 return DAG.getConstant(0, VT);
1549 // X % undef -> undef
1550 if (N1.getOpcode() == ISD::UNDEF)
1556 SDValue DAGCombiner::visitMULHS(SDNode *N) {
1557 SDValue N0 = N->getOperand(0);
1558 SDValue N1 = N->getOperand(1);
1559 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1560 EVT VT = N->getValueType(0);
1562 // fold (mulhs x, 0) -> 0
1563 if (N1C && N1C->isNullValue())
1565 // fold (mulhs x, 1) -> (sra x, size(x)-1)
1566 if (N1C && N1C->getAPIntValue() == 1)
1567 return DAG.getNode(ISD::SRA, N->getDebugLoc(), N0.getValueType(), N0,
1568 DAG.getConstant(N0.getValueType().getSizeInBits() - 1,
1569 getShiftAmountTy()));
1570 // fold (mulhs x, undef) -> 0
1571 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
1572 return DAG.getConstant(0, VT);
1577 SDValue DAGCombiner::visitMULHU(SDNode *N) {
1578 SDValue N0 = N->getOperand(0);
1579 SDValue N1 = N->getOperand(1);
1580 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1581 EVT VT = N->getValueType(0);
1583 // fold (mulhu x, 0) -> 0
1584 if (N1C && N1C->isNullValue())
1586 // fold (mulhu x, 1) -> 0
1587 if (N1C && N1C->getAPIntValue() == 1)
1588 return DAG.getConstant(0, N0.getValueType());
1589 // fold (mulhu x, undef) -> 0
1590 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
1591 return DAG.getConstant(0, VT);
1596 /// SimplifyNodeWithTwoResults - Perform optimizations common to nodes that
1597 /// compute two values. LoOp and HiOp give the opcodes for the two computations
1598 /// that are being performed. Return true if a simplification was made.
1600 SDValue DAGCombiner::SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
1602 // If the high half is not needed, just compute the low half.
1603 bool HiExists = N->hasAnyUseOfValue(1);
1605 (!LegalOperations ||
1606 TLI.isOperationLegal(LoOp, N->getValueType(0)))) {
1607 SDValue Res = DAG.getNode(LoOp, N->getDebugLoc(), N->getValueType(0),
1608 N->op_begin(), N->getNumOperands());
1609 return CombineTo(N, Res, Res);
1612 // If the low half is not needed, just compute the high half.
1613 bool LoExists = N->hasAnyUseOfValue(0);
1615 (!LegalOperations ||
1616 TLI.isOperationLegal(HiOp, N->getValueType(1)))) {
1617 SDValue Res = DAG.getNode(HiOp, N->getDebugLoc(), N->getValueType(1),
1618 N->op_begin(), N->getNumOperands());
1619 return CombineTo(N, Res, Res);
1622 // If both halves are used, return as it is.
1623 if (LoExists && HiExists)
1626 // If the two computed results can be simplified separately, separate them.
1628 SDValue Lo = DAG.getNode(LoOp, N->getDebugLoc(), N->getValueType(0),
1629 N->op_begin(), N->getNumOperands());
1630 AddToWorkList(Lo.getNode());
1631 SDValue LoOpt = combine(Lo.getNode());
1632 if (LoOpt.getNode() && LoOpt.getNode() != Lo.getNode() &&
1633 (!LegalOperations ||
1634 TLI.isOperationLegal(LoOpt.getOpcode(), LoOpt.getValueType())))
1635 return CombineTo(N, LoOpt, LoOpt);
1639 SDValue Hi = DAG.getNode(HiOp, N->getDebugLoc(), N->getValueType(1),
1640 N->op_begin(), N->getNumOperands());
1641 AddToWorkList(Hi.getNode());
1642 SDValue HiOpt = combine(Hi.getNode());
1643 if (HiOpt.getNode() && HiOpt != Hi &&
1644 (!LegalOperations ||
1645 TLI.isOperationLegal(HiOpt.getOpcode(), HiOpt.getValueType())))
1646 return CombineTo(N, HiOpt, HiOpt);
1652 SDValue DAGCombiner::visitSMUL_LOHI(SDNode *N) {
1653 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHS);
1654 if (Res.getNode()) return Res;
1659 SDValue DAGCombiner::visitUMUL_LOHI(SDNode *N) {
1660 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHU);
1661 if (Res.getNode()) return Res;
1666 SDValue DAGCombiner::visitSDIVREM(SDNode *N) {
1667 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::SDIV, ISD::SREM);
1668 if (Res.getNode()) return Res;
1673 SDValue DAGCombiner::visitUDIVREM(SDNode *N) {
1674 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::UDIV, ISD::UREM);
1675 if (Res.getNode()) return Res;
1680 /// SimplifyBinOpWithSameOpcodeHands - If this is a binary operator with
1681 /// two operands of the same opcode, try to simplify it.
1682 SDValue DAGCombiner::SimplifyBinOpWithSameOpcodeHands(SDNode *N) {
1683 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
1684 EVT VT = N0.getValueType();
1685 assert(N0.getOpcode() == N1.getOpcode() && "Bad input!");
1687 // For each of OP in AND/OR/XOR:
1688 // fold (OP (zext x), (zext y)) -> (zext (OP x, y))
1689 // fold (OP (sext x), (sext y)) -> (sext (OP x, y))
1690 // fold (OP (aext x), (aext y)) -> (aext (OP x, y))
1691 // fold (OP (trunc x), (trunc y)) -> (trunc (OP x, y))
1693 // do not sink logical op inside of a vector extend, since it may combine
1695 EVT Op0VT = N0.getOperand(0).getValueType();
1696 if ((N0.getOpcode() == ISD::ZERO_EXTEND ||
1697 N0.getOpcode() == ISD::ANY_EXTEND ||
1698 N0.getOpcode() == ISD::SIGN_EXTEND ||
1699 (N0.getOpcode() == ISD::TRUNCATE && TLI.isTypeLegal(Op0VT))) &&
1701 Op0VT == N1.getOperand(0).getValueType() &&
1702 (!LegalOperations || TLI.isOperationLegal(N->getOpcode(), Op0VT))) {
1703 SDValue ORNode = DAG.getNode(N->getOpcode(), N0.getDebugLoc(),
1704 N0.getOperand(0).getValueType(),
1705 N0.getOperand(0), N1.getOperand(0));
1706 AddToWorkList(ORNode.getNode());
1707 return DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT, ORNode);
1710 // For each of OP in SHL/SRL/SRA/AND...
1711 // fold (and (OP x, z), (OP y, z)) -> (OP (and x, y), z)
1712 // fold (or (OP x, z), (OP y, z)) -> (OP (or x, y), z)
1713 // fold (xor (OP x, z), (OP y, z)) -> (OP (xor x, y), z)
1714 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL ||
1715 N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::AND) &&
1716 N0.getOperand(1) == N1.getOperand(1)) {
1717 SDValue ORNode = DAG.getNode(N->getOpcode(), N0.getDebugLoc(),
1718 N0.getOperand(0).getValueType(),
1719 N0.getOperand(0), N1.getOperand(0));
1720 AddToWorkList(ORNode.getNode());
1721 return DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT,
1722 ORNode, N0.getOperand(1));
1728 SDValue DAGCombiner::visitAND(SDNode *N) {
1729 SDValue N0 = N->getOperand(0);
1730 SDValue N1 = N->getOperand(1);
1731 SDValue LL, LR, RL, RR, CC0, CC1;
1732 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1733 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1734 EVT VT = N1.getValueType();
1735 unsigned BitWidth = VT.getSizeInBits();
1738 if (VT.isVector()) {
1739 SDValue FoldedVOp = SimplifyVBinOp(N);
1740 if (FoldedVOp.getNode()) return FoldedVOp;
1743 // fold (and x, undef) -> 0
1744 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
1745 return DAG.getConstant(0, VT);
1746 // fold (and c1, c2) -> c1&c2
1748 return DAG.FoldConstantArithmetic(ISD::AND, VT, N0C, N1C);
1749 // canonicalize constant to RHS
1751 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N1, N0);
1752 // fold (and x, -1) -> x
1753 if (N1C && N1C->isAllOnesValue())
1755 // if (and x, c) is known to be zero, return 0
1756 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
1757 APInt::getAllOnesValue(BitWidth)))
1758 return DAG.getConstant(0, VT);
1760 SDValue RAND = ReassociateOps(ISD::AND, N->getDebugLoc(), N0, N1);
1761 if (RAND.getNode() != 0)
1763 // fold (and (or x, 0xFFFF), 0xFF) -> 0xFF
1764 if (N1C && N0.getOpcode() == ISD::OR)
1765 if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
1766 if ((ORI->getAPIntValue() & N1C->getAPIntValue()) == N1C->getAPIntValue())
1768 // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits.
1769 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
1770 SDValue N0Op0 = N0.getOperand(0);
1771 APInt Mask = ~N1C->getAPIntValue();
1772 Mask.trunc(N0Op0.getValueSizeInBits());
1773 if (DAG.MaskedValueIsZero(N0Op0, Mask)) {
1774 SDValue Zext = DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(),
1775 N0.getValueType(), N0Op0);
1777 // Replace uses of the AND with uses of the Zero extend node.
1780 // We actually want to replace all uses of the any_extend with the
1781 // zero_extend, to avoid duplicating things. This will later cause this
1782 // AND to be folded.
1783 CombineTo(N0.getNode(), Zext);
1784 return SDValue(N, 0); // Return N so it doesn't get rechecked!
1787 // fold (and (setcc x), (setcc y)) -> (setcc (and x, y))
1788 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
1789 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
1790 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
1792 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
1793 LL.getValueType().isInteger()) {
1794 // fold (and (seteq X, 0), (seteq Y, 0)) -> (seteq (or X, Y), 0)
1795 if (cast<ConstantSDNode>(LR)->isNullValue() && Op1 == ISD::SETEQ) {
1796 SDValue ORNode = DAG.getNode(ISD::OR, N0.getDebugLoc(),
1797 LR.getValueType(), LL, RL);
1798 AddToWorkList(ORNode.getNode());
1799 return DAG.getSetCC(N->getDebugLoc(), VT, ORNode, LR, Op1);
1801 // fold (and (seteq X, -1), (seteq Y, -1)) -> (seteq (and X, Y), -1)
1802 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETEQ) {
1803 SDValue ANDNode = DAG.getNode(ISD::AND, N0.getDebugLoc(),
1804 LR.getValueType(), LL, RL);
1805 AddToWorkList(ANDNode.getNode());
1806 return DAG.getSetCC(N->getDebugLoc(), VT, ANDNode, LR, Op1);
1808 // fold (and (setgt X, -1), (setgt Y, -1)) -> (setgt (or X, Y), -1)
1809 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETGT) {
1810 SDValue ORNode = DAG.getNode(ISD::OR, N0.getDebugLoc(),
1811 LR.getValueType(), LL, RL);
1812 AddToWorkList(ORNode.getNode());
1813 return DAG.getSetCC(N->getDebugLoc(), VT, ORNode, LR, Op1);
1816 // canonicalize equivalent to ll == rl
1817 if (LL == RR && LR == RL) {
1818 Op1 = ISD::getSetCCSwappedOperands(Op1);
1821 if (LL == RL && LR == RR) {
1822 bool isInteger = LL.getValueType().isInteger();
1823 ISD::CondCode Result = ISD::getSetCCAndOperation(Op0, Op1, isInteger);
1824 if (Result != ISD::SETCC_INVALID &&
1825 (!LegalOperations || TLI.isCondCodeLegal(Result, LL.getValueType())))
1826 return DAG.getSetCC(N->getDebugLoc(), N0.getValueType(),
1831 // Simplify: (and (op x...), (op y...)) -> (op (and x, y))
1832 if (N0.getOpcode() == N1.getOpcode()) {
1833 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
1834 if (Tmp.getNode()) return Tmp;
1837 // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1)
1838 // fold (and (sra)) -> (and (srl)) when possible.
1839 if (!VT.isVector() &&
1840 SimplifyDemandedBits(SDValue(N, 0)))
1841 return SDValue(N, 0);
1843 // fold (zext_inreg (extload x)) -> (zextload x)
1844 if (ISD::isEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode())) {
1845 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
1846 EVT MemVT = LN0->getMemoryVT();
1847 // If we zero all the possible extended bits, then we can turn this into
1848 // a zextload if we are running before legalize or the operation is legal.
1849 unsigned BitWidth = N1.getValueSizeInBits();
1850 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
1851 BitWidth - MemVT.getSizeInBits())) &&
1852 ((!LegalOperations && !LN0->isVolatile()) ||
1853 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
1854 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N0.getDebugLoc(), VT,
1855 LN0->getChain(), LN0->getBasePtr(),
1857 LN0->getSrcValueOffset(), MemVT,
1858 LN0->isVolatile(), LN0->getAlignment());
1860 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
1861 return SDValue(N, 0); // Return N so it doesn't get rechecked!
1864 // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use
1865 if (ISD::isSEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
1867 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
1868 EVT MemVT = LN0->getMemoryVT();
1869 // If we zero all the possible extended bits, then we can turn this into
1870 // a zextload if we are running before legalize or the operation is legal.
1871 unsigned BitWidth = N1.getValueSizeInBits();
1872 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
1873 BitWidth - MemVT.getSizeInBits())) &&
1874 ((!LegalOperations && !LN0->isVolatile()) ||
1875 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
1876 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N0.getDebugLoc(), VT,
1878 LN0->getBasePtr(), LN0->getSrcValue(),
1879 LN0->getSrcValueOffset(), MemVT,
1880 LN0->isVolatile(), LN0->getAlignment());
1882 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
1883 return SDValue(N, 0); // Return N so it doesn't get rechecked!
1887 // fold (and (load x), 255) -> (zextload x, i8)
1888 // fold (and (extload x, i16), 255) -> (zextload x, i8)
1889 // fold (and (any_ext (extload x, i16)), 255) -> (zextload x, i8)
1890 if (N1C && (N0.getOpcode() == ISD::LOAD ||
1891 (N0.getOpcode() == ISD::ANY_EXTEND &&
1892 N0.getOperand(0).getOpcode() == ISD::LOAD))) {
1893 bool HasAnyExt = N0.getOpcode() == ISD::ANY_EXTEND;
1894 LoadSDNode *LN0 = HasAnyExt
1895 ? cast<LoadSDNode>(N0.getOperand(0))
1896 : cast<LoadSDNode>(N0);
1897 if (LN0->getExtensionType() != ISD::SEXTLOAD &&
1898 LN0->isUnindexed() && N0.hasOneUse() && LN0->hasOneUse()) {
1899 uint32_t ActiveBits = N1C->getAPIntValue().getActiveBits();
1900 if (ActiveBits > 0 && APIntOps::isMask(ActiveBits, N1C->getAPIntValue())){
1901 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits);
1902 EVT LoadedVT = LN0->getMemoryVT();
1904 if (ExtVT == LoadedVT &&
1905 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) {
1906 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
1909 DAG.getExtLoad(ISD::ZEXTLOAD, LN0->getDebugLoc(), LoadResultTy,
1910 LN0->getChain(), LN0->getBasePtr(),
1911 LN0->getSrcValue(), LN0->getSrcValueOffset(),
1912 ExtVT, LN0->isVolatile(), LN0->getAlignment());
1914 CombineTo(LN0, NewLoad, NewLoad.getValue(1));
1915 return SDValue(N, 0); // Return N so it doesn't get rechecked!
1918 // Do not change the width of a volatile load.
1919 // Do not generate loads of non-round integer types since these can
1920 // be expensive (and would be wrong if the type is not byte sized).
1921 if (!LN0->isVolatile() && LoadedVT.bitsGT(ExtVT) && ExtVT.isRound() &&
1922 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) {
1923 EVT PtrType = LN0->getOperand(1).getValueType();
1925 unsigned Alignment = LN0->getAlignment();
1926 SDValue NewPtr = LN0->getBasePtr();
1928 // For big endian targets, we need to add an offset to the pointer
1929 // to load the correct bytes. For little endian systems, we merely
1930 // need to read fewer bytes from the same pointer.
1931 if (TLI.isBigEndian()) {
1932 unsigned LVTStoreBytes = LoadedVT.getStoreSize();
1933 unsigned EVTStoreBytes = ExtVT.getStoreSize();
1934 unsigned PtrOff = LVTStoreBytes - EVTStoreBytes;
1935 NewPtr = DAG.getNode(ISD::ADD, LN0->getDebugLoc(), PtrType,
1936 NewPtr, DAG.getConstant(PtrOff, PtrType));
1937 Alignment = MinAlign(Alignment, PtrOff);
1940 AddToWorkList(NewPtr.getNode());
1942 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
1944 DAG.getExtLoad(ISD::ZEXTLOAD, LN0->getDebugLoc(), LoadResultTy,
1945 LN0->getChain(), NewPtr,
1946 LN0->getSrcValue(), LN0->getSrcValueOffset(),
1947 ExtVT, LN0->isVolatile(), Alignment);
1949 CombineTo(LN0, Load, Load.getValue(1));
1950 return SDValue(N, 0); // Return N so it doesn't get rechecked!
1959 SDValue DAGCombiner::visitOR(SDNode *N) {
1960 SDValue N0 = N->getOperand(0);
1961 SDValue N1 = N->getOperand(1);
1962 SDValue LL, LR, RL, RR, CC0, CC1;
1963 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1964 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1965 EVT VT = N1.getValueType();
1968 if (VT.isVector()) {
1969 SDValue FoldedVOp = SimplifyVBinOp(N);
1970 if (FoldedVOp.getNode()) return FoldedVOp;
1973 // fold (or x, undef) -> -1
1974 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF) {
1975 EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT;
1976 return DAG.getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()), VT);
1978 // fold (or c1, c2) -> c1|c2
1980 return DAG.FoldConstantArithmetic(ISD::OR, VT, N0C, N1C);
1981 // canonicalize constant to RHS
1983 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N1, N0);
1984 // fold (or x, 0) -> x
1985 if (N1C && N1C->isNullValue())
1987 // fold (or x, -1) -> -1
1988 if (N1C && N1C->isAllOnesValue())
1990 // fold (or x, c) -> c iff (x & ~c) == 0
1991 if (N1C && DAG.MaskedValueIsZero(N0, ~N1C->getAPIntValue()))
1994 SDValue ROR = ReassociateOps(ISD::OR, N->getDebugLoc(), N0, N1);
1995 if (ROR.getNode() != 0)
1997 // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2)
1998 if (N1C && N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
1999 isa<ConstantSDNode>(N0.getOperand(1))) {
2000 ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1));
2001 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
2002 DAG.getNode(ISD::OR, N0.getDebugLoc(), VT,
2003 N0.getOperand(0), N1),
2004 DAG.FoldConstantArithmetic(ISD::OR, VT, N1C, C1));
2006 // fold (or (setcc x), (setcc y)) -> (setcc (or x, y))
2007 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
2008 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
2009 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
2011 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
2012 LL.getValueType().isInteger()) {
2013 // fold (or (setne X, 0), (setne Y, 0)) -> (setne (or X, Y), 0)
2014 // fold (or (setlt X, 0), (setlt Y, 0)) -> (setne (or X, Y), 0)
2015 if (cast<ConstantSDNode>(LR)->isNullValue() &&
2016 (Op1 == ISD::SETNE || Op1 == ISD::SETLT)) {
2017 SDValue ORNode = DAG.getNode(ISD::OR, LR.getDebugLoc(),
2018 LR.getValueType(), LL, RL);
2019 AddToWorkList(ORNode.getNode());
2020 return DAG.getSetCC(N->getDebugLoc(), VT, ORNode, LR, Op1);
2022 // fold (or (setne X, -1), (setne Y, -1)) -> (setne (and X, Y), -1)
2023 // fold (or (setgt X, -1), (setgt Y -1)) -> (setgt (and X, Y), -1)
2024 if (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
2025 (Op1 == ISD::SETNE || Op1 == ISD::SETGT)) {
2026 SDValue ANDNode = DAG.getNode(ISD::AND, LR.getDebugLoc(),
2027 LR.getValueType(), LL, RL);
2028 AddToWorkList(ANDNode.getNode());
2029 return DAG.getSetCC(N->getDebugLoc(), VT, ANDNode, LR, Op1);
2032 // canonicalize equivalent to ll == rl
2033 if (LL == RR && LR == RL) {
2034 Op1 = ISD::getSetCCSwappedOperands(Op1);
2037 if (LL == RL && LR == RR) {
2038 bool isInteger = LL.getValueType().isInteger();
2039 ISD::CondCode Result = ISD::getSetCCOrOperation(Op0, Op1, isInteger);
2040 if (Result != ISD::SETCC_INVALID &&
2041 (!LegalOperations || TLI.isCondCodeLegal(Result, LL.getValueType())))
2042 return DAG.getSetCC(N->getDebugLoc(), N0.getValueType(),
2047 // Simplify: (or (op x...), (op y...)) -> (op (or x, y))
2048 if (N0.getOpcode() == N1.getOpcode()) {
2049 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
2050 if (Tmp.getNode()) return Tmp;
2053 // (or (and X, C1), (and Y, C2)) -> (and (or X, Y), C3) if possible.
2054 if (N0.getOpcode() == ISD::AND &&
2055 N1.getOpcode() == ISD::AND &&
2056 N0.getOperand(1).getOpcode() == ISD::Constant &&
2057 N1.getOperand(1).getOpcode() == ISD::Constant &&
2058 // Don't increase # computations.
2059 (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) {
2060 // We can only do this xform if we know that bits from X that are set in C2
2061 // but not in C1 are already zero. Likewise for Y.
2062 const APInt &LHSMask =
2063 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
2064 const APInt &RHSMask =
2065 cast<ConstantSDNode>(N1.getOperand(1))->getAPIntValue();
2067 if (DAG.MaskedValueIsZero(N0.getOperand(0), RHSMask&~LHSMask) &&
2068 DAG.MaskedValueIsZero(N1.getOperand(0), LHSMask&~RHSMask)) {
2069 SDValue X = DAG.getNode(ISD::OR, N0.getDebugLoc(), VT,
2070 N0.getOperand(0), N1.getOperand(0));
2071 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, X,
2072 DAG.getConstant(LHSMask | RHSMask, VT));
2076 // See if this is some rotate idiom.
2077 if (SDNode *Rot = MatchRotate(N0, N1, N->getDebugLoc()))
2078 return SDValue(Rot, 0);
2083 /// MatchRotateHalf - Match "(X shl/srl V1) & V2" where V2 may not be present.
2084 static bool MatchRotateHalf(SDValue Op, SDValue &Shift, SDValue &Mask) {
2085 if (Op.getOpcode() == ISD::AND) {
2086 if (isa<ConstantSDNode>(Op.getOperand(1))) {
2087 Mask = Op.getOperand(1);
2088 Op = Op.getOperand(0);
2094 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SHL) {
2102 // MatchRotate - Handle an 'or' of two operands. If this is one of the many
2103 // idioms for rotate, and if the target supports rotation instructions, generate
2105 SDNode *DAGCombiner::MatchRotate(SDValue LHS, SDValue RHS, DebugLoc DL) {
2106 // Must be a legal type. Expanded 'n promoted things won't work with rotates.
2107 EVT VT = LHS.getValueType();
2108 if (!TLI.isTypeLegal(VT)) return 0;
2110 // The target must have at least one rotate flavor.
2111 bool HasROTL = TLI.isOperationLegalOrCustom(ISD::ROTL, VT);
2112 bool HasROTR = TLI.isOperationLegalOrCustom(ISD::ROTR, VT);
2113 if (!HasROTL && !HasROTR) return 0;
2115 // Match "(X shl/srl V1) & V2" where V2 may not be present.
2116 SDValue LHSShift; // The shift.
2117 SDValue LHSMask; // AND value if any.
2118 if (!MatchRotateHalf(LHS, LHSShift, LHSMask))
2119 return 0; // Not part of a rotate.
2121 SDValue RHSShift; // The shift.
2122 SDValue RHSMask; // AND value if any.
2123 if (!MatchRotateHalf(RHS, RHSShift, RHSMask))
2124 return 0; // Not part of a rotate.
2126 if (LHSShift.getOperand(0) != RHSShift.getOperand(0))
2127 return 0; // Not shifting the same value.
2129 if (LHSShift.getOpcode() == RHSShift.getOpcode())
2130 return 0; // Shifts must disagree.
2132 // Canonicalize shl to left side in a shl/srl pair.
2133 if (RHSShift.getOpcode() == ISD::SHL) {
2134 std::swap(LHS, RHS);
2135 std::swap(LHSShift, RHSShift);
2136 std::swap(LHSMask , RHSMask );
2139 unsigned OpSizeInBits = VT.getSizeInBits();
2140 SDValue LHSShiftArg = LHSShift.getOperand(0);
2141 SDValue LHSShiftAmt = LHSShift.getOperand(1);
2142 SDValue RHSShiftAmt = RHSShift.getOperand(1);
2144 // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1)
2145 // fold (or (shl x, C1), (srl x, C2)) -> (rotr x, C2)
2146 if (LHSShiftAmt.getOpcode() == ISD::Constant &&
2147 RHSShiftAmt.getOpcode() == ISD::Constant) {
2148 uint64_t LShVal = cast<ConstantSDNode>(LHSShiftAmt)->getZExtValue();
2149 uint64_t RShVal = cast<ConstantSDNode>(RHSShiftAmt)->getZExtValue();
2150 if ((LShVal + RShVal) != OpSizeInBits)
2155 Rot = DAG.getNode(ISD::ROTL, DL, VT, LHSShiftArg, LHSShiftAmt);
2157 Rot = DAG.getNode(ISD::ROTR, DL, VT, LHSShiftArg, RHSShiftAmt);
2159 // If there is an AND of either shifted operand, apply it to the result.
2160 if (LHSMask.getNode() || RHSMask.getNode()) {
2161 APInt Mask = APInt::getAllOnesValue(OpSizeInBits);
2163 if (LHSMask.getNode()) {
2164 APInt RHSBits = APInt::getLowBitsSet(OpSizeInBits, LShVal);
2165 Mask &= cast<ConstantSDNode>(LHSMask)->getAPIntValue() | RHSBits;
2167 if (RHSMask.getNode()) {
2168 APInt LHSBits = APInt::getHighBitsSet(OpSizeInBits, RShVal);
2169 Mask &= cast<ConstantSDNode>(RHSMask)->getAPIntValue() | LHSBits;
2172 Rot = DAG.getNode(ISD::AND, DL, VT, Rot, DAG.getConstant(Mask, VT));
2175 return Rot.getNode();
2178 // If there is a mask here, and we have a variable shift, we can't be sure
2179 // that we're masking out the right stuff.
2180 if (LHSMask.getNode() || RHSMask.getNode())
2183 // fold (or (shl x, y), (srl x, (sub 32, y))) -> (rotl x, y)
2184 // fold (or (shl x, y), (srl x, (sub 32, y))) -> (rotr x, (sub 32, y))
2185 if (RHSShiftAmt.getOpcode() == ISD::SUB &&
2186 LHSShiftAmt == RHSShiftAmt.getOperand(1)) {
2187 if (ConstantSDNode *SUBC =
2188 dyn_cast<ConstantSDNode>(RHSShiftAmt.getOperand(0))) {
2189 if (SUBC->getAPIntValue() == OpSizeInBits) {
2191 return DAG.getNode(ISD::ROTL, DL, VT,
2192 LHSShiftArg, LHSShiftAmt).getNode();
2194 return DAG.getNode(ISD::ROTR, DL, VT,
2195 LHSShiftArg, RHSShiftAmt).getNode();
2200 // fold (or (shl x, (sub 32, y)), (srl x, r)) -> (rotr x, y)
2201 // fold (or (shl x, (sub 32, y)), (srl x, r)) -> (rotl x, (sub 32, y))
2202 if (LHSShiftAmt.getOpcode() == ISD::SUB &&
2203 RHSShiftAmt == LHSShiftAmt.getOperand(1)) {
2204 if (ConstantSDNode *SUBC =
2205 dyn_cast<ConstantSDNode>(LHSShiftAmt.getOperand(0))) {
2206 if (SUBC->getAPIntValue() == OpSizeInBits) {
2208 return DAG.getNode(ISD::ROTR, DL, VT,
2209 LHSShiftArg, RHSShiftAmt).getNode();
2211 return DAG.getNode(ISD::ROTL, DL, VT,
2212 LHSShiftArg, LHSShiftAmt).getNode();
2217 // Look for sign/zext/any-extended or truncate cases:
2218 if ((LHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND
2219 || LHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND
2220 || LHSShiftAmt.getOpcode() == ISD::ANY_EXTEND
2221 || LHSShiftAmt.getOpcode() == ISD::TRUNCATE) &&
2222 (RHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND
2223 || RHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND
2224 || RHSShiftAmt.getOpcode() == ISD::ANY_EXTEND
2225 || RHSShiftAmt.getOpcode() == ISD::TRUNCATE)) {
2226 SDValue LExtOp0 = LHSShiftAmt.getOperand(0);
2227 SDValue RExtOp0 = RHSShiftAmt.getOperand(0);
2228 if (RExtOp0.getOpcode() == ISD::SUB &&
2229 RExtOp0.getOperand(1) == LExtOp0) {
2230 // fold (or (shl x, (*ext y)), (srl x, (*ext (sub 32, y)))) ->
2232 // fold (or (shl x, (*ext y)), (srl x, (*ext (sub 32, y)))) ->
2233 // (rotr x, (sub 32, y))
2234 if (ConstantSDNode *SUBC =
2235 dyn_cast<ConstantSDNode>(RExtOp0.getOperand(0))) {
2236 if (SUBC->getAPIntValue() == OpSizeInBits) {
2237 return DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT,
2239 HasROTL ? LHSShiftAmt : RHSShiftAmt).getNode();
2242 } else if (LExtOp0.getOpcode() == ISD::SUB &&
2243 RExtOp0 == LExtOp0.getOperand(1)) {
2244 // fold (or (shl x, (*ext (sub 32, y))), (srl x, (*ext y))) ->
2246 // fold (or (shl x, (*ext (sub 32, y))), (srl x, (*ext y))) ->
2247 // (rotl x, (sub 32, y))
2248 if (ConstantSDNode *SUBC =
2249 dyn_cast<ConstantSDNode>(LExtOp0.getOperand(0))) {
2250 if (SUBC->getAPIntValue() == OpSizeInBits) {
2251 return DAG.getNode(HasROTR ? ISD::ROTR : ISD::ROTL, DL, VT,
2253 HasROTR ? RHSShiftAmt : LHSShiftAmt).getNode();
2262 SDValue DAGCombiner::visitXOR(SDNode *N) {
2263 SDValue N0 = N->getOperand(0);
2264 SDValue N1 = N->getOperand(1);
2265 SDValue LHS, RHS, CC;
2266 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2267 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2268 EVT VT = N0.getValueType();
2271 if (VT.isVector()) {
2272 SDValue FoldedVOp = SimplifyVBinOp(N);
2273 if (FoldedVOp.getNode()) return FoldedVOp;
2276 // fold (xor undef, undef) -> 0. This is a common idiom (misuse).
2277 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
2278 return DAG.getConstant(0, VT);
2279 // fold (xor x, undef) -> undef
2280 if (N0.getOpcode() == ISD::UNDEF)
2282 if (N1.getOpcode() == ISD::UNDEF)
2284 // fold (xor c1, c2) -> c1^c2
2286 return DAG.FoldConstantArithmetic(ISD::XOR, VT, N0C, N1C);
2287 // canonicalize constant to RHS
2289 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N1, N0);
2290 // fold (xor x, 0) -> x
2291 if (N1C && N1C->isNullValue())
2294 SDValue RXOR = ReassociateOps(ISD::XOR, N->getDebugLoc(), N0, N1);
2295 if (RXOR.getNode() != 0)
2298 // fold !(x cc y) -> (x !cc y)
2299 if (N1C && N1C->getAPIntValue() == 1 && isSetCCEquivalent(N0, LHS, RHS, CC)) {
2300 bool isInt = LHS.getValueType().isInteger();
2301 ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
2304 if (!LegalOperations || TLI.isCondCodeLegal(NotCC, LHS.getValueType())) {
2305 switch (N0.getOpcode()) {
2307 llvm_unreachable("Unhandled SetCC Equivalent!");
2309 return DAG.getSetCC(N->getDebugLoc(), VT, LHS, RHS, NotCC);
2310 case ISD::SELECT_CC:
2311 return DAG.getSelectCC(N->getDebugLoc(), LHS, RHS, N0.getOperand(2),
2312 N0.getOperand(3), NotCC);
2317 // fold (not (zext (setcc x, y))) -> (zext (not (setcc x, y)))
2318 if (N1C && N1C->getAPIntValue() == 1 && N0.getOpcode() == ISD::ZERO_EXTEND &&
2319 N0.getNode()->hasOneUse() &&
2320 isSetCCEquivalent(N0.getOperand(0), LHS, RHS, CC)){
2321 SDValue V = N0.getOperand(0);
2322 V = DAG.getNode(ISD::XOR, N0.getDebugLoc(), V.getValueType(), V,
2323 DAG.getConstant(1, V.getValueType()));
2324 AddToWorkList(V.getNode());
2325 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, V);
2328 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are setcc
2329 if (N1C && N1C->getAPIntValue() == 1 && VT == MVT::i1 &&
2330 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
2331 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
2332 if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) {
2333 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
2334 LHS = DAG.getNode(ISD::XOR, LHS.getDebugLoc(), VT, LHS, N1); // LHS = ~LHS
2335 RHS = DAG.getNode(ISD::XOR, RHS.getDebugLoc(), VT, RHS, N1); // RHS = ~RHS
2336 AddToWorkList(LHS.getNode()); AddToWorkList(RHS.getNode());
2337 return DAG.getNode(NewOpcode, N->getDebugLoc(), VT, LHS, RHS);
2340 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are constants
2341 if (N1C && N1C->isAllOnesValue() &&
2342 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
2343 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
2344 if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) {
2345 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
2346 LHS = DAG.getNode(ISD::XOR, LHS.getDebugLoc(), VT, LHS, N1); // LHS = ~LHS
2347 RHS = DAG.getNode(ISD::XOR, RHS.getDebugLoc(), VT, RHS, N1); // RHS = ~RHS
2348 AddToWorkList(LHS.getNode()); AddToWorkList(RHS.getNode());
2349 return DAG.getNode(NewOpcode, N->getDebugLoc(), VT, LHS, RHS);
2352 // fold (xor (xor x, c1), c2) -> (xor x, (xor c1, c2))
2353 if (N1C && N0.getOpcode() == ISD::XOR) {
2354 ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0));
2355 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
2357 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N0.getOperand(1),
2358 DAG.getConstant(N1C->getAPIntValue() ^
2359 N00C->getAPIntValue(), VT));
2361 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N0.getOperand(0),
2362 DAG.getConstant(N1C->getAPIntValue() ^
2363 N01C->getAPIntValue(), VT));
2365 // fold (xor x, x) -> 0
2367 if (!VT.isVector()) {
2368 return DAG.getConstant(0, VT);
2369 } else if (!LegalOperations || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT)){
2370 // Produce a vector of zeros.
2371 SDValue El = DAG.getConstant(0, VT.getVectorElementType());
2372 std::vector<SDValue> Ops(VT.getVectorNumElements(), El);
2373 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), VT,
2374 &Ops[0], Ops.size());
2378 // Simplify: xor (op x...), (op y...) -> (op (xor x, y))
2379 if (N0.getOpcode() == N1.getOpcode()) {
2380 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
2381 if (Tmp.getNode()) return Tmp;
2384 // Simplify the expression using non-local knowledge.
2385 if (!VT.isVector() &&
2386 SimplifyDemandedBits(SDValue(N, 0)))
2387 return SDValue(N, 0);
2392 /// visitShiftByConstant - Handle transforms common to the three shifts, when
2393 /// the shift amount is a constant.
2394 SDValue DAGCombiner::visitShiftByConstant(SDNode *N, unsigned Amt) {
2395 SDNode *LHS = N->getOperand(0).getNode();
2396 if (!LHS->hasOneUse()) return SDValue();
2398 // We want to pull some binops through shifts, so that we have (and (shift))
2399 // instead of (shift (and)), likewise for add, or, xor, etc. This sort of
2400 // thing happens with address calculations, so it's important to canonicalize
2402 bool HighBitSet = false; // Can we transform this if the high bit is set?
2404 switch (LHS->getOpcode()) {
2405 default: return SDValue();
2408 HighBitSet = false; // We can only transform sra if the high bit is clear.
2411 HighBitSet = true; // We can only transform sra if the high bit is set.
2414 if (N->getOpcode() != ISD::SHL)
2415 return SDValue(); // only shl(add) not sr[al](add).
2416 HighBitSet = false; // We can only transform sra if the high bit is clear.
2420 // We require the RHS of the binop to be a constant as well.
2421 ConstantSDNode *BinOpCst = dyn_cast<ConstantSDNode>(LHS->getOperand(1));
2422 if (!BinOpCst) return SDValue();
2424 // FIXME: disable this unless the input to the binop is a shift by a constant.
2425 // If it is not a shift, it pessimizes some common cases like:
2427 // void foo(int *X, int i) { X[i & 1235] = 1; }
2428 // int bar(int *X, int i) { return X[i & 255]; }
2429 SDNode *BinOpLHSVal = LHS->getOperand(0).getNode();
2430 if ((BinOpLHSVal->getOpcode() != ISD::SHL &&
2431 BinOpLHSVal->getOpcode() != ISD::SRA &&
2432 BinOpLHSVal->getOpcode() != ISD::SRL) ||
2433 !isa<ConstantSDNode>(BinOpLHSVal->getOperand(1)))
2436 EVT VT = N->getValueType(0);
2438 // If this is a signed shift right, and the high bit is modified by the
2439 // logical operation, do not perform the transformation. The highBitSet
2440 // boolean indicates the value of the high bit of the constant which would
2441 // cause it to be modified for this operation.
2442 if (N->getOpcode() == ISD::SRA) {
2443 bool BinOpRHSSignSet = BinOpCst->getAPIntValue().isNegative();
2444 if (BinOpRHSSignSet != HighBitSet)
2448 // Fold the constants, shifting the binop RHS by the shift amount.
2449 SDValue NewRHS = DAG.getNode(N->getOpcode(), LHS->getOperand(1).getDebugLoc(),
2451 LHS->getOperand(1), N->getOperand(1));
2453 // Create the new shift.
2454 SDValue NewShift = DAG.getNode(N->getOpcode(), LHS->getOperand(0).getDebugLoc(),
2455 VT, LHS->getOperand(0), N->getOperand(1));
2457 // Create the new binop.
2458 return DAG.getNode(LHS->getOpcode(), N->getDebugLoc(), VT, NewShift, NewRHS);
2461 SDValue DAGCombiner::visitSHL(SDNode *N) {
2462 SDValue N0 = N->getOperand(0);
2463 SDValue N1 = N->getOperand(1);
2464 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2465 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2466 EVT VT = N0.getValueType();
2467 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
2469 // fold (shl c1, c2) -> c1<<c2
2471 return DAG.FoldConstantArithmetic(ISD::SHL, VT, N0C, N1C);
2472 // fold (shl 0, x) -> 0
2473 if (N0C && N0C->isNullValue())
2475 // fold (shl x, c >= size(x)) -> undef
2476 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
2477 return DAG.getUNDEF(VT);
2478 // fold (shl x, 0) -> x
2479 if (N1C && N1C->isNullValue())
2481 // if (shl x, c) is known to be zero, return 0
2482 if (DAG.MaskedValueIsZero(SDValue(N, 0),
2483 APInt::getAllOnesValue(OpSizeInBits)))
2484 return DAG.getConstant(0, VT);
2485 // fold (shl x, (trunc (and y, c))) -> (shl x, (and (trunc y), (trunc c))).
2486 if (N1.getOpcode() == ISD::TRUNCATE &&
2487 N1.getOperand(0).getOpcode() == ISD::AND &&
2488 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
2489 SDValue N101 = N1.getOperand(0).getOperand(1);
2490 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
2491 EVT TruncVT = N1.getValueType();
2492 SDValue N100 = N1.getOperand(0).getOperand(0);
2493 APInt TruncC = N101C->getAPIntValue();
2494 TruncC.trunc(TruncVT.getSizeInBits());
2495 return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0,
2496 DAG.getNode(ISD::AND, N->getDebugLoc(), TruncVT,
2497 DAG.getNode(ISD::TRUNCATE,
2500 DAG.getConstant(TruncC, TruncVT)));
2504 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
2505 return SDValue(N, 0);
2507 // fold (shl (shl x, c1), c2) -> 0 or (shl x, (add c1, c2))
2508 if (N1C && N0.getOpcode() == ISD::SHL &&
2509 N0.getOperand(1).getOpcode() == ISD::Constant) {
2510 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
2511 uint64_t c2 = N1C->getZExtValue();
2512 if (c1 + c2 > OpSizeInBits)
2513 return DAG.getConstant(0, VT);
2514 return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0.getOperand(0),
2515 DAG.getConstant(c1 + c2, N1.getValueType()));
2517 // fold (shl (srl x, c1), c2) -> (shl (and x, (shl -1, c1)), (sub c2, c1)) or
2518 // (srl (and x, (shl -1, c1)), (sub c1, c2))
2519 if (N1C && N0.getOpcode() == ISD::SRL &&
2520 N0.getOperand(1).getOpcode() == ISD::Constant) {
2521 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
2522 if (c1 < VT.getSizeInBits()) {
2523 uint64_t c2 = N1C->getZExtValue();
2524 SDValue HiBitsMask =
2525 DAG.getConstant(APInt::getHighBitsSet(VT.getSizeInBits(),
2526 VT.getSizeInBits() - c1),
2528 SDValue Mask = DAG.getNode(ISD::AND, N0.getDebugLoc(), VT,
2532 return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, Mask,
2533 DAG.getConstant(c2-c1, N1.getValueType()));
2535 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, Mask,
2536 DAG.getConstant(c1-c2, N1.getValueType()));
2539 // fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1))
2540 if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1)) {
2541 SDValue HiBitsMask =
2542 DAG.getConstant(APInt::getHighBitsSet(VT.getSizeInBits(),
2543 VT.getSizeInBits() -
2544 N1C->getZExtValue()),
2546 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0.getOperand(0),
2550 return N1C ? visitShiftByConstant(N, N1C->getZExtValue()) : SDValue();
2553 SDValue DAGCombiner::visitSRA(SDNode *N) {
2554 SDValue N0 = N->getOperand(0);
2555 SDValue N1 = N->getOperand(1);
2556 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2557 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2558 EVT VT = N0.getValueType();
2559 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
2561 // fold (sra c1, c2) -> (sra c1, c2)
2563 return DAG.FoldConstantArithmetic(ISD::SRA, VT, N0C, N1C);
2564 // fold (sra 0, x) -> 0
2565 if (N0C && N0C->isNullValue())
2567 // fold (sra -1, x) -> -1
2568 if (N0C && N0C->isAllOnesValue())
2570 // fold (sra x, (setge c, size(x))) -> undef
2571 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
2572 return DAG.getUNDEF(VT);
2573 // fold (sra x, 0) -> x
2574 if (N1C && N1C->isNullValue())
2576 // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports
2578 if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) {
2579 unsigned LowBits = OpSizeInBits - (unsigned)N1C->getZExtValue();
2580 EVT EVT = EVT::getIntegerVT(*DAG.getContext(), LowBits);
2581 if ((!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, EVT)))
2582 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT,
2583 N0.getOperand(0), DAG.getValueType(EVT));
2586 // fold (sra (sra x, c1), c2) -> (sra x, (add c1, c2))
2587 if (N1C && N0.getOpcode() == ISD::SRA) {
2588 if (ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
2589 unsigned Sum = N1C->getZExtValue() + C1->getZExtValue();
2590 if (Sum >= OpSizeInBits) Sum = OpSizeInBits-1;
2591 return DAG.getNode(ISD::SRA, N->getDebugLoc(), VT, N0.getOperand(0),
2592 DAG.getConstant(Sum, N1C->getValueType(0)));
2596 // fold (sra (shl X, m), (sub result_size, n))
2597 // -> (sign_extend (trunc (shl X, (sub (sub result_size, n), m)))) for
2598 // result_size - n != m.
2599 // If truncate is free for the target sext(shl) is likely to result in better
2601 if (N0.getOpcode() == ISD::SHL) {
2602 // Get the two constanst of the shifts, CN0 = m, CN = n.
2603 const ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
2605 // Determine what the truncate's result bitsize and type would be.
2607 EVT::getIntegerVT(*DAG.getContext(), OpSizeInBits - N1C->getZExtValue());
2608 // Determine the residual right-shift amount.
2609 signed ShiftAmt = N1C->getZExtValue() - N01C->getZExtValue();
2611 // If the shift is not a no-op (in which case this should be just a sign
2612 // extend already), the truncated to type is legal, sign_extend is legal
2613 // on that type, and the the truncate to that type is both legal and free,
2614 // perform the transform.
2615 if ((ShiftAmt > 0) &&
2616 TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND, TruncVT) &&
2617 TLI.isOperationLegalOrCustom(ISD::TRUNCATE, VT) &&
2618 TLI.isTruncateFree(VT, TruncVT)) {
2620 SDValue Amt = DAG.getConstant(ShiftAmt, getShiftAmountTy());
2621 SDValue Shift = DAG.getNode(ISD::SRL, N0.getDebugLoc(), VT,
2622 N0.getOperand(0), Amt);
2623 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(), TruncVT,
2625 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(),
2626 N->getValueType(0), Trunc);
2631 // fold (sra x, (trunc (and y, c))) -> (sra x, (and (trunc y), (trunc c))).
2632 if (N1.getOpcode() == ISD::TRUNCATE &&
2633 N1.getOperand(0).getOpcode() == ISD::AND &&
2634 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
2635 SDValue N101 = N1.getOperand(0).getOperand(1);
2636 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
2637 EVT TruncVT = N1.getValueType();
2638 SDValue N100 = N1.getOperand(0).getOperand(0);
2639 APInt TruncC = N101C->getAPIntValue();
2640 TruncC.trunc(TruncVT.getScalarType().getSizeInBits());
2641 return DAG.getNode(ISD::SRA, N->getDebugLoc(), VT, N0,
2642 DAG.getNode(ISD::AND, N->getDebugLoc(),
2644 DAG.getNode(ISD::TRUNCATE,
2647 DAG.getConstant(TruncC, TruncVT)));
2651 // Simplify, based on bits shifted out of the LHS.
2652 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
2653 return SDValue(N, 0);
2656 // If the sign bit is known to be zero, switch this to a SRL.
2657 if (DAG.SignBitIsZero(N0))
2658 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0, N1);
2660 return N1C ? visitShiftByConstant(N, N1C->getZExtValue()) : SDValue();
2663 SDValue DAGCombiner::visitSRL(SDNode *N) {
2664 SDValue N0 = N->getOperand(0);
2665 SDValue N1 = N->getOperand(1);
2666 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2667 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2668 EVT VT = N0.getValueType();
2669 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
2671 // fold (srl c1, c2) -> c1 >>u c2
2673 return DAG.FoldConstantArithmetic(ISD::SRL, VT, N0C, N1C);
2674 // fold (srl 0, x) -> 0
2675 if (N0C && N0C->isNullValue())
2677 // fold (srl x, c >= size(x)) -> undef
2678 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
2679 return DAG.getUNDEF(VT);
2680 // fold (srl x, 0) -> x
2681 if (N1C && N1C->isNullValue())
2683 // if (srl x, c) is known to be zero, return 0
2684 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
2685 APInt::getAllOnesValue(OpSizeInBits)))
2686 return DAG.getConstant(0, VT);
2688 // fold (srl (srl x, c1), c2) -> 0 or (srl x, (add c1, c2))
2689 if (N1C && N0.getOpcode() == ISD::SRL &&
2690 N0.getOperand(1).getOpcode() == ISD::Constant) {
2691 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
2692 uint64_t c2 = N1C->getZExtValue();
2693 if (c1 + c2 > OpSizeInBits)
2694 return DAG.getConstant(0, VT);
2695 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0.getOperand(0),
2696 DAG.getConstant(c1 + c2, N1.getValueType()));
2699 // fold (srl (anyextend x), c) -> (anyextend (srl x, c))
2700 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
2701 // Shifting in all undef bits?
2702 EVT SmallVT = N0.getOperand(0).getValueType();
2703 if (N1C->getZExtValue() >= SmallVT.getSizeInBits())
2704 return DAG.getUNDEF(VT);
2706 SDValue SmallShift = DAG.getNode(ISD::SRL, N0.getDebugLoc(), SmallVT,
2707 N0.getOperand(0), N1);
2708 AddToWorkList(SmallShift.getNode());
2709 return DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, SmallShift);
2712 // fold (srl (sra X, Y), 31) -> (srl X, 31). This srl only looks at the sign
2713 // bit, which is unmodified by sra.
2714 if (N1C && N1C->getZExtValue() + 1 == VT.getSizeInBits()) {
2715 if (N0.getOpcode() == ISD::SRA)
2716 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0.getOperand(0), N1);
2719 // fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit).
2720 if (N1C && N0.getOpcode() == ISD::CTLZ &&
2721 N1C->getAPIntValue() == Log2_32(VT.getSizeInBits())) {
2722 APInt KnownZero, KnownOne;
2723 APInt Mask = APInt::getAllOnesValue(VT.getSizeInBits());
2724 DAG.ComputeMaskedBits(N0.getOperand(0), Mask, KnownZero, KnownOne);
2726 // If any of the input bits are KnownOne, then the input couldn't be all
2727 // zeros, thus the result of the srl will always be zero.
2728 if (KnownOne.getBoolValue()) return DAG.getConstant(0, VT);
2730 // If all of the bits input the to ctlz node are known to be zero, then
2731 // the result of the ctlz is "32" and the result of the shift is one.
2732 APInt UnknownBits = ~KnownZero & Mask;
2733 if (UnknownBits == 0) return DAG.getConstant(1, VT);
2735 // Otherwise, check to see if there is exactly one bit input to the ctlz.
2736 if ((UnknownBits & (UnknownBits - 1)) == 0) {
2737 // Okay, we know that only that the single bit specified by UnknownBits
2738 // could be set on input to the CTLZ node. If this bit is set, the SRL
2739 // will return 0, if it is clear, it returns 1. Change the CTLZ/SRL pair
2740 // to an SRL/XOR pair, which is likely to simplify more.
2741 unsigned ShAmt = UnknownBits.countTrailingZeros();
2742 SDValue Op = N0.getOperand(0);
2745 Op = DAG.getNode(ISD::SRL, N0.getDebugLoc(), VT, Op,
2746 DAG.getConstant(ShAmt, getShiftAmountTy()));
2747 AddToWorkList(Op.getNode());
2750 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT,
2751 Op, DAG.getConstant(1, VT));
2755 // fold (srl x, (trunc (and y, c))) -> (srl x, (and (trunc y), (trunc c))).
2756 if (N1.getOpcode() == ISD::TRUNCATE &&
2757 N1.getOperand(0).getOpcode() == ISD::AND &&
2758 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
2759 SDValue N101 = N1.getOperand(0).getOperand(1);
2760 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
2761 EVT TruncVT = N1.getValueType();
2762 SDValue N100 = N1.getOperand(0).getOperand(0);
2763 APInt TruncC = N101C->getAPIntValue();
2764 TruncC.trunc(TruncVT.getSizeInBits());
2765 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0,
2766 DAG.getNode(ISD::AND, N->getDebugLoc(),
2768 DAG.getNode(ISD::TRUNCATE,
2771 DAG.getConstant(TruncC, TruncVT)));
2775 // fold operands of srl based on knowledge that the low bits are not
2777 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
2778 return SDValue(N, 0);
2781 SDValue NewSRL = visitShiftByConstant(N, N1C->getZExtValue());
2782 if (NewSRL.getNode())
2786 // Here is a common situation. We want to optimize:
2789 // %b = and i32 %a, 2
2790 // %c = srl i32 %b, 1
2791 // brcond i32 %c ...
2797 // %c = setcc eq %b, 0
2800 // However when after the source operand of SRL is optimized into AND, the SRL
2801 // itself may not be optimized further. Look for it and add the BRCOND into
2803 if (N->hasOneUse()) {
2804 SDNode *Use = *N->use_begin();
2805 if (Use->getOpcode() == ISD::BRCOND)
2807 else if (Use->getOpcode() == ISD::TRUNCATE && Use->hasOneUse()) {
2808 // Also look pass the truncate.
2809 Use = *Use->use_begin();
2810 if (Use->getOpcode() == ISD::BRCOND)
2818 SDValue DAGCombiner::visitCTLZ(SDNode *N) {
2819 SDValue N0 = N->getOperand(0);
2820 EVT VT = N->getValueType(0);
2822 // fold (ctlz c1) -> c2
2823 if (isa<ConstantSDNode>(N0))
2824 return DAG.getNode(ISD::CTLZ, N->getDebugLoc(), VT, N0);
2828 SDValue DAGCombiner::visitCTTZ(SDNode *N) {
2829 SDValue N0 = N->getOperand(0);
2830 EVT VT = N->getValueType(0);
2832 // fold (cttz c1) -> c2
2833 if (isa<ConstantSDNode>(N0))
2834 return DAG.getNode(ISD::CTTZ, N->getDebugLoc(), VT, N0);
2838 SDValue DAGCombiner::visitCTPOP(SDNode *N) {
2839 SDValue N0 = N->getOperand(0);
2840 EVT VT = N->getValueType(0);
2842 // fold (ctpop c1) -> c2
2843 if (isa<ConstantSDNode>(N0))
2844 return DAG.getNode(ISD::CTPOP, N->getDebugLoc(), VT, N0);
2848 SDValue DAGCombiner::visitSELECT(SDNode *N) {
2849 SDValue N0 = N->getOperand(0);
2850 SDValue N1 = N->getOperand(1);
2851 SDValue N2 = N->getOperand(2);
2852 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2853 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2854 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
2855 EVT VT = N->getValueType(0);
2856 EVT VT0 = N0.getValueType();
2858 // fold (select C, X, X) -> X
2861 // fold (select true, X, Y) -> X
2862 if (N0C && !N0C->isNullValue())
2864 // fold (select false, X, Y) -> Y
2865 if (N0C && N0C->isNullValue())
2867 // fold (select C, 1, X) -> (or C, X)
2868 if (VT == MVT::i1 && N1C && N1C->getAPIntValue() == 1)
2869 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N0, N2);
2870 // fold (select C, 0, 1) -> (xor C, 1)
2871 if (VT.isInteger() &&
2874 TLI.getBooleanContents() == TargetLowering::ZeroOrOneBooleanContent)) &&
2875 N1C && N2C && N1C->isNullValue() && N2C->getAPIntValue() == 1) {
2878 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT0,
2879 N0, DAG.getConstant(1, VT0));
2880 XORNode = DAG.getNode(ISD::XOR, N0.getDebugLoc(), VT0,
2881 N0, DAG.getConstant(1, VT0));
2882 AddToWorkList(XORNode.getNode());
2884 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, XORNode);
2885 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, XORNode);
2887 // fold (select C, 0, X) -> (and (not C), X)
2888 if (VT == VT0 && VT == MVT::i1 && N1C && N1C->isNullValue()) {
2889 SDValue NOTNode = DAG.getNOT(N0.getDebugLoc(), N0, VT);
2890 AddToWorkList(NOTNode.getNode());
2891 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, NOTNode, N2);
2893 // fold (select C, X, 1) -> (or (not C), X)
2894 if (VT == VT0 && VT == MVT::i1 && N2C && N2C->getAPIntValue() == 1) {
2895 SDValue NOTNode = DAG.getNOT(N0.getDebugLoc(), N0, VT);
2896 AddToWorkList(NOTNode.getNode());
2897 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, NOTNode, N1);
2899 // fold (select C, X, 0) -> (and C, X)
2900 if (VT == MVT::i1 && N2C && N2C->isNullValue())
2901 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0, N1);
2902 // fold (select X, X, Y) -> (or X, Y)
2903 // fold (select X, 1, Y) -> (or X, Y)
2904 if (VT == MVT::i1 && (N0 == N1 || (N1C && N1C->getAPIntValue() == 1)))
2905 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N0, N2);
2906 // fold (select X, Y, X) -> (and X, Y)
2907 // fold (select X, Y, 0) -> (and X, Y)
2908 if (VT == MVT::i1 && (N0 == N2 || (N2C && N2C->getAPIntValue() == 0)))
2909 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0, N1);
2911 // If we can fold this based on the true/false value, do so.
2912 if (SimplifySelectOps(N, N1, N2))
2913 return SDValue(N, 0); // Don't revisit N.
2915 // fold selects based on a setcc into other things, such as min/max/abs
2916 if (N0.getOpcode() == ISD::SETCC) {
2918 // Check against MVT::Other for SELECT_CC, which is a workaround for targets
2919 // having to say they don't support SELECT_CC on every type the DAG knows
2920 // about, since there is no way to mark an opcode illegal at all value types
2921 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, MVT::Other) &&
2922 TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT))
2923 return DAG.getNode(ISD::SELECT_CC, N->getDebugLoc(), VT,
2924 N0.getOperand(0), N0.getOperand(1),
2925 N1, N2, N0.getOperand(2));
2926 return SimplifySelect(N->getDebugLoc(), N0, N1, N2);
2932 SDValue DAGCombiner::visitSELECT_CC(SDNode *N) {
2933 SDValue N0 = N->getOperand(0);
2934 SDValue N1 = N->getOperand(1);
2935 SDValue N2 = N->getOperand(2);
2936 SDValue N3 = N->getOperand(3);
2937 SDValue N4 = N->getOperand(4);
2938 ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get();
2940 // fold select_cc lhs, rhs, x, x, cc -> x
2944 // Determine if the condition we're dealing with is constant
2945 SDValue SCC = SimplifySetCC(TLI.getSetCCResultType(N0.getValueType()),
2946 N0, N1, CC, N->getDebugLoc(), false);
2947 if (SCC.getNode()) AddToWorkList(SCC.getNode());
2949 if (ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode())) {
2950 if (!SCCC->isNullValue())
2951 return N2; // cond always true -> true val
2953 return N3; // cond always false -> false val
2956 // Fold to a simpler select_cc
2957 if (SCC.getNode() && SCC.getOpcode() == ISD::SETCC)
2958 return DAG.getNode(ISD::SELECT_CC, N->getDebugLoc(), N2.getValueType(),
2959 SCC.getOperand(0), SCC.getOperand(1), N2, N3,
2962 // If we can fold this based on the true/false value, do so.
2963 if (SimplifySelectOps(N, N2, N3))
2964 return SDValue(N, 0); // Don't revisit N.
2966 // fold select_cc into other things, such as min/max/abs
2967 return SimplifySelectCC(N->getDebugLoc(), N0, N1, N2, N3, CC);
2970 SDValue DAGCombiner::visitSETCC(SDNode *N) {
2971 return SimplifySetCC(N->getValueType(0), N->getOperand(0), N->getOperand(1),
2972 cast<CondCodeSDNode>(N->getOperand(2))->get(),
2976 // ExtendUsesToFormExtLoad - Trying to extend uses of a load to enable this:
2977 // "fold ({s|z|a}ext (load x)) -> ({s|z|a}ext (truncate ({s|z|a}extload x)))"
2978 // transformation. Returns true if extension are possible and the above
2979 // mentioned transformation is profitable.
2980 static bool ExtendUsesToFormExtLoad(SDNode *N, SDValue N0,
2982 SmallVector<SDNode*, 4> &ExtendNodes,
2983 const TargetLowering &TLI) {
2984 bool HasCopyToRegUses = false;
2985 bool isTruncFree = TLI.isTruncateFree(N->getValueType(0), N0.getValueType());
2986 for (SDNode::use_iterator UI = N0.getNode()->use_begin(),
2987 UE = N0.getNode()->use_end();
2992 if (UI.getUse().getResNo() != N0.getResNo())
2994 // FIXME: Only extend SETCC N, N and SETCC N, c for now.
2995 if (ExtOpc != ISD::ANY_EXTEND && User->getOpcode() == ISD::SETCC) {
2996 ISD::CondCode CC = cast<CondCodeSDNode>(User->getOperand(2))->get();
2997 if (ExtOpc == ISD::ZERO_EXTEND && ISD::isSignedIntSetCC(CC))
2998 // Sign bits will be lost after a zext.
3001 for (unsigned i = 0; i != 2; ++i) {
3002 SDValue UseOp = User->getOperand(i);
3005 if (!isa<ConstantSDNode>(UseOp))
3010 ExtendNodes.push_back(User);
3013 // If truncates aren't free and there are users we can't
3014 // extend, it isn't worthwhile.
3017 // Remember if this value is live-out.
3018 if (User->getOpcode() == ISD::CopyToReg)
3019 HasCopyToRegUses = true;
3022 if (HasCopyToRegUses) {
3023 bool BothLiveOut = false;
3024 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
3026 SDUse &Use = UI.getUse();
3027 if (Use.getResNo() == 0 && Use.getUser()->getOpcode() == ISD::CopyToReg) {
3033 // Both unextended and extended values are live out. There had better be
3034 // good a reason for the transformation.
3035 return ExtendNodes.size();
3040 SDValue DAGCombiner::visitSIGN_EXTEND(SDNode *N) {
3041 SDValue N0 = N->getOperand(0);
3042 EVT VT = N->getValueType(0);
3044 // fold (sext c1) -> c1
3045 if (isa<ConstantSDNode>(N0))
3046 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, N0);
3048 // fold (sext (sext x)) -> (sext x)
3049 // fold (sext (aext x)) -> (sext x)
3050 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
3051 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT,
3054 if (N0.getOpcode() == ISD::TRUNCATE) {
3055 // fold (sext (truncate (load x))) -> (sext (smaller load x))
3056 // fold (sext (truncate (srl (load x), c))) -> (sext (smaller load (x+c/n)))
3057 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
3058 if (NarrowLoad.getNode()) {
3059 if (NarrowLoad.getNode() != N0.getNode())
3060 CombineTo(N0.getNode(), NarrowLoad);
3061 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3064 // See if the value being truncated is already sign extended. If so, just
3065 // eliminate the trunc/sext pair.
3066 SDValue Op = N0.getOperand(0);
3067 unsigned OpBits = Op.getValueType().getSizeInBits();
3068 unsigned MidBits = N0.getValueType().getSizeInBits();
3069 unsigned DestBits = VT.getSizeInBits();
3070 unsigned NumSignBits = DAG.ComputeNumSignBits(Op);
3072 if (OpBits == DestBits) {
3073 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
3074 // bits, it is already ready.
3075 if (NumSignBits > DestBits-MidBits)
3077 } else if (OpBits < DestBits) {
3078 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
3079 // bits, just sext from i32.
3080 if (NumSignBits > OpBits-MidBits)
3081 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, Op);
3083 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
3084 // bits, just truncate to i32.
3085 if (NumSignBits > OpBits-MidBits)
3086 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, Op);
3089 // fold (sext (truncate x)) -> (sextinreg x).
3090 if (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG,
3091 N0.getValueType())) {
3092 if (Op.getValueType().bitsLT(VT))
3093 Op = DAG.getNode(ISD::ANY_EXTEND, N0.getDebugLoc(), VT, Op);
3094 else if (Op.getValueType().bitsGT(VT))
3095 Op = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(), VT, Op);
3096 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT, Op,
3097 DAG.getValueType(N0.getValueType().getScalarType()));
3101 // fold (sext (load x)) -> (sext (truncate (sextload x)))
3102 if (ISD::isNON_EXTLoad(N0.getNode()) &&
3103 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
3104 TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()))) {
3105 bool DoXform = true;
3106 SmallVector<SDNode*, 4> SetCCs;
3107 if (!N0.hasOneUse())
3108 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::SIGN_EXTEND, SetCCs, TLI);
3110 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
3111 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
3113 LN0->getBasePtr(), LN0->getSrcValue(),
3114 LN0->getSrcValueOffset(),
3116 LN0->isVolatile(), LN0->getAlignment());
3117 CombineTo(N, ExtLoad);
3118 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
3119 N0.getValueType(), ExtLoad);
3120 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
3122 // Extend SetCC uses if necessary.
3123 for (unsigned i = 0, e = SetCCs.size(); i != e; ++i) {
3124 SDNode *SetCC = SetCCs[i];
3125 SmallVector<SDValue, 4> Ops;
3127 for (unsigned j = 0; j != 2; ++j) {
3128 SDValue SOp = SetCC->getOperand(j);
3130 Ops.push_back(ExtLoad);
3132 Ops.push_back(DAG.getNode(ISD::SIGN_EXTEND,
3133 N->getDebugLoc(), VT, SOp));
3136 Ops.push_back(SetCC->getOperand(2));
3137 CombineTo(SetCC, DAG.getNode(ISD::SETCC, N->getDebugLoc(),
3138 SetCC->getValueType(0),
3139 &Ops[0], Ops.size()));
3142 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3146 // fold (sext (sextload x)) -> (sext (truncate (sextload x)))
3147 // fold (sext ( extload x)) -> (sext (truncate (sextload x)))
3148 if ((ISD::isSEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
3149 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
3150 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
3151 EVT MemVT = LN0->getMemoryVT();
3152 if ((!LegalOperations && !LN0->isVolatile()) ||
3153 TLI.isLoadExtLegal(ISD::SEXTLOAD, MemVT)) {
3154 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
3156 LN0->getBasePtr(), LN0->getSrcValue(),
3157 LN0->getSrcValueOffset(), MemVT,
3158 LN0->isVolatile(), LN0->getAlignment());
3159 CombineTo(N, ExtLoad);
3160 CombineTo(N0.getNode(),
3161 DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
3162 N0.getValueType(), ExtLoad),
3163 ExtLoad.getValue(1));
3164 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3168 if (N0.getOpcode() == ISD::SETCC) {
3169 // sext(setcc) -> sext_in_reg(vsetcc) for vectors.
3170 if (VT.isVector() &&
3171 // We know that the # elements of the results is the same as the
3172 // # elements of the compare (and the # elements of the compare result
3173 // for that matter). Check to see that they are the same size. If so,
3174 // we know that the element size of the sext'd result matches the
3175 // element size of the compare operands.
3176 VT.getSizeInBits() == N0.getOperand(0).getValueType().getSizeInBits() &&
3178 // Only do this before legalize for now.
3180 return DAG.getVSetCC(N->getDebugLoc(), VT, N0.getOperand(0),
3182 cast<CondCodeSDNode>(N0.getOperand(2))->get());
3185 // sext(setcc x, y, cc) -> (select_cc x, y, -1, 0, cc)
3187 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT);
3189 SimplifySelectCC(N->getDebugLoc(), N0.getOperand(0), N0.getOperand(1),
3190 NegOne, DAG.getConstant(0, VT),
3191 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
3192 if (SCC.getNode()) return SCC;
3197 // fold (sext x) -> (zext x) if the sign bit is known zero.
3198 if ((!LegalOperations || TLI.isOperationLegal(ISD::ZERO_EXTEND, VT)) &&
3199 DAG.SignBitIsZero(N0))
3200 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, N0);
3205 SDValue DAGCombiner::visitZERO_EXTEND(SDNode *N) {
3206 SDValue N0 = N->getOperand(0);
3207 EVT VT = N->getValueType(0);
3209 // fold (zext c1) -> c1
3210 if (isa<ConstantSDNode>(N0))
3211 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, N0);
3212 // fold (zext (zext x)) -> (zext x)
3213 // fold (zext (aext x)) -> (zext x)
3214 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
3215 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT,
3218 // fold (zext (truncate (load x))) -> (zext (smaller load x))
3219 // fold (zext (truncate (srl (load x), c))) -> (zext (small load (x+c/n)))
3220 if (N0.getOpcode() == ISD::TRUNCATE) {
3221 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
3222 if (NarrowLoad.getNode()) {
3223 if (NarrowLoad.getNode() != N0.getNode())
3224 CombineTo(N0.getNode(), NarrowLoad);
3225 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, NarrowLoad);
3229 // fold (zext (truncate x)) -> (and x, mask)
3230 if (N0.getOpcode() == ISD::TRUNCATE &&
3231 (!LegalOperations || TLI.isOperationLegal(ISD::AND, VT)) &&
3232 (!TLI.isTruncateFree(N0.getOperand(0).getValueType(),
3233 N0.getValueType()) ||
3234 !TLI.isZExtFree(N0.getValueType(), VT))) {
3235 SDValue Op = N0.getOperand(0);
3236 if (Op.getValueType().bitsLT(VT)) {
3237 Op = DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, Op);
3238 } else if (Op.getValueType().bitsGT(VT)) {
3239 Op = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, Op);
3241 return DAG.getZeroExtendInReg(Op, N->getDebugLoc(),
3242 N0.getValueType().getScalarType());
3245 // Fold (zext (and (trunc x), cst)) -> (and x, cst),
3246 // if either of the casts is not free.
3247 if (N0.getOpcode() == ISD::AND &&
3248 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
3249 N0.getOperand(1).getOpcode() == ISD::Constant &&
3250 (!TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
3251 N0.getValueType()) ||
3252 !TLI.isZExtFree(N0.getValueType(), VT))) {
3253 SDValue X = N0.getOperand(0).getOperand(0);
3254 if (X.getValueType().bitsLT(VT)) {
3255 X = DAG.getNode(ISD::ANY_EXTEND, X.getDebugLoc(), VT, X);
3256 } else if (X.getValueType().bitsGT(VT)) {
3257 X = DAG.getNode(ISD::TRUNCATE, X.getDebugLoc(), VT, X);
3259 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
3260 Mask.zext(VT.getSizeInBits());
3261 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
3262 X, DAG.getConstant(Mask, VT));
3265 // fold (zext (load x)) -> (zext (truncate (zextload x)))
3266 if (ISD::isNON_EXTLoad(N0.getNode()) &&
3267 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
3268 TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()))) {
3269 bool DoXform = true;
3270 SmallVector<SDNode*, 4> SetCCs;
3271 if (!N0.hasOneUse())
3272 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ZERO_EXTEND, SetCCs, TLI);
3274 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
3275 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N->getDebugLoc(), VT,
3277 LN0->getBasePtr(), LN0->getSrcValue(),
3278 LN0->getSrcValueOffset(),
3280 LN0->isVolatile(), LN0->getAlignment());
3281 CombineTo(N, ExtLoad);
3282 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
3283 N0.getValueType(), ExtLoad);
3284 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
3286 // Extend SetCC uses if necessary.
3287 for (unsigned i = 0, e = SetCCs.size(); i != e; ++i) {
3288 SDNode *SetCC = SetCCs[i];
3289 SmallVector<SDValue, 4> Ops;
3291 for (unsigned j = 0; j != 2; ++j) {
3292 SDValue SOp = SetCC->getOperand(j);
3294 Ops.push_back(ExtLoad);
3296 Ops.push_back(DAG.getNode(ISD::ZERO_EXTEND,
3297 N->getDebugLoc(), VT, SOp));
3300 Ops.push_back(SetCC->getOperand(2));
3301 CombineTo(SetCC, DAG.getNode(ISD::SETCC, N->getDebugLoc(),
3302 SetCC->getValueType(0),
3303 &Ops[0], Ops.size()));
3306 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3310 // fold (zext (zextload x)) -> (zext (truncate (zextload x)))
3311 // fold (zext ( extload x)) -> (zext (truncate (zextload x)))
3312 if ((ISD::isZEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
3313 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
3314 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
3315 EVT MemVT = LN0->getMemoryVT();
3316 if ((!LegalOperations && !LN0->isVolatile()) ||
3317 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT)) {
3318 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N->getDebugLoc(), VT,
3320 LN0->getBasePtr(), LN0->getSrcValue(),
3321 LN0->getSrcValueOffset(), MemVT,
3322 LN0->isVolatile(), LN0->getAlignment());
3323 CombineTo(N, ExtLoad);
3324 CombineTo(N0.getNode(),
3325 DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(), N0.getValueType(),
3327 ExtLoad.getValue(1));
3328 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3332 // zext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
3333 if (N0.getOpcode() == ISD::SETCC) {
3335 SimplifySelectCC(N->getDebugLoc(), N0.getOperand(0), N0.getOperand(1),
3336 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
3337 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
3338 if (SCC.getNode()) return SCC;
3341 // (zext (shl (zext x), cst)) -> (shl (zext x), cst)
3342 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL) &&
3343 isa<ConstantSDNode>(N0.getOperand(1)) &&
3344 N0.getOperand(0).getOpcode() == ISD::ZERO_EXTEND &&
3346 if (N0.getOpcode() == ISD::SHL) {
3347 // If the original shl may be shifting out bits, do not perform this
3349 unsigned ShAmt = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
3350 unsigned KnownZeroBits = N0.getOperand(0).getValueType().getSizeInBits() -
3351 N0.getOperand(0).getOperand(0).getValueType().getSizeInBits();
3352 if (ShAmt > KnownZeroBits)
3355 DebugLoc dl = N->getDebugLoc();
3356 return DAG.getNode(N0.getOpcode(), dl, VT,
3357 DAG.getNode(ISD::ZERO_EXTEND, dl, VT, N0.getOperand(0)),
3358 DAG.getNode(ISD::ZERO_EXTEND, dl,
3359 N0.getOperand(1).getValueType(),
3366 SDValue DAGCombiner::visitANY_EXTEND(SDNode *N) {
3367 SDValue N0 = N->getOperand(0);
3368 EVT VT = N->getValueType(0);
3370 // fold (aext c1) -> c1
3371 if (isa<ConstantSDNode>(N0))
3372 return DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, N0);
3373 // fold (aext (aext x)) -> (aext x)
3374 // fold (aext (zext x)) -> (zext x)
3375 // fold (aext (sext x)) -> (sext x)
3376 if (N0.getOpcode() == ISD::ANY_EXTEND ||
3377 N0.getOpcode() == ISD::ZERO_EXTEND ||
3378 N0.getOpcode() == ISD::SIGN_EXTEND)
3379 return DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT, N0.getOperand(0));
3381 // fold (aext (truncate (load x))) -> (aext (smaller load x))
3382 // fold (aext (truncate (srl (load x), c))) -> (aext (small load (x+c/n)))
3383 if (N0.getOpcode() == ISD::TRUNCATE) {
3384 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
3385 if (NarrowLoad.getNode()) {
3386 if (NarrowLoad.getNode() != N0.getNode())
3387 CombineTo(N0.getNode(), NarrowLoad);
3388 return DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, NarrowLoad);
3392 // fold (aext (truncate x))
3393 if (N0.getOpcode() == ISD::TRUNCATE) {
3394 SDValue TruncOp = N0.getOperand(0);
3395 if (TruncOp.getValueType() == VT)
3396 return TruncOp; // x iff x size == zext size.
3397 if (TruncOp.getValueType().bitsGT(VT))
3398 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, TruncOp);
3399 return DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, TruncOp);
3402 // Fold (aext (and (trunc x), cst)) -> (and x, cst)
3403 // if the trunc is not free.
3404 if (N0.getOpcode() == ISD::AND &&
3405 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
3406 N0.getOperand(1).getOpcode() == ISD::Constant &&
3407 !TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
3408 N0.getValueType())) {
3409 SDValue X = N0.getOperand(0).getOperand(0);
3410 if (X.getValueType().bitsLT(VT)) {
3411 X = DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, X);
3412 } else if (X.getValueType().bitsGT(VT)) {
3413 X = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, X);
3415 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
3416 Mask.zext(VT.getSizeInBits());
3417 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
3418 X, DAG.getConstant(Mask, VT));
3421 // fold (aext (load x)) -> (aext (truncate (extload x)))
3422 if (ISD::isNON_EXTLoad(N0.getNode()) &&
3423 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
3424 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType()))) {
3425 bool DoXform = true;
3426 SmallVector<SDNode*, 4> SetCCs;
3427 if (!N0.hasOneUse())
3428 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ANY_EXTEND, SetCCs, TLI);
3430 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
3431 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, N->getDebugLoc(), VT,
3433 LN0->getBasePtr(), LN0->getSrcValue(),
3434 LN0->getSrcValueOffset(),
3436 LN0->isVolatile(), LN0->getAlignment());
3437 CombineTo(N, ExtLoad);
3438 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
3439 N0.getValueType(), ExtLoad);
3440 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
3442 // Extend SetCC uses if necessary.
3443 for (unsigned i = 0, e = SetCCs.size(); i != e; ++i) {
3444 SDNode *SetCC = SetCCs[i];
3445 SmallVector<SDValue, 4> Ops;
3447 for (unsigned j = 0; j != 2; ++j) {
3448 SDValue SOp = SetCC->getOperand(j);
3450 Ops.push_back(ExtLoad);
3452 Ops.push_back(DAG.getNode(ISD::ANY_EXTEND,
3453 N->getDebugLoc(), VT, SOp));
3456 Ops.push_back(SetCC->getOperand(2));
3457 CombineTo(SetCC, DAG.getNode(ISD::SETCC, N->getDebugLoc(),
3458 SetCC->getValueType(0),
3459 &Ops[0], Ops.size()));
3462 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3466 // fold (aext (zextload x)) -> (aext (truncate (zextload x)))
3467 // fold (aext (sextload x)) -> (aext (truncate (sextload x)))
3468 // fold (aext ( extload x)) -> (aext (truncate (extload x)))
3469 if (N0.getOpcode() == ISD::LOAD &&
3470 !ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
3472 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
3473 EVT MemVT = LN0->getMemoryVT();
3474 SDValue ExtLoad = DAG.getExtLoad(LN0->getExtensionType(), N->getDebugLoc(),
3475 VT, LN0->getChain(), LN0->getBasePtr(),
3477 LN0->getSrcValueOffset(), MemVT,
3478 LN0->isVolatile(), LN0->getAlignment());
3479 CombineTo(N, ExtLoad);
3480 CombineTo(N0.getNode(),
3481 DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
3482 N0.getValueType(), ExtLoad),
3483 ExtLoad.getValue(1));
3484 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3487 // aext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
3488 if (N0.getOpcode() == ISD::SETCC) {
3490 SimplifySelectCC(N->getDebugLoc(), N0.getOperand(0), N0.getOperand(1),
3491 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
3492 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
3500 /// GetDemandedBits - See if the specified operand can be simplified with the
3501 /// knowledge that only the bits specified by Mask are used. If so, return the
3502 /// simpler operand, otherwise return a null SDValue.
3503 SDValue DAGCombiner::GetDemandedBits(SDValue V, const APInt &Mask) {
3504 switch (V.getOpcode()) {
3508 // If the LHS or RHS don't contribute bits to the or, drop them.
3509 if (DAG.MaskedValueIsZero(V.getOperand(0), Mask))
3510 return V.getOperand(1);
3511 if (DAG.MaskedValueIsZero(V.getOperand(1), Mask))
3512 return V.getOperand(0);
3515 // Only look at single-use SRLs.
3516 if (!V.getNode()->hasOneUse())
3518 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(V.getOperand(1))) {
3519 // See if we can recursively simplify the LHS.
3520 unsigned Amt = RHSC->getZExtValue();
3522 // Watch out for shift count overflow though.
3523 if (Amt >= Mask.getBitWidth()) break;
3524 APInt NewMask = Mask << Amt;
3525 SDValue SimplifyLHS = GetDemandedBits(V.getOperand(0), NewMask);
3526 if (SimplifyLHS.getNode())
3527 return DAG.getNode(ISD::SRL, V.getDebugLoc(), V.getValueType(),
3528 SimplifyLHS, V.getOperand(1));
3534 /// ReduceLoadWidth - If the result of a wider load is shifted to right of N
3535 /// bits and then truncated to a narrower type and where N is a multiple
3536 /// of number of bits of the narrower type, transform it to a narrower load
3537 /// from address + N / num of bits of new type. If the result is to be
3538 /// extended, also fold the extension to form a extending load.
3539 SDValue DAGCombiner::ReduceLoadWidth(SDNode *N) {
3540 unsigned Opc = N->getOpcode();
3541 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
3542 SDValue N0 = N->getOperand(0);
3543 EVT VT = N->getValueType(0);
3546 // This transformation isn't valid for vector loads.
3550 // Special case: SIGN_EXTEND_INREG is basically truncating to EVT then
3552 if (Opc == ISD::SIGN_EXTEND_INREG) {
3553 ExtType = ISD::SEXTLOAD;
3554 ExtVT = cast<VTSDNode>(N->getOperand(1))->getVT();
3555 if (LegalOperations && !TLI.isLoadExtLegal(ISD::SEXTLOAD, ExtVT))
3559 unsigned EVTBits = ExtVT.getSizeInBits();
3561 if (N0.getOpcode() == ISD::SRL && N0.hasOneUse() && ExtVT.isRound()) {
3562 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
3563 ShAmt = N01->getZExtValue();
3564 // Is the shift amount a multiple of size of VT?
3565 if ((ShAmt & (EVTBits-1)) == 0) {
3566 N0 = N0.getOperand(0);
3567 // Is the load width a multiple of size of VT?
3568 if ((N0.getValueType().getSizeInBits() & (EVTBits-1)) != 0)
3574 // Do not generate loads of non-round integer types since these can
3575 // be expensive (and would be wrong if the type is not byte sized).
3576 if (isa<LoadSDNode>(N0) && N0.hasOneUse() && ExtVT.isRound() &&
3577 cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits() > EVTBits &&
3578 // Do not change the width of a volatile load.
3579 !cast<LoadSDNode>(N0)->isVolatile()) {
3580 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
3581 EVT PtrType = N0.getOperand(1).getValueType();
3583 // For big endian targets, we need to adjust the offset to the pointer to
3584 // load the correct bytes.
3585 if (TLI.isBigEndian()) {
3586 unsigned LVTStoreBits = LN0->getMemoryVT().getStoreSizeInBits();
3587 unsigned EVTStoreBits = ExtVT.getStoreSizeInBits();
3588 ShAmt = LVTStoreBits - EVTStoreBits - ShAmt;
3591 uint64_t PtrOff = ShAmt / 8;
3592 unsigned NewAlign = MinAlign(LN0->getAlignment(), PtrOff);
3593 SDValue NewPtr = DAG.getNode(ISD::ADD, LN0->getDebugLoc(),
3594 PtrType, LN0->getBasePtr(),
3595 DAG.getConstant(PtrOff, PtrType));
3596 AddToWorkList(NewPtr.getNode());
3598 SDValue Load = (ExtType == ISD::NON_EXTLOAD)
3599 ? DAG.getLoad(VT, N0.getDebugLoc(), LN0->getChain(), NewPtr,
3600 LN0->getSrcValue(), LN0->getSrcValueOffset() + PtrOff,
3601 LN0->isVolatile(), NewAlign)
3602 : DAG.getExtLoad(ExtType, N0.getDebugLoc(), VT, LN0->getChain(), NewPtr,
3603 LN0->getSrcValue(), LN0->getSrcValueOffset() + PtrOff,
3604 ExtVT, LN0->isVolatile(), NewAlign);
3606 // Replace the old load's chain with the new load's chain.
3607 WorkListRemover DeadNodes(*this);
3608 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1),
3611 // Return the new loaded value.
3618 SDValue DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
3619 SDValue N0 = N->getOperand(0);
3620 SDValue N1 = N->getOperand(1);
3621 EVT VT = N->getValueType(0);
3622 EVT EVT = cast<VTSDNode>(N1)->getVT();
3623 unsigned VTBits = VT.getScalarType().getSizeInBits();
3624 unsigned EVTBits = EVT.getSizeInBits();
3626 // fold (sext_in_reg c1) -> c1
3627 if (isa<ConstantSDNode>(N0) || N0.getOpcode() == ISD::UNDEF)
3628 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT, N0, N1);
3630 // If the input is already sign extended, just drop the extension.
3631 if (DAG.ComputeNumSignBits(N0) >= VTBits-EVTBits+1)
3634 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2
3635 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
3636 EVT.bitsLT(cast<VTSDNode>(N0.getOperand(1))->getVT())) {
3637 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT,
3638 N0.getOperand(0), N1);
3641 // fold (sext_in_reg (sext x)) -> (sext x)
3642 // fold (sext_in_reg (aext x)) -> (sext x)
3643 // if x is small enough.
3644 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) {
3645 SDValue N00 = N0.getOperand(0);
3646 if (N00.getValueType().getScalarType().getSizeInBits() < EVTBits)
3647 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, N00, N1);
3650 // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is known zero.
3651 if (DAG.MaskedValueIsZero(N0, APInt::getBitsSet(VTBits, EVTBits-1, EVTBits)))
3652 return DAG.getZeroExtendInReg(N0, N->getDebugLoc(), EVT);
3654 // fold operands of sext_in_reg based on knowledge that the top bits are not
3656 if (SimplifyDemandedBits(SDValue(N, 0)))
3657 return SDValue(N, 0);
3659 // fold (sext_in_reg (load x)) -> (smaller sextload x)
3660 // fold (sext_in_reg (srl (load x), c)) -> (smaller sextload (x+c/evtbits))
3661 SDValue NarrowLoad = ReduceLoadWidth(N);
3662 if (NarrowLoad.getNode())
3665 // fold (sext_in_reg (srl X, 24), i8) -> (sra X, 24)
3666 // fold (sext_in_reg (srl X, 23), i8) -> (sra X, 23) iff possible.
3667 // We already fold "(sext_in_reg (srl X, 25), i8) -> srl X, 25" above.
3668 if (N0.getOpcode() == ISD::SRL) {
3669 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
3670 if (ShAmt->getZExtValue()+EVTBits <= VTBits) {
3671 // We can turn this into an SRA iff the input to the SRL is already sign
3673 unsigned InSignBits = DAG.ComputeNumSignBits(N0.getOperand(0));
3674 if (VTBits-(ShAmt->getZExtValue()+EVTBits) < InSignBits)
3675 return DAG.getNode(ISD::SRA, N->getDebugLoc(), VT,
3676 N0.getOperand(0), N0.getOperand(1));
3680 // fold (sext_inreg (extload x)) -> (sextload x)
3681 if (ISD::isEXTLoad(N0.getNode()) &&
3682 ISD::isUNINDEXEDLoad(N0.getNode()) &&
3683 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
3684 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
3685 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) {
3686 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
3687 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
3689 LN0->getBasePtr(), LN0->getSrcValue(),
3690 LN0->getSrcValueOffset(), EVT,
3691 LN0->isVolatile(), LN0->getAlignment());
3692 CombineTo(N, ExtLoad);
3693 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
3694 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3696 // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use
3697 if (ISD::isZEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
3699 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
3700 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
3701 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) {
3702 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
3703 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
3705 LN0->getBasePtr(), LN0->getSrcValue(),
3706 LN0->getSrcValueOffset(), EVT,
3707 LN0->isVolatile(), LN0->getAlignment());
3708 CombineTo(N, ExtLoad);
3709 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
3710 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3715 SDValue DAGCombiner::visitTRUNCATE(SDNode *N) {
3716 SDValue N0 = N->getOperand(0);
3717 EVT VT = N->getValueType(0);
3720 if (N0.getValueType() == N->getValueType(0))
3722 // fold (truncate c1) -> c1
3723 if (isa<ConstantSDNode>(N0))
3724 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, N0);
3725 // fold (truncate (truncate x)) -> (truncate x)
3726 if (N0.getOpcode() == ISD::TRUNCATE)
3727 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, N0.getOperand(0));
3728 // fold (truncate (ext x)) -> (ext x) or (truncate x) or x
3729 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::SIGN_EXTEND||
3730 N0.getOpcode() == ISD::ANY_EXTEND) {
3731 if (N0.getOperand(0).getValueType().bitsLT(VT))
3732 // if the source is smaller than the dest, we still need an extend
3733 return DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT,
3735 else if (N0.getOperand(0).getValueType().bitsGT(VT))
3736 // if the source is larger than the dest, than we just need the truncate
3737 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, N0.getOperand(0));
3739 // if the source and dest are the same type, we can drop both the extend
3740 // and the truncate.
3741 return N0.getOperand(0);
3744 // See if we can simplify the input to this truncate through knowledge that
3745 // only the low bits are being used. For example "trunc (or (shl x, 8), y)"
3748 GetDemandedBits(N0, APInt::getLowBitsSet(N0.getValueSizeInBits(),
3749 VT.getSizeInBits()));
3750 if (Shorter.getNode())
3751 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, Shorter);
3753 // fold (truncate (load x)) -> (smaller load x)
3754 // fold (truncate (srl (load x), c)) -> (smaller load (x+c/evtbits))
3755 return ReduceLoadWidth(N);
3758 static SDNode *getBuildPairElt(SDNode *N, unsigned i) {
3759 SDValue Elt = N->getOperand(i);
3760 if (Elt.getOpcode() != ISD::MERGE_VALUES)
3761 return Elt.getNode();
3762 return Elt.getOperand(Elt.getResNo()).getNode();
3765 /// CombineConsecutiveLoads - build_pair (load, load) -> load
3766 /// if load locations are consecutive.
3767 SDValue DAGCombiner::CombineConsecutiveLoads(SDNode *N, EVT VT) {
3768 assert(N->getOpcode() == ISD::BUILD_PAIR);
3770 LoadSDNode *LD1 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 0));
3771 LoadSDNode *LD2 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 1));
3772 if (!LD1 || !LD2 || !ISD::isNON_EXTLoad(LD1) || !LD1->hasOneUse())
3774 EVT LD1VT = LD1->getValueType(0);
3776 if (ISD::isNON_EXTLoad(LD2) &&
3778 // If both are volatile this would reduce the number of volatile loads.
3779 // If one is volatile it might be ok, but play conservative and bail out.
3780 !LD1->isVolatile() &&
3781 !LD2->isVolatile() &&
3782 DAG.isConsecutiveLoad(LD2, LD1, LD1VT.getSizeInBits()/8, 1)) {
3783 unsigned Align = LD1->getAlignment();
3784 unsigned NewAlign = TLI.getTargetData()->
3785 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
3787 if (NewAlign <= Align &&
3788 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)))
3789 return DAG.getLoad(VT, N->getDebugLoc(), LD1->getChain(),
3790 LD1->getBasePtr(), LD1->getSrcValue(),
3791 LD1->getSrcValueOffset(), false, Align);
3797 SDValue DAGCombiner::visitBIT_CONVERT(SDNode *N) {
3798 SDValue N0 = N->getOperand(0);
3799 EVT VT = N->getValueType(0);
3801 // If the input is a BUILD_VECTOR with all constant elements, fold this now.
3802 // Only do this before legalize, since afterward the target may be depending
3803 // on the bitconvert.
3804 // First check to see if this is all constant.
3806 N0.getOpcode() == ISD::BUILD_VECTOR && N0.getNode()->hasOneUse() &&
3808 bool isSimple = true;
3809 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i)
3810 if (N0.getOperand(i).getOpcode() != ISD::UNDEF &&
3811 N0.getOperand(i).getOpcode() != ISD::Constant &&
3812 N0.getOperand(i).getOpcode() != ISD::ConstantFP) {
3817 EVT DestEltVT = N->getValueType(0).getVectorElementType();
3818 assert(!DestEltVT.isVector() &&
3819 "Element type of vector ValueType must not be vector!");
3821 return ConstantFoldBIT_CONVERTofBUILD_VECTOR(N0.getNode(), DestEltVT);
3824 // If the input is a constant, let getNode fold it.
3825 if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) {
3826 SDValue Res = DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, N0);
3827 if (Res.getNode() != N) {
3828 if (!LegalOperations ||
3829 TLI.isOperationLegal(Res.getNode()->getOpcode(), VT))
3832 // Folding it resulted in an illegal node, and it's too late to
3833 // do that. Clean up the old node and forego the transformation.
3834 // Ideally this won't happen very often, because instcombine
3835 // and the earlier dagcombine runs (where illegal nodes are
3836 // permitted) should have folded most of them already.
3837 DAG.DeleteNode(Res.getNode());
3841 // (conv (conv x, t1), t2) -> (conv x, t2)
3842 if (N0.getOpcode() == ISD::BIT_CONVERT)
3843 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT,
3846 // fold (conv (load x)) -> (load (conv*)x)
3847 // If the resultant load doesn't need a higher alignment than the original!
3848 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
3849 // Do not change the width of a volatile load.
3850 !cast<LoadSDNode>(N0)->isVolatile() &&
3851 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT))) {
3852 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
3853 unsigned Align = TLI.getTargetData()->
3854 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
3855 unsigned OrigAlign = LN0->getAlignment();
3857 if (Align <= OrigAlign) {
3858 SDValue Load = DAG.getLoad(VT, N->getDebugLoc(), LN0->getChain(),
3860 LN0->getSrcValue(), LN0->getSrcValueOffset(),
3861 LN0->isVolatile(), OrigAlign);
3863 CombineTo(N0.getNode(),
3864 DAG.getNode(ISD::BIT_CONVERT, N0.getDebugLoc(),
3865 N0.getValueType(), Load),
3871 // fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit)
3872 // fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit))
3873 // This often reduces constant pool loads.
3874 if ((N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FABS) &&
3875 N0.getNode()->hasOneUse() && VT.isInteger() && !VT.isVector()) {
3876 SDValue NewConv = DAG.getNode(ISD::BIT_CONVERT, N0.getDebugLoc(), VT,
3878 AddToWorkList(NewConv.getNode());
3880 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
3881 if (N0.getOpcode() == ISD::FNEG)
3882 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT,
3883 NewConv, DAG.getConstant(SignBit, VT));
3884 assert(N0.getOpcode() == ISD::FABS);
3885 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
3886 NewConv, DAG.getConstant(~SignBit, VT));
3889 // fold (bitconvert (fcopysign cst, x)) ->
3890 // (or (and (bitconvert x), sign), (and cst, (not sign)))
3891 // Note that we don't handle (copysign x, cst) because this can always be
3892 // folded to an fneg or fabs.
3893 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse() &&
3894 isa<ConstantFPSDNode>(N0.getOperand(0)) &&
3895 VT.isInteger() && !VT.isVector()) {
3896 unsigned OrigXWidth = N0.getOperand(1).getValueType().getSizeInBits();
3897 EVT IntXVT = EVT::getIntegerVT(*DAG.getContext(), OrigXWidth);
3898 if (TLI.isTypeLegal(IntXVT) || !LegalTypes) {
3899 SDValue X = DAG.getNode(ISD::BIT_CONVERT, N0.getDebugLoc(),
3900 IntXVT, N0.getOperand(1));
3901 AddToWorkList(X.getNode());
3903 // If X has a different width than the result/lhs, sext it or truncate it.
3904 unsigned VTWidth = VT.getSizeInBits();
3905 if (OrigXWidth < VTWidth) {
3906 X = DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, X);
3907 AddToWorkList(X.getNode());
3908 } else if (OrigXWidth > VTWidth) {
3909 // To get the sign bit in the right place, we have to shift it right
3910 // before truncating.
3911 X = DAG.getNode(ISD::SRL, X.getDebugLoc(),
3912 X.getValueType(), X,
3913 DAG.getConstant(OrigXWidth-VTWidth, X.getValueType()));
3914 AddToWorkList(X.getNode());
3915 X = DAG.getNode(ISD::TRUNCATE, X.getDebugLoc(), VT, X);
3916 AddToWorkList(X.getNode());
3919 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
3920 X = DAG.getNode(ISD::AND, X.getDebugLoc(), VT,
3921 X, DAG.getConstant(SignBit, VT));
3922 AddToWorkList(X.getNode());
3924 SDValue Cst = DAG.getNode(ISD::BIT_CONVERT, N0.getDebugLoc(),
3925 VT, N0.getOperand(0));
3926 Cst = DAG.getNode(ISD::AND, Cst.getDebugLoc(), VT,
3927 Cst, DAG.getConstant(~SignBit, VT));
3928 AddToWorkList(Cst.getNode());
3930 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, X, Cst);
3934 // bitconvert(build_pair(ld, ld)) -> ld iff load locations are consecutive.
3935 if (N0.getOpcode() == ISD::BUILD_PAIR) {
3936 SDValue CombineLD = CombineConsecutiveLoads(N0.getNode(), VT);
3937 if (CombineLD.getNode())
3944 SDValue DAGCombiner::visitBUILD_PAIR(SDNode *N) {
3945 EVT VT = N->getValueType(0);
3946 return CombineConsecutiveLoads(N, VT);
3949 /// ConstantFoldBIT_CONVERTofBUILD_VECTOR - We know that BV is a build_vector
3950 /// node with Constant, ConstantFP or Undef operands. DstEltVT indicates the
3951 /// destination element value type.
3952 SDValue DAGCombiner::
3953 ConstantFoldBIT_CONVERTofBUILD_VECTOR(SDNode *BV, EVT DstEltVT) {
3954 EVT SrcEltVT = BV->getValueType(0).getVectorElementType();
3956 // If this is already the right type, we're done.
3957 if (SrcEltVT == DstEltVT) return SDValue(BV, 0);
3959 unsigned SrcBitSize = SrcEltVT.getSizeInBits();
3960 unsigned DstBitSize = DstEltVT.getSizeInBits();
3962 // If this is a conversion of N elements of one type to N elements of another
3963 // type, convert each element. This handles FP<->INT cases.
3964 if (SrcBitSize == DstBitSize) {
3965 SmallVector<SDValue, 8> Ops;
3966 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
3967 SDValue Op = BV->getOperand(i);
3968 // If the vector element type is not legal, the BUILD_VECTOR operands
3969 // are promoted and implicitly truncated. Make that explicit here.
3970 if (Op.getValueType() != SrcEltVT)
3971 Op = DAG.getNode(ISD::TRUNCATE, BV->getDebugLoc(), SrcEltVT, Op);
3972 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, BV->getDebugLoc(),
3974 AddToWorkList(Ops.back().getNode());
3976 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
3977 BV->getValueType(0).getVectorNumElements());
3978 return DAG.getNode(ISD::BUILD_VECTOR, BV->getDebugLoc(), VT,
3979 &Ops[0], Ops.size());
3982 // Otherwise, we're growing or shrinking the elements. To avoid having to
3983 // handle annoying details of growing/shrinking FP values, we convert them to
3985 if (SrcEltVT.isFloatingPoint()) {
3986 // Convert the input float vector to a int vector where the elements are the
3988 assert((SrcEltVT == MVT::f32 || SrcEltVT == MVT::f64) && "Unknown FP VT!");
3989 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), SrcEltVT.getSizeInBits());
3990 BV = ConstantFoldBIT_CONVERTofBUILD_VECTOR(BV, IntVT).getNode();
3994 // Now we know the input is an integer vector. If the output is a FP type,
3995 // convert to integer first, then to FP of the right size.
3996 if (DstEltVT.isFloatingPoint()) {
3997 assert((DstEltVT == MVT::f32 || DstEltVT == MVT::f64) && "Unknown FP VT!");
3998 EVT TmpVT = EVT::getIntegerVT(*DAG.getContext(), DstEltVT.getSizeInBits());
3999 SDNode *Tmp = ConstantFoldBIT_CONVERTofBUILD_VECTOR(BV, TmpVT).getNode();
4001 // Next, convert to FP elements of the same size.
4002 return ConstantFoldBIT_CONVERTofBUILD_VECTOR(Tmp, DstEltVT);
4005 // Okay, we know the src/dst types are both integers of differing types.
4006 // Handling growing first.
4007 assert(SrcEltVT.isInteger() && DstEltVT.isInteger());
4008 if (SrcBitSize < DstBitSize) {
4009 unsigned NumInputsPerOutput = DstBitSize/SrcBitSize;
4011 SmallVector<SDValue, 8> Ops;
4012 for (unsigned i = 0, e = BV->getNumOperands(); i != e;
4013 i += NumInputsPerOutput) {
4014 bool isLE = TLI.isLittleEndian();
4015 APInt NewBits = APInt(DstBitSize, 0);
4016 bool EltIsUndef = true;
4017 for (unsigned j = 0; j != NumInputsPerOutput; ++j) {
4018 // Shift the previously computed bits over.
4019 NewBits <<= SrcBitSize;
4020 SDValue Op = BV->getOperand(i+ (isLE ? (NumInputsPerOutput-j-1) : j));
4021 if (Op.getOpcode() == ISD::UNDEF) continue;
4024 NewBits |= (APInt(cast<ConstantSDNode>(Op)->getAPIntValue()).
4025 zextOrTrunc(SrcBitSize).zext(DstBitSize));
4029 Ops.push_back(DAG.getUNDEF(DstEltVT));
4031 Ops.push_back(DAG.getConstant(NewBits, DstEltVT));
4034 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, Ops.size());
4035 return DAG.getNode(ISD::BUILD_VECTOR, BV->getDebugLoc(), VT,
4036 &Ops[0], Ops.size());
4039 // Finally, this must be the case where we are shrinking elements: each input
4040 // turns into multiple outputs.
4041 bool isS2V = ISD::isScalarToVector(BV);
4042 unsigned NumOutputsPerInput = SrcBitSize/DstBitSize;
4043 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
4044 NumOutputsPerInput*BV->getNumOperands());
4045 SmallVector<SDValue, 8> Ops;
4047 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
4048 if (BV->getOperand(i).getOpcode() == ISD::UNDEF) {
4049 for (unsigned j = 0; j != NumOutputsPerInput; ++j)
4050 Ops.push_back(DAG.getUNDEF(DstEltVT));
4054 APInt OpVal = APInt(cast<ConstantSDNode>(BV->getOperand(i))->
4055 getAPIntValue()).zextOrTrunc(SrcBitSize);
4057 for (unsigned j = 0; j != NumOutputsPerInput; ++j) {
4058 APInt ThisVal = APInt(OpVal).trunc(DstBitSize);
4059 Ops.push_back(DAG.getConstant(ThisVal, DstEltVT));
4060 if (isS2V && i == 0 && j == 0 && APInt(ThisVal).zext(SrcBitSize) == OpVal)
4061 // Simply turn this into a SCALAR_TO_VECTOR of the new type.
4062 return DAG.getNode(ISD::SCALAR_TO_VECTOR, BV->getDebugLoc(), VT,
4064 OpVal = OpVal.lshr(DstBitSize);
4067 // For big endian targets, swap the order of the pieces of each element.
4068 if (TLI.isBigEndian())
4069 std::reverse(Ops.end()-NumOutputsPerInput, Ops.end());
4072 return DAG.getNode(ISD::BUILD_VECTOR, BV->getDebugLoc(), VT,
4073 &Ops[0], Ops.size());
4076 SDValue DAGCombiner::visitFADD(SDNode *N) {
4077 SDValue N0 = N->getOperand(0);
4078 SDValue N1 = N->getOperand(1);
4079 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
4080 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
4081 EVT VT = N->getValueType(0);
4084 if (VT.isVector()) {
4085 SDValue FoldedVOp = SimplifyVBinOp(N);
4086 if (FoldedVOp.getNode()) return FoldedVOp;
4089 // fold (fadd c1, c2) -> (fadd c1, c2)
4090 if (N0CFP && N1CFP && VT != MVT::ppcf128)
4091 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0, N1);
4092 // canonicalize constant to RHS
4093 if (N0CFP && !N1CFP)
4094 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N1, N0);
4095 // fold (fadd A, 0) -> A
4096 if (UnsafeFPMath && N1CFP && N1CFP->getValueAPF().isZero())
4098 // fold (fadd A, (fneg B)) -> (fsub A, B)
4099 if (isNegatibleForFree(N1, LegalOperations) == 2)
4100 return DAG.getNode(ISD::FSUB, N->getDebugLoc(), VT, N0,
4101 GetNegatedExpression(N1, DAG, LegalOperations));
4102 // fold (fadd (fneg A), B) -> (fsub B, A)
4103 if (isNegatibleForFree(N0, LegalOperations) == 2)
4104 return DAG.getNode(ISD::FSUB, N->getDebugLoc(), VT, N1,
4105 GetNegatedExpression(N0, DAG, LegalOperations));
4107 // If allowed, fold (fadd (fadd x, c1), c2) -> (fadd x, (fadd c1, c2))
4108 if (UnsafeFPMath && N1CFP && N0.getOpcode() == ISD::FADD &&
4109 N0.getNode()->hasOneUse() && isa<ConstantFPSDNode>(N0.getOperand(1)))
4110 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0.getOperand(0),
4111 DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
4112 N0.getOperand(1), N1));
4117 SDValue DAGCombiner::visitFSUB(SDNode *N) {
4118 SDValue N0 = N->getOperand(0);
4119 SDValue N1 = N->getOperand(1);
4120 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
4121 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
4122 EVT VT = N->getValueType(0);
4125 if (VT.isVector()) {
4126 SDValue FoldedVOp = SimplifyVBinOp(N);
4127 if (FoldedVOp.getNode()) return FoldedVOp;
4130 // fold (fsub c1, c2) -> c1-c2
4131 if (N0CFP && N1CFP && VT != MVT::ppcf128)
4132 return DAG.getNode(ISD::FSUB, N->getDebugLoc(), VT, N0, N1);
4133 // fold (fsub A, 0) -> A
4134 if (UnsafeFPMath && N1CFP && N1CFP->getValueAPF().isZero())
4136 // fold (fsub 0, B) -> -B
4137 if (UnsafeFPMath && N0CFP && N0CFP->getValueAPF().isZero()) {
4138 if (isNegatibleForFree(N1, LegalOperations))
4139 return GetNegatedExpression(N1, DAG, LegalOperations);
4140 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
4141 return DAG.getNode(ISD::FNEG, N->getDebugLoc(), VT, N1);
4143 // fold (fsub A, (fneg B)) -> (fadd A, B)
4144 if (isNegatibleForFree(N1, LegalOperations))
4145 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0,
4146 GetNegatedExpression(N1, DAG, LegalOperations));
4151 SDValue DAGCombiner::visitFMUL(SDNode *N) {
4152 SDValue N0 = N->getOperand(0);
4153 SDValue N1 = N->getOperand(1);
4154 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
4155 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
4156 EVT VT = N->getValueType(0);
4159 if (VT.isVector()) {
4160 SDValue FoldedVOp = SimplifyVBinOp(N);
4161 if (FoldedVOp.getNode()) return FoldedVOp;
4164 // fold (fmul c1, c2) -> c1*c2
4165 if (N0CFP && N1CFP && VT != MVT::ppcf128)
4166 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT, N0, N1);
4167 // canonicalize constant to RHS
4168 if (N0CFP && !N1CFP)
4169 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT, N1, N0);
4170 // fold (fmul A, 0) -> 0
4171 if (UnsafeFPMath && N1CFP && N1CFP->getValueAPF().isZero())
4173 // fold (fmul A, 0) -> 0, vector edition.
4174 if (UnsafeFPMath && ISD::isBuildVectorAllZeros(N1.getNode()))
4176 // fold (fmul X, 2.0) -> (fadd X, X)
4177 if (N1CFP && N1CFP->isExactlyValue(+2.0))
4178 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0, N0);
4179 // fold (fmul X, -1.0) -> (fneg X)
4180 if (N1CFP && N1CFP->isExactlyValue(-1.0))
4181 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
4182 return DAG.getNode(ISD::FNEG, N->getDebugLoc(), VT, N0);
4184 // fold (fmul (fneg X), (fneg Y)) -> (fmul X, Y)
4185 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations)) {
4186 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations)) {
4187 // Both can be negated for free, check to see if at least one is cheaper
4189 if (LHSNeg == 2 || RHSNeg == 2)
4190 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
4191 GetNegatedExpression(N0, DAG, LegalOperations),
4192 GetNegatedExpression(N1, DAG, LegalOperations));
4196 // If allowed, fold (fmul (fmul x, c1), c2) -> (fmul x, (fmul c1, c2))
4197 if (UnsafeFPMath && N1CFP && N0.getOpcode() == ISD::FMUL &&
4198 N0.getNode()->hasOneUse() && isa<ConstantFPSDNode>(N0.getOperand(1)))
4199 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT, N0.getOperand(0),
4200 DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
4201 N0.getOperand(1), N1));
4206 SDValue DAGCombiner::visitFDIV(SDNode *N) {
4207 SDValue N0 = N->getOperand(0);
4208 SDValue N1 = N->getOperand(1);
4209 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
4210 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
4211 EVT VT = N->getValueType(0);
4214 if (VT.isVector()) {
4215 SDValue FoldedVOp = SimplifyVBinOp(N);
4216 if (FoldedVOp.getNode()) return FoldedVOp;
4219 // fold (fdiv c1, c2) -> c1/c2
4220 if (N0CFP && N1CFP && VT != MVT::ppcf128)
4221 return DAG.getNode(ISD::FDIV, N->getDebugLoc(), VT, N0, N1);
4224 // (fdiv (fneg X), (fneg Y)) -> (fdiv X, Y)
4225 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations)) {
4226 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations)) {
4227 // Both can be negated for free, check to see if at least one is cheaper
4229 if (LHSNeg == 2 || RHSNeg == 2)
4230 return DAG.getNode(ISD::FDIV, N->getDebugLoc(), VT,
4231 GetNegatedExpression(N0, DAG, LegalOperations),
4232 GetNegatedExpression(N1, DAG, LegalOperations));
4239 SDValue DAGCombiner::visitFREM(SDNode *N) {
4240 SDValue N0 = N->getOperand(0);
4241 SDValue N1 = N->getOperand(1);
4242 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
4243 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
4244 EVT VT = N->getValueType(0);
4246 // fold (frem c1, c2) -> fmod(c1,c2)
4247 if (N0CFP && N1CFP && VT != MVT::ppcf128)
4248 return DAG.getNode(ISD::FREM, N->getDebugLoc(), VT, N0, N1);
4253 SDValue DAGCombiner::visitFCOPYSIGN(SDNode *N) {
4254 SDValue N0 = N->getOperand(0);
4255 SDValue N1 = N->getOperand(1);
4256 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
4257 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
4258 EVT VT = N->getValueType(0);
4260 if (N0CFP && N1CFP && VT != MVT::ppcf128) // Constant fold
4261 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT, N0, N1);
4264 const APFloat& V = N1CFP->getValueAPF();
4265 // copysign(x, c1) -> fabs(x) iff ispos(c1)
4266 // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1)
4267 if (!V.isNegative()) {
4268 if (!LegalOperations || TLI.isOperationLegal(ISD::FABS, VT))
4269 return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0);
4271 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
4272 return DAG.getNode(ISD::FNEG, N->getDebugLoc(), VT,
4273 DAG.getNode(ISD::FABS, N0.getDebugLoc(), VT, N0));
4277 // copysign(fabs(x), y) -> copysign(x, y)
4278 // copysign(fneg(x), y) -> copysign(x, y)
4279 // copysign(copysign(x,z), y) -> copysign(x, y)
4280 if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG ||
4281 N0.getOpcode() == ISD::FCOPYSIGN)
4282 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
4283 N0.getOperand(0), N1);
4285 // copysign(x, abs(y)) -> abs(x)
4286 if (N1.getOpcode() == ISD::FABS)
4287 return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0);
4289 // copysign(x, copysign(y,z)) -> copysign(x, z)
4290 if (N1.getOpcode() == ISD::FCOPYSIGN)
4291 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
4292 N0, N1.getOperand(1));
4294 // copysign(x, fp_extend(y)) -> copysign(x, y)
4295 // copysign(x, fp_round(y)) -> copysign(x, y)
4296 if (N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND)
4297 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
4298 N0, N1.getOperand(0));
4303 SDValue DAGCombiner::visitSINT_TO_FP(SDNode *N) {
4304 SDValue N0 = N->getOperand(0);
4305 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4306 EVT VT = N->getValueType(0);
4307 EVT OpVT = N0.getValueType();
4309 // fold (sint_to_fp c1) -> c1fp
4310 if (N0C && OpVT != MVT::ppcf128)
4311 return DAG.getNode(ISD::SINT_TO_FP, N->getDebugLoc(), VT, N0);
4313 // If the input is a legal type, and SINT_TO_FP is not legal on this target,
4314 // but UINT_TO_FP is legal on this target, try to convert.
4315 if (!TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT) &&
4316 TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT)) {
4317 // If the sign bit is known to be zero, we can change this to UINT_TO_FP.
4318 if (DAG.SignBitIsZero(N0))
4319 return DAG.getNode(ISD::UINT_TO_FP, N->getDebugLoc(), VT, N0);
4325 SDValue DAGCombiner::visitUINT_TO_FP(SDNode *N) {
4326 SDValue N0 = N->getOperand(0);
4327 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4328 EVT VT = N->getValueType(0);
4329 EVT OpVT = N0.getValueType();
4331 // fold (uint_to_fp c1) -> c1fp
4332 if (N0C && OpVT != MVT::ppcf128)
4333 return DAG.getNode(ISD::UINT_TO_FP, N->getDebugLoc(), VT, N0);
4335 // If the input is a legal type, and UINT_TO_FP is not legal on this target,
4336 // but SINT_TO_FP is legal on this target, try to convert.
4337 if (!TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT) &&
4338 TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT)) {
4339 // If the sign bit is known to be zero, we can change this to SINT_TO_FP.
4340 if (DAG.SignBitIsZero(N0))
4341 return DAG.getNode(ISD::SINT_TO_FP, N->getDebugLoc(), VT, N0);
4347 SDValue DAGCombiner::visitFP_TO_SINT(SDNode *N) {
4348 SDValue N0 = N->getOperand(0);
4349 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
4350 EVT VT = N->getValueType(0);
4352 // fold (fp_to_sint c1fp) -> c1
4354 return DAG.getNode(ISD::FP_TO_SINT, N->getDebugLoc(), VT, N0);
4359 SDValue DAGCombiner::visitFP_TO_UINT(SDNode *N) {
4360 SDValue N0 = N->getOperand(0);
4361 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
4362 EVT VT = N->getValueType(0);
4364 // fold (fp_to_uint c1fp) -> c1
4365 if (N0CFP && VT != MVT::ppcf128)
4366 return DAG.getNode(ISD::FP_TO_UINT, N->getDebugLoc(), VT, N0);
4371 SDValue DAGCombiner::visitFP_ROUND(SDNode *N) {
4372 SDValue N0 = N->getOperand(0);
4373 SDValue N1 = N->getOperand(1);
4374 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
4375 EVT VT = N->getValueType(0);
4377 // fold (fp_round c1fp) -> c1fp
4378 if (N0CFP && N0.getValueType() != MVT::ppcf128)
4379 return DAG.getNode(ISD::FP_ROUND, N->getDebugLoc(), VT, N0, N1);
4381 // fold (fp_round (fp_extend x)) -> x
4382 if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType())
4383 return N0.getOperand(0);
4385 // fold (fp_round (fp_round x)) -> (fp_round x)
4386 if (N0.getOpcode() == ISD::FP_ROUND) {
4387 // This is a value preserving truncation if both round's are.
4388 bool IsTrunc = N->getConstantOperandVal(1) == 1 &&
4389 N0.getNode()->getConstantOperandVal(1) == 1;
4390 return DAG.getNode(ISD::FP_ROUND, N->getDebugLoc(), VT, N0.getOperand(0),
4391 DAG.getIntPtrConstant(IsTrunc));
4394 // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y)
4395 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse()) {
4396 SDValue Tmp = DAG.getNode(ISD::FP_ROUND, N0.getDebugLoc(), VT,
4397 N0.getOperand(0), N1);
4398 AddToWorkList(Tmp.getNode());
4399 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
4400 Tmp, N0.getOperand(1));
4406 SDValue DAGCombiner::visitFP_ROUND_INREG(SDNode *N) {
4407 SDValue N0 = N->getOperand(0);
4408 EVT VT = N->getValueType(0);
4409 EVT EVT = cast<VTSDNode>(N->getOperand(1))->getVT();
4410 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
4412 // fold (fp_round_inreg c1fp) -> c1fp
4413 if (N0CFP && (TLI.isTypeLegal(EVT) || !LegalTypes)) {
4414 SDValue Round = DAG.getConstantFP(*N0CFP->getConstantFPValue(), EVT);
4415 return DAG.getNode(ISD::FP_EXTEND, N->getDebugLoc(), VT, Round);
4421 SDValue DAGCombiner::visitFP_EXTEND(SDNode *N) {
4422 SDValue N0 = N->getOperand(0);
4423 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
4424 EVT VT = N->getValueType(0);
4426 // If this is fp_round(fpextend), don't fold it, allow ourselves to be folded.
4427 if (N->hasOneUse() &&
4428 N->use_begin()->getOpcode() == ISD::FP_ROUND)
4431 // fold (fp_extend c1fp) -> c1fp
4432 if (N0CFP && VT != MVT::ppcf128)
4433 return DAG.getNode(ISD::FP_EXTEND, N->getDebugLoc(), VT, N0);
4435 // Turn fp_extend(fp_round(X, 1)) -> x since the fp_round doesn't affect the
4437 if (N0.getOpcode() == ISD::FP_ROUND
4438 && N0.getNode()->getConstantOperandVal(1) == 1) {
4439 SDValue In = N0.getOperand(0);
4440 if (In.getValueType() == VT) return In;
4441 if (VT.bitsLT(In.getValueType()))
4442 return DAG.getNode(ISD::FP_ROUND, N->getDebugLoc(), VT,
4443 In, N0.getOperand(1));
4444 return DAG.getNode(ISD::FP_EXTEND, N->getDebugLoc(), VT, In);
4447 // fold (fpext (load x)) -> (fpext (fptrunc (extload x)))
4448 if (ISD::isNON_EXTLoad(N0.getNode()) && N0.hasOneUse() &&
4449 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
4450 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType()))) {
4451 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4452 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, N->getDebugLoc(), VT,
4454 LN0->getBasePtr(), LN0->getSrcValue(),
4455 LN0->getSrcValueOffset(),
4457 LN0->isVolatile(), LN0->getAlignment());
4458 CombineTo(N, ExtLoad);
4459 CombineTo(N0.getNode(),
4460 DAG.getNode(ISD::FP_ROUND, N0.getDebugLoc(),
4461 N0.getValueType(), ExtLoad, DAG.getIntPtrConstant(1)),
4462 ExtLoad.getValue(1));
4463 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4469 SDValue DAGCombiner::visitFNEG(SDNode *N) {
4470 SDValue N0 = N->getOperand(0);
4471 EVT VT = N->getValueType(0);
4473 if (isNegatibleForFree(N0, LegalOperations))
4474 return GetNegatedExpression(N0, DAG, LegalOperations);
4476 // Transform fneg(bitconvert(x)) -> bitconvert(x^sign) to avoid loading
4477 // constant pool values.
4478 if (N0.getOpcode() == ISD::BIT_CONVERT &&
4480 N0.getNode()->hasOneUse() &&
4481 N0.getOperand(0).getValueType().isInteger()) {
4482 SDValue Int = N0.getOperand(0);
4483 EVT IntVT = Int.getValueType();
4484 if (IntVT.isInteger() && !IntVT.isVector()) {
4485 Int = DAG.getNode(ISD::XOR, N0.getDebugLoc(), IntVT, Int,
4486 DAG.getConstant(APInt::getSignBit(IntVT.getSizeInBits()), IntVT));
4487 AddToWorkList(Int.getNode());
4488 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(),
4496 SDValue DAGCombiner::visitFABS(SDNode *N) {
4497 SDValue N0 = N->getOperand(0);
4498 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
4499 EVT VT = N->getValueType(0);
4501 // fold (fabs c1) -> fabs(c1)
4502 if (N0CFP && VT != MVT::ppcf128)
4503 return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0);
4504 // fold (fabs (fabs x)) -> (fabs x)
4505 if (N0.getOpcode() == ISD::FABS)
4506 return N->getOperand(0);
4507 // fold (fabs (fneg x)) -> (fabs x)
4508 // fold (fabs (fcopysign x, y)) -> (fabs x)
4509 if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN)
4510 return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0.getOperand(0));
4512 // Transform fabs(bitconvert(x)) -> bitconvert(x&~sign) to avoid loading
4513 // constant pool values.
4514 if (N0.getOpcode() == ISD::BIT_CONVERT && N0.getNode()->hasOneUse() &&
4515 N0.getOperand(0).getValueType().isInteger() &&
4516 !N0.getOperand(0).getValueType().isVector()) {
4517 SDValue Int = N0.getOperand(0);
4518 EVT IntVT = Int.getValueType();
4519 if (IntVT.isInteger() && !IntVT.isVector()) {
4520 Int = DAG.getNode(ISD::AND, N0.getDebugLoc(), IntVT, Int,
4521 DAG.getConstant(~APInt::getSignBit(IntVT.getSizeInBits()), IntVT));
4522 AddToWorkList(Int.getNode());
4523 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(),
4524 N->getValueType(0), Int);
4531 SDValue DAGCombiner::visitBRCOND(SDNode *N) {
4532 SDValue Chain = N->getOperand(0);
4533 SDValue N1 = N->getOperand(1);
4534 SDValue N2 = N->getOperand(2);
4536 // If N is a constant we could fold this into a fallthrough or unconditional
4537 // branch. However that doesn't happen very often in normal code, because
4538 // Instcombine/SimplifyCFG should have handled the available opportunities.
4539 // If we did this folding here, it would be necessary to update the
4540 // MachineBasicBlock CFG, which is awkward.
4542 // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal
4544 if (N1.getOpcode() == ISD::SETCC &&
4545 TLI.isOperationLegalOrCustom(ISD::BR_CC, MVT::Other)) {
4546 return DAG.getNode(ISD::BR_CC, N->getDebugLoc(), MVT::Other,
4547 Chain, N1.getOperand(2),
4548 N1.getOperand(0), N1.getOperand(1), N2);
4552 if (N1.getOpcode() == ISD::TRUNCATE && N1.hasOneUse()) {
4553 // Look pass truncate.
4554 Trunc = N1.getNode();
4555 N1 = N1.getOperand(0);
4558 if (N1.hasOneUse() && N1.getOpcode() == ISD::SRL) {
4559 // Match this pattern so that we can generate simpler code:
4562 // %b = and i32 %a, 2
4563 // %c = srl i32 %b, 1
4564 // brcond i32 %c ...
4569 // %b = and i32 %a, 2
4570 // %c = setcc eq %b, 0
4573 // This applies only when the AND constant value has one bit set and the
4574 // SRL constant is equal to the log2 of the AND constant. The back-end is
4575 // smart enough to convert the result into a TEST/JMP sequence.
4576 SDValue Op0 = N1.getOperand(0);
4577 SDValue Op1 = N1.getOperand(1);
4579 if (Op0.getOpcode() == ISD::AND &&
4580 Op1.getOpcode() == ISD::Constant) {
4581 SDValue AndOp1 = Op0.getOperand(1);
4583 if (AndOp1.getOpcode() == ISD::Constant) {
4584 const APInt &AndConst = cast<ConstantSDNode>(AndOp1)->getAPIntValue();
4586 if (AndConst.isPowerOf2() &&
4587 cast<ConstantSDNode>(Op1)->getAPIntValue()==AndConst.logBase2()) {
4589 DAG.getSetCC(N->getDebugLoc(),
4590 TLI.getSetCCResultType(Op0.getValueType()),
4591 Op0, DAG.getConstant(0, Op0.getValueType()),
4594 SDValue NewBRCond = DAG.getNode(ISD::BRCOND, N->getDebugLoc(),
4595 MVT::Other, Chain, SetCC, N2);
4596 // Don't add the new BRCond into the worklist or else SimplifySelectCC
4597 // will convert it back to (X & C1) >> C2.
4598 CombineTo(N, NewBRCond, false);
4599 // Truncate is dead.
4601 removeFromWorkList(Trunc);
4602 DAG.DeleteNode(Trunc);
4604 // Replace the uses of SRL with SETCC
4605 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
4606 removeFromWorkList(N1.getNode());
4607 DAG.DeleteNode(N1.getNode());
4608 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4617 // Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB.
4619 SDValue DAGCombiner::visitBR_CC(SDNode *N) {
4620 CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1));
4621 SDValue CondLHS = N->getOperand(2), CondRHS = N->getOperand(3);
4623 // If N is a constant we could fold this into a fallthrough or unconditional
4624 // branch. However that doesn't happen very often in normal code, because
4625 // Instcombine/SimplifyCFG should have handled the available opportunities.
4626 // If we did this folding here, it would be necessary to update the
4627 // MachineBasicBlock CFG, which is awkward.
4629 // Use SimplifySetCC to simplify SETCC's.
4630 SDValue Simp = SimplifySetCC(TLI.getSetCCResultType(CondLHS.getValueType()),
4631 CondLHS, CondRHS, CC->get(), N->getDebugLoc(),
4633 if (Simp.getNode()) AddToWorkList(Simp.getNode());
4635 // fold to a simpler setcc
4636 if (Simp.getNode() && Simp.getOpcode() == ISD::SETCC)
4637 return DAG.getNode(ISD::BR_CC, N->getDebugLoc(), MVT::Other,
4638 N->getOperand(0), Simp.getOperand(2),
4639 Simp.getOperand(0), Simp.getOperand(1),
4645 /// CombineToPreIndexedLoadStore - Try turning a load / store into a
4646 /// pre-indexed load / store when the base pointer is an add or subtract
4647 /// and it has other uses besides the load / store. After the
4648 /// transformation, the new indexed load / store has effectively folded
4649 /// the add / subtract in and all of its other uses are redirected to the
4650 /// new load / store.
4651 bool DAGCombiner::CombineToPreIndexedLoadStore(SDNode *N) {
4652 if (!LegalOperations)
4658 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4659 if (LD->isIndexed())
4661 VT = LD->getMemoryVT();
4662 if (!TLI.isIndexedLoadLegal(ISD::PRE_INC, VT) &&
4663 !TLI.isIndexedLoadLegal(ISD::PRE_DEC, VT))
4665 Ptr = LD->getBasePtr();
4666 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4667 if (ST->isIndexed())
4669 VT = ST->getMemoryVT();
4670 if (!TLI.isIndexedStoreLegal(ISD::PRE_INC, VT) &&
4671 !TLI.isIndexedStoreLegal(ISD::PRE_DEC, VT))
4673 Ptr = ST->getBasePtr();
4679 // If the pointer is not an add/sub, or if it doesn't have multiple uses, bail
4680 // out. There is no reason to make this a preinc/predec.
4681 if ((Ptr.getOpcode() != ISD::ADD && Ptr.getOpcode() != ISD::SUB) ||
4682 Ptr.getNode()->hasOneUse())
4685 // Ask the target to do addressing mode selection.
4688 ISD::MemIndexedMode AM = ISD::UNINDEXED;
4689 if (!TLI.getPreIndexedAddressParts(N, BasePtr, Offset, AM, DAG))
4691 // Don't create a indexed load / store with zero offset.
4692 if (isa<ConstantSDNode>(Offset) &&
4693 cast<ConstantSDNode>(Offset)->isNullValue())
4696 // Try turning it into a pre-indexed load / store except when:
4697 // 1) The new base ptr is a frame index.
4698 // 2) If N is a store and the new base ptr is either the same as or is a
4699 // predecessor of the value being stored.
4700 // 3) Another use of old base ptr is a predecessor of N. If ptr is folded
4701 // that would create a cycle.
4702 // 4) All uses are load / store ops that use it as old base ptr.
4704 // Check #1. Preinc'ing a frame index would require copying the stack pointer
4705 // (plus the implicit offset) to a register to preinc anyway.
4706 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
4711 SDValue Val = cast<StoreSDNode>(N)->getValue();
4712 if (Val == BasePtr || BasePtr.getNode()->isPredecessorOf(Val.getNode()))
4716 // Now check for #3 and #4.
4717 bool RealUse = false;
4718 for (SDNode::use_iterator I = Ptr.getNode()->use_begin(),
4719 E = Ptr.getNode()->use_end(); I != E; ++I) {
4723 if (Use->isPredecessorOf(N))
4726 if (!((Use->getOpcode() == ISD::LOAD &&
4727 cast<LoadSDNode>(Use)->getBasePtr() == Ptr) ||
4728 (Use->getOpcode() == ISD::STORE &&
4729 cast<StoreSDNode>(Use)->getBasePtr() == Ptr)))
4738 Result = DAG.getIndexedLoad(SDValue(N,0), N->getDebugLoc(),
4739 BasePtr, Offset, AM);
4741 Result = DAG.getIndexedStore(SDValue(N,0), N->getDebugLoc(),
4742 BasePtr, Offset, AM);
4745 DEBUG(dbgs() << "\nReplacing.4 ";
4747 dbgs() << "\nWith: ";
4748 Result.getNode()->dump(&DAG);
4750 WorkListRemover DeadNodes(*this);
4752 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0),
4754 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2),
4757 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1),
4761 // Finally, since the node is now dead, remove it from the graph.
4764 // Replace the uses of Ptr with uses of the updated base value.
4765 DAG.ReplaceAllUsesOfValueWith(Ptr, Result.getValue(isLoad ? 1 : 0),
4767 removeFromWorkList(Ptr.getNode());
4768 DAG.DeleteNode(Ptr.getNode());
4773 /// CombineToPostIndexedLoadStore - Try to combine a load / store with a
4774 /// add / sub of the base pointer node into a post-indexed load / store.
4775 /// The transformation folded the add / subtract into the new indexed
4776 /// load / store effectively and all of its uses are redirected to the
4777 /// new load / store.
4778 bool DAGCombiner::CombineToPostIndexedLoadStore(SDNode *N) {
4779 if (!LegalOperations)
4785 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4786 if (LD->isIndexed())
4788 VT = LD->getMemoryVT();
4789 if (!TLI.isIndexedLoadLegal(ISD::POST_INC, VT) &&
4790 !TLI.isIndexedLoadLegal(ISD::POST_DEC, VT))
4792 Ptr = LD->getBasePtr();
4793 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4794 if (ST->isIndexed())
4796 VT = ST->getMemoryVT();
4797 if (!TLI.isIndexedStoreLegal(ISD::POST_INC, VT) &&
4798 !TLI.isIndexedStoreLegal(ISD::POST_DEC, VT))
4800 Ptr = ST->getBasePtr();
4806 if (Ptr.getNode()->hasOneUse())
4809 for (SDNode::use_iterator I = Ptr.getNode()->use_begin(),
4810 E = Ptr.getNode()->use_end(); I != E; ++I) {
4813 (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB))
4818 ISD::MemIndexedMode AM = ISD::UNINDEXED;
4819 if (TLI.getPostIndexedAddressParts(N, Op, BasePtr, Offset, AM, DAG)) {
4820 if (Ptr == Offset && Op->getOpcode() == ISD::ADD)
4821 std::swap(BasePtr, Offset);
4824 // Don't create a indexed load / store with zero offset.
4825 if (isa<ConstantSDNode>(Offset) &&
4826 cast<ConstantSDNode>(Offset)->isNullValue())
4829 // Try turning it into a post-indexed load / store except when
4830 // 1) All uses are load / store ops that use it as base ptr.
4831 // 2) Op must be independent of N, i.e. Op is neither a predecessor
4832 // nor a successor of N. Otherwise, if Op is folded that would
4835 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
4839 bool TryNext = false;
4840 for (SDNode::use_iterator II = BasePtr.getNode()->use_begin(),
4841 EE = BasePtr.getNode()->use_end(); II != EE; ++II) {
4843 if (Use == Ptr.getNode())
4846 // If all the uses are load / store addresses, then don't do the
4848 if (Use->getOpcode() == ISD::ADD || Use->getOpcode() == ISD::SUB){
4849 bool RealUse = false;
4850 for (SDNode::use_iterator III = Use->use_begin(),
4851 EEE = Use->use_end(); III != EEE; ++III) {
4852 SDNode *UseUse = *III;
4853 if (!((UseUse->getOpcode() == ISD::LOAD &&
4854 cast<LoadSDNode>(UseUse)->getBasePtr().getNode() == Use) ||
4855 (UseUse->getOpcode() == ISD::STORE &&
4856 cast<StoreSDNode>(UseUse)->getBasePtr().getNode() == Use)))
4871 if (!Op->isPredecessorOf(N) && !N->isPredecessorOf(Op)) {
4872 SDValue Result = isLoad
4873 ? DAG.getIndexedLoad(SDValue(N,0), N->getDebugLoc(),
4874 BasePtr, Offset, AM)
4875 : DAG.getIndexedStore(SDValue(N,0), N->getDebugLoc(),
4876 BasePtr, Offset, AM);
4879 DEBUG(dbgs() << "\nReplacing.5 ";
4881 dbgs() << "\nWith: ";
4882 Result.getNode()->dump(&DAG);
4884 WorkListRemover DeadNodes(*this);
4886 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0),
4888 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2),
4891 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1),
4895 // Finally, since the node is now dead, remove it from the graph.
4898 // Replace the uses of Use with uses of the updated base value.
4899 DAG.ReplaceAllUsesOfValueWith(SDValue(Op, 0),
4900 Result.getValue(isLoad ? 1 : 0),
4902 removeFromWorkList(Op);
4912 SDValue DAGCombiner::visitLOAD(SDNode *N) {
4913 LoadSDNode *LD = cast<LoadSDNode>(N);
4914 SDValue Chain = LD->getChain();
4915 SDValue Ptr = LD->getBasePtr();
4917 // Try to infer better alignment information than the load already has.
4918 if (OptLevel != CodeGenOpt::None && LD->isUnindexed()) {
4919 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
4920 if (Align > LD->getAlignment())
4921 return DAG.getExtLoad(LD->getExtensionType(), N->getDebugLoc(),
4922 LD->getValueType(0),
4923 Chain, Ptr, LD->getSrcValue(),
4924 LD->getSrcValueOffset(), LD->getMemoryVT(),
4925 LD->isVolatile(), Align);
4929 // If load is not volatile and there are no uses of the loaded value (and
4930 // the updated indexed value in case of indexed loads), change uses of the
4931 // chain value into uses of the chain input (i.e. delete the dead load).
4932 if (!LD->isVolatile()) {
4933 if (N->getValueType(1) == MVT::Other) {
4935 if (N->hasNUsesOfValue(0, 0)) {
4936 // It's not safe to use the two value CombineTo variant here. e.g.
4937 // v1, chain2 = load chain1, loc
4938 // v2, chain3 = load chain2, loc
4940 // Now we replace use of chain2 with chain1. This makes the second load
4941 // isomorphic to the one we are deleting, and thus makes this load live.
4942 DEBUG(dbgs() << "\nReplacing.6 ";
4944 dbgs() << "\nWith chain: ";
4945 Chain.getNode()->dump(&DAG);
4947 WorkListRemover DeadNodes(*this);
4948 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain, &DeadNodes);
4950 if (N->use_empty()) {
4951 removeFromWorkList(N);
4955 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4959 assert(N->getValueType(2) == MVT::Other && "Malformed indexed loads?");
4960 if (N->hasNUsesOfValue(0, 0) && N->hasNUsesOfValue(0, 1)) {
4961 SDValue Undef = DAG.getUNDEF(N->getValueType(0));
4962 DEBUG(dbgs() << "\nReplacing.6 ";
4964 dbgs() << "\nWith: ";
4965 Undef.getNode()->dump(&DAG);
4966 dbgs() << " and 2 other values\n");
4967 WorkListRemover DeadNodes(*this);
4968 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Undef, &DeadNodes);
4969 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1),
4970 DAG.getUNDEF(N->getValueType(1)),
4972 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 2), Chain, &DeadNodes);
4973 removeFromWorkList(N);
4975 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4980 // If this load is directly stored, replace the load value with the stored
4982 // TODO: Handle store large -> read small portion.
4983 // TODO: Handle TRUNCSTORE/LOADEXT
4984 if (LD->getExtensionType() == ISD::NON_EXTLOAD &&
4985 !LD->isVolatile()) {
4986 if (ISD::isNON_TRUNCStore(Chain.getNode())) {
4987 StoreSDNode *PrevST = cast<StoreSDNode>(Chain);
4988 if (PrevST->getBasePtr() == Ptr &&
4989 PrevST->getValue().getValueType() == N->getValueType(0))
4990 return CombineTo(N, Chain.getOperand(1), Chain);
4995 // Walk up chain skipping non-aliasing memory nodes.
4996 SDValue BetterChain = FindBetterChain(N, Chain);
4998 // If there is a better chain.
4999 if (Chain != BetterChain) {
5002 // Replace the chain to void dependency.
5003 if (LD->getExtensionType() == ISD::NON_EXTLOAD) {
5004 ReplLoad = DAG.getLoad(N->getValueType(0), LD->getDebugLoc(),
5006 LD->getSrcValue(), LD->getSrcValueOffset(),
5007 LD->isVolatile(), LD->getAlignment());
5009 ReplLoad = DAG.getExtLoad(LD->getExtensionType(), LD->getDebugLoc(),
5010 LD->getValueType(0),
5011 BetterChain, Ptr, LD->getSrcValue(),
5012 LD->getSrcValueOffset(),
5015 LD->getAlignment());
5018 // Create token factor to keep old chain connected.
5019 SDValue Token = DAG.getNode(ISD::TokenFactor, N->getDebugLoc(),
5020 MVT::Other, Chain, ReplLoad.getValue(1));
5022 // Make sure the new and old chains are cleaned up.
5023 AddToWorkList(Token.getNode());
5025 // Replace uses with load result and token factor. Don't add users
5027 return CombineTo(N, ReplLoad.getValue(0), Token, false);
5031 // Try transforming N to an indexed load.
5032 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
5033 return SDValue(N, 0);
5039 /// ReduceLoadOpStoreWidth - Look for sequence of load / op / store where op is
5040 /// one of 'or', 'xor', and 'and' of immediates. If 'op' is only touching some
5041 /// of the loaded bits, try narrowing the load and store if it would end up
5042 /// being a win for performance or code size.
5043 SDValue DAGCombiner::ReduceLoadOpStoreWidth(SDNode *N) {
5044 StoreSDNode *ST = cast<StoreSDNode>(N);
5045 if (ST->isVolatile())
5048 SDValue Chain = ST->getChain();
5049 SDValue Value = ST->getValue();
5050 SDValue Ptr = ST->getBasePtr();
5051 EVT VT = Value.getValueType();
5053 if (ST->isTruncatingStore() || VT.isVector() || !Value.hasOneUse())
5056 unsigned Opc = Value.getOpcode();
5057 if ((Opc != ISD::OR && Opc != ISD::XOR && Opc != ISD::AND) ||
5058 Value.getOperand(1).getOpcode() != ISD::Constant)
5061 SDValue N0 = Value.getOperand(0);
5062 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse()) {
5063 LoadSDNode *LD = cast<LoadSDNode>(N0);
5064 if (LD->getBasePtr() != Ptr)
5067 // Find the type to narrow it the load / op / store to.
5068 SDValue N1 = Value.getOperand(1);
5069 unsigned BitWidth = N1.getValueSizeInBits();
5070 APInt Imm = cast<ConstantSDNode>(N1)->getAPIntValue();
5071 if (Opc == ISD::AND)
5072 Imm ^= APInt::getAllOnesValue(BitWidth);
5073 if (Imm == 0 || Imm.isAllOnesValue())
5075 unsigned ShAmt = Imm.countTrailingZeros();
5076 unsigned MSB = BitWidth - Imm.countLeadingZeros() - 1;
5077 unsigned NewBW = NextPowerOf2(MSB - ShAmt);
5078 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
5079 while (NewBW < BitWidth &&
5080 !(TLI.isOperationLegalOrCustom(Opc, NewVT) &&
5081 TLI.isNarrowingProfitable(VT, NewVT))) {
5082 NewBW = NextPowerOf2(NewBW);
5083 NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
5085 if (NewBW >= BitWidth)
5088 // If the lsb changed does not start at the type bitwidth boundary,
5089 // start at the previous one.
5091 ShAmt = (((ShAmt + NewBW - 1) / NewBW) * NewBW) - NewBW;
5092 APInt Mask = APInt::getBitsSet(BitWidth, ShAmt, ShAmt + NewBW);
5093 if ((Imm & Mask) == Imm) {
5094 APInt NewImm = (Imm & Mask).lshr(ShAmt).trunc(NewBW);
5095 if (Opc == ISD::AND)
5096 NewImm ^= APInt::getAllOnesValue(NewBW);
5097 uint64_t PtrOff = ShAmt / 8;
5098 // For big endian targets, we need to adjust the offset to the pointer to
5099 // load the correct bytes.
5100 if (TLI.isBigEndian())
5101 PtrOff = (BitWidth + 7 - NewBW) / 8 - PtrOff;
5103 unsigned NewAlign = MinAlign(LD->getAlignment(), PtrOff);
5105 TLI.getTargetData()->getABITypeAlignment(NewVT.getTypeForEVT(*DAG.getContext())))
5108 SDValue NewPtr = DAG.getNode(ISD::ADD, LD->getDebugLoc(),
5109 Ptr.getValueType(), Ptr,
5110 DAG.getConstant(PtrOff, Ptr.getValueType()));
5111 SDValue NewLD = DAG.getLoad(NewVT, N0.getDebugLoc(),
5112 LD->getChain(), NewPtr,
5113 LD->getSrcValue(), LD->getSrcValueOffset(),
5114 LD->isVolatile(), NewAlign);
5115 SDValue NewVal = DAG.getNode(Opc, Value.getDebugLoc(), NewVT, NewLD,
5116 DAG.getConstant(NewImm, NewVT));
5117 SDValue NewST = DAG.getStore(Chain, N->getDebugLoc(),
5119 ST->getSrcValue(), ST->getSrcValueOffset(),
5122 AddToWorkList(NewPtr.getNode());
5123 AddToWorkList(NewLD.getNode());
5124 AddToWorkList(NewVal.getNode());
5125 WorkListRemover DeadNodes(*this);
5126 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLD.getValue(1),
5136 SDValue DAGCombiner::visitSTORE(SDNode *N) {
5137 StoreSDNode *ST = cast<StoreSDNode>(N);
5138 SDValue Chain = ST->getChain();
5139 SDValue Value = ST->getValue();
5140 SDValue Ptr = ST->getBasePtr();
5142 // Try to infer better alignment information than the store already has.
5143 if (OptLevel != CodeGenOpt::None && ST->isUnindexed()) {
5144 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
5145 if (Align > ST->getAlignment())
5146 return DAG.getTruncStore(Chain, N->getDebugLoc(), Value,
5147 Ptr, ST->getSrcValue(),
5148 ST->getSrcValueOffset(), ST->getMemoryVT(),
5149 ST->isVolatile(), Align);
5153 // If this is a store of a bit convert, store the input value if the
5154 // resultant store does not need a higher alignment than the original.
5155 if (Value.getOpcode() == ISD::BIT_CONVERT && !ST->isTruncatingStore() &&
5156 ST->isUnindexed()) {
5157 unsigned OrigAlign = ST->getAlignment();
5158 EVT SVT = Value.getOperand(0).getValueType();
5159 unsigned Align = TLI.getTargetData()->
5160 getABITypeAlignment(SVT.getTypeForEVT(*DAG.getContext()));
5161 if (Align <= OrigAlign &&
5162 ((!LegalOperations && !ST->isVolatile()) ||
5163 TLI.isOperationLegalOrCustom(ISD::STORE, SVT)))
5164 return DAG.getStore(Chain, N->getDebugLoc(), Value.getOperand(0),
5165 Ptr, ST->getSrcValue(),
5166 ST->getSrcValueOffset(), ST->isVolatile(), OrigAlign);
5169 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
5170 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Value)) {
5171 // NOTE: If the original store is volatile, this transform must not increase
5172 // the number of stores. For example, on x86-32 an f64 can be stored in one
5173 // processor operation but an i64 (which is not legal) requires two. So the
5174 // transform should not be done in this case.
5175 if (Value.getOpcode() != ISD::TargetConstantFP) {
5177 switch (CFP->getValueType(0).getSimpleVT().SimpleTy) {
5178 default: llvm_unreachable("Unknown FP type");
5179 case MVT::f80: // We don't do this for these yet.
5184 if (((TLI.isTypeLegal(MVT::i32) || !LegalTypes) && !LegalOperations &&
5185 !ST->isVolatile()) ||
5186 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
5187 Tmp = DAG.getConstant((uint32_t)CFP->getValueAPF().
5188 bitcastToAPInt().getZExtValue(), MVT::i32);
5189 return DAG.getStore(Chain, N->getDebugLoc(), Tmp,
5190 Ptr, ST->getSrcValue(),
5191 ST->getSrcValueOffset(), ST->isVolatile(),
5192 ST->getAlignment());
5196 if (((TLI.isTypeLegal(MVT::i64) || !LegalTypes) && !LegalOperations &&
5197 !ST->isVolatile()) ||
5198 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i64)) {
5199 Tmp = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
5200 getZExtValue(), MVT::i64);
5201 return DAG.getStore(Chain, N->getDebugLoc(), Tmp,
5202 Ptr, ST->getSrcValue(),
5203 ST->getSrcValueOffset(), ST->isVolatile(),
5204 ST->getAlignment());
5205 } else if (!ST->isVolatile() &&
5206 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
5207 // Many FP stores are not made apparent until after legalize, e.g. for
5208 // argument passing. Since this is so common, custom legalize the
5209 // 64-bit integer store into two 32-bit stores.
5210 uint64_t Val = CFP->getValueAPF().bitcastToAPInt().getZExtValue();
5211 SDValue Lo = DAG.getConstant(Val & 0xFFFFFFFF, MVT::i32);
5212 SDValue Hi = DAG.getConstant(Val >> 32, MVT::i32);
5213 if (TLI.isBigEndian()) std::swap(Lo, Hi);
5215 int SVOffset = ST->getSrcValueOffset();
5216 unsigned Alignment = ST->getAlignment();
5217 bool isVolatile = ST->isVolatile();
5219 SDValue St0 = DAG.getStore(Chain, ST->getDebugLoc(), Lo,
5220 Ptr, ST->getSrcValue(),
5221 ST->getSrcValueOffset(),
5222 isVolatile, ST->getAlignment());
5223 Ptr = DAG.getNode(ISD::ADD, N->getDebugLoc(), Ptr.getValueType(), Ptr,
5224 DAG.getConstant(4, Ptr.getValueType()));
5226 Alignment = MinAlign(Alignment, 4U);
5227 SDValue St1 = DAG.getStore(Chain, ST->getDebugLoc(), Hi,
5228 Ptr, ST->getSrcValue(),
5229 SVOffset, isVolatile, Alignment);
5230 return DAG.getNode(ISD::TokenFactor, N->getDebugLoc(), MVT::Other,
5240 // Walk up chain skipping non-aliasing memory nodes.
5241 SDValue BetterChain = FindBetterChain(N, Chain);
5243 // If there is a better chain.
5244 if (Chain != BetterChain) {
5247 // Replace the chain to avoid dependency.
5248 if (ST->isTruncatingStore()) {
5249 ReplStore = DAG.getTruncStore(BetterChain, N->getDebugLoc(), Value, Ptr,
5250 ST->getSrcValue(),ST->getSrcValueOffset(),
5252 ST->isVolatile(), ST->getAlignment());
5254 ReplStore = DAG.getStore(BetterChain, N->getDebugLoc(), Value, Ptr,
5255 ST->getSrcValue(), ST->getSrcValueOffset(),
5256 ST->isVolatile(), ST->getAlignment());
5259 // Create token to keep both nodes around.
5260 SDValue Token = DAG.getNode(ISD::TokenFactor, N->getDebugLoc(),
5261 MVT::Other, Chain, ReplStore);
5263 // Make sure the new and old chains are cleaned up.
5264 AddToWorkList(Token.getNode());
5266 // Don't add users to work list.
5267 return CombineTo(N, Token, false);
5271 // Try transforming N to an indexed store.
5272 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
5273 return SDValue(N, 0);
5275 // FIXME: is there such a thing as a truncating indexed store?
5276 if (ST->isTruncatingStore() && ST->isUnindexed() &&
5277 Value.getValueType().isInteger()) {
5278 // See if we can simplify the input to this truncstore with knowledge that
5279 // only the low bits are being used. For example:
5280 // "truncstore (or (shl x, 8), y), i8" -> "truncstore y, i8"
5282 GetDemandedBits(Value,
5283 APInt::getLowBitsSet(Value.getValueSizeInBits(),
5284 ST->getMemoryVT().getSizeInBits()));
5285 AddToWorkList(Value.getNode());
5286 if (Shorter.getNode())
5287 return DAG.getTruncStore(Chain, N->getDebugLoc(), Shorter,
5288 Ptr, ST->getSrcValue(),
5289 ST->getSrcValueOffset(), ST->getMemoryVT(),
5290 ST->isVolatile(), ST->getAlignment());
5292 // Otherwise, see if we can simplify the operation with
5293 // SimplifyDemandedBits, which only works if the value has a single use.
5294 if (SimplifyDemandedBits(Value,
5295 APInt::getLowBitsSet(
5296 Value.getValueType().getScalarType().getSizeInBits(),
5297 ST->getMemoryVT().getSizeInBits())))
5298 return SDValue(N, 0);
5301 // If this is a load followed by a store to the same location, then the store
5303 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Value)) {
5304 if (Ld->getBasePtr() == Ptr && ST->getMemoryVT() == Ld->getMemoryVT() &&
5305 ST->isUnindexed() && !ST->isVolatile() &&
5306 // There can't be any side effects between the load and store, such as
5308 Chain.reachesChainWithoutSideEffects(SDValue(Ld, 1))) {
5309 // The store is dead, remove it.
5314 // If this is an FP_ROUND or TRUNC followed by a store, fold this into a
5315 // truncating store. We can do this even if this is already a truncstore.
5316 if ((Value.getOpcode() == ISD::FP_ROUND || Value.getOpcode() == ISD::TRUNCATE)
5317 && Value.getNode()->hasOneUse() && ST->isUnindexed() &&
5318 TLI.isTruncStoreLegal(Value.getOperand(0).getValueType(),
5319 ST->getMemoryVT())) {
5320 return DAG.getTruncStore(Chain, N->getDebugLoc(), Value.getOperand(0),
5321 Ptr, ST->getSrcValue(),
5322 ST->getSrcValueOffset(), ST->getMemoryVT(),
5323 ST->isVolatile(), ST->getAlignment());
5326 return ReduceLoadOpStoreWidth(N);
5329 SDValue DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) {
5330 SDValue InVec = N->getOperand(0);
5331 SDValue InVal = N->getOperand(1);
5332 SDValue EltNo = N->getOperand(2);
5334 // If the invec is a BUILD_VECTOR and if EltNo is a constant, build a new
5335 // vector with the inserted element.
5336 if (InVec.getOpcode() == ISD::BUILD_VECTOR && isa<ConstantSDNode>(EltNo)) {
5337 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
5338 SmallVector<SDValue, 8> Ops(InVec.getNode()->op_begin(),
5339 InVec.getNode()->op_end());
5340 if (Elt < Ops.size())
5342 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
5343 InVec.getValueType(), &Ops[0], Ops.size());
5345 // If the invec is an UNDEF and if EltNo is a constant, create a new
5346 // BUILD_VECTOR with undef elements and the inserted element.
5347 if (!LegalOperations && InVec.getOpcode() == ISD::UNDEF &&
5348 isa<ConstantSDNode>(EltNo)) {
5349 EVT VT = InVec.getValueType();
5350 EVT EltVT = VT.getVectorElementType();
5351 unsigned NElts = VT.getVectorNumElements();
5352 SmallVector<SDValue, 8> Ops(NElts, DAG.getUNDEF(EltVT));
5354 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
5355 if (Elt < Ops.size())
5357 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
5358 InVec.getValueType(), &Ops[0], Ops.size());
5363 SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) {
5364 // (vextract (scalar_to_vector val, 0) -> val
5365 SDValue InVec = N->getOperand(0);
5367 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5368 // If the operand is wider than the vector element type then it is implicitly
5369 // truncated. Make that explicit here.
5370 EVT EltVT = InVec.getValueType().getVectorElementType();
5371 SDValue InOp = InVec.getOperand(0);
5372 if (InOp.getValueType() != EltVT)
5373 return DAG.getNode(ISD::TRUNCATE, InVec.getDebugLoc(), EltVT, InOp);
5377 // Perform only after legalization to ensure build_vector / vector_shuffle
5378 // optimizations have already been done.
5379 if (!LegalOperations) return SDValue();
5381 // (vextract (v4f32 load $addr), c) -> (f32 load $addr+c*size)
5382 // (vextract (v4f32 s2v (f32 load $addr)), c) -> (f32 load $addr+c*size)
5383 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), 0) -> (f32 load $addr)
5384 SDValue EltNo = N->getOperand(1);
5386 if (isa<ConstantSDNode>(EltNo)) {
5387 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
5388 bool NewLoad = false;
5389 bool BCNumEltsChanged = false;
5390 EVT VT = InVec.getValueType();
5391 EVT ExtVT = VT.getVectorElementType();
5394 if (InVec.getOpcode() == ISD::BIT_CONVERT) {
5395 EVT BCVT = InVec.getOperand(0).getValueType();
5396 if (!BCVT.isVector() || ExtVT.bitsGT(BCVT.getVectorElementType()))
5398 if (VT.getVectorNumElements() != BCVT.getVectorNumElements())
5399 BCNumEltsChanged = true;
5400 InVec = InVec.getOperand(0);
5401 ExtVT = BCVT.getVectorElementType();
5405 LoadSDNode *LN0 = NULL;
5406 const ShuffleVectorSDNode *SVN = NULL;
5407 if (ISD::isNormalLoad(InVec.getNode())) {
5408 LN0 = cast<LoadSDNode>(InVec);
5409 } else if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR &&
5410 InVec.getOperand(0).getValueType() == ExtVT &&
5411 ISD::isNormalLoad(InVec.getOperand(0).getNode())) {
5412 LN0 = cast<LoadSDNode>(InVec.getOperand(0));
5413 } else if ((SVN = dyn_cast<ShuffleVectorSDNode>(InVec))) {
5414 // (vextract (vector_shuffle (load $addr), v2, <1, u, u, u>), 1)
5416 // (load $addr+1*size)
5418 // If the bit convert changed the number of elements, it is unsafe
5419 // to examine the mask.
5420 if (BCNumEltsChanged)
5423 // Select the input vector, guarding against out of range extract vector.
5424 unsigned NumElems = VT.getVectorNumElements();
5425 int Idx = (Elt > NumElems) ? -1 : SVN->getMaskElt(Elt);
5426 InVec = (Idx < (int)NumElems) ? InVec.getOperand(0) : InVec.getOperand(1);
5428 if (InVec.getOpcode() == ISD::BIT_CONVERT)
5429 InVec = InVec.getOperand(0);
5430 if (ISD::isNormalLoad(InVec.getNode())) {
5431 LN0 = cast<LoadSDNode>(InVec);
5432 Elt = (Idx < (int)NumElems) ? Idx : Idx - NumElems;
5436 if (!LN0 || !LN0->hasOneUse() || LN0->isVolatile())
5439 unsigned Align = LN0->getAlignment();
5441 // Check the resultant load doesn't need a higher alignment than the
5444 TLI.getTargetData()->getABITypeAlignment(LVT.getTypeForEVT(*DAG.getContext()));
5446 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, LVT))
5452 SDValue NewPtr = LN0->getBasePtr();
5454 unsigned PtrOff = LVT.getSizeInBits() * Elt / 8;
5455 EVT PtrType = NewPtr.getValueType();
5456 if (TLI.isBigEndian())
5457 PtrOff = VT.getSizeInBits() / 8 - PtrOff;
5458 NewPtr = DAG.getNode(ISD::ADD, N->getDebugLoc(), PtrType, NewPtr,
5459 DAG.getConstant(PtrOff, PtrType));
5462 return DAG.getLoad(LVT, N->getDebugLoc(), LN0->getChain(), NewPtr,
5463 LN0->getSrcValue(), LN0->getSrcValueOffset(),
5464 LN0->isVolatile(), Align);
5470 SDValue DAGCombiner::visitBUILD_VECTOR(SDNode *N) {
5471 unsigned NumInScalars = N->getNumOperands();
5472 EVT VT = N->getValueType(0);
5474 // Check to see if this is a BUILD_VECTOR of a bunch of EXTRACT_VECTOR_ELT
5475 // operations. If so, and if the EXTRACT_VECTOR_ELT vector inputs come from
5476 // at most two distinct vectors, turn this into a shuffle node.
5477 SDValue VecIn1, VecIn2;
5478 for (unsigned i = 0; i != NumInScalars; ++i) {
5479 // Ignore undef inputs.
5480 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
5482 // If this input is something other than a EXTRACT_VECTOR_ELT with a
5483 // constant index, bail out.
5484 if (N->getOperand(i).getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5485 !isa<ConstantSDNode>(N->getOperand(i).getOperand(1))) {
5486 VecIn1 = VecIn2 = SDValue(0, 0);
5490 // If the input vector type disagrees with the result of the build_vector,
5491 // we can't make a shuffle.
5492 SDValue ExtractedFromVec = N->getOperand(i).getOperand(0);
5493 if (ExtractedFromVec.getValueType() != VT) {
5494 VecIn1 = VecIn2 = SDValue(0, 0);
5498 // Otherwise, remember this. We allow up to two distinct input vectors.
5499 if (ExtractedFromVec == VecIn1 || ExtractedFromVec == VecIn2)
5502 if (VecIn1.getNode() == 0) {
5503 VecIn1 = ExtractedFromVec;
5504 } else if (VecIn2.getNode() == 0) {
5505 VecIn2 = ExtractedFromVec;
5508 VecIn1 = VecIn2 = SDValue(0, 0);
5513 // If everything is good, we can make a shuffle operation.
5514 if (VecIn1.getNode()) {
5515 SmallVector<int, 8> Mask;
5516 for (unsigned i = 0; i != NumInScalars; ++i) {
5517 if (N->getOperand(i).getOpcode() == ISD::UNDEF) {
5522 // If extracting from the first vector, just use the index directly.
5523 SDValue Extract = N->getOperand(i);
5524 SDValue ExtVal = Extract.getOperand(1);
5525 if (Extract.getOperand(0) == VecIn1) {
5526 unsigned ExtIndex = cast<ConstantSDNode>(ExtVal)->getZExtValue();
5527 if (ExtIndex > VT.getVectorNumElements())
5530 Mask.push_back(ExtIndex);
5534 // Otherwise, use InIdx + VecSize
5535 unsigned Idx = cast<ConstantSDNode>(ExtVal)->getZExtValue();
5536 Mask.push_back(Idx+NumInScalars);
5539 // Add count and size info.
5540 if (!TLI.isTypeLegal(VT) && LegalTypes)
5543 // Return the new VECTOR_SHUFFLE node.
5546 Ops[1] = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
5547 return DAG.getVectorShuffle(VT, N->getDebugLoc(), Ops[0], Ops[1], &Mask[0]);
5553 SDValue DAGCombiner::visitCONCAT_VECTORS(SDNode *N) {
5554 // TODO: Check to see if this is a CONCAT_VECTORS of a bunch of
5555 // EXTRACT_SUBVECTOR operations. If so, and if the EXTRACT_SUBVECTOR vector
5556 // inputs come from at most two distinct vectors, turn this into a shuffle
5559 // If we only have one input vector, we don't need to do any concatenation.
5560 if (N->getNumOperands() == 1)
5561 return N->getOperand(0);
5566 SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
5569 EVT VT = N->getValueType(0);
5570 unsigned NumElts = VT.getVectorNumElements();
5572 SDValue N0 = N->getOperand(0);
5574 assert(N0.getValueType().getVectorNumElements() == NumElts &&
5575 "Vector shuffle must be normalized in DAG");
5577 // FIXME: implement canonicalizations from DAG.getVectorShuffle()
5579 // If it is a splat, check if the argument vector is a build_vector with
5580 // all scalar elements the same.
5581 if (cast<ShuffleVectorSDNode>(N)->isSplat()) {
5582 SDNode *V = N0.getNode();
5585 // If this is a bit convert that changes the element type of the vector but
5586 // not the number of vector elements, look through it. Be careful not to
5587 // look though conversions that change things like v4f32 to v2f64.
5588 if (V->getOpcode() == ISD::BIT_CONVERT) {
5589 SDValue ConvInput = V->getOperand(0);
5590 if (ConvInput.getValueType().isVector() &&
5591 ConvInput.getValueType().getVectorNumElements() == NumElts)
5592 V = ConvInput.getNode();
5595 if (V->getOpcode() == ISD::BUILD_VECTOR) {
5596 unsigned NumElems = V->getNumOperands();
5597 unsigned BaseIdx = cast<ShuffleVectorSDNode>(N)->getSplatIndex();
5598 if (NumElems > BaseIdx) {
5600 bool AllSame = true;
5601 for (unsigned i = 0; i != NumElems; ++i) {
5602 if (V->getOperand(i).getOpcode() != ISD::UNDEF) {
5603 Base = V->getOperand(i);
5607 // Splat of <u, u, u, u>, return <u, u, u, u>
5608 if (!Base.getNode())
5610 for (unsigned i = 0; i != NumElems; ++i) {
5611 if (V->getOperand(i) != Base) {
5616 // Splat of <x, x, x, x>, return <x, x, x, x>
5625 /// XformToShuffleWithZero - Returns a vector_shuffle if it able to transform
5626 /// an AND to a vector_shuffle with the destination vector and a zero vector.
5627 /// e.g. AND V, <0xffffffff, 0, 0xffffffff, 0>. ==>
5628 /// vector_shuffle V, Zero, <0, 4, 2, 4>
5629 SDValue DAGCombiner::XformToShuffleWithZero(SDNode *N) {
5630 EVT VT = N->getValueType(0);
5631 DebugLoc dl = N->getDebugLoc();
5632 SDValue LHS = N->getOperand(0);
5633 SDValue RHS = N->getOperand(1);
5634 if (N->getOpcode() == ISD::AND) {
5635 if (RHS.getOpcode() == ISD::BIT_CONVERT)
5636 RHS = RHS.getOperand(0);
5637 if (RHS.getOpcode() == ISD::BUILD_VECTOR) {
5638 SmallVector<int, 8> Indices;
5639 unsigned NumElts = RHS.getNumOperands();
5640 for (unsigned i = 0; i != NumElts; ++i) {
5641 SDValue Elt = RHS.getOperand(i);
5642 if (!isa<ConstantSDNode>(Elt))
5644 else if (cast<ConstantSDNode>(Elt)->isAllOnesValue())
5645 Indices.push_back(i);
5646 else if (cast<ConstantSDNode>(Elt)->isNullValue())
5647 Indices.push_back(NumElts);
5652 // Let's see if the target supports this vector_shuffle.
5653 EVT RVT = RHS.getValueType();
5654 if (!TLI.isVectorClearMaskLegal(Indices, RVT))
5657 // Return the new VECTOR_SHUFFLE node.
5658 EVT EltVT = RVT.getVectorElementType();
5659 SmallVector<SDValue,8> ZeroOps(RVT.getVectorNumElements(),
5660 DAG.getConstant(0, EltVT));
5661 SDValue Zero = DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
5662 RVT, &ZeroOps[0], ZeroOps.size());
5663 LHS = DAG.getNode(ISD::BIT_CONVERT, dl, RVT, LHS);
5664 SDValue Shuf = DAG.getVectorShuffle(RVT, dl, LHS, Zero, &Indices[0]);
5665 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Shuf);
5672 /// SimplifyVBinOp - Visit a binary vector operation, like ADD.
5673 SDValue DAGCombiner::SimplifyVBinOp(SDNode *N) {
5674 // After legalize, the target may be depending on adds and other
5675 // binary ops to provide legal ways to construct constants or other
5676 // things. Simplifying them may result in a loss of legality.
5677 if (LegalOperations) return SDValue();
5679 EVT VT = N->getValueType(0);
5680 assert(VT.isVector() && "SimplifyVBinOp only works on vectors!");
5682 EVT EltType = VT.getVectorElementType();
5683 SDValue LHS = N->getOperand(0);
5684 SDValue RHS = N->getOperand(1);
5685 SDValue Shuffle = XformToShuffleWithZero(N);
5686 if (Shuffle.getNode()) return Shuffle;
5688 // If the LHS and RHS are BUILD_VECTOR nodes, see if we can constant fold
5690 if (LHS.getOpcode() == ISD::BUILD_VECTOR &&
5691 RHS.getOpcode() == ISD::BUILD_VECTOR) {
5692 SmallVector<SDValue, 8> Ops;
5693 for (unsigned i = 0, e = LHS.getNumOperands(); i != e; ++i) {
5694 SDValue LHSOp = LHS.getOperand(i);
5695 SDValue RHSOp = RHS.getOperand(i);
5696 // If these two elements can't be folded, bail out.
5697 if ((LHSOp.getOpcode() != ISD::UNDEF &&
5698 LHSOp.getOpcode() != ISD::Constant &&
5699 LHSOp.getOpcode() != ISD::ConstantFP) ||
5700 (RHSOp.getOpcode() != ISD::UNDEF &&
5701 RHSOp.getOpcode() != ISD::Constant &&
5702 RHSOp.getOpcode() != ISD::ConstantFP))
5705 // Can't fold divide by zero.
5706 if (N->getOpcode() == ISD::SDIV || N->getOpcode() == ISD::UDIV ||
5707 N->getOpcode() == ISD::FDIV) {
5708 if ((RHSOp.getOpcode() == ISD::Constant &&
5709 cast<ConstantSDNode>(RHSOp.getNode())->isNullValue()) ||
5710 (RHSOp.getOpcode() == ISD::ConstantFP &&
5711 cast<ConstantFPSDNode>(RHSOp.getNode())->getValueAPF().isZero()))
5715 Ops.push_back(DAG.getNode(N->getOpcode(), LHS.getDebugLoc(),
5716 EltType, LHSOp, RHSOp));
5717 AddToWorkList(Ops.back().getNode());
5718 assert((Ops.back().getOpcode() == ISD::UNDEF ||
5719 Ops.back().getOpcode() == ISD::Constant ||
5720 Ops.back().getOpcode() == ISD::ConstantFP) &&
5721 "Scalar binop didn't fold!");
5724 if (Ops.size() == LHS.getNumOperands()) {
5725 EVT VT = LHS.getValueType();
5726 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), VT,
5727 &Ops[0], Ops.size());
5734 SDValue DAGCombiner::SimplifySelect(DebugLoc DL, SDValue N0,
5735 SDValue N1, SDValue N2){
5736 assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!");
5738 SDValue SCC = SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1), N1, N2,
5739 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5741 // If we got a simplified select_cc node back from SimplifySelectCC, then
5742 // break it down into a new SETCC node, and a new SELECT node, and then return
5743 // the SELECT node, since we were called with a SELECT node.
5744 if (SCC.getNode()) {
5745 // Check to see if we got a select_cc back (to turn into setcc/select).
5746 // Otherwise, just return whatever node we got back, like fabs.
5747 if (SCC.getOpcode() == ISD::SELECT_CC) {
5748 SDValue SETCC = DAG.getNode(ISD::SETCC, N0.getDebugLoc(),
5750 SCC.getOperand(0), SCC.getOperand(1),
5752 AddToWorkList(SETCC.getNode());
5753 return DAG.getNode(ISD::SELECT, SCC.getDebugLoc(), SCC.getValueType(),
5754 SCC.getOperand(2), SCC.getOperand(3), SETCC);
5762 /// SimplifySelectOps - Given a SELECT or a SELECT_CC node, where LHS and RHS
5763 /// are the two values being selected between, see if we can simplify the
5764 /// select. Callers of this should assume that TheSelect is deleted if this
5765 /// returns true. As such, they should return the appropriate thing (e.g. the
5766 /// node) back to the top-level of the DAG combiner loop to avoid it being
5768 bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDValue LHS,
5771 // If this is a select from two identical things, try to pull the operation
5772 // through the select.
5773 if (LHS.getOpcode() == RHS.getOpcode() && LHS.hasOneUse() && RHS.hasOneUse()){
5774 // If this is a load and the token chain is identical, replace the select
5775 // of two loads with a load through a select of the address to load from.
5776 // This triggers in things like "select bool X, 10.0, 123.0" after the FP
5777 // constants have been dropped into the constant pool.
5778 if (LHS.getOpcode() == ISD::LOAD &&
5779 // Do not let this transformation reduce the number of volatile loads.
5780 !cast<LoadSDNode>(LHS)->isVolatile() &&
5781 !cast<LoadSDNode>(RHS)->isVolatile() &&
5782 // Token chains must be identical.
5783 LHS.getOperand(0) == RHS.getOperand(0)) {
5784 LoadSDNode *LLD = cast<LoadSDNode>(LHS);
5785 LoadSDNode *RLD = cast<LoadSDNode>(RHS);
5787 // If this is an EXTLOAD, the VT's must match.
5788 if (LLD->getMemoryVT() == RLD->getMemoryVT()) {
5789 // FIXME: this discards src value information. This is
5790 // over-conservative. It would be beneficial to be able to remember
5791 // both potential memory locations.
5793 if (TheSelect->getOpcode() == ISD::SELECT) {
5794 // Check that the condition doesn't reach either load. If so, folding
5795 // this will induce a cycle into the DAG.
5796 if ((!LLD->hasAnyUseOfValue(1) ||
5797 !LLD->isPredecessorOf(TheSelect->getOperand(0).getNode())) &&
5798 (!RLD->hasAnyUseOfValue(1) ||
5799 !RLD->isPredecessorOf(TheSelect->getOperand(0).getNode()))) {
5800 Addr = DAG.getNode(ISD::SELECT, TheSelect->getDebugLoc(),
5801 LLD->getBasePtr().getValueType(),
5802 TheSelect->getOperand(0), LLD->getBasePtr(),
5806 // Check that the condition doesn't reach either load. If so, folding
5807 // this will induce a cycle into the DAG.
5808 if ((!LLD->hasAnyUseOfValue(1) ||
5809 (!LLD->isPredecessorOf(TheSelect->getOperand(0).getNode()) &&
5810 !LLD->isPredecessorOf(TheSelect->getOperand(1).getNode()))) &&
5811 (!RLD->hasAnyUseOfValue(1) ||
5812 (!RLD->isPredecessorOf(TheSelect->getOperand(0).getNode()) &&
5813 !RLD->isPredecessorOf(TheSelect->getOperand(1).getNode())))) {
5814 Addr = DAG.getNode(ISD::SELECT_CC, TheSelect->getDebugLoc(),
5815 LLD->getBasePtr().getValueType(),
5816 TheSelect->getOperand(0),
5817 TheSelect->getOperand(1),
5818 LLD->getBasePtr(), RLD->getBasePtr(),
5819 TheSelect->getOperand(4));
5823 if (Addr.getNode()) {
5825 if (LLD->getExtensionType() == ISD::NON_EXTLOAD) {
5826 Load = DAG.getLoad(TheSelect->getValueType(0),
5827 TheSelect->getDebugLoc(),
5831 LLD->getAlignment());
5833 Load = DAG.getExtLoad(LLD->getExtensionType(),
5834 TheSelect->getDebugLoc(),
5835 TheSelect->getValueType(0),
5836 LLD->getChain(), Addr, 0, 0,
5839 LLD->getAlignment());
5842 // Users of the select now use the result of the load.
5843 CombineTo(TheSelect, Load);
5845 // Users of the old loads now use the new load's chain. We know the
5846 // old-load value is dead now.
5847 CombineTo(LHS.getNode(), Load.getValue(0), Load.getValue(1));
5848 CombineTo(RHS.getNode(), Load.getValue(0), Load.getValue(1));
5858 /// SimplifySelectCC - Simplify an expression of the form (N0 cond N1) ? N2 : N3
5859 /// where 'cond' is the comparison specified by CC.
5860 SDValue DAGCombiner::SimplifySelectCC(DebugLoc DL, SDValue N0, SDValue N1,
5861 SDValue N2, SDValue N3,
5862 ISD::CondCode CC, bool NotExtCompare) {
5863 // (x ? y : y) -> y.
5864 if (N2 == N3) return N2;
5866 EVT VT = N2.getValueType();
5867 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
5868 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
5869 ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N3.getNode());
5871 // Determine if the condition we're dealing with is constant
5872 SDValue SCC = SimplifySetCC(TLI.getSetCCResultType(N0.getValueType()),
5873 N0, N1, CC, DL, false);
5874 if (SCC.getNode()) AddToWorkList(SCC.getNode());
5875 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode());
5877 // fold select_cc true, x, y -> x
5878 if (SCCC && !SCCC->isNullValue())
5880 // fold select_cc false, x, y -> y
5881 if (SCCC && SCCC->isNullValue())
5884 // Check to see if we can simplify the select into an fabs node
5885 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) {
5886 // Allow either -0.0 or 0.0
5887 if (CFP->getValueAPF().isZero()) {
5888 // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs
5889 if ((CC == ISD::SETGE || CC == ISD::SETGT) &&
5890 N0 == N2 && N3.getOpcode() == ISD::FNEG &&
5891 N2 == N3.getOperand(0))
5892 return DAG.getNode(ISD::FABS, DL, VT, N0);
5894 // select (setl[te] X, +/-0.0), fneg(X), X -> fabs
5895 if ((CC == ISD::SETLT || CC == ISD::SETLE) &&
5896 N0 == N3 && N2.getOpcode() == ISD::FNEG &&
5897 N2.getOperand(0) == N3)
5898 return DAG.getNode(ISD::FABS, DL, VT, N3);
5902 // Turn "(a cond b) ? 1.0f : 2.0f" into "load (tmp + ((a cond b) ? 0 : 4)"
5903 // where "tmp" is a constant pool entry containing an array with 1.0 and 2.0
5904 // in it. This is a win when the constant is not otherwise available because
5905 // it replaces two constant pool loads with one. We only do this if the FP
5906 // type is known to be legal, because if it isn't, then we are before legalize
5907 // types an we want the other legalization to happen first (e.g. to avoid
5908 // messing with soft float) and if the ConstantFP is not legal, because if
5909 // it is legal, we may not need to store the FP constant in a constant pool.
5910 if (ConstantFPSDNode *TV = dyn_cast<ConstantFPSDNode>(N2))
5911 if (ConstantFPSDNode *FV = dyn_cast<ConstantFPSDNode>(N3)) {
5912 if (TLI.isTypeLegal(N2.getValueType()) &&
5913 (TLI.getOperationAction(ISD::ConstantFP, N2.getValueType()) !=
5914 TargetLowering::Legal) &&
5915 // If both constants have multiple uses, then we won't need to do an
5916 // extra load, they are likely around in registers for other users.
5917 (TV->hasOneUse() || FV->hasOneUse())) {
5918 Constant *Elts[] = {
5919 const_cast<ConstantFP*>(FV->getConstantFPValue()),
5920 const_cast<ConstantFP*>(TV->getConstantFPValue())
5922 const Type *FPTy = Elts[0]->getType();
5923 const TargetData &TD = *TLI.getTargetData();
5925 // Create a ConstantArray of the two constants.
5926 Constant *CA = ConstantArray::get(ArrayType::get(FPTy, 2), Elts, 2);
5927 SDValue CPIdx = DAG.getConstantPool(CA, TLI.getPointerTy(),
5928 TD.getPrefTypeAlignment(FPTy));
5929 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
5931 // Get the offsets to the 0 and 1 element of the array so that we can
5932 // select between them.
5933 SDValue Zero = DAG.getIntPtrConstant(0);
5934 unsigned EltSize = (unsigned)TD.getTypeAllocSize(Elts[0]->getType());
5935 SDValue One = DAG.getIntPtrConstant(EltSize);
5937 SDValue Cond = DAG.getSetCC(DL,
5938 TLI.getSetCCResultType(N0.getValueType()),
5940 SDValue CstOffset = DAG.getNode(ISD::SELECT, DL, Zero.getValueType(),
5942 CPIdx = DAG.getNode(ISD::ADD, DL, TLI.getPointerTy(), CPIdx,
5944 return DAG.getLoad(TV->getValueType(0), DL, DAG.getEntryNode(), CPIdx,
5945 PseudoSourceValue::getConstantPool(), 0, false,
5951 // Check to see if we can perform the "gzip trick", transforming
5952 // (select_cc setlt X, 0, A, 0) -> (and (sra X, (sub size(X), 1), A)
5953 if (N1C && N3C && N3C->isNullValue() && CC == ISD::SETLT &&
5954 N0.getValueType().isInteger() &&
5955 N2.getValueType().isInteger() &&
5956 (N1C->isNullValue() || // (a < 0) ? b : 0
5957 (N1C->getAPIntValue() == 1 && N0 == N2))) { // (a < 1) ? a : 0
5958 EVT XType = N0.getValueType();
5959 EVT AType = N2.getValueType();
5960 if (XType.bitsGE(AType)) {
5961 // and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a
5962 // single-bit constant.
5963 if (N2C && ((N2C->getAPIntValue() & (N2C->getAPIntValue()-1)) == 0)) {
5964 unsigned ShCtV = N2C->getAPIntValue().logBase2();
5965 ShCtV = XType.getSizeInBits()-ShCtV-1;
5966 SDValue ShCt = DAG.getConstant(ShCtV, getShiftAmountTy());
5967 SDValue Shift = DAG.getNode(ISD::SRL, N0.getDebugLoc(),
5969 AddToWorkList(Shift.getNode());
5971 if (XType.bitsGT(AType)) {
5972 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
5973 AddToWorkList(Shift.getNode());
5976 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
5979 SDValue Shift = DAG.getNode(ISD::SRA, N0.getDebugLoc(),
5981 DAG.getConstant(XType.getSizeInBits()-1,
5982 getShiftAmountTy()));
5983 AddToWorkList(Shift.getNode());
5985 if (XType.bitsGT(AType)) {
5986 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
5987 AddToWorkList(Shift.getNode());
5990 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
5994 // fold select C, 16, 0 -> shl C, 4
5995 if (N2C && N3C && N3C->isNullValue() && N2C->getAPIntValue().isPowerOf2() &&
5996 TLI.getBooleanContents() == TargetLowering::ZeroOrOneBooleanContent) {
5998 // If the caller doesn't want us to simplify this into a zext of a compare,
6000 if (NotExtCompare && N2C->getAPIntValue() == 1)
6003 // Get a SetCC of the condition
6004 // FIXME: Should probably make sure that setcc is legal if we ever have a
6005 // target where it isn't.
6007 // cast from setcc result type to select result type
6009 SCC = DAG.getSetCC(DL, TLI.getSetCCResultType(N0.getValueType()),
6011 if (N2.getValueType().bitsLT(SCC.getValueType()))
6012 Temp = DAG.getZeroExtendInReg(SCC, N2.getDebugLoc(), N2.getValueType());
6014 Temp = DAG.getNode(ISD::ZERO_EXTEND, N2.getDebugLoc(),
6015 N2.getValueType(), SCC);
6017 SCC = DAG.getSetCC(N0.getDebugLoc(), MVT::i1, N0, N1, CC);
6018 Temp = DAG.getNode(ISD::ZERO_EXTEND, N2.getDebugLoc(),
6019 N2.getValueType(), SCC);
6022 AddToWorkList(SCC.getNode());
6023 AddToWorkList(Temp.getNode());
6025 if (N2C->getAPIntValue() == 1)
6028 // shl setcc result by log2 n2c
6029 return DAG.getNode(ISD::SHL, DL, N2.getValueType(), Temp,
6030 DAG.getConstant(N2C->getAPIntValue().logBase2(),
6031 getShiftAmountTy()));
6034 // Check to see if this is the equivalent of setcc
6035 // FIXME: Turn all of these into setcc if setcc if setcc is legal
6036 // otherwise, go ahead with the folds.
6037 if (0 && N3C && N3C->isNullValue() && N2C && (N2C->getAPIntValue() == 1ULL)) {
6038 EVT XType = N0.getValueType();
6039 if (!LegalOperations ||
6040 TLI.isOperationLegal(ISD::SETCC, TLI.getSetCCResultType(XType))) {
6041 SDValue Res = DAG.getSetCC(DL, TLI.getSetCCResultType(XType), N0, N1, CC);
6042 if (Res.getValueType() != VT)
6043 Res = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Res);
6047 // fold (seteq X, 0) -> (srl (ctlz X, log2(size(X))))
6048 if (N1C && N1C->isNullValue() && CC == ISD::SETEQ &&
6049 (!LegalOperations ||
6050 TLI.isOperationLegal(ISD::CTLZ, XType))) {
6051 SDValue Ctlz = DAG.getNode(ISD::CTLZ, N0.getDebugLoc(), XType, N0);
6052 return DAG.getNode(ISD::SRL, DL, XType, Ctlz,
6053 DAG.getConstant(Log2_32(XType.getSizeInBits()),
6054 getShiftAmountTy()));
6056 // fold (setgt X, 0) -> (srl (and (-X, ~X), size(X)-1))
6057 if (N1C && N1C->isNullValue() && CC == ISD::SETGT) {
6058 SDValue NegN0 = DAG.getNode(ISD::SUB, N0.getDebugLoc(),
6059 XType, DAG.getConstant(0, XType), N0);
6060 SDValue NotN0 = DAG.getNOT(N0.getDebugLoc(), N0, XType);
6061 return DAG.getNode(ISD::SRL, DL, XType,
6062 DAG.getNode(ISD::AND, DL, XType, NegN0, NotN0),
6063 DAG.getConstant(XType.getSizeInBits()-1,
6064 getShiftAmountTy()));
6066 // fold (setgt X, -1) -> (xor (srl (X, size(X)-1), 1))
6067 if (N1C && N1C->isAllOnesValue() && CC == ISD::SETGT) {
6068 SDValue Sign = DAG.getNode(ISD::SRL, N0.getDebugLoc(), XType, N0,
6069 DAG.getConstant(XType.getSizeInBits()-1,
6070 getShiftAmountTy()));
6071 return DAG.getNode(ISD::XOR, DL, XType, Sign, DAG.getConstant(1, XType));
6075 // Check to see if this is an integer abs. select_cc setl[te] X, 0, -X, X ->
6076 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
6077 if (N1C && N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE) &&
6078 N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1) &&
6079 N2.getOperand(0) == N1 && N0.getValueType().isInteger()) {
6080 EVT XType = N0.getValueType();
6081 SDValue Shift = DAG.getNode(ISD::SRA, N0.getDebugLoc(), XType, N0,
6082 DAG.getConstant(XType.getSizeInBits()-1,
6083 getShiftAmountTy()));
6084 SDValue Add = DAG.getNode(ISD::ADD, N0.getDebugLoc(), XType,
6086 AddToWorkList(Shift.getNode());
6087 AddToWorkList(Add.getNode());
6088 return DAG.getNode(ISD::XOR, DL, XType, Add, Shift);
6090 // Check to see if this is an integer abs. select_cc setgt X, -1, X, -X ->
6091 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
6092 if (N1C && N1C->isAllOnesValue() && CC == ISD::SETGT &&
6093 N0 == N2 && N3.getOpcode() == ISD::SUB && N0 == N3.getOperand(1)) {
6094 if (ConstantSDNode *SubC = dyn_cast<ConstantSDNode>(N3.getOperand(0))) {
6095 EVT XType = N0.getValueType();
6096 if (SubC->isNullValue() && XType.isInteger()) {
6097 SDValue Shift = DAG.getNode(ISD::SRA, N0.getDebugLoc(), XType,
6099 DAG.getConstant(XType.getSizeInBits()-1,
6100 getShiftAmountTy()));
6101 SDValue Add = DAG.getNode(ISD::ADD, N0.getDebugLoc(),
6103 AddToWorkList(Shift.getNode());
6104 AddToWorkList(Add.getNode());
6105 return DAG.getNode(ISD::XOR, DL, XType, Add, Shift);
6113 /// SimplifySetCC - This is a stub for TargetLowering::SimplifySetCC.
6114 SDValue DAGCombiner::SimplifySetCC(EVT VT, SDValue N0,
6115 SDValue N1, ISD::CondCode Cond,
6116 DebugLoc DL, bool foldBooleans) {
6117 TargetLowering::DAGCombinerInfo
6118 DagCombineInfo(DAG, !LegalTypes, !LegalOperations, false, this);
6119 return TLI.SimplifySetCC(VT, N0, N1, Cond, foldBooleans, DagCombineInfo, DL);
6122 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
6123 /// return a DAG expression to select that will generate the same value by
6124 /// multiplying by a magic number. See:
6125 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
6126 SDValue DAGCombiner::BuildSDIV(SDNode *N) {
6127 std::vector<SDNode*> Built;
6128 SDValue S = TLI.BuildSDIV(N, DAG, &Built);
6130 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end();
6136 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
6137 /// return a DAG expression to select that will generate the same value by
6138 /// multiplying by a magic number. See:
6139 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
6140 SDValue DAGCombiner::BuildUDIV(SDNode *N) {
6141 std::vector<SDNode*> Built;
6142 SDValue S = TLI.BuildUDIV(N, DAG, &Built);
6144 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end();
6150 /// FindBaseOffset - Return true if base is a frame index, which is known not
6151 // to alias with anything but itself. Provides base object and offset as results.
6152 static bool FindBaseOffset(SDValue Ptr, SDValue &Base, int64_t &Offset,
6153 GlobalValue *&GV, void *&CV) {
6154 // Assume it is a primitive operation.
6155 Base = Ptr; Offset = 0; GV = 0; CV = 0;
6157 // If it's an adding a simple constant then integrate the offset.
6158 if (Base.getOpcode() == ISD::ADD) {
6159 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Base.getOperand(1))) {
6160 Base = Base.getOperand(0);
6161 Offset += C->getZExtValue();
6165 // Return the underlying GlobalValue, and update the Offset. Return false
6166 // for GlobalAddressSDNode since the same GlobalAddress may be represented
6167 // by multiple nodes with different offsets.
6168 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Base)) {
6169 GV = G->getGlobal();
6170 Offset += G->getOffset();
6174 // Return the underlying Constant value, and update the Offset. Return false
6175 // for ConstantSDNodes since the same constant pool entry may be represented
6176 // by multiple nodes with different offsets.
6177 if (ConstantPoolSDNode *C = dyn_cast<ConstantPoolSDNode>(Base)) {
6178 CV = C->isMachineConstantPoolEntry() ? (void *)C->getMachineCPVal()
6179 : (void *)C->getConstVal();
6180 Offset += C->getOffset();
6183 // If it's any of the following then it can't alias with anything but itself.
6184 return isa<FrameIndexSDNode>(Base);
6187 /// isAlias - Return true if there is any possibility that the two addresses
6189 bool DAGCombiner::isAlias(SDValue Ptr1, int64_t Size1,
6190 const Value *SrcValue1, int SrcValueOffset1,
6191 unsigned SrcValueAlign1,
6192 SDValue Ptr2, int64_t Size2,
6193 const Value *SrcValue2, int SrcValueOffset2,
6194 unsigned SrcValueAlign2) const {
6195 // If they are the same then they must be aliases.
6196 if (Ptr1 == Ptr2) return true;
6198 // Gather base node and offset information.
6199 SDValue Base1, Base2;
6200 int64_t Offset1, Offset2;
6201 GlobalValue *GV1, *GV2;
6203 bool isFrameIndex1 = FindBaseOffset(Ptr1, Base1, Offset1, GV1, CV1);
6204 bool isFrameIndex2 = FindBaseOffset(Ptr2, Base2, Offset2, GV2, CV2);
6206 // If they have a same base address then check to see if they overlap.
6207 if (Base1 == Base2 || (GV1 && (GV1 == GV2)) || (CV1 && (CV1 == CV2)))
6208 return !((Offset1 + Size1) <= Offset2 || (Offset2 + Size2) <= Offset1);
6210 // If we know what the bases are, and they aren't identical, then we know they
6212 if ((isFrameIndex1 || CV1 || GV1) && (isFrameIndex2 || CV2 || GV2))
6215 // If we know required SrcValue1 and SrcValue2 have relatively large alignment
6216 // compared to the size and offset of the access, we may be able to prove they
6217 // do not alias. This check is conservative for now to catch cases created by
6218 // splitting vector types.
6219 if ((SrcValueAlign1 == SrcValueAlign2) &&
6220 (SrcValueOffset1 != SrcValueOffset2) &&
6221 (Size1 == Size2) && (SrcValueAlign1 > Size1)) {
6222 int64_t OffAlign1 = SrcValueOffset1 % SrcValueAlign1;
6223 int64_t OffAlign2 = SrcValueOffset2 % SrcValueAlign1;
6225 // There is no overlap between these relatively aligned accesses of similar
6226 // size, return no alias.
6227 if ((OffAlign1 + Size1) <= OffAlign2 || (OffAlign2 + Size2) <= OffAlign1)
6231 if (CombinerGlobalAA) {
6232 // Use alias analysis information.
6233 int64_t MinOffset = std::min(SrcValueOffset1, SrcValueOffset2);
6234 int64_t Overlap1 = Size1 + SrcValueOffset1 - MinOffset;
6235 int64_t Overlap2 = Size2 + SrcValueOffset2 - MinOffset;
6236 AliasAnalysis::AliasResult AAResult =
6237 AA.alias(SrcValue1, Overlap1, SrcValue2, Overlap2);
6238 if (AAResult == AliasAnalysis::NoAlias)
6242 // Otherwise we have to assume they alias.
6246 /// FindAliasInfo - Extracts the relevant alias information from the memory
6247 /// node. Returns true if the operand was a load.
6248 bool DAGCombiner::FindAliasInfo(SDNode *N,
6249 SDValue &Ptr, int64_t &Size,
6250 const Value *&SrcValue,
6251 int &SrcValueOffset,
6252 unsigned &SrcValueAlign) const {
6253 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
6254 Ptr = LD->getBasePtr();
6255 Size = LD->getMemoryVT().getSizeInBits() >> 3;
6256 SrcValue = LD->getSrcValue();
6257 SrcValueOffset = LD->getSrcValueOffset();
6258 SrcValueAlign = LD->getOriginalAlignment();
6260 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
6261 Ptr = ST->getBasePtr();
6262 Size = ST->getMemoryVT().getSizeInBits() >> 3;
6263 SrcValue = ST->getSrcValue();
6264 SrcValueOffset = ST->getSrcValueOffset();
6265 SrcValueAlign = ST->getOriginalAlignment();
6267 llvm_unreachable("FindAliasInfo expected a memory operand");
6273 /// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes,
6274 /// looking for aliasing nodes and adding them to the Aliases vector.
6275 void DAGCombiner::GatherAllAliases(SDNode *N, SDValue OriginalChain,
6276 SmallVector<SDValue, 8> &Aliases) {
6277 SmallVector<SDValue, 8> Chains; // List of chains to visit.
6278 SmallPtrSet<SDNode *, 16> Visited; // Visited node set.
6280 // Get alias information for node.
6283 const Value *SrcValue;
6285 unsigned SrcValueAlign;
6286 bool IsLoad = FindAliasInfo(N, Ptr, Size, SrcValue, SrcValueOffset,
6290 Chains.push_back(OriginalChain);
6293 // Look at each chain and determine if it is an alias. If so, add it to the
6294 // aliases list. If not, then continue up the chain looking for the next
6296 while (!Chains.empty()) {
6297 SDValue Chain = Chains.back();
6300 // For TokenFactor nodes, look at each operand and only continue up the
6301 // chain until we find two aliases. If we've seen two aliases, assume we'll
6302 // find more and revert to original chain since the xform is unlikely to be
6305 // FIXME: The depth check could be made to return the last non-aliasing
6306 // chain we found before we hit a tokenfactor rather than the original
6308 if (Depth > 6 || Aliases.size() == 2) {
6310 Aliases.push_back(OriginalChain);
6314 // Don't bother if we've been before.
6315 if (!Visited.insert(Chain.getNode()))
6318 switch (Chain.getOpcode()) {
6319 case ISD::EntryToken:
6320 // Entry token is ideal chain operand, but handled in FindBetterChain.
6325 // Get alias information for Chain.
6328 const Value *OpSrcValue;
6329 int OpSrcValueOffset;
6330 unsigned OpSrcValueAlign;
6331 bool IsOpLoad = FindAliasInfo(Chain.getNode(), OpPtr, OpSize,
6332 OpSrcValue, OpSrcValueOffset,
6335 // If chain is alias then stop here.
6336 if (!(IsLoad && IsOpLoad) &&
6337 isAlias(Ptr, Size, SrcValue, SrcValueOffset, SrcValueAlign,
6338 OpPtr, OpSize, OpSrcValue, OpSrcValueOffset,
6340 Aliases.push_back(Chain);
6342 // Look further up the chain.
6343 Chains.push_back(Chain.getOperand(0));
6349 case ISD::TokenFactor:
6350 // We have to check each of the operands of the token factor for "small"
6351 // token factors, so we queue them up. Adding the operands to the queue
6352 // (stack) in reverse order maintains the original order and increases the
6353 // likelihood that getNode will find a matching token factor (CSE.)
6354 if (Chain.getNumOperands() > 16) {
6355 Aliases.push_back(Chain);
6358 for (unsigned n = Chain.getNumOperands(); n;)
6359 Chains.push_back(Chain.getOperand(--n));
6364 // For all other instructions we will just have to take what we can get.
6365 Aliases.push_back(Chain);
6371 /// FindBetterChain - Walk up chain skipping non-aliasing memory nodes, looking
6372 /// for a better chain (aliasing node.)
6373 SDValue DAGCombiner::FindBetterChain(SDNode *N, SDValue OldChain) {
6374 SmallVector<SDValue, 8> Aliases; // Ops for replacing token factor.
6376 // Accumulate all the aliases to this node.
6377 GatherAllAliases(N, OldChain, Aliases);
6379 if (Aliases.size() == 0) {
6380 // If no operands then chain to entry token.
6381 return DAG.getEntryNode();
6382 } else if (Aliases.size() == 1) {
6383 // If a single operand then chain to it. We don't need to revisit it.
6387 // Construct a custom tailored token factor.
6388 return DAG.getNode(ISD::TokenFactor, N->getDebugLoc(), MVT::Other,
6389 &Aliases[0], Aliases.size());
6392 // SelectionDAG::Combine - This is the entry point for the file.
6394 void SelectionDAG::Combine(CombineLevel Level, AliasAnalysis &AA,
6395 CodeGenOpt::Level OptLevel) {
6396 /// run - This is the main entry point to this class.
6398 DAGCombiner(*this, AA, OptLevel).Run(Level);