1 //===-- DAGCombiner.cpp - Implement a DAG node combiner -------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass combines dag nodes to form fewer, simpler DAG nodes. It can be run
11 // both before and after the DAG is legalized.
13 // This pass is not a substitute for the LLVM IR instcombine pass. This pass is
14 // primarily intended to handle simplification opportunities that are implicit
15 // in the LLVM IR and exposed by the various codegen lowering phases.
17 //===----------------------------------------------------------------------===//
19 #define DEBUG_TYPE "dagcombine"
20 #include "llvm/CodeGen/SelectionDAG.h"
21 #include "llvm/DerivedTypes.h"
22 #include "llvm/LLVMContext.h"
23 #include "llvm/CodeGen/MachineFunction.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/Analysis/AliasAnalysis.h"
26 #include "llvm/Target/TargetData.h"
27 #include "llvm/Target/TargetLowering.h"
28 #include "llvm/Target/TargetMachine.h"
29 #include "llvm/Target/TargetOptions.h"
30 #include "llvm/ADT/SmallPtrSet.h"
31 #include "llvm/ADT/Statistic.h"
32 #include "llvm/Support/CommandLine.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Support/ErrorHandling.h"
35 #include "llvm/Support/MathExtras.h"
36 #include "llvm/Support/raw_ostream.h"
40 STATISTIC(NodesCombined , "Number of dag nodes combined");
41 STATISTIC(PreIndexedNodes , "Number of pre-indexed nodes created");
42 STATISTIC(PostIndexedNodes, "Number of post-indexed nodes created");
43 STATISTIC(OpsNarrowed , "Number of load/op/store narrowed");
44 STATISTIC(LdStFP2Int , "Number of fp load/store pairs transformed to int");
48 CombinerAA("combiner-alias-analysis", cl::Hidden,
49 cl::desc("Turn on alias analysis during testing"));
52 CombinerGlobalAA("combiner-global-alias-analysis", cl::Hidden,
53 cl::desc("Include global information in alias analysis"));
55 //------------------------------ DAGCombiner ---------------------------------//
59 const TargetLowering &TLI;
61 CodeGenOpt::Level OptLevel;
65 // Worklist of all of the nodes that need to be simplified.
67 // This has the semantics that when adding to the worklist,
68 // the item added must be next to be processed. It should
69 // also only appear once. The naive approach to this takes
72 // To reduce the insert/remove time to logarithmic, we use
73 // a set and a vector to maintain our worklist.
75 // The set contains the items on the worklist, but does not
76 // maintain the order they should be visited.
78 // The vector maintains the order nodes should be visited, but may
79 // contain duplicate or removed nodes. When choosing a node to
80 // visit, we pop off the order stack until we find an item that is
81 // also in the contents set. All operations are O(log N).
82 SmallPtrSet<SDNode*, 64> WorkListContents;
83 SmallVector<SDNode*, 64> WorkListOrder;
85 // AA - Used for DAG load/store alias analysis.
88 /// AddUsersToWorkList - When an instruction is simplified, add all users of
89 /// the instruction to the work lists because they might get more simplified
92 void AddUsersToWorkList(SDNode *N) {
93 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
98 /// visit - call the node-specific routine that knows how to fold each
99 /// particular type of node.
100 SDValue visit(SDNode *N);
103 /// AddToWorkList - Add to the work list making sure its instance is at the
104 /// back (next to be processed.)
105 void AddToWorkList(SDNode *N) {
106 WorkListContents.insert(N);
107 WorkListOrder.push_back(N);
110 /// removeFromWorkList - remove all instances of N from the worklist.
112 void removeFromWorkList(SDNode *N) {
113 WorkListContents.erase(N);
116 SDValue CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
119 SDValue CombineTo(SDNode *N, SDValue Res, bool AddTo = true) {
120 return CombineTo(N, &Res, 1, AddTo);
123 SDValue CombineTo(SDNode *N, SDValue Res0, SDValue Res1,
125 SDValue To[] = { Res0, Res1 };
126 return CombineTo(N, To, 2, AddTo);
129 void CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO);
133 /// SimplifyDemandedBits - Check the specified integer node value to see if
134 /// it can be simplified or if things it uses can be simplified by bit
135 /// propagation. If so, return true.
136 bool SimplifyDemandedBits(SDValue Op) {
137 unsigned BitWidth = Op.getValueType().getScalarType().getSizeInBits();
138 APInt Demanded = APInt::getAllOnesValue(BitWidth);
139 return SimplifyDemandedBits(Op, Demanded);
142 bool SimplifyDemandedBits(SDValue Op, const APInt &Demanded);
144 bool CombineToPreIndexedLoadStore(SDNode *N);
145 bool CombineToPostIndexedLoadStore(SDNode *N);
147 void ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad);
148 SDValue PromoteOperand(SDValue Op, EVT PVT, bool &Replace);
149 SDValue SExtPromoteOperand(SDValue Op, EVT PVT);
150 SDValue ZExtPromoteOperand(SDValue Op, EVT PVT);
151 SDValue PromoteIntBinOp(SDValue Op);
152 SDValue PromoteIntShiftOp(SDValue Op);
153 SDValue PromoteExtend(SDValue Op);
154 bool PromoteLoad(SDValue Op);
156 void ExtendSetCCUses(SmallVector<SDNode*, 4> SetCCs,
157 SDValue Trunc, SDValue ExtLoad, DebugLoc DL,
158 ISD::NodeType ExtType);
160 /// combine - call the node-specific routine that knows how to fold each
161 /// particular type of node. If that doesn't do anything, try the
162 /// target-specific DAG combines.
163 SDValue combine(SDNode *N);
165 // Visitation implementation - Implement dag node combining for different
166 // node types. The semantics are as follows:
168 // SDValue.getNode() == 0 - No change was made
169 // SDValue.getNode() == N - N was replaced, is dead and has been handled.
170 // otherwise - N should be replaced by the returned Operand.
172 SDValue visitTokenFactor(SDNode *N);
173 SDValue visitMERGE_VALUES(SDNode *N);
174 SDValue visitADD(SDNode *N);
175 SDValue visitSUB(SDNode *N);
176 SDValue visitADDC(SDNode *N);
177 SDValue visitSUBC(SDNode *N);
178 SDValue visitADDE(SDNode *N);
179 SDValue visitSUBE(SDNode *N);
180 SDValue visitMUL(SDNode *N);
181 SDValue visitSDIV(SDNode *N);
182 SDValue visitUDIV(SDNode *N);
183 SDValue visitSREM(SDNode *N);
184 SDValue visitUREM(SDNode *N);
185 SDValue visitMULHU(SDNode *N);
186 SDValue visitMULHS(SDNode *N);
187 SDValue visitSMUL_LOHI(SDNode *N);
188 SDValue visitUMUL_LOHI(SDNode *N);
189 SDValue visitSMULO(SDNode *N);
190 SDValue visitUMULO(SDNode *N);
191 SDValue visitSDIVREM(SDNode *N);
192 SDValue visitUDIVREM(SDNode *N);
193 SDValue visitAND(SDNode *N);
194 SDValue visitOR(SDNode *N);
195 SDValue visitXOR(SDNode *N);
196 SDValue SimplifyVBinOp(SDNode *N);
197 SDValue visitSHL(SDNode *N);
198 SDValue visitSRA(SDNode *N);
199 SDValue visitSRL(SDNode *N);
200 SDValue visitCTLZ(SDNode *N);
201 SDValue visitCTLZ_ZERO_UNDEF(SDNode *N);
202 SDValue visitCTTZ(SDNode *N);
203 SDValue visitCTTZ_ZERO_UNDEF(SDNode *N);
204 SDValue visitCTPOP(SDNode *N);
205 SDValue visitSELECT(SDNode *N);
206 SDValue visitSELECT_CC(SDNode *N);
207 SDValue visitSETCC(SDNode *N);
208 SDValue visitSIGN_EXTEND(SDNode *N);
209 SDValue visitZERO_EXTEND(SDNode *N);
210 SDValue visitANY_EXTEND(SDNode *N);
211 SDValue visitSIGN_EXTEND_INREG(SDNode *N);
212 SDValue visitTRUNCATE(SDNode *N);
213 SDValue visitBITCAST(SDNode *N);
214 SDValue visitBUILD_PAIR(SDNode *N);
215 SDValue visitFADD(SDNode *N);
216 SDValue visitFSUB(SDNode *N);
217 SDValue visitFMUL(SDNode *N);
218 SDValue visitFMA(SDNode *N);
219 SDValue visitFDIV(SDNode *N);
220 SDValue visitFREM(SDNode *N);
221 SDValue visitFCOPYSIGN(SDNode *N);
222 SDValue visitSINT_TO_FP(SDNode *N);
223 SDValue visitUINT_TO_FP(SDNode *N);
224 SDValue visitFP_TO_SINT(SDNode *N);
225 SDValue visitFP_TO_UINT(SDNode *N);
226 SDValue visitFP_ROUND(SDNode *N);
227 SDValue visitFP_ROUND_INREG(SDNode *N);
228 SDValue visitFP_EXTEND(SDNode *N);
229 SDValue visitFNEG(SDNode *N);
230 SDValue visitFABS(SDNode *N);
231 SDValue visitBRCOND(SDNode *N);
232 SDValue visitBR_CC(SDNode *N);
233 SDValue visitLOAD(SDNode *N);
234 SDValue visitSTORE(SDNode *N);
235 SDValue visitINSERT_VECTOR_ELT(SDNode *N);
236 SDValue visitEXTRACT_VECTOR_ELT(SDNode *N);
237 SDValue visitBUILD_VECTOR(SDNode *N);
238 SDValue visitCONCAT_VECTORS(SDNode *N);
239 SDValue visitEXTRACT_SUBVECTOR(SDNode *N);
240 SDValue visitVECTOR_SHUFFLE(SDNode *N);
241 SDValue visitMEMBARRIER(SDNode *N);
243 SDValue XformToShuffleWithZero(SDNode *N);
244 SDValue ReassociateOps(unsigned Opc, DebugLoc DL, SDValue LHS, SDValue RHS);
246 SDValue visitShiftByConstant(SDNode *N, unsigned Amt);
248 bool SimplifySelectOps(SDNode *SELECT, SDValue LHS, SDValue RHS);
249 SDValue SimplifyBinOpWithSameOpcodeHands(SDNode *N);
250 SDValue SimplifySelect(DebugLoc DL, SDValue N0, SDValue N1, SDValue N2);
251 SDValue SimplifySelectCC(DebugLoc DL, SDValue N0, SDValue N1, SDValue N2,
252 SDValue N3, ISD::CondCode CC,
253 bool NotExtCompare = false);
254 SDValue SimplifySetCC(EVT VT, SDValue N0, SDValue N1, ISD::CondCode Cond,
255 DebugLoc DL, bool foldBooleans = true);
256 SDValue SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
258 SDValue CombineConsecutiveLoads(SDNode *N, EVT VT);
259 SDValue ConstantFoldBITCASTofBUILD_VECTOR(SDNode *, EVT);
260 SDValue BuildSDIV(SDNode *N);
261 SDValue BuildUDIV(SDNode *N);
262 SDValue MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
263 bool DemandHighBits = true);
264 SDValue MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1);
265 SDNode *MatchRotate(SDValue LHS, SDValue RHS, DebugLoc DL);
266 SDValue ReduceLoadWidth(SDNode *N);
267 SDValue ReduceLoadOpStoreWidth(SDNode *N);
268 SDValue TransformFPLoadStorePair(SDNode *N);
270 SDValue GetDemandedBits(SDValue V, const APInt &Mask);
272 /// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes,
273 /// looking for aliasing nodes and adding them to the Aliases vector.
274 void GatherAllAliases(SDNode *N, SDValue OriginalChain,
275 SmallVector<SDValue, 8> &Aliases);
277 /// isAlias - Return true if there is any possibility that the two addresses
279 bool isAlias(SDValue Ptr1, int64_t Size1,
280 const Value *SrcValue1, int SrcValueOffset1,
281 unsigned SrcValueAlign1,
282 const MDNode *TBAAInfo1,
283 SDValue Ptr2, int64_t Size2,
284 const Value *SrcValue2, int SrcValueOffset2,
285 unsigned SrcValueAlign2,
286 const MDNode *TBAAInfo2) const;
288 /// FindAliasInfo - Extracts the relevant alias information from the memory
289 /// node. Returns true if the operand was a load.
290 bool FindAliasInfo(SDNode *N,
291 SDValue &Ptr, int64_t &Size,
292 const Value *&SrcValue, int &SrcValueOffset,
293 unsigned &SrcValueAlignment,
294 const MDNode *&TBAAInfo) const;
296 /// FindBetterChain - Walk up chain skipping non-aliasing memory nodes,
297 /// looking for a better chain (aliasing node.)
298 SDValue FindBetterChain(SDNode *N, SDValue Chain);
301 DAGCombiner(SelectionDAG &D, AliasAnalysis &A, CodeGenOpt::Level OL)
302 : DAG(D), TLI(D.getTargetLoweringInfo()), Level(BeforeLegalizeTypes),
303 OptLevel(OL), LegalOperations(false), LegalTypes(false), AA(A) {}
305 /// Run - runs the dag combiner on all nodes in the work list
306 void Run(CombineLevel AtLevel);
308 SelectionDAG &getDAG() const { return DAG; }
310 /// getShiftAmountTy - Returns a type large enough to hold any valid
311 /// shift amount - before type legalization these can be huge.
312 EVT getShiftAmountTy(EVT LHSTy) {
313 return LegalTypes ? TLI.getShiftAmountTy(LHSTy) : TLI.getPointerTy();
316 /// isTypeLegal - This method returns true if we are running before type
317 /// legalization or if the specified VT is legal.
318 bool isTypeLegal(const EVT &VT) {
319 if (!LegalTypes) return true;
320 return TLI.isTypeLegal(VT);
327 /// WorkListRemover - This class is a DAGUpdateListener that removes any deleted
328 /// nodes from the worklist.
329 class WorkListRemover : public SelectionDAG::DAGUpdateListener {
332 explicit WorkListRemover(DAGCombiner &dc)
333 : SelectionDAG::DAGUpdateListener(dc.getDAG()), DC(dc) {}
335 virtual void NodeDeleted(SDNode *N, SDNode *E) {
336 DC.removeFromWorkList(N);
341 //===----------------------------------------------------------------------===//
342 // TargetLowering::DAGCombinerInfo implementation
343 //===----------------------------------------------------------------------===//
345 void TargetLowering::DAGCombinerInfo::AddToWorklist(SDNode *N) {
346 ((DAGCombiner*)DC)->AddToWorkList(N);
349 void TargetLowering::DAGCombinerInfo::RemoveFromWorklist(SDNode *N) {
350 ((DAGCombiner*)DC)->removeFromWorkList(N);
353 SDValue TargetLowering::DAGCombinerInfo::
354 CombineTo(SDNode *N, const std::vector<SDValue> &To, bool AddTo) {
355 return ((DAGCombiner*)DC)->CombineTo(N, &To[0], To.size(), AddTo);
358 SDValue TargetLowering::DAGCombinerInfo::
359 CombineTo(SDNode *N, SDValue Res, bool AddTo) {
360 return ((DAGCombiner*)DC)->CombineTo(N, Res, AddTo);
364 SDValue TargetLowering::DAGCombinerInfo::
365 CombineTo(SDNode *N, SDValue Res0, SDValue Res1, bool AddTo) {
366 return ((DAGCombiner*)DC)->CombineTo(N, Res0, Res1, AddTo);
369 void TargetLowering::DAGCombinerInfo::
370 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
371 return ((DAGCombiner*)DC)->CommitTargetLoweringOpt(TLO);
374 //===----------------------------------------------------------------------===//
376 //===----------------------------------------------------------------------===//
378 /// isNegatibleForFree - Return 1 if we can compute the negated form of the
379 /// specified expression for the same cost as the expression itself, or 2 if we
380 /// can compute the negated form more cheaply than the expression itself.
381 static char isNegatibleForFree(SDValue Op, bool LegalOperations,
382 const TargetLowering &TLI,
383 const TargetOptions *Options,
384 unsigned Depth = 0) {
385 // No compile time optimizations on this type.
386 if (Op.getValueType() == MVT::ppcf128)
389 // fneg is removable even if it has multiple uses.
390 if (Op.getOpcode() == ISD::FNEG) return 2;
392 // Don't allow anything with multiple uses.
393 if (!Op.hasOneUse()) return 0;
395 // Don't recurse exponentially.
396 if (Depth > 6) return 0;
398 switch (Op.getOpcode()) {
399 default: return false;
400 case ISD::ConstantFP:
401 // Don't invert constant FP values after legalize. The negated constant
402 // isn't necessarily legal.
403 return LegalOperations ? 0 : 1;
405 // FIXME: determine better conditions for this xform.
406 if (!Options->UnsafeFPMath) return 0;
408 // After operation legalization, it might not be legal to create new FSUBs.
409 if (LegalOperations &&
410 !TLI.isOperationLegalOrCustom(ISD::FSUB, Op.getValueType()))
413 // fold (fsub (fadd A, B)) -> (fsub (fneg A), B)
414 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI,
417 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
418 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options,
421 // We can't turn -(A-B) into B-A when we honor signed zeros.
422 if (!Options->UnsafeFPMath) return 0;
424 // fold (fneg (fsub A, B)) -> (fsub B, A)
429 if (Options->HonorSignDependentRoundingFPMath()) return 0;
431 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y) or (fmul X, (fneg Y))
432 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI,
436 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options,
442 return isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI, Options,
447 /// GetNegatedExpression - If isNegatibleForFree returns true, this function
448 /// returns the newly negated expression.
449 static SDValue GetNegatedExpression(SDValue Op, SelectionDAG &DAG,
450 bool LegalOperations, unsigned Depth = 0) {
451 // fneg is removable even if it has multiple uses.
452 if (Op.getOpcode() == ISD::FNEG) return Op.getOperand(0);
454 // Don't allow anything with multiple uses.
455 assert(Op.hasOneUse() && "Unknown reuse!");
457 assert(Depth <= 6 && "GetNegatedExpression doesn't match isNegatibleForFree");
458 switch (Op.getOpcode()) {
459 default: llvm_unreachable("Unknown code");
460 case ISD::ConstantFP: {
461 APFloat V = cast<ConstantFPSDNode>(Op)->getValueAPF();
463 return DAG.getConstantFP(V, Op.getValueType());
466 // FIXME: determine better conditions for this xform.
467 assert(DAG.getTarget().Options.UnsafeFPMath);
469 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B)
470 if (isNegatibleForFree(Op.getOperand(0), LegalOperations,
471 DAG.getTargetLoweringInfo(),
472 &DAG.getTarget().Options, Depth+1))
473 return DAG.getNode(ISD::FSUB, Op.getDebugLoc(), Op.getValueType(),
474 GetNegatedExpression(Op.getOperand(0), DAG,
475 LegalOperations, Depth+1),
477 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
478 return DAG.getNode(ISD::FSUB, Op.getDebugLoc(), Op.getValueType(),
479 GetNegatedExpression(Op.getOperand(1), DAG,
480 LegalOperations, Depth+1),
483 // We can't turn -(A-B) into B-A when we honor signed zeros.
484 assert(DAG.getTarget().Options.UnsafeFPMath);
486 // fold (fneg (fsub 0, B)) -> B
487 if (ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(Op.getOperand(0)))
488 if (N0CFP->getValueAPF().isZero())
489 return Op.getOperand(1);
491 // fold (fneg (fsub A, B)) -> (fsub B, A)
492 return DAG.getNode(ISD::FSUB, Op.getDebugLoc(), Op.getValueType(),
493 Op.getOperand(1), Op.getOperand(0));
497 assert(!DAG.getTarget().Options.HonorSignDependentRoundingFPMath());
499 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y)
500 if (isNegatibleForFree(Op.getOperand(0), LegalOperations,
501 DAG.getTargetLoweringInfo(),
502 &DAG.getTarget().Options, Depth+1))
503 return DAG.getNode(Op.getOpcode(), Op.getDebugLoc(), Op.getValueType(),
504 GetNegatedExpression(Op.getOperand(0), DAG,
505 LegalOperations, Depth+1),
508 // fold (fneg (fmul X, Y)) -> (fmul X, (fneg Y))
509 return DAG.getNode(Op.getOpcode(), Op.getDebugLoc(), Op.getValueType(),
511 GetNegatedExpression(Op.getOperand(1), DAG,
512 LegalOperations, Depth+1));
516 return DAG.getNode(Op.getOpcode(), Op.getDebugLoc(), Op.getValueType(),
517 GetNegatedExpression(Op.getOperand(0), DAG,
518 LegalOperations, Depth+1));
520 return DAG.getNode(ISD::FP_ROUND, Op.getDebugLoc(), Op.getValueType(),
521 GetNegatedExpression(Op.getOperand(0), DAG,
522 LegalOperations, Depth+1),
528 // isSetCCEquivalent - Return true if this node is a setcc, or is a select_cc
529 // that selects between the values 1 and 0, making it equivalent to a setcc.
530 // Also, set the incoming LHS, RHS, and CC references to the appropriate
531 // nodes based on the type of node we are checking. This simplifies life a
532 // bit for the callers.
533 static bool isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS,
535 if (N.getOpcode() == ISD::SETCC) {
536 LHS = N.getOperand(0);
537 RHS = N.getOperand(1);
538 CC = N.getOperand(2);
541 if (N.getOpcode() == ISD::SELECT_CC &&
542 N.getOperand(2).getOpcode() == ISD::Constant &&
543 N.getOperand(3).getOpcode() == ISD::Constant &&
544 cast<ConstantSDNode>(N.getOperand(2))->getAPIntValue() == 1 &&
545 cast<ConstantSDNode>(N.getOperand(3))->isNullValue()) {
546 LHS = N.getOperand(0);
547 RHS = N.getOperand(1);
548 CC = N.getOperand(4);
554 // isOneUseSetCC - Return true if this is a SetCC-equivalent operation with only
555 // one use. If this is true, it allows the users to invert the operation for
556 // free when it is profitable to do so.
557 static bool isOneUseSetCC(SDValue N) {
559 if (isSetCCEquivalent(N, N0, N1, N2) && N.getNode()->hasOneUse())
564 SDValue DAGCombiner::ReassociateOps(unsigned Opc, DebugLoc DL,
565 SDValue N0, SDValue N1) {
566 EVT VT = N0.getValueType();
567 if (N0.getOpcode() == Opc && isa<ConstantSDNode>(N0.getOperand(1))) {
568 if (isa<ConstantSDNode>(N1)) {
569 // reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2))
571 DAG.FoldConstantArithmetic(Opc, VT,
572 cast<ConstantSDNode>(N0.getOperand(1)),
573 cast<ConstantSDNode>(N1));
574 return DAG.getNode(Opc, DL, VT, N0.getOperand(0), OpNode);
576 if (N0.hasOneUse()) {
577 // reassoc. (op (op x, c1), y) -> (op (op x, y), c1) iff x+c1 has one use
578 SDValue OpNode = DAG.getNode(Opc, N0.getDebugLoc(), VT,
579 N0.getOperand(0), N1);
580 AddToWorkList(OpNode.getNode());
581 return DAG.getNode(Opc, DL, VT, OpNode, N0.getOperand(1));
585 if (N1.getOpcode() == Opc && isa<ConstantSDNode>(N1.getOperand(1))) {
586 if (isa<ConstantSDNode>(N0)) {
587 // reassoc. (op c2, (op x, c1)) -> (op x, (op c1, c2))
589 DAG.FoldConstantArithmetic(Opc, VT,
590 cast<ConstantSDNode>(N1.getOperand(1)),
591 cast<ConstantSDNode>(N0));
592 return DAG.getNode(Opc, DL, VT, N1.getOperand(0), OpNode);
594 if (N1.hasOneUse()) {
595 // reassoc. (op y, (op x, c1)) -> (op (op x, y), c1) iff x+c1 has one use
596 SDValue OpNode = DAG.getNode(Opc, N0.getDebugLoc(), VT,
597 N1.getOperand(0), N0);
598 AddToWorkList(OpNode.getNode());
599 return DAG.getNode(Opc, DL, VT, OpNode, N1.getOperand(1));
606 SDValue DAGCombiner::CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
608 assert(N->getNumValues() == NumTo && "Broken CombineTo call!");
610 DEBUG(dbgs() << "\nReplacing.1 ";
612 dbgs() << "\nWith: ";
613 To[0].getNode()->dump(&DAG);
614 dbgs() << " and " << NumTo-1 << " other values\n";
615 for (unsigned i = 0, e = NumTo; i != e; ++i)
616 assert((!To[i].getNode() ||
617 N->getValueType(i) == To[i].getValueType()) &&
618 "Cannot combine value to value of different type!"));
619 WorkListRemover DeadNodes(*this);
620 DAG.ReplaceAllUsesWith(N, To);
622 // Push the new nodes and any users onto the worklist
623 for (unsigned i = 0, e = NumTo; i != e; ++i) {
624 if (To[i].getNode()) {
625 AddToWorkList(To[i].getNode());
626 AddUsersToWorkList(To[i].getNode());
631 // Finally, if the node is now dead, remove it from the graph. The node
632 // may not be dead if the replacement process recursively simplified to
633 // something else needing this node.
634 if (N->use_empty()) {
635 // Nodes can be reintroduced into the worklist. Make sure we do not
636 // process a node that has been replaced.
637 removeFromWorkList(N);
639 // Finally, since the node is now dead, remove it from the graph.
642 return SDValue(N, 0);
646 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
647 // Replace all uses. If any nodes become isomorphic to other nodes and
648 // are deleted, make sure to remove them from our worklist.
649 WorkListRemover DeadNodes(*this);
650 DAG.ReplaceAllUsesOfValueWith(TLO.Old, TLO.New);
652 // Push the new node and any (possibly new) users onto the worklist.
653 AddToWorkList(TLO.New.getNode());
654 AddUsersToWorkList(TLO.New.getNode());
656 // Finally, if the node is now dead, remove it from the graph. The node
657 // may not be dead if the replacement process recursively simplified to
658 // something else needing this node.
659 if (TLO.Old.getNode()->use_empty()) {
660 removeFromWorkList(TLO.Old.getNode());
662 // If the operands of this node are only used by the node, they will now
663 // be dead. Make sure to visit them first to delete dead nodes early.
664 for (unsigned i = 0, e = TLO.Old.getNode()->getNumOperands(); i != e; ++i)
665 if (TLO.Old.getNode()->getOperand(i).getNode()->hasOneUse())
666 AddToWorkList(TLO.Old.getNode()->getOperand(i).getNode());
668 DAG.DeleteNode(TLO.Old.getNode());
672 /// SimplifyDemandedBits - Check the specified integer node value to see if
673 /// it can be simplified or if things it uses can be simplified by bit
674 /// propagation. If so, return true.
675 bool DAGCombiner::SimplifyDemandedBits(SDValue Op, const APInt &Demanded) {
676 TargetLowering::TargetLoweringOpt TLO(DAG, LegalTypes, LegalOperations);
677 APInt KnownZero, KnownOne;
678 if (!TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
682 AddToWorkList(Op.getNode());
684 // Replace the old value with the new one.
686 DEBUG(dbgs() << "\nReplacing.2 ";
687 TLO.Old.getNode()->dump(&DAG);
688 dbgs() << "\nWith: ";
689 TLO.New.getNode()->dump(&DAG);
692 CommitTargetLoweringOpt(TLO);
696 void DAGCombiner::ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad) {
697 DebugLoc dl = Load->getDebugLoc();
698 EVT VT = Load->getValueType(0);
699 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, VT, SDValue(ExtLoad, 0));
701 DEBUG(dbgs() << "\nReplacing.9 ";
703 dbgs() << "\nWith: ";
704 Trunc.getNode()->dump(&DAG);
706 WorkListRemover DeadNodes(*this);
707 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 0), Trunc);
708 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), SDValue(ExtLoad, 1));
709 removeFromWorkList(Load);
710 DAG.DeleteNode(Load);
711 AddToWorkList(Trunc.getNode());
714 SDValue DAGCombiner::PromoteOperand(SDValue Op, EVT PVT, bool &Replace) {
716 DebugLoc dl = Op.getDebugLoc();
717 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Op)) {
718 EVT MemVT = LD->getMemoryVT();
719 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
720 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT) ? ISD::ZEXTLOAD
722 : LD->getExtensionType();
724 return DAG.getExtLoad(ExtType, dl, PVT,
725 LD->getChain(), LD->getBasePtr(),
726 LD->getPointerInfo(),
727 MemVT, LD->isVolatile(),
728 LD->isNonTemporal(), LD->getAlignment());
731 unsigned Opc = Op.getOpcode();
734 case ISD::AssertSext:
735 return DAG.getNode(ISD::AssertSext, dl, PVT,
736 SExtPromoteOperand(Op.getOperand(0), PVT),
738 case ISD::AssertZext:
739 return DAG.getNode(ISD::AssertZext, dl, PVT,
740 ZExtPromoteOperand(Op.getOperand(0), PVT),
742 case ISD::Constant: {
744 Op.getValueType().isByteSized() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
745 return DAG.getNode(ExtOpc, dl, PVT, Op);
749 if (!TLI.isOperationLegal(ISD::ANY_EXTEND, PVT))
751 return DAG.getNode(ISD::ANY_EXTEND, dl, PVT, Op);
754 SDValue DAGCombiner::SExtPromoteOperand(SDValue Op, EVT PVT) {
755 if (!TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, PVT))
757 EVT OldVT = Op.getValueType();
758 DebugLoc dl = Op.getDebugLoc();
759 bool Replace = false;
760 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
761 if (NewOp.getNode() == 0)
763 AddToWorkList(NewOp.getNode());
766 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
767 return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, NewOp.getValueType(), NewOp,
768 DAG.getValueType(OldVT));
771 SDValue DAGCombiner::ZExtPromoteOperand(SDValue Op, EVT PVT) {
772 EVT OldVT = Op.getValueType();
773 DebugLoc dl = Op.getDebugLoc();
774 bool Replace = false;
775 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
776 if (NewOp.getNode() == 0)
778 AddToWorkList(NewOp.getNode());
781 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
782 return DAG.getZeroExtendInReg(NewOp, dl, OldVT);
785 /// PromoteIntBinOp - Promote the specified integer binary operation if the
786 /// target indicates it is beneficial. e.g. On x86, it's usually better to
787 /// promote i16 operations to i32 since i16 instructions are longer.
788 SDValue DAGCombiner::PromoteIntBinOp(SDValue Op) {
789 if (!LegalOperations)
792 EVT VT = Op.getValueType();
793 if (VT.isVector() || !VT.isInteger())
796 // If operation type is 'undesirable', e.g. i16 on x86, consider
798 unsigned Opc = Op.getOpcode();
799 if (TLI.isTypeDesirableForOp(Opc, VT))
803 // Consult target whether it is a good idea to promote this operation and
804 // what's the right type to promote it to.
805 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
806 assert(PVT != VT && "Don't know what type to promote to!");
808 bool Replace0 = false;
809 SDValue N0 = Op.getOperand(0);
810 SDValue NN0 = PromoteOperand(N0, PVT, Replace0);
811 if (NN0.getNode() == 0)
814 bool Replace1 = false;
815 SDValue N1 = Op.getOperand(1);
820 NN1 = PromoteOperand(N1, PVT, Replace1);
821 if (NN1.getNode() == 0)
825 AddToWorkList(NN0.getNode());
827 AddToWorkList(NN1.getNode());
830 ReplaceLoadWithPromotedLoad(N0.getNode(), NN0.getNode());
832 ReplaceLoadWithPromotedLoad(N1.getNode(), NN1.getNode());
834 DEBUG(dbgs() << "\nPromoting ";
835 Op.getNode()->dump(&DAG));
836 DebugLoc dl = Op.getDebugLoc();
837 return DAG.getNode(ISD::TRUNCATE, dl, VT,
838 DAG.getNode(Opc, dl, PVT, NN0, NN1));
843 /// PromoteIntShiftOp - Promote the specified integer shift operation if the
844 /// target indicates it is beneficial. e.g. On x86, it's usually better to
845 /// promote i16 operations to i32 since i16 instructions are longer.
846 SDValue DAGCombiner::PromoteIntShiftOp(SDValue Op) {
847 if (!LegalOperations)
850 EVT VT = Op.getValueType();
851 if (VT.isVector() || !VT.isInteger())
854 // If operation type is 'undesirable', e.g. i16 on x86, consider
856 unsigned Opc = Op.getOpcode();
857 if (TLI.isTypeDesirableForOp(Opc, VT))
861 // Consult target whether it is a good idea to promote this operation and
862 // what's the right type to promote it to.
863 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
864 assert(PVT != VT && "Don't know what type to promote to!");
866 bool Replace = false;
867 SDValue N0 = Op.getOperand(0);
869 N0 = SExtPromoteOperand(Op.getOperand(0), PVT);
870 else if (Opc == ISD::SRL)
871 N0 = ZExtPromoteOperand(Op.getOperand(0), PVT);
873 N0 = PromoteOperand(N0, PVT, Replace);
874 if (N0.getNode() == 0)
877 AddToWorkList(N0.getNode());
879 ReplaceLoadWithPromotedLoad(Op.getOperand(0).getNode(), N0.getNode());
881 DEBUG(dbgs() << "\nPromoting ";
882 Op.getNode()->dump(&DAG));
883 DebugLoc dl = Op.getDebugLoc();
884 return DAG.getNode(ISD::TRUNCATE, dl, VT,
885 DAG.getNode(Opc, dl, PVT, N0, Op.getOperand(1)));
890 SDValue DAGCombiner::PromoteExtend(SDValue Op) {
891 if (!LegalOperations)
894 EVT VT = Op.getValueType();
895 if (VT.isVector() || !VT.isInteger())
898 // If operation type is 'undesirable', e.g. i16 on x86, consider
900 unsigned Opc = Op.getOpcode();
901 if (TLI.isTypeDesirableForOp(Opc, VT))
905 // Consult target whether it is a good idea to promote this operation and
906 // what's the right type to promote it to.
907 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
908 assert(PVT != VT && "Don't know what type to promote to!");
909 // fold (aext (aext x)) -> (aext x)
910 // fold (aext (zext x)) -> (zext x)
911 // fold (aext (sext x)) -> (sext x)
912 DEBUG(dbgs() << "\nPromoting ";
913 Op.getNode()->dump(&DAG));
914 return DAG.getNode(Op.getOpcode(), Op.getDebugLoc(), VT, Op.getOperand(0));
919 bool DAGCombiner::PromoteLoad(SDValue Op) {
920 if (!LegalOperations)
923 EVT VT = Op.getValueType();
924 if (VT.isVector() || !VT.isInteger())
927 // If operation type is 'undesirable', e.g. i16 on x86, consider
929 unsigned Opc = Op.getOpcode();
930 if (TLI.isTypeDesirableForOp(Opc, VT))
934 // Consult target whether it is a good idea to promote this operation and
935 // what's the right type to promote it to.
936 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
937 assert(PVT != VT && "Don't know what type to promote to!");
939 DebugLoc dl = Op.getDebugLoc();
940 SDNode *N = Op.getNode();
941 LoadSDNode *LD = cast<LoadSDNode>(N);
942 EVT MemVT = LD->getMemoryVT();
943 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
944 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT) ? ISD::ZEXTLOAD
946 : LD->getExtensionType();
947 SDValue NewLD = DAG.getExtLoad(ExtType, dl, PVT,
948 LD->getChain(), LD->getBasePtr(),
949 LD->getPointerInfo(),
950 MemVT, LD->isVolatile(),
951 LD->isNonTemporal(), LD->getAlignment());
952 SDValue Result = DAG.getNode(ISD::TRUNCATE, dl, VT, NewLD);
954 DEBUG(dbgs() << "\nPromoting ";
957 Result.getNode()->dump(&DAG);
959 WorkListRemover DeadNodes(*this);
960 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result);
961 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), NewLD.getValue(1));
962 removeFromWorkList(N);
964 AddToWorkList(Result.getNode());
971 //===----------------------------------------------------------------------===//
972 // Main DAG Combiner implementation
973 //===----------------------------------------------------------------------===//
975 void DAGCombiner::Run(CombineLevel AtLevel) {
976 // set the instance variables, so that the various visit routines may use it.
978 LegalOperations = Level >= AfterLegalizeVectorOps;
979 LegalTypes = Level >= AfterLegalizeTypes;
981 // Add all the dag nodes to the worklist.
982 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
983 E = DAG.allnodes_end(); I != E; ++I)
986 // Create a dummy node (which is not added to allnodes), that adds a reference
987 // to the root node, preventing it from being deleted, and tracking any
988 // changes of the root.
989 HandleSDNode Dummy(DAG.getRoot());
991 // The root of the dag may dangle to deleted nodes until the dag combiner is
992 // done. Set it to null to avoid confusion.
993 DAG.setRoot(SDValue());
995 // while the worklist isn't empty, find a node and
996 // try and combine it.
997 while (!WorkListContents.empty()) {
999 // The WorkListOrder holds the SDNodes in order, but it may contain duplicates.
1000 // In order to avoid a linear scan, we use a set (O(log N)) to hold what the
1001 // worklist *should* contain, and check the node we want to visit is should
1002 // actually be visited.
1004 N = WorkListOrder.pop_back_val();
1005 } while (!WorkListContents.erase(N));
1007 // If N has no uses, it is dead. Make sure to revisit all N's operands once
1008 // N is deleted from the DAG, since they too may now be dead or may have a
1009 // reduced number of uses, allowing other xforms.
1010 if (N->use_empty() && N != &Dummy) {
1011 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1012 AddToWorkList(N->getOperand(i).getNode());
1018 SDValue RV = combine(N);
1020 if (RV.getNode() == 0)
1025 // If we get back the same node we passed in, rather than a new node or
1026 // zero, we know that the node must have defined multiple values and
1027 // CombineTo was used. Since CombineTo takes care of the worklist
1028 // mechanics for us, we have no work to do in this case.
1029 if (RV.getNode() == N)
1032 assert(N->getOpcode() != ISD::DELETED_NODE &&
1033 RV.getNode()->getOpcode() != ISD::DELETED_NODE &&
1034 "Node was deleted but visit returned new node!");
1036 DEBUG(dbgs() << "\nReplacing.3 ";
1038 dbgs() << "\nWith: ";
1039 RV.getNode()->dump(&DAG);
1042 // Transfer debug value.
1043 DAG.TransferDbgValues(SDValue(N, 0), RV);
1044 WorkListRemover DeadNodes(*this);
1045 if (N->getNumValues() == RV.getNode()->getNumValues())
1046 DAG.ReplaceAllUsesWith(N, RV.getNode());
1048 assert(N->getValueType(0) == RV.getValueType() &&
1049 N->getNumValues() == 1 && "Type mismatch");
1051 DAG.ReplaceAllUsesWith(N, &OpV);
1054 // Push the new node and any users onto the worklist
1055 AddToWorkList(RV.getNode());
1056 AddUsersToWorkList(RV.getNode());
1058 // Add any uses of the old node to the worklist in case this node is the
1059 // last one that uses them. They may become dead after this node is
1061 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1062 AddToWorkList(N->getOperand(i).getNode());
1064 // Finally, if the node is now dead, remove it from the graph. The node
1065 // may not be dead if the replacement process recursively simplified to
1066 // something else needing this node.
1067 if (N->use_empty()) {
1068 // Nodes can be reintroduced into the worklist. Make sure we do not
1069 // process a node that has been replaced.
1070 removeFromWorkList(N);
1072 // Finally, since the node is now dead, remove it from the graph.
1077 // If the root changed (e.g. it was a dead load, update the root).
1078 DAG.setRoot(Dummy.getValue());
1079 DAG.RemoveDeadNodes();
1082 SDValue DAGCombiner::visit(SDNode *N) {
1083 switch (N->getOpcode()) {
1085 case ISD::TokenFactor: return visitTokenFactor(N);
1086 case ISD::MERGE_VALUES: return visitMERGE_VALUES(N);
1087 case ISD::ADD: return visitADD(N);
1088 case ISD::SUB: return visitSUB(N);
1089 case ISD::ADDC: return visitADDC(N);
1090 case ISD::SUBC: return visitSUBC(N);
1091 case ISD::ADDE: return visitADDE(N);
1092 case ISD::SUBE: return visitSUBE(N);
1093 case ISD::MUL: return visitMUL(N);
1094 case ISD::SDIV: return visitSDIV(N);
1095 case ISD::UDIV: return visitUDIV(N);
1096 case ISD::SREM: return visitSREM(N);
1097 case ISD::UREM: return visitUREM(N);
1098 case ISD::MULHU: return visitMULHU(N);
1099 case ISD::MULHS: return visitMULHS(N);
1100 case ISD::SMUL_LOHI: return visitSMUL_LOHI(N);
1101 case ISD::UMUL_LOHI: return visitUMUL_LOHI(N);
1102 case ISD::SMULO: return visitSMULO(N);
1103 case ISD::UMULO: return visitUMULO(N);
1104 case ISD::SDIVREM: return visitSDIVREM(N);
1105 case ISD::UDIVREM: return visitUDIVREM(N);
1106 case ISD::AND: return visitAND(N);
1107 case ISD::OR: return visitOR(N);
1108 case ISD::XOR: return visitXOR(N);
1109 case ISD::SHL: return visitSHL(N);
1110 case ISD::SRA: return visitSRA(N);
1111 case ISD::SRL: return visitSRL(N);
1112 case ISD::CTLZ: return visitCTLZ(N);
1113 case ISD::CTLZ_ZERO_UNDEF: return visitCTLZ_ZERO_UNDEF(N);
1114 case ISD::CTTZ: return visitCTTZ(N);
1115 case ISD::CTTZ_ZERO_UNDEF: return visitCTTZ_ZERO_UNDEF(N);
1116 case ISD::CTPOP: return visitCTPOP(N);
1117 case ISD::SELECT: return visitSELECT(N);
1118 case ISD::SELECT_CC: return visitSELECT_CC(N);
1119 case ISD::SETCC: return visitSETCC(N);
1120 case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N);
1121 case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N);
1122 case ISD::ANY_EXTEND: return visitANY_EXTEND(N);
1123 case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N);
1124 case ISD::TRUNCATE: return visitTRUNCATE(N);
1125 case ISD::BITCAST: return visitBITCAST(N);
1126 case ISD::BUILD_PAIR: return visitBUILD_PAIR(N);
1127 case ISD::FADD: return visitFADD(N);
1128 case ISD::FSUB: return visitFSUB(N);
1129 case ISD::FMUL: return visitFMUL(N);
1130 case ISD::FMA: return visitFMA(N);
1131 case ISD::FDIV: return visitFDIV(N);
1132 case ISD::FREM: return visitFREM(N);
1133 case ISD::FCOPYSIGN: return visitFCOPYSIGN(N);
1134 case ISD::SINT_TO_FP: return visitSINT_TO_FP(N);
1135 case ISD::UINT_TO_FP: return visitUINT_TO_FP(N);
1136 case ISD::FP_TO_SINT: return visitFP_TO_SINT(N);
1137 case ISD::FP_TO_UINT: return visitFP_TO_UINT(N);
1138 case ISD::FP_ROUND: return visitFP_ROUND(N);
1139 case ISD::FP_ROUND_INREG: return visitFP_ROUND_INREG(N);
1140 case ISD::FP_EXTEND: return visitFP_EXTEND(N);
1141 case ISD::FNEG: return visitFNEG(N);
1142 case ISD::FABS: return visitFABS(N);
1143 case ISD::BRCOND: return visitBRCOND(N);
1144 case ISD::BR_CC: return visitBR_CC(N);
1145 case ISD::LOAD: return visitLOAD(N);
1146 case ISD::STORE: return visitSTORE(N);
1147 case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N);
1148 case ISD::EXTRACT_VECTOR_ELT: return visitEXTRACT_VECTOR_ELT(N);
1149 case ISD::BUILD_VECTOR: return visitBUILD_VECTOR(N);
1150 case ISD::CONCAT_VECTORS: return visitCONCAT_VECTORS(N);
1151 case ISD::EXTRACT_SUBVECTOR: return visitEXTRACT_SUBVECTOR(N);
1152 case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N);
1153 case ISD::MEMBARRIER: return visitMEMBARRIER(N);
1158 SDValue DAGCombiner::combine(SDNode *N) {
1159 SDValue RV = visit(N);
1161 // If nothing happened, try a target-specific DAG combine.
1162 if (RV.getNode() == 0) {
1163 assert(N->getOpcode() != ISD::DELETED_NODE &&
1164 "Node was deleted but visit returned NULL!");
1166 if (N->getOpcode() >= ISD::BUILTIN_OP_END ||
1167 TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode())) {
1169 // Expose the DAG combiner to the target combiner impls.
1170 TargetLowering::DAGCombinerInfo
1171 DagCombineInfo(DAG, !LegalTypes, !LegalOperations, false, this);
1173 RV = TLI.PerformDAGCombine(N, DagCombineInfo);
1177 // If nothing happened still, try promoting the operation.
1178 if (RV.getNode() == 0) {
1179 switch (N->getOpcode()) {
1187 RV = PromoteIntBinOp(SDValue(N, 0));
1192 RV = PromoteIntShiftOp(SDValue(N, 0));
1194 case ISD::SIGN_EXTEND:
1195 case ISD::ZERO_EXTEND:
1196 case ISD::ANY_EXTEND:
1197 RV = PromoteExtend(SDValue(N, 0));
1200 if (PromoteLoad(SDValue(N, 0)))
1206 // If N is a commutative binary node, try commuting it to enable more
1208 if (RV.getNode() == 0 &&
1209 SelectionDAG::isCommutativeBinOp(N->getOpcode()) &&
1210 N->getNumValues() == 1) {
1211 SDValue N0 = N->getOperand(0);
1212 SDValue N1 = N->getOperand(1);
1214 // Constant operands are canonicalized to RHS.
1215 if (isa<ConstantSDNode>(N0) || !isa<ConstantSDNode>(N1)) {
1216 SDValue Ops[] = { N1, N0 };
1217 SDNode *CSENode = DAG.getNodeIfExists(N->getOpcode(), N->getVTList(),
1220 return SDValue(CSENode, 0);
1227 /// getInputChainForNode - Given a node, return its input chain if it has one,
1228 /// otherwise return a null sd operand.
1229 static SDValue getInputChainForNode(SDNode *N) {
1230 if (unsigned NumOps = N->getNumOperands()) {
1231 if (N->getOperand(0).getValueType() == MVT::Other)
1232 return N->getOperand(0);
1233 else if (N->getOperand(NumOps-1).getValueType() == MVT::Other)
1234 return N->getOperand(NumOps-1);
1235 for (unsigned i = 1; i < NumOps-1; ++i)
1236 if (N->getOperand(i).getValueType() == MVT::Other)
1237 return N->getOperand(i);
1242 SDValue DAGCombiner::visitTokenFactor(SDNode *N) {
1243 // If N has two operands, where one has an input chain equal to the other,
1244 // the 'other' chain is redundant.
1245 if (N->getNumOperands() == 2) {
1246 if (getInputChainForNode(N->getOperand(0).getNode()) == N->getOperand(1))
1247 return N->getOperand(0);
1248 if (getInputChainForNode(N->getOperand(1).getNode()) == N->getOperand(0))
1249 return N->getOperand(1);
1252 SmallVector<SDNode *, 8> TFs; // List of token factors to visit.
1253 SmallVector<SDValue, 8> Ops; // Ops for replacing token factor.
1254 SmallPtrSet<SDNode*, 16> SeenOps;
1255 bool Changed = false; // If we should replace this token factor.
1257 // Start out with this token factor.
1260 // Iterate through token factors. The TFs grows when new token factors are
1262 for (unsigned i = 0; i < TFs.size(); ++i) {
1263 SDNode *TF = TFs[i];
1265 // Check each of the operands.
1266 for (unsigned i = 0, ie = TF->getNumOperands(); i != ie; ++i) {
1267 SDValue Op = TF->getOperand(i);
1269 switch (Op.getOpcode()) {
1270 case ISD::EntryToken:
1271 // Entry tokens don't need to be added to the list. They are
1276 case ISD::TokenFactor:
1277 if (Op.hasOneUse() &&
1278 std::find(TFs.begin(), TFs.end(), Op.getNode()) == TFs.end()) {
1279 // Queue up for processing.
1280 TFs.push_back(Op.getNode());
1281 // Clean up in case the token factor is removed.
1282 AddToWorkList(Op.getNode());
1289 // Only add if it isn't already in the list.
1290 if (SeenOps.insert(Op.getNode()))
1301 // If we've change things around then replace token factor.
1304 // The entry token is the only possible outcome.
1305 Result = DAG.getEntryNode();
1307 // New and improved token factor.
1308 Result = DAG.getNode(ISD::TokenFactor, N->getDebugLoc(),
1309 MVT::Other, &Ops[0], Ops.size());
1312 // Don't add users to work list.
1313 return CombineTo(N, Result, false);
1319 /// MERGE_VALUES can always be eliminated.
1320 SDValue DAGCombiner::visitMERGE_VALUES(SDNode *N) {
1321 WorkListRemover DeadNodes(*this);
1322 // Replacing results may cause a different MERGE_VALUES to suddenly
1323 // be CSE'd with N, and carry its uses with it. Iterate until no
1324 // uses remain, to ensure that the node can be safely deleted.
1326 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1327 DAG.ReplaceAllUsesOfValueWith(SDValue(N, i), N->getOperand(i));
1328 } while (!N->use_empty());
1329 removeFromWorkList(N);
1331 return SDValue(N, 0); // Return N so it doesn't get rechecked!
1335 SDValue combineShlAddConstant(DebugLoc DL, SDValue N0, SDValue N1,
1336 SelectionDAG &DAG) {
1337 EVT VT = N0.getValueType();
1338 SDValue N00 = N0.getOperand(0);
1339 SDValue N01 = N0.getOperand(1);
1340 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N01);
1342 if (N01C && N00.getOpcode() == ISD::ADD && N00.getNode()->hasOneUse() &&
1343 isa<ConstantSDNode>(N00.getOperand(1))) {
1344 // fold (add (shl (add x, c1), c2), ) -> (add (add (shl x, c2), c1<<c2), )
1345 N0 = DAG.getNode(ISD::ADD, N0.getDebugLoc(), VT,
1346 DAG.getNode(ISD::SHL, N00.getDebugLoc(), VT,
1347 N00.getOperand(0), N01),
1348 DAG.getNode(ISD::SHL, N01.getDebugLoc(), VT,
1349 N00.getOperand(1), N01));
1350 return DAG.getNode(ISD::ADD, DL, VT, N0, N1);
1356 SDValue DAGCombiner::visitADD(SDNode *N) {
1357 SDValue N0 = N->getOperand(0);
1358 SDValue N1 = N->getOperand(1);
1359 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1360 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1361 EVT VT = N0.getValueType();
1364 if (VT.isVector()) {
1365 SDValue FoldedVOp = SimplifyVBinOp(N);
1366 if (FoldedVOp.getNode()) return FoldedVOp;
1369 // fold (add x, undef) -> undef
1370 if (N0.getOpcode() == ISD::UNDEF)
1372 if (N1.getOpcode() == ISD::UNDEF)
1374 // fold (add c1, c2) -> c1+c2
1376 return DAG.FoldConstantArithmetic(ISD::ADD, VT, N0C, N1C);
1377 // canonicalize constant to RHS
1379 return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N1, N0);
1380 // fold (add x, 0) -> x
1381 if (N1C && N1C->isNullValue())
1383 // fold (add Sym, c) -> Sym+c
1384 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1385 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA) && N1C &&
1386 GA->getOpcode() == ISD::GlobalAddress)
1387 return DAG.getGlobalAddress(GA->getGlobal(), N1C->getDebugLoc(), VT,
1389 (uint64_t)N1C->getSExtValue());
1390 // fold ((c1-A)+c2) -> (c1+c2)-A
1391 if (N1C && N0.getOpcode() == ISD::SUB)
1392 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getOperand(0)))
1393 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1394 DAG.getConstant(N1C->getAPIntValue()+
1395 N0C->getAPIntValue(), VT),
1398 SDValue RADD = ReassociateOps(ISD::ADD, N->getDebugLoc(), N0, N1);
1399 if (RADD.getNode() != 0)
1401 // fold ((0-A) + B) -> B-A
1402 if (N0.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N0.getOperand(0)) &&
1403 cast<ConstantSDNode>(N0.getOperand(0))->isNullValue())
1404 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N1, N0.getOperand(1));
1405 // fold (A + (0-B)) -> A-B
1406 if (N1.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N1.getOperand(0)) &&
1407 cast<ConstantSDNode>(N1.getOperand(0))->isNullValue())
1408 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N0, N1.getOperand(1));
1409 // fold (A+(B-A)) -> B
1410 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1))
1411 return N1.getOperand(0);
1412 // fold ((B-A)+A) -> B
1413 if (N0.getOpcode() == ISD::SUB && N1 == N0.getOperand(1))
1414 return N0.getOperand(0);
1415 // fold (A+(B-(A+C))) to (B-C)
1416 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1417 N0 == N1.getOperand(1).getOperand(0))
1418 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N1.getOperand(0),
1419 N1.getOperand(1).getOperand(1));
1420 // fold (A+(B-(C+A))) to (B-C)
1421 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1422 N0 == N1.getOperand(1).getOperand(1))
1423 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N1.getOperand(0),
1424 N1.getOperand(1).getOperand(0));
1425 // fold (A+((B-A)+or-C)) to (B+or-C)
1426 if ((N1.getOpcode() == ISD::SUB || N1.getOpcode() == ISD::ADD) &&
1427 N1.getOperand(0).getOpcode() == ISD::SUB &&
1428 N0 == N1.getOperand(0).getOperand(1))
1429 return DAG.getNode(N1.getOpcode(), N->getDebugLoc(), VT,
1430 N1.getOperand(0).getOperand(0), N1.getOperand(1));
1432 // fold (A-B)+(C-D) to (A+C)-(B+D) when A or C is constant
1433 if (N0.getOpcode() == ISD::SUB && N1.getOpcode() == ISD::SUB) {
1434 SDValue N00 = N0.getOperand(0);
1435 SDValue N01 = N0.getOperand(1);
1436 SDValue N10 = N1.getOperand(0);
1437 SDValue N11 = N1.getOperand(1);
1439 if (isa<ConstantSDNode>(N00) || isa<ConstantSDNode>(N10))
1440 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1441 DAG.getNode(ISD::ADD, N0.getDebugLoc(), VT, N00, N10),
1442 DAG.getNode(ISD::ADD, N1.getDebugLoc(), VT, N01, N11));
1445 if (!VT.isVector() && SimplifyDemandedBits(SDValue(N, 0)))
1446 return SDValue(N, 0);
1448 // fold (a+b) -> (a|b) iff a and b share no bits.
1449 if (VT.isInteger() && !VT.isVector()) {
1450 APInt LHSZero, LHSOne;
1451 APInt RHSZero, RHSOne;
1452 DAG.ComputeMaskedBits(N0, LHSZero, LHSOne);
1454 if (LHSZero.getBoolValue()) {
1455 DAG.ComputeMaskedBits(N1, RHSZero, RHSOne);
1457 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1458 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1459 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero)
1460 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N0, N1);
1464 // fold (add (shl (add x, c1), c2), ) -> (add (add (shl x, c2), c1<<c2), )
1465 if (N0.getOpcode() == ISD::SHL && N0.getNode()->hasOneUse()) {
1466 SDValue Result = combineShlAddConstant(N->getDebugLoc(), N0, N1, DAG);
1467 if (Result.getNode()) return Result;
1469 if (N1.getOpcode() == ISD::SHL && N1.getNode()->hasOneUse()) {
1470 SDValue Result = combineShlAddConstant(N->getDebugLoc(), N1, N0, DAG);
1471 if (Result.getNode()) return Result;
1474 // fold (add x, shl(0 - y, n)) -> sub(x, shl(y, n))
1475 if (N1.getOpcode() == ISD::SHL &&
1476 N1.getOperand(0).getOpcode() == ISD::SUB)
1477 if (ConstantSDNode *C =
1478 dyn_cast<ConstantSDNode>(N1.getOperand(0).getOperand(0)))
1479 if (C->getAPIntValue() == 0)
1480 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N0,
1481 DAG.getNode(ISD::SHL, N->getDebugLoc(), VT,
1482 N1.getOperand(0).getOperand(1),
1484 if (N0.getOpcode() == ISD::SHL &&
1485 N0.getOperand(0).getOpcode() == ISD::SUB)
1486 if (ConstantSDNode *C =
1487 dyn_cast<ConstantSDNode>(N0.getOperand(0).getOperand(0)))
1488 if (C->getAPIntValue() == 0)
1489 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N1,
1490 DAG.getNode(ISD::SHL, N->getDebugLoc(), VT,
1491 N0.getOperand(0).getOperand(1),
1494 if (N1.getOpcode() == ISD::AND) {
1495 SDValue AndOp0 = N1.getOperand(0);
1496 ConstantSDNode *AndOp1 = dyn_cast<ConstantSDNode>(N1->getOperand(1));
1497 unsigned NumSignBits = DAG.ComputeNumSignBits(AndOp0);
1498 unsigned DestBits = VT.getScalarType().getSizeInBits();
1500 // (add z, (and (sbbl x, x), 1)) -> (sub z, (sbbl x, x))
1501 // and similar xforms where the inner op is either ~0 or 0.
1502 if (NumSignBits == DestBits && AndOp1 && AndOp1->isOne()) {
1503 DebugLoc DL = N->getDebugLoc();
1504 return DAG.getNode(ISD::SUB, DL, VT, N->getOperand(0), AndOp0);
1508 // add (sext i1), X -> sub X, (zext i1)
1509 if (N0.getOpcode() == ISD::SIGN_EXTEND &&
1510 N0.getOperand(0).getValueType() == MVT::i1 &&
1511 !TLI.isOperationLegal(ISD::SIGN_EXTEND, MVT::i1)) {
1512 DebugLoc DL = N->getDebugLoc();
1513 SDValue ZExt = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0));
1514 return DAG.getNode(ISD::SUB, DL, VT, N1, ZExt);
1520 SDValue DAGCombiner::visitADDC(SDNode *N) {
1521 SDValue N0 = N->getOperand(0);
1522 SDValue N1 = N->getOperand(1);
1523 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1524 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1525 EVT VT = N0.getValueType();
1527 // If the flag result is dead, turn this into an ADD.
1528 if (!N->hasAnyUseOfValue(1))
1529 return CombineTo(N, DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N0, N1),
1530 DAG.getNode(ISD::CARRY_FALSE,
1531 N->getDebugLoc(), MVT::Glue));
1533 // canonicalize constant to RHS.
1535 return DAG.getNode(ISD::ADDC, N->getDebugLoc(), N->getVTList(), N1, N0);
1537 // fold (addc x, 0) -> x + no carry out
1538 if (N1C && N1C->isNullValue())
1539 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE,
1540 N->getDebugLoc(), MVT::Glue));
1542 // fold (addc a, b) -> (or a, b), CARRY_FALSE iff a and b share no bits.
1543 APInt LHSZero, LHSOne;
1544 APInt RHSZero, RHSOne;
1545 DAG.ComputeMaskedBits(N0, LHSZero, LHSOne);
1547 if (LHSZero.getBoolValue()) {
1548 DAG.ComputeMaskedBits(N1, RHSZero, RHSOne);
1550 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1551 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1552 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero)
1553 return CombineTo(N, DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N0, N1),
1554 DAG.getNode(ISD::CARRY_FALSE,
1555 N->getDebugLoc(), MVT::Glue));
1561 SDValue DAGCombiner::visitADDE(SDNode *N) {
1562 SDValue N0 = N->getOperand(0);
1563 SDValue N1 = N->getOperand(1);
1564 SDValue CarryIn = N->getOperand(2);
1565 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1566 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1568 // canonicalize constant to RHS
1570 return DAG.getNode(ISD::ADDE, N->getDebugLoc(), N->getVTList(),
1573 // fold (adde x, y, false) -> (addc x, y)
1574 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1575 return DAG.getNode(ISD::ADDC, N->getDebugLoc(), N->getVTList(), N0, N1);
1580 // Since it may not be valid to emit a fold to zero for vector initializers
1581 // check if we can before folding.
1582 static SDValue tryFoldToZero(DebugLoc DL, const TargetLowering &TLI, EVT VT,
1583 SelectionDAG &DAG, bool LegalOperations) {
1584 if (!VT.isVector()) {
1585 return DAG.getConstant(0, VT);
1587 if (!LegalOperations || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT)) {
1588 // Produce a vector of zeros.
1589 SDValue El = DAG.getConstant(0, VT.getVectorElementType());
1590 std::vector<SDValue> Ops(VT.getVectorNumElements(), El);
1591 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT,
1592 &Ops[0], Ops.size());
1597 SDValue DAGCombiner::visitSUB(SDNode *N) {
1598 SDValue N0 = N->getOperand(0);
1599 SDValue N1 = N->getOperand(1);
1600 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1601 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1602 ConstantSDNode *N1C1 = N1.getOpcode() != ISD::ADD ? 0 :
1603 dyn_cast<ConstantSDNode>(N1.getOperand(1).getNode());
1604 EVT VT = N0.getValueType();
1607 if (VT.isVector()) {
1608 SDValue FoldedVOp = SimplifyVBinOp(N);
1609 if (FoldedVOp.getNode()) return FoldedVOp;
1612 // fold (sub x, x) -> 0
1613 // FIXME: Refactor this and xor and other similar operations together.
1615 return tryFoldToZero(N->getDebugLoc(), TLI, VT, DAG, LegalOperations);
1616 // fold (sub c1, c2) -> c1-c2
1618 return DAG.FoldConstantArithmetic(ISD::SUB, VT, N0C, N1C);
1619 // fold (sub x, c) -> (add x, -c)
1621 return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N0,
1622 DAG.getConstant(-N1C->getAPIntValue(), VT));
1623 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1)
1624 if (N0C && N0C->isAllOnesValue())
1625 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N1, N0);
1626 // fold A-(A-B) -> B
1627 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(0))
1628 return N1.getOperand(1);
1629 // fold (A+B)-A -> B
1630 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1)
1631 return N0.getOperand(1);
1632 // fold (A+B)-B -> A
1633 if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1)
1634 return N0.getOperand(0);
1635 // fold C2-(A+C1) -> (C2-C1)-A
1636 if (N1.getOpcode() == ISD::ADD && N0C && N1C1) {
1637 SDValue NewC = DAG.getConstant((N0C->getAPIntValue() - N1C1->getAPIntValue()), VT);
1638 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, NewC,
1641 // fold ((A+(B+or-C))-B) -> A+or-C
1642 if (N0.getOpcode() == ISD::ADD &&
1643 (N0.getOperand(1).getOpcode() == ISD::SUB ||
1644 N0.getOperand(1).getOpcode() == ISD::ADD) &&
1645 N0.getOperand(1).getOperand(0) == N1)
1646 return DAG.getNode(N0.getOperand(1).getOpcode(), N->getDebugLoc(), VT,
1647 N0.getOperand(0), N0.getOperand(1).getOperand(1));
1648 // fold ((A+(C+B))-B) -> A+C
1649 if (N0.getOpcode() == ISD::ADD &&
1650 N0.getOperand(1).getOpcode() == ISD::ADD &&
1651 N0.getOperand(1).getOperand(1) == N1)
1652 return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT,
1653 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1654 // fold ((A-(B-C))-C) -> A-B
1655 if (N0.getOpcode() == ISD::SUB &&
1656 N0.getOperand(1).getOpcode() == ISD::SUB &&
1657 N0.getOperand(1).getOperand(1) == N1)
1658 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1659 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1661 // If either operand of a sub is undef, the result is undef
1662 if (N0.getOpcode() == ISD::UNDEF)
1664 if (N1.getOpcode() == ISD::UNDEF)
1667 // If the relocation model supports it, consider symbol offsets.
1668 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1669 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA)) {
1670 // fold (sub Sym, c) -> Sym-c
1671 if (N1C && GA->getOpcode() == ISD::GlobalAddress)
1672 return DAG.getGlobalAddress(GA->getGlobal(), N1C->getDebugLoc(), VT,
1674 (uint64_t)N1C->getSExtValue());
1675 // fold (sub Sym+c1, Sym+c2) -> c1-c2
1676 if (GlobalAddressSDNode *GB = dyn_cast<GlobalAddressSDNode>(N1))
1677 if (GA->getGlobal() == GB->getGlobal())
1678 return DAG.getConstant((uint64_t)GA->getOffset() - GB->getOffset(),
1685 SDValue DAGCombiner::visitSUBC(SDNode *N) {
1686 SDValue N0 = N->getOperand(0);
1687 SDValue N1 = N->getOperand(1);
1688 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1689 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1690 EVT VT = N0.getValueType();
1692 // If the flag result is dead, turn this into an SUB.
1693 if (!N->hasAnyUseOfValue(1))
1694 return CombineTo(N, DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N0, N1),
1695 DAG.getNode(ISD::CARRY_FALSE, N->getDebugLoc(),
1698 // fold (subc x, x) -> 0 + no borrow
1700 return CombineTo(N, DAG.getConstant(0, VT),
1701 DAG.getNode(ISD::CARRY_FALSE, N->getDebugLoc(),
1704 // fold (subc x, 0) -> x + no borrow
1705 if (N1C && N1C->isNullValue())
1706 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE, N->getDebugLoc(),
1709 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1) + no borrow
1710 if (N0C && N0C->isAllOnesValue())
1711 return CombineTo(N, DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N1, N0),
1712 DAG.getNode(ISD::CARRY_FALSE, N->getDebugLoc(),
1718 SDValue DAGCombiner::visitSUBE(SDNode *N) {
1719 SDValue N0 = N->getOperand(0);
1720 SDValue N1 = N->getOperand(1);
1721 SDValue CarryIn = N->getOperand(2);
1723 // fold (sube x, y, false) -> (subc x, y)
1724 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1725 return DAG.getNode(ISD::SUBC, N->getDebugLoc(), N->getVTList(), N0, N1);
1730 SDValue DAGCombiner::visitMUL(SDNode *N) {
1731 SDValue N0 = N->getOperand(0);
1732 SDValue N1 = N->getOperand(1);
1733 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1734 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1735 EVT VT = N0.getValueType();
1738 if (VT.isVector()) {
1739 SDValue FoldedVOp = SimplifyVBinOp(N);
1740 if (FoldedVOp.getNode()) return FoldedVOp;
1743 // fold (mul x, undef) -> 0
1744 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
1745 return DAG.getConstant(0, VT);
1746 // fold (mul c1, c2) -> c1*c2
1748 return DAG.FoldConstantArithmetic(ISD::MUL, VT, N0C, N1C);
1749 // canonicalize constant to RHS
1751 return DAG.getNode(ISD::MUL, N->getDebugLoc(), VT, N1, N0);
1752 // fold (mul x, 0) -> 0
1753 if (N1C && N1C->isNullValue())
1755 // fold (mul x, -1) -> 0-x
1756 if (N1C && N1C->isAllOnesValue())
1757 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1758 DAG.getConstant(0, VT), N0);
1759 // fold (mul x, (1 << c)) -> x << c
1760 if (N1C && N1C->getAPIntValue().isPowerOf2())
1761 return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0,
1762 DAG.getConstant(N1C->getAPIntValue().logBase2(),
1763 getShiftAmountTy(N0.getValueType())));
1764 // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c
1765 if (N1C && (-N1C->getAPIntValue()).isPowerOf2()) {
1766 unsigned Log2Val = (-N1C->getAPIntValue()).logBase2();
1767 // FIXME: If the input is something that is easily negated (e.g. a
1768 // single-use add), we should put the negate there.
1769 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1770 DAG.getConstant(0, VT),
1771 DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0,
1772 DAG.getConstant(Log2Val,
1773 getShiftAmountTy(N0.getValueType()))));
1775 // (mul (shl X, c1), c2) -> (mul X, c2 << c1)
1776 if (N1C && N0.getOpcode() == ISD::SHL &&
1777 isa<ConstantSDNode>(N0.getOperand(1))) {
1778 SDValue C3 = DAG.getNode(ISD::SHL, N->getDebugLoc(), VT,
1779 N1, N0.getOperand(1));
1780 AddToWorkList(C3.getNode());
1781 return DAG.getNode(ISD::MUL, N->getDebugLoc(), VT,
1782 N0.getOperand(0), C3);
1785 // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one
1788 SDValue Sh(0,0), Y(0,0);
1789 // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)).
1790 if (N0.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N0.getOperand(1)) &&
1791 N0.getNode()->hasOneUse()) {
1793 } else if (N1.getOpcode() == ISD::SHL &&
1794 isa<ConstantSDNode>(N1.getOperand(1)) &&
1795 N1.getNode()->hasOneUse()) {
1800 SDValue Mul = DAG.getNode(ISD::MUL, N->getDebugLoc(), VT,
1801 Sh.getOperand(0), Y);
1802 return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT,
1803 Mul, Sh.getOperand(1));
1807 // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2)
1808 if (N1C && N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() &&
1809 isa<ConstantSDNode>(N0.getOperand(1)))
1810 return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT,
1811 DAG.getNode(ISD::MUL, N0.getDebugLoc(), VT,
1812 N0.getOperand(0), N1),
1813 DAG.getNode(ISD::MUL, N1.getDebugLoc(), VT,
1814 N0.getOperand(1), N1));
1817 SDValue RMUL = ReassociateOps(ISD::MUL, N->getDebugLoc(), N0, N1);
1818 if (RMUL.getNode() != 0)
1824 SDValue DAGCombiner::visitSDIV(SDNode *N) {
1825 SDValue N0 = N->getOperand(0);
1826 SDValue N1 = N->getOperand(1);
1827 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1828 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1829 EVT VT = N->getValueType(0);
1832 if (VT.isVector()) {
1833 SDValue FoldedVOp = SimplifyVBinOp(N);
1834 if (FoldedVOp.getNode()) return FoldedVOp;
1837 // fold (sdiv c1, c2) -> c1/c2
1838 if (N0C && N1C && !N1C->isNullValue())
1839 return DAG.FoldConstantArithmetic(ISD::SDIV, VT, N0C, N1C);
1840 // fold (sdiv X, 1) -> X
1841 if (N1C && N1C->getAPIntValue() == 1LL)
1843 // fold (sdiv X, -1) -> 0-X
1844 if (N1C && N1C->isAllOnesValue())
1845 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1846 DAG.getConstant(0, VT), N0);
1847 // If we know the sign bits of both operands are zero, strength reduce to a
1848 // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2
1849 if (!VT.isVector()) {
1850 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
1851 return DAG.getNode(ISD::UDIV, N->getDebugLoc(), N1.getValueType(),
1854 // fold (sdiv X, pow2) -> simple ops after legalize
1855 if (N1C && !N1C->isNullValue() &&
1856 (N1C->getAPIntValue().isPowerOf2() ||
1857 (-N1C->getAPIntValue()).isPowerOf2())) {
1858 // If dividing by powers of two is cheap, then don't perform the following
1860 if (TLI.isPow2DivCheap())
1863 unsigned lg2 = N1C->getAPIntValue().countTrailingZeros();
1865 // Splat the sign bit into the register
1866 SDValue SGN = DAG.getNode(ISD::SRA, N->getDebugLoc(), VT, N0,
1867 DAG.getConstant(VT.getSizeInBits()-1,
1868 getShiftAmountTy(N0.getValueType())));
1869 AddToWorkList(SGN.getNode());
1871 // Add (N0 < 0) ? abs2 - 1 : 0;
1872 SDValue SRL = DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, SGN,
1873 DAG.getConstant(VT.getSizeInBits() - lg2,
1874 getShiftAmountTy(SGN.getValueType())));
1875 SDValue ADD = DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N0, SRL);
1876 AddToWorkList(SRL.getNode());
1877 AddToWorkList(ADD.getNode()); // Divide by pow2
1878 SDValue SRA = DAG.getNode(ISD::SRA, N->getDebugLoc(), VT, ADD,
1879 DAG.getConstant(lg2, getShiftAmountTy(ADD.getValueType())));
1881 // If we're dividing by a positive value, we're done. Otherwise, we must
1882 // negate the result.
1883 if (N1C->getAPIntValue().isNonNegative())
1886 AddToWorkList(SRA.getNode());
1887 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1888 DAG.getConstant(0, VT), SRA);
1891 // if integer divide is expensive and we satisfy the requirements, emit an
1892 // alternate sequence.
1893 if (N1C && !N1C->isNullValue() && !TLI.isIntDivCheap()) {
1894 SDValue Op = BuildSDIV(N);
1895 if (Op.getNode()) return Op;
1899 if (N0.getOpcode() == ISD::UNDEF)
1900 return DAG.getConstant(0, VT);
1901 // X / undef -> undef
1902 if (N1.getOpcode() == ISD::UNDEF)
1908 SDValue DAGCombiner::visitUDIV(SDNode *N) {
1909 SDValue N0 = N->getOperand(0);
1910 SDValue N1 = N->getOperand(1);
1911 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1912 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1913 EVT VT = N->getValueType(0);
1916 if (VT.isVector()) {
1917 SDValue FoldedVOp = SimplifyVBinOp(N);
1918 if (FoldedVOp.getNode()) return FoldedVOp;
1921 // fold (udiv c1, c2) -> c1/c2
1922 if (N0C && N1C && !N1C->isNullValue())
1923 return DAG.FoldConstantArithmetic(ISD::UDIV, VT, N0C, N1C);
1924 // fold (udiv x, (1 << c)) -> x >>u c
1925 if (N1C && N1C->getAPIntValue().isPowerOf2())
1926 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0,
1927 DAG.getConstant(N1C->getAPIntValue().logBase2(),
1928 getShiftAmountTy(N0.getValueType())));
1929 // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2
1930 if (N1.getOpcode() == ISD::SHL) {
1931 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
1932 if (SHC->getAPIntValue().isPowerOf2()) {
1933 EVT ADDVT = N1.getOperand(1).getValueType();
1934 SDValue Add = DAG.getNode(ISD::ADD, N->getDebugLoc(), ADDVT,
1936 DAG.getConstant(SHC->getAPIntValue()
1939 AddToWorkList(Add.getNode());
1940 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0, Add);
1944 // fold (udiv x, c) -> alternate
1945 if (N1C && !N1C->isNullValue() && !TLI.isIntDivCheap()) {
1946 SDValue Op = BuildUDIV(N);
1947 if (Op.getNode()) return Op;
1951 if (N0.getOpcode() == ISD::UNDEF)
1952 return DAG.getConstant(0, VT);
1953 // X / undef -> undef
1954 if (N1.getOpcode() == ISD::UNDEF)
1960 SDValue DAGCombiner::visitSREM(SDNode *N) {
1961 SDValue N0 = N->getOperand(0);
1962 SDValue N1 = N->getOperand(1);
1963 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1964 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1965 EVT VT = N->getValueType(0);
1967 // fold (srem c1, c2) -> c1%c2
1968 if (N0C && N1C && !N1C->isNullValue())
1969 return DAG.FoldConstantArithmetic(ISD::SREM, VT, N0C, N1C);
1970 // If we know the sign bits of both operands are zero, strength reduce to a
1971 // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15
1972 if (!VT.isVector()) {
1973 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
1974 return DAG.getNode(ISD::UREM, N->getDebugLoc(), VT, N0, N1);
1977 // If X/C can be simplified by the division-by-constant logic, lower
1978 // X%C to the equivalent of X-X/C*C.
1979 if (N1C && !N1C->isNullValue()) {
1980 SDValue Div = DAG.getNode(ISD::SDIV, N->getDebugLoc(), VT, N0, N1);
1981 AddToWorkList(Div.getNode());
1982 SDValue OptimizedDiv = combine(Div.getNode());
1983 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
1984 SDValue Mul = DAG.getNode(ISD::MUL, N->getDebugLoc(), VT,
1986 SDValue Sub = DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N0, Mul);
1987 AddToWorkList(Mul.getNode());
1993 if (N0.getOpcode() == ISD::UNDEF)
1994 return DAG.getConstant(0, VT);
1995 // X % undef -> undef
1996 if (N1.getOpcode() == ISD::UNDEF)
2002 SDValue DAGCombiner::visitUREM(SDNode *N) {
2003 SDValue N0 = N->getOperand(0);
2004 SDValue N1 = N->getOperand(1);
2005 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2006 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2007 EVT VT = N->getValueType(0);
2009 // fold (urem c1, c2) -> c1%c2
2010 if (N0C && N1C && !N1C->isNullValue())
2011 return DAG.FoldConstantArithmetic(ISD::UREM, VT, N0C, N1C);
2012 // fold (urem x, pow2) -> (and x, pow2-1)
2013 if (N1C && !N1C->isNullValue() && N1C->getAPIntValue().isPowerOf2())
2014 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0,
2015 DAG.getConstant(N1C->getAPIntValue()-1,VT));
2016 // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1))
2017 if (N1.getOpcode() == ISD::SHL) {
2018 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
2019 if (SHC->getAPIntValue().isPowerOf2()) {
2021 DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N1,
2022 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()),
2024 AddToWorkList(Add.getNode());
2025 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0, Add);
2030 // If X/C can be simplified by the division-by-constant logic, lower
2031 // X%C to the equivalent of X-X/C*C.
2032 if (N1C && !N1C->isNullValue()) {
2033 SDValue Div = DAG.getNode(ISD::UDIV, N->getDebugLoc(), VT, N0, N1);
2034 AddToWorkList(Div.getNode());
2035 SDValue OptimizedDiv = combine(Div.getNode());
2036 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2037 SDValue Mul = DAG.getNode(ISD::MUL, N->getDebugLoc(), VT,
2039 SDValue Sub = DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N0, Mul);
2040 AddToWorkList(Mul.getNode());
2046 if (N0.getOpcode() == ISD::UNDEF)
2047 return DAG.getConstant(0, VT);
2048 // X % undef -> undef
2049 if (N1.getOpcode() == ISD::UNDEF)
2055 SDValue DAGCombiner::visitMULHS(SDNode *N) {
2056 SDValue N0 = N->getOperand(0);
2057 SDValue N1 = N->getOperand(1);
2058 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2059 EVT VT = N->getValueType(0);
2060 DebugLoc DL = N->getDebugLoc();
2062 // fold (mulhs x, 0) -> 0
2063 if (N1C && N1C->isNullValue())
2065 // fold (mulhs x, 1) -> (sra x, size(x)-1)
2066 if (N1C && N1C->getAPIntValue() == 1)
2067 return DAG.getNode(ISD::SRA, N->getDebugLoc(), N0.getValueType(), N0,
2068 DAG.getConstant(N0.getValueType().getSizeInBits() - 1,
2069 getShiftAmountTy(N0.getValueType())));
2070 // fold (mulhs x, undef) -> 0
2071 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2072 return DAG.getConstant(0, VT);
2074 // If the type twice as wide is legal, transform the mulhs to a wider multiply
2076 if (VT.isSimple() && !VT.isVector()) {
2077 MVT Simple = VT.getSimpleVT();
2078 unsigned SimpleSize = Simple.getSizeInBits();
2079 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2080 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2081 N0 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N0);
2082 N1 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N1);
2083 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2084 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2085 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2086 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2093 SDValue DAGCombiner::visitMULHU(SDNode *N) {
2094 SDValue N0 = N->getOperand(0);
2095 SDValue N1 = N->getOperand(1);
2096 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2097 EVT VT = N->getValueType(0);
2098 DebugLoc DL = N->getDebugLoc();
2100 // fold (mulhu x, 0) -> 0
2101 if (N1C && N1C->isNullValue())
2103 // fold (mulhu x, 1) -> 0
2104 if (N1C && N1C->getAPIntValue() == 1)
2105 return DAG.getConstant(0, N0.getValueType());
2106 // fold (mulhu x, undef) -> 0
2107 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2108 return DAG.getConstant(0, VT);
2110 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2112 if (VT.isSimple() && !VT.isVector()) {
2113 MVT Simple = VT.getSimpleVT();
2114 unsigned SimpleSize = Simple.getSizeInBits();
2115 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2116 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2117 N0 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N0);
2118 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N1);
2119 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2120 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2121 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2122 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2129 /// SimplifyNodeWithTwoResults - Perform optimizations common to nodes that
2130 /// compute two values. LoOp and HiOp give the opcodes for the two computations
2131 /// that are being performed. Return true if a simplification was made.
2133 SDValue DAGCombiner::SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
2135 // If the high half is not needed, just compute the low half.
2136 bool HiExists = N->hasAnyUseOfValue(1);
2138 (!LegalOperations ||
2139 TLI.isOperationLegal(LoOp, N->getValueType(0)))) {
2140 SDValue Res = DAG.getNode(LoOp, N->getDebugLoc(), N->getValueType(0),
2141 N->op_begin(), N->getNumOperands());
2142 return CombineTo(N, Res, Res);
2145 // If the low half is not needed, just compute the high half.
2146 bool LoExists = N->hasAnyUseOfValue(0);
2148 (!LegalOperations ||
2149 TLI.isOperationLegal(HiOp, N->getValueType(1)))) {
2150 SDValue Res = DAG.getNode(HiOp, N->getDebugLoc(), N->getValueType(1),
2151 N->op_begin(), N->getNumOperands());
2152 return CombineTo(N, Res, Res);
2155 // If both halves are used, return as it is.
2156 if (LoExists && HiExists)
2159 // If the two computed results can be simplified separately, separate them.
2161 SDValue Lo = DAG.getNode(LoOp, N->getDebugLoc(), N->getValueType(0),
2162 N->op_begin(), N->getNumOperands());
2163 AddToWorkList(Lo.getNode());
2164 SDValue LoOpt = combine(Lo.getNode());
2165 if (LoOpt.getNode() && LoOpt.getNode() != Lo.getNode() &&
2166 (!LegalOperations ||
2167 TLI.isOperationLegal(LoOpt.getOpcode(), LoOpt.getValueType())))
2168 return CombineTo(N, LoOpt, LoOpt);
2172 SDValue Hi = DAG.getNode(HiOp, N->getDebugLoc(), N->getValueType(1),
2173 N->op_begin(), N->getNumOperands());
2174 AddToWorkList(Hi.getNode());
2175 SDValue HiOpt = combine(Hi.getNode());
2176 if (HiOpt.getNode() && HiOpt != Hi &&
2177 (!LegalOperations ||
2178 TLI.isOperationLegal(HiOpt.getOpcode(), HiOpt.getValueType())))
2179 return CombineTo(N, HiOpt, HiOpt);
2185 SDValue DAGCombiner::visitSMUL_LOHI(SDNode *N) {
2186 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHS);
2187 if (Res.getNode()) return Res;
2189 EVT VT = N->getValueType(0);
2190 DebugLoc DL = N->getDebugLoc();
2192 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2194 if (VT.isSimple() && !VT.isVector()) {
2195 MVT Simple = VT.getSimpleVT();
2196 unsigned SimpleSize = Simple.getSizeInBits();
2197 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2198 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2199 SDValue Lo = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(0));
2200 SDValue Hi = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(1));
2201 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2202 // Compute the high part as N1.
2203 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2204 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2205 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2206 // Compute the low part as N0.
2207 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2208 return CombineTo(N, Lo, Hi);
2215 SDValue DAGCombiner::visitUMUL_LOHI(SDNode *N) {
2216 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHU);
2217 if (Res.getNode()) return Res;
2219 EVT VT = N->getValueType(0);
2220 DebugLoc DL = N->getDebugLoc();
2222 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2224 if (VT.isSimple() && !VT.isVector()) {
2225 MVT Simple = VT.getSimpleVT();
2226 unsigned SimpleSize = Simple.getSizeInBits();
2227 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2228 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2229 SDValue Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(0));
2230 SDValue Hi = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(1));
2231 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2232 // Compute the high part as N1.
2233 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2234 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2235 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2236 // Compute the low part as N0.
2237 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2238 return CombineTo(N, Lo, Hi);
2245 SDValue DAGCombiner::visitSMULO(SDNode *N) {
2246 // (smulo x, 2) -> (saddo x, x)
2247 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2248 if (C2->getAPIntValue() == 2)
2249 return DAG.getNode(ISD::SADDO, N->getDebugLoc(), N->getVTList(),
2250 N->getOperand(0), N->getOperand(0));
2255 SDValue DAGCombiner::visitUMULO(SDNode *N) {
2256 // (umulo x, 2) -> (uaddo x, x)
2257 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2258 if (C2->getAPIntValue() == 2)
2259 return DAG.getNode(ISD::UADDO, N->getDebugLoc(), N->getVTList(),
2260 N->getOperand(0), N->getOperand(0));
2265 SDValue DAGCombiner::visitSDIVREM(SDNode *N) {
2266 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::SDIV, ISD::SREM);
2267 if (Res.getNode()) return Res;
2272 SDValue DAGCombiner::visitUDIVREM(SDNode *N) {
2273 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::UDIV, ISD::UREM);
2274 if (Res.getNode()) return Res;
2279 /// SimplifyBinOpWithSameOpcodeHands - If this is a binary operator with
2280 /// two operands of the same opcode, try to simplify it.
2281 SDValue DAGCombiner::SimplifyBinOpWithSameOpcodeHands(SDNode *N) {
2282 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
2283 EVT VT = N0.getValueType();
2284 assert(N0.getOpcode() == N1.getOpcode() && "Bad input!");
2286 // Bail early if none of these transforms apply.
2287 if (N0.getNode()->getNumOperands() == 0) return SDValue();
2289 // For each of OP in AND/OR/XOR:
2290 // fold (OP (zext x), (zext y)) -> (zext (OP x, y))
2291 // fold (OP (sext x), (sext y)) -> (sext (OP x, y))
2292 // fold (OP (aext x), (aext y)) -> (aext (OP x, y))
2293 // fold (OP (trunc x), (trunc y)) -> (trunc (OP x, y)) (if trunc isn't free)
2295 // do not sink logical op inside of a vector extend, since it may combine
2297 EVT Op0VT = N0.getOperand(0).getValueType();
2298 if ((N0.getOpcode() == ISD::ZERO_EXTEND ||
2299 N0.getOpcode() == ISD::SIGN_EXTEND ||
2300 // Avoid infinite looping with PromoteIntBinOp.
2301 (N0.getOpcode() == ISD::ANY_EXTEND &&
2302 (!LegalTypes || TLI.isTypeDesirableForOp(N->getOpcode(), Op0VT))) ||
2303 (N0.getOpcode() == ISD::TRUNCATE &&
2304 (!TLI.isZExtFree(VT, Op0VT) ||
2305 !TLI.isTruncateFree(Op0VT, VT)) &&
2306 TLI.isTypeLegal(Op0VT))) &&
2308 Op0VT == N1.getOperand(0).getValueType() &&
2309 (!LegalOperations || TLI.isOperationLegal(N->getOpcode(), Op0VT))) {
2310 SDValue ORNode = DAG.getNode(N->getOpcode(), N0.getDebugLoc(),
2311 N0.getOperand(0).getValueType(),
2312 N0.getOperand(0), N1.getOperand(0));
2313 AddToWorkList(ORNode.getNode());
2314 return DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT, ORNode);
2317 // For each of OP in SHL/SRL/SRA/AND...
2318 // fold (and (OP x, z), (OP y, z)) -> (OP (and x, y), z)
2319 // fold (or (OP x, z), (OP y, z)) -> (OP (or x, y), z)
2320 // fold (xor (OP x, z), (OP y, z)) -> (OP (xor x, y), z)
2321 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL ||
2322 N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::AND) &&
2323 N0.getOperand(1) == N1.getOperand(1)) {
2324 SDValue ORNode = DAG.getNode(N->getOpcode(), N0.getDebugLoc(),
2325 N0.getOperand(0).getValueType(),
2326 N0.getOperand(0), N1.getOperand(0));
2327 AddToWorkList(ORNode.getNode());
2328 return DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT,
2329 ORNode, N0.getOperand(1));
2332 // Simplify xor/and/or (bitcast(A), bitcast(B)) -> bitcast(op (A,B))
2333 // Only perform this optimization after type legalization and before
2334 // LegalizeVectorOprs. LegalizeVectorOprs promotes vector operations by
2335 // adding bitcasts. For example (xor v4i32) is promoted to (v2i64), and
2336 // we don't want to undo this promotion.
2337 // We also handle SCALAR_TO_VECTOR because xor/or/and operations are cheaper
2339 if ((N0.getOpcode() == ISD::BITCAST || N0.getOpcode() == ISD::SCALAR_TO_VECTOR)
2340 && Level == AfterLegalizeVectorOps) {
2341 SDValue In0 = N0.getOperand(0);
2342 SDValue In1 = N1.getOperand(0);
2343 EVT In0Ty = In0.getValueType();
2344 EVT In1Ty = In1.getValueType();
2345 // If both incoming values are integers, and the original types are the same.
2346 if (In0Ty.isInteger() && In1Ty.isInteger() && In0Ty == In1Ty) {
2347 SDValue Op = DAG.getNode(N->getOpcode(), N->getDebugLoc(), In0Ty, In0, In1);
2348 SDValue BC = DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT, Op);
2349 AddToWorkList(Op.getNode());
2354 // Xor/and/or are indifferent to the swizzle operation (shuffle of one value).
2355 // Simplify xor/and/or (shuff(A), shuff(B)) -> shuff(op (A,B))
2356 // If both shuffles use the same mask, and both shuffle within a single
2357 // vector, then it is worthwhile to move the swizzle after the operation.
2358 // The type-legalizer generates this pattern when loading illegal
2359 // vector types from memory. In many cases this allows additional shuffle
2361 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
2362 N0.getOperand(1).getOpcode() == ISD::UNDEF &&
2363 N1.getOperand(1).getOpcode() == ISD::UNDEF) {
2364 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(N0);
2365 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(N1);
2367 assert(N0.getOperand(0).getValueType() == N1.getOperand(1).getValueType() &&
2368 "Inputs to shuffles are not the same type");
2370 unsigned NumElts = VT.getVectorNumElements();
2372 // Check that both shuffles use the same mask. The masks are known to be of
2373 // the same length because the result vector type is the same.
2374 bool SameMask = true;
2375 for (unsigned i = 0; i != NumElts; ++i) {
2376 int Idx0 = SVN0->getMaskElt(i);
2377 int Idx1 = SVN1->getMaskElt(i);
2385 SDValue Op = DAG.getNode(N->getOpcode(), N->getDebugLoc(), VT,
2386 N0.getOperand(0), N1.getOperand(0));
2387 AddToWorkList(Op.getNode());
2388 return DAG.getVectorShuffle(VT, N->getDebugLoc(), Op,
2389 DAG.getUNDEF(VT), &SVN0->getMask()[0]);
2396 SDValue DAGCombiner::visitAND(SDNode *N) {
2397 SDValue N0 = N->getOperand(0);
2398 SDValue N1 = N->getOperand(1);
2399 SDValue LL, LR, RL, RR, CC0, CC1;
2400 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2401 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2402 EVT VT = N1.getValueType();
2403 unsigned BitWidth = VT.getScalarType().getSizeInBits();
2406 if (VT.isVector()) {
2407 SDValue FoldedVOp = SimplifyVBinOp(N);
2408 if (FoldedVOp.getNode()) return FoldedVOp;
2411 // fold (and x, undef) -> 0
2412 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2413 return DAG.getConstant(0, VT);
2414 // fold (and c1, c2) -> c1&c2
2416 return DAG.FoldConstantArithmetic(ISD::AND, VT, N0C, N1C);
2417 // canonicalize constant to RHS
2419 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N1, N0);
2420 // fold (and x, -1) -> x
2421 if (N1C && N1C->isAllOnesValue())
2423 // if (and x, c) is known to be zero, return 0
2424 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
2425 APInt::getAllOnesValue(BitWidth)))
2426 return DAG.getConstant(0, VT);
2428 SDValue RAND = ReassociateOps(ISD::AND, N->getDebugLoc(), N0, N1);
2429 if (RAND.getNode() != 0)
2431 // fold (and (or x, C), D) -> D if (C & D) == D
2432 if (N1C && N0.getOpcode() == ISD::OR)
2433 if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
2434 if ((ORI->getAPIntValue() & N1C->getAPIntValue()) == N1C->getAPIntValue())
2436 // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits.
2437 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
2438 SDValue N0Op0 = N0.getOperand(0);
2439 APInt Mask = ~N1C->getAPIntValue();
2440 Mask = Mask.trunc(N0Op0.getValueSizeInBits());
2441 if (DAG.MaskedValueIsZero(N0Op0, Mask)) {
2442 SDValue Zext = DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(),
2443 N0.getValueType(), N0Op0);
2445 // Replace uses of the AND with uses of the Zero extend node.
2448 // We actually want to replace all uses of the any_extend with the
2449 // zero_extend, to avoid duplicating things. This will later cause this
2450 // AND to be folded.
2451 CombineTo(N0.getNode(), Zext);
2452 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2455 // similarly fold (and (X (load ([non_ext|any_ext|zero_ext] V))), c) ->
2456 // (X (load ([non_ext|zero_ext] V))) if 'and' only clears top bits which must
2457 // already be zero by virtue of the width of the base type of the load.
2459 // the 'X' node here can either be nothing or an extract_vector_elt to catch
2461 if ((N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
2462 N0.getOperand(0).getOpcode() == ISD::LOAD) ||
2463 N0.getOpcode() == ISD::LOAD) {
2464 LoadSDNode *Load = cast<LoadSDNode>( (N0.getOpcode() == ISD::LOAD) ?
2465 N0 : N0.getOperand(0) );
2467 // Get the constant (if applicable) the zero'th operand is being ANDed with.
2468 // This can be a pure constant or a vector splat, in which case we treat the
2469 // vector as a scalar and use the splat value.
2470 APInt Constant = APInt::getNullValue(1);
2471 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
2472 Constant = C->getAPIntValue();
2473 } else if (BuildVectorSDNode *Vector = dyn_cast<BuildVectorSDNode>(N1)) {
2474 APInt SplatValue, SplatUndef;
2475 unsigned SplatBitSize;
2477 bool IsSplat = Vector->isConstantSplat(SplatValue, SplatUndef,
2478 SplatBitSize, HasAnyUndefs);
2480 // Undef bits can contribute to a possible optimisation if set, so
2482 SplatValue |= SplatUndef;
2484 // The splat value may be something like "0x00FFFFFF", which means 0 for
2485 // the first vector value and FF for the rest, repeating. We need a mask
2486 // that will apply equally to all members of the vector, so AND all the
2487 // lanes of the constant together.
2488 EVT VT = Vector->getValueType(0);
2489 unsigned BitWidth = VT.getVectorElementType().getSizeInBits();
2490 Constant = APInt::getAllOnesValue(BitWidth);
2491 for (unsigned i = 0, n = VT.getVectorNumElements(); i < n; ++i)
2492 Constant &= SplatValue.lshr(i*BitWidth).zextOrTrunc(BitWidth);
2496 // If we want to change an EXTLOAD to a ZEXTLOAD, ensure a ZEXTLOAD is
2497 // actually legal and isn't going to get expanded, else this is a false
2499 bool CanZextLoadProfitably = TLI.isLoadExtLegal(ISD::ZEXTLOAD,
2500 Load->getMemoryVT());
2502 // Resize the constant to the same size as the original memory access before
2503 // extension. If it is still the AllOnesValue then this AND is completely
2506 Constant.zextOrTrunc(Load->getMemoryVT().getScalarType().getSizeInBits());
2509 switch (Load->getExtensionType()) {
2510 default: B = false; break;
2511 case ISD::EXTLOAD: B = CanZextLoadProfitably; break;
2513 case ISD::NON_EXTLOAD: B = true; break;
2516 if (B && Constant.isAllOnesValue()) {
2517 // If the load type was an EXTLOAD, convert to ZEXTLOAD in order to
2518 // preserve semantics once we get rid of the AND.
2519 SDValue NewLoad(Load, 0);
2520 if (Load->getExtensionType() == ISD::EXTLOAD) {
2521 NewLoad = DAG.getLoad(Load->getAddressingMode(), ISD::ZEXTLOAD,
2522 Load->getValueType(0), Load->getDebugLoc(),
2523 Load->getChain(), Load->getBasePtr(),
2524 Load->getOffset(), Load->getMemoryVT(),
2525 Load->getMemOperand());
2526 // Replace uses of the EXTLOAD with the new ZEXTLOAD.
2527 if (Load->getNumValues() == 3) {
2528 // PRE/POST_INC loads have 3 values.
2529 SDValue To[] = { NewLoad.getValue(0), NewLoad.getValue(1),
2530 NewLoad.getValue(2) };
2531 CombineTo(Load, To, 3, true);
2533 CombineTo(Load, NewLoad.getValue(0), NewLoad.getValue(1));
2537 // Fold the AND away, taking care not to fold to the old load node if we
2539 CombineTo(N, (N0.getNode() == Load) ? NewLoad : N0);
2541 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2544 // fold (and (setcc x), (setcc y)) -> (setcc (and x, y))
2545 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
2546 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
2547 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
2549 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
2550 LL.getValueType().isInteger()) {
2551 // fold (and (seteq X, 0), (seteq Y, 0)) -> (seteq (or X, Y), 0)
2552 if (cast<ConstantSDNode>(LR)->isNullValue() && Op1 == ISD::SETEQ) {
2553 SDValue ORNode = DAG.getNode(ISD::OR, N0.getDebugLoc(),
2554 LR.getValueType(), LL, RL);
2555 AddToWorkList(ORNode.getNode());
2556 return DAG.getSetCC(N->getDebugLoc(), VT, ORNode, LR, Op1);
2558 // fold (and (seteq X, -1), (seteq Y, -1)) -> (seteq (and X, Y), -1)
2559 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETEQ) {
2560 SDValue ANDNode = DAG.getNode(ISD::AND, N0.getDebugLoc(),
2561 LR.getValueType(), LL, RL);
2562 AddToWorkList(ANDNode.getNode());
2563 return DAG.getSetCC(N->getDebugLoc(), VT, ANDNode, LR, Op1);
2565 // fold (and (setgt X, -1), (setgt Y, -1)) -> (setgt (or X, Y), -1)
2566 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETGT) {
2567 SDValue ORNode = DAG.getNode(ISD::OR, N0.getDebugLoc(),
2568 LR.getValueType(), LL, RL);
2569 AddToWorkList(ORNode.getNode());
2570 return DAG.getSetCC(N->getDebugLoc(), VT, ORNode, LR, Op1);
2573 // canonicalize equivalent to ll == rl
2574 if (LL == RR && LR == RL) {
2575 Op1 = ISD::getSetCCSwappedOperands(Op1);
2578 if (LL == RL && LR == RR) {
2579 bool isInteger = LL.getValueType().isInteger();
2580 ISD::CondCode Result = ISD::getSetCCAndOperation(Op0, Op1, isInteger);
2581 if (Result != ISD::SETCC_INVALID &&
2582 (!LegalOperations || TLI.isCondCodeLegal(Result, LL.getValueType())))
2583 return DAG.getSetCC(N->getDebugLoc(), N0.getValueType(),
2588 // Simplify: (and (op x...), (op y...)) -> (op (and x, y))
2589 if (N0.getOpcode() == N1.getOpcode()) {
2590 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
2591 if (Tmp.getNode()) return Tmp;
2594 // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1)
2595 // fold (and (sra)) -> (and (srl)) when possible.
2596 if (!VT.isVector() &&
2597 SimplifyDemandedBits(SDValue(N, 0)))
2598 return SDValue(N, 0);
2600 // fold (zext_inreg (extload x)) -> (zextload x)
2601 if (ISD::isEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode())) {
2602 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2603 EVT MemVT = LN0->getMemoryVT();
2604 // If we zero all the possible extended bits, then we can turn this into
2605 // a zextload if we are running before legalize or the operation is legal.
2606 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
2607 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
2608 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
2609 ((!LegalOperations && !LN0->isVolatile()) ||
2610 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
2611 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N0.getDebugLoc(), VT,
2612 LN0->getChain(), LN0->getBasePtr(),
2613 LN0->getPointerInfo(), MemVT,
2614 LN0->isVolatile(), LN0->isNonTemporal(),
2615 LN0->getAlignment());
2617 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
2618 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2621 // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use
2622 if (ISD::isSEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
2624 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2625 EVT MemVT = LN0->getMemoryVT();
2626 // If we zero all the possible extended bits, then we can turn this into
2627 // a zextload if we are running before legalize or the operation is legal.
2628 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
2629 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
2630 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
2631 ((!LegalOperations && !LN0->isVolatile()) ||
2632 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
2633 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N0.getDebugLoc(), VT,
2635 LN0->getBasePtr(), LN0->getPointerInfo(),
2637 LN0->isVolatile(), LN0->isNonTemporal(),
2638 LN0->getAlignment());
2640 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
2641 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2645 // fold (and (load x), 255) -> (zextload x, i8)
2646 // fold (and (extload x, i16), 255) -> (zextload x, i8)
2647 // fold (and (any_ext (extload x, i16)), 255) -> (zextload x, i8)
2648 if (N1C && (N0.getOpcode() == ISD::LOAD ||
2649 (N0.getOpcode() == ISD::ANY_EXTEND &&
2650 N0.getOperand(0).getOpcode() == ISD::LOAD))) {
2651 bool HasAnyExt = N0.getOpcode() == ISD::ANY_EXTEND;
2652 LoadSDNode *LN0 = HasAnyExt
2653 ? cast<LoadSDNode>(N0.getOperand(0))
2654 : cast<LoadSDNode>(N0);
2655 if (LN0->getExtensionType() != ISD::SEXTLOAD &&
2656 LN0->isUnindexed() && N0.hasOneUse() && LN0->hasOneUse()) {
2657 uint32_t ActiveBits = N1C->getAPIntValue().getActiveBits();
2658 if (ActiveBits > 0 && APIntOps::isMask(ActiveBits, N1C->getAPIntValue())){
2659 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits);
2660 EVT LoadedVT = LN0->getMemoryVT();
2662 if (ExtVT == LoadedVT &&
2663 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) {
2664 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
2667 DAG.getExtLoad(ISD::ZEXTLOAD, LN0->getDebugLoc(), LoadResultTy,
2668 LN0->getChain(), LN0->getBasePtr(),
2669 LN0->getPointerInfo(),
2670 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
2671 LN0->getAlignment());
2673 CombineTo(LN0, NewLoad, NewLoad.getValue(1));
2674 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2677 // Do not change the width of a volatile load.
2678 // Do not generate loads of non-round integer types since these can
2679 // be expensive (and would be wrong if the type is not byte sized).
2680 if (!LN0->isVolatile() && LoadedVT.bitsGT(ExtVT) && ExtVT.isRound() &&
2681 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) {
2682 EVT PtrType = LN0->getOperand(1).getValueType();
2684 unsigned Alignment = LN0->getAlignment();
2685 SDValue NewPtr = LN0->getBasePtr();
2687 // For big endian targets, we need to add an offset to the pointer
2688 // to load the correct bytes. For little endian systems, we merely
2689 // need to read fewer bytes from the same pointer.
2690 if (TLI.isBigEndian()) {
2691 unsigned LVTStoreBytes = LoadedVT.getStoreSize();
2692 unsigned EVTStoreBytes = ExtVT.getStoreSize();
2693 unsigned PtrOff = LVTStoreBytes - EVTStoreBytes;
2694 NewPtr = DAG.getNode(ISD::ADD, LN0->getDebugLoc(), PtrType,
2695 NewPtr, DAG.getConstant(PtrOff, PtrType));
2696 Alignment = MinAlign(Alignment, PtrOff);
2699 AddToWorkList(NewPtr.getNode());
2701 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
2703 DAG.getExtLoad(ISD::ZEXTLOAD, LN0->getDebugLoc(), LoadResultTy,
2704 LN0->getChain(), NewPtr,
2705 LN0->getPointerInfo(),
2706 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
2709 CombineTo(LN0, Load, Load.getValue(1));
2710 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2719 /// MatchBSwapHWord - Match (a >> 8) | (a << 8) as (bswap a) >> 16
2721 SDValue DAGCombiner::MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
2722 bool DemandHighBits) {
2723 if (!LegalOperations)
2726 EVT VT = N->getValueType(0);
2727 if (VT != MVT::i64 && VT != MVT::i32 && VT != MVT::i16)
2729 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
2732 // Recognize (and (shl a, 8), 0xff), (and (srl a, 8), 0xff00)
2733 bool LookPassAnd0 = false;
2734 bool LookPassAnd1 = false;
2735 if (N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode() == ISD::SRL)
2737 if (N1.getOpcode() == ISD::AND && N1.getOperand(0).getOpcode() == ISD::SHL)
2739 if (N0.getOpcode() == ISD::AND) {
2740 if (!N0.getNode()->hasOneUse())
2742 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
2743 if (!N01C || N01C->getZExtValue() != 0xFF00)
2745 N0 = N0.getOperand(0);
2746 LookPassAnd0 = true;
2749 if (N1.getOpcode() == ISD::AND) {
2750 if (!N1.getNode()->hasOneUse())
2752 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
2753 if (!N11C || N11C->getZExtValue() != 0xFF)
2755 N1 = N1.getOperand(0);
2756 LookPassAnd1 = true;
2759 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
2761 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
2763 if (!N0.getNode()->hasOneUse() ||
2764 !N1.getNode()->hasOneUse())
2767 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
2768 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
2771 if (N01C->getZExtValue() != 8 || N11C->getZExtValue() != 8)
2774 // Look for (shl (and a, 0xff), 8), (srl (and a, 0xff00), 8)
2775 SDValue N00 = N0->getOperand(0);
2776 if (!LookPassAnd0 && N00.getOpcode() == ISD::AND) {
2777 if (!N00.getNode()->hasOneUse())
2779 ConstantSDNode *N001C = dyn_cast<ConstantSDNode>(N00.getOperand(1));
2780 if (!N001C || N001C->getZExtValue() != 0xFF)
2782 N00 = N00.getOperand(0);
2783 LookPassAnd0 = true;
2786 SDValue N10 = N1->getOperand(0);
2787 if (!LookPassAnd1 && N10.getOpcode() == ISD::AND) {
2788 if (!N10.getNode()->hasOneUse())
2790 ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N10.getOperand(1));
2791 if (!N101C || N101C->getZExtValue() != 0xFF00)
2793 N10 = N10.getOperand(0);
2794 LookPassAnd1 = true;
2800 // Make sure everything beyond the low halfword is zero since the SRL 16
2801 // will clear the top bits.
2802 unsigned OpSizeInBits = VT.getSizeInBits();
2803 if (DemandHighBits && OpSizeInBits > 16 &&
2804 (!LookPassAnd0 || !LookPassAnd1) &&
2805 !DAG.MaskedValueIsZero(N10, APInt::getHighBitsSet(OpSizeInBits, 16)))
2808 SDValue Res = DAG.getNode(ISD::BSWAP, N->getDebugLoc(), VT, N00);
2809 if (OpSizeInBits > 16)
2810 Res = DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, Res,
2811 DAG.getConstant(OpSizeInBits-16, getShiftAmountTy(VT)));
2815 /// isBSwapHWordElement - Return true if the specified node is an element
2816 /// that makes up a 32-bit packed halfword byteswap. i.e.
2817 /// ((x&0xff)<<8)|((x&0xff00)>>8)|((x&0x00ff0000)<<8)|((x&0xff000000)>>8)
2818 static bool isBSwapHWordElement(SDValue N, SmallVector<SDNode*,4> &Parts) {
2819 if (!N.getNode()->hasOneUse())
2822 unsigned Opc = N.getOpcode();
2823 if (Opc != ISD::AND && Opc != ISD::SHL && Opc != ISD::SRL)
2826 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N.getOperand(1));
2831 switch (N1C->getZExtValue()) {
2834 case 0xFF: Num = 0; break;
2835 case 0xFF00: Num = 1; break;
2836 case 0xFF0000: Num = 2; break;
2837 case 0xFF000000: Num = 3; break;
2840 // Look for (x & 0xff) << 8 as well as ((x << 8) & 0xff00).
2841 SDValue N0 = N.getOperand(0);
2842 if (Opc == ISD::AND) {
2843 if (Num == 0 || Num == 2) {
2845 // (x >> 8) & 0xff0000
2846 if (N0.getOpcode() != ISD::SRL)
2848 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
2849 if (!C || C->getZExtValue() != 8)
2852 // (x << 8) & 0xff00
2853 // (x << 8) & 0xff000000
2854 if (N0.getOpcode() != ISD::SHL)
2856 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
2857 if (!C || C->getZExtValue() != 8)
2860 } else if (Opc == ISD::SHL) {
2862 // (x & 0xff0000) << 8
2863 if (Num != 0 && Num != 2)
2865 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
2866 if (!C || C->getZExtValue() != 8)
2868 } else { // Opc == ISD::SRL
2869 // (x & 0xff00) >> 8
2870 // (x & 0xff000000) >> 8
2871 if (Num != 1 && Num != 3)
2873 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
2874 if (!C || C->getZExtValue() != 8)
2881 Parts[Num] = N0.getOperand(0).getNode();
2885 /// MatchBSwapHWord - Match a 32-bit packed halfword bswap. That is
2886 /// ((x&0xff)<<8)|((x&0xff00)>>8)|((x&0x00ff0000)<<8)|((x&0xff000000)>>8)
2887 /// => (rotl (bswap x), 16)
2888 SDValue DAGCombiner::MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1) {
2889 if (!LegalOperations)
2892 EVT VT = N->getValueType(0);
2895 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
2898 SmallVector<SDNode*,4> Parts(4, (SDNode*)0);
2900 // (or (or (and), (and)), (or (and), (and)))
2901 // (or (or (or (and), (and)), (and)), (and))
2902 if (N0.getOpcode() != ISD::OR)
2904 SDValue N00 = N0.getOperand(0);
2905 SDValue N01 = N0.getOperand(1);
2907 if (N1.getOpcode() == ISD::OR) {
2908 // (or (or (and), (and)), (or (and), (and)))
2909 SDValue N000 = N00.getOperand(0);
2910 if (!isBSwapHWordElement(N000, Parts))
2913 SDValue N001 = N00.getOperand(1);
2914 if (!isBSwapHWordElement(N001, Parts))
2916 SDValue N010 = N01.getOperand(0);
2917 if (!isBSwapHWordElement(N010, Parts))
2919 SDValue N011 = N01.getOperand(1);
2920 if (!isBSwapHWordElement(N011, Parts))
2923 // (or (or (or (and), (and)), (and)), (and))
2924 if (!isBSwapHWordElement(N1, Parts))
2926 if (!isBSwapHWordElement(N01, Parts))
2928 if (N00.getOpcode() != ISD::OR)
2930 SDValue N000 = N00.getOperand(0);
2931 if (!isBSwapHWordElement(N000, Parts))
2933 SDValue N001 = N00.getOperand(1);
2934 if (!isBSwapHWordElement(N001, Parts))
2938 // Make sure the parts are all coming from the same node.
2939 if (Parts[0] != Parts[1] || Parts[0] != Parts[2] || Parts[0] != Parts[3])
2942 SDValue BSwap = DAG.getNode(ISD::BSWAP, N->getDebugLoc(), VT,
2943 SDValue(Parts[0],0));
2945 // Result of the bswap should be rotated by 16. If it's not legal, than
2946 // do (x << 16) | (x >> 16).
2947 SDValue ShAmt = DAG.getConstant(16, getShiftAmountTy(VT));
2948 if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT))
2949 return DAG.getNode(ISD::ROTL, N->getDebugLoc(), VT, BSwap, ShAmt);
2950 else if (TLI.isOperationLegalOrCustom(ISD::ROTR, VT))
2951 return DAG.getNode(ISD::ROTR, N->getDebugLoc(), VT, BSwap, ShAmt);
2952 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT,
2953 DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, BSwap, ShAmt),
2954 DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, BSwap, ShAmt));
2957 SDValue DAGCombiner::visitOR(SDNode *N) {
2958 SDValue N0 = N->getOperand(0);
2959 SDValue N1 = N->getOperand(1);
2960 SDValue LL, LR, RL, RR, CC0, CC1;
2961 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2962 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2963 EVT VT = N1.getValueType();
2966 if (VT.isVector()) {
2967 SDValue FoldedVOp = SimplifyVBinOp(N);
2968 if (FoldedVOp.getNode()) return FoldedVOp;
2971 // fold (or x, undef) -> -1
2972 if (!LegalOperations &&
2973 (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)) {
2974 EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT;
2975 return DAG.getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()), VT);
2977 // fold (or c1, c2) -> c1|c2
2979 return DAG.FoldConstantArithmetic(ISD::OR, VT, N0C, N1C);
2980 // canonicalize constant to RHS
2982 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N1, N0);
2983 // fold (or x, 0) -> x
2984 if (N1C && N1C->isNullValue())
2986 // fold (or x, -1) -> -1
2987 if (N1C && N1C->isAllOnesValue())
2989 // fold (or x, c) -> c iff (x & ~c) == 0
2990 if (N1C && DAG.MaskedValueIsZero(N0, ~N1C->getAPIntValue()))
2993 // Recognize halfword bswaps as (bswap + rotl 16) or (bswap + shl 16)
2994 SDValue BSwap = MatchBSwapHWord(N, N0, N1);
2995 if (BSwap.getNode() != 0)
2997 BSwap = MatchBSwapHWordLow(N, N0, N1);
2998 if (BSwap.getNode() != 0)
3002 SDValue ROR = ReassociateOps(ISD::OR, N->getDebugLoc(), N0, N1);
3003 if (ROR.getNode() != 0)
3005 // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2)
3006 // iff (c1 & c2) == 0.
3007 if (N1C && N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
3008 isa<ConstantSDNode>(N0.getOperand(1))) {
3009 ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1));
3010 if ((C1->getAPIntValue() & N1C->getAPIntValue()) != 0)
3011 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
3012 DAG.getNode(ISD::OR, N0.getDebugLoc(), VT,
3013 N0.getOperand(0), N1),
3014 DAG.FoldConstantArithmetic(ISD::OR, VT, N1C, C1));
3016 // fold (or (setcc x), (setcc y)) -> (setcc (or x, y))
3017 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
3018 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
3019 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
3021 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
3022 LL.getValueType().isInteger()) {
3023 // fold (or (setne X, 0), (setne Y, 0)) -> (setne (or X, Y), 0)
3024 // fold (or (setlt X, 0), (setlt Y, 0)) -> (setne (or X, Y), 0)
3025 if (cast<ConstantSDNode>(LR)->isNullValue() &&
3026 (Op1 == ISD::SETNE || Op1 == ISD::SETLT)) {
3027 SDValue ORNode = DAG.getNode(ISD::OR, LR.getDebugLoc(),
3028 LR.getValueType(), LL, RL);
3029 AddToWorkList(ORNode.getNode());
3030 return DAG.getSetCC(N->getDebugLoc(), VT, ORNode, LR, Op1);
3032 // fold (or (setne X, -1), (setne Y, -1)) -> (setne (and X, Y), -1)
3033 // fold (or (setgt X, -1), (setgt Y -1)) -> (setgt (and X, Y), -1)
3034 if (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
3035 (Op1 == ISD::SETNE || Op1 == ISD::SETGT)) {
3036 SDValue ANDNode = DAG.getNode(ISD::AND, LR.getDebugLoc(),
3037 LR.getValueType(), LL, RL);
3038 AddToWorkList(ANDNode.getNode());
3039 return DAG.getSetCC(N->getDebugLoc(), VT, ANDNode, LR, Op1);
3042 // canonicalize equivalent to ll == rl
3043 if (LL == RR && LR == RL) {
3044 Op1 = ISD::getSetCCSwappedOperands(Op1);
3047 if (LL == RL && LR == RR) {
3048 bool isInteger = LL.getValueType().isInteger();
3049 ISD::CondCode Result = ISD::getSetCCOrOperation(Op0, Op1, isInteger);
3050 if (Result != ISD::SETCC_INVALID &&
3051 (!LegalOperations || TLI.isCondCodeLegal(Result, LL.getValueType())))
3052 return DAG.getSetCC(N->getDebugLoc(), N0.getValueType(),
3057 // Simplify: (or (op x...), (op y...)) -> (op (or x, y))
3058 if (N0.getOpcode() == N1.getOpcode()) {
3059 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3060 if (Tmp.getNode()) return Tmp;
3063 // (or (and X, C1), (and Y, C2)) -> (and (or X, Y), C3) if possible.
3064 if (N0.getOpcode() == ISD::AND &&
3065 N1.getOpcode() == ISD::AND &&
3066 N0.getOperand(1).getOpcode() == ISD::Constant &&
3067 N1.getOperand(1).getOpcode() == ISD::Constant &&
3068 // Don't increase # computations.
3069 (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) {
3070 // We can only do this xform if we know that bits from X that are set in C2
3071 // but not in C1 are already zero. Likewise for Y.
3072 const APInt &LHSMask =
3073 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
3074 const APInt &RHSMask =
3075 cast<ConstantSDNode>(N1.getOperand(1))->getAPIntValue();
3077 if (DAG.MaskedValueIsZero(N0.getOperand(0), RHSMask&~LHSMask) &&
3078 DAG.MaskedValueIsZero(N1.getOperand(0), LHSMask&~RHSMask)) {
3079 SDValue X = DAG.getNode(ISD::OR, N0.getDebugLoc(), VT,
3080 N0.getOperand(0), N1.getOperand(0));
3081 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, X,
3082 DAG.getConstant(LHSMask | RHSMask, VT));
3086 // See if this is some rotate idiom.
3087 if (SDNode *Rot = MatchRotate(N0, N1, N->getDebugLoc()))
3088 return SDValue(Rot, 0);
3090 // Simplify the operands using demanded-bits information.
3091 if (!VT.isVector() &&
3092 SimplifyDemandedBits(SDValue(N, 0)))
3093 return SDValue(N, 0);
3098 /// MatchRotateHalf - Match "(X shl/srl V1) & V2" where V2 may not be present.
3099 static bool MatchRotateHalf(SDValue Op, SDValue &Shift, SDValue &Mask) {
3100 if (Op.getOpcode() == ISD::AND) {
3101 if (isa<ConstantSDNode>(Op.getOperand(1))) {
3102 Mask = Op.getOperand(1);
3103 Op = Op.getOperand(0);
3109 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SHL) {
3117 // MatchRotate - Handle an 'or' of two operands. If this is one of the many
3118 // idioms for rotate, and if the target supports rotation instructions, generate
3120 SDNode *DAGCombiner::MatchRotate(SDValue LHS, SDValue RHS, DebugLoc DL) {
3121 // Must be a legal type. Expanded 'n promoted things won't work with rotates.
3122 EVT VT = LHS.getValueType();
3123 if (!TLI.isTypeLegal(VT)) return 0;
3125 // The target must have at least one rotate flavor.
3126 bool HasROTL = TLI.isOperationLegalOrCustom(ISD::ROTL, VT);
3127 bool HasROTR = TLI.isOperationLegalOrCustom(ISD::ROTR, VT);
3128 if (!HasROTL && !HasROTR) return 0;
3130 // Match "(X shl/srl V1) & V2" where V2 may not be present.
3131 SDValue LHSShift; // The shift.
3132 SDValue LHSMask; // AND value if any.
3133 if (!MatchRotateHalf(LHS, LHSShift, LHSMask))
3134 return 0; // Not part of a rotate.
3136 SDValue RHSShift; // The shift.
3137 SDValue RHSMask; // AND value if any.
3138 if (!MatchRotateHalf(RHS, RHSShift, RHSMask))
3139 return 0; // Not part of a rotate.
3141 if (LHSShift.getOperand(0) != RHSShift.getOperand(0))
3142 return 0; // Not shifting the same value.
3144 if (LHSShift.getOpcode() == RHSShift.getOpcode())
3145 return 0; // Shifts must disagree.
3147 // Canonicalize shl to left side in a shl/srl pair.
3148 if (RHSShift.getOpcode() == ISD::SHL) {
3149 std::swap(LHS, RHS);
3150 std::swap(LHSShift, RHSShift);
3151 std::swap(LHSMask , RHSMask );
3154 unsigned OpSizeInBits = VT.getSizeInBits();
3155 SDValue LHSShiftArg = LHSShift.getOperand(0);
3156 SDValue LHSShiftAmt = LHSShift.getOperand(1);
3157 SDValue RHSShiftAmt = RHSShift.getOperand(1);
3159 // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1)
3160 // fold (or (shl x, C1), (srl x, C2)) -> (rotr x, C2)
3161 if (LHSShiftAmt.getOpcode() == ISD::Constant &&
3162 RHSShiftAmt.getOpcode() == ISD::Constant) {
3163 uint64_t LShVal = cast<ConstantSDNode>(LHSShiftAmt)->getZExtValue();
3164 uint64_t RShVal = cast<ConstantSDNode>(RHSShiftAmt)->getZExtValue();
3165 if ((LShVal + RShVal) != OpSizeInBits)
3170 Rot = DAG.getNode(ISD::ROTL, DL, VT, LHSShiftArg, LHSShiftAmt);
3172 Rot = DAG.getNode(ISD::ROTR, DL, VT, LHSShiftArg, RHSShiftAmt);
3174 // If there is an AND of either shifted operand, apply it to the result.
3175 if (LHSMask.getNode() || RHSMask.getNode()) {
3176 APInt Mask = APInt::getAllOnesValue(OpSizeInBits);
3178 if (LHSMask.getNode()) {
3179 APInt RHSBits = APInt::getLowBitsSet(OpSizeInBits, LShVal);
3180 Mask &= cast<ConstantSDNode>(LHSMask)->getAPIntValue() | RHSBits;
3182 if (RHSMask.getNode()) {
3183 APInt LHSBits = APInt::getHighBitsSet(OpSizeInBits, RShVal);
3184 Mask &= cast<ConstantSDNode>(RHSMask)->getAPIntValue() | LHSBits;
3187 Rot = DAG.getNode(ISD::AND, DL, VT, Rot, DAG.getConstant(Mask, VT));
3190 return Rot.getNode();
3193 // If there is a mask here, and we have a variable shift, we can't be sure
3194 // that we're masking out the right stuff.
3195 if (LHSMask.getNode() || RHSMask.getNode())
3198 // fold (or (shl x, y), (srl x, (sub 32, y))) -> (rotl x, y)
3199 // fold (or (shl x, y), (srl x, (sub 32, y))) -> (rotr x, (sub 32, y))
3200 if (RHSShiftAmt.getOpcode() == ISD::SUB &&
3201 LHSShiftAmt == RHSShiftAmt.getOperand(1)) {
3202 if (ConstantSDNode *SUBC =
3203 dyn_cast<ConstantSDNode>(RHSShiftAmt.getOperand(0))) {
3204 if (SUBC->getAPIntValue() == OpSizeInBits) {
3206 return DAG.getNode(ISD::ROTL, DL, VT,
3207 LHSShiftArg, LHSShiftAmt).getNode();
3209 return DAG.getNode(ISD::ROTR, DL, VT,
3210 LHSShiftArg, RHSShiftAmt).getNode();
3215 // fold (or (shl x, (sub 32, y)), (srl x, r)) -> (rotr x, y)
3216 // fold (or (shl x, (sub 32, y)), (srl x, r)) -> (rotl x, (sub 32, y))
3217 if (LHSShiftAmt.getOpcode() == ISD::SUB &&
3218 RHSShiftAmt == LHSShiftAmt.getOperand(1)) {
3219 if (ConstantSDNode *SUBC =
3220 dyn_cast<ConstantSDNode>(LHSShiftAmt.getOperand(0))) {
3221 if (SUBC->getAPIntValue() == OpSizeInBits) {
3223 return DAG.getNode(ISD::ROTR, DL, VT,
3224 LHSShiftArg, RHSShiftAmt).getNode();
3226 return DAG.getNode(ISD::ROTL, DL, VT,
3227 LHSShiftArg, LHSShiftAmt).getNode();
3232 // Look for sign/zext/any-extended or truncate cases:
3233 if ((LHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND
3234 || LHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND
3235 || LHSShiftAmt.getOpcode() == ISD::ANY_EXTEND
3236 || LHSShiftAmt.getOpcode() == ISD::TRUNCATE) &&
3237 (RHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND
3238 || RHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND
3239 || RHSShiftAmt.getOpcode() == ISD::ANY_EXTEND
3240 || RHSShiftAmt.getOpcode() == ISD::TRUNCATE)) {
3241 SDValue LExtOp0 = LHSShiftAmt.getOperand(0);
3242 SDValue RExtOp0 = RHSShiftAmt.getOperand(0);
3243 if (RExtOp0.getOpcode() == ISD::SUB &&
3244 RExtOp0.getOperand(1) == LExtOp0) {
3245 // fold (or (shl x, (*ext y)), (srl x, (*ext (sub 32, y)))) ->
3247 // fold (or (shl x, (*ext y)), (srl x, (*ext (sub 32, y)))) ->
3248 // (rotr x, (sub 32, y))
3249 if (ConstantSDNode *SUBC =
3250 dyn_cast<ConstantSDNode>(RExtOp0.getOperand(0))) {
3251 if (SUBC->getAPIntValue() == OpSizeInBits) {
3252 return DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT,
3254 HasROTL ? LHSShiftAmt : RHSShiftAmt).getNode();
3257 } else if (LExtOp0.getOpcode() == ISD::SUB &&
3258 RExtOp0 == LExtOp0.getOperand(1)) {
3259 // fold (or (shl x, (*ext (sub 32, y))), (srl x, (*ext y))) ->
3261 // fold (or (shl x, (*ext (sub 32, y))), (srl x, (*ext y))) ->
3262 // (rotl x, (sub 32, y))
3263 if (ConstantSDNode *SUBC =
3264 dyn_cast<ConstantSDNode>(LExtOp0.getOperand(0))) {
3265 if (SUBC->getAPIntValue() == OpSizeInBits) {
3266 return DAG.getNode(HasROTR ? ISD::ROTR : ISD::ROTL, DL, VT,
3268 HasROTR ? RHSShiftAmt : LHSShiftAmt).getNode();
3277 SDValue DAGCombiner::visitXOR(SDNode *N) {
3278 SDValue N0 = N->getOperand(0);
3279 SDValue N1 = N->getOperand(1);
3280 SDValue LHS, RHS, CC;
3281 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3282 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3283 EVT VT = N0.getValueType();
3286 if (VT.isVector()) {
3287 SDValue FoldedVOp = SimplifyVBinOp(N);
3288 if (FoldedVOp.getNode()) return FoldedVOp;
3291 // fold (xor undef, undef) -> 0. This is a common idiom (misuse).
3292 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
3293 return DAG.getConstant(0, VT);
3294 // fold (xor x, undef) -> undef
3295 if (N0.getOpcode() == ISD::UNDEF)
3297 if (N1.getOpcode() == ISD::UNDEF)
3299 // fold (xor c1, c2) -> c1^c2
3301 return DAG.FoldConstantArithmetic(ISD::XOR, VT, N0C, N1C);
3302 // canonicalize constant to RHS
3304 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N1, N0);
3305 // fold (xor x, 0) -> x
3306 if (N1C && N1C->isNullValue())
3309 SDValue RXOR = ReassociateOps(ISD::XOR, N->getDebugLoc(), N0, N1);
3310 if (RXOR.getNode() != 0)
3313 // fold !(x cc y) -> (x !cc y)
3314 if (N1C && N1C->getAPIntValue() == 1 && isSetCCEquivalent(N0, LHS, RHS, CC)) {
3315 bool isInt = LHS.getValueType().isInteger();
3316 ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
3319 if (!LegalOperations || TLI.isCondCodeLegal(NotCC, LHS.getValueType())) {
3320 switch (N0.getOpcode()) {
3322 llvm_unreachable("Unhandled SetCC Equivalent!");
3324 return DAG.getSetCC(N->getDebugLoc(), VT, LHS, RHS, NotCC);
3325 case ISD::SELECT_CC:
3326 return DAG.getSelectCC(N->getDebugLoc(), LHS, RHS, N0.getOperand(2),
3327 N0.getOperand(3), NotCC);
3332 // fold (not (zext (setcc x, y))) -> (zext (not (setcc x, y)))
3333 if (N1C && N1C->getAPIntValue() == 1 && N0.getOpcode() == ISD::ZERO_EXTEND &&
3334 N0.getNode()->hasOneUse() &&
3335 isSetCCEquivalent(N0.getOperand(0), LHS, RHS, CC)){
3336 SDValue V = N0.getOperand(0);
3337 V = DAG.getNode(ISD::XOR, N0.getDebugLoc(), V.getValueType(), V,
3338 DAG.getConstant(1, V.getValueType()));
3339 AddToWorkList(V.getNode());
3340 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, V);
3343 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are setcc
3344 if (N1C && N1C->getAPIntValue() == 1 && VT == MVT::i1 &&
3345 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3346 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3347 if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) {
3348 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3349 LHS = DAG.getNode(ISD::XOR, LHS.getDebugLoc(), VT, LHS, N1); // LHS = ~LHS
3350 RHS = DAG.getNode(ISD::XOR, RHS.getDebugLoc(), VT, RHS, N1); // RHS = ~RHS
3351 AddToWorkList(LHS.getNode()); AddToWorkList(RHS.getNode());
3352 return DAG.getNode(NewOpcode, N->getDebugLoc(), VT, LHS, RHS);
3355 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are constants
3356 if (N1C && N1C->isAllOnesValue() &&
3357 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3358 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3359 if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) {
3360 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3361 LHS = DAG.getNode(ISD::XOR, LHS.getDebugLoc(), VT, LHS, N1); // LHS = ~LHS
3362 RHS = DAG.getNode(ISD::XOR, RHS.getDebugLoc(), VT, RHS, N1); // RHS = ~RHS
3363 AddToWorkList(LHS.getNode()); AddToWorkList(RHS.getNode());
3364 return DAG.getNode(NewOpcode, N->getDebugLoc(), VT, LHS, RHS);
3367 // fold (xor (xor x, c1), c2) -> (xor x, (xor c1, c2))
3368 if (N1C && N0.getOpcode() == ISD::XOR) {
3369 ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0));
3370 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3372 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N0.getOperand(1),
3373 DAG.getConstant(N1C->getAPIntValue() ^
3374 N00C->getAPIntValue(), VT));
3376 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N0.getOperand(0),
3377 DAG.getConstant(N1C->getAPIntValue() ^
3378 N01C->getAPIntValue(), VT));
3380 // fold (xor x, x) -> 0
3382 return tryFoldToZero(N->getDebugLoc(), TLI, VT, DAG, LegalOperations);
3384 // Simplify: xor (op x...), (op y...) -> (op (xor x, y))
3385 if (N0.getOpcode() == N1.getOpcode()) {
3386 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3387 if (Tmp.getNode()) return Tmp;
3390 // Simplify the expression using non-local knowledge.
3391 if (!VT.isVector() &&
3392 SimplifyDemandedBits(SDValue(N, 0)))
3393 return SDValue(N, 0);
3398 /// visitShiftByConstant - Handle transforms common to the three shifts, when
3399 /// the shift amount is a constant.
3400 SDValue DAGCombiner::visitShiftByConstant(SDNode *N, unsigned Amt) {
3401 SDNode *LHS = N->getOperand(0).getNode();
3402 if (!LHS->hasOneUse()) return SDValue();
3404 // We want to pull some binops through shifts, so that we have (and (shift))
3405 // instead of (shift (and)), likewise for add, or, xor, etc. This sort of
3406 // thing happens with address calculations, so it's important to canonicalize
3408 bool HighBitSet = false; // Can we transform this if the high bit is set?
3410 switch (LHS->getOpcode()) {
3411 default: return SDValue();
3414 HighBitSet = false; // We can only transform sra if the high bit is clear.
3417 HighBitSet = true; // We can only transform sra if the high bit is set.
3420 if (N->getOpcode() != ISD::SHL)
3421 return SDValue(); // only shl(add) not sr[al](add).
3422 HighBitSet = false; // We can only transform sra if the high bit is clear.
3426 // We require the RHS of the binop to be a constant as well.
3427 ConstantSDNode *BinOpCst = dyn_cast<ConstantSDNode>(LHS->getOperand(1));
3428 if (!BinOpCst) return SDValue();
3430 // FIXME: disable this unless the input to the binop is a shift by a constant.
3431 // If it is not a shift, it pessimizes some common cases like:
3433 // void foo(int *X, int i) { X[i & 1235] = 1; }
3434 // int bar(int *X, int i) { return X[i & 255]; }
3435 SDNode *BinOpLHSVal = LHS->getOperand(0).getNode();
3436 if ((BinOpLHSVal->getOpcode() != ISD::SHL &&
3437 BinOpLHSVal->getOpcode() != ISD::SRA &&
3438 BinOpLHSVal->getOpcode() != ISD::SRL) ||
3439 !isa<ConstantSDNode>(BinOpLHSVal->getOperand(1)))
3442 EVT VT = N->getValueType(0);
3444 // If this is a signed shift right, and the high bit is modified by the
3445 // logical operation, do not perform the transformation. The highBitSet
3446 // boolean indicates the value of the high bit of the constant which would
3447 // cause it to be modified for this operation.
3448 if (N->getOpcode() == ISD::SRA) {
3449 bool BinOpRHSSignSet = BinOpCst->getAPIntValue().isNegative();
3450 if (BinOpRHSSignSet != HighBitSet)
3454 // Fold the constants, shifting the binop RHS by the shift amount.
3455 SDValue NewRHS = DAG.getNode(N->getOpcode(), LHS->getOperand(1).getDebugLoc(),
3457 LHS->getOperand(1), N->getOperand(1));
3459 // Create the new shift.
3460 SDValue NewShift = DAG.getNode(N->getOpcode(),
3461 LHS->getOperand(0).getDebugLoc(),
3462 VT, LHS->getOperand(0), N->getOperand(1));
3464 // Create the new binop.
3465 return DAG.getNode(LHS->getOpcode(), N->getDebugLoc(), VT, NewShift, NewRHS);
3468 SDValue DAGCombiner::visitSHL(SDNode *N) {
3469 SDValue N0 = N->getOperand(0);
3470 SDValue N1 = N->getOperand(1);
3471 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3472 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3473 EVT VT = N0.getValueType();
3474 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
3476 // fold (shl c1, c2) -> c1<<c2
3478 return DAG.FoldConstantArithmetic(ISD::SHL, VT, N0C, N1C);
3479 // fold (shl 0, x) -> 0
3480 if (N0C && N0C->isNullValue())
3482 // fold (shl x, c >= size(x)) -> undef
3483 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
3484 return DAG.getUNDEF(VT);
3485 // fold (shl x, 0) -> x
3486 if (N1C && N1C->isNullValue())
3488 // fold (shl undef, x) -> 0
3489 if (N0.getOpcode() == ISD::UNDEF)
3490 return DAG.getConstant(0, VT);
3491 // if (shl x, c) is known to be zero, return 0
3492 if (DAG.MaskedValueIsZero(SDValue(N, 0),
3493 APInt::getAllOnesValue(OpSizeInBits)))
3494 return DAG.getConstant(0, VT);
3495 // fold (shl x, (trunc (and y, c))) -> (shl x, (and (trunc y), (trunc c))).
3496 if (N1.getOpcode() == ISD::TRUNCATE &&
3497 N1.getOperand(0).getOpcode() == ISD::AND &&
3498 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
3499 SDValue N101 = N1.getOperand(0).getOperand(1);
3500 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
3501 EVT TruncVT = N1.getValueType();
3502 SDValue N100 = N1.getOperand(0).getOperand(0);
3503 APInt TruncC = N101C->getAPIntValue();
3504 TruncC = TruncC.trunc(TruncVT.getSizeInBits());
3505 return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0,
3506 DAG.getNode(ISD::AND, N->getDebugLoc(), TruncVT,
3507 DAG.getNode(ISD::TRUNCATE,
3510 DAG.getConstant(TruncC, TruncVT)));
3514 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
3515 return SDValue(N, 0);
3517 // fold (shl (shl x, c1), c2) -> 0 or (shl x, (add c1, c2))
3518 if (N1C && N0.getOpcode() == ISD::SHL &&
3519 N0.getOperand(1).getOpcode() == ISD::Constant) {
3520 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
3521 uint64_t c2 = N1C->getZExtValue();
3522 if (c1 + c2 >= OpSizeInBits)
3523 return DAG.getConstant(0, VT);
3524 return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0.getOperand(0),
3525 DAG.getConstant(c1 + c2, N1.getValueType()));
3528 // fold (shl (ext (shl x, c1)), c2) -> (ext (shl x, (add c1, c2)))
3529 // For this to be valid, the second form must not preserve any of the bits
3530 // that are shifted out by the inner shift in the first form. This means
3531 // the outer shift size must be >= the number of bits added by the ext.
3532 // As a corollary, we don't care what kind of ext it is.
3533 if (N1C && (N0.getOpcode() == ISD::ZERO_EXTEND ||
3534 N0.getOpcode() == ISD::ANY_EXTEND ||
3535 N0.getOpcode() == ISD::SIGN_EXTEND) &&
3536 N0.getOperand(0).getOpcode() == ISD::SHL &&
3537 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
3539 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
3540 uint64_t c2 = N1C->getZExtValue();
3541 EVT InnerShiftVT = N0.getOperand(0).getValueType();
3542 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
3543 if (c2 >= OpSizeInBits - InnerShiftSize) {
3544 if (c1 + c2 >= OpSizeInBits)
3545 return DAG.getConstant(0, VT);
3546 return DAG.getNode(ISD::SHL, N0->getDebugLoc(), VT,
3547 DAG.getNode(N0.getOpcode(), N0->getDebugLoc(), VT,
3548 N0.getOperand(0)->getOperand(0)),
3549 DAG.getConstant(c1 + c2, N1.getValueType()));
3553 // fold (shl (srl x, c1), c2) -> (and (shl x, (sub c2, c1), MASK) or
3554 // (and (srl x, (sub c1, c2), MASK)
3555 // Only fold this if the inner shift has no other uses -- if it does, folding
3556 // this will increase the total number of instructions.
3557 if (N1C && N0.getOpcode() == ISD::SRL && N0.hasOneUse() &&
3558 N0.getOperand(1).getOpcode() == ISD::Constant) {
3559 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
3560 if (c1 < VT.getSizeInBits()) {
3561 uint64_t c2 = N1C->getZExtValue();
3562 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(),
3563 VT.getSizeInBits() - c1);
3566 Mask = Mask.shl(c2-c1);
3567 Shift = DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0.getOperand(0),
3568 DAG.getConstant(c2-c1, N1.getValueType()));
3570 Mask = Mask.lshr(c1-c2);
3571 Shift = DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0.getOperand(0),
3572 DAG.getConstant(c1-c2, N1.getValueType()));
3574 return DAG.getNode(ISD::AND, N0.getDebugLoc(), VT, Shift,
3575 DAG.getConstant(Mask, VT));
3578 // fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1))
3579 if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1)) {
3580 SDValue HiBitsMask =
3581 DAG.getConstant(APInt::getHighBitsSet(VT.getSizeInBits(),
3582 VT.getSizeInBits() -
3583 N1C->getZExtValue()),
3585 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0.getOperand(0),
3590 SDValue NewSHL = visitShiftByConstant(N, N1C->getZExtValue());
3591 if (NewSHL.getNode())
3598 SDValue DAGCombiner::visitSRA(SDNode *N) {
3599 SDValue N0 = N->getOperand(0);
3600 SDValue N1 = N->getOperand(1);
3601 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3602 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3603 EVT VT = N0.getValueType();
3604 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
3606 // fold (sra c1, c2) -> (sra c1, c2)
3608 return DAG.FoldConstantArithmetic(ISD::SRA, VT, N0C, N1C);
3609 // fold (sra 0, x) -> 0
3610 if (N0C && N0C->isNullValue())
3612 // fold (sra -1, x) -> -1
3613 if (N0C && N0C->isAllOnesValue())
3615 // fold (sra x, (setge c, size(x))) -> undef
3616 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
3617 return DAG.getUNDEF(VT);
3618 // fold (sra x, 0) -> x
3619 if (N1C && N1C->isNullValue())
3621 // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports
3623 if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) {
3624 unsigned LowBits = OpSizeInBits - (unsigned)N1C->getZExtValue();
3625 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), LowBits);
3627 ExtVT = EVT::getVectorVT(*DAG.getContext(),
3628 ExtVT, VT.getVectorNumElements());
3629 if ((!LegalOperations ||
3630 TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, ExtVT)))
3631 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT,
3632 N0.getOperand(0), DAG.getValueType(ExtVT));
3635 // fold (sra (sra x, c1), c2) -> (sra x, (add c1, c2))
3636 if (N1C && N0.getOpcode() == ISD::SRA) {
3637 if (ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
3638 unsigned Sum = N1C->getZExtValue() + C1->getZExtValue();
3639 if (Sum >= OpSizeInBits) Sum = OpSizeInBits-1;
3640 return DAG.getNode(ISD::SRA, N->getDebugLoc(), VT, N0.getOperand(0),
3641 DAG.getConstant(Sum, N1C->getValueType(0)));
3645 // fold (sra (shl X, m), (sub result_size, n))
3646 // -> (sign_extend (trunc (shl X, (sub (sub result_size, n), m)))) for
3647 // result_size - n != m.
3648 // If truncate is free for the target sext(shl) is likely to result in better
3650 if (N0.getOpcode() == ISD::SHL) {
3651 // Get the two constanst of the shifts, CN0 = m, CN = n.
3652 const ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3654 // Determine what the truncate's result bitsize and type would be.
3656 EVT::getIntegerVT(*DAG.getContext(),
3657 OpSizeInBits - N1C->getZExtValue());
3658 // Determine the residual right-shift amount.
3659 signed ShiftAmt = N1C->getZExtValue() - N01C->getZExtValue();
3661 // If the shift is not a no-op (in which case this should be just a sign
3662 // extend already), the truncated to type is legal, sign_extend is legal
3663 // on that type, and the truncate to that type is both legal and free,
3664 // perform the transform.
3665 if ((ShiftAmt > 0) &&
3666 TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND, TruncVT) &&
3667 TLI.isOperationLegalOrCustom(ISD::TRUNCATE, VT) &&
3668 TLI.isTruncateFree(VT, TruncVT)) {
3670 SDValue Amt = DAG.getConstant(ShiftAmt,
3671 getShiftAmountTy(N0.getOperand(0).getValueType()));
3672 SDValue Shift = DAG.getNode(ISD::SRL, N0.getDebugLoc(), VT,
3673 N0.getOperand(0), Amt);
3674 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(), TruncVT,
3676 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(),
3677 N->getValueType(0), Trunc);
3682 // fold (sra x, (trunc (and y, c))) -> (sra x, (and (trunc y), (trunc c))).
3683 if (N1.getOpcode() == ISD::TRUNCATE &&
3684 N1.getOperand(0).getOpcode() == ISD::AND &&
3685 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
3686 SDValue N101 = N1.getOperand(0).getOperand(1);
3687 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
3688 EVT TruncVT = N1.getValueType();
3689 SDValue N100 = N1.getOperand(0).getOperand(0);
3690 APInt TruncC = N101C->getAPIntValue();
3691 TruncC = TruncC.trunc(TruncVT.getScalarType().getSizeInBits());
3692 return DAG.getNode(ISD::SRA, N->getDebugLoc(), VT, N0,
3693 DAG.getNode(ISD::AND, N->getDebugLoc(),
3695 DAG.getNode(ISD::TRUNCATE,
3698 DAG.getConstant(TruncC, TruncVT)));
3702 // fold (sra (trunc (sr x, c1)), c2) -> (trunc (sra x, c1+c2))
3703 // if c1 is equal to the number of bits the trunc removes
3704 if (N0.getOpcode() == ISD::TRUNCATE &&
3705 (N0.getOperand(0).getOpcode() == ISD::SRL ||
3706 N0.getOperand(0).getOpcode() == ISD::SRA) &&
3707 N0.getOperand(0).hasOneUse() &&
3708 N0.getOperand(0).getOperand(1).hasOneUse() &&
3709 N1C && isa<ConstantSDNode>(N0.getOperand(0).getOperand(1))) {
3710 EVT LargeVT = N0.getOperand(0).getValueType();
3711 ConstantSDNode *LargeShiftAmt =
3712 cast<ConstantSDNode>(N0.getOperand(0).getOperand(1));
3714 if (LargeVT.getScalarType().getSizeInBits() - OpSizeInBits ==
3715 LargeShiftAmt->getZExtValue()) {
3717 DAG.getConstant(LargeShiftAmt->getZExtValue() + N1C->getZExtValue(),
3718 getShiftAmountTy(N0.getOperand(0).getOperand(0).getValueType()));
3719 SDValue SRA = DAG.getNode(ISD::SRA, N->getDebugLoc(), LargeVT,
3720 N0.getOperand(0).getOperand(0), Amt);
3721 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, SRA);
3725 // Simplify, based on bits shifted out of the LHS.
3726 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
3727 return SDValue(N, 0);
3730 // If the sign bit is known to be zero, switch this to a SRL.
3731 if (DAG.SignBitIsZero(N0))
3732 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0, N1);
3735 SDValue NewSRA = visitShiftByConstant(N, N1C->getZExtValue());
3736 if (NewSRA.getNode())
3743 SDValue DAGCombiner::visitSRL(SDNode *N) {
3744 SDValue N0 = N->getOperand(0);
3745 SDValue N1 = N->getOperand(1);
3746 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3747 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3748 EVT VT = N0.getValueType();
3749 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
3751 // fold (srl c1, c2) -> c1 >>u c2
3753 return DAG.FoldConstantArithmetic(ISD::SRL, VT, N0C, N1C);
3754 // fold (srl 0, x) -> 0
3755 if (N0C && N0C->isNullValue())
3757 // fold (srl x, c >= size(x)) -> undef
3758 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
3759 return DAG.getUNDEF(VT);
3760 // fold (srl x, 0) -> x
3761 if (N1C && N1C->isNullValue())
3763 // if (srl x, c) is known to be zero, return 0
3764 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
3765 APInt::getAllOnesValue(OpSizeInBits)))
3766 return DAG.getConstant(0, VT);
3768 // fold (srl (srl x, c1), c2) -> 0 or (srl x, (add c1, c2))
3769 if (N1C && N0.getOpcode() == ISD::SRL &&
3770 N0.getOperand(1).getOpcode() == ISD::Constant) {
3771 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
3772 uint64_t c2 = N1C->getZExtValue();
3773 if (c1 + c2 >= OpSizeInBits)
3774 return DAG.getConstant(0, VT);
3775 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0.getOperand(0),
3776 DAG.getConstant(c1 + c2, N1.getValueType()));
3779 // fold (srl (trunc (srl x, c1)), c2) -> 0 or (trunc (srl x, (add c1, c2)))
3780 if (N1C && N0.getOpcode() == ISD::TRUNCATE &&
3781 N0.getOperand(0).getOpcode() == ISD::SRL &&
3782 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
3784 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
3785 uint64_t c2 = N1C->getZExtValue();
3786 EVT InnerShiftVT = N0.getOperand(0).getValueType();
3787 EVT ShiftCountVT = N0.getOperand(0)->getOperand(1).getValueType();
3788 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
3789 // This is only valid if the OpSizeInBits + c1 = size of inner shift.
3790 if (c1 + OpSizeInBits == InnerShiftSize) {
3791 if (c1 + c2 >= InnerShiftSize)
3792 return DAG.getConstant(0, VT);
3793 return DAG.getNode(ISD::TRUNCATE, N0->getDebugLoc(), VT,
3794 DAG.getNode(ISD::SRL, N0->getDebugLoc(), InnerShiftVT,
3795 N0.getOperand(0)->getOperand(0),
3796 DAG.getConstant(c1 + c2, ShiftCountVT)));
3800 // fold (srl (shl x, c), c) -> (and x, cst2)
3801 if (N1C && N0.getOpcode() == ISD::SHL && N0.getOperand(1) == N1 &&
3802 N0.getValueSizeInBits() <= 64) {
3803 uint64_t ShAmt = N1C->getZExtValue()+64-N0.getValueSizeInBits();
3804 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0.getOperand(0),
3805 DAG.getConstant(~0ULL >> ShAmt, VT));
3809 // fold (srl (anyextend x), c) -> (anyextend (srl x, c))
3810 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
3811 // Shifting in all undef bits?
3812 EVT SmallVT = N0.getOperand(0).getValueType();
3813 if (N1C->getZExtValue() >= SmallVT.getSizeInBits())
3814 return DAG.getUNDEF(VT);
3816 if (!LegalTypes || TLI.isTypeDesirableForOp(ISD::SRL, SmallVT)) {
3817 uint64_t ShiftAmt = N1C->getZExtValue();
3818 SDValue SmallShift = DAG.getNode(ISD::SRL, N0.getDebugLoc(), SmallVT,
3820 DAG.getConstant(ShiftAmt, getShiftAmountTy(SmallVT)));
3821 AddToWorkList(SmallShift.getNode());
3822 return DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, SmallShift);
3826 // fold (srl (sra X, Y), 31) -> (srl X, 31). This srl only looks at the sign
3827 // bit, which is unmodified by sra.
3828 if (N1C && N1C->getZExtValue() + 1 == VT.getSizeInBits()) {
3829 if (N0.getOpcode() == ISD::SRA)
3830 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0.getOperand(0), N1);
3833 // fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit).
3834 if (N1C && N0.getOpcode() == ISD::CTLZ &&
3835 N1C->getAPIntValue() == Log2_32(VT.getSizeInBits())) {
3836 APInt KnownZero, KnownOne;
3837 DAG.ComputeMaskedBits(N0.getOperand(0), KnownZero, KnownOne);
3839 // If any of the input bits are KnownOne, then the input couldn't be all
3840 // zeros, thus the result of the srl will always be zero.
3841 if (KnownOne.getBoolValue()) return DAG.getConstant(0, VT);
3843 // If all of the bits input the to ctlz node are known to be zero, then
3844 // the result of the ctlz is "32" and the result of the shift is one.
3845 APInt UnknownBits = ~KnownZero;
3846 if (UnknownBits == 0) return DAG.getConstant(1, VT);
3848 // Otherwise, check to see if there is exactly one bit input to the ctlz.
3849 if ((UnknownBits & (UnknownBits - 1)) == 0) {
3850 // Okay, we know that only that the single bit specified by UnknownBits
3851 // could be set on input to the CTLZ node. If this bit is set, the SRL
3852 // will return 0, if it is clear, it returns 1. Change the CTLZ/SRL pair
3853 // to an SRL/XOR pair, which is likely to simplify more.
3854 unsigned ShAmt = UnknownBits.countTrailingZeros();
3855 SDValue Op = N0.getOperand(0);
3858 Op = DAG.getNode(ISD::SRL, N0.getDebugLoc(), VT, Op,
3859 DAG.getConstant(ShAmt, getShiftAmountTy(Op.getValueType())));
3860 AddToWorkList(Op.getNode());
3863 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT,
3864 Op, DAG.getConstant(1, VT));
3868 // fold (srl x, (trunc (and y, c))) -> (srl x, (and (trunc y), (trunc c))).
3869 if (N1.getOpcode() == ISD::TRUNCATE &&
3870 N1.getOperand(0).getOpcode() == ISD::AND &&
3871 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
3872 SDValue N101 = N1.getOperand(0).getOperand(1);
3873 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
3874 EVT TruncVT = N1.getValueType();
3875 SDValue N100 = N1.getOperand(0).getOperand(0);
3876 APInt TruncC = N101C->getAPIntValue();
3877 TruncC = TruncC.trunc(TruncVT.getSizeInBits());
3878 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0,
3879 DAG.getNode(ISD::AND, N->getDebugLoc(),
3881 DAG.getNode(ISD::TRUNCATE,
3884 DAG.getConstant(TruncC, TruncVT)));
3888 // fold operands of srl based on knowledge that the low bits are not
3890 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
3891 return SDValue(N, 0);
3894 SDValue NewSRL = visitShiftByConstant(N, N1C->getZExtValue());
3895 if (NewSRL.getNode())
3899 // Attempt to convert a srl of a load into a narrower zero-extending load.
3900 SDValue NarrowLoad = ReduceLoadWidth(N);
3901 if (NarrowLoad.getNode())
3904 // Here is a common situation. We want to optimize:
3907 // %b = and i32 %a, 2
3908 // %c = srl i32 %b, 1
3909 // brcond i32 %c ...
3915 // %c = setcc eq %b, 0
3918 // However when after the source operand of SRL is optimized into AND, the SRL
3919 // itself may not be optimized further. Look for it and add the BRCOND into
3921 if (N->hasOneUse()) {
3922 SDNode *Use = *N->use_begin();
3923 if (Use->getOpcode() == ISD::BRCOND)
3925 else if (Use->getOpcode() == ISD::TRUNCATE && Use->hasOneUse()) {
3926 // Also look pass the truncate.
3927 Use = *Use->use_begin();
3928 if (Use->getOpcode() == ISD::BRCOND)
3936 SDValue DAGCombiner::visitCTLZ(SDNode *N) {
3937 SDValue N0 = N->getOperand(0);
3938 EVT VT = N->getValueType(0);
3940 // fold (ctlz c1) -> c2
3941 if (isa<ConstantSDNode>(N0))
3942 return DAG.getNode(ISD::CTLZ, N->getDebugLoc(), VT, N0);
3946 SDValue DAGCombiner::visitCTLZ_ZERO_UNDEF(SDNode *N) {
3947 SDValue N0 = N->getOperand(0);
3948 EVT VT = N->getValueType(0);
3950 // fold (ctlz_zero_undef c1) -> c2
3951 if (isa<ConstantSDNode>(N0))
3952 return DAG.getNode(ISD::CTLZ_ZERO_UNDEF, N->getDebugLoc(), VT, N0);
3956 SDValue DAGCombiner::visitCTTZ(SDNode *N) {
3957 SDValue N0 = N->getOperand(0);
3958 EVT VT = N->getValueType(0);
3960 // fold (cttz c1) -> c2
3961 if (isa<ConstantSDNode>(N0))
3962 return DAG.getNode(ISD::CTTZ, N->getDebugLoc(), VT, N0);
3966 SDValue DAGCombiner::visitCTTZ_ZERO_UNDEF(SDNode *N) {
3967 SDValue N0 = N->getOperand(0);
3968 EVT VT = N->getValueType(0);
3970 // fold (cttz_zero_undef c1) -> c2
3971 if (isa<ConstantSDNode>(N0))
3972 return DAG.getNode(ISD::CTTZ_ZERO_UNDEF, N->getDebugLoc(), VT, N0);
3976 SDValue DAGCombiner::visitCTPOP(SDNode *N) {
3977 SDValue N0 = N->getOperand(0);
3978 EVT VT = N->getValueType(0);
3980 // fold (ctpop c1) -> c2
3981 if (isa<ConstantSDNode>(N0))
3982 return DAG.getNode(ISD::CTPOP, N->getDebugLoc(), VT, N0);
3986 SDValue DAGCombiner::visitSELECT(SDNode *N) {
3987 SDValue N0 = N->getOperand(0);
3988 SDValue N1 = N->getOperand(1);
3989 SDValue N2 = N->getOperand(2);
3990 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3991 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3992 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
3993 EVT VT = N->getValueType(0);
3994 EVT VT0 = N0.getValueType();
3996 // fold (select C, X, X) -> X
3999 // fold (select true, X, Y) -> X
4000 if (N0C && !N0C->isNullValue())
4002 // fold (select false, X, Y) -> Y
4003 if (N0C && N0C->isNullValue())
4005 // fold (select C, 1, X) -> (or C, X)
4006 if (VT == MVT::i1 && N1C && N1C->getAPIntValue() == 1)
4007 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N0, N2);
4008 // fold (select C, 0, 1) -> (xor C, 1)
4009 if (VT.isInteger() &&
4012 TLI.getBooleanContents(false) == TargetLowering::ZeroOrOneBooleanContent)) &&
4013 N1C && N2C && N1C->isNullValue() && N2C->getAPIntValue() == 1) {
4016 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT0,
4017 N0, DAG.getConstant(1, VT0));
4018 XORNode = DAG.getNode(ISD::XOR, N0.getDebugLoc(), VT0,
4019 N0, DAG.getConstant(1, VT0));
4020 AddToWorkList(XORNode.getNode());
4022 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, XORNode);
4023 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, XORNode);
4025 // fold (select C, 0, X) -> (and (not C), X)
4026 if (VT == VT0 && VT == MVT::i1 && N1C && N1C->isNullValue()) {
4027 SDValue NOTNode = DAG.getNOT(N0.getDebugLoc(), N0, VT);
4028 AddToWorkList(NOTNode.getNode());
4029 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, NOTNode, N2);
4031 // fold (select C, X, 1) -> (or (not C), X)
4032 if (VT == VT0 && VT == MVT::i1 && N2C && N2C->getAPIntValue() == 1) {
4033 SDValue NOTNode = DAG.getNOT(N0.getDebugLoc(), N0, VT);
4034 AddToWorkList(NOTNode.getNode());
4035 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, NOTNode, N1);
4037 // fold (select C, X, 0) -> (and C, X)
4038 if (VT == MVT::i1 && N2C && N2C->isNullValue())
4039 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0, N1);
4040 // fold (select X, X, Y) -> (or X, Y)
4041 // fold (select X, 1, Y) -> (or X, Y)
4042 if (VT == MVT::i1 && (N0 == N1 || (N1C && N1C->getAPIntValue() == 1)))
4043 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N0, N2);
4044 // fold (select X, Y, X) -> (and X, Y)
4045 // fold (select X, Y, 0) -> (and X, Y)
4046 if (VT == MVT::i1 && (N0 == N2 || (N2C && N2C->getAPIntValue() == 0)))
4047 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0, N1);
4049 // If we can fold this based on the true/false value, do so.
4050 if (SimplifySelectOps(N, N1, N2))
4051 return SDValue(N, 0); // Don't revisit N.
4053 // fold selects based on a setcc into other things, such as min/max/abs
4054 if (N0.getOpcode() == ISD::SETCC) {
4056 // Check against MVT::Other for SELECT_CC, which is a workaround for targets
4057 // having to say they don't support SELECT_CC on every type the DAG knows
4058 // about, since there is no way to mark an opcode illegal at all value types
4059 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, MVT::Other) &&
4060 TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT))
4061 return DAG.getNode(ISD::SELECT_CC, N->getDebugLoc(), VT,
4062 N0.getOperand(0), N0.getOperand(1),
4063 N1, N2, N0.getOperand(2));
4064 return SimplifySelect(N->getDebugLoc(), N0, N1, N2);
4070 SDValue DAGCombiner::visitSELECT_CC(SDNode *N) {
4071 SDValue N0 = N->getOperand(0);
4072 SDValue N1 = N->getOperand(1);
4073 SDValue N2 = N->getOperand(2);
4074 SDValue N3 = N->getOperand(3);
4075 SDValue N4 = N->getOperand(4);
4076 ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get();
4078 // fold select_cc lhs, rhs, x, x, cc -> x
4082 // Determine if the condition we're dealing with is constant
4083 SDValue SCC = SimplifySetCC(TLI.getSetCCResultType(N0.getValueType()),
4084 N0, N1, CC, N->getDebugLoc(), false);
4085 if (SCC.getNode()) AddToWorkList(SCC.getNode());
4087 if (ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode())) {
4088 if (!SCCC->isNullValue())
4089 return N2; // cond always true -> true val
4091 return N3; // cond always false -> false val
4094 // Fold to a simpler select_cc
4095 if (SCC.getNode() && SCC.getOpcode() == ISD::SETCC)
4096 return DAG.getNode(ISD::SELECT_CC, N->getDebugLoc(), N2.getValueType(),
4097 SCC.getOperand(0), SCC.getOperand(1), N2, N3,
4100 // If we can fold this based on the true/false value, do so.
4101 if (SimplifySelectOps(N, N2, N3))
4102 return SDValue(N, 0); // Don't revisit N.
4104 // fold select_cc into other things, such as min/max/abs
4105 return SimplifySelectCC(N->getDebugLoc(), N0, N1, N2, N3, CC);
4108 SDValue DAGCombiner::visitSETCC(SDNode *N) {
4109 return SimplifySetCC(N->getValueType(0), N->getOperand(0), N->getOperand(1),
4110 cast<CondCodeSDNode>(N->getOperand(2))->get(),
4114 // ExtendUsesToFormExtLoad - Trying to extend uses of a load to enable this:
4115 // "fold ({s|z|a}ext (load x)) -> ({s|z|a}ext (truncate ({s|z|a}extload x)))"
4116 // transformation. Returns true if extension are possible and the above
4117 // mentioned transformation is profitable.
4118 static bool ExtendUsesToFormExtLoad(SDNode *N, SDValue N0,
4120 SmallVector<SDNode*, 4> &ExtendNodes,
4121 const TargetLowering &TLI) {
4122 bool HasCopyToRegUses = false;
4123 bool isTruncFree = TLI.isTruncateFree(N->getValueType(0), N0.getValueType());
4124 for (SDNode::use_iterator UI = N0.getNode()->use_begin(),
4125 UE = N0.getNode()->use_end();
4130 if (UI.getUse().getResNo() != N0.getResNo())
4132 // FIXME: Only extend SETCC N, N and SETCC N, c for now.
4133 if (ExtOpc != ISD::ANY_EXTEND && User->getOpcode() == ISD::SETCC) {
4134 ISD::CondCode CC = cast<CondCodeSDNode>(User->getOperand(2))->get();
4135 if (ExtOpc == ISD::ZERO_EXTEND && ISD::isSignedIntSetCC(CC))
4136 // Sign bits will be lost after a zext.
4139 for (unsigned i = 0; i != 2; ++i) {
4140 SDValue UseOp = User->getOperand(i);
4143 if (!isa<ConstantSDNode>(UseOp))
4148 ExtendNodes.push_back(User);
4151 // If truncates aren't free and there are users we can't
4152 // extend, it isn't worthwhile.
4155 // Remember if this value is live-out.
4156 if (User->getOpcode() == ISD::CopyToReg)
4157 HasCopyToRegUses = true;
4160 if (HasCopyToRegUses) {
4161 bool BothLiveOut = false;
4162 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
4164 SDUse &Use = UI.getUse();
4165 if (Use.getResNo() == 0 && Use.getUser()->getOpcode() == ISD::CopyToReg) {
4171 // Both unextended and extended values are live out. There had better be
4172 // a good reason for the transformation.
4173 return ExtendNodes.size();
4178 void DAGCombiner::ExtendSetCCUses(SmallVector<SDNode*, 4> SetCCs,
4179 SDValue Trunc, SDValue ExtLoad, DebugLoc DL,
4180 ISD::NodeType ExtType) {
4181 // Extend SetCC uses if necessary.
4182 for (unsigned i = 0, e = SetCCs.size(); i != e; ++i) {
4183 SDNode *SetCC = SetCCs[i];
4184 SmallVector<SDValue, 4> Ops;
4186 for (unsigned j = 0; j != 2; ++j) {
4187 SDValue SOp = SetCC->getOperand(j);
4189 Ops.push_back(ExtLoad);
4191 Ops.push_back(DAG.getNode(ExtType, DL, ExtLoad->getValueType(0), SOp));
4194 Ops.push_back(SetCC->getOperand(2));
4195 CombineTo(SetCC, DAG.getNode(ISD::SETCC, DL, SetCC->getValueType(0),
4196 &Ops[0], Ops.size()));
4200 SDValue DAGCombiner::visitSIGN_EXTEND(SDNode *N) {
4201 SDValue N0 = N->getOperand(0);
4202 EVT VT = N->getValueType(0);
4204 // fold (sext c1) -> c1
4205 if (isa<ConstantSDNode>(N0))
4206 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, N0);
4208 // fold (sext (sext x)) -> (sext x)
4209 // fold (sext (aext x)) -> (sext x)
4210 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
4211 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT,
4214 if (N0.getOpcode() == ISD::TRUNCATE) {
4215 // fold (sext (truncate (load x))) -> (sext (smaller load x))
4216 // fold (sext (truncate (srl (load x), c))) -> (sext (smaller load (x+c/n)))
4217 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
4218 if (NarrowLoad.getNode()) {
4219 SDNode* oye = N0.getNode()->getOperand(0).getNode();
4220 if (NarrowLoad.getNode() != N0.getNode()) {
4221 CombineTo(N0.getNode(), NarrowLoad);
4222 // CombineTo deleted the truncate, if needed, but not what's under it.
4225 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4228 // See if the value being truncated is already sign extended. If so, just
4229 // eliminate the trunc/sext pair.
4230 SDValue Op = N0.getOperand(0);
4231 unsigned OpBits = Op.getValueType().getScalarType().getSizeInBits();
4232 unsigned MidBits = N0.getValueType().getScalarType().getSizeInBits();
4233 unsigned DestBits = VT.getScalarType().getSizeInBits();
4234 unsigned NumSignBits = DAG.ComputeNumSignBits(Op);
4236 if (OpBits == DestBits) {
4237 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
4238 // bits, it is already ready.
4239 if (NumSignBits > DestBits-MidBits)
4241 } else if (OpBits < DestBits) {
4242 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
4243 // bits, just sext from i32.
4244 if (NumSignBits > OpBits-MidBits)
4245 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, Op);
4247 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
4248 // bits, just truncate to i32.
4249 if (NumSignBits > OpBits-MidBits)
4250 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, Op);
4253 // fold (sext (truncate x)) -> (sextinreg x).
4254 if (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG,
4255 N0.getValueType())) {
4256 if (OpBits < DestBits)
4257 Op = DAG.getNode(ISD::ANY_EXTEND, N0.getDebugLoc(), VT, Op);
4258 else if (OpBits > DestBits)
4259 Op = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(), VT, Op);
4260 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT, Op,
4261 DAG.getValueType(N0.getValueType()));
4265 // fold (sext (load x)) -> (sext (truncate (sextload x)))
4266 // None of the supported targets knows how to perform load and sign extend
4267 // on vectors in one instruction. We only perform this transformation on
4269 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
4270 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
4271 TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()))) {
4272 bool DoXform = true;
4273 SmallVector<SDNode*, 4> SetCCs;
4274 if (!N0.hasOneUse())
4275 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::SIGN_EXTEND, SetCCs, TLI);
4277 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4278 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
4280 LN0->getBasePtr(), LN0->getPointerInfo(),
4282 LN0->isVolatile(), LN0->isNonTemporal(),
4283 LN0->getAlignment());
4284 CombineTo(N, ExtLoad);
4285 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4286 N0.getValueType(), ExtLoad);
4287 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
4288 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, N->getDebugLoc(),
4290 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4294 // fold (sext (sextload x)) -> (sext (truncate (sextload x)))
4295 // fold (sext ( extload x)) -> (sext (truncate (sextload x)))
4296 if ((ISD::isSEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
4297 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
4298 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4299 EVT MemVT = LN0->getMemoryVT();
4300 if ((!LegalOperations && !LN0->isVolatile()) ||
4301 TLI.isLoadExtLegal(ISD::SEXTLOAD, MemVT)) {
4302 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
4304 LN0->getBasePtr(), LN0->getPointerInfo(),
4306 LN0->isVolatile(), LN0->isNonTemporal(),
4307 LN0->getAlignment());
4308 CombineTo(N, ExtLoad);
4309 CombineTo(N0.getNode(),
4310 DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4311 N0.getValueType(), ExtLoad),
4312 ExtLoad.getValue(1));
4313 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4317 // fold (sext (and/or/xor (load x), cst)) ->
4318 // (and/or/xor (sextload x), (sext cst))
4319 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
4320 N0.getOpcode() == ISD::XOR) &&
4321 isa<LoadSDNode>(N0.getOperand(0)) &&
4322 N0.getOperand(1).getOpcode() == ISD::Constant &&
4323 TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()) &&
4324 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
4325 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
4326 if (LN0->getExtensionType() != ISD::ZEXTLOAD) {
4327 bool DoXform = true;
4328 SmallVector<SDNode*, 4> SetCCs;
4329 if (!N0.hasOneUse())
4330 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::SIGN_EXTEND,
4333 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, LN0->getDebugLoc(), VT,
4334 LN0->getChain(), LN0->getBasePtr(),
4335 LN0->getPointerInfo(),
4338 LN0->isNonTemporal(),
4339 LN0->getAlignment());
4340 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
4341 Mask = Mask.sext(VT.getSizeInBits());
4342 SDValue And = DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT,
4343 ExtLoad, DAG.getConstant(Mask, VT));
4344 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
4345 N0.getOperand(0).getDebugLoc(),
4346 N0.getOperand(0).getValueType(), ExtLoad);
4348 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
4349 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, N->getDebugLoc(),
4351 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4356 if (N0.getOpcode() == ISD::SETCC) {
4357 // sext(setcc) -> sext_in_reg(vsetcc) for vectors.
4358 // Only do this before legalize for now.
4359 if (VT.isVector() && !LegalOperations) {
4360 EVT N0VT = N0.getOperand(0).getValueType();
4361 // On some architectures (such as SSE/NEON/etc) the SETCC result type is
4362 // of the same size as the compared operands. Only optimize sext(setcc())
4363 // if this is the case.
4364 EVT SVT = TLI.getSetCCResultType(N0VT);
4366 // We know that the # elements of the results is the same as the
4367 // # elements of the compare (and the # elements of the compare result
4368 // for that matter). Check to see that they are the same size. If so,
4369 // we know that the element size of the sext'd result matches the
4370 // element size of the compare operands.
4371 if (VT.getSizeInBits() == SVT.getSizeInBits())
4372 return DAG.getSetCC(N->getDebugLoc(), VT, N0.getOperand(0),
4374 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4375 // If the desired elements are smaller or larger than the source
4376 // elements we can use a matching integer vector type and then
4377 // truncate/sign extend
4379 EVT MatchingElementType =
4380 EVT::getIntegerVT(*DAG.getContext(),
4381 N0VT.getScalarType().getSizeInBits());
4382 EVT MatchingVectorType =
4383 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
4384 N0VT.getVectorNumElements());
4386 if (SVT == MatchingVectorType) {
4387 SDValue VsetCC = DAG.getSetCC(N->getDebugLoc(), MatchingVectorType,
4388 N0.getOperand(0), N0.getOperand(1),
4389 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4390 return DAG.getSExtOrTrunc(VsetCC, N->getDebugLoc(), VT);
4395 // sext(setcc x, y, cc) -> (select_cc x, y, -1, 0, cc)
4396 unsigned ElementWidth = VT.getScalarType().getSizeInBits();
4398 DAG.getConstant(APInt::getAllOnesValue(ElementWidth), VT);
4400 SimplifySelectCC(N->getDebugLoc(), N0.getOperand(0), N0.getOperand(1),
4401 NegOne, DAG.getConstant(0, VT),
4402 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
4403 if (SCC.getNode()) return SCC;
4404 if (!LegalOperations ||
4405 TLI.isOperationLegal(ISD::SETCC, TLI.getSetCCResultType(VT)))
4406 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
4407 DAG.getSetCC(N->getDebugLoc(),
4408 TLI.getSetCCResultType(VT),
4409 N0.getOperand(0), N0.getOperand(1),
4410 cast<CondCodeSDNode>(N0.getOperand(2))->get()),
4411 NegOne, DAG.getConstant(0, VT));
4414 // fold (sext x) -> (zext x) if the sign bit is known zero.
4415 if ((!LegalOperations || TLI.isOperationLegal(ISD::ZERO_EXTEND, VT)) &&
4416 DAG.SignBitIsZero(N0))
4417 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, N0);
4422 // isTruncateOf - If N is a truncate of some other value, return true, record
4423 // the value being truncated in Op and which of Op's bits are zero in KnownZero.
4424 // This function computes KnownZero to avoid a duplicated call to
4425 // ComputeMaskedBits in the caller.
4426 static bool isTruncateOf(SelectionDAG &DAG, SDValue N, SDValue &Op,
4429 if (N->getOpcode() == ISD::TRUNCATE) {
4430 Op = N->getOperand(0);
4431 DAG.ComputeMaskedBits(Op, KnownZero, KnownOne);
4435 if (N->getOpcode() != ISD::SETCC || N->getValueType(0) != MVT::i1 ||
4436 cast<CondCodeSDNode>(N->getOperand(2))->get() != ISD::SETNE)
4439 SDValue Op0 = N->getOperand(0);
4440 SDValue Op1 = N->getOperand(1);
4441 assert(Op0.getValueType() == Op1.getValueType());
4443 ConstantSDNode *COp0 = dyn_cast<ConstantSDNode>(Op0);
4444 ConstantSDNode *COp1 = dyn_cast<ConstantSDNode>(Op1);
4445 if (COp0 && COp0->isNullValue())
4447 else if (COp1 && COp1->isNullValue())
4452 DAG.ComputeMaskedBits(Op, KnownZero, KnownOne);
4454 if (!(KnownZero | APInt(Op.getValueSizeInBits(), 1)).isAllOnesValue())
4460 SDValue DAGCombiner::visitZERO_EXTEND(SDNode *N) {
4461 SDValue N0 = N->getOperand(0);
4462 EVT VT = N->getValueType(0);
4464 // fold (zext c1) -> c1
4465 if (isa<ConstantSDNode>(N0))
4466 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, N0);
4467 // fold (zext (zext x)) -> (zext x)
4468 // fold (zext (aext x)) -> (zext x)
4469 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
4470 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT,
4473 // fold (zext (truncate x)) -> (zext x) or
4474 // (zext (truncate x)) -> (truncate x)
4475 // This is valid when the truncated bits of x are already zero.
4476 // FIXME: We should extend this to work for vectors too.
4479 if (!VT.isVector() && isTruncateOf(DAG, N0, Op, KnownZero)) {
4480 APInt TruncatedBits =
4481 (Op.getValueSizeInBits() == N0.getValueSizeInBits()) ?
4482 APInt(Op.getValueSizeInBits(), 0) :
4483 APInt::getBitsSet(Op.getValueSizeInBits(),
4484 N0.getValueSizeInBits(),
4485 std::min(Op.getValueSizeInBits(),
4486 VT.getSizeInBits()));
4487 if (TruncatedBits == (KnownZero & TruncatedBits)) {
4488 if (VT.bitsGT(Op.getValueType()))
4489 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, Op);
4490 if (VT.bitsLT(Op.getValueType()))
4491 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, Op);
4497 // fold (zext (truncate (load x))) -> (zext (smaller load x))
4498 // fold (zext (truncate (srl (load x), c))) -> (zext (small load (x+c/n)))
4499 if (N0.getOpcode() == ISD::TRUNCATE) {
4500 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
4501 if (NarrowLoad.getNode()) {
4502 SDNode* oye = N0.getNode()->getOperand(0).getNode();
4503 if (NarrowLoad.getNode() != N0.getNode()) {
4504 CombineTo(N0.getNode(), NarrowLoad);
4505 // CombineTo deleted the truncate, if needed, but not what's under it.
4508 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4512 // fold (zext (truncate x)) -> (and x, mask)
4513 if (N0.getOpcode() == ISD::TRUNCATE &&
4514 (!LegalOperations || TLI.isOperationLegal(ISD::AND, VT))) {
4516 // fold (zext (truncate (load x))) -> (zext (smaller load x))
4517 // fold (zext (truncate (srl (load x), c))) -> (zext (smaller load (x+c/n)))
4518 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
4519 if (NarrowLoad.getNode()) {
4520 SDNode* oye = N0.getNode()->getOperand(0).getNode();
4521 if (NarrowLoad.getNode() != N0.getNode()) {
4522 CombineTo(N0.getNode(), NarrowLoad);
4523 // CombineTo deleted the truncate, if needed, but not what's under it.
4526 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4529 SDValue Op = N0.getOperand(0);
4530 if (Op.getValueType().bitsLT(VT)) {
4531 Op = DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, Op);
4532 AddToWorkList(Op.getNode());
4533 } else if (Op.getValueType().bitsGT(VT)) {
4534 Op = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, Op);
4535 AddToWorkList(Op.getNode());
4537 return DAG.getZeroExtendInReg(Op, N->getDebugLoc(),
4538 N0.getValueType().getScalarType());
4541 // Fold (zext (and (trunc x), cst)) -> (and x, cst),
4542 // if either of the casts is not free.
4543 if (N0.getOpcode() == ISD::AND &&
4544 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
4545 N0.getOperand(1).getOpcode() == ISD::Constant &&
4546 (!TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
4547 N0.getValueType()) ||
4548 !TLI.isZExtFree(N0.getValueType(), VT))) {
4549 SDValue X = N0.getOperand(0).getOperand(0);
4550 if (X.getValueType().bitsLT(VT)) {
4551 X = DAG.getNode(ISD::ANY_EXTEND, X.getDebugLoc(), VT, X);
4552 } else if (X.getValueType().bitsGT(VT)) {
4553 X = DAG.getNode(ISD::TRUNCATE, X.getDebugLoc(), VT, X);
4555 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
4556 Mask = Mask.zext(VT.getSizeInBits());
4557 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
4558 X, DAG.getConstant(Mask, VT));
4561 // fold (zext (load x)) -> (zext (truncate (zextload x)))
4562 // None of the supported targets knows how to perform load and vector_zext
4563 // on vectors in one instruction. We only perform this transformation on
4565 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
4566 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
4567 TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()))) {
4568 bool DoXform = true;
4569 SmallVector<SDNode*, 4> SetCCs;
4570 if (!N0.hasOneUse())
4571 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ZERO_EXTEND, SetCCs, TLI);
4573 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4574 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N->getDebugLoc(), VT,
4576 LN0->getBasePtr(), LN0->getPointerInfo(),
4578 LN0->isVolatile(), LN0->isNonTemporal(),
4579 LN0->getAlignment());
4580 CombineTo(N, ExtLoad);
4581 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4582 N0.getValueType(), ExtLoad);
4583 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
4585 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, N->getDebugLoc(),
4587 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4591 // fold (zext (and/or/xor (load x), cst)) ->
4592 // (and/or/xor (zextload x), (zext cst))
4593 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
4594 N0.getOpcode() == ISD::XOR) &&
4595 isa<LoadSDNode>(N0.getOperand(0)) &&
4596 N0.getOperand(1).getOpcode() == ISD::Constant &&
4597 TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()) &&
4598 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
4599 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
4600 if (LN0->getExtensionType() != ISD::SEXTLOAD) {
4601 bool DoXform = true;
4602 SmallVector<SDNode*, 4> SetCCs;
4603 if (!N0.hasOneUse())
4604 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::ZERO_EXTEND,
4607 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, LN0->getDebugLoc(), VT,
4608 LN0->getChain(), LN0->getBasePtr(),
4609 LN0->getPointerInfo(),
4612 LN0->isNonTemporal(),
4613 LN0->getAlignment());
4614 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
4615 Mask = Mask.zext(VT.getSizeInBits());
4616 SDValue And = DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT,
4617 ExtLoad, DAG.getConstant(Mask, VT));
4618 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
4619 N0.getOperand(0).getDebugLoc(),
4620 N0.getOperand(0).getValueType(), ExtLoad);
4622 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
4623 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, N->getDebugLoc(),
4625 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4630 // fold (zext (zextload x)) -> (zext (truncate (zextload x)))
4631 // fold (zext ( extload x)) -> (zext (truncate (zextload x)))
4632 if ((ISD::isZEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
4633 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
4634 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4635 EVT MemVT = LN0->getMemoryVT();
4636 if ((!LegalOperations && !LN0->isVolatile()) ||
4637 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT)) {
4638 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N->getDebugLoc(), VT,
4640 LN0->getBasePtr(), LN0->getPointerInfo(),
4642 LN0->isVolatile(), LN0->isNonTemporal(),
4643 LN0->getAlignment());
4644 CombineTo(N, ExtLoad);
4645 CombineTo(N0.getNode(),
4646 DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(), N0.getValueType(),
4648 ExtLoad.getValue(1));
4649 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4653 if (N0.getOpcode() == ISD::SETCC) {
4654 if (!LegalOperations && VT.isVector()) {
4655 // zext(setcc) -> (and (vsetcc), (1, 1, ...) for vectors.
4656 // Only do this before legalize for now.
4657 EVT N0VT = N0.getOperand(0).getValueType();
4658 EVT EltVT = VT.getVectorElementType();
4659 SmallVector<SDValue,8> OneOps(VT.getVectorNumElements(),
4660 DAG.getConstant(1, EltVT));
4661 if (VT.getSizeInBits() == N0VT.getSizeInBits())
4662 // We know that the # elements of the results is the same as the
4663 // # elements of the compare (and the # elements of the compare result
4664 // for that matter). Check to see that they are the same size. If so,
4665 // we know that the element size of the sext'd result matches the
4666 // element size of the compare operands.
4667 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
4668 DAG.getSetCC(N->getDebugLoc(), VT, N0.getOperand(0),
4670 cast<CondCodeSDNode>(N0.getOperand(2))->get()),
4671 DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), VT,
4672 &OneOps[0], OneOps.size()));
4674 // If the desired elements are smaller or larger than the source
4675 // elements we can use a matching integer vector type and then
4676 // truncate/sign extend
4677 EVT MatchingElementType =
4678 EVT::getIntegerVT(*DAG.getContext(),
4679 N0VT.getScalarType().getSizeInBits());
4680 EVT MatchingVectorType =
4681 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
4682 N0VT.getVectorNumElements());
4684 DAG.getSetCC(N->getDebugLoc(), MatchingVectorType, N0.getOperand(0),
4686 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4687 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
4688 DAG.getSExtOrTrunc(VsetCC, N->getDebugLoc(), VT),
4689 DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), VT,
4690 &OneOps[0], OneOps.size()));
4693 // zext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
4695 SimplifySelectCC(N->getDebugLoc(), N0.getOperand(0), N0.getOperand(1),
4696 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
4697 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
4698 if (SCC.getNode()) return SCC;
4701 // (zext (shl (zext x), cst)) -> (shl (zext x), cst)
4702 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL) &&
4703 isa<ConstantSDNode>(N0.getOperand(1)) &&
4704 N0.getOperand(0).getOpcode() == ISD::ZERO_EXTEND &&
4706 SDValue ShAmt = N0.getOperand(1);
4707 unsigned ShAmtVal = cast<ConstantSDNode>(ShAmt)->getZExtValue();
4708 if (N0.getOpcode() == ISD::SHL) {
4709 SDValue InnerZExt = N0.getOperand(0);
4710 // If the original shl may be shifting out bits, do not perform this
4712 unsigned KnownZeroBits = InnerZExt.getValueType().getSizeInBits() -
4713 InnerZExt.getOperand(0).getValueType().getSizeInBits();
4714 if (ShAmtVal > KnownZeroBits)
4718 DebugLoc DL = N->getDebugLoc();
4720 // Ensure that the shift amount is wide enough for the shifted value.
4721 if (VT.getSizeInBits() >= 256)
4722 ShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, ShAmt);
4724 return DAG.getNode(N0.getOpcode(), DL, VT,
4725 DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0)),
4732 SDValue DAGCombiner::visitANY_EXTEND(SDNode *N) {
4733 SDValue N0 = N->getOperand(0);
4734 EVT VT = N->getValueType(0);
4736 // fold (aext c1) -> c1
4737 if (isa<ConstantSDNode>(N0))
4738 return DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, N0);
4739 // fold (aext (aext x)) -> (aext x)
4740 // fold (aext (zext x)) -> (zext x)
4741 // fold (aext (sext x)) -> (sext x)
4742 if (N0.getOpcode() == ISD::ANY_EXTEND ||
4743 N0.getOpcode() == ISD::ZERO_EXTEND ||
4744 N0.getOpcode() == ISD::SIGN_EXTEND)
4745 return DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT, N0.getOperand(0));
4747 // fold (aext (truncate (load x))) -> (aext (smaller load x))
4748 // fold (aext (truncate (srl (load x), c))) -> (aext (small load (x+c/n)))
4749 if (N0.getOpcode() == ISD::TRUNCATE) {
4750 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
4751 if (NarrowLoad.getNode()) {
4752 SDNode* oye = N0.getNode()->getOperand(0).getNode();
4753 if (NarrowLoad.getNode() != N0.getNode()) {
4754 CombineTo(N0.getNode(), NarrowLoad);
4755 // CombineTo deleted the truncate, if needed, but not what's under it.
4758 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4762 // fold (aext (truncate x))
4763 if (N0.getOpcode() == ISD::TRUNCATE) {
4764 SDValue TruncOp = N0.getOperand(0);
4765 if (TruncOp.getValueType() == VT)
4766 return TruncOp; // x iff x size == zext size.
4767 if (TruncOp.getValueType().bitsGT(VT))
4768 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, TruncOp);
4769 return DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, TruncOp);
4772 // Fold (aext (and (trunc x), cst)) -> (and x, cst)
4773 // if the trunc is not free.
4774 if (N0.getOpcode() == ISD::AND &&
4775 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
4776 N0.getOperand(1).getOpcode() == ISD::Constant &&
4777 !TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
4778 N0.getValueType())) {
4779 SDValue X = N0.getOperand(0).getOperand(0);
4780 if (X.getValueType().bitsLT(VT)) {
4781 X = DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, X);
4782 } else if (X.getValueType().bitsGT(VT)) {
4783 X = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, X);
4785 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
4786 Mask = Mask.zext(VT.getSizeInBits());
4787 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
4788 X, DAG.getConstant(Mask, VT));
4791 // fold (aext (load x)) -> (aext (truncate (extload x)))
4792 // None of the supported targets knows how to perform load and any_ext
4793 // on vectors in one instruction. We only perform this transformation on
4795 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
4796 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
4797 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType()))) {
4798 bool DoXform = true;
4799 SmallVector<SDNode*, 4> SetCCs;
4800 if (!N0.hasOneUse())
4801 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ANY_EXTEND, SetCCs, TLI);
4803 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4804 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, N->getDebugLoc(), VT,
4806 LN0->getBasePtr(), LN0->getPointerInfo(),
4808 LN0->isVolatile(), LN0->isNonTemporal(),
4809 LN0->getAlignment());
4810 CombineTo(N, ExtLoad);
4811 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4812 N0.getValueType(), ExtLoad);
4813 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
4814 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, N->getDebugLoc(),
4816 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4820 // fold (aext (zextload x)) -> (aext (truncate (zextload x)))
4821 // fold (aext (sextload x)) -> (aext (truncate (sextload x)))
4822 // fold (aext ( extload x)) -> (aext (truncate (extload x)))
4823 if (N0.getOpcode() == ISD::LOAD &&
4824 !ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
4826 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4827 EVT MemVT = LN0->getMemoryVT();
4828 SDValue ExtLoad = DAG.getExtLoad(LN0->getExtensionType(), N->getDebugLoc(),
4829 VT, LN0->getChain(), LN0->getBasePtr(),
4830 LN0->getPointerInfo(), MemVT,
4831 LN0->isVolatile(), LN0->isNonTemporal(),
4832 LN0->getAlignment());
4833 CombineTo(N, ExtLoad);
4834 CombineTo(N0.getNode(),
4835 DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4836 N0.getValueType(), ExtLoad),
4837 ExtLoad.getValue(1));
4838 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4841 if (N0.getOpcode() == ISD::SETCC) {
4842 // aext(setcc) -> sext_in_reg(vsetcc) for vectors.
4843 // Only do this before legalize for now.
4844 if (VT.isVector() && !LegalOperations) {
4845 EVT N0VT = N0.getOperand(0).getValueType();
4846 // We know that the # elements of the results is the same as the
4847 // # elements of the compare (and the # elements of the compare result
4848 // for that matter). Check to see that they are the same size. If so,
4849 // we know that the element size of the sext'd result matches the
4850 // element size of the compare operands.
4851 if (VT.getSizeInBits() == N0VT.getSizeInBits())
4852 return DAG.getSetCC(N->getDebugLoc(), VT, N0.getOperand(0),
4854 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4855 // If the desired elements are smaller or larger than the source
4856 // elements we can use a matching integer vector type and then
4857 // truncate/sign extend
4859 EVT MatchingElementType =
4860 EVT::getIntegerVT(*DAG.getContext(),
4861 N0VT.getScalarType().getSizeInBits());
4862 EVT MatchingVectorType =
4863 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
4864 N0VT.getVectorNumElements());
4866 DAG.getSetCC(N->getDebugLoc(), MatchingVectorType, N0.getOperand(0),
4868 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4869 return DAG.getSExtOrTrunc(VsetCC, N->getDebugLoc(), VT);
4873 // aext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
4875 SimplifySelectCC(N->getDebugLoc(), N0.getOperand(0), N0.getOperand(1),
4876 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
4877 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
4885 /// GetDemandedBits - See if the specified operand can be simplified with the
4886 /// knowledge that only the bits specified by Mask are used. If so, return the
4887 /// simpler operand, otherwise return a null SDValue.
4888 SDValue DAGCombiner::GetDemandedBits(SDValue V, const APInt &Mask) {
4889 switch (V.getOpcode()) {
4891 case ISD::Constant: {
4892 const ConstantSDNode *CV = cast<ConstantSDNode>(V.getNode());
4893 assert(CV != 0 && "Const value should be ConstSDNode.");
4894 const APInt &CVal = CV->getAPIntValue();
4895 APInt NewVal = CVal & Mask;
4896 if (NewVal != CVal) {
4897 return DAG.getConstant(NewVal, V.getValueType());
4903 // If the LHS or RHS don't contribute bits to the or, drop them.
4904 if (DAG.MaskedValueIsZero(V.getOperand(0), Mask))
4905 return V.getOperand(1);
4906 if (DAG.MaskedValueIsZero(V.getOperand(1), Mask))
4907 return V.getOperand(0);
4910 // Only look at single-use SRLs.
4911 if (!V.getNode()->hasOneUse())
4913 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(V.getOperand(1))) {
4914 // See if we can recursively simplify the LHS.
4915 unsigned Amt = RHSC->getZExtValue();
4917 // Watch out for shift count overflow though.
4918 if (Amt >= Mask.getBitWidth()) break;
4919 APInt NewMask = Mask << Amt;
4920 SDValue SimplifyLHS = GetDemandedBits(V.getOperand(0), NewMask);
4921 if (SimplifyLHS.getNode())
4922 return DAG.getNode(ISD::SRL, V.getDebugLoc(), V.getValueType(),
4923 SimplifyLHS, V.getOperand(1));
4929 /// ReduceLoadWidth - If the result of a wider load is shifted to right of N
4930 /// bits and then truncated to a narrower type and where N is a multiple
4931 /// of number of bits of the narrower type, transform it to a narrower load
4932 /// from address + N / num of bits of new type. If the result is to be
4933 /// extended, also fold the extension to form a extending load.
4934 SDValue DAGCombiner::ReduceLoadWidth(SDNode *N) {
4935 unsigned Opc = N->getOpcode();
4937 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
4938 SDValue N0 = N->getOperand(0);
4939 EVT VT = N->getValueType(0);
4942 // This transformation isn't valid for vector loads.
4946 // Special case: SIGN_EXTEND_INREG is basically truncating to ExtVT then
4948 if (Opc == ISD::SIGN_EXTEND_INREG) {
4949 ExtType = ISD::SEXTLOAD;
4950 ExtVT = cast<VTSDNode>(N->getOperand(1))->getVT();
4951 } else if (Opc == ISD::SRL) {
4952 // Another special-case: SRL is basically zero-extending a narrower value.
4953 ExtType = ISD::ZEXTLOAD;
4955 ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1));
4956 if (!N01) return SDValue();
4957 ExtVT = EVT::getIntegerVT(*DAG.getContext(),
4958 VT.getSizeInBits() - N01->getZExtValue());
4960 if (LegalOperations && !TLI.isLoadExtLegal(ExtType, ExtVT))
4963 unsigned EVTBits = ExtVT.getSizeInBits();
4965 // Do not generate loads of non-round integer types since these can
4966 // be expensive (and would be wrong if the type is not byte sized).
4967 if (!ExtVT.isRound())
4971 if (N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
4972 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
4973 ShAmt = N01->getZExtValue();
4974 // Is the shift amount a multiple of size of VT?
4975 if ((ShAmt & (EVTBits-1)) == 0) {
4976 N0 = N0.getOperand(0);
4977 // Is the load width a multiple of size of VT?
4978 if ((N0.getValueType().getSizeInBits() & (EVTBits-1)) != 0)
4982 // At this point, we must have a load or else we can't do the transform.
4983 if (!isa<LoadSDNode>(N0)) return SDValue();
4985 // If the shift amount is larger than the input type then we're not
4986 // accessing any of the loaded bytes. If the load was a zextload/extload
4987 // then the result of the shift+trunc is zero/undef (handled elsewhere).
4988 // If the load was a sextload then the result is a splat of the sign bit
4989 // of the extended byte. This is not worth optimizing for.
4990 if (ShAmt >= cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits())
4995 // If the load is shifted left (and the result isn't shifted back right),
4996 // we can fold the truncate through the shift.
4997 unsigned ShLeftAmt = 0;
4998 if (ShAmt == 0 && N0.getOpcode() == ISD::SHL && N0.hasOneUse() &&
4999 ExtVT == VT && TLI.isNarrowingProfitable(N0.getValueType(), VT)) {
5000 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
5001 ShLeftAmt = N01->getZExtValue();
5002 N0 = N0.getOperand(0);
5006 // If we haven't found a load, we can't narrow it. Don't transform one with
5007 // multiple uses, this would require adding a new load.
5008 if (!isa<LoadSDNode>(N0) || !N0.hasOneUse() ||
5009 // Don't change the width of a volatile load.
5010 cast<LoadSDNode>(N0)->isVolatile())
5013 // Verify that we are actually reducing a load width here.
5014 if (cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits() < EVTBits)
5017 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5018 EVT PtrType = N0.getOperand(1).getValueType();
5020 // For big endian targets, we need to adjust the offset to the pointer to
5021 // load the correct bytes.
5022 if (TLI.isBigEndian()) {
5023 unsigned LVTStoreBits = LN0->getMemoryVT().getStoreSizeInBits();
5024 unsigned EVTStoreBits = ExtVT.getStoreSizeInBits();
5025 ShAmt = LVTStoreBits - EVTStoreBits - ShAmt;
5028 uint64_t PtrOff = ShAmt / 8;
5029 unsigned NewAlign = MinAlign(LN0->getAlignment(), PtrOff);
5030 SDValue NewPtr = DAG.getNode(ISD::ADD, LN0->getDebugLoc(),
5031 PtrType, LN0->getBasePtr(),
5032 DAG.getConstant(PtrOff, PtrType));
5033 AddToWorkList(NewPtr.getNode());
5036 if (ExtType == ISD::NON_EXTLOAD)
5037 Load = DAG.getLoad(VT, N0.getDebugLoc(), LN0->getChain(), NewPtr,
5038 LN0->getPointerInfo().getWithOffset(PtrOff),
5039 LN0->isVolatile(), LN0->isNonTemporal(),
5040 LN0->isInvariant(), NewAlign);
5042 Load = DAG.getExtLoad(ExtType, N0.getDebugLoc(), VT, LN0->getChain(),NewPtr,
5043 LN0->getPointerInfo().getWithOffset(PtrOff),
5044 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
5047 // Replace the old load's chain with the new load's chain.
5048 WorkListRemover DeadNodes(*this);
5049 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1));
5051 // Shift the result left, if we've swallowed a left shift.
5052 SDValue Result = Load;
5053 if (ShLeftAmt != 0) {
5054 EVT ShImmTy = getShiftAmountTy(Result.getValueType());
5055 if (!isUIntN(ShImmTy.getSizeInBits(), ShLeftAmt))
5057 Result = DAG.getNode(ISD::SHL, N0.getDebugLoc(), VT,
5058 Result, DAG.getConstant(ShLeftAmt, ShImmTy));
5061 // Return the new loaded value.
5065 SDValue DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
5066 SDValue N0 = N->getOperand(0);
5067 SDValue N1 = N->getOperand(1);
5068 EVT VT = N->getValueType(0);
5069 EVT EVT = cast<VTSDNode>(N1)->getVT();
5070 unsigned VTBits = VT.getScalarType().getSizeInBits();
5071 unsigned EVTBits = EVT.getScalarType().getSizeInBits();
5073 // fold (sext_in_reg c1) -> c1
5074 if (isa<ConstantSDNode>(N0) || N0.getOpcode() == ISD::UNDEF)
5075 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT, N0, N1);
5077 // If the input is already sign extended, just drop the extension.
5078 if (DAG.ComputeNumSignBits(N0) >= VTBits-EVTBits+1)
5081 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2
5082 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
5083 EVT.bitsLT(cast<VTSDNode>(N0.getOperand(1))->getVT())) {
5084 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT,
5085 N0.getOperand(0), N1);
5088 // fold (sext_in_reg (sext x)) -> (sext x)
5089 // fold (sext_in_reg (aext x)) -> (sext x)
5090 // if x is small enough.
5091 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) {
5092 SDValue N00 = N0.getOperand(0);
5093 if (N00.getValueType().getScalarType().getSizeInBits() <= EVTBits &&
5094 (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND, VT)))
5095 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, N00, N1);
5098 // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is known zero.
5099 if (DAG.MaskedValueIsZero(N0, APInt::getBitsSet(VTBits, EVTBits-1, EVTBits)))
5100 return DAG.getZeroExtendInReg(N0, N->getDebugLoc(), EVT);
5102 // fold operands of sext_in_reg based on knowledge that the top bits are not
5104 if (SimplifyDemandedBits(SDValue(N, 0)))
5105 return SDValue(N, 0);
5107 // fold (sext_in_reg (load x)) -> (smaller sextload x)
5108 // fold (sext_in_reg (srl (load x), c)) -> (smaller sextload (x+c/evtbits))
5109 SDValue NarrowLoad = ReduceLoadWidth(N);
5110 if (NarrowLoad.getNode())
5113 // fold (sext_in_reg (srl X, 24), i8) -> (sra X, 24)
5114 // fold (sext_in_reg (srl X, 23), i8) -> (sra X, 23) iff possible.
5115 // We already fold "(sext_in_reg (srl X, 25), i8) -> srl X, 25" above.
5116 if (N0.getOpcode() == ISD::SRL) {
5117 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
5118 if (ShAmt->getZExtValue()+EVTBits <= VTBits) {
5119 // We can turn this into an SRA iff the input to the SRL is already sign
5121 unsigned InSignBits = DAG.ComputeNumSignBits(N0.getOperand(0));
5122 if (VTBits-(ShAmt->getZExtValue()+EVTBits) < InSignBits)
5123 return DAG.getNode(ISD::SRA, N->getDebugLoc(), VT,
5124 N0.getOperand(0), N0.getOperand(1));
5128 // fold (sext_inreg (extload x)) -> (sextload x)
5129 if (ISD::isEXTLoad(N0.getNode()) &&
5130 ISD::isUNINDEXEDLoad(N0.getNode()) &&
5131 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
5132 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5133 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) {
5134 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5135 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
5137 LN0->getBasePtr(), LN0->getPointerInfo(),
5139 LN0->isVolatile(), LN0->isNonTemporal(),
5140 LN0->getAlignment());
5141 CombineTo(N, ExtLoad);
5142 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
5143 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5145 // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use
5146 if (ISD::isZEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
5148 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
5149 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5150 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) {
5151 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5152 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
5154 LN0->getBasePtr(), LN0->getPointerInfo(),
5156 LN0->isVolatile(), LN0->isNonTemporal(),
5157 LN0->getAlignment());
5158 CombineTo(N, ExtLoad);
5159 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
5160 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5163 // Form (sext_inreg (bswap >> 16)) or (sext_inreg (rotl (bswap) 16))
5164 if (EVTBits <= 16 && N0.getOpcode() == ISD::OR) {
5165 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
5166 N0.getOperand(1), false);
5167 if (BSwap.getNode() != 0)
5168 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT,
5175 SDValue DAGCombiner::visitTRUNCATE(SDNode *N) {
5176 SDValue N0 = N->getOperand(0);
5177 EVT VT = N->getValueType(0);
5178 bool isLE = TLI.isLittleEndian();
5181 if (N0.getValueType() == N->getValueType(0))
5183 // fold (truncate c1) -> c1
5184 if (isa<ConstantSDNode>(N0))
5185 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, N0);
5186 // fold (truncate (truncate x)) -> (truncate x)
5187 if (N0.getOpcode() == ISD::TRUNCATE)
5188 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, N0.getOperand(0));
5189 // fold (truncate (ext x)) -> (ext x) or (truncate x) or x
5190 if (N0.getOpcode() == ISD::ZERO_EXTEND ||
5191 N0.getOpcode() == ISD::SIGN_EXTEND ||
5192 N0.getOpcode() == ISD::ANY_EXTEND) {
5193 if (N0.getOperand(0).getValueType().bitsLT(VT))
5194 // if the source is smaller than the dest, we still need an extend
5195 return DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT,
5197 else if (N0.getOperand(0).getValueType().bitsGT(VT))
5198 // if the source is larger than the dest, than we just need the truncate
5199 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, N0.getOperand(0));
5201 // if the source and dest are the same type, we can drop both the extend
5202 // and the truncate.
5203 return N0.getOperand(0);
5206 // Fold extract-and-trunc into a narrow extract. For example:
5207 // i64 x = EXTRACT_VECTOR_ELT(v2i64 val, i32 1)
5208 // i32 y = TRUNCATE(i64 x)
5210 // v16i8 b = BITCAST (v2i64 val)
5211 // i8 x = EXTRACT_VECTOR_ELT(v16i8 b, i32 8)
5213 // Note: We only run this optimization after type legalization (which often
5214 // creates this pattern) and before operation legalization after which
5215 // we need to be more careful about the vector instructions that we generate.
5216 if (N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5217 LegalTypes && !LegalOperations && N0->hasOneUse()) {
5219 EVT VecTy = N0.getOperand(0).getValueType();
5220 EVT ExTy = N0.getValueType();
5221 EVT TrTy = N->getValueType(0);
5223 unsigned NumElem = VecTy.getVectorNumElements();
5224 unsigned SizeRatio = ExTy.getSizeInBits()/TrTy.getSizeInBits();
5226 EVT NVT = EVT::getVectorVT(*DAG.getContext(), TrTy, SizeRatio * NumElem);
5227 assert(NVT.getSizeInBits() == VecTy.getSizeInBits() && "Invalid Size");
5229 SDValue EltNo = N0->getOperand(1);
5230 if (isa<ConstantSDNode>(EltNo) && isTypeLegal(NVT)) {
5231 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
5232 EVT IndexTy = N0->getOperand(1).getValueType();
5233 int Index = isLE ? (Elt*SizeRatio) : (Elt*SizeRatio + (SizeRatio-1));
5235 SDValue V = DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
5236 NVT, N0.getOperand(0));
5238 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
5239 N->getDebugLoc(), TrTy, V,
5240 DAG.getConstant(Index, IndexTy));
5244 // See if we can simplify the input to this truncate through knowledge that
5245 // only the low bits are being used.
5246 // For example "trunc (or (shl x, 8), y)" // -> trunc y
5247 // Currently we only perform this optimization on scalars because vectors
5248 // may have different active low bits.
5249 if (!VT.isVector()) {
5251 GetDemandedBits(N0, APInt::getLowBitsSet(N0.getValueSizeInBits(),
5252 VT.getSizeInBits()));
5253 if (Shorter.getNode())
5254 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, Shorter);
5256 // fold (truncate (load x)) -> (smaller load x)
5257 // fold (truncate (srl (load x), c)) -> (smaller load (x+c/evtbits))
5258 if (!LegalTypes || TLI.isTypeDesirableForOp(N0.getOpcode(), VT)) {
5259 SDValue Reduced = ReduceLoadWidth(N);
5260 if (Reduced.getNode())
5264 // Simplify the operands using demanded-bits information.
5265 if (!VT.isVector() &&
5266 SimplifyDemandedBits(SDValue(N, 0)))
5267 return SDValue(N, 0);
5272 static SDNode *getBuildPairElt(SDNode *N, unsigned i) {
5273 SDValue Elt = N->getOperand(i);
5274 if (Elt.getOpcode() != ISD::MERGE_VALUES)
5275 return Elt.getNode();
5276 return Elt.getOperand(Elt.getResNo()).getNode();
5279 /// CombineConsecutiveLoads - build_pair (load, load) -> load
5280 /// if load locations are consecutive.
5281 SDValue DAGCombiner::CombineConsecutiveLoads(SDNode *N, EVT VT) {
5282 assert(N->getOpcode() == ISD::BUILD_PAIR);
5284 LoadSDNode *LD1 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 0));
5285 LoadSDNode *LD2 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 1));
5286 if (!LD1 || !LD2 || !ISD::isNON_EXTLoad(LD1) || !LD1->hasOneUse() ||
5287 LD1->getPointerInfo().getAddrSpace() !=
5288 LD2->getPointerInfo().getAddrSpace())
5290 EVT LD1VT = LD1->getValueType(0);
5292 if (ISD::isNON_EXTLoad(LD2) &&
5294 // If both are volatile this would reduce the number of volatile loads.
5295 // If one is volatile it might be ok, but play conservative and bail out.
5296 !LD1->isVolatile() &&
5297 !LD2->isVolatile() &&
5298 DAG.isConsecutiveLoad(LD2, LD1, LD1VT.getSizeInBits()/8, 1)) {
5299 unsigned Align = LD1->getAlignment();
5300 unsigned NewAlign = TLI.getTargetData()->
5301 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
5303 if (NewAlign <= Align &&
5304 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)))
5305 return DAG.getLoad(VT, N->getDebugLoc(), LD1->getChain(),
5306 LD1->getBasePtr(), LD1->getPointerInfo(),
5307 false, false, false, Align);
5313 SDValue DAGCombiner::visitBITCAST(SDNode *N) {
5314 SDValue N0 = N->getOperand(0);
5315 EVT VT = N->getValueType(0);
5317 // If the input is a BUILD_VECTOR with all constant elements, fold this now.
5318 // Only do this before legalize, since afterward the target may be depending
5319 // on the bitconvert.
5320 // First check to see if this is all constant.
5322 N0.getOpcode() == ISD::BUILD_VECTOR && N0.getNode()->hasOneUse() &&
5324 bool isSimple = true;
5325 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i)
5326 if (N0.getOperand(i).getOpcode() != ISD::UNDEF &&
5327 N0.getOperand(i).getOpcode() != ISD::Constant &&
5328 N0.getOperand(i).getOpcode() != ISD::ConstantFP) {
5333 EVT DestEltVT = N->getValueType(0).getVectorElementType();
5334 assert(!DestEltVT.isVector() &&
5335 "Element type of vector ValueType must not be vector!");
5337 return ConstantFoldBITCASTofBUILD_VECTOR(N0.getNode(), DestEltVT);
5340 // If the input is a constant, let getNode fold it.
5341 if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) {
5342 SDValue Res = DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, N0);
5343 if (Res.getNode() != N) {
5344 if (!LegalOperations ||
5345 TLI.isOperationLegal(Res.getNode()->getOpcode(), VT))
5348 // Folding it resulted in an illegal node, and it's too late to
5349 // do that. Clean up the old node and forego the transformation.
5350 // Ideally this won't happen very often, because instcombine
5351 // and the earlier dagcombine runs (where illegal nodes are
5352 // permitted) should have folded most of them already.
5353 DAG.DeleteNode(Res.getNode());
5357 // (conv (conv x, t1), t2) -> (conv x, t2)
5358 if (N0.getOpcode() == ISD::BITCAST)
5359 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT,
5362 // fold (conv (load x)) -> (load (conv*)x)
5363 // If the resultant load doesn't need a higher alignment than the original!
5364 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
5365 // Do not change the width of a volatile load.
5366 !cast<LoadSDNode>(N0)->isVolatile() &&
5367 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT))) {
5368 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5369 unsigned Align = TLI.getTargetData()->
5370 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
5371 unsigned OrigAlign = LN0->getAlignment();
5373 if (Align <= OrigAlign) {
5374 SDValue Load = DAG.getLoad(VT, N->getDebugLoc(), LN0->getChain(),
5375 LN0->getBasePtr(), LN0->getPointerInfo(),
5376 LN0->isVolatile(), LN0->isNonTemporal(),
5377 LN0->isInvariant(), OrigAlign);
5379 CombineTo(N0.getNode(),
5380 DAG.getNode(ISD::BITCAST, N0.getDebugLoc(),
5381 N0.getValueType(), Load),
5387 // fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit)
5388 // fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit))
5389 // This often reduces constant pool loads.
5390 if (((N0.getOpcode() == ISD::FNEG && !TLI.isFNegFree(VT)) ||
5391 (N0.getOpcode() == ISD::FABS && !TLI.isFAbsFree(VT))) &&
5392 N0.getNode()->hasOneUse() && VT.isInteger() && !VT.isVector()) {
5393 SDValue NewConv = DAG.getNode(ISD::BITCAST, N0.getDebugLoc(), VT,
5395 AddToWorkList(NewConv.getNode());
5397 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
5398 if (N0.getOpcode() == ISD::FNEG)
5399 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT,
5400 NewConv, DAG.getConstant(SignBit, VT));
5401 assert(N0.getOpcode() == ISD::FABS);
5402 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
5403 NewConv, DAG.getConstant(~SignBit, VT));
5406 // fold (bitconvert (fcopysign cst, x)) ->
5407 // (or (and (bitconvert x), sign), (and cst, (not sign)))
5408 // Note that we don't handle (copysign x, cst) because this can always be
5409 // folded to an fneg or fabs.
5410 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse() &&
5411 isa<ConstantFPSDNode>(N0.getOperand(0)) &&
5412 VT.isInteger() && !VT.isVector()) {
5413 unsigned OrigXWidth = N0.getOperand(1).getValueType().getSizeInBits();
5414 EVT IntXVT = EVT::getIntegerVT(*DAG.getContext(), OrigXWidth);
5415 if (isTypeLegal(IntXVT)) {
5416 SDValue X = DAG.getNode(ISD::BITCAST, N0.getDebugLoc(),
5417 IntXVT, N0.getOperand(1));
5418 AddToWorkList(X.getNode());
5420 // If X has a different width than the result/lhs, sext it or truncate it.
5421 unsigned VTWidth = VT.getSizeInBits();
5422 if (OrigXWidth < VTWidth) {
5423 X = DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, X);
5424 AddToWorkList(X.getNode());
5425 } else if (OrigXWidth > VTWidth) {
5426 // To get the sign bit in the right place, we have to shift it right
5427 // before truncating.
5428 X = DAG.getNode(ISD::SRL, X.getDebugLoc(),
5429 X.getValueType(), X,
5430 DAG.getConstant(OrigXWidth-VTWidth, X.getValueType()));
5431 AddToWorkList(X.getNode());
5432 X = DAG.getNode(ISD::TRUNCATE, X.getDebugLoc(), VT, X);
5433 AddToWorkList(X.getNode());
5436 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
5437 X = DAG.getNode(ISD::AND, X.getDebugLoc(), VT,
5438 X, DAG.getConstant(SignBit, VT));
5439 AddToWorkList(X.getNode());
5441 SDValue Cst = DAG.getNode(ISD::BITCAST, N0.getDebugLoc(),
5442 VT, N0.getOperand(0));
5443 Cst = DAG.getNode(ISD::AND, Cst.getDebugLoc(), VT,
5444 Cst, DAG.getConstant(~SignBit, VT));
5445 AddToWorkList(Cst.getNode());
5447 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, X, Cst);
5451 // bitconvert(build_pair(ld, ld)) -> ld iff load locations are consecutive.
5452 if (N0.getOpcode() == ISD::BUILD_PAIR) {
5453 SDValue CombineLD = CombineConsecutiveLoads(N0.getNode(), VT);
5454 if (CombineLD.getNode())
5461 SDValue DAGCombiner::visitBUILD_PAIR(SDNode *N) {
5462 EVT VT = N->getValueType(0);
5463 return CombineConsecutiveLoads(N, VT);
5466 /// ConstantFoldBITCASTofBUILD_VECTOR - We know that BV is a build_vector
5467 /// node with Constant, ConstantFP or Undef operands. DstEltVT indicates the
5468 /// destination element value type.
5469 SDValue DAGCombiner::
5470 ConstantFoldBITCASTofBUILD_VECTOR(SDNode *BV, EVT DstEltVT) {
5471 EVT SrcEltVT = BV->getValueType(0).getVectorElementType();
5473 // If this is already the right type, we're done.
5474 if (SrcEltVT == DstEltVT) return SDValue(BV, 0);
5476 unsigned SrcBitSize = SrcEltVT.getSizeInBits();
5477 unsigned DstBitSize = DstEltVT.getSizeInBits();
5479 // If this is a conversion of N elements of one type to N elements of another
5480 // type, convert each element. This handles FP<->INT cases.
5481 if (SrcBitSize == DstBitSize) {
5482 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
5483 BV->getValueType(0).getVectorNumElements());
5485 // Due to the FP element handling below calling this routine recursively,
5486 // we can end up with a scalar-to-vector node here.
5487 if (BV->getOpcode() == ISD::SCALAR_TO_VECTOR)
5488 return DAG.getNode(ISD::SCALAR_TO_VECTOR, BV->getDebugLoc(), VT,
5489 DAG.getNode(ISD::BITCAST, BV->getDebugLoc(),
5490 DstEltVT, BV->getOperand(0)));
5492 SmallVector<SDValue, 8> Ops;
5493 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
5494 SDValue Op = BV->getOperand(i);
5495 // If the vector element type is not legal, the BUILD_VECTOR operands
5496 // are promoted and implicitly truncated. Make that explicit here.
5497 if (Op.getValueType() != SrcEltVT)
5498 Op = DAG.getNode(ISD::TRUNCATE, BV->getDebugLoc(), SrcEltVT, Op);
5499 Ops.push_back(DAG.getNode(ISD::BITCAST, BV->getDebugLoc(),
5501 AddToWorkList(Ops.back().getNode());
5503 return DAG.getNode(ISD::BUILD_VECTOR, BV->getDebugLoc(), VT,
5504 &Ops[0], Ops.size());
5507 // Otherwise, we're growing or shrinking the elements. To avoid having to
5508 // handle annoying details of growing/shrinking FP values, we convert them to
5510 if (SrcEltVT.isFloatingPoint()) {
5511 // Convert the input float vector to a int vector where the elements are the
5513 assert((SrcEltVT == MVT::f32 || SrcEltVT == MVT::f64) && "Unknown FP VT!");
5514 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), SrcEltVT.getSizeInBits());
5515 BV = ConstantFoldBITCASTofBUILD_VECTOR(BV, IntVT).getNode();
5519 // Now we know the input is an integer vector. If the output is a FP type,
5520 // convert to integer first, then to FP of the right size.
5521 if (DstEltVT.isFloatingPoint()) {
5522 assert((DstEltVT == MVT::f32 || DstEltVT == MVT::f64) && "Unknown FP VT!");
5523 EVT TmpVT = EVT::getIntegerVT(*DAG.getContext(), DstEltVT.getSizeInBits());
5524 SDNode *Tmp = ConstantFoldBITCASTofBUILD_VECTOR(BV, TmpVT).getNode();
5526 // Next, convert to FP elements of the same size.
5527 return ConstantFoldBITCASTofBUILD_VECTOR(Tmp, DstEltVT);
5530 // Okay, we know the src/dst types are both integers of differing types.
5531 // Handling growing first.
5532 assert(SrcEltVT.isInteger() && DstEltVT.isInteger());
5533 if (SrcBitSize < DstBitSize) {
5534 unsigned NumInputsPerOutput = DstBitSize/SrcBitSize;
5536 SmallVector<SDValue, 8> Ops;
5537 for (unsigned i = 0, e = BV->getNumOperands(); i != e;
5538 i += NumInputsPerOutput) {
5539 bool isLE = TLI.isLittleEndian();
5540 APInt NewBits = APInt(DstBitSize, 0);
5541 bool EltIsUndef = true;
5542 for (unsigned j = 0; j != NumInputsPerOutput; ++j) {
5543 // Shift the previously computed bits over.
5544 NewBits <<= SrcBitSize;
5545 SDValue Op = BV->getOperand(i+ (isLE ? (NumInputsPerOutput-j-1) : j));
5546 if (Op.getOpcode() == ISD::UNDEF) continue;
5549 NewBits |= cast<ConstantSDNode>(Op)->getAPIntValue().
5550 zextOrTrunc(SrcBitSize).zext(DstBitSize);
5554 Ops.push_back(DAG.getUNDEF(DstEltVT));
5556 Ops.push_back(DAG.getConstant(NewBits, DstEltVT));
5559 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, Ops.size());
5560 return DAG.getNode(ISD::BUILD_VECTOR, BV->getDebugLoc(), VT,
5561 &Ops[0], Ops.size());
5564 // Finally, this must be the case where we are shrinking elements: each input
5565 // turns into multiple outputs.
5566 bool isS2V = ISD::isScalarToVector(BV);
5567 unsigned NumOutputsPerInput = SrcBitSize/DstBitSize;
5568 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
5569 NumOutputsPerInput*BV->getNumOperands());
5570 SmallVector<SDValue, 8> Ops;
5572 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
5573 if (BV->getOperand(i).getOpcode() == ISD::UNDEF) {
5574 for (unsigned j = 0; j != NumOutputsPerInput; ++j)
5575 Ops.push_back(DAG.getUNDEF(DstEltVT));
5579 APInt OpVal = cast<ConstantSDNode>(BV->getOperand(i))->
5580 getAPIntValue().zextOrTrunc(SrcBitSize);
5582 for (unsigned j = 0; j != NumOutputsPerInput; ++j) {
5583 APInt ThisVal = OpVal.trunc(DstBitSize);
5584 Ops.push_back(DAG.getConstant(ThisVal, DstEltVT));
5585 if (isS2V && i == 0 && j == 0 && ThisVal.zext(SrcBitSize) == OpVal)
5586 // Simply turn this into a SCALAR_TO_VECTOR of the new type.
5587 return DAG.getNode(ISD::SCALAR_TO_VECTOR, BV->getDebugLoc(), VT,
5589 OpVal = OpVal.lshr(DstBitSize);
5592 // For big endian targets, swap the order of the pieces of each element.
5593 if (TLI.isBigEndian())
5594 std::reverse(Ops.end()-NumOutputsPerInput, Ops.end());
5597 return DAG.getNode(ISD::BUILD_VECTOR, BV->getDebugLoc(), VT,
5598 &Ops[0], Ops.size());
5601 SDValue DAGCombiner::visitFADD(SDNode *N) {
5602 SDValue N0 = N->getOperand(0);
5603 SDValue N1 = N->getOperand(1);
5604 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5605 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
5606 EVT VT = N->getValueType(0);
5609 if (VT.isVector()) {
5610 SDValue FoldedVOp = SimplifyVBinOp(N);
5611 if (FoldedVOp.getNode()) return FoldedVOp;
5614 // fold (fadd c1, c2) -> c1 + c2
5615 if (N0CFP && N1CFP && VT != MVT::ppcf128)
5616 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0, N1);
5617 // canonicalize constant to RHS
5618 if (N0CFP && !N1CFP)
5619 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N1, N0);
5620 // fold (fadd A, 0) -> A
5621 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
5622 N1CFP->getValueAPF().isZero())
5624 // fold (fadd A, (fneg B)) -> (fsub A, B)
5625 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
5626 isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options) == 2)
5627 return DAG.getNode(ISD::FSUB, N->getDebugLoc(), VT, N0,
5628 GetNegatedExpression(N1, DAG, LegalOperations));
5629 // fold (fadd (fneg A), B) -> (fsub B, A)
5630 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
5631 isNegatibleForFree(N0, LegalOperations, TLI, &DAG.getTarget().Options) == 2)
5632 return DAG.getNode(ISD::FSUB, N->getDebugLoc(), VT, N1,
5633 GetNegatedExpression(N0, DAG, LegalOperations));
5635 // If allowed, fold (fadd (fadd x, c1), c2) -> (fadd x, (fadd c1, c2))
5636 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
5637 N0.getOpcode() == ISD::FADD && N0.getNode()->hasOneUse() &&
5638 isa<ConstantFPSDNode>(N0.getOperand(1)))
5639 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0.getOperand(0),
5640 DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
5641 N0.getOperand(1), N1));
5643 // FADD -> FMA combines:
5644 if ((DAG.getTarget().Options.AllowExcessFPPrecision ||
5645 DAG.getTarget().Options.UnsafeFPMath) &&
5646 DAG.getTarget().getTargetLowering()->isFMAFasterThanMulAndAdd(VT) &&
5647 TLI.isOperationLegal(ISD::FMA, VT)) {
5649 // fold (fadd (fmul x, y), z) -> (fma x, y, z)
5650 if (N0.getOpcode() == ISD::FMUL && N0->hasOneUse()) {
5651 return DAG.getNode(ISD::FMA, N->getDebugLoc(), VT,
5652 N0.getOperand(0), N0.getOperand(1), N1);
5655 // fold (fadd x, (fmul y, z)) -> (fma x, y, z)
5656 // Note: Commutes FADD operands.
5657 if (N1.getOpcode() == ISD::FMUL && N1->hasOneUse()) {
5658 return DAG.getNode(ISD::FMA, N->getDebugLoc(), VT,
5659 N1.getOperand(0), N1.getOperand(1), N0);
5666 SDValue DAGCombiner::visitFSUB(SDNode *N) {
5667 SDValue N0 = N->getOperand(0);
5668 SDValue N1 = N->getOperand(1);
5669 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5670 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
5671 EVT VT = N->getValueType(0);
5674 if (VT.isVector()) {
5675 SDValue FoldedVOp = SimplifyVBinOp(N);
5676 if (FoldedVOp.getNode()) return FoldedVOp;
5679 // fold (fsub c1, c2) -> c1-c2
5680 if (N0CFP && N1CFP && VT != MVT::ppcf128)
5681 return DAG.getNode(ISD::FSUB, N->getDebugLoc(), VT, N0, N1);
5682 // fold (fsub A, 0) -> A
5683 if (DAG.getTarget().Options.UnsafeFPMath &&
5684 N1CFP && N1CFP->getValueAPF().isZero())
5686 // fold (fsub 0, B) -> -B
5687 if (DAG.getTarget().Options.UnsafeFPMath &&
5688 N0CFP && N0CFP->getValueAPF().isZero()) {
5689 if (isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options))
5690 return GetNegatedExpression(N1, DAG, LegalOperations);
5691 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
5692 return DAG.getNode(ISD::FNEG, N->getDebugLoc(), VT, N1);
5694 // fold (fsub A, (fneg B)) -> (fadd A, B)
5695 if (isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options))
5696 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0,
5697 GetNegatedExpression(N1, DAG, LegalOperations));
5699 // If 'unsafe math' is enabled, fold
5700 // (fsub x, x) -> 0.0 &
5701 // (fsub x, (fadd x, y)) -> (fneg y) &
5702 // (fsub x, (fadd y, x)) -> (fneg y)
5703 if (DAG.getTarget().Options.UnsafeFPMath) {
5705 return DAG.getConstantFP(0.0f, VT);
5707 if (N1.getOpcode() == ISD::FADD) {
5708 SDValue N10 = N1->getOperand(0);
5709 SDValue N11 = N1->getOperand(1);
5711 if (N10 == N0 && isNegatibleForFree(N11, LegalOperations, TLI,
5712 &DAG.getTarget().Options))
5713 return GetNegatedExpression(N11, DAG, LegalOperations);
5714 else if (N11 == N0 && isNegatibleForFree(N10, LegalOperations, TLI,
5715 &DAG.getTarget().Options))
5716 return GetNegatedExpression(N10, DAG, LegalOperations);
5720 // FSUB -> FMA combines:
5721 if ((DAG.getTarget().Options.AllowExcessFPPrecision ||
5722 DAG.getTarget().Options.UnsafeFPMath) &&
5723 DAG.getTarget().getTargetLowering()->isFMAFasterThanMulAndAdd(VT) &&
5724 TLI.isOperationLegal(ISD::FMA, VT)) {
5726 // fold (fsub (fmul x, y), z) -> (fma x, y, (fneg z))
5727 if (N0.getOpcode() == ISD::FMUL && N0->hasOneUse()) {
5728 return DAG.getNode(ISD::FMA, N->getDebugLoc(), VT,
5729 N0.getOperand(0), N0.getOperand(1),
5730 DAG.getNode(ISD::FNEG, N1->getDebugLoc(), VT, N1));
5733 // fold (fsub x, (fmul y, z)) -> (fma (fneg y), z, x)
5734 // Note: Commutes FSUB operands.
5735 if (N1.getOpcode() == ISD::FMUL && N1->hasOneUse()) {
5736 return DAG.getNode(ISD::FMA, N->getDebugLoc(), VT,
5737 DAG.getNode(ISD::FNEG, N1->getDebugLoc(), VT,
5739 N1.getOperand(1), N0);
5746 SDValue DAGCombiner::visitFMUL(SDNode *N) {
5747 SDValue N0 = N->getOperand(0);
5748 SDValue N1 = N->getOperand(1);
5749 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5750 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
5751 EVT VT = N->getValueType(0);
5752 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5755 if (VT.isVector()) {
5756 SDValue FoldedVOp = SimplifyVBinOp(N);
5757 if (FoldedVOp.getNode()) return FoldedVOp;
5760 // fold (fmul c1, c2) -> c1*c2
5761 if (N0CFP && N1CFP && VT != MVT::ppcf128)
5762 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT, N0, N1);
5763 // canonicalize constant to RHS
5764 if (N0CFP && !N1CFP)
5765 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT, N1, N0);
5766 // fold (fmul A, 0) -> 0
5767 if (DAG.getTarget().Options.UnsafeFPMath &&
5768 N1CFP && N1CFP->getValueAPF().isZero())
5770 // fold (fmul A, 0) -> 0, vector edition.
5771 if (DAG.getTarget().Options.UnsafeFPMath &&
5772 ISD::isBuildVectorAllZeros(N1.getNode()))
5774 // fold (fmul A, 1.0) -> A
5775 if (N1CFP && N1CFP->isExactlyValue(1.0))
5777 // fold (fmul X, 2.0) -> (fadd X, X)
5778 if (N1CFP && N1CFP->isExactlyValue(+2.0))
5779 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0, N0);
5780 // fold (fmul X, -1.0) -> (fneg X)
5781 if (N1CFP && N1CFP->isExactlyValue(-1.0))
5782 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
5783 return DAG.getNode(ISD::FNEG, N->getDebugLoc(), VT, N0);
5785 // fold (fmul (fneg X), (fneg Y)) -> (fmul X, Y)
5786 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI,
5787 &DAG.getTarget().Options)) {
5788 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI,
5789 &DAG.getTarget().Options)) {
5790 // Both can be negated for free, check to see if at least one is cheaper
5792 if (LHSNeg == 2 || RHSNeg == 2)
5793 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5794 GetNegatedExpression(N0, DAG, LegalOperations),
5795 GetNegatedExpression(N1, DAG, LegalOperations));
5799 // If allowed, fold (fmul (fmul x, c1), c2) -> (fmul x, (fmul c1, c2))
5800 if (DAG.getTarget().Options.UnsafeFPMath &&
5801 N1CFP && N0.getOpcode() == ISD::FMUL &&
5802 N0.getNode()->hasOneUse() && isa<ConstantFPSDNode>(N0.getOperand(1)))
5803 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT, N0.getOperand(0),
5804 DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5805 N0.getOperand(1), N1));
5810 SDValue DAGCombiner::visitFMA(SDNode *N) {
5811 SDValue N0 = N->getOperand(0);
5812 SDValue N1 = N->getOperand(1);
5813 SDValue N2 = N->getOperand(2);
5814 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5815 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
5816 EVT VT = N->getValueType(0);
5818 if (N0CFP && N0CFP->isExactlyValue(1.0))
5819 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N1, N2);
5820 if (N1CFP && N1CFP->isExactlyValue(1.0))
5821 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0, N2);
5823 // Canonicalize (fma c, x, y) -> (fma x, c, y)
5824 if (N0CFP && !N1CFP)
5825 return DAG.getNode(ISD::FMA, N->getDebugLoc(), VT, N1, N0, N2);
5830 SDValue DAGCombiner::visitFDIV(SDNode *N) {
5831 SDValue N0 = N->getOperand(0);
5832 SDValue N1 = N->getOperand(1);
5833 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5834 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
5835 EVT VT = N->getValueType(0);
5836 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5839 if (VT.isVector()) {
5840 SDValue FoldedVOp = SimplifyVBinOp(N);
5841 if (FoldedVOp.getNode()) return FoldedVOp;
5844 // fold (fdiv c1, c2) -> c1/c2
5845 if (N0CFP && N1CFP && VT != MVT::ppcf128)
5846 return DAG.getNode(ISD::FDIV, N->getDebugLoc(), VT, N0, N1);
5848 // fold (fdiv X, c2) -> fmul X, 1/c2 if losing precision is acceptable.
5849 if (N1CFP && VT != MVT::ppcf128 && DAG.getTarget().Options.UnsafeFPMath) {
5850 // Compute the reciprocal 1.0 / c2.
5851 APFloat N1APF = N1CFP->getValueAPF();
5852 APFloat Recip(N1APF.getSemantics(), 1); // 1.0
5853 APFloat::opStatus st = Recip.divide(N1APF, APFloat::rmNearestTiesToEven);
5854 // Only do the transform if the reciprocal is a legal fp immediate that
5855 // isn't too nasty (eg NaN, denormal, ...).
5856 if ((st == APFloat::opOK || st == APFloat::opInexact) && // Not too nasty
5857 (!LegalOperations ||
5858 // FIXME: custom lowering of ConstantFP might fail (see e.g. ARM
5859 // backend)... we should handle this gracefully after Legalize.
5860 // TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT) ||
5861 TLI.isOperationLegal(llvm::ISD::ConstantFP, VT) ||
5862 TLI.isFPImmLegal(Recip, VT)))
5863 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT, N0,
5864 DAG.getConstantFP(Recip, VT));
5867 // (fdiv (fneg X), (fneg Y)) -> (fdiv X, Y)
5868 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI,
5869 &DAG.getTarget().Options)) {
5870 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI,
5871 &DAG.getTarget().Options)) {
5872 // Both can be negated for free, check to see if at least one is cheaper
5874 if (LHSNeg == 2 || RHSNeg == 2)
5875 return DAG.getNode(ISD::FDIV, N->getDebugLoc(), VT,
5876 GetNegatedExpression(N0, DAG, LegalOperations),
5877 GetNegatedExpression(N1, DAG, LegalOperations));
5884 SDValue DAGCombiner::visitFREM(SDNode *N) {
5885 SDValue N0 = N->getOperand(0);
5886 SDValue N1 = N->getOperand(1);
5887 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5888 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
5889 EVT VT = N->getValueType(0);
5891 // fold (frem c1, c2) -> fmod(c1,c2)
5892 if (N0CFP && N1CFP && VT != MVT::ppcf128)
5893 return DAG.getNode(ISD::FREM, N->getDebugLoc(), VT, N0, N1);
5898 SDValue DAGCombiner::visitFCOPYSIGN(SDNode *N) {
5899 SDValue N0 = N->getOperand(0);
5900 SDValue N1 = N->getOperand(1);
5901 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5902 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
5903 EVT VT = N->getValueType(0);
5905 if (N0CFP && N1CFP && VT != MVT::ppcf128) // Constant fold
5906 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT, N0, N1);
5909 const APFloat& V = N1CFP->getValueAPF();
5910 // copysign(x, c1) -> fabs(x) iff ispos(c1)
5911 // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1)
5912 if (!V.isNegative()) {
5913 if (!LegalOperations || TLI.isOperationLegal(ISD::FABS, VT))
5914 return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0);
5916 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
5917 return DAG.getNode(ISD::FNEG, N->getDebugLoc(), VT,
5918 DAG.getNode(ISD::FABS, N0.getDebugLoc(), VT, N0));
5922 // copysign(fabs(x), y) -> copysign(x, y)
5923 // copysign(fneg(x), y) -> copysign(x, y)
5924 // copysign(copysign(x,z), y) -> copysign(x, y)
5925 if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG ||
5926 N0.getOpcode() == ISD::FCOPYSIGN)
5927 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
5928 N0.getOperand(0), N1);
5930 // copysign(x, abs(y)) -> abs(x)
5931 if (N1.getOpcode() == ISD::FABS)
5932 return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0);
5934 // copysign(x, copysign(y,z)) -> copysign(x, z)
5935 if (N1.getOpcode() == ISD::FCOPYSIGN)
5936 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
5937 N0, N1.getOperand(1));
5939 // copysign(x, fp_extend(y)) -> copysign(x, y)
5940 // copysign(x, fp_round(y)) -> copysign(x, y)
5941 if (N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND)
5942 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
5943 N0, N1.getOperand(0));
5948 SDValue DAGCombiner::visitSINT_TO_FP(SDNode *N) {
5949 SDValue N0 = N->getOperand(0);
5950 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
5951 EVT VT = N->getValueType(0);
5952 EVT OpVT = N0.getValueType();
5954 // fold (sint_to_fp c1) -> c1fp
5955 if (N0C && OpVT != MVT::ppcf128 &&
5956 // ...but only if the target supports immediate floating-point values
5957 (!LegalOperations ||
5958 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
5959 return DAG.getNode(ISD::SINT_TO_FP, N->getDebugLoc(), VT, N0);
5961 // If the input is a legal type, and SINT_TO_FP is not legal on this target,
5962 // but UINT_TO_FP is legal on this target, try to convert.
5963 if (!TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT) &&
5964 TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT)) {
5965 // If the sign bit is known to be zero, we can change this to UINT_TO_FP.
5966 if (DAG.SignBitIsZero(N0))
5967 return DAG.getNode(ISD::UINT_TO_FP, N->getDebugLoc(), VT, N0);
5973 SDValue DAGCombiner::visitUINT_TO_FP(SDNode *N) {
5974 SDValue N0 = N->getOperand(0);
5975 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
5976 EVT VT = N->getValueType(0);
5977 EVT OpVT = N0.getValueType();
5979 // fold (uint_to_fp c1) -> c1fp
5980 if (N0C && OpVT != MVT::ppcf128 &&
5981 // ...but only if the target supports immediate floating-point values
5982 (!LegalOperations ||
5983 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
5984 return DAG.getNode(ISD::UINT_TO_FP, N->getDebugLoc(), VT, N0);
5986 // If the input is a legal type, and UINT_TO_FP is not legal on this target,
5987 // but SINT_TO_FP is legal on this target, try to convert.
5988 if (!TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT) &&
5989 TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT)) {
5990 // If the sign bit is known to be zero, we can change this to SINT_TO_FP.
5991 if (DAG.SignBitIsZero(N0))
5992 return DAG.getNode(ISD::SINT_TO_FP, N->getDebugLoc(), VT, N0);
5998 SDValue DAGCombiner::visitFP_TO_SINT(SDNode *N) {
5999 SDValue N0 = N->getOperand(0);
6000 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6001 EVT VT = N->getValueType(0);
6003 // fold (fp_to_sint c1fp) -> c1
6005 return DAG.getNode(ISD::FP_TO_SINT, N->getDebugLoc(), VT, N0);
6010 SDValue DAGCombiner::visitFP_TO_UINT(SDNode *N) {
6011 SDValue N0 = N->getOperand(0);
6012 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6013 EVT VT = N->getValueType(0);
6015 // fold (fp_to_uint c1fp) -> c1
6016 if (N0CFP && VT != MVT::ppcf128)
6017 return DAG.getNode(ISD::FP_TO_UINT, N->getDebugLoc(), VT, N0);
6022 SDValue DAGCombiner::visitFP_ROUND(SDNode *N) {
6023 SDValue N0 = N->getOperand(0);
6024 SDValue N1 = N->getOperand(1);
6025 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6026 EVT VT = N->getValueType(0);
6028 // fold (fp_round c1fp) -> c1fp
6029 if (N0CFP && N0.getValueType() != MVT::ppcf128)
6030 return DAG.getNode(ISD::FP_ROUND, N->getDebugLoc(), VT, N0, N1);
6032 // fold (fp_round (fp_extend x)) -> x
6033 if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType())
6034 return N0.getOperand(0);
6036 // fold (fp_round (fp_round x)) -> (fp_round x)
6037 if (N0.getOpcode() == ISD::FP_ROUND) {
6038 // This is a value preserving truncation if both round's are.
6039 bool IsTrunc = N->getConstantOperandVal(1) == 1 &&
6040 N0.getNode()->getConstantOperandVal(1) == 1;
6041 return DAG.getNode(ISD::FP_ROUND, N->getDebugLoc(), VT, N0.getOperand(0),
6042 DAG.getIntPtrConstant(IsTrunc));
6045 // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y)
6046 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse()) {
6047 SDValue Tmp = DAG.getNode(ISD::FP_ROUND, N0.getDebugLoc(), VT,
6048 N0.getOperand(0), N1);
6049 AddToWorkList(Tmp.getNode());
6050 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
6051 Tmp, N0.getOperand(1));
6057 SDValue DAGCombiner::visitFP_ROUND_INREG(SDNode *N) {
6058 SDValue N0 = N->getOperand(0);
6059 EVT VT = N->getValueType(0);
6060 EVT EVT = cast<VTSDNode>(N->getOperand(1))->getVT();
6061 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6063 // fold (fp_round_inreg c1fp) -> c1fp
6064 if (N0CFP && isTypeLegal(EVT)) {
6065 SDValue Round = DAG.getConstantFP(*N0CFP->getConstantFPValue(), EVT);
6066 return DAG.getNode(ISD::FP_EXTEND, N->getDebugLoc(), VT, Round);
6072 SDValue DAGCombiner::visitFP_EXTEND(SDNode *N) {
6073 SDValue N0 = N->getOperand(0);
6074 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6075 EVT VT = N->getValueType(0);
6077 // If this is fp_round(fpextend), don't fold it, allow ourselves to be folded.
6078 if (N->hasOneUse() &&
6079 N->use_begin()->getOpcode() == ISD::FP_ROUND)
6082 // fold (fp_extend c1fp) -> c1fp
6083 if (N0CFP && VT != MVT::ppcf128)
6084 return DAG.getNode(ISD::FP_EXTEND, N->getDebugLoc(), VT, N0);
6086 // Turn fp_extend(fp_round(X, 1)) -> x since the fp_round doesn't affect the
6088 if (N0.getOpcode() == ISD::FP_ROUND
6089 && N0.getNode()->getConstantOperandVal(1) == 1) {
6090 SDValue In = N0.getOperand(0);
6091 if (In.getValueType() == VT) return In;
6092 if (VT.bitsLT(In.getValueType()))
6093 return DAG.getNode(ISD::FP_ROUND, N->getDebugLoc(), VT,
6094 In, N0.getOperand(1));
6095 return DAG.getNode(ISD::FP_EXTEND, N->getDebugLoc(), VT, In);
6098 // fold (fpext (load x)) -> (fpext (fptrunc (extload x)))
6099 if (ISD::isNON_EXTLoad(N0.getNode()) && N0.hasOneUse() &&
6100 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
6101 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType()))) {
6102 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6103 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, N->getDebugLoc(), VT,
6105 LN0->getBasePtr(), LN0->getPointerInfo(),
6107 LN0->isVolatile(), LN0->isNonTemporal(),
6108 LN0->getAlignment());
6109 CombineTo(N, ExtLoad);
6110 CombineTo(N0.getNode(),
6111 DAG.getNode(ISD::FP_ROUND, N0.getDebugLoc(),
6112 N0.getValueType(), ExtLoad, DAG.getIntPtrConstant(1)),
6113 ExtLoad.getValue(1));
6114 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6120 SDValue DAGCombiner::visitFNEG(SDNode *N) {
6121 SDValue N0 = N->getOperand(0);
6122 EVT VT = N->getValueType(0);
6124 if (isNegatibleForFree(N0, LegalOperations, DAG.getTargetLoweringInfo(),
6125 &DAG.getTarget().Options))
6126 return GetNegatedExpression(N0, DAG, LegalOperations);
6128 // Transform fneg(bitconvert(x)) -> bitconvert(x^sign) to avoid loading
6129 // constant pool values.
6130 if (!TLI.isFNegFree(VT) && N0.getOpcode() == ISD::BITCAST &&
6132 N0.getNode()->hasOneUse() &&
6133 N0.getOperand(0).getValueType().isInteger()) {
6134 SDValue Int = N0.getOperand(0);
6135 EVT IntVT = Int.getValueType();
6136 if (IntVT.isInteger() && !IntVT.isVector()) {
6137 Int = DAG.getNode(ISD::XOR, N0.getDebugLoc(), IntVT, Int,
6138 DAG.getConstant(APInt::getSignBit(IntVT.getSizeInBits()), IntVT));
6139 AddToWorkList(Int.getNode());
6140 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
6148 SDValue DAGCombiner::visitFABS(SDNode *N) {
6149 SDValue N0 = N->getOperand(0);
6150 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6151 EVT VT = N->getValueType(0);
6153 // fold (fabs c1) -> fabs(c1)
6154 if (N0CFP && VT != MVT::ppcf128)
6155 return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0);
6156 // fold (fabs (fabs x)) -> (fabs x)
6157 if (N0.getOpcode() == ISD::FABS)
6158 return N->getOperand(0);
6159 // fold (fabs (fneg x)) -> (fabs x)
6160 // fold (fabs (fcopysign x, y)) -> (fabs x)
6161 if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN)
6162 return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0.getOperand(0));
6164 // Transform fabs(bitconvert(x)) -> bitconvert(x&~sign) to avoid loading
6165 // constant pool values.
6166 if (!TLI.isFAbsFree(VT) &&
6167 N0.getOpcode() == ISD::BITCAST && N0.getNode()->hasOneUse() &&
6168 N0.getOperand(0).getValueType().isInteger() &&
6169 !N0.getOperand(0).getValueType().isVector()) {
6170 SDValue Int = N0.getOperand(0);
6171 EVT IntVT = Int.getValueType();
6172 if (IntVT.isInteger() && !IntVT.isVector()) {
6173 Int = DAG.getNode(ISD::AND, N0.getDebugLoc(), IntVT, Int,
6174 DAG.getConstant(~APInt::getSignBit(IntVT.getSizeInBits()), IntVT));
6175 AddToWorkList(Int.getNode());
6176 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
6177 N->getValueType(0), Int);
6184 SDValue DAGCombiner::visitBRCOND(SDNode *N) {
6185 SDValue Chain = N->getOperand(0);
6186 SDValue N1 = N->getOperand(1);
6187 SDValue N2 = N->getOperand(2);
6189 // If N is a constant we could fold this into a fallthrough or unconditional
6190 // branch. However that doesn't happen very often in normal code, because
6191 // Instcombine/SimplifyCFG should have handled the available opportunities.
6192 // If we did this folding here, it would be necessary to update the
6193 // MachineBasicBlock CFG, which is awkward.
6195 // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal
6197 if (N1.getOpcode() == ISD::SETCC &&
6198 TLI.isOperationLegalOrCustom(ISD::BR_CC, MVT::Other)) {
6199 return DAG.getNode(ISD::BR_CC, N->getDebugLoc(), MVT::Other,
6200 Chain, N1.getOperand(2),
6201 N1.getOperand(0), N1.getOperand(1), N2);
6204 if ((N1.hasOneUse() && N1.getOpcode() == ISD::SRL) ||
6205 ((N1.getOpcode() == ISD::TRUNCATE && N1.hasOneUse()) &&
6206 (N1.getOperand(0).hasOneUse() &&
6207 N1.getOperand(0).getOpcode() == ISD::SRL))) {
6209 if (N1.getOpcode() == ISD::TRUNCATE) {
6210 // Look pass the truncate.
6211 Trunc = N1.getNode();
6212 N1 = N1.getOperand(0);
6215 // Match this pattern so that we can generate simpler code:
6218 // %b = and i32 %a, 2
6219 // %c = srl i32 %b, 1
6220 // brcond i32 %c ...
6225 // %b = and i32 %a, 2
6226 // %c = setcc eq %b, 0
6229 // This applies only when the AND constant value has one bit set and the
6230 // SRL constant is equal to the log2 of the AND constant. The back-end is
6231 // smart enough to convert the result into a TEST/JMP sequence.
6232 SDValue Op0 = N1.getOperand(0);
6233 SDValue Op1 = N1.getOperand(1);
6235 if (Op0.getOpcode() == ISD::AND &&
6236 Op1.getOpcode() == ISD::Constant) {
6237 SDValue AndOp1 = Op0.getOperand(1);
6239 if (AndOp1.getOpcode() == ISD::Constant) {
6240 const APInt &AndConst = cast<ConstantSDNode>(AndOp1)->getAPIntValue();
6242 if (AndConst.isPowerOf2() &&
6243 cast<ConstantSDNode>(Op1)->getAPIntValue()==AndConst.logBase2()) {
6245 DAG.getSetCC(N->getDebugLoc(),
6246 TLI.getSetCCResultType(Op0.getValueType()),
6247 Op0, DAG.getConstant(0, Op0.getValueType()),
6250 SDValue NewBRCond = DAG.getNode(ISD::BRCOND, N->getDebugLoc(),
6251 MVT::Other, Chain, SetCC, N2);
6252 // Don't add the new BRCond into the worklist or else SimplifySelectCC
6253 // will convert it back to (X & C1) >> C2.
6254 CombineTo(N, NewBRCond, false);
6255 // Truncate is dead.
6257 removeFromWorkList(Trunc);
6258 DAG.DeleteNode(Trunc);
6260 // Replace the uses of SRL with SETCC
6261 WorkListRemover DeadNodes(*this);
6262 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
6263 removeFromWorkList(N1.getNode());
6264 DAG.DeleteNode(N1.getNode());
6265 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6271 // Restore N1 if the above transformation doesn't match.
6272 N1 = N->getOperand(1);
6275 // Transform br(xor(x, y)) -> br(x != y)
6276 // Transform br(xor(xor(x,y), 1)) -> br (x == y)
6277 if (N1.hasOneUse() && N1.getOpcode() == ISD::XOR) {
6278 SDNode *TheXor = N1.getNode();
6279 SDValue Op0 = TheXor->getOperand(0);
6280 SDValue Op1 = TheXor->getOperand(1);
6281 if (Op0.getOpcode() == Op1.getOpcode()) {
6282 // Avoid missing important xor optimizations.
6283 SDValue Tmp = visitXOR(TheXor);
6284 if (Tmp.getNode() && Tmp.getNode() != TheXor) {
6285 DEBUG(dbgs() << "\nReplacing.8 ";
6287 dbgs() << "\nWith: ";
6288 Tmp.getNode()->dump(&DAG);
6290 WorkListRemover DeadNodes(*this);
6291 DAG.ReplaceAllUsesOfValueWith(N1, Tmp);
6292 removeFromWorkList(TheXor);
6293 DAG.DeleteNode(TheXor);
6294 return DAG.getNode(ISD::BRCOND, N->getDebugLoc(),
6295 MVT::Other, Chain, Tmp, N2);
6299 if (Op0.getOpcode() != ISD::SETCC && Op1.getOpcode() != ISD::SETCC) {
6301 if (ConstantSDNode *RHSCI = dyn_cast<ConstantSDNode>(Op0))
6302 if (RHSCI->getAPIntValue() == 1 && Op0.hasOneUse() &&
6303 Op0.getOpcode() == ISD::XOR) {
6304 TheXor = Op0.getNode();
6308 EVT SetCCVT = N1.getValueType();
6310 SetCCVT = TLI.getSetCCResultType(SetCCVT);
6311 SDValue SetCC = DAG.getSetCC(TheXor->getDebugLoc(),
6314 Equal ? ISD::SETEQ : ISD::SETNE);
6315 // Replace the uses of XOR with SETCC
6316 WorkListRemover DeadNodes(*this);
6317 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
6318 removeFromWorkList(N1.getNode());
6319 DAG.DeleteNode(N1.getNode());
6320 return DAG.getNode(ISD::BRCOND, N->getDebugLoc(),
6321 MVT::Other, Chain, SetCC, N2);
6328 // Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB.
6330 SDValue DAGCombiner::visitBR_CC(SDNode *N) {
6331 CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1));
6332 SDValue CondLHS = N->getOperand(2), CondRHS = N->getOperand(3);
6334 // If N is a constant we could fold this into a fallthrough or unconditional
6335 // branch. However that doesn't happen very often in normal code, because
6336 // Instcombine/SimplifyCFG should have handled the available opportunities.
6337 // If we did this folding here, it would be necessary to update the
6338 // MachineBasicBlock CFG, which is awkward.
6340 // Use SimplifySetCC to simplify SETCC's.
6341 SDValue Simp = SimplifySetCC(TLI.getSetCCResultType(CondLHS.getValueType()),
6342 CondLHS, CondRHS, CC->get(), N->getDebugLoc(),
6344 if (Simp.getNode()) AddToWorkList(Simp.getNode());
6346 // fold to a simpler setcc
6347 if (Simp.getNode() && Simp.getOpcode() == ISD::SETCC)
6348 return DAG.getNode(ISD::BR_CC, N->getDebugLoc(), MVT::Other,
6349 N->getOperand(0), Simp.getOperand(2),
6350 Simp.getOperand(0), Simp.getOperand(1),
6356 /// canFoldInAddressingMode - Return true if 'Use' is a load or a store that
6357 /// uses N as its base pointer and that N may be folded in the load / store
6358 /// addressing mode.
6359 static bool canFoldInAddressingMode(SDNode *N, SDNode *Use,
6361 const TargetLowering &TLI) {
6363 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Use)) {
6364 if (LD->isIndexed() || LD->getBasePtr().getNode() != N)
6366 VT = Use->getValueType(0);
6367 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(Use)) {
6368 if (ST->isIndexed() || ST->getBasePtr().getNode() != N)
6370 VT = ST->getValue().getValueType();
6374 TargetLowering::AddrMode AM;
6375 if (N->getOpcode() == ISD::ADD) {
6376 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
6379 AM.BaseOffs = Offset->getSExtValue();
6383 } else if (N->getOpcode() == ISD::SUB) {
6384 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
6387 AM.BaseOffs = -Offset->getSExtValue();
6394 return TLI.isLegalAddressingMode(AM, VT.getTypeForEVT(*DAG.getContext()));
6397 /// CombineToPreIndexedLoadStore - Try turning a load / store into a
6398 /// pre-indexed load / store when the base pointer is an add or subtract
6399 /// and it has other uses besides the load / store. After the
6400 /// transformation, the new indexed load / store has effectively folded
6401 /// the add / subtract in and all of its other uses are redirected to the
6402 /// new load / store.
6403 bool DAGCombiner::CombineToPreIndexedLoadStore(SDNode *N) {
6404 if (Level < AfterLegalizeDAG)
6410 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
6411 if (LD->isIndexed())
6413 VT = LD->getMemoryVT();
6414 if (!TLI.isIndexedLoadLegal(ISD::PRE_INC, VT) &&
6415 !TLI.isIndexedLoadLegal(ISD::PRE_DEC, VT))
6417 Ptr = LD->getBasePtr();
6418 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
6419 if (ST->isIndexed())
6421 VT = ST->getMemoryVT();
6422 if (!TLI.isIndexedStoreLegal(ISD::PRE_INC, VT) &&
6423 !TLI.isIndexedStoreLegal(ISD::PRE_DEC, VT))
6425 Ptr = ST->getBasePtr();
6431 // If the pointer is not an add/sub, or if it doesn't have multiple uses, bail
6432 // out. There is no reason to make this a preinc/predec.
6433 if ((Ptr.getOpcode() != ISD::ADD && Ptr.getOpcode() != ISD::SUB) ||
6434 Ptr.getNode()->hasOneUse())
6437 // Ask the target to do addressing mode selection.
6440 ISD::MemIndexedMode AM = ISD::UNINDEXED;
6441 if (!TLI.getPreIndexedAddressParts(N, BasePtr, Offset, AM, DAG))
6443 // Don't create a indexed load / store with zero offset.
6444 if (isa<ConstantSDNode>(Offset) &&
6445 cast<ConstantSDNode>(Offset)->isNullValue())
6448 // Try turning it into a pre-indexed load / store except when:
6449 // 1) The new base ptr is a frame index.
6450 // 2) If N is a store and the new base ptr is either the same as or is a
6451 // predecessor of the value being stored.
6452 // 3) Another use of old base ptr is a predecessor of N. If ptr is folded
6453 // that would create a cycle.
6454 // 4) All uses are load / store ops that use it as old base ptr.
6456 // Check #1. Preinc'ing a frame index would require copying the stack pointer
6457 // (plus the implicit offset) to a register to preinc anyway.
6458 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
6463 SDValue Val = cast<StoreSDNode>(N)->getValue();
6464 if (Val == BasePtr || BasePtr.getNode()->isPredecessorOf(Val.getNode()))
6468 // Now check for #3 and #4.
6469 bool RealUse = false;
6471 // Caches for hasPredecessorHelper
6472 SmallPtrSet<const SDNode *, 32> Visited;
6473 SmallVector<const SDNode *, 16> Worklist;
6475 for (SDNode::use_iterator I = Ptr.getNode()->use_begin(),
6476 E = Ptr.getNode()->use_end(); I != E; ++I) {
6480 if (N->hasPredecessorHelper(Use, Visited, Worklist))
6483 // If Ptr may be folded in addressing mode of other use, then it's
6484 // not profitable to do this transformation.
6485 if (!canFoldInAddressingMode(Ptr.getNode(), Use, DAG, TLI))
6494 Result = DAG.getIndexedLoad(SDValue(N,0), N->getDebugLoc(),
6495 BasePtr, Offset, AM);
6497 Result = DAG.getIndexedStore(SDValue(N,0), N->getDebugLoc(),
6498 BasePtr, Offset, AM);
6501 DEBUG(dbgs() << "\nReplacing.4 ";
6503 dbgs() << "\nWith: ";
6504 Result.getNode()->dump(&DAG);
6506 WorkListRemover DeadNodes(*this);
6508 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
6509 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
6511 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
6514 // Finally, since the node is now dead, remove it from the graph.
6517 // Replace the uses of Ptr with uses of the updated base value.
6518 DAG.ReplaceAllUsesOfValueWith(Ptr, Result.getValue(isLoad ? 1 : 0));
6519 removeFromWorkList(Ptr.getNode());
6520 DAG.DeleteNode(Ptr.getNode());
6525 /// CombineToPostIndexedLoadStore - Try to combine a load / store with a
6526 /// add / sub of the base pointer node into a post-indexed load / store.
6527 /// The transformation folded the add / subtract into the new indexed
6528 /// load / store effectively and all of its uses are redirected to the
6529 /// new load / store.
6530 bool DAGCombiner::CombineToPostIndexedLoadStore(SDNode *N) {
6531 if (Level < AfterLegalizeDAG)
6537 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
6538 if (LD->isIndexed())
6540 VT = LD->getMemoryVT();
6541 if (!TLI.isIndexedLoadLegal(ISD::POST_INC, VT) &&
6542 !TLI.isIndexedLoadLegal(ISD::POST_DEC, VT))
6544 Ptr = LD->getBasePtr();
6545 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
6546 if (ST->isIndexed())
6548 VT = ST->getMemoryVT();
6549 if (!TLI.isIndexedStoreLegal(ISD::POST_INC, VT) &&
6550 !TLI.isIndexedStoreLegal(ISD::POST_DEC, VT))
6552 Ptr = ST->getBasePtr();
6558 if (Ptr.getNode()->hasOneUse())
6561 for (SDNode::use_iterator I = Ptr.getNode()->use_begin(),
6562 E = Ptr.getNode()->use_end(); I != E; ++I) {
6565 (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB))
6570 ISD::MemIndexedMode AM = ISD::UNINDEXED;
6571 if (TLI.getPostIndexedAddressParts(N, Op, BasePtr, Offset, AM, DAG)) {
6572 // Don't create a indexed load / store with zero offset.
6573 if (isa<ConstantSDNode>(Offset) &&
6574 cast<ConstantSDNode>(Offset)->isNullValue())
6577 // Try turning it into a post-indexed load / store except when
6578 // 1) All uses are load / store ops that use it as base ptr (and
6579 // it may be folded as addressing mmode).
6580 // 2) Op must be independent of N, i.e. Op is neither a predecessor
6581 // nor a successor of N. Otherwise, if Op is folded that would
6584 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
6588 bool TryNext = false;
6589 for (SDNode::use_iterator II = BasePtr.getNode()->use_begin(),
6590 EE = BasePtr.getNode()->use_end(); II != EE; ++II) {
6592 if (Use == Ptr.getNode())
6595 // If all the uses are load / store addresses, then don't do the
6597 if (Use->getOpcode() == ISD::ADD || Use->getOpcode() == ISD::SUB){
6598 bool RealUse = false;
6599 for (SDNode::use_iterator III = Use->use_begin(),
6600 EEE = Use->use_end(); III != EEE; ++III) {
6601 SDNode *UseUse = *III;
6602 if (!canFoldInAddressingMode(Use, UseUse, DAG, TLI))
6617 if (!Op->isPredecessorOf(N) && !N->isPredecessorOf(Op)) {
6618 SDValue Result = isLoad
6619 ? DAG.getIndexedLoad(SDValue(N,0), N->getDebugLoc(),
6620 BasePtr, Offset, AM)
6621 : DAG.getIndexedStore(SDValue(N,0), N->getDebugLoc(),
6622 BasePtr, Offset, AM);
6625 DEBUG(dbgs() << "\nReplacing.5 ";
6627 dbgs() << "\nWith: ";
6628 Result.getNode()->dump(&DAG);
6630 WorkListRemover DeadNodes(*this);
6632 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
6633 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
6635 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
6638 // Finally, since the node is now dead, remove it from the graph.
6641 // Replace the uses of Use with uses of the updated base value.
6642 DAG.ReplaceAllUsesOfValueWith(SDValue(Op, 0),
6643 Result.getValue(isLoad ? 1 : 0));
6644 removeFromWorkList(Op);
6654 SDValue DAGCombiner::visitLOAD(SDNode *N) {
6655 LoadSDNode *LD = cast<LoadSDNode>(N);
6656 SDValue Chain = LD->getChain();
6657 SDValue Ptr = LD->getBasePtr();
6659 // If load is not volatile and there are no uses of the loaded value (and
6660 // the updated indexed value in case of indexed loads), change uses of the
6661 // chain value into uses of the chain input (i.e. delete the dead load).
6662 if (!LD->isVolatile()) {
6663 if (N->getValueType(1) == MVT::Other) {
6665 if (!N->hasAnyUseOfValue(0)) {
6666 // It's not safe to use the two value CombineTo variant here. e.g.
6667 // v1, chain2 = load chain1, loc
6668 // v2, chain3 = load chain2, loc
6670 // Now we replace use of chain2 with chain1. This makes the second load
6671 // isomorphic to the one we are deleting, and thus makes this load live.
6672 DEBUG(dbgs() << "\nReplacing.6 ";
6674 dbgs() << "\nWith chain: ";
6675 Chain.getNode()->dump(&DAG);
6677 WorkListRemover DeadNodes(*this);
6678 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
6680 if (N->use_empty()) {
6681 removeFromWorkList(N);
6685 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6689 assert(N->getValueType(2) == MVT::Other && "Malformed indexed loads?");
6690 if (!N->hasAnyUseOfValue(0) && !N->hasAnyUseOfValue(1)) {
6691 SDValue Undef = DAG.getUNDEF(N->getValueType(0));
6692 DEBUG(dbgs() << "\nReplacing.7 ";
6694 dbgs() << "\nWith: ";
6695 Undef.getNode()->dump(&DAG);
6696 dbgs() << " and 2 other values\n");
6697 WorkListRemover DeadNodes(*this);
6698 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Undef);
6699 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1),
6700 DAG.getUNDEF(N->getValueType(1)));
6701 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 2), Chain);
6702 removeFromWorkList(N);
6704 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6709 // If this load is directly stored, replace the load value with the stored
6711 // TODO: Handle store large -> read small portion.
6712 // TODO: Handle TRUNCSTORE/LOADEXT
6713 if (ISD::isNormalLoad(N) && !LD->isVolatile()) {
6714 if (ISD::isNON_TRUNCStore(Chain.getNode())) {
6715 StoreSDNode *PrevST = cast<StoreSDNode>(Chain);
6716 if (PrevST->getBasePtr() == Ptr &&
6717 PrevST->getValue().getValueType() == N->getValueType(0))
6718 return CombineTo(N, Chain.getOperand(1), Chain);
6722 // Try to infer better alignment information than the load already has.
6723 if (OptLevel != CodeGenOpt::None && LD->isUnindexed()) {
6724 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
6725 if (Align > LD->getAlignment())
6726 return DAG.getExtLoad(LD->getExtensionType(), N->getDebugLoc(),
6727 LD->getValueType(0),
6728 Chain, Ptr, LD->getPointerInfo(),
6730 LD->isVolatile(), LD->isNonTemporal(), Align);
6735 // Walk up chain skipping non-aliasing memory nodes.
6736 SDValue BetterChain = FindBetterChain(N, Chain);
6738 // If there is a better chain.
6739 if (Chain != BetterChain) {
6742 // Replace the chain to void dependency.
6743 if (LD->getExtensionType() == ISD::NON_EXTLOAD) {
6744 ReplLoad = DAG.getLoad(N->getValueType(0), LD->getDebugLoc(),
6745 BetterChain, Ptr, LD->getPointerInfo(),
6746 LD->isVolatile(), LD->isNonTemporal(),
6747 LD->isInvariant(), LD->getAlignment());
6749 ReplLoad = DAG.getExtLoad(LD->getExtensionType(), LD->getDebugLoc(),
6750 LD->getValueType(0),
6751 BetterChain, Ptr, LD->getPointerInfo(),
6754 LD->isNonTemporal(),
6755 LD->getAlignment());
6758 // Create token factor to keep old chain connected.
6759 SDValue Token = DAG.getNode(ISD::TokenFactor, N->getDebugLoc(),
6760 MVT::Other, Chain, ReplLoad.getValue(1));
6762 // Make sure the new and old chains are cleaned up.
6763 AddToWorkList(Token.getNode());
6765 // Replace uses with load result and token factor. Don't add users
6767 return CombineTo(N, ReplLoad.getValue(0), Token, false);
6771 // Try transforming N to an indexed load.
6772 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
6773 return SDValue(N, 0);
6778 /// CheckForMaskedLoad - Check to see if V is (and load (ptr), imm), where the
6779 /// load is having specific bytes cleared out. If so, return the byte size
6780 /// being masked out and the shift amount.
6781 static std::pair<unsigned, unsigned>
6782 CheckForMaskedLoad(SDValue V, SDValue Ptr, SDValue Chain) {
6783 std::pair<unsigned, unsigned> Result(0, 0);
6785 // Check for the structure we're looking for.
6786 if (V->getOpcode() != ISD::AND ||
6787 !isa<ConstantSDNode>(V->getOperand(1)) ||
6788 !ISD::isNormalLoad(V->getOperand(0).getNode()))
6791 // Check the chain and pointer.
6792 LoadSDNode *LD = cast<LoadSDNode>(V->getOperand(0));
6793 if (LD->getBasePtr() != Ptr) return Result; // Not from same pointer.
6795 // The store should be chained directly to the load or be an operand of a
6797 if (LD == Chain.getNode())
6799 else if (Chain->getOpcode() != ISD::TokenFactor)
6800 return Result; // Fail.
6803 for (unsigned i = 0, e = Chain->getNumOperands(); i != e; ++i)
6804 if (Chain->getOperand(i).getNode() == LD) {
6808 if (!isOk) return Result;
6811 // This only handles simple types.
6812 if (V.getValueType() != MVT::i16 &&
6813 V.getValueType() != MVT::i32 &&
6814 V.getValueType() != MVT::i64)
6817 // Check the constant mask. Invert it so that the bits being masked out are
6818 // 0 and the bits being kept are 1. Use getSExtValue so that leading bits
6819 // follow the sign bit for uniformity.
6820 uint64_t NotMask = ~cast<ConstantSDNode>(V->getOperand(1))->getSExtValue();
6821 unsigned NotMaskLZ = CountLeadingZeros_64(NotMask);
6822 if (NotMaskLZ & 7) return Result; // Must be multiple of a byte.
6823 unsigned NotMaskTZ = CountTrailingZeros_64(NotMask);
6824 if (NotMaskTZ & 7) return Result; // Must be multiple of a byte.
6825 if (NotMaskLZ == 64) return Result; // All zero mask.
6827 // See if we have a continuous run of bits. If so, we have 0*1+0*
6828 if (CountTrailingOnes_64(NotMask >> NotMaskTZ)+NotMaskTZ+NotMaskLZ != 64)
6831 // Adjust NotMaskLZ down to be from the actual size of the int instead of i64.
6832 if (V.getValueType() != MVT::i64 && NotMaskLZ)
6833 NotMaskLZ -= 64-V.getValueSizeInBits();
6835 unsigned MaskedBytes = (V.getValueSizeInBits()-NotMaskLZ-NotMaskTZ)/8;
6836 switch (MaskedBytes) {
6840 default: return Result; // All one mask, or 5-byte mask.
6843 // Verify that the first bit starts at a multiple of mask so that the access
6844 // is aligned the same as the access width.
6845 if (NotMaskTZ && NotMaskTZ/8 % MaskedBytes) return Result;
6847 Result.first = MaskedBytes;
6848 Result.second = NotMaskTZ/8;
6853 /// ShrinkLoadReplaceStoreWithStore - Check to see if IVal is something that
6854 /// provides a value as specified by MaskInfo. If so, replace the specified
6855 /// store with a narrower store of truncated IVal.
6857 ShrinkLoadReplaceStoreWithStore(const std::pair<unsigned, unsigned> &MaskInfo,
6858 SDValue IVal, StoreSDNode *St,
6860 unsigned NumBytes = MaskInfo.first;
6861 unsigned ByteShift = MaskInfo.second;
6862 SelectionDAG &DAG = DC->getDAG();
6864 // Check to see if IVal is all zeros in the part being masked in by the 'or'
6865 // that uses this. If not, this is not a replacement.
6866 APInt Mask = ~APInt::getBitsSet(IVal.getValueSizeInBits(),
6867 ByteShift*8, (ByteShift+NumBytes)*8);
6868 if (!DAG.MaskedValueIsZero(IVal, Mask)) return 0;
6870 // Check that it is legal on the target to do this. It is legal if the new
6871 // VT we're shrinking to (i8/i16/i32) is legal or we're still before type
6873 MVT VT = MVT::getIntegerVT(NumBytes*8);
6874 if (!DC->isTypeLegal(VT))
6877 // Okay, we can do this! Replace the 'St' store with a store of IVal that is
6878 // shifted by ByteShift and truncated down to NumBytes.
6880 IVal = DAG.getNode(ISD::SRL, IVal->getDebugLoc(), IVal.getValueType(), IVal,
6881 DAG.getConstant(ByteShift*8,
6882 DC->getShiftAmountTy(IVal.getValueType())));
6884 // Figure out the offset for the store and the alignment of the access.
6886 unsigned NewAlign = St->getAlignment();
6888 if (DAG.getTargetLoweringInfo().isLittleEndian())
6889 StOffset = ByteShift;
6891 StOffset = IVal.getValueType().getStoreSize() - ByteShift - NumBytes;
6893 SDValue Ptr = St->getBasePtr();
6895 Ptr = DAG.getNode(ISD::ADD, IVal->getDebugLoc(), Ptr.getValueType(),
6896 Ptr, DAG.getConstant(StOffset, Ptr.getValueType()));
6897 NewAlign = MinAlign(NewAlign, StOffset);
6900 // Truncate down to the new size.
6901 IVal = DAG.getNode(ISD::TRUNCATE, IVal->getDebugLoc(), VT, IVal);
6904 return DAG.getStore(St->getChain(), St->getDebugLoc(), IVal, Ptr,
6905 St->getPointerInfo().getWithOffset(StOffset),
6906 false, false, NewAlign).getNode();
6910 /// ReduceLoadOpStoreWidth - Look for sequence of load / op / store where op is
6911 /// one of 'or', 'xor', and 'and' of immediates. If 'op' is only touching some
6912 /// of the loaded bits, try narrowing the load and store if it would end up
6913 /// being a win for performance or code size.
6914 SDValue DAGCombiner::ReduceLoadOpStoreWidth(SDNode *N) {
6915 StoreSDNode *ST = cast<StoreSDNode>(N);
6916 if (ST->isVolatile())
6919 SDValue Chain = ST->getChain();
6920 SDValue Value = ST->getValue();
6921 SDValue Ptr = ST->getBasePtr();
6922 EVT VT = Value.getValueType();
6924 if (ST->isTruncatingStore() || VT.isVector() || !Value.hasOneUse())
6927 unsigned Opc = Value.getOpcode();
6929 // If this is "store (or X, Y), P" and X is "(and (load P), cst)", where cst
6930 // is a byte mask indicating a consecutive number of bytes, check to see if
6931 // Y is known to provide just those bytes. If so, we try to replace the
6932 // load + replace + store sequence with a single (narrower) store, which makes
6934 if (Opc == ISD::OR) {
6935 std::pair<unsigned, unsigned> MaskedLoad;
6936 MaskedLoad = CheckForMaskedLoad(Value.getOperand(0), Ptr, Chain);
6937 if (MaskedLoad.first)
6938 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
6939 Value.getOperand(1), ST,this))
6940 return SDValue(NewST, 0);
6942 // Or is commutative, so try swapping X and Y.
6943 MaskedLoad = CheckForMaskedLoad(Value.getOperand(1), Ptr, Chain);
6944 if (MaskedLoad.first)
6945 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
6946 Value.getOperand(0), ST,this))
6947 return SDValue(NewST, 0);
6950 if ((Opc != ISD::OR && Opc != ISD::XOR && Opc != ISD::AND) ||
6951 Value.getOperand(1).getOpcode() != ISD::Constant)
6954 SDValue N0 = Value.getOperand(0);
6955 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
6956 Chain == SDValue(N0.getNode(), 1)) {
6957 LoadSDNode *LD = cast<LoadSDNode>(N0);
6958 if (LD->getBasePtr() != Ptr ||
6959 LD->getPointerInfo().getAddrSpace() !=
6960 ST->getPointerInfo().getAddrSpace())
6963 // Find the type to narrow it the load / op / store to.
6964 SDValue N1 = Value.getOperand(1);
6965 unsigned BitWidth = N1.getValueSizeInBits();
6966 APInt Imm = cast<ConstantSDNode>(N1)->getAPIntValue();
6967 if (Opc == ISD::AND)
6968 Imm ^= APInt::getAllOnesValue(BitWidth);
6969 if (Imm == 0 || Imm.isAllOnesValue())
6971 unsigned ShAmt = Imm.countTrailingZeros();
6972 unsigned MSB = BitWidth - Imm.countLeadingZeros() - 1;
6973 unsigned NewBW = NextPowerOf2(MSB - ShAmt);
6974 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
6975 while (NewBW < BitWidth &&
6976 !(TLI.isOperationLegalOrCustom(Opc, NewVT) &&
6977 TLI.isNarrowingProfitable(VT, NewVT))) {
6978 NewBW = NextPowerOf2(NewBW);
6979 NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
6981 if (NewBW >= BitWidth)
6984 // If the lsb changed does not start at the type bitwidth boundary,
6985 // start at the previous one.
6987 ShAmt = (((ShAmt + NewBW - 1) / NewBW) * NewBW) - NewBW;
6988 APInt Mask = APInt::getBitsSet(BitWidth, ShAmt, ShAmt + NewBW);
6989 if ((Imm & Mask) == Imm) {
6990 APInt NewImm = (Imm & Mask).lshr(ShAmt).trunc(NewBW);
6991 if (Opc == ISD::AND)
6992 NewImm ^= APInt::getAllOnesValue(NewBW);
6993 uint64_t PtrOff = ShAmt / 8;
6994 // For big endian targets, we need to adjust the offset to the pointer to
6995 // load the correct bytes.
6996 if (TLI.isBigEndian())
6997 PtrOff = (BitWidth + 7 - NewBW) / 8 - PtrOff;
6999 unsigned NewAlign = MinAlign(LD->getAlignment(), PtrOff);
7000 Type *NewVTTy = NewVT.getTypeForEVT(*DAG.getContext());
7001 if (NewAlign < TLI.getTargetData()->getABITypeAlignment(NewVTTy))
7004 SDValue NewPtr = DAG.getNode(ISD::ADD, LD->getDebugLoc(),
7005 Ptr.getValueType(), Ptr,
7006 DAG.getConstant(PtrOff, Ptr.getValueType()));
7007 SDValue NewLD = DAG.getLoad(NewVT, N0.getDebugLoc(),
7008 LD->getChain(), NewPtr,
7009 LD->getPointerInfo().getWithOffset(PtrOff),
7010 LD->isVolatile(), LD->isNonTemporal(),
7011 LD->isInvariant(), NewAlign);
7012 SDValue NewVal = DAG.getNode(Opc, Value.getDebugLoc(), NewVT, NewLD,
7013 DAG.getConstant(NewImm, NewVT));
7014 SDValue NewST = DAG.getStore(Chain, N->getDebugLoc(),
7016 ST->getPointerInfo().getWithOffset(PtrOff),
7017 false, false, NewAlign);
7019 AddToWorkList(NewPtr.getNode());
7020 AddToWorkList(NewLD.getNode());
7021 AddToWorkList(NewVal.getNode());
7022 WorkListRemover DeadNodes(*this);
7023 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLD.getValue(1));
7032 /// TransformFPLoadStorePair - For a given floating point load / store pair,
7033 /// if the load value isn't used by any other operations, then consider
7034 /// transforming the pair to integer load / store operations if the target
7035 /// deems the transformation profitable.
7036 SDValue DAGCombiner::TransformFPLoadStorePair(SDNode *N) {
7037 StoreSDNode *ST = cast<StoreSDNode>(N);
7038 SDValue Chain = ST->getChain();
7039 SDValue Value = ST->getValue();
7040 if (ISD::isNormalStore(ST) && ISD::isNormalLoad(Value.getNode()) &&
7041 Value.hasOneUse() &&
7042 Chain == SDValue(Value.getNode(), 1)) {
7043 LoadSDNode *LD = cast<LoadSDNode>(Value);
7044 EVT VT = LD->getMemoryVT();
7045 if (!VT.isFloatingPoint() ||
7046 VT != ST->getMemoryVT() ||
7047 LD->isNonTemporal() ||
7048 ST->isNonTemporal() ||
7049 LD->getPointerInfo().getAddrSpace() != 0 ||
7050 ST->getPointerInfo().getAddrSpace() != 0)
7053 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
7054 if (!TLI.isOperationLegal(ISD::LOAD, IntVT) ||
7055 !TLI.isOperationLegal(ISD::STORE, IntVT) ||
7056 !TLI.isDesirableToTransformToIntegerOp(ISD::LOAD, VT) ||
7057 !TLI.isDesirableToTransformToIntegerOp(ISD::STORE, VT))
7060 unsigned LDAlign = LD->getAlignment();
7061 unsigned STAlign = ST->getAlignment();
7062 Type *IntVTTy = IntVT.getTypeForEVT(*DAG.getContext());
7063 unsigned ABIAlign = TLI.getTargetData()->getABITypeAlignment(IntVTTy);
7064 if (LDAlign < ABIAlign || STAlign < ABIAlign)
7067 SDValue NewLD = DAG.getLoad(IntVT, Value.getDebugLoc(),
7068 LD->getChain(), LD->getBasePtr(),
7069 LD->getPointerInfo(),
7070 false, false, false, LDAlign);
7072 SDValue NewST = DAG.getStore(NewLD.getValue(1), N->getDebugLoc(),
7073 NewLD, ST->getBasePtr(),
7074 ST->getPointerInfo(),
7075 false, false, STAlign);
7077 AddToWorkList(NewLD.getNode());
7078 AddToWorkList(NewST.getNode());
7079 WorkListRemover DeadNodes(*this);
7080 DAG.ReplaceAllUsesOfValueWith(Value.getValue(1), NewLD.getValue(1));
7088 SDValue DAGCombiner::visitSTORE(SDNode *N) {
7089 StoreSDNode *ST = cast<StoreSDNode>(N);
7090 SDValue Chain = ST->getChain();
7091 SDValue Value = ST->getValue();
7092 SDValue Ptr = ST->getBasePtr();
7094 // If this is a store of a bit convert, store the input value if the
7095 // resultant store does not need a higher alignment than the original.
7096 if (Value.getOpcode() == ISD::BITCAST && !ST->isTruncatingStore() &&
7097 ST->isUnindexed()) {
7098 unsigned OrigAlign = ST->getAlignment();
7099 EVT SVT = Value.getOperand(0).getValueType();
7100 unsigned Align = TLI.getTargetData()->
7101 getABITypeAlignment(SVT.getTypeForEVT(*DAG.getContext()));
7102 if (Align <= OrigAlign &&
7103 ((!LegalOperations && !ST->isVolatile()) ||
7104 TLI.isOperationLegalOrCustom(ISD::STORE, SVT)))
7105 return DAG.getStore(Chain, N->getDebugLoc(), Value.getOperand(0),
7106 Ptr, ST->getPointerInfo(), ST->isVolatile(),
7107 ST->isNonTemporal(), OrigAlign);
7110 // Turn 'store undef, Ptr' -> nothing.
7111 if (Value.getOpcode() == ISD::UNDEF && ST->isUnindexed())
7114 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
7115 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Value)) {
7116 // NOTE: If the original store is volatile, this transform must not increase
7117 // the number of stores. For example, on x86-32 an f64 can be stored in one
7118 // processor operation but an i64 (which is not legal) requires two. So the
7119 // transform should not be done in this case.
7120 if (Value.getOpcode() != ISD::TargetConstantFP) {
7122 switch (CFP->getValueType(0).getSimpleVT().SimpleTy) {
7123 default: llvm_unreachable("Unknown FP type");
7124 case MVT::f80: // We don't do this for these yet.
7129 if ((isTypeLegal(MVT::i32) && !LegalOperations && !ST->isVolatile()) ||
7130 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
7131 Tmp = DAG.getConstant((uint32_t)CFP->getValueAPF().
7132 bitcastToAPInt().getZExtValue(), MVT::i32);
7133 return DAG.getStore(Chain, N->getDebugLoc(), Tmp,
7134 Ptr, ST->getPointerInfo(), ST->isVolatile(),
7135 ST->isNonTemporal(), ST->getAlignment());
7139 if ((TLI.isTypeLegal(MVT::i64) && !LegalOperations &&
7140 !ST->isVolatile()) ||
7141 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i64)) {
7142 Tmp = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
7143 getZExtValue(), MVT::i64);
7144 return DAG.getStore(Chain, N->getDebugLoc(), Tmp,
7145 Ptr, ST->getPointerInfo(), ST->isVolatile(),
7146 ST->isNonTemporal(), ST->getAlignment());
7149 if (!ST->isVolatile() &&
7150 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
7151 // Many FP stores are not made apparent until after legalize, e.g. for
7152 // argument passing. Since this is so common, custom legalize the
7153 // 64-bit integer store into two 32-bit stores.
7154 uint64_t Val = CFP->getValueAPF().bitcastToAPInt().getZExtValue();
7155 SDValue Lo = DAG.getConstant(Val & 0xFFFFFFFF, MVT::i32);
7156 SDValue Hi = DAG.getConstant(Val >> 32, MVT::i32);
7157 if (TLI.isBigEndian()) std::swap(Lo, Hi);
7159 unsigned Alignment = ST->getAlignment();
7160 bool isVolatile = ST->isVolatile();
7161 bool isNonTemporal = ST->isNonTemporal();
7163 SDValue St0 = DAG.getStore(Chain, ST->getDebugLoc(), Lo,
7164 Ptr, ST->getPointerInfo(),
7165 isVolatile, isNonTemporal,
7166 ST->getAlignment());
7167 Ptr = DAG.getNode(ISD::ADD, N->getDebugLoc(), Ptr.getValueType(), Ptr,
7168 DAG.getConstant(4, Ptr.getValueType()));
7169 Alignment = MinAlign(Alignment, 4U);
7170 SDValue St1 = DAG.getStore(Chain, ST->getDebugLoc(), Hi,
7171 Ptr, ST->getPointerInfo().getWithOffset(4),
7172 isVolatile, isNonTemporal,
7174 return DAG.getNode(ISD::TokenFactor, N->getDebugLoc(), MVT::Other,
7183 // Try to infer better alignment information than the store already has.
7184 if (OptLevel != CodeGenOpt::None && ST->isUnindexed()) {
7185 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
7186 if (Align > ST->getAlignment())
7187 return DAG.getTruncStore(Chain, N->getDebugLoc(), Value,
7188 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
7189 ST->isVolatile(), ST->isNonTemporal(), Align);
7193 // Try transforming a pair floating point load / store ops to integer
7194 // load / store ops.
7195 SDValue NewST = TransformFPLoadStorePair(N);
7196 if (NewST.getNode())
7200 // Walk up chain skipping non-aliasing memory nodes.
7201 SDValue BetterChain = FindBetterChain(N, Chain);
7203 // If there is a better chain.
7204 if (Chain != BetterChain) {
7207 // Replace the chain to avoid dependency.
7208 if (ST->isTruncatingStore()) {
7209 ReplStore = DAG.getTruncStore(BetterChain, N->getDebugLoc(), Value, Ptr,
7210 ST->getPointerInfo(),
7211 ST->getMemoryVT(), ST->isVolatile(),
7212 ST->isNonTemporal(), ST->getAlignment());
7214 ReplStore = DAG.getStore(BetterChain, N->getDebugLoc(), Value, Ptr,
7215 ST->getPointerInfo(),
7216 ST->isVolatile(), ST->isNonTemporal(),
7217 ST->getAlignment());
7220 // Create token to keep both nodes around.
7221 SDValue Token = DAG.getNode(ISD::TokenFactor, N->getDebugLoc(),
7222 MVT::Other, Chain, ReplStore);
7224 // Make sure the new and old chains are cleaned up.
7225 AddToWorkList(Token.getNode());
7227 // Don't add users to work list.
7228 return CombineTo(N, Token, false);
7232 // Try transforming N to an indexed store.
7233 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
7234 return SDValue(N, 0);
7236 // FIXME: is there such a thing as a truncating indexed store?
7237 if (ST->isTruncatingStore() && ST->isUnindexed() &&
7238 Value.getValueType().isInteger()) {
7239 // See if we can simplify the input to this truncstore with knowledge that
7240 // only the low bits are being used. For example:
7241 // "truncstore (or (shl x, 8), y), i8" -> "truncstore y, i8"
7243 GetDemandedBits(Value,
7244 APInt::getLowBitsSet(
7245 Value.getValueType().getScalarType().getSizeInBits(),
7246 ST->getMemoryVT().getScalarType().getSizeInBits()));
7247 AddToWorkList(Value.getNode());
7248 if (Shorter.getNode())
7249 return DAG.getTruncStore(Chain, N->getDebugLoc(), Shorter,
7250 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
7251 ST->isVolatile(), ST->isNonTemporal(),
7252 ST->getAlignment());
7254 // Otherwise, see if we can simplify the operation with
7255 // SimplifyDemandedBits, which only works if the value has a single use.
7256 if (SimplifyDemandedBits(Value,
7257 APInt::getLowBitsSet(
7258 Value.getValueType().getScalarType().getSizeInBits(),
7259 ST->getMemoryVT().getScalarType().getSizeInBits())))
7260 return SDValue(N, 0);
7263 // If this is a load followed by a store to the same location, then the store
7265 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Value)) {
7266 if (Ld->getBasePtr() == Ptr && ST->getMemoryVT() == Ld->getMemoryVT() &&
7267 ST->isUnindexed() && !ST->isVolatile() &&
7268 // There can't be any side effects between the load and store, such as
7270 Chain.reachesChainWithoutSideEffects(SDValue(Ld, 1))) {
7271 // The store is dead, remove it.
7276 // If this is an FP_ROUND or TRUNC followed by a store, fold this into a
7277 // truncating store. We can do this even if this is already a truncstore.
7278 if ((Value.getOpcode() == ISD::FP_ROUND || Value.getOpcode() == ISD::TRUNCATE)
7279 && Value.getNode()->hasOneUse() && ST->isUnindexed() &&
7280 TLI.isTruncStoreLegal(Value.getOperand(0).getValueType(),
7281 ST->getMemoryVT())) {
7282 return DAG.getTruncStore(Chain, N->getDebugLoc(), Value.getOperand(0),
7283 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
7284 ST->isVolatile(), ST->isNonTemporal(),
7285 ST->getAlignment());
7288 return ReduceLoadOpStoreWidth(N);
7291 SDValue DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) {
7292 SDValue InVec = N->getOperand(0);
7293 SDValue InVal = N->getOperand(1);
7294 SDValue EltNo = N->getOperand(2);
7295 DebugLoc dl = N->getDebugLoc();
7297 // If the inserted element is an UNDEF, just use the input vector.
7298 if (InVal.getOpcode() == ISD::UNDEF)
7301 EVT VT = InVec.getValueType();
7303 // If we can't generate a legal BUILD_VECTOR, exit
7304 if (LegalOperations && !TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
7307 // Check that we know which element is being inserted
7308 if (!isa<ConstantSDNode>(EltNo))
7310 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
7312 // Check that the operand is a BUILD_VECTOR (or UNDEF, which can essentially
7313 // be converted to a BUILD_VECTOR). Fill in the Ops vector with the
7315 SmallVector<SDValue, 8> Ops;
7316 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
7317 Ops.append(InVec.getNode()->op_begin(),
7318 InVec.getNode()->op_end());
7319 } else if (InVec.getOpcode() == ISD::UNDEF) {
7320 unsigned NElts = VT.getVectorNumElements();
7321 Ops.append(NElts, DAG.getUNDEF(InVal.getValueType()));
7326 // Insert the element
7327 if (Elt < Ops.size()) {
7328 // All the operands of BUILD_VECTOR must have the same type;
7329 // we enforce that here.
7330 EVT OpVT = Ops[0].getValueType();
7331 if (InVal.getValueType() != OpVT)
7332 InVal = OpVT.bitsGT(InVal.getValueType()) ?
7333 DAG.getNode(ISD::ANY_EXTEND, dl, OpVT, InVal) :
7334 DAG.getNode(ISD::TRUNCATE, dl, OpVT, InVal);
7338 // Return the new vector
7339 return DAG.getNode(ISD::BUILD_VECTOR, dl,
7340 VT, &Ops[0], Ops.size());
7343 SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) {
7344 // (vextract (scalar_to_vector val, 0) -> val
7345 SDValue InVec = N->getOperand(0);
7346 EVT VT = InVec.getValueType();
7347 EVT NVT = N->getValueType(0);
7349 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
7350 // Check if the result type doesn't match the inserted element type. A
7351 // SCALAR_TO_VECTOR may truncate the inserted element and the
7352 // EXTRACT_VECTOR_ELT may widen the extracted vector.
7353 SDValue InOp = InVec.getOperand(0);
7354 if (InOp.getValueType() != NVT) {
7355 assert(InOp.getValueType().isInteger() && NVT.isInteger());
7356 return DAG.getSExtOrTrunc(InOp, InVec.getDebugLoc(), NVT);
7361 SDValue EltNo = N->getOperand(1);
7362 bool ConstEltNo = isa<ConstantSDNode>(EltNo);
7364 // Transform: (EXTRACT_VECTOR_ELT( VECTOR_SHUFFLE )) -> EXTRACT_VECTOR_ELT.
7365 // We only perform this optimization before the op legalization phase because
7366 // we may introduce new vector instructions which are not backed by TD patterns.
7367 // For example on AVX, extracting elements from a wide vector without using
7368 // extract_subvector.
7369 if (InVec.getOpcode() == ISD::VECTOR_SHUFFLE
7370 && ConstEltNo && !LegalOperations) {
7371 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
7372 int NumElem = VT.getVectorNumElements();
7373 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(InVec);
7374 // Find the new index to extract from.
7375 int OrigElt = SVOp->getMaskElt(Elt);
7377 // Extracting an undef index is undef.
7379 return DAG.getUNDEF(NVT);
7381 // Select the right vector half to extract from.
7382 if (OrigElt < NumElem) {
7383 InVec = InVec->getOperand(0);
7385 InVec = InVec->getOperand(1);
7389 EVT IndexTy = N->getOperand(1).getValueType();
7390 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, N->getDebugLoc(), NVT,
7391 InVec, DAG.getConstant(OrigElt, IndexTy));
7394 // Perform only after legalization to ensure build_vector / vector_shuffle
7395 // optimizations have already been done.
7396 if (!LegalOperations) return SDValue();
7398 // (vextract (v4f32 load $addr), c) -> (f32 load $addr+c*size)
7399 // (vextract (v4f32 s2v (f32 load $addr)), c) -> (f32 load $addr+c*size)
7400 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), 0) -> (f32 load $addr)
7403 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
7404 bool NewLoad = false;
7405 bool BCNumEltsChanged = false;
7406 EVT ExtVT = VT.getVectorElementType();
7409 // If the result of load has to be truncated, then it's not necessarily
7411 if (NVT.bitsLT(LVT) && !TLI.isTruncateFree(LVT, NVT))
7414 if (InVec.getOpcode() == ISD::BITCAST) {
7415 // Don't duplicate a load with other uses.
7416 if (!InVec.hasOneUse())
7419 EVT BCVT = InVec.getOperand(0).getValueType();
7420 if (!BCVT.isVector() || ExtVT.bitsGT(BCVT.getVectorElementType()))
7422 if (VT.getVectorNumElements() != BCVT.getVectorNumElements())
7423 BCNumEltsChanged = true;
7424 InVec = InVec.getOperand(0);
7425 ExtVT = BCVT.getVectorElementType();
7429 LoadSDNode *LN0 = NULL;
7430 const ShuffleVectorSDNode *SVN = NULL;
7431 if (ISD::isNormalLoad(InVec.getNode())) {
7432 LN0 = cast<LoadSDNode>(InVec);
7433 } else if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR &&
7434 InVec.getOperand(0).getValueType() == ExtVT &&
7435 ISD::isNormalLoad(InVec.getOperand(0).getNode())) {
7436 // Don't duplicate a load with other uses.
7437 if (!InVec.hasOneUse())
7440 LN0 = cast<LoadSDNode>(InVec.getOperand(0));
7441 } else if ((SVN = dyn_cast<ShuffleVectorSDNode>(InVec))) {
7442 // (vextract (vector_shuffle (load $addr), v2, <1, u, u, u>), 1)
7444 // (load $addr+1*size)
7446 // Don't duplicate a load with other uses.
7447 if (!InVec.hasOneUse())
7450 // If the bit convert changed the number of elements, it is unsafe
7451 // to examine the mask.
7452 if (BCNumEltsChanged)
7455 // Select the input vector, guarding against out of range extract vector.
7456 unsigned NumElems = VT.getVectorNumElements();
7457 int Idx = (Elt > (int)NumElems) ? -1 : SVN->getMaskElt(Elt);
7458 InVec = (Idx < (int)NumElems) ? InVec.getOperand(0) : InVec.getOperand(1);
7460 if (InVec.getOpcode() == ISD::BITCAST) {
7461 // Don't duplicate a load with other uses.
7462 if (!InVec.hasOneUse())
7465 InVec = InVec.getOperand(0);
7467 if (ISD::isNormalLoad(InVec.getNode())) {
7468 LN0 = cast<LoadSDNode>(InVec);
7469 Elt = (Idx < (int)NumElems) ? Idx : Idx - (int)NumElems;
7473 // Make sure we found a non-volatile load and the extractelement is
7475 if (!LN0 || !LN0->hasNUsesOfValue(1,0) || LN0->isVolatile())
7478 // If Idx was -1 above, Elt is going to be -1, so just return undef.
7480 return DAG.getUNDEF(LVT);
7482 unsigned Align = LN0->getAlignment();
7484 // Check the resultant load doesn't need a higher alignment than the
7488 ->getABITypeAlignment(LVT.getTypeForEVT(*DAG.getContext()));
7490 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, LVT))
7496 SDValue NewPtr = LN0->getBasePtr();
7497 unsigned PtrOff = 0;
7500 PtrOff = LVT.getSizeInBits() * Elt / 8;
7501 EVT PtrType = NewPtr.getValueType();
7502 if (TLI.isBigEndian())
7503 PtrOff = VT.getSizeInBits() / 8 - PtrOff;
7504 NewPtr = DAG.getNode(ISD::ADD, N->getDebugLoc(), PtrType, NewPtr,
7505 DAG.getConstant(PtrOff, PtrType));
7508 // The replacement we need to do here is a little tricky: we need to
7509 // replace an extractelement of a load with a load.
7510 // Use ReplaceAllUsesOfValuesWith to do the replacement.
7511 // Note that this replacement assumes that the extractvalue is the only
7512 // use of the load; that's okay because we don't want to perform this
7513 // transformation in other cases anyway.
7516 if (NVT.bitsGT(LVT)) {
7517 // If the result type of vextract is wider than the load, then issue an
7518 // extending load instead.
7519 ISD::LoadExtType ExtType = TLI.isLoadExtLegal(ISD::ZEXTLOAD, LVT)
7520 ? ISD::ZEXTLOAD : ISD::EXTLOAD;
7521 Load = DAG.getExtLoad(ExtType, N->getDebugLoc(), NVT, LN0->getChain(),
7522 NewPtr, LN0->getPointerInfo().getWithOffset(PtrOff),
7523 LVT, LN0->isVolatile(), LN0->isNonTemporal(),Align);
7524 Chain = Load.getValue(1);
7526 Load = DAG.getLoad(LVT, N->getDebugLoc(), LN0->getChain(), NewPtr,
7527 LN0->getPointerInfo().getWithOffset(PtrOff),
7528 LN0->isVolatile(), LN0->isNonTemporal(),
7529 LN0->isInvariant(), Align);
7530 Chain = Load.getValue(1);
7531 if (NVT.bitsLT(LVT))
7532 Load = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), NVT, Load);
7534 Load = DAG.getNode(ISD::BITCAST, N->getDebugLoc(), NVT, Load);
7536 WorkListRemover DeadNodes(*this);
7537 SDValue From[] = { SDValue(N, 0), SDValue(LN0,1) };
7538 SDValue To[] = { Load, Chain };
7539 DAG.ReplaceAllUsesOfValuesWith(From, To, 2);
7540 // Since we're explcitly calling ReplaceAllUses, add the new node to the
7541 // worklist explicitly as well.
7542 AddToWorkList(Load.getNode());
7543 AddUsersToWorkList(Load.getNode()); // Add users too
7544 // Make sure to revisit this node to clean it up; it will usually be dead.
7546 return SDValue(N, 0);
7552 SDValue DAGCombiner::visitBUILD_VECTOR(SDNode *N) {
7553 unsigned NumInScalars = N->getNumOperands();
7554 DebugLoc dl = N->getDebugLoc();
7555 EVT VT = N->getValueType(0);
7556 // Check to see if this is a BUILD_VECTOR of a bunch of values
7557 // which come from any_extend or zero_extend nodes. If so, we can create
7558 // a new BUILD_VECTOR using bit-casts which may enable other BUILD_VECTOR
7559 // optimizations. We do not handle sign-extend because we can't fill the sign
7561 EVT SourceType = MVT::Other;
7562 bool AllAnyExt = true;
7563 bool AllUndef = true;
7564 for (unsigned i = 0; i != NumInScalars; ++i) {
7565 SDValue In = N->getOperand(i);
7566 // Ignore undef inputs.
7567 if (In.getOpcode() == ISD::UNDEF) continue;
7570 bool AnyExt = In.getOpcode() == ISD::ANY_EXTEND;
7571 bool ZeroExt = In.getOpcode() == ISD::ZERO_EXTEND;
7573 // Abort if the element is not an extension.
7574 if (!ZeroExt && !AnyExt) {
7575 SourceType = MVT::Other;
7579 // The input is a ZeroExt or AnyExt. Check the original type.
7580 EVT InTy = In.getOperand(0).getValueType();
7582 // Check that all of the widened source types are the same.
7583 if (SourceType == MVT::Other)
7586 else if (InTy != SourceType) {
7587 // Multiple income types. Abort.
7588 SourceType = MVT::Other;
7592 // Check if all of the extends are ANY_EXTENDs.
7593 AllAnyExt &= AnyExt;
7597 return DAG.getUNDEF(VT);
7599 // In order to have valid types, all of the inputs must be extended from the
7600 // same source type and all of the inputs must be any or zero extend.
7601 // Scalar sizes must be a power of two.
7602 EVT OutScalarTy = N->getValueType(0).getScalarType();
7603 bool ValidTypes = SourceType != MVT::Other &&
7604 isPowerOf2_32(OutScalarTy.getSizeInBits()) &&
7605 isPowerOf2_32(SourceType.getSizeInBits());
7607 // We perform this optimization post type-legalization because
7608 // the type-legalizer often scalarizes integer-promoted vectors.
7609 // Performing this optimization before may create bit-casts which
7610 // will be type-legalized to complex code sequences.
7611 // We perform this optimization only before the operation legalizer because we
7612 // may introduce illegal operations.
7613 // Create a new simpler BUILD_VECTOR sequence which other optimizations can
7614 // turn into a single shuffle instruction.
7615 if ((Level == AfterLegalizeVectorOps || Level == AfterLegalizeTypes) &&
7617 bool isLE = TLI.isLittleEndian();
7618 unsigned ElemRatio = OutScalarTy.getSizeInBits()/SourceType.getSizeInBits();
7619 assert(ElemRatio > 1 && "Invalid element size ratio");
7620 SDValue Filler = AllAnyExt ? DAG.getUNDEF(SourceType):
7621 DAG.getConstant(0, SourceType);
7623 unsigned NewBVElems = ElemRatio * N->getValueType(0).getVectorNumElements();
7624 SmallVector<SDValue, 8> Ops(NewBVElems, Filler);
7626 // Populate the new build_vector
7627 for (unsigned i=0; i < N->getNumOperands(); ++i) {
7628 SDValue Cast = N->getOperand(i);
7629 assert((Cast.getOpcode() == ISD::ANY_EXTEND ||
7630 Cast.getOpcode() == ISD::ZERO_EXTEND ||
7631 Cast.getOpcode() == ISD::UNDEF) && "Invalid cast opcode");
7633 if (Cast.getOpcode() == ISD::UNDEF)
7634 In = DAG.getUNDEF(SourceType);
7636 In = Cast->getOperand(0);
7637 unsigned Index = isLE ? (i * ElemRatio) :
7638 (i * ElemRatio + (ElemRatio - 1));
7640 assert(Index < Ops.size() && "Invalid index");
7644 // The type of the new BUILD_VECTOR node.
7645 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), SourceType, NewBVElems);
7646 assert(VecVT.getSizeInBits() == N->getValueType(0).getSizeInBits() &&
7647 "Invalid vector size");
7648 // Check if the new vector type is legal.
7649 if (!isTypeLegal(VecVT)) return SDValue();
7651 // Make the new BUILD_VECTOR.
7652 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
7653 VecVT, &Ops[0], Ops.size());
7655 // The new BUILD_VECTOR node has the potential to be further optimized.
7656 AddToWorkList(BV.getNode());
7657 // Bitcast to the desired type.
7658 return DAG.getNode(ISD::BITCAST, dl, N->getValueType(0), BV);
7661 // Check to see if this is a BUILD_VECTOR of a bunch of EXTRACT_VECTOR_ELT
7662 // operations. If so, and if the EXTRACT_VECTOR_ELT vector inputs come from
7663 // at most two distinct vectors, turn this into a shuffle node.
7665 // May only combine to shuffle after legalize if shuffle is legal.
7666 if (LegalOperations &&
7667 !TLI.isOperationLegalOrCustom(ISD::VECTOR_SHUFFLE, VT))
7670 SDValue VecIn1, VecIn2;
7671 for (unsigned i = 0; i != NumInScalars; ++i) {
7672 // Ignore undef inputs.
7673 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
7675 // If this input is something other than a EXTRACT_VECTOR_ELT with a
7676 // constant index, bail out.
7677 if (N->getOperand(i).getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
7678 !isa<ConstantSDNode>(N->getOperand(i).getOperand(1))) {
7679 VecIn1 = VecIn2 = SDValue(0, 0);
7683 // We allow up to two distinct input vectors.
7684 SDValue ExtractedFromVec = N->getOperand(i).getOperand(0);
7685 if (ExtractedFromVec == VecIn1 || ExtractedFromVec == VecIn2)
7688 if (VecIn1.getNode() == 0) {
7689 VecIn1 = ExtractedFromVec;
7690 } else if (VecIn2.getNode() == 0) {
7691 VecIn2 = ExtractedFromVec;
7694 VecIn1 = VecIn2 = SDValue(0, 0);
7699 // If everything is good, we can make a shuffle operation.
7700 if (VecIn1.getNode()) {
7701 SmallVector<int, 8> Mask;
7702 for (unsigned i = 0; i != NumInScalars; ++i) {
7703 if (N->getOperand(i).getOpcode() == ISD::UNDEF) {
7708 // If extracting from the first vector, just use the index directly.
7709 SDValue Extract = N->getOperand(i);
7710 SDValue ExtVal = Extract.getOperand(1);
7711 if (Extract.getOperand(0) == VecIn1) {
7712 unsigned ExtIndex = cast<ConstantSDNode>(ExtVal)->getZExtValue();
7713 if (ExtIndex > VT.getVectorNumElements())
7716 Mask.push_back(ExtIndex);
7720 // Otherwise, use InIdx + VecSize
7721 unsigned Idx = cast<ConstantSDNode>(ExtVal)->getZExtValue();
7722 Mask.push_back(Idx+NumInScalars);
7725 // We can't generate a shuffle node with mismatched input and output types.
7726 // Attempt to transform a single input vector to the correct type.
7727 if ((VT != VecIn1.getValueType())) {
7728 // We don't support shuffeling between TWO values of different types.
7729 if (VecIn2.getNode() != 0)
7732 // We only support widening of vectors which are half the size of the
7733 // output registers. For example XMM->YMM widening on X86 with AVX.
7734 if (VecIn1.getValueType().getSizeInBits()*2 != VT.getSizeInBits())
7737 // Widen the input vector by adding undef values.
7738 VecIn1 = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
7739 VecIn1, DAG.getUNDEF(VecIn1.getValueType()));
7742 // If VecIn2 is unused then change it to undef.
7743 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
7745 // Check that we were able to transform all incoming values to the same type.
7746 if (VecIn2.getValueType() != VecIn1.getValueType() ||
7747 VecIn1.getValueType() != VT)
7750 // Only type-legal BUILD_VECTOR nodes are converted to shuffle nodes.
7751 if (!isTypeLegal(VT))
7754 // Return the new VECTOR_SHUFFLE node.
7758 return DAG.getVectorShuffle(VT, N->getDebugLoc(), Ops[0], Ops[1], &Mask[0]);
7764 SDValue DAGCombiner::visitCONCAT_VECTORS(SDNode *N) {
7765 // TODO: Check to see if this is a CONCAT_VECTORS of a bunch of
7766 // EXTRACT_SUBVECTOR operations. If so, and if the EXTRACT_SUBVECTOR vector
7767 // inputs come from at most two distinct vectors, turn this into a shuffle
7770 // If we only have one input vector, we don't need to do any concatenation.
7771 if (N->getNumOperands() == 1)
7772 return N->getOperand(0);
7777 SDValue DAGCombiner::visitEXTRACT_SUBVECTOR(SDNode* N) {
7778 EVT NVT = N->getValueType(0);
7779 SDValue V = N->getOperand(0);
7781 if (V->getOpcode() == ISD::INSERT_SUBVECTOR) {
7782 // Handle only simple case where vector being inserted and vector
7783 // being extracted are of same type, and are half size of larger vectors.
7784 EVT BigVT = V->getOperand(0).getValueType();
7785 EVT SmallVT = V->getOperand(1).getValueType();
7786 if (NVT != SmallVT || NVT.getSizeInBits()*2 != BigVT.getSizeInBits())
7789 // Only handle cases where both indexes are constants with the same type.
7790 ConstantSDNode *InsIdx = dyn_cast<ConstantSDNode>(N->getOperand(1));
7791 ConstantSDNode *ExtIdx = dyn_cast<ConstantSDNode>(V->getOperand(2));
7793 if (InsIdx && ExtIdx &&
7794 InsIdx->getValueType(0).getSizeInBits() <= 64 &&
7795 ExtIdx->getValueType(0).getSizeInBits() <= 64) {
7797 // (extract_subvec (insert_subvec V1, V2, InsIdx), ExtIdx)
7799 // indices are equal => V1
7800 // otherwise => (extract_subvec V1, ExtIdx)
7801 if (InsIdx->getZExtValue() == ExtIdx->getZExtValue())
7802 return V->getOperand(1);
7803 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, N->getDebugLoc(), NVT,
7804 V->getOperand(0), N->getOperand(1));
7811 SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
7812 EVT VT = N->getValueType(0);
7813 unsigned NumElts = VT.getVectorNumElements();
7815 SDValue N0 = N->getOperand(0);
7816 SDValue N1 = N->getOperand(1);
7818 assert(N0.getValueType() == VT && "Vector shuffle must be normalized in DAG");
7820 // Canonicalize shuffle undef, undef -> undef
7821 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
7822 return DAG.getUNDEF(VT);
7824 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
7826 // Canonicalize shuffle v, v -> v, undef
7828 SmallVector<int, 8> NewMask;
7829 for (unsigned i = 0; i != NumElts; ++i) {
7830 int Idx = SVN->getMaskElt(i);
7831 if (Idx >= (int)NumElts) Idx -= NumElts;
7832 NewMask.push_back(Idx);
7834 return DAG.getVectorShuffle(VT, N->getDebugLoc(), N0, DAG.getUNDEF(VT),
7838 // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask.
7839 if (N0.getOpcode() == ISD::UNDEF) {
7840 SmallVector<int, 8> NewMask;
7841 for (unsigned i = 0; i != NumElts; ++i) {
7842 int Idx = SVN->getMaskElt(i);
7844 if (Idx < (int)NumElts)
7849 NewMask.push_back(Idx);
7851 return DAG.getVectorShuffle(VT, N->getDebugLoc(), N1, DAG.getUNDEF(VT),
7855 // Remove references to rhs if it is undef
7856 if (N1.getOpcode() == ISD::UNDEF) {
7857 bool Changed = false;
7858 SmallVector<int, 8> NewMask;
7859 for (unsigned i = 0; i != NumElts; ++i) {
7860 int Idx = SVN->getMaskElt(i);
7861 if (Idx >= (int)NumElts) {
7865 NewMask.push_back(Idx);
7868 return DAG.getVectorShuffle(VT, N->getDebugLoc(), N0, N1, &NewMask[0]);
7871 // If it is a splat, check if the argument vector is another splat or a
7872 // build_vector with all scalar elements the same.
7873 if (SVN->isSplat() && SVN->getSplatIndex() < (int)NumElts) {
7874 SDNode *V = N0.getNode();
7876 // If this is a bit convert that changes the element type of the vector but
7877 // not the number of vector elements, look through it. Be careful not to
7878 // look though conversions that change things like v4f32 to v2f64.
7879 if (V->getOpcode() == ISD::BITCAST) {
7880 SDValue ConvInput = V->getOperand(0);
7881 if (ConvInput.getValueType().isVector() &&
7882 ConvInput.getValueType().getVectorNumElements() == NumElts)
7883 V = ConvInput.getNode();
7886 if (V->getOpcode() == ISD::BUILD_VECTOR) {
7887 assert(V->getNumOperands() == NumElts &&
7888 "BUILD_VECTOR has wrong number of operands");
7890 bool AllSame = true;
7891 for (unsigned i = 0; i != NumElts; ++i) {
7892 if (V->getOperand(i).getOpcode() != ISD::UNDEF) {
7893 Base = V->getOperand(i);
7897 // Splat of <u, u, u, u>, return <u, u, u, u>
7898 if (!Base.getNode())
7900 for (unsigned i = 0; i != NumElts; ++i) {
7901 if (V->getOperand(i) != Base) {
7906 // Splat of <x, x, x, x>, return <x, x, x, x>
7912 // If this shuffle node is simply a swizzle of another shuffle node,
7913 // and it reverses the swizzle of the previous shuffle then we can
7914 // optimize shuffle(shuffle(x, undef), undef) -> x.
7915 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
7916 N1.getOpcode() == ISD::UNDEF) {
7918 ShuffleVectorSDNode *OtherSV = cast<ShuffleVectorSDNode>(N0);
7920 // Shuffle nodes can only reverse shuffles with a single non-undef value.
7921 if (N0.getOperand(1).getOpcode() != ISD::UNDEF)
7924 // The incoming shuffle must be of the same type as the result of the
7926 assert(OtherSV->getOperand(0).getValueType() == VT &&
7927 "Shuffle types don't match");
7929 for (unsigned i = 0; i != NumElts; ++i) {
7930 int Idx = SVN->getMaskElt(i);
7931 assert(Idx < (int)NumElts && "Index references undef operand");
7932 // Next, this index comes from the first value, which is the incoming
7933 // shuffle. Adopt the incoming index.
7935 Idx = OtherSV->getMaskElt(Idx);
7937 // The combined shuffle must map each index to itself.
7938 if (Idx >= 0 && (unsigned)Idx != i)
7942 return OtherSV->getOperand(0);
7948 SDValue DAGCombiner::visitMEMBARRIER(SDNode* N) {
7949 if (!TLI.getShouldFoldAtomicFences())
7952 SDValue atomic = N->getOperand(0);
7953 switch (atomic.getOpcode()) {
7954 case ISD::ATOMIC_CMP_SWAP:
7955 case ISD::ATOMIC_SWAP:
7956 case ISD::ATOMIC_LOAD_ADD:
7957 case ISD::ATOMIC_LOAD_SUB:
7958 case ISD::ATOMIC_LOAD_AND:
7959 case ISD::ATOMIC_LOAD_OR:
7960 case ISD::ATOMIC_LOAD_XOR:
7961 case ISD::ATOMIC_LOAD_NAND:
7962 case ISD::ATOMIC_LOAD_MIN:
7963 case ISD::ATOMIC_LOAD_MAX:
7964 case ISD::ATOMIC_LOAD_UMIN:
7965 case ISD::ATOMIC_LOAD_UMAX:
7971 SDValue fence = atomic.getOperand(0);
7972 if (fence.getOpcode() != ISD::MEMBARRIER)
7975 switch (atomic.getOpcode()) {
7976 case ISD::ATOMIC_CMP_SWAP:
7977 return SDValue(DAG.UpdateNodeOperands(atomic.getNode(),
7978 fence.getOperand(0),
7979 atomic.getOperand(1), atomic.getOperand(2),
7980 atomic.getOperand(3)), atomic.getResNo());
7981 case ISD::ATOMIC_SWAP:
7982 case ISD::ATOMIC_LOAD_ADD:
7983 case ISD::ATOMIC_LOAD_SUB:
7984 case ISD::ATOMIC_LOAD_AND:
7985 case ISD::ATOMIC_LOAD_OR:
7986 case ISD::ATOMIC_LOAD_XOR:
7987 case ISD::ATOMIC_LOAD_NAND:
7988 case ISD::ATOMIC_LOAD_MIN:
7989 case ISD::ATOMIC_LOAD_MAX:
7990 case ISD::ATOMIC_LOAD_UMIN:
7991 case ISD::ATOMIC_LOAD_UMAX:
7992 return SDValue(DAG.UpdateNodeOperands(atomic.getNode(),
7993 fence.getOperand(0),
7994 atomic.getOperand(1), atomic.getOperand(2)),
8001 /// XformToShuffleWithZero - Returns a vector_shuffle if it able to transform
8002 /// an AND to a vector_shuffle with the destination vector and a zero vector.
8003 /// e.g. AND V, <0xffffffff, 0, 0xffffffff, 0>. ==>
8004 /// vector_shuffle V, Zero, <0, 4, 2, 4>
8005 SDValue DAGCombiner::XformToShuffleWithZero(SDNode *N) {
8006 EVT VT = N->getValueType(0);
8007 DebugLoc dl = N->getDebugLoc();
8008 SDValue LHS = N->getOperand(0);
8009 SDValue RHS = N->getOperand(1);
8010 if (N->getOpcode() == ISD::AND) {
8011 if (RHS.getOpcode() == ISD::BITCAST)
8012 RHS = RHS.getOperand(0);
8013 if (RHS.getOpcode() == ISD::BUILD_VECTOR) {
8014 SmallVector<int, 8> Indices;
8015 unsigned NumElts = RHS.getNumOperands();
8016 for (unsigned i = 0; i != NumElts; ++i) {
8017 SDValue Elt = RHS.getOperand(i);
8018 if (!isa<ConstantSDNode>(Elt))
8021 if (cast<ConstantSDNode>(Elt)->isAllOnesValue())
8022 Indices.push_back(i);
8023 else if (cast<ConstantSDNode>(Elt)->isNullValue())
8024 Indices.push_back(NumElts);
8029 // Let's see if the target supports this vector_shuffle.
8030 EVT RVT = RHS.getValueType();
8031 if (!TLI.isVectorClearMaskLegal(Indices, RVT))
8034 // Return the new VECTOR_SHUFFLE node.
8035 EVT EltVT = RVT.getVectorElementType();
8036 SmallVector<SDValue,8> ZeroOps(RVT.getVectorNumElements(),
8037 DAG.getConstant(0, EltVT));
8038 SDValue Zero = DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
8039 RVT, &ZeroOps[0], ZeroOps.size());
8040 LHS = DAG.getNode(ISD::BITCAST, dl, RVT, LHS);
8041 SDValue Shuf = DAG.getVectorShuffle(RVT, dl, LHS, Zero, &Indices[0]);
8042 return DAG.getNode(ISD::BITCAST, dl, VT, Shuf);
8049 /// SimplifyVBinOp - Visit a binary vector operation, like ADD.
8050 SDValue DAGCombiner::SimplifyVBinOp(SDNode *N) {
8051 // After legalize, the target may be depending on adds and other
8052 // binary ops to provide legal ways to construct constants or other
8053 // things. Simplifying them may result in a loss of legality.
8054 if (LegalOperations) return SDValue();
8056 assert(N->getValueType(0).isVector() &&
8057 "SimplifyVBinOp only works on vectors!");
8059 SDValue LHS = N->getOperand(0);
8060 SDValue RHS = N->getOperand(1);
8061 SDValue Shuffle = XformToShuffleWithZero(N);
8062 if (Shuffle.getNode()) return Shuffle;
8064 // If the LHS and RHS are BUILD_VECTOR nodes, see if we can constant fold
8066 if (LHS.getOpcode() == ISD::BUILD_VECTOR &&
8067 RHS.getOpcode() == ISD::BUILD_VECTOR) {
8068 SmallVector<SDValue, 8> Ops;
8069 for (unsigned i = 0, e = LHS.getNumOperands(); i != e; ++i) {
8070 SDValue LHSOp = LHS.getOperand(i);
8071 SDValue RHSOp = RHS.getOperand(i);
8072 // If these two elements can't be folded, bail out.
8073 if ((LHSOp.getOpcode() != ISD::UNDEF &&
8074 LHSOp.getOpcode() != ISD::Constant &&
8075 LHSOp.getOpcode() != ISD::ConstantFP) ||
8076 (RHSOp.getOpcode() != ISD::UNDEF &&
8077 RHSOp.getOpcode() != ISD::Constant &&
8078 RHSOp.getOpcode() != ISD::ConstantFP))
8081 // Can't fold divide by zero.
8082 if (N->getOpcode() == ISD::SDIV || N->getOpcode() == ISD::UDIV ||
8083 N->getOpcode() == ISD::FDIV) {
8084 if ((RHSOp.getOpcode() == ISD::Constant &&
8085 cast<ConstantSDNode>(RHSOp.getNode())->isNullValue()) ||
8086 (RHSOp.getOpcode() == ISD::ConstantFP &&
8087 cast<ConstantFPSDNode>(RHSOp.getNode())->getValueAPF().isZero()))
8091 EVT VT = LHSOp.getValueType();
8092 EVT RVT = RHSOp.getValueType();
8094 // Integer BUILD_VECTOR operands may have types larger than the element
8095 // size (e.g., when the element type is not legal). Prior to type
8096 // legalization, the types may not match between the two BUILD_VECTORS.
8097 // Truncate one of the operands to make them match.
8098 if (RVT.getSizeInBits() > VT.getSizeInBits()) {
8099 RHSOp = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, RHSOp);
8101 LHSOp = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), RVT, LHSOp);
8105 SDValue FoldOp = DAG.getNode(N->getOpcode(), LHS.getDebugLoc(), VT,
8107 if (FoldOp.getOpcode() != ISD::UNDEF &&
8108 FoldOp.getOpcode() != ISD::Constant &&
8109 FoldOp.getOpcode() != ISD::ConstantFP)
8111 Ops.push_back(FoldOp);
8112 AddToWorkList(FoldOp.getNode());
8115 if (Ops.size() == LHS.getNumOperands())
8116 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
8117 LHS.getValueType(), &Ops[0], Ops.size());
8123 SDValue DAGCombiner::SimplifySelect(DebugLoc DL, SDValue N0,
8124 SDValue N1, SDValue N2){
8125 assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!");
8127 SDValue SCC = SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1), N1, N2,
8128 cast<CondCodeSDNode>(N0.getOperand(2))->get());
8130 // If we got a simplified select_cc node back from SimplifySelectCC, then
8131 // break it down into a new SETCC node, and a new SELECT node, and then return
8132 // the SELECT node, since we were called with a SELECT node.
8133 if (SCC.getNode()) {
8134 // Check to see if we got a select_cc back (to turn into setcc/select).
8135 // Otherwise, just return whatever node we got back, like fabs.
8136 if (SCC.getOpcode() == ISD::SELECT_CC) {
8137 SDValue SETCC = DAG.getNode(ISD::SETCC, N0.getDebugLoc(),
8139 SCC.getOperand(0), SCC.getOperand(1),
8141 AddToWorkList(SETCC.getNode());
8142 return DAG.getNode(ISD::SELECT, SCC.getDebugLoc(), SCC.getValueType(),
8143 SCC.getOperand(2), SCC.getOperand(3), SETCC);
8151 /// SimplifySelectOps - Given a SELECT or a SELECT_CC node, where LHS and RHS
8152 /// are the two values being selected between, see if we can simplify the
8153 /// select. Callers of this should assume that TheSelect is deleted if this
8154 /// returns true. As such, they should return the appropriate thing (e.g. the
8155 /// node) back to the top-level of the DAG combiner loop to avoid it being
8157 bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDValue LHS,
8160 // Cannot simplify select with vector condition
8161 if (TheSelect->getOperand(0).getValueType().isVector()) return false;
8163 // If this is a select from two identical things, try to pull the operation
8164 // through the select.
8165 if (LHS.getOpcode() != RHS.getOpcode() ||
8166 !LHS.hasOneUse() || !RHS.hasOneUse())
8169 // If this is a load and the token chain is identical, replace the select
8170 // of two loads with a load through a select of the address to load from.
8171 // This triggers in things like "select bool X, 10.0, 123.0" after the FP
8172 // constants have been dropped into the constant pool.
8173 if (LHS.getOpcode() == ISD::LOAD) {
8174 LoadSDNode *LLD = cast<LoadSDNode>(LHS);
8175 LoadSDNode *RLD = cast<LoadSDNode>(RHS);
8177 // Token chains must be identical.
8178 if (LHS.getOperand(0) != RHS.getOperand(0) ||
8179 // Do not let this transformation reduce the number of volatile loads.
8180 LLD->isVolatile() || RLD->isVolatile() ||
8181 // If this is an EXTLOAD, the VT's must match.
8182 LLD->getMemoryVT() != RLD->getMemoryVT() ||
8183 // If this is an EXTLOAD, the kind of extension must match.
8184 (LLD->getExtensionType() != RLD->getExtensionType() &&
8185 // The only exception is if one of the extensions is anyext.
8186 LLD->getExtensionType() != ISD::EXTLOAD &&
8187 RLD->getExtensionType() != ISD::EXTLOAD) ||
8188 // FIXME: this discards src value information. This is
8189 // over-conservative. It would be beneficial to be able to remember
8190 // both potential memory locations. Since we are discarding
8191 // src value info, don't do the transformation if the memory
8192 // locations are not in the default address space.
8193 LLD->getPointerInfo().getAddrSpace() != 0 ||
8194 RLD->getPointerInfo().getAddrSpace() != 0)
8197 // Check that the select condition doesn't reach either load. If so,
8198 // folding this will induce a cycle into the DAG. If not, this is safe to
8199 // xform, so create a select of the addresses.
8201 if (TheSelect->getOpcode() == ISD::SELECT) {
8202 SDNode *CondNode = TheSelect->getOperand(0).getNode();
8203 if ((LLD->hasAnyUseOfValue(1) && LLD->isPredecessorOf(CondNode)) ||
8204 (RLD->hasAnyUseOfValue(1) && RLD->isPredecessorOf(CondNode)))
8206 Addr = DAG.getNode(ISD::SELECT, TheSelect->getDebugLoc(),
8207 LLD->getBasePtr().getValueType(),
8208 TheSelect->getOperand(0), LLD->getBasePtr(),
8210 } else { // Otherwise SELECT_CC
8211 SDNode *CondLHS = TheSelect->getOperand(0).getNode();
8212 SDNode *CondRHS = TheSelect->getOperand(1).getNode();
8214 if ((LLD->hasAnyUseOfValue(1) &&
8215 (LLD->isPredecessorOf(CondLHS) || LLD->isPredecessorOf(CondRHS))) ||
8216 (RLD->hasAnyUseOfValue(1) &&
8217 (RLD->isPredecessorOf(CondLHS) || RLD->isPredecessorOf(CondRHS))))
8220 Addr = DAG.getNode(ISD::SELECT_CC, TheSelect->getDebugLoc(),
8221 LLD->getBasePtr().getValueType(),
8222 TheSelect->getOperand(0),
8223 TheSelect->getOperand(1),
8224 LLD->getBasePtr(), RLD->getBasePtr(),
8225 TheSelect->getOperand(4));
8229 if (LLD->getExtensionType() == ISD::NON_EXTLOAD) {
8230 Load = DAG.getLoad(TheSelect->getValueType(0),
8231 TheSelect->getDebugLoc(),
8232 // FIXME: Discards pointer info.
8233 LLD->getChain(), Addr, MachinePointerInfo(),
8234 LLD->isVolatile(), LLD->isNonTemporal(),
8235 LLD->isInvariant(), LLD->getAlignment());
8237 Load = DAG.getExtLoad(LLD->getExtensionType() == ISD::EXTLOAD ?
8238 RLD->getExtensionType() : LLD->getExtensionType(),
8239 TheSelect->getDebugLoc(),
8240 TheSelect->getValueType(0),
8241 // FIXME: Discards pointer info.
8242 LLD->getChain(), Addr, MachinePointerInfo(),
8243 LLD->getMemoryVT(), LLD->isVolatile(),
8244 LLD->isNonTemporal(), LLD->getAlignment());
8247 // Users of the select now use the result of the load.
8248 CombineTo(TheSelect, Load);
8250 // Users of the old loads now use the new load's chain. We know the
8251 // old-load value is dead now.
8252 CombineTo(LHS.getNode(), Load.getValue(0), Load.getValue(1));
8253 CombineTo(RHS.getNode(), Load.getValue(0), Load.getValue(1));
8260 /// SimplifySelectCC - Simplify an expression of the form (N0 cond N1) ? N2 : N3
8261 /// where 'cond' is the comparison specified by CC.
8262 SDValue DAGCombiner::SimplifySelectCC(DebugLoc DL, SDValue N0, SDValue N1,
8263 SDValue N2, SDValue N3,
8264 ISD::CondCode CC, bool NotExtCompare) {
8265 // (x ? y : y) -> y.
8266 if (N2 == N3) return N2;
8268 EVT VT = N2.getValueType();
8269 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
8270 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
8271 ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N3.getNode());
8273 // Determine if the condition we're dealing with is constant
8274 SDValue SCC = SimplifySetCC(TLI.getSetCCResultType(N0.getValueType()),
8275 N0, N1, CC, DL, false);
8276 if (SCC.getNode()) AddToWorkList(SCC.getNode());
8277 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode());
8279 // fold select_cc true, x, y -> x
8280 if (SCCC && !SCCC->isNullValue())
8282 // fold select_cc false, x, y -> y
8283 if (SCCC && SCCC->isNullValue())
8286 // Check to see if we can simplify the select into an fabs node
8287 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) {
8288 // Allow either -0.0 or 0.0
8289 if (CFP->getValueAPF().isZero()) {
8290 // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs
8291 if ((CC == ISD::SETGE || CC == ISD::SETGT) &&
8292 N0 == N2 && N3.getOpcode() == ISD::FNEG &&
8293 N2 == N3.getOperand(0))
8294 return DAG.getNode(ISD::FABS, DL, VT, N0);
8296 // select (setl[te] X, +/-0.0), fneg(X), X -> fabs
8297 if ((CC == ISD::SETLT || CC == ISD::SETLE) &&
8298 N0 == N3 && N2.getOpcode() == ISD::FNEG &&
8299 N2.getOperand(0) == N3)
8300 return DAG.getNode(ISD::FABS, DL, VT, N3);
8304 // Turn "(a cond b) ? 1.0f : 2.0f" into "load (tmp + ((a cond b) ? 0 : 4)"
8305 // where "tmp" is a constant pool entry containing an array with 1.0 and 2.0
8306 // in it. This is a win when the constant is not otherwise available because
8307 // it replaces two constant pool loads with one. We only do this if the FP
8308 // type is known to be legal, because if it isn't, then we are before legalize
8309 // types an we want the other legalization to happen first (e.g. to avoid
8310 // messing with soft float) and if the ConstantFP is not legal, because if
8311 // it is legal, we may not need to store the FP constant in a constant pool.
8312 if (ConstantFPSDNode *TV = dyn_cast<ConstantFPSDNode>(N2))
8313 if (ConstantFPSDNode *FV = dyn_cast<ConstantFPSDNode>(N3)) {
8314 if (TLI.isTypeLegal(N2.getValueType()) &&
8315 (TLI.getOperationAction(ISD::ConstantFP, N2.getValueType()) !=
8316 TargetLowering::Legal) &&
8317 // If both constants have multiple uses, then we won't need to do an
8318 // extra load, they are likely around in registers for other users.
8319 (TV->hasOneUse() || FV->hasOneUse())) {
8320 Constant *Elts[] = {
8321 const_cast<ConstantFP*>(FV->getConstantFPValue()),
8322 const_cast<ConstantFP*>(TV->getConstantFPValue())
8324 Type *FPTy = Elts[0]->getType();
8325 const TargetData &TD = *TLI.getTargetData();
8327 // Create a ConstantArray of the two constants.
8328 Constant *CA = ConstantArray::get(ArrayType::get(FPTy, 2), Elts);
8329 SDValue CPIdx = DAG.getConstantPool(CA, TLI.getPointerTy(),
8330 TD.getPrefTypeAlignment(FPTy));
8331 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
8333 // Get the offsets to the 0 and 1 element of the array so that we can
8334 // select between them.
8335 SDValue Zero = DAG.getIntPtrConstant(0);
8336 unsigned EltSize = (unsigned)TD.getTypeAllocSize(Elts[0]->getType());
8337 SDValue One = DAG.getIntPtrConstant(EltSize);
8339 SDValue Cond = DAG.getSetCC(DL,
8340 TLI.getSetCCResultType(N0.getValueType()),
8342 AddToWorkList(Cond.getNode());
8343 SDValue CstOffset = DAG.getNode(ISD::SELECT, DL, Zero.getValueType(),
8345 AddToWorkList(CstOffset.getNode());
8346 CPIdx = DAG.getNode(ISD::ADD, DL, TLI.getPointerTy(), CPIdx,
8348 AddToWorkList(CPIdx.getNode());
8349 return DAG.getLoad(TV->getValueType(0), DL, DAG.getEntryNode(), CPIdx,
8350 MachinePointerInfo::getConstantPool(), false,
8351 false, false, Alignment);
8356 // Check to see if we can perform the "gzip trick", transforming
8357 // (select_cc setlt X, 0, A, 0) -> (and (sra X, (sub size(X), 1), A)
8358 if (N1C && N3C && N3C->isNullValue() && CC == ISD::SETLT &&
8359 (N1C->isNullValue() || // (a < 0) ? b : 0
8360 (N1C->getAPIntValue() == 1 && N0 == N2))) { // (a < 1) ? a : 0
8361 EVT XType = N0.getValueType();
8362 EVT AType = N2.getValueType();
8363 if (XType.bitsGE(AType)) {
8364 // and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a
8365 // single-bit constant.
8366 if (N2C && ((N2C->getAPIntValue() & (N2C->getAPIntValue()-1)) == 0)) {
8367 unsigned ShCtV = N2C->getAPIntValue().logBase2();
8368 ShCtV = XType.getSizeInBits()-ShCtV-1;
8369 SDValue ShCt = DAG.getConstant(ShCtV,
8370 getShiftAmountTy(N0.getValueType()));
8371 SDValue Shift = DAG.getNode(ISD::SRL, N0.getDebugLoc(),
8373 AddToWorkList(Shift.getNode());
8375 if (XType.bitsGT(AType)) {
8376 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
8377 AddToWorkList(Shift.getNode());
8380 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
8383 SDValue Shift = DAG.getNode(ISD::SRA, N0.getDebugLoc(),
8385 DAG.getConstant(XType.getSizeInBits()-1,
8386 getShiftAmountTy(N0.getValueType())));
8387 AddToWorkList(Shift.getNode());
8389 if (XType.bitsGT(AType)) {
8390 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
8391 AddToWorkList(Shift.getNode());
8394 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
8398 // fold (select_cc seteq (and x, y), 0, 0, A) -> (and (shr (shl x)) A)
8399 // where y is has a single bit set.
8400 // A plaintext description would be, we can turn the SELECT_CC into an AND
8401 // when the condition can be materialized as an all-ones register. Any
8402 // single bit-test can be materialized as an all-ones register with
8403 // shift-left and shift-right-arith.
8404 if (CC == ISD::SETEQ && N0->getOpcode() == ISD::AND &&
8405 N0->getValueType(0) == VT &&
8406 N1C && N1C->isNullValue() &&
8407 N2C && N2C->isNullValue()) {
8408 SDValue AndLHS = N0->getOperand(0);
8409 ConstantSDNode *ConstAndRHS = dyn_cast<ConstantSDNode>(N0->getOperand(1));
8410 if (ConstAndRHS && ConstAndRHS->getAPIntValue().countPopulation() == 1) {
8411 // Shift the tested bit over the sign bit.
8412 APInt AndMask = ConstAndRHS->getAPIntValue();
8414 DAG.getConstant(AndMask.countLeadingZeros(),
8415 getShiftAmountTy(AndLHS.getValueType()));
8416 SDValue Shl = DAG.getNode(ISD::SHL, N0.getDebugLoc(), VT, AndLHS, ShlAmt);
8418 // Now arithmetic right shift it all the way over, so the result is either
8419 // all-ones, or zero.
8421 DAG.getConstant(AndMask.getBitWidth()-1,
8422 getShiftAmountTy(Shl.getValueType()));
8423 SDValue Shr = DAG.getNode(ISD::SRA, N0.getDebugLoc(), VT, Shl, ShrAmt);
8425 return DAG.getNode(ISD::AND, DL, VT, Shr, N3);
8429 // fold select C, 16, 0 -> shl C, 4
8430 if (N2C && N3C && N3C->isNullValue() && N2C->getAPIntValue().isPowerOf2() &&
8431 TLI.getBooleanContents(N0.getValueType().isVector()) ==
8432 TargetLowering::ZeroOrOneBooleanContent) {
8434 // If the caller doesn't want us to simplify this into a zext of a compare,
8436 if (NotExtCompare && N2C->getAPIntValue() == 1)
8439 // Get a SetCC of the condition
8440 // FIXME: Should probably make sure that setcc is legal if we ever have a
8441 // target where it isn't.
8443 // cast from setcc result type to select result type
8445 SCC = DAG.getSetCC(DL, TLI.getSetCCResultType(N0.getValueType()),
8447 if (N2.getValueType().bitsLT(SCC.getValueType()))
8448 Temp = DAG.getZeroExtendInReg(SCC, N2.getDebugLoc(), N2.getValueType());
8450 Temp = DAG.getNode(ISD::ZERO_EXTEND, N2.getDebugLoc(),
8451 N2.getValueType(), SCC);
8453 SCC = DAG.getSetCC(N0.getDebugLoc(), MVT::i1, N0, N1, CC);
8454 Temp = DAG.getNode(ISD::ZERO_EXTEND, N2.getDebugLoc(),
8455 N2.getValueType(), SCC);
8458 AddToWorkList(SCC.getNode());
8459 AddToWorkList(Temp.getNode());
8461 if (N2C->getAPIntValue() == 1)
8464 // shl setcc result by log2 n2c
8465 return DAG.getNode(ISD::SHL, DL, N2.getValueType(), Temp,
8466 DAG.getConstant(N2C->getAPIntValue().logBase2(),
8467 getShiftAmountTy(Temp.getValueType())));
8470 // Check to see if this is the equivalent of setcc
8471 // FIXME: Turn all of these into setcc if setcc if setcc is legal
8472 // otherwise, go ahead with the folds.
8473 if (0 && N3C && N3C->isNullValue() && N2C && (N2C->getAPIntValue() == 1ULL)) {
8474 EVT XType = N0.getValueType();
8475 if (!LegalOperations ||
8476 TLI.isOperationLegal(ISD::SETCC, TLI.getSetCCResultType(XType))) {
8477 SDValue Res = DAG.getSetCC(DL, TLI.getSetCCResultType(XType), N0, N1, CC);
8478 if (Res.getValueType() != VT)
8479 Res = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Res);
8483 // fold (seteq X, 0) -> (srl (ctlz X, log2(size(X))))
8484 if (N1C && N1C->isNullValue() && CC == ISD::SETEQ &&
8485 (!LegalOperations ||
8486 TLI.isOperationLegal(ISD::CTLZ, XType))) {
8487 SDValue Ctlz = DAG.getNode(ISD::CTLZ, N0.getDebugLoc(), XType, N0);
8488 return DAG.getNode(ISD::SRL, DL, XType, Ctlz,
8489 DAG.getConstant(Log2_32(XType.getSizeInBits()),
8490 getShiftAmountTy(Ctlz.getValueType())));
8492 // fold (setgt X, 0) -> (srl (and (-X, ~X), size(X)-1))
8493 if (N1C && N1C->isNullValue() && CC == ISD::SETGT) {
8494 SDValue NegN0 = DAG.getNode(ISD::SUB, N0.getDebugLoc(),
8495 XType, DAG.getConstant(0, XType), N0);
8496 SDValue NotN0 = DAG.getNOT(N0.getDebugLoc(), N0, XType);
8497 return DAG.getNode(ISD::SRL, DL, XType,
8498 DAG.getNode(ISD::AND, DL, XType, NegN0, NotN0),
8499 DAG.getConstant(XType.getSizeInBits()-1,
8500 getShiftAmountTy(XType)));
8502 // fold (setgt X, -1) -> (xor (srl (X, size(X)-1), 1))
8503 if (N1C && N1C->isAllOnesValue() && CC == ISD::SETGT) {
8504 SDValue Sign = DAG.getNode(ISD::SRL, N0.getDebugLoc(), XType, N0,
8505 DAG.getConstant(XType.getSizeInBits()-1,
8506 getShiftAmountTy(N0.getValueType())));
8507 return DAG.getNode(ISD::XOR, DL, XType, Sign, DAG.getConstant(1, XType));
8511 // Check to see if this is an integer abs.
8512 // select_cc setg[te] X, 0, X, -X ->
8513 // select_cc setgt X, -1, X, -X ->
8514 // select_cc setl[te] X, 0, -X, X ->
8515 // select_cc setlt X, 1, -X, X ->
8516 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
8518 ConstantSDNode *SubC = NULL;
8519 if (((N1C->isNullValue() && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
8520 (N1C->isAllOnesValue() && CC == ISD::SETGT)) &&
8521 N0 == N2 && N3.getOpcode() == ISD::SUB && N0 == N3.getOperand(1))
8522 SubC = dyn_cast<ConstantSDNode>(N3.getOperand(0));
8523 else if (((N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE)) ||
8524 (N1C->isOne() && CC == ISD::SETLT)) &&
8525 N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1))
8526 SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0));
8528 EVT XType = N0.getValueType();
8529 if (SubC && SubC->isNullValue() && XType.isInteger()) {
8530 SDValue Shift = DAG.getNode(ISD::SRA, N0.getDebugLoc(), XType,
8532 DAG.getConstant(XType.getSizeInBits()-1,
8533 getShiftAmountTy(N0.getValueType())));
8534 SDValue Add = DAG.getNode(ISD::ADD, N0.getDebugLoc(),
8536 AddToWorkList(Shift.getNode());
8537 AddToWorkList(Add.getNode());
8538 return DAG.getNode(ISD::XOR, DL, XType, Add, Shift);
8545 /// SimplifySetCC - This is a stub for TargetLowering::SimplifySetCC.
8546 SDValue DAGCombiner::SimplifySetCC(EVT VT, SDValue N0,
8547 SDValue N1, ISD::CondCode Cond,
8548 DebugLoc DL, bool foldBooleans) {
8549 TargetLowering::DAGCombinerInfo
8550 DagCombineInfo(DAG, !LegalTypes, !LegalOperations, false, this);
8551 return TLI.SimplifySetCC(VT, N0, N1, Cond, foldBooleans, DagCombineInfo, DL);
8554 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
8555 /// return a DAG expression to select that will generate the same value by
8556 /// multiplying by a magic number. See:
8557 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
8558 SDValue DAGCombiner::BuildSDIV(SDNode *N) {
8559 std::vector<SDNode*> Built;
8560 SDValue S = TLI.BuildSDIV(N, DAG, LegalOperations, &Built);
8562 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end();
8568 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
8569 /// return a DAG expression to select that will generate the same value by
8570 /// multiplying by a magic number. See:
8571 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
8572 SDValue DAGCombiner::BuildUDIV(SDNode *N) {
8573 std::vector<SDNode*> Built;
8574 SDValue S = TLI.BuildUDIV(N, DAG, LegalOperations, &Built);
8576 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end();
8582 /// FindBaseOffset - Return true if base is a frame index, which is known not
8583 // to alias with anything but itself. Provides base object and offset as
8585 static bool FindBaseOffset(SDValue Ptr, SDValue &Base, int64_t &Offset,
8586 const GlobalValue *&GV, void *&CV) {
8587 // Assume it is a primitive operation.
8588 Base = Ptr; Offset = 0; GV = 0; CV = 0;
8590 // If it's an adding a simple constant then integrate the offset.
8591 if (Base.getOpcode() == ISD::ADD) {
8592 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Base.getOperand(1))) {
8593 Base = Base.getOperand(0);
8594 Offset += C->getZExtValue();
8598 // Return the underlying GlobalValue, and update the Offset. Return false
8599 // for GlobalAddressSDNode since the same GlobalAddress may be represented
8600 // by multiple nodes with different offsets.
8601 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Base)) {
8602 GV = G->getGlobal();
8603 Offset += G->getOffset();
8607 // Return the underlying Constant value, and update the Offset. Return false
8608 // for ConstantSDNodes since the same constant pool entry may be represented
8609 // by multiple nodes with different offsets.
8610 if (ConstantPoolSDNode *C = dyn_cast<ConstantPoolSDNode>(Base)) {
8611 CV = C->isMachineConstantPoolEntry() ? (void *)C->getMachineCPVal()
8612 : (void *)C->getConstVal();
8613 Offset += C->getOffset();
8616 // If it's any of the following then it can't alias with anything but itself.
8617 return isa<FrameIndexSDNode>(Base);
8620 /// isAlias - Return true if there is any possibility that the two addresses
8622 bool DAGCombiner::isAlias(SDValue Ptr1, int64_t Size1,
8623 const Value *SrcValue1, int SrcValueOffset1,
8624 unsigned SrcValueAlign1,
8625 const MDNode *TBAAInfo1,
8626 SDValue Ptr2, int64_t Size2,
8627 const Value *SrcValue2, int SrcValueOffset2,
8628 unsigned SrcValueAlign2,
8629 const MDNode *TBAAInfo2) const {
8630 // If they are the same then they must be aliases.
8631 if (Ptr1 == Ptr2) return true;
8633 // Gather base node and offset information.
8634 SDValue Base1, Base2;
8635 int64_t Offset1, Offset2;
8636 const GlobalValue *GV1, *GV2;
8638 bool isFrameIndex1 = FindBaseOffset(Ptr1, Base1, Offset1, GV1, CV1);
8639 bool isFrameIndex2 = FindBaseOffset(Ptr2, Base2, Offset2, GV2, CV2);
8641 // If they have a same base address then check to see if they overlap.
8642 if (Base1 == Base2 || (GV1 && (GV1 == GV2)) || (CV1 && (CV1 == CV2)))
8643 return !((Offset1 + Size1) <= Offset2 || (Offset2 + Size2) <= Offset1);
8645 // It is possible for different frame indices to alias each other, mostly
8646 // when tail call optimization reuses return address slots for arguments.
8647 // To catch this case, look up the actual index of frame indices to compute
8648 // the real alias relationship.
8649 if (isFrameIndex1 && isFrameIndex2) {
8650 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8651 Offset1 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base1)->getIndex());
8652 Offset2 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base2)->getIndex());
8653 return !((Offset1 + Size1) <= Offset2 || (Offset2 + Size2) <= Offset1);
8656 // Otherwise, if we know what the bases are, and they aren't identical, then
8657 // we know they cannot alias.
8658 if ((isFrameIndex1 || CV1 || GV1) && (isFrameIndex2 || CV2 || GV2))
8661 // If we know required SrcValue1 and SrcValue2 have relatively large alignment
8662 // compared to the size and offset of the access, we may be able to prove they
8663 // do not alias. This check is conservative for now to catch cases created by
8664 // splitting vector types.
8665 if ((SrcValueAlign1 == SrcValueAlign2) &&
8666 (SrcValueOffset1 != SrcValueOffset2) &&
8667 (Size1 == Size2) && (SrcValueAlign1 > Size1)) {
8668 int64_t OffAlign1 = SrcValueOffset1 % SrcValueAlign1;
8669 int64_t OffAlign2 = SrcValueOffset2 % SrcValueAlign1;
8671 // There is no overlap between these relatively aligned accesses of similar
8672 // size, return no alias.
8673 if ((OffAlign1 + Size1) <= OffAlign2 || (OffAlign2 + Size2) <= OffAlign1)
8677 if (CombinerGlobalAA) {
8678 // Use alias analysis information.
8679 int64_t MinOffset = std::min(SrcValueOffset1, SrcValueOffset2);
8680 int64_t Overlap1 = Size1 + SrcValueOffset1 - MinOffset;
8681 int64_t Overlap2 = Size2 + SrcValueOffset2 - MinOffset;
8682 AliasAnalysis::AliasResult AAResult =
8683 AA.alias(AliasAnalysis::Location(SrcValue1, Overlap1, TBAAInfo1),
8684 AliasAnalysis::Location(SrcValue2, Overlap2, TBAAInfo2));
8685 if (AAResult == AliasAnalysis::NoAlias)
8689 // Otherwise we have to assume they alias.
8693 /// FindAliasInfo - Extracts the relevant alias information from the memory
8694 /// node. Returns true if the operand was a load.
8695 bool DAGCombiner::FindAliasInfo(SDNode *N,
8696 SDValue &Ptr, int64_t &Size,
8697 const Value *&SrcValue,
8698 int &SrcValueOffset,
8699 unsigned &SrcValueAlign,
8700 const MDNode *&TBAAInfo) const {
8701 LSBaseSDNode *LS = cast<LSBaseSDNode>(N);
8703 Ptr = LS->getBasePtr();
8704 Size = LS->getMemoryVT().getSizeInBits() >> 3;
8705 SrcValue = LS->getSrcValue();
8706 SrcValueOffset = LS->getSrcValueOffset();
8707 SrcValueAlign = LS->getOriginalAlignment();
8708 TBAAInfo = LS->getTBAAInfo();
8709 return isa<LoadSDNode>(LS);
8712 /// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes,
8713 /// looking for aliasing nodes and adding them to the Aliases vector.
8714 void DAGCombiner::GatherAllAliases(SDNode *N, SDValue OriginalChain,
8715 SmallVector<SDValue, 8> &Aliases) {
8716 SmallVector<SDValue, 8> Chains; // List of chains to visit.
8717 SmallPtrSet<SDNode *, 16> Visited; // Visited node set.
8719 // Get alias information for node.
8722 const Value *SrcValue;
8724 unsigned SrcValueAlign;
8725 const MDNode *SrcTBAAInfo;
8726 bool IsLoad = FindAliasInfo(N, Ptr, Size, SrcValue, SrcValueOffset,
8727 SrcValueAlign, SrcTBAAInfo);
8730 Chains.push_back(OriginalChain);
8733 // Look at each chain and determine if it is an alias. If so, add it to the
8734 // aliases list. If not, then continue up the chain looking for the next
8736 while (!Chains.empty()) {
8737 SDValue Chain = Chains.back();
8740 // For TokenFactor nodes, look at each operand and only continue up the
8741 // chain until we find two aliases. If we've seen two aliases, assume we'll
8742 // find more and revert to original chain since the xform is unlikely to be
8745 // FIXME: The depth check could be made to return the last non-aliasing
8746 // chain we found before we hit a tokenfactor rather than the original
8748 if (Depth > 6 || Aliases.size() == 2) {
8750 Aliases.push_back(OriginalChain);
8754 // Don't bother if we've been before.
8755 if (!Visited.insert(Chain.getNode()))
8758 switch (Chain.getOpcode()) {
8759 case ISD::EntryToken:
8760 // Entry token is ideal chain operand, but handled in FindBetterChain.
8765 // Get alias information for Chain.
8768 const Value *OpSrcValue;
8769 int OpSrcValueOffset;
8770 unsigned OpSrcValueAlign;
8771 const MDNode *OpSrcTBAAInfo;
8772 bool IsOpLoad = FindAliasInfo(Chain.getNode(), OpPtr, OpSize,
8773 OpSrcValue, OpSrcValueOffset,
8777 // If chain is alias then stop here.
8778 if (!(IsLoad && IsOpLoad) &&
8779 isAlias(Ptr, Size, SrcValue, SrcValueOffset, SrcValueAlign,
8781 OpPtr, OpSize, OpSrcValue, OpSrcValueOffset,
8782 OpSrcValueAlign, OpSrcTBAAInfo)) {
8783 Aliases.push_back(Chain);
8785 // Look further up the chain.
8786 Chains.push_back(Chain.getOperand(0));
8792 case ISD::TokenFactor:
8793 // We have to check each of the operands of the token factor for "small"
8794 // token factors, so we queue them up. Adding the operands to the queue
8795 // (stack) in reverse order maintains the original order and increases the
8796 // likelihood that getNode will find a matching token factor (CSE.)
8797 if (Chain.getNumOperands() > 16) {
8798 Aliases.push_back(Chain);
8801 for (unsigned n = Chain.getNumOperands(); n;)
8802 Chains.push_back(Chain.getOperand(--n));
8807 // For all other instructions we will just have to take what we can get.
8808 Aliases.push_back(Chain);
8814 /// FindBetterChain - Walk up chain skipping non-aliasing memory nodes, looking
8815 /// for a better chain (aliasing node.)
8816 SDValue DAGCombiner::FindBetterChain(SDNode *N, SDValue OldChain) {
8817 SmallVector<SDValue, 8> Aliases; // Ops for replacing token factor.
8819 // Accumulate all the aliases to this node.
8820 GatherAllAliases(N, OldChain, Aliases);
8822 // If no operands then chain to entry token.
8823 if (Aliases.size() == 0)
8824 return DAG.getEntryNode();
8826 // If a single operand then chain to it. We don't need to revisit it.
8827 if (Aliases.size() == 1)
8830 // Construct a custom tailored token factor.
8831 return DAG.getNode(ISD::TokenFactor, N->getDebugLoc(), MVT::Other,
8832 &Aliases[0], Aliases.size());
8835 // SelectionDAG::Combine - This is the entry point for the file.
8837 void SelectionDAG::Combine(CombineLevel Level, AliasAnalysis &AA,
8838 CodeGenOpt::Level OptLevel) {
8839 /// run - This is the main entry point to this class.
8841 DAGCombiner(*this, AA, OptLevel).Run(Level);