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 SimplifyVUnaryOp(SDNode *N);
198 SDValue visitSHL(SDNode *N);
199 SDValue visitSRA(SDNode *N);
200 SDValue visitSRL(SDNode *N);
201 SDValue visitCTLZ(SDNode *N);
202 SDValue visitCTLZ_ZERO_UNDEF(SDNode *N);
203 SDValue visitCTTZ(SDNode *N);
204 SDValue visitCTTZ_ZERO_UNDEF(SDNode *N);
205 SDValue visitCTPOP(SDNode *N);
206 SDValue visitSELECT(SDNode *N);
207 SDValue visitSELECT_CC(SDNode *N);
208 SDValue visitSETCC(SDNode *N);
209 SDValue visitSIGN_EXTEND(SDNode *N);
210 SDValue visitZERO_EXTEND(SDNode *N);
211 SDValue visitANY_EXTEND(SDNode *N);
212 SDValue visitSIGN_EXTEND_INREG(SDNode *N);
213 SDValue visitTRUNCATE(SDNode *N);
214 SDValue visitBITCAST(SDNode *N);
215 SDValue visitBUILD_PAIR(SDNode *N);
216 SDValue visitFADD(SDNode *N);
217 SDValue visitFSUB(SDNode *N);
218 SDValue visitFMUL(SDNode *N);
219 SDValue visitFMA(SDNode *N);
220 SDValue visitFDIV(SDNode *N);
221 SDValue visitFREM(SDNode *N);
222 SDValue visitFCOPYSIGN(SDNode *N);
223 SDValue visitSINT_TO_FP(SDNode *N);
224 SDValue visitUINT_TO_FP(SDNode *N);
225 SDValue visitFP_TO_SINT(SDNode *N);
226 SDValue visitFP_TO_UINT(SDNode *N);
227 SDValue visitFP_ROUND(SDNode *N);
228 SDValue visitFP_ROUND_INREG(SDNode *N);
229 SDValue visitFP_EXTEND(SDNode *N);
230 SDValue visitFNEG(SDNode *N);
231 SDValue visitFABS(SDNode *N);
232 SDValue visitFCEIL(SDNode *N);
233 SDValue visitFTRUNC(SDNode *N);
234 SDValue visitFFLOOR(SDNode *N);
235 SDValue visitBRCOND(SDNode *N);
236 SDValue visitBR_CC(SDNode *N);
237 SDValue visitLOAD(SDNode *N);
238 SDValue visitSTORE(SDNode *N);
239 SDValue visitINSERT_VECTOR_ELT(SDNode *N);
240 SDValue visitEXTRACT_VECTOR_ELT(SDNode *N);
241 SDValue visitBUILD_VECTOR(SDNode *N);
242 SDValue visitCONCAT_VECTORS(SDNode *N);
243 SDValue visitEXTRACT_SUBVECTOR(SDNode *N);
244 SDValue visitVECTOR_SHUFFLE(SDNode *N);
245 SDValue visitMEMBARRIER(SDNode *N);
247 SDValue XformToShuffleWithZero(SDNode *N);
248 SDValue ReassociateOps(unsigned Opc, DebugLoc DL, SDValue LHS, SDValue RHS);
250 SDValue visitShiftByConstant(SDNode *N, unsigned Amt);
252 bool SimplifySelectOps(SDNode *SELECT, SDValue LHS, SDValue RHS);
253 SDValue SimplifyBinOpWithSameOpcodeHands(SDNode *N);
254 SDValue SimplifySelect(DebugLoc DL, SDValue N0, SDValue N1, SDValue N2);
255 SDValue SimplifySelectCC(DebugLoc DL, SDValue N0, SDValue N1, SDValue N2,
256 SDValue N3, ISD::CondCode CC,
257 bool NotExtCompare = false);
258 SDValue SimplifySetCC(EVT VT, SDValue N0, SDValue N1, ISD::CondCode Cond,
259 DebugLoc DL, bool foldBooleans = true);
260 SDValue SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
262 SDValue CombineConsecutiveLoads(SDNode *N, EVT VT);
263 SDValue ConstantFoldBITCASTofBUILD_VECTOR(SDNode *, EVT);
264 SDValue BuildSDIV(SDNode *N);
265 SDValue BuildUDIV(SDNode *N);
266 SDValue MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
267 bool DemandHighBits = true);
268 SDValue MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1);
269 SDNode *MatchRotate(SDValue LHS, SDValue RHS, DebugLoc DL);
270 SDValue ReduceLoadWidth(SDNode *N);
271 SDValue ReduceLoadOpStoreWidth(SDNode *N);
272 SDValue TransformFPLoadStorePair(SDNode *N);
274 SDValue GetDemandedBits(SDValue V, const APInt &Mask);
276 /// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes,
277 /// looking for aliasing nodes and adding them to the Aliases vector.
278 void GatherAllAliases(SDNode *N, SDValue OriginalChain,
279 SmallVector<SDValue, 8> &Aliases);
281 /// isAlias - Return true if there is any possibility that the two addresses
283 bool isAlias(SDValue Ptr1, int64_t Size1,
284 const Value *SrcValue1, int SrcValueOffset1,
285 unsigned SrcValueAlign1,
286 const MDNode *TBAAInfo1,
287 SDValue Ptr2, int64_t Size2,
288 const Value *SrcValue2, int SrcValueOffset2,
289 unsigned SrcValueAlign2,
290 const MDNode *TBAAInfo2) const;
292 /// FindAliasInfo - Extracts the relevant alias information from the memory
293 /// node. Returns true if the operand was a load.
294 bool FindAliasInfo(SDNode *N,
295 SDValue &Ptr, int64_t &Size,
296 const Value *&SrcValue, int &SrcValueOffset,
297 unsigned &SrcValueAlignment,
298 const MDNode *&TBAAInfo) const;
300 /// FindBetterChain - Walk up chain skipping non-aliasing memory nodes,
301 /// looking for a better chain (aliasing node.)
302 SDValue FindBetterChain(SDNode *N, SDValue Chain);
304 /// Merge consecutive store operations into a wide store.
305 /// \return True if some memory operations were changed.
306 bool MergeConsecutiveStores(StoreSDNode *N);
309 DAGCombiner(SelectionDAG &D, AliasAnalysis &A, CodeGenOpt::Level OL)
310 : DAG(D), TLI(D.getTargetLoweringInfo()), Level(BeforeLegalizeTypes),
311 OptLevel(OL), LegalOperations(false), LegalTypes(false), AA(A) {}
313 /// Run - runs the dag combiner on all nodes in the work list
314 void Run(CombineLevel AtLevel);
316 SelectionDAG &getDAG() const { return DAG; }
318 /// getShiftAmountTy - Returns a type large enough to hold any valid
319 /// shift amount - before type legalization these can be huge.
320 EVT getShiftAmountTy(EVT LHSTy) {
321 return LegalTypes ? TLI.getShiftAmountTy(LHSTy) : TLI.getPointerTy();
324 /// isTypeLegal - This method returns true if we are running before type
325 /// legalization or if the specified VT is legal.
326 bool isTypeLegal(const EVT &VT) {
327 if (!LegalTypes) return true;
328 return TLI.isTypeLegal(VT);
335 /// WorkListRemover - This class is a DAGUpdateListener that removes any deleted
336 /// nodes from the worklist.
337 class WorkListRemover : public SelectionDAG::DAGUpdateListener {
340 explicit WorkListRemover(DAGCombiner &dc)
341 : SelectionDAG::DAGUpdateListener(dc.getDAG()), DC(dc) {}
343 virtual void NodeDeleted(SDNode *N, SDNode *E) {
344 DC.removeFromWorkList(N);
349 //===----------------------------------------------------------------------===//
350 // TargetLowering::DAGCombinerInfo implementation
351 //===----------------------------------------------------------------------===//
353 void TargetLowering::DAGCombinerInfo::AddToWorklist(SDNode *N) {
354 ((DAGCombiner*)DC)->AddToWorkList(N);
357 void TargetLowering::DAGCombinerInfo::RemoveFromWorklist(SDNode *N) {
358 ((DAGCombiner*)DC)->removeFromWorkList(N);
361 SDValue TargetLowering::DAGCombinerInfo::
362 CombineTo(SDNode *N, const std::vector<SDValue> &To, bool AddTo) {
363 return ((DAGCombiner*)DC)->CombineTo(N, &To[0], To.size(), AddTo);
366 SDValue TargetLowering::DAGCombinerInfo::
367 CombineTo(SDNode *N, SDValue Res, bool AddTo) {
368 return ((DAGCombiner*)DC)->CombineTo(N, Res, AddTo);
372 SDValue TargetLowering::DAGCombinerInfo::
373 CombineTo(SDNode *N, SDValue Res0, SDValue Res1, bool AddTo) {
374 return ((DAGCombiner*)DC)->CombineTo(N, Res0, Res1, AddTo);
377 void TargetLowering::DAGCombinerInfo::
378 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
379 return ((DAGCombiner*)DC)->CommitTargetLoweringOpt(TLO);
382 //===----------------------------------------------------------------------===//
384 //===----------------------------------------------------------------------===//
386 /// isNegatibleForFree - Return 1 if we can compute the negated form of the
387 /// specified expression for the same cost as the expression itself, or 2 if we
388 /// can compute the negated form more cheaply than the expression itself.
389 static char isNegatibleForFree(SDValue Op, bool LegalOperations,
390 const TargetLowering &TLI,
391 const TargetOptions *Options,
392 unsigned Depth = 0) {
393 // No compile time optimizations on this type.
394 if (Op.getValueType() == MVT::ppcf128)
397 // fneg is removable even if it has multiple uses.
398 if (Op.getOpcode() == ISD::FNEG) return 2;
400 // Don't allow anything with multiple uses.
401 if (!Op.hasOneUse()) return 0;
403 // Don't recurse exponentially.
404 if (Depth > 6) return 0;
406 switch (Op.getOpcode()) {
407 default: return false;
408 case ISD::ConstantFP:
409 // Don't invert constant FP values after legalize. The negated constant
410 // isn't necessarily legal.
411 return LegalOperations ? 0 : 1;
413 // FIXME: determine better conditions for this xform.
414 if (!Options->UnsafeFPMath) return 0;
416 // After operation legalization, it might not be legal to create new FSUBs.
417 if (LegalOperations &&
418 !TLI.isOperationLegalOrCustom(ISD::FSUB, Op.getValueType()))
421 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B)
422 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI,
425 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
426 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options,
429 // We can't turn -(A-B) into B-A when we honor signed zeros.
430 if (!Options->UnsafeFPMath) return 0;
432 // fold (fneg (fsub A, B)) -> (fsub B, A)
437 if (Options->HonorSignDependentRoundingFPMath()) return 0;
439 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y) or (fmul X, (fneg Y))
440 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI,
444 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options,
450 return isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI, Options,
455 /// GetNegatedExpression - If isNegatibleForFree returns true, this function
456 /// returns the newly negated expression.
457 static SDValue GetNegatedExpression(SDValue Op, SelectionDAG &DAG,
458 bool LegalOperations, unsigned Depth = 0) {
459 // fneg is removable even if it has multiple uses.
460 if (Op.getOpcode() == ISD::FNEG) return Op.getOperand(0);
462 // Don't allow anything with multiple uses.
463 assert(Op.hasOneUse() && "Unknown reuse!");
465 assert(Depth <= 6 && "GetNegatedExpression doesn't match isNegatibleForFree");
466 switch (Op.getOpcode()) {
467 default: llvm_unreachable("Unknown code");
468 case ISD::ConstantFP: {
469 APFloat V = cast<ConstantFPSDNode>(Op)->getValueAPF();
471 return DAG.getConstantFP(V, Op.getValueType());
474 // FIXME: determine better conditions for this xform.
475 assert(DAG.getTarget().Options.UnsafeFPMath);
477 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B)
478 if (isNegatibleForFree(Op.getOperand(0), LegalOperations,
479 DAG.getTargetLoweringInfo(),
480 &DAG.getTarget().Options, Depth+1))
481 return DAG.getNode(ISD::FSUB, Op.getDebugLoc(), Op.getValueType(),
482 GetNegatedExpression(Op.getOperand(0), DAG,
483 LegalOperations, Depth+1),
485 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
486 return DAG.getNode(ISD::FSUB, Op.getDebugLoc(), Op.getValueType(),
487 GetNegatedExpression(Op.getOperand(1), DAG,
488 LegalOperations, Depth+1),
491 // We can't turn -(A-B) into B-A when we honor signed zeros.
492 assert(DAG.getTarget().Options.UnsafeFPMath);
494 // fold (fneg (fsub 0, B)) -> B
495 if (ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(Op.getOperand(0)))
496 if (N0CFP->getValueAPF().isZero())
497 return Op.getOperand(1);
499 // fold (fneg (fsub A, B)) -> (fsub B, A)
500 return DAG.getNode(ISD::FSUB, Op.getDebugLoc(), Op.getValueType(),
501 Op.getOperand(1), Op.getOperand(0));
505 assert(!DAG.getTarget().Options.HonorSignDependentRoundingFPMath());
507 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y)
508 if (isNegatibleForFree(Op.getOperand(0), LegalOperations,
509 DAG.getTargetLoweringInfo(),
510 &DAG.getTarget().Options, Depth+1))
511 return DAG.getNode(Op.getOpcode(), Op.getDebugLoc(), Op.getValueType(),
512 GetNegatedExpression(Op.getOperand(0), DAG,
513 LegalOperations, Depth+1),
516 // fold (fneg (fmul X, Y)) -> (fmul X, (fneg Y))
517 return DAG.getNode(Op.getOpcode(), Op.getDebugLoc(), Op.getValueType(),
519 GetNegatedExpression(Op.getOperand(1), DAG,
520 LegalOperations, Depth+1));
524 return DAG.getNode(Op.getOpcode(), Op.getDebugLoc(), Op.getValueType(),
525 GetNegatedExpression(Op.getOperand(0), DAG,
526 LegalOperations, Depth+1));
528 return DAG.getNode(ISD::FP_ROUND, Op.getDebugLoc(), Op.getValueType(),
529 GetNegatedExpression(Op.getOperand(0), DAG,
530 LegalOperations, Depth+1),
536 // isSetCCEquivalent - Return true if this node is a setcc, or is a select_cc
537 // that selects between the values 1 and 0, making it equivalent to a setcc.
538 // Also, set the incoming LHS, RHS, and CC references to the appropriate
539 // nodes based on the type of node we are checking. This simplifies life a
540 // bit for the callers.
541 static bool isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS,
543 if (N.getOpcode() == ISD::SETCC) {
544 LHS = N.getOperand(0);
545 RHS = N.getOperand(1);
546 CC = N.getOperand(2);
549 if (N.getOpcode() == ISD::SELECT_CC &&
550 N.getOperand(2).getOpcode() == ISD::Constant &&
551 N.getOperand(3).getOpcode() == ISD::Constant &&
552 cast<ConstantSDNode>(N.getOperand(2))->getAPIntValue() == 1 &&
553 cast<ConstantSDNode>(N.getOperand(3))->isNullValue()) {
554 LHS = N.getOperand(0);
555 RHS = N.getOperand(1);
556 CC = N.getOperand(4);
562 // isOneUseSetCC - Return true if this is a SetCC-equivalent operation with only
563 // one use. If this is true, it allows the users to invert the operation for
564 // free when it is profitable to do so.
565 static bool isOneUseSetCC(SDValue N) {
567 if (isSetCCEquivalent(N, N0, N1, N2) && N.getNode()->hasOneUse())
572 SDValue DAGCombiner::ReassociateOps(unsigned Opc, DebugLoc DL,
573 SDValue N0, SDValue N1) {
574 EVT VT = N0.getValueType();
575 if (N0.getOpcode() == Opc && isa<ConstantSDNode>(N0.getOperand(1))) {
576 if (isa<ConstantSDNode>(N1)) {
577 // reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2))
579 DAG.FoldConstantArithmetic(Opc, VT,
580 cast<ConstantSDNode>(N0.getOperand(1)),
581 cast<ConstantSDNode>(N1));
582 return DAG.getNode(Opc, DL, VT, N0.getOperand(0), OpNode);
584 if (N0.hasOneUse()) {
585 // reassoc. (op (op x, c1), y) -> (op (op x, y), c1) iff x+c1 has one use
586 SDValue OpNode = DAG.getNode(Opc, N0.getDebugLoc(), VT,
587 N0.getOperand(0), N1);
588 AddToWorkList(OpNode.getNode());
589 return DAG.getNode(Opc, DL, VT, OpNode, N0.getOperand(1));
593 if (N1.getOpcode() == Opc && isa<ConstantSDNode>(N1.getOperand(1))) {
594 if (isa<ConstantSDNode>(N0)) {
595 // reassoc. (op c2, (op x, c1)) -> (op x, (op c1, c2))
597 DAG.FoldConstantArithmetic(Opc, VT,
598 cast<ConstantSDNode>(N1.getOperand(1)),
599 cast<ConstantSDNode>(N0));
600 return DAG.getNode(Opc, DL, VT, N1.getOperand(0), OpNode);
602 if (N1.hasOneUse()) {
603 // reassoc. (op y, (op x, c1)) -> (op (op x, y), c1) iff x+c1 has one use
604 SDValue OpNode = DAG.getNode(Opc, N0.getDebugLoc(), VT,
605 N1.getOperand(0), N0);
606 AddToWorkList(OpNode.getNode());
607 return DAG.getNode(Opc, DL, VT, OpNode, N1.getOperand(1));
614 SDValue DAGCombiner::CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
616 assert(N->getNumValues() == NumTo && "Broken CombineTo call!");
618 DEBUG(dbgs() << "\nReplacing.1 ";
620 dbgs() << "\nWith: ";
621 To[0].getNode()->dump(&DAG);
622 dbgs() << " and " << NumTo-1 << " other values\n";
623 for (unsigned i = 0, e = NumTo; i != e; ++i)
624 assert((!To[i].getNode() ||
625 N->getValueType(i) == To[i].getValueType()) &&
626 "Cannot combine value to value of different type!"));
627 WorkListRemover DeadNodes(*this);
628 DAG.ReplaceAllUsesWith(N, To);
630 // Push the new nodes and any users onto the worklist
631 for (unsigned i = 0, e = NumTo; i != e; ++i) {
632 if (To[i].getNode()) {
633 AddToWorkList(To[i].getNode());
634 AddUsersToWorkList(To[i].getNode());
639 // Finally, if the node is now dead, remove it from the graph. The node
640 // may not be dead if the replacement process recursively simplified to
641 // something else needing this node.
642 if (N->use_empty()) {
643 // Nodes can be reintroduced into the worklist. Make sure we do not
644 // process a node that has been replaced.
645 removeFromWorkList(N);
647 // Finally, since the node is now dead, remove it from the graph.
650 return SDValue(N, 0);
654 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
655 // Replace all uses. If any nodes become isomorphic to other nodes and
656 // are deleted, make sure to remove them from our worklist.
657 WorkListRemover DeadNodes(*this);
658 DAG.ReplaceAllUsesOfValueWith(TLO.Old, TLO.New);
660 // Push the new node and any (possibly new) users onto the worklist.
661 AddToWorkList(TLO.New.getNode());
662 AddUsersToWorkList(TLO.New.getNode());
664 // Finally, if the node is now dead, remove it from the graph. The node
665 // may not be dead if the replacement process recursively simplified to
666 // something else needing this node.
667 if (TLO.Old.getNode()->use_empty()) {
668 removeFromWorkList(TLO.Old.getNode());
670 // If the operands of this node are only used by the node, they will now
671 // be dead. Make sure to visit them first to delete dead nodes early.
672 for (unsigned i = 0, e = TLO.Old.getNode()->getNumOperands(); i != e; ++i)
673 if (TLO.Old.getNode()->getOperand(i).getNode()->hasOneUse())
674 AddToWorkList(TLO.Old.getNode()->getOperand(i).getNode());
676 DAG.DeleteNode(TLO.Old.getNode());
680 /// SimplifyDemandedBits - Check the specified integer node value to see if
681 /// it can be simplified or if things it uses can be simplified by bit
682 /// propagation. If so, return true.
683 bool DAGCombiner::SimplifyDemandedBits(SDValue Op, const APInt &Demanded) {
684 TargetLowering::TargetLoweringOpt TLO(DAG, LegalTypes, LegalOperations);
685 APInt KnownZero, KnownOne;
686 if (!TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
690 AddToWorkList(Op.getNode());
692 // Replace the old value with the new one.
694 DEBUG(dbgs() << "\nReplacing.2 ";
695 TLO.Old.getNode()->dump(&DAG);
696 dbgs() << "\nWith: ";
697 TLO.New.getNode()->dump(&DAG);
700 CommitTargetLoweringOpt(TLO);
704 void DAGCombiner::ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad) {
705 DebugLoc dl = Load->getDebugLoc();
706 EVT VT = Load->getValueType(0);
707 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, VT, SDValue(ExtLoad, 0));
709 DEBUG(dbgs() << "\nReplacing.9 ";
711 dbgs() << "\nWith: ";
712 Trunc.getNode()->dump(&DAG);
714 WorkListRemover DeadNodes(*this);
715 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 0), Trunc);
716 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), SDValue(ExtLoad, 1));
717 removeFromWorkList(Load);
718 DAG.DeleteNode(Load);
719 AddToWorkList(Trunc.getNode());
722 SDValue DAGCombiner::PromoteOperand(SDValue Op, EVT PVT, bool &Replace) {
724 DebugLoc dl = Op.getDebugLoc();
725 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Op)) {
726 EVT MemVT = LD->getMemoryVT();
727 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
728 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT) ? ISD::ZEXTLOAD
730 : LD->getExtensionType();
732 return DAG.getExtLoad(ExtType, dl, PVT,
733 LD->getChain(), LD->getBasePtr(),
734 LD->getPointerInfo(),
735 MemVT, LD->isVolatile(),
736 LD->isNonTemporal(), LD->getAlignment());
739 unsigned Opc = Op.getOpcode();
742 case ISD::AssertSext:
743 return DAG.getNode(ISD::AssertSext, dl, PVT,
744 SExtPromoteOperand(Op.getOperand(0), PVT),
746 case ISD::AssertZext:
747 return DAG.getNode(ISD::AssertZext, dl, PVT,
748 ZExtPromoteOperand(Op.getOperand(0), PVT),
750 case ISD::Constant: {
752 Op.getValueType().isByteSized() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
753 return DAG.getNode(ExtOpc, dl, PVT, Op);
757 if (!TLI.isOperationLegal(ISD::ANY_EXTEND, PVT))
759 return DAG.getNode(ISD::ANY_EXTEND, dl, PVT, Op);
762 SDValue DAGCombiner::SExtPromoteOperand(SDValue Op, EVT PVT) {
763 if (!TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, PVT))
765 EVT OldVT = Op.getValueType();
766 DebugLoc dl = Op.getDebugLoc();
767 bool Replace = false;
768 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
769 if (NewOp.getNode() == 0)
771 AddToWorkList(NewOp.getNode());
774 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
775 return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, NewOp.getValueType(), NewOp,
776 DAG.getValueType(OldVT));
779 SDValue DAGCombiner::ZExtPromoteOperand(SDValue Op, EVT PVT) {
780 EVT OldVT = Op.getValueType();
781 DebugLoc dl = Op.getDebugLoc();
782 bool Replace = false;
783 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
784 if (NewOp.getNode() == 0)
786 AddToWorkList(NewOp.getNode());
789 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
790 return DAG.getZeroExtendInReg(NewOp, dl, OldVT);
793 /// PromoteIntBinOp - Promote the specified integer binary operation if the
794 /// target indicates it is beneficial. e.g. On x86, it's usually better to
795 /// promote i16 operations to i32 since i16 instructions are longer.
796 SDValue DAGCombiner::PromoteIntBinOp(SDValue Op) {
797 if (!LegalOperations)
800 EVT VT = Op.getValueType();
801 if (VT.isVector() || !VT.isInteger())
804 // If operation type is 'undesirable', e.g. i16 on x86, consider
806 unsigned Opc = Op.getOpcode();
807 if (TLI.isTypeDesirableForOp(Opc, VT))
811 // Consult target whether it is a good idea to promote this operation and
812 // what's the right type to promote it to.
813 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
814 assert(PVT != VT && "Don't know what type to promote to!");
816 bool Replace0 = false;
817 SDValue N0 = Op.getOperand(0);
818 SDValue NN0 = PromoteOperand(N0, PVT, Replace0);
819 if (NN0.getNode() == 0)
822 bool Replace1 = false;
823 SDValue N1 = Op.getOperand(1);
828 NN1 = PromoteOperand(N1, PVT, Replace1);
829 if (NN1.getNode() == 0)
833 AddToWorkList(NN0.getNode());
835 AddToWorkList(NN1.getNode());
838 ReplaceLoadWithPromotedLoad(N0.getNode(), NN0.getNode());
840 ReplaceLoadWithPromotedLoad(N1.getNode(), NN1.getNode());
842 DEBUG(dbgs() << "\nPromoting ";
843 Op.getNode()->dump(&DAG));
844 DebugLoc dl = Op.getDebugLoc();
845 return DAG.getNode(ISD::TRUNCATE, dl, VT,
846 DAG.getNode(Opc, dl, PVT, NN0, NN1));
851 /// PromoteIntShiftOp - Promote the specified integer shift operation if the
852 /// target indicates it is beneficial. e.g. On x86, it's usually better to
853 /// promote i16 operations to i32 since i16 instructions are longer.
854 SDValue DAGCombiner::PromoteIntShiftOp(SDValue Op) {
855 if (!LegalOperations)
858 EVT VT = Op.getValueType();
859 if (VT.isVector() || !VT.isInteger())
862 // If operation type is 'undesirable', e.g. i16 on x86, consider
864 unsigned Opc = Op.getOpcode();
865 if (TLI.isTypeDesirableForOp(Opc, VT))
869 // Consult target whether it is a good idea to promote this operation and
870 // what's the right type to promote it to.
871 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
872 assert(PVT != VT && "Don't know what type to promote to!");
874 bool Replace = false;
875 SDValue N0 = Op.getOperand(0);
877 N0 = SExtPromoteOperand(Op.getOperand(0), PVT);
878 else if (Opc == ISD::SRL)
879 N0 = ZExtPromoteOperand(Op.getOperand(0), PVT);
881 N0 = PromoteOperand(N0, PVT, Replace);
882 if (N0.getNode() == 0)
885 AddToWorkList(N0.getNode());
887 ReplaceLoadWithPromotedLoad(Op.getOperand(0).getNode(), N0.getNode());
889 DEBUG(dbgs() << "\nPromoting ";
890 Op.getNode()->dump(&DAG));
891 DebugLoc dl = Op.getDebugLoc();
892 return DAG.getNode(ISD::TRUNCATE, dl, VT,
893 DAG.getNode(Opc, dl, PVT, N0, Op.getOperand(1)));
898 SDValue DAGCombiner::PromoteExtend(SDValue Op) {
899 if (!LegalOperations)
902 EVT VT = Op.getValueType();
903 if (VT.isVector() || !VT.isInteger())
906 // If operation type is 'undesirable', e.g. i16 on x86, consider
908 unsigned Opc = Op.getOpcode();
909 if (TLI.isTypeDesirableForOp(Opc, VT))
913 // Consult target whether it is a good idea to promote this operation and
914 // what's the right type to promote it to.
915 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
916 assert(PVT != VT && "Don't know what type to promote to!");
917 // fold (aext (aext x)) -> (aext x)
918 // fold (aext (zext x)) -> (zext x)
919 // fold (aext (sext x)) -> (sext x)
920 DEBUG(dbgs() << "\nPromoting ";
921 Op.getNode()->dump(&DAG));
922 return DAG.getNode(Op.getOpcode(), Op.getDebugLoc(), VT, Op.getOperand(0));
927 bool DAGCombiner::PromoteLoad(SDValue Op) {
928 if (!LegalOperations)
931 EVT VT = Op.getValueType();
932 if (VT.isVector() || !VT.isInteger())
935 // If operation type is 'undesirable', e.g. i16 on x86, consider
937 unsigned Opc = Op.getOpcode();
938 if (TLI.isTypeDesirableForOp(Opc, VT))
942 // Consult target whether it is a good idea to promote this operation and
943 // what's the right type to promote it to.
944 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
945 assert(PVT != VT && "Don't know what type to promote to!");
947 DebugLoc dl = Op.getDebugLoc();
948 SDNode *N = Op.getNode();
949 LoadSDNode *LD = cast<LoadSDNode>(N);
950 EVT MemVT = LD->getMemoryVT();
951 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
952 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT) ? ISD::ZEXTLOAD
954 : LD->getExtensionType();
955 SDValue NewLD = DAG.getExtLoad(ExtType, dl, PVT,
956 LD->getChain(), LD->getBasePtr(),
957 LD->getPointerInfo(),
958 MemVT, LD->isVolatile(),
959 LD->isNonTemporal(), LD->getAlignment());
960 SDValue Result = DAG.getNode(ISD::TRUNCATE, dl, VT, NewLD);
962 DEBUG(dbgs() << "\nPromoting ";
965 Result.getNode()->dump(&DAG);
967 WorkListRemover DeadNodes(*this);
968 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result);
969 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), NewLD.getValue(1));
970 removeFromWorkList(N);
972 AddToWorkList(Result.getNode());
979 //===----------------------------------------------------------------------===//
980 // Main DAG Combiner implementation
981 //===----------------------------------------------------------------------===//
983 void DAGCombiner::Run(CombineLevel AtLevel) {
984 // set the instance variables, so that the various visit routines may use it.
986 LegalOperations = Level >= AfterLegalizeVectorOps;
987 LegalTypes = Level >= AfterLegalizeTypes;
989 // Add all the dag nodes to the worklist.
990 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
991 E = DAG.allnodes_end(); I != E; ++I)
994 // Create a dummy node (which is not added to allnodes), that adds a reference
995 // to the root node, preventing it from being deleted, and tracking any
996 // changes of the root.
997 HandleSDNode Dummy(DAG.getRoot());
999 // The root of the dag may dangle to deleted nodes until the dag combiner is
1000 // done. Set it to null to avoid confusion.
1001 DAG.setRoot(SDValue());
1003 // while the worklist isn't empty, find a node and
1004 // try and combine it.
1005 while (!WorkListContents.empty()) {
1007 // The WorkListOrder holds the SDNodes in order, but it may contain duplicates.
1008 // In order to avoid a linear scan, we use a set (O(log N)) to hold what the
1009 // worklist *should* contain, and check the node we want to visit is should
1010 // actually be visited.
1012 N = WorkListOrder.pop_back_val();
1013 } while (!WorkListContents.erase(N));
1015 // If N has no uses, it is dead. Make sure to revisit all N's operands once
1016 // N is deleted from the DAG, since they too may now be dead or may have a
1017 // reduced number of uses, allowing other xforms.
1018 if (N->use_empty() && N != &Dummy) {
1019 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1020 AddToWorkList(N->getOperand(i).getNode());
1026 SDValue RV = combine(N);
1028 if (RV.getNode() == 0)
1033 // If we get back the same node we passed in, rather than a new node or
1034 // zero, we know that the node must have defined multiple values and
1035 // CombineTo was used. Since CombineTo takes care of the worklist
1036 // mechanics for us, we have no work to do in this case.
1037 if (RV.getNode() == N)
1040 assert(N->getOpcode() != ISD::DELETED_NODE &&
1041 RV.getNode()->getOpcode() != ISD::DELETED_NODE &&
1042 "Node was deleted but visit returned new node!");
1044 DEBUG(dbgs() << "\nReplacing.3 ";
1046 dbgs() << "\nWith: ";
1047 RV.getNode()->dump(&DAG);
1050 // Transfer debug value.
1051 DAG.TransferDbgValues(SDValue(N, 0), RV);
1052 WorkListRemover DeadNodes(*this);
1053 if (N->getNumValues() == RV.getNode()->getNumValues())
1054 DAG.ReplaceAllUsesWith(N, RV.getNode());
1056 assert(N->getValueType(0) == RV.getValueType() &&
1057 N->getNumValues() == 1 && "Type mismatch");
1059 DAG.ReplaceAllUsesWith(N, &OpV);
1062 // Push the new node and any users onto the worklist
1063 AddToWorkList(RV.getNode());
1064 AddUsersToWorkList(RV.getNode());
1066 // Add any uses of the old node to the worklist in case this node is the
1067 // last one that uses them. They may become dead after this node is
1069 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1070 AddToWorkList(N->getOperand(i).getNode());
1072 // Finally, if the node is now dead, remove it from the graph. The node
1073 // may not be dead if the replacement process recursively simplified to
1074 // something else needing this node.
1075 if (N->use_empty()) {
1076 // Nodes can be reintroduced into the worklist. Make sure we do not
1077 // process a node that has been replaced.
1078 removeFromWorkList(N);
1080 // Finally, since the node is now dead, remove it from the graph.
1085 // If the root changed (e.g. it was a dead load, update the root).
1086 DAG.setRoot(Dummy.getValue());
1087 DAG.RemoveDeadNodes();
1090 SDValue DAGCombiner::visit(SDNode *N) {
1091 switch (N->getOpcode()) {
1093 case ISD::TokenFactor: return visitTokenFactor(N);
1094 case ISD::MERGE_VALUES: return visitMERGE_VALUES(N);
1095 case ISD::ADD: return visitADD(N);
1096 case ISD::SUB: return visitSUB(N);
1097 case ISD::ADDC: return visitADDC(N);
1098 case ISD::SUBC: return visitSUBC(N);
1099 case ISD::ADDE: return visitADDE(N);
1100 case ISD::SUBE: return visitSUBE(N);
1101 case ISD::MUL: return visitMUL(N);
1102 case ISD::SDIV: return visitSDIV(N);
1103 case ISD::UDIV: return visitUDIV(N);
1104 case ISD::SREM: return visitSREM(N);
1105 case ISD::UREM: return visitUREM(N);
1106 case ISD::MULHU: return visitMULHU(N);
1107 case ISD::MULHS: return visitMULHS(N);
1108 case ISD::SMUL_LOHI: return visitSMUL_LOHI(N);
1109 case ISD::UMUL_LOHI: return visitUMUL_LOHI(N);
1110 case ISD::SMULO: return visitSMULO(N);
1111 case ISD::UMULO: return visitUMULO(N);
1112 case ISD::SDIVREM: return visitSDIVREM(N);
1113 case ISD::UDIVREM: return visitUDIVREM(N);
1114 case ISD::AND: return visitAND(N);
1115 case ISD::OR: return visitOR(N);
1116 case ISD::XOR: return visitXOR(N);
1117 case ISD::SHL: return visitSHL(N);
1118 case ISD::SRA: return visitSRA(N);
1119 case ISD::SRL: return visitSRL(N);
1120 case ISD::CTLZ: return visitCTLZ(N);
1121 case ISD::CTLZ_ZERO_UNDEF: return visitCTLZ_ZERO_UNDEF(N);
1122 case ISD::CTTZ: return visitCTTZ(N);
1123 case ISD::CTTZ_ZERO_UNDEF: return visitCTTZ_ZERO_UNDEF(N);
1124 case ISD::CTPOP: return visitCTPOP(N);
1125 case ISD::SELECT: return visitSELECT(N);
1126 case ISD::SELECT_CC: return visitSELECT_CC(N);
1127 case ISD::SETCC: return visitSETCC(N);
1128 case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N);
1129 case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N);
1130 case ISD::ANY_EXTEND: return visitANY_EXTEND(N);
1131 case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N);
1132 case ISD::TRUNCATE: return visitTRUNCATE(N);
1133 case ISD::BITCAST: return visitBITCAST(N);
1134 case ISD::BUILD_PAIR: return visitBUILD_PAIR(N);
1135 case ISD::FADD: return visitFADD(N);
1136 case ISD::FSUB: return visitFSUB(N);
1137 case ISD::FMUL: return visitFMUL(N);
1138 case ISD::FMA: return visitFMA(N);
1139 case ISD::FDIV: return visitFDIV(N);
1140 case ISD::FREM: return visitFREM(N);
1141 case ISD::FCOPYSIGN: return visitFCOPYSIGN(N);
1142 case ISD::SINT_TO_FP: return visitSINT_TO_FP(N);
1143 case ISD::UINT_TO_FP: return visitUINT_TO_FP(N);
1144 case ISD::FP_TO_SINT: return visitFP_TO_SINT(N);
1145 case ISD::FP_TO_UINT: return visitFP_TO_UINT(N);
1146 case ISD::FP_ROUND: return visitFP_ROUND(N);
1147 case ISD::FP_ROUND_INREG: return visitFP_ROUND_INREG(N);
1148 case ISD::FP_EXTEND: return visitFP_EXTEND(N);
1149 case ISD::FNEG: return visitFNEG(N);
1150 case ISD::FABS: return visitFABS(N);
1151 case ISD::FFLOOR: return visitFFLOOR(N);
1152 case ISD::FCEIL: return visitFCEIL(N);
1153 case ISD::FTRUNC: return visitFTRUNC(N);
1154 case ISD::BRCOND: return visitBRCOND(N);
1155 case ISD::BR_CC: return visitBR_CC(N);
1156 case ISD::LOAD: return visitLOAD(N);
1157 case ISD::STORE: return visitSTORE(N);
1158 case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N);
1159 case ISD::EXTRACT_VECTOR_ELT: return visitEXTRACT_VECTOR_ELT(N);
1160 case ISD::BUILD_VECTOR: return visitBUILD_VECTOR(N);
1161 case ISD::CONCAT_VECTORS: return visitCONCAT_VECTORS(N);
1162 case ISD::EXTRACT_SUBVECTOR: return visitEXTRACT_SUBVECTOR(N);
1163 case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N);
1164 case ISD::MEMBARRIER: return visitMEMBARRIER(N);
1169 SDValue DAGCombiner::combine(SDNode *N) {
1170 SDValue RV = visit(N);
1172 // If nothing happened, try a target-specific DAG combine.
1173 if (RV.getNode() == 0) {
1174 assert(N->getOpcode() != ISD::DELETED_NODE &&
1175 "Node was deleted but visit returned NULL!");
1177 if (N->getOpcode() >= ISD::BUILTIN_OP_END ||
1178 TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode())) {
1180 // Expose the DAG combiner to the target combiner impls.
1181 TargetLowering::DAGCombinerInfo
1182 DagCombineInfo(DAG, !LegalTypes, !LegalOperations, false, this);
1184 RV = TLI.PerformDAGCombine(N, DagCombineInfo);
1188 // If nothing happened still, try promoting the operation.
1189 if (RV.getNode() == 0) {
1190 switch (N->getOpcode()) {
1198 RV = PromoteIntBinOp(SDValue(N, 0));
1203 RV = PromoteIntShiftOp(SDValue(N, 0));
1205 case ISD::SIGN_EXTEND:
1206 case ISD::ZERO_EXTEND:
1207 case ISD::ANY_EXTEND:
1208 RV = PromoteExtend(SDValue(N, 0));
1211 if (PromoteLoad(SDValue(N, 0)))
1217 // If N is a commutative binary node, try commuting it to enable more
1219 if (RV.getNode() == 0 &&
1220 SelectionDAG::isCommutativeBinOp(N->getOpcode()) &&
1221 N->getNumValues() == 1) {
1222 SDValue N0 = N->getOperand(0);
1223 SDValue N1 = N->getOperand(1);
1225 // Constant operands are canonicalized to RHS.
1226 if (isa<ConstantSDNode>(N0) || !isa<ConstantSDNode>(N1)) {
1227 SDValue Ops[] = { N1, N0 };
1228 SDNode *CSENode = DAG.getNodeIfExists(N->getOpcode(), N->getVTList(),
1231 return SDValue(CSENode, 0);
1238 /// getInputChainForNode - Given a node, return its input chain if it has one,
1239 /// otherwise return a null sd operand.
1240 static SDValue getInputChainForNode(SDNode *N) {
1241 if (unsigned NumOps = N->getNumOperands()) {
1242 if (N->getOperand(0).getValueType() == MVT::Other)
1243 return N->getOperand(0);
1244 else if (N->getOperand(NumOps-1).getValueType() == MVT::Other)
1245 return N->getOperand(NumOps-1);
1246 for (unsigned i = 1; i < NumOps-1; ++i)
1247 if (N->getOperand(i).getValueType() == MVT::Other)
1248 return N->getOperand(i);
1253 SDValue DAGCombiner::visitTokenFactor(SDNode *N) {
1254 // If N has two operands, where one has an input chain equal to the other,
1255 // the 'other' chain is redundant.
1256 if (N->getNumOperands() == 2) {
1257 if (getInputChainForNode(N->getOperand(0).getNode()) == N->getOperand(1))
1258 return N->getOperand(0);
1259 if (getInputChainForNode(N->getOperand(1).getNode()) == N->getOperand(0))
1260 return N->getOperand(1);
1263 SmallVector<SDNode *, 8> TFs; // List of token factors to visit.
1264 SmallVector<SDValue, 8> Ops; // Ops for replacing token factor.
1265 SmallPtrSet<SDNode*, 16> SeenOps;
1266 bool Changed = false; // If we should replace this token factor.
1268 // Start out with this token factor.
1271 // Iterate through token factors. The TFs grows when new token factors are
1273 for (unsigned i = 0; i < TFs.size(); ++i) {
1274 SDNode *TF = TFs[i];
1276 // Check each of the operands.
1277 for (unsigned i = 0, ie = TF->getNumOperands(); i != ie; ++i) {
1278 SDValue Op = TF->getOperand(i);
1280 switch (Op.getOpcode()) {
1281 case ISD::EntryToken:
1282 // Entry tokens don't need to be added to the list. They are
1287 case ISD::TokenFactor:
1288 if (Op.hasOneUse() &&
1289 std::find(TFs.begin(), TFs.end(), Op.getNode()) == TFs.end()) {
1290 // Queue up for processing.
1291 TFs.push_back(Op.getNode());
1292 // Clean up in case the token factor is removed.
1293 AddToWorkList(Op.getNode());
1300 // Only add if it isn't already in the list.
1301 if (SeenOps.insert(Op.getNode()))
1312 // If we've change things around then replace token factor.
1315 // The entry token is the only possible outcome.
1316 Result = DAG.getEntryNode();
1318 // New and improved token factor.
1319 Result = DAG.getNode(ISD::TokenFactor, N->getDebugLoc(),
1320 MVT::Other, &Ops[0], Ops.size());
1323 // Don't add users to work list.
1324 return CombineTo(N, Result, false);
1330 /// MERGE_VALUES can always be eliminated.
1331 SDValue DAGCombiner::visitMERGE_VALUES(SDNode *N) {
1332 WorkListRemover DeadNodes(*this);
1333 // Replacing results may cause a different MERGE_VALUES to suddenly
1334 // be CSE'd with N, and carry its uses with it. Iterate until no
1335 // uses remain, to ensure that the node can be safely deleted.
1336 // First add the users of this node to the work list so that they
1337 // can be tried again once they have new operands.
1338 AddUsersToWorkList(N);
1340 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1341 DAG.ReplaceAllUsesOfValueWith(SDValue(N, i), N->getOperand(i));
1342 } while (!N->use_empty());
1343 removeFromWorkList(N);
1345 return SDValue(N, 0); // Return N so it doesn't get rechecked!
1349 SDValue combineShlAddConstant(DebugLoc DL, SDValue N0, SDValue N1,
1350 SelectionDAG &DAG) {
1351 EVT VT = N0.getValueType();
1352 SDValue N00 = N0.getOperand(0);
1353 SDValue N01 = N0.getOperand(1);
1354 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N01);
1356 if (N01C && N00.getOpcode() == ISD::ADD && N00.getNode()->hasOneUse() &&
1357 isa<ConstantSDNode>(N00.getOperand(1))) {
1358 // fold (add (shl (add x, c1), c2), ) -> (add (add (shl x, c2), c1<<c2), )
1359 N0 = DAG.getNode(ISD::ADD, N0.getDebugLoc(), VT,
1360 DAG.getNode(ISD::SHL, N00.getDebugLoc(), VT,
1361 N00.getOperand(0), N01),
1362 DAG.getNode(ISD::SHL, N01.getDebugLoc(), VT,
1363 N00.getOperand(1), N01));
1364 return DAG.getNode(ISD::ADD, DL, VT, N0, N1);
1370 SDValue DAGCombiner::visitADD(SDNode *N) {
1371 SDValue N0 = N->getOperand(0);
1372 SDValue N1 = N->getOperand(1);
1373 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1374 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1375 EVT VT = N0.getValueType();
1378 if (VT.isVector()) {
1379 SDValue FoldedVOp = SimplifyVBinOp(N);
1380 if (FoldedVOp.getNode()) return FoldedVOp;
1383 // fold (add x, undef) -> undef
1384 if (N0.getOpcode() == ISD::UNDEF)
1386 if (N1.getOpcode() == ISD::UNDEF)
1388 // fold (add c1, c2) -> c1+c2
1390 return DAG.FoldConstantArithmetic(ISD::ADD, VT, N0C, N1C);
1391 // canonicalize constant to RHS
1393 return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N1, N0);
1394 // fold (add x, 0) -> x
1395 if (N1C && N1C->isNullValue())
1397 // fold (add Sym, c) -> Sym+c
1398 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1399 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA) && N1C &&
1400 GA->getOpcode() == ISD::GlobalAddress)
1401 return DAG.getGlobalAddress(GA->getGlobal(), N1C->getDebugLoc(), VT,
1403 (uint64_t)N1C->getSExtValue());
1404 // fold ((c1-A)+c2) -> (c1+c2)-A
1405 if (N1C && N0.getOpcode() == ISD::SUB)
1406 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getOperand(0)))
1407 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1408 DAG.getConstant(N1C->getAPIntValue()+
1409 N0C->getAPIntValue(), VT),
1412 SDValue RADD = ReassociateOps(ISD::ADD, N->getDebugLoc(), N0, N1);
1413 if (RADD.getNode() != 0)
1415 // fold ((0-A) + B) -> B-A
1416 if (N0.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N0.getOperand(0)) &&
1417 cast<ConstantSDNode>(N0.getOperand(0))->isNullValue())
1418 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N1, N0.getOperand(1));
1419 // fold (A + (0-B)) -> A-B
1420 if (N1.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N1.getOperand(0)) &&
1421 cast<ConstantSDNode>(N1.getOperand(0))->isNullValue())
1422 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N0, N1.getOperand(1));
1423 // fold (A+(B-A)) -> B
1424 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1))
1425 return N1.getOperand(0);
1426 // fold ((B-A)+A) -> B
1427 if (N0.getOpcode() == ISD::SUB && N1 == N0.getOperand(1))
1428 return N0.getOperand(0);
1429 // fold (A+(B-(A+C))) to (B-C)
1430 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1431 N0 == N1.getOperand(1).getOperand(0))
1432 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N1.getOperand(0),
1433 N1.getOperand(1).getOperand(1));
1434 // fold (A+(B-(C+A))) to (B-C)
1435 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1436 N0 == N1.getOperand(1).getOperand(1))
1437 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N1.getOperand(0),
1438 N1.getOperand(1).getOperand(0));
1439 // fold (A+((B-A)+or-C)) to (B+or-C)
1440 if ((N1.getOpcode() == ISD::SUB || N1.getOpcode() == ISD::ADD) &&
1441 N1.getOperand(0).getOpcode() == ISD::SUB &&
1442 N0 == N1.getOperand(0).getOperand(1))
1443 return DAG.getNode(N1.getOpcode(), N->getDebugLoc(), VT,
1444 N1.getOperand(0).getOperand(0), N1.getOperand(1));
1446 // fold (A-B)+(C-D) to (A+C)-(B+D) when A or C is constant
1447 if (N0.getOpcode() == ISD::SUB && N1.getOpcode() == ISD::SUB) {
1448 SDValue N00 = N0.getOperand(0);
1449 SDValue N01 = N0.getOperand(1);
1450 SDValue N10 = N1.getOperand(0);
1451 SDValue N11 = N1.getOperand(1);
1453 if (isa<ConstantSDNode>(N00) || isa<ConstantSDNode>(N10))
1454 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1455 DAG.getNode(ISD::ADD, N0.getDebugLoc(), VT, N00, N10),
1456 DAG.getNode(ISD::ADD, N1.getDebugLoc(), VT, N01, N11));
1459 if (!VT.isVector() && SimplifyDemandedBits(SDValue(N, 0)))
1460 return SDValue(N, 0);
1462 // fold (a+b) -> (a|b) iff a and b share no bits.
1463 if (VT.isInteger() && !VT.isVector()) {
1464 APInt LHSZero, LHSOne;
1465 APInt RHSZero, RHSOne;
1466 DAG.ComputeMaskedBits(N0, LHSZero, LHSOne);
1468 if (LHSZero.getBoolValue()) {
1469 DAG.ComputeMaskedBits(N1, RHSZero, RHSOne);
1471 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1472 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1473 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero)
1474 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N0, N1);
1478 // fold (add (shl (add x, c1), c2), ) -> (add (add (shl x, c2), c1<<c2), )
1479 if (N0.getOpcode() == ISD::SHL && N0.getNode()->hasOneUse()) {
1480 SDValue Result = combineShlAddConstant(N->getDebugLoc(), N0, N1, DAG);
1481 if (Result.getNode()) return Result;
1483 if (N1.getOpcode() == ISD::SHL && N1.getNode()->hasOneUse()) {
1484 SDValue Result = combineShlAddConstant(N->getDebugLoc(), N1, N0, DAG);
1485 if (Result.getNode()) return Result;
1488 // fold (add x, shl(0 - y, n)) -> sub(x, shl(y, n))
1489 if (N1.getOpcode() == ISD::SHL &&
1490 N1.getOperand(0).getOpcode() == ISD::SUB)
1491 if (ConstantSDNode *C =
1492 dyn_cast<ConstantSDNode>(N1.getOperand(0).getOperand(0)))
1493 if (C->getAPIntValue() == 0)
1494 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N0,
1495 DAG.getNode(ISD::SHL, N->getDebugLoc(), VT,
1496 N1.getOperand(0).getOperand(1),
1498 if (N0.getOpcode() == ISD::SHL &&
1499 N0.getOperand(0).getOpcode() == ISD::SUB)
1500 if (ConstantSDNode *C =
1501 dyn_cast<ConstantSDNode>(N0.getOperand(0).getOperand(0)))
1502 if (C->getAPIntValue() == 0)
1503 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N1,
1504 DAG.getNode(ISD::SHL, N->getDebugLoc(), VT,
1505 N0.getOperand(0).getOperand(1),
1508 if (N1.getOpcode() == ISD::AND) {
1509 SDValue AndOp0 = N1.getOperand(0);
1510 ConstantSDNode *AndOp1 = dyn_cast<ConstantSDNode>(N1->getOperand(1));
1511 unsigned NumSignBits = DAG.ComputeNumSignBits(AndOp0);
1512 unsigned DestBits = VT.getScalarType().getSizeInBits();
1514 // (add z, (and (sbbl x, x), 1)) -> (sub z, (sbbl x, x))
1515 // and similar xforms where the inner op is either ~0 or 0.
1516 if (NumSignBits == DestBits && AndOp1 && AndOp1->isOne()) {
1517 DebugLoc DL = N->getDebugLoc();
1518 return DAG.getNode(ISD::SUB, DL, VT, N->getOperand(0), AndOp0);
1522 // add (sext i1), X -> sub X, (zext i1)
1523 if (N0.getOpcode() == ISD::SIGN_EXTEND &&
1524 N0.getOperand(0).getValueType() == MVT::i1 &&
1525 !TLI.isOperationLegal(ISD::SIGN_EXTEND, MVT::i1)) {
1526 DebugLoc DL = N->getDebugLoc();
1527 SDValue ZExt = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0));
1528 return DAG.getNode(ISD::SUB, DL, VT, N1, ZExt);
1534 SDValue DAGCombiner::visitADDC(SDNode *N) {
1535 SDValue N0 = N->getOperand(0);
1536 SDValue N1 = N->getOperand(1);
1537 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1538 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1539 EVT VT = N0.getValueType();
1541 // If the flag result is dead, turn this into an ADD.
1542 if (!N->hasAnyUseOfValue(1))
1543 return CombineTo(N, DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N0, N1),
1544 DAG.getNode(ISD::CARRY_FALSE,
1545 N->getDebugLoc(), MVT::Glue));
1547 // canonicalize constant to RHS.
1549 return DAG.getNode(ISD::ADDC, N->getDebugLoc(), N->getVTList(), N1, N0);
1551 // fold (addc x, 0) -> x + no carry out
1552 if (N1C && N1C->isNullValue())
1553 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE,
1554 N->getDebugLoc(), MVT::Glue));
1556 // fold (addc a, b) -> (or a, b), CARRY_FALSE iff a and b share no bits.
1557 APInt LHSZero, LHSOne;
1558 APInt RHSZero, RHSOne;
1559 DAG.ComputeMaskedBits(N0, LHSZero, LHSOne);
1561 if (LHSZero.getBoolValue()) {
1562 DAG.ComputeMaskedBits(N1, RHSZero, RHSOne);
1564 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1565 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1566 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero)
1567 return CombineTo(N, DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N0, N1),
1568 DAG.getNode(ISD::CARRY_FALSE,
1569 N->getDebugLoc(), MVT::Glue));
1575 SDValue DAGCombiner::visitADDE(SDNode *N) {
1576 SDValue N0 = N->getOperand(0);
1577 SDValue N1 = N->getOperand(1);
1578 SDValue CarryIn = N->getOperand(2);
1579 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1580 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1582 // canonicalize constant to RHS
1584 return DAG.getNode(ISD::ADDE, N->getDebugLoc(), N->getVTList(),
1587 // fold (adde x, y, false) -> (addc x, y)
1588 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1589 return DAG.getNode(ISD::ADDC, N->getDebugLoc(), N->getVTList(), N0, N1);
1594 // Since it may not be valid to emit a fold to zero for vector initializers
1595 // check if we can before folding.
1596 static SDValue tryFoldToZero(DebugLoc DL, const TargetLowering &TLI, EVT VT,
1597 SelectionDAG &DAG, bool LegalOperations) {
1598 if (!VT.isVector()) {
1599 return DAG.getConstant(0, VT);
1601 if (!LegalOperations || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT)) {
1602 // Produce a vector of zeros.
1603 SDValue El = DAG.getConstant(0, VT.getVectorElementType());
1604 std::vector<SDValue> Ops(VT.getVectorNumElements(), El);
1605 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT,
1606 &Ops[0], Ops.size());
1611 SDValue DAGCombiner::visitSUB(SDNode *N) {
1612 SDValue N0 = N->getOperand(0);
1613 SDValue N1 = N->getOperand(1);
1614 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1615 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1616 ConstantSDNode *N1C1 = N1.getOpcode() != ISD::ADD ? 0 :
1617 dyn_cast<ConstantSDNode>(N1.getOperand(1).getNode());
1618 EVT VT = N0.getValueType();
1621 if (VT.isVector()) {
1622 SDValue FoldedVOp = SimplifyVBinOp(N);
1623 if (FoldedVOp.getNode()) return FoldedVOp;
1626 // fold (sub x, x) -> 0
1627 // FIXME: Refactor this and xor and other similar operations together.
1629 return tryFoldToZero(N->getDebugLoc(), TLI, VT, DAG, LegalOperations);
1630 // fold (sub c1, c2) -> c1-c2
1632 return DAG.FoldConstantArithmetic(ISD::SUB, VT, N0C, N1C);
1633 // fold (sub x, c) -> (add x, -c)
1635 return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N0,
1636 DAG.getConstant(-N1C->getAPIntValue(), VT));
1637 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1)
1638 if (N0C && N0C->isAllOnesValue())
1639 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N1, N0);
1640 // fold A-(A-B) -> B
1641 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(0))
1642 return N1.getOperand(1);
1643 // fold (A+B)-A -> B
1644 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1)
1645 return N0.getOperand(1);
1646 // fold (A+B)-B -> A
1647 if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1)
1648 return N0.getOperand(0);
1649 // fold C2-(A+C1) -> (C2-C1)-A
1650 if (N1.getOpcode() == ISD::ADD && N0C && N1C1) {
1651 SDValue NewC = DAG.getConstant(N0C->getAPIntValue() - N1C1->getAPIntValue(),
1653 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, NewC,
1656 // fold ((A+(B+or-C))-B) -> A+or-C
1657 if (N0.getOpcode() == ISD::ADD &&
1658 (N0.getOperand(1).getOpcode() == ISD::SUB ||
1659 N0.getOperand(1).getOpcode() == ISD::ADD) &&
1660 N0.getOperand(1).getOperand(0) == N1)
1661 return DAG.getNode(N0.getOperand(1).getOpcode(), N->getDebugLoc(), VT,
1662 N0.getOperand(0), N0.getOperand(1).getOperand(1));
1663 // fold ((A+(C+B))-B) -> A+C
1664 if (N0.getOpcode() == ISD::ADD &&
1665 N0.getOperand(1).getOpcode() == ISD::ADD &&
1666 N0.getOperand(1).getOperand(1) == N1)
1667 return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT,
1668 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1669 // fold ((A-(B-C))-C) -> A-B
1670 if (N0.getOpcode() == ISD::SUB &&
1671 N0.getOperand(1).getOpcode() == ISD::SUB &&
1672 N0.getOperand(1).getOperand(1) == N1)
1673 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1674 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1676 // If either operand of a sub is undef, the result is undef
1677 if (N0.getOpcode() == ISD::UNDEF)
1679 if (N1.getOpcode() == ISD::UNDEF)
1682 // If the relocation model supports it, consider symbol offsets.
1683 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1684 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA)) {
1685 // fold (sub Sym, c) -> Sym-c
1686 if (N1C && GA->getOpcode() == ISD::GlobalAddress)
1687 return DAG.getGlobalAddress(GA->getGlobal(), N1C->getDebugLoc(), VT,
1689 (uint64_t)N1C->getSExtValue());
1690 // fold (sub Sym+c1, Sym+c2) -> c1-c2
1691 if (GlobalAddressSDNode *GB = dyn_cast<GlobalAddressSDNode>(N1))
1692 if (GA->getGlobal() == GB->getGlobal())
1693 return DAG.getConstant((uint64_t)GA->getOffset() - GB->getOffset(),
1700 SDValue DAGCombiner::visitSUBC(SDNode *N) {
1701 SDValue N0 = N->getOperand(0);
1702 SDValue N1 = N->getOperand(1);
1703 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1704 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1705 EVT VT = N0.getValueType();
1707 // If the flag result is dead, turn this into an SUB.
1708 if (!N->hasAnyUseOfValue(1))
1709 return CombineTo(N, DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N0, N1),
1710 DAG.getNode(ISD::CARRY_FALSE, N->getDebugLoc(),
1713 // fold (subc x, x) -> 0 + no borrow
1715 return CombineTo(N, DAG.getConstant(0, VT),
1716 DAG.getNode(ISD::CARRY_FALSE, N->getDebugLoc(),
1719 // fold (subc x, 0) -> x + no borrow
1720 if (N1C && N1C->isNullValue())
1721 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE, N->getDebugLoc(),
1724 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1) + no borrow
1725 if (N0C && N0C->isAllOnesValue())
1726 return CombineTo(N, DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N1, N0),
1727 DAG.getNode(ISD::CARRY_FALSE, N->getDebugLoc(),
1733 SDValue DAGCombiner::visitSUBE(SDNode *N) {
1734 SDValue N0 = N->getOperand(0);
1735 SDValue N1 = N->getOperand(1);
1736 SDValue CarryIn = N->getOperand(2);
1738 // fold (sube x, y, false) -> (subc x, y)
1739 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1740 return DAG.getNode(ISD::SUBC, N->getDebugLoc(), N->getVTList(), N0, N1);
1745 SDValue DAGCombiner::visitMUL(SDNode *N) {
1746 SDValue N0 = N->getOperand(0);
1747 SDValue N1 = N->getOperand(1);
1748 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1749 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1750 EVT VT = N0.getValueType();
1753 if (VT.isVector()) {
1754 SDValue FoldedVOp = SimplifyVBinOp(N);
1755 if (FoldedVOp.getNode()) return FoldedVOp;
1758 // fold (mul x, undef) -> 0
1759 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
1760 return DAG.getConstant(0, VT);
1761 // fold (mul c1, c2) -> c1*c2
1763 return DAG.FoldConstantArithmetic(ISD::MUL, VT, N0C, N1C);
1764 // canonicalize constant to RHS
1766 return DAG.getNode(ISD::MUL, N->getDebugLoc(), VT, N1, N0);
1767 // fold (mul x, 0) -> 0
1768 if (N1C && N1C->isNullValue())
1770 // fold (mul x, -1) -> 0-x
1771 if (N1C && N1C->isAllOnesValue())
1772 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1773 DAG.getConstant(0, VT), N0);
1774 // fold (mul x, (1 << c)) -> x << c
1775 if (N1C && N1C->getAPIntValue().isPowerOf2())
1776 return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0,
1777 DAG.getConstant(N1C->getAPIntValue().logBase2(),
1778 getShiftAmountTy(N0.getValueType())));
1779 // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c
1780 if (N1C && (-N1C->getAPIntValue()).isPowerOf2()) {
1781 unsigned Log2Val = (-N1C->getAPIntValue()).logBase2();
1782 // FIXME: If the input is something that is easily negated (e.g. a
1783 // single-use add), we should put the negate there.
1784 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1785 DAG.getConstant(0, VT),
1786 DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0,
1787 DAG.getConstant(Log2Val,
1788 getShiftAmountTy(N0.getValueType()))));
1790 // (mul (shl X, c1), c2) -> (mul X, c2 << c1)
1791 if (N1C && N0.getOpcode() == ISD::SHL &&
1792 isa<ConstantSDNode>(N0.getOperand(1))) {
1793 SDValue C3 = DAG.getNode(ISD::SHL, N->getDebugLoc(), VT,
1794 N1, N0.getOperand(1));
1795 AddToWorkList(C3.getNode());
1796 return DAG.getNode(ISD::MUL, N->getDebugLoc(), VT,
1797 N0.getOperand(0), C3);
1800 // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one
1803 SDValue Sh(0,0), Y(0,0);
1804 // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)).
1805 if (N0.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N0.getOperand(1)) &&
1806 N0.getNode()->hasOneUse()) {
1808 } else if (N1.getOpcode() == ISD::SHL &&
1809 isa<ConstantSDNode>(N1.getOperand(1)) &&
1810 N1.getNode()->hasOneUse()) {
1815 SDValue Mul = DAG.getNode(ISD::MUL, N->getDebugLoc(), VT,
1816 Sh.getOperand(0), Y);
1817 return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT,
1818 Mul, Sh.getOperand(1));
1822 // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2)
1823 if (N1C && N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() &&
1824 isa<ConstantSDNode>(N0.getOperand(1)))
1825 return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT,
1826 DAG.getNode(ISD::MUL, N0.getDebugLoc(), VT,
1827 N0.getOperand(0), N1),
1828 DAG.getNode(ISD::MUL, N1.getDebugLoc(), VT,
1829 N0.getOperand(1), N1));
1832 SDValue RMUL = ReassociateOps(ISD::MUL, N->getDebugLoc(), N0, N1);
1833 if (RMUL.getNode() != 0)
1839 SDValue DAGCombiner::visitSDIV(SDNode *N) {
1840 SDValue N0 = N->getOperand(0);
1841 SDValue N1 = N->getOperand(1);
1842 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1843 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1844 EVT VT = N->getValueType(0);
1847 if (VT.isVector()) {
1848 SDValue FoldedVOp = SimplifyVBinOp(N);
1849 if (FoldedVOp.getNode()) return FoldedVOp;
1852 // fold (sdiv c1, c2) -> c1/c2
1853 if (N0C && N1C && !N1C->isNullValue())
1854 return DAG.FoldConstantArithmetic(ISD::SDIV, VT, N0C, N1C);
1855 // fold (sdiv X, 1) -> X
1856 if (N1C && N1C->getAPIntValue() == 1LL)
1858 // fold (sdiv X, -1) -> 0-X
1859 if (N1C && N1C->isAllOnesValue())
1860 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1861 DAG.getConstant(0, VT), N0);
1862 // If we know the sign bits of both operands are zero, strength reduce to a
1863 // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2
1864 if (!VT.isVector()) {
1865 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
1866 return DAG.getNode(ISD::UDIV, N->getDebugLoc(), N1.getValueType(),
1869 // fold (sdiv X, pow2) -> simple ops after legalize
1870 if (N1C && !N1C->isNullValue() &&
1871 (N1C->getAPIntValue().isPowerOf2() ||
1872 (-N1C->getAPIntValue()).isPowerOf2())) {
1873 // If dividing by powers of two is cheap, then don't perform the following
1875 if (TLI.isPow2DivCheap())
1878 unsigned lg2 = N1C->getAPIntValue().countTrailingZeros();
1880 // Splat the sign bit into the register
1881 SDValue SGN = DAG.getNode(ISD::SRA, N->getDebugLoc(), VT, N0,
1882 DAG.getConstant(VT.getSizeInBits()-1,
1883 getShiftAmountTy(N0.getValueType())));
1884 AddToWorkList(SGN.getNode());
1886 // Add (N0 < 0) ? abs2 - 1 : 0;
1887 SDValue SRL = DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, SGN,
1888 DAG.getConstant(VT.getSizeInBits() - lg2,
1889 getShiftAmountTy(SGN.getValueType())));
1890 SDValue ADD = DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N0, SRL);
1891 AddToWorkList(SRL.getNode());
1892 AddToWorkList(ADD.getNode()); // Divide by pow2
1893 SDValue SRA = DAG.getNode(ISD::SRA, N->getDebugLoc(), VT, ADD,
1894 DAG.getConstant(lg2, getShiftAmountTy(ADD.getValueType())));
1896 // If we're dividing by a positive value, we're done. Otherwise, we must
1897 // negate the result.
1898 if (N1C->getAPIntValue().isNonNegative())
1901 AddToWorkList(SRA.getNode());
1902 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1903 DAG.getConstant(0, VT), SRA);
1906 // if integer divide is expensive and we satisfy the requirements, emit an
1907 // alternate sequence.
1908 if (N1C && !N1C->isNullValue() && !TLI.isIntDivCheap()) {
1909 SDValue Op = BuildSDIV(N);
1910 if (Op.getNode()) return Op;
1914 if (N0.getOpcode() == ISD::UNDEF)
1915 return DAG.getConstant(0, VT);
1916 // X / undef -> undef
1917 if (N1.getOpcode() == ISD::UNDEF)
1923 SDValue DAGCombiner::visitUDIV(SDNode *N) {
1924 SDValue N0 = N->getOperand(0);
1925 SDValue N1 = N->getOperand(1);
1926 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1927 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1928 EVT VT = N->getValueType(0);
1931 if (VT.isVector()) {
1932 SDValue FoldedVOp = SimplifyVBinOp(N);
1933 if (FoldedVOp.getNode()) return FoldedVOp;
1936 // fold (udiv c1, c2) -> c1/c2
1937 if (N0C && N1C && !N1C->isNullValue())
1938 return DAG.FoldConstantArithmetic(ISD::UDIV, VT, N0C, N1C);
1939 // fold (udiv x, (1 << c)) -> x >>u c
1940 if (N1C && N1C->getAPIntValue().isPowerOf2())
1941 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0,
1942 DAG.getConstant(N1C->getAPIntValue().logBase2(),
1943 getShiftAmountTy(N0.getValueType())));
1944 // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2
1945 if (N1.getOpcode() == ISD::SHL) {
1946 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
1947 if (SHC->getAPIntValue().isPowerOf2()) {
1948 EVT ADDVT = N1.getOperand(1).getValueType();
1949 SDValue Add = DAG.getNode(ISD::ADD, N->getDebugLoc(), ADDVT,
1951 DAG.getConstant(SHC->getAPIntValue()
1954 AddToWorkList(Add.getNode());
1955 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0, Add);
1959 // fold (udiv x, c) -> alternate
1960 if (N1C && !N1C->isNullValue() && !TLI.isIntDivCheap()) {
1961 SDValue Op = BuildUDIV(N);
1962 if (Op.getNode()) return Op;
1966 if (N0.getOpcode() == ISD::UNDEF)
1967 return DAG.getConstant(0, VT);
1968 // X / undef -> undef
1969 if (N1.getOpcode() == ISD::UNDEF)
1975 SDValue DAGCombiner::visitSREM(SDNode *N) {
1976 SDValue N0 = N->getOperand(0);
1977 SDValue N1 = N->getOperand(1);
1978 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1979 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1980 EVT VT = N->getValueType(0);
1982 // fold (srem c1, c2) -> c1%c2
1983 if (N0C && N1C && !N1C->isNullValue())
1984 return DAG.FoldConstantArithmetic(ISD::SREM, VT, N0C, N1C);
1985 // If we know the sign bits of both operands are zero, strength reduce to a
1986 // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15
1987 if (!VT.isVector()) {
1988 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
1989 return DAG.getNode(ISD::UREM, N->getDebugLoc(), VT, N0, N1);
1992 // If X/C can be simplified by the division-by-constant logic, lower
1993 // X%C to the equivalent of X-X/C*C.
1994 if (N1C && !N1C->isNullValue()) {
1995 SDValue Div = DAG.getNode(ISD::SDIV, N->getDebugLoc(), VT, N0, N1);
1996 AddToWorkList(Div.getNode());
1997 SDValue OptimizedDiv = combine(Div.getNode());
1998 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
1999 SDValue Mul = DAG.getNode(ISD::MUL, N->getDebugLoc(), VT,
2001 SDValue Sub = DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N0, Mul);
2002 AddToWorkList(Mul.getNode());
2008 if (N0.getOpcode() == ISD::UNDEF)
2009 return DAG.getConstant(0, VT);
2010 // X % undef -> undef
2011 if (N1.getOpcode() == ISD::UNDEF)
2017 SDValue DAGCombiner::visitUREM(SDNode *N) {
2018 SDValue N0 = N->getOperand(0);
2019 SDValue N1 = N->getOperand(1);
2020 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2021 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2022 EVT VT = N->getValueType(0);
2024 // fold (urem c1, c2) -> c1%c2
2025 if (N0C && N1C && !N1C->isNullValue())
2026 return DAG.FoldConstantArithmetic(ISD::UREM, VT, N0C, N1C);
2027 // fold (urem x, pow2) -> (and x, pow2-1)
2028 if (N1C && !N1C->isNullValue() && N1C->getAPIntValue().isPowerOf2())
2029 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0,
2030 DAG.getConstant(N1C->getAPIntValue()-1,VT));
2031 // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1))
2032 if (N1.getOpcode() == ISD::SHL) {
2033 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
2034 if (SHC->getAPIntValue().isPowerOf2()) {
2036 DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N1,
2037 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()),
2039 AddToWorkList(Add.getNode());
2040 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0, Add);
2045 // If X/C can be simplified by the division-by-constant logic, lower
2046 // X%C to the equivalent of X-X/C*C.
2047 if (N1C && !N1C->isNullValue()) {
2048 SDValue Div = DAG.getNode(ISD::UDIV, N->getDebugLoc(), VT, N0, N1);
2049 AddToWorkList(Div.getNode());
2050 SDValue OptimizedDiv = combine(Div.getNode());
2051 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2052 SDValue Mul = DAG.getNode(ISD::MUL, N->getDebugLoc(), VT,
2054 SDValue Sub = DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N0, Mul);
2055 AddToWorkList(Mul.getNode());
2061 if (N0.getOpcode() == ISD::UNDEF)
2062 return DAG.getConstant(0, VT);
2063 // X % undef -> undef
2064 if (N1.getOpcode() == ISD::UNDEF)
2070 SDValue DAGCombiner::visitMULHS(SDNode *N) {
2071 SDValue N0 = N->getOperand(0);
2072 SDValue N1 = N->getOperand(1);
2073 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2074 EVT VT = N->getValueType(0);
2075 DebugLoc DL = N->getDebugLoc();
2077 // fold (mulhs x, 0) -> 0
2078 if (N1C && N1C->isNullValue())
2080 // fold (mulhs x, 1) -> (sra x, size(x)-1)
2081 if (N1C && N1C->getAPIntValue() == 1)
2082 return DAG.getNode(ISD::SRA, N->getDebugLoc(), N0.getValueType(), N0,
2083 DAG.getConstant(N0.getValueType().getSizeInBits() - 1,
2084 getShiftAmountTy(N0.getValueType())));
2085 // fold (mulhs x, undef) -> 0
2086 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2087 return DAG.getConstant(0, VT);
2089 // If the type twice as wide is legal, transform the mulhs to a wider multiply
2091 if (VT.isSimple() && !VT.isVector()) {
2092 MVT Simple = VT.getSimpleVT();
2093 unsigned SimpleSize = Simple.getSizeInBits();
2094 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2095 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2096 N0 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N0);
2097 N1 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N1);
2098 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2099 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2100 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2101 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2108 SDValue DAGCombiner::visitMULHU(SDNode *N) {
2109 SDValue N0 = N->getOperand(0);
2110 SDValue N1 = N->getOperand(1);
2111 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2112 EVT VT = N->getValueType(0);
2113 DebugLoc DL = N->getDebugLoc();
2115 // fold (mulhu x, 0) -> 0
2116 if (N1C && N1C->isNullValue())
2118 // fold (mulhu x, 1) -> 0
2119 if (N1C && N1C->getAPIntValue() == 1)
2120 return DAG.getConstant(0, N0.getValueType());
2121 // fold (mulhu x, undef) -> 0
2122 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2123 return DAG.getConstant(0, VT);
2125 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2127 if (VT.isSimple() && !VT.isVector()) {
2128 MVT Simple = VT.getSimpleVT();
2129 unsigned SimpleSize = Simple.getSizeInBits();
2130 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2131 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2132 N0 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N0);
2133 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N1);
2134 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2135 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2136 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2137 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2144 /// SimplifyNodeWithTwoResults - Perform optimizations common to nodes that
2145 /// compute two values. LoOp and HiOp give the opcodes for the two computations
2146 /// that are being performed. Return true if a simplification was made.
2148 SDValue DAGCombiner::SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
2150 // If the high half is not needed, just compute the low half.
2151 bool HiExists = N->hasAnyUseOfValue(1);
2153 (!LegalOperations ||
2154 TLI.isOperationLegal(LoOp, N->getValueType(0)))) {
2155 SDValue Res = DAG.getNode(LoOp, N->getDebugLoc(), N->getValueType(0),
2156 N->op_begin(), N->getNumOperands());
2157 return CombineTo(N, Res, Res);
2160 // If the low half is not needed, just compute the high half.
2161 bool LoExists = N->hasAnyUseOfValue(0);
2163 (!LegalOperations ||
2164 TLI.isOperationLegal(HiOp, N->getValueType(1)))) {
2165 SDValue Res = DAG.getNode(HiOp, N->getDebugLoc(), N->getValueType(1),
2166 N->op_begin(), N->getNumOperands());
2167 return CombineTo(N, Res, Res);
2170 // If both halves are used, return as it is.
2171 if (LoExists && HiExists)
2174 // If the two computed results can be simplified separately, separate them.
2176 SDValue Lo = DAG.getNode(LoOp, N->getDebugLoc(), N->getValueType(0),
2177 N->op_begin(), N->getNumOperands());
2178 AddToWorkList(Lo.getNode());
2179 SDValue LoOpt = combine(Lo.getNode());
2180 if (LoOpt.getNode() && LoOpt.getNode() != Lo.getNode() &&
2181 (!LegalOperations ||
2182 TLI.isOperationLegal(LoOpt.getOpcode(), LoOpt.getValueType())))
2183 return CombineTo(N, LoOpt, LoOpt);
2187 SDValue Hi = DAG.getNode(HiOp, N->getDebugLoc(), N->getValueType(1),
2188 N->op_begin(), N->getNumOperands());
2189 AddToWorkList(Hi.getNode());
2190 SDValue HiOpt = combine(Hi.getNode());
2191 if (HiOpt.getNode() && HiOpt != Hi &&
2192 (!LegalOperations ||
2193 TLI.isOperationLegal(HiOpt.getOpcode(), HiOpt.getValueType())))
2194 return CombineTo(N, HiOpt, HiOpt);
2200 SDValue DAGCombiner::visitSMUL_LOHI(SDNode *N) {
2201 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHS);
2202 if (Res.getNode()) return Res;
2204 EVT VT = N->getValueType(0);
2205 DebugLoc DL = N->getDebugLoc();
2207 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2209 if (VT.isSimple() && !VT.isVector()) {
2210 MVT Simple = VT.getSimpleVT();
2211 unsigned SimpleSize = Simple.getSizeInBits();
2212 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2213 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2214 SDValue Lo = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(0));
2215 SDValue Hi = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(1));
2216 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2217 // Compute the high part as N1.
2218 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2219 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2220 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2221 // Compute the low part as N0.
2222 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2223 return CombineTo(N, Lo, Hi);
2230 SDValue DAGCombiner::visitUMUL_LOHI(SDNode *N) {
2231 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHU);
2232 if (Res.getNode()) return Res;
2234 EVT VT = N->getValueType(0);
2235 DebugLoc DL = N->getDebugLoc();
2237 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2239 if (VT.isSimple() && !VT.isVector()) {
2240 MVT Simple = VT.getSimpleVT();
2241 unsigned SimpleSize = Simple.getSizeInBits();
2242 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2243 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2244 SDValue Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(0));
2245 SDValue Hi = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(1));
2246 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2247 // Compute the high part as N1.
2248 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2249 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2250 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2251 // Compute the low part as N0.
2252 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2253 return CombineTo(N, Lo, Hi);
2260 SDValue DAGCombiner::visitSMULO(SDNode *N) {
2261 // (smulo x, 2) -> (saddo x, x)
2262 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2263 if (C2->getAPIntValue() == 2)
2264 return DAG.getNode(ISD::SADDO, N->getDebugLoc(), N->getVTList(),
2265 N->getOperand(0), N->getOperand(0));
2270 SDValue DAGCombiner::visitUMULO(SDNode *N) {
2271 // (umulo x, 2) -> (uaddo x, x)
2272 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2273 if (C2->getAPIntValue() == 2)
2274 return DAG.getNode(ISD::UADDO, N->getDebugLoc(), N->getVTList(),
2275 N->getOperand(0), N->getOperand(0));
2280 SDValue DAGCombiner::visitSDIVREM(SDNode *N) {
2281 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::SDIV, ISD::SREM);
2282 if (Res.getNode()) return Res;
2287 SDValue DAGCombiner::visitUDIVREM(SDNode *N) {
2288 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::UDIV, ISD::UREM);
2289 if (Res.getNode()) return Res;
2294 /// SimplifyBinOpWithSameOpcodeHands - If this is a binary operator with
2295 /// two operands of the same opcode, try to simplify it.
2296 SDValue DAGCombiner::SimplifyBinOpWithSameOpcodeHands(SDNode *N) {
2297 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
2298 EVT VT = N0.getValueType();
2299 assert(N0.getOpcode() == N1.getOpcode() && "Bad input!");
2301 // Bail early if none of these transforms apply.
2302 if (N0.getNode()->getNumOperands() == 0) return SDValue();
2304 // For each of OP in AND/OR/XOR:
2305 // fold (OP (zext x), (zext y)) -> (zext (OP x, y))
2306 // fold (OP (sext x), (sext y)) -> (sext (OP x, y))
2307 // fold (OP (aext x), (aext y)) -> (aext (OP x, y))
2308 // fold (OP (trunc x), (trunc y)) -> (trunc (OP x, y)) (if trunc isn't free)
2310 // do not sink logical op inside of a vector extend, since it may combine
2312 EVT Op0VT = N0.getOperand(0).getValueType();
2313 if ((N0.getOpcode() == ISD::ZERO_EXTEND ||
2314 N0.getOpcode() == ISD::SIGN_EXTEND ||
2315 // Avoid infinite looping with PromoteIntBinOp.
2316 (N0.getOpcode() == ISD::ANY_EXTEND &&
2317 (!LegalTypes || TLI.isTypeDesirableForOp(N->getOpcode(), Op0VT))) ||
2318 (N0.getOpcode() == ISD::TRUNCATE &&
2319 (!TLI.isZExtFree(VT, Op0VT) ||
2320 !TLI.isTruncateFree(Op0VT, VT)) &&
2321 TLI.isTypeLegal(Op0VT))) &&
2323 Op0VT == N1.getOperand(0).getValueType() &&
2324 (!LegalOperations || TLI.isOperationLegal(N->getOpcode(), Op0VT))) {
2325 SDValue ORNode = DAG.getNode(N->getOpcode(), N0.getDebugLoc(),
2326 N0.getOperand(0).getValueType(),
2327 N0.getOperand(0), N1.getOperand(0));
2328 AddToWorkList(ORNode.getNode());
2329 return DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT, ORNode);
2332 // For each of OP in SHL/SRL/SRA/AND...
2333 // fold (and (OP x, z), (OP y, z)) -> (OP (and x, y), z)
2334 // fold (or (OP x, z), (OP y, z)) -> (OP (or x, y), z)
2335 // fold (xor (OP x, z), (OP y, z)) -> (OP (xor x, y), z)
2336 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL ||
2337 N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::AND) &&
2338 N0.getOperand(1) == N1.getOperand(1)) {
2339 SDValue ORNode = DAG.getNode(N->getOpcode(), N0.getDebugLoc(),
2340 N0.getOperand(0).getValueType(),
2341 N0.getOperand(0), N1.getOperand(0));
2342 AddToWorkList(ORNode.getNode());
2343 return DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT,
2344 ORNode, N0.getOperand(1));
2347 // Simplify xor/and/or (bitcast(A), bitcast(B)) -> bitcast(op (A,B))
2348 // Only perform this optimization after type legalization and before
2349 // LegalizeVectorOprs. LegalizeVectorOprs promotes vector operations by
2350 // adding bitcasts. For example (xor v4i32) is promoted to (v2i64), and
2351 // we don't want to undo this promotion.
2352 // We also handle SCALAR_TO_VECTOR because xor/or/and operations are cheaper
2354 if ((N0.getOpcode() == ISD::BITCAST ||
2355 N0.getOpcode() == ISD::SCALAR_TO_VECTOR) &&
2356 Level == AfterLegalizeTypes) {
2357 SDValue In0 = N0.getOperand(0);
2358 SDValue In1 = N1.getOperand(0);
2359 EVT In0Ty = In0.getValueType();
2360 EVT In1Ty = In1.getValueType();
2361 DebugLoc DL = N->getDebugLoc();
2362 // If both incoming values are integers, and the original types are the
2364 if (In0Ty.isInteger() && In1Ty.isInteger() && In0Ty == In1Ty) {
2365 SDValue Op = DAG.getNode(N->getOpcode(), DL, In0Ty, In0, In1);
2366 SDValue BC = DAG.getNode(N0.getOpcode(), DL, VT, Op);
2367 AddToWorkList(Op.getNode());
2372 // Xor/and/or are indifferent to the swizzle operation (shuffle of one value).
2373 // Simplify xor/and/or (shuff(A), shuff(B)) -> shuff(op (A,B))
2374 // If both shuffles use the same mask, and both shuffle within a single
2375 // vector, then it is worthwhile to move the swizzle after the operation.
2376 // The type-legalizer generates this pattern when loading illegal
2377 // vector types from memory. In many cases this allows additional shuffle
2379 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
2380 N0.getOperand(1).getOpcode() == ISD::UNDEF &&
2381 N1.getOperand(1).getOpcode() == ISD::UNDEF) {
2382 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(N0);
2383 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(N1);
2385 assert(N0.getOperand(0).getValueType() == N1.getOperand(1).getValueType() &&
2386 "Inputs to shuffles are not the same type");
2388 unsigned NumElts = VT.getVectorNumElements();
2390 // Check that both shuffles use the same mask. The masks are known to be of
2391 // the same length because the result vector type is the same.
2392 bool SameMask = true;
2393 for (unsigned i = 0; i != NumElts; ++i) {
2394 int Idx0 = SVN0->getMaskElt(i);
2395 int Idx1 = SVN1->getMaskElt(i);
2403 SDValue Op = DAG.getNode(N->getOpcode(), N->getDebugLoc(), VT,
2404 N0.getOperand(0), N1.getOperand(0));
2405 AddToWorkList(Op.getNode());
2406 return DAG.getVectorShuffle(VT, N->getDebugLoc(), Op,
2407 DAG.getUNDEF(VT), &SVN0->getMask()[0]);
2414 SDValue DAGCombiner::visitAND(SDNode *N) {
2415 SDValue N0 = N->getOperand(0);
2416 SDValue N1 = N->getOperand(1);
2417 SDValue LL, LR, RL, RR, CC0, CC1;
2418 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2419 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2420 EVT VT = N1.getValueType();
2421 unsigned BitWidth = VT.getScalarType().getSizeInBits();
2424 if (VT.isVector()) {
2425 SDValue FoldedVOp = SimplifyVBinOp(N);
2426 if (FoldedVOp.getNode()) return FoldedVOp;
2429 // fold (and x, undef) -> 0
2430 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2431 return DAG.getConstant(0, VT);
2432 // fold (and c1, c2) -> c1&c2
2434 return DAG.FoldConstantArithmetic(ISD::AND, VT, N0C, N1C);
2435 // canonicalize constant to RHS
2437 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N1, N0);
2438 // fold (and x, -1) -> x
2439 if (N1C && N1C->isAllOnesValue())
2441 // if (and x, c) is known to be zero, return 0
2442 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
2443 APInt::getAllOnesValue(BitWidth)))
2444 return DAG.getConstant(0, VT);
2446 SDValue RAND = ReassociateOps(ISD::AND, N->getDebugLoc(), N0, N1);
2447 if (RAND.getNode() != 0)
2449 // fold (and (or x, C), D) -> D if (C & D) == D
2450 if (N1C && N0.getOpcode() == ISD::OR)
2451 if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
2452 if ((ORI->getAPIntValue() & N1C->getAPIntValue()) == N1C->getAPIntValue())
2454 // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits.
2455 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
2456 SDValue N0Op0 = N0.getOperand(0);
2457 APInt Mask = ~N1C->getAPIntValue();
2458 Mask = Mask.trunc(N0Op0.getValueSizeInBits());
2459 if (DAG.MaskedValueIsZero(N0Op0, Mask)) {
2460 SDValue Zext = DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(),
2461 N0.getValueType(), N0Op0);
2463 // Replace uses of the AND with uses of the Zero extend node.
2466 // We actually want to replace all uses of the any_extend with the
2467 // zero_extend, to avoid duplicating things. This will later cause this
2468 // AND to be folded.
2469 CombineTo(N0.getNode(), Zext);
2470 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2473 // similarly fold (and (X (load ([non_ext|any_ext|zero_ext] V))), c) ->
2474 // (X (load ([non_ext|zero_ext] V))) if 'and' only clears top bits which must
2475 // already be zero by virtue of the width of the base type of the load.
2477 // the 'X' node here can either be nothing or an extract_vector_elt to catch
2479 if ((N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
2480 N0.getOperand(0).getOpcode() == ISD::LOAD) ||
2481 N0.getOpcode() == ISD::LOAD) {
2482 LoadSDNode *Load = cast<LoadSDNode>( (N0.getOpcode() == ISD::LOAD) ?
2483 N0 : N0.getOperand(0) );
2485 // Get the constant (if applicable) the zero'th operand is being ANDed with.
2486 // This can be a pure constant or a vector splat, in which case we treat the
2487 // vector as a scalar and use the splat value.
2488 APInt Constant = APInt::getNullValue(1);
2489 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
2490 Constant = C->getAPIntValue();
2491 } else if (BuildVectorSDNode *Vector = dyn_cast<BuildVectorSDNode>(N1)) {
2492 APInt SplatValue, SplatUndef;
2493 unsigned SplatBitSize;
2495 bool IsSplat = Vector->isConstantSplat(SplatValue, SplatUndef,
2496 SplatBitSize, HasAnyUndefs);
2498 // Undef bits can contribute to a possible optimisation if set, so
2500 SplatValue |= SplatUndef;
2502 // The splat value may be something like "0x00FFFFFF", which means 0 for
2503 // the first vector value and FF for the rest, repeating. We need a mask
2504 // that will apply equally to all members of the vector, so AND all the
2505 // lanes of the constant together.
2506 EVT VT = Vector->getValueType(0);
2507 unsigned BitWidth = VT.getVectorElementType().getSizeInBits();
2509 // If the splat value has been compressed to a bitlength lower
2510 // than the size of the vector lane, we need to re-expand it to
2512 if (BitWidth > SplatBitSize)
2513 for (SplatValue = SplatValue.zextOrTrunc(BitWidth);
2514 SplatBitSize < BitWidth;
2515 SplatBitSize = SplatBitSize * 2)
2516 SplatValue |= SplatValue.shl(SplatBitSize);
2518 Constant = APInt::getAllOnesValue(BitWidth);
2519 for (unsigned i = 0, n = SplatBitSize/BitWidth; i < n; ++i)
2520 Constant &= SplatValue.lshr(i*BitWidth).zextOrTrunc(BitWidth);
2524 // If we want to change an EXTLOAD to a ZEXTLOAD, ensure a ZEXTLOAD is
2525 // actually legal and isn't going to get expanded, else this is a false
2527 bool CanZextLoadProfitably = TLI.isLoadExtLegal(ISD::ZEXTLOAD,
2528 Load->getMemoryVT());
2530 // Resize the constant to the same size as the original memory access before
2531 // extension. If it is still the AllOnesValue then this AND is completely
2534 Constant.zextOrTrunc(Load->getMemoryVT().getScalarType().getSizeInBits());
2537 switch (Load->getExtensionType()) {
2538 default: B = false; break;
2539 case ISD::EXTLOAD: B = CanZextLoadProfitably; break;
2541 case ISD::NON_EXTLOAD: B = true; break;
2544 if (B && Constant.isAllOnesValue()) {
2545 // If the load type was an EXTLOAD, convert to ZEXTLOAD in order to
2546 // preserve semantics once we get rid of the AND.
2547 SDValue NewLoad(Load, 0);
2548 if (Load->getExtensionType() == ISD::EXTLOAD) {
2549 NewLoad = DAG.getLoad(Load->getAddressingMode(), ISD::ZEXTLOAD,
2550 Load->getValueType(0), Load->getDebugLoc(),
2551 Load->getChain(), Load->getBasePtr(),
2552 Load->getOffset(), Load->getMemoryVT(),
2553 Load->getMemOperand());
2554 // Replace uses of the EXTLOAD with the new ZEXTLOAD.
2555 if (Load->getNumValues() == 3) {
2556 // PRE/POST_INC loads have 3 values.
2557 SDValue To[] = { NewLoad.getValue(0), NewLoad.getValue(1),
2558 NewLoad.getValue(2) };
2559 CombineTo(Load, To, 3, true);
2561 CombineTo(Load, NewLoad.getValue(0), NewLoad.getValue(1));
2565 // Fold the AND away, taking care not to fold to the old load node if we
2567 CombineTo(N, (N0.getNode() == Load) ? NewLoad : N0);
2569 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2572 // fold (and (setcc x), (setcc y)) -> (setcc (and x, y))
2573 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
2574 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
2575 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
2577 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
2578 LL.getValueType().isInteger()) {
2579 // fold (and (seteq X, 0), (seteq Y, 0)) -> (seteq (or X, Y), 0)
2580 if (cast<ConstantSDNode>(LR)->isNullValue() && Op1 == ISD::SETEQ) {
2581 SDValue ORNode = DAG.getNode(ISD::OR, N0.getDebugLoc(),
2582 LR.getValueType(), LL, RL);
2583 AddToWorkList(ORNode.getNode());
2584 return DAG.getSetCC(N->getDebugLoc(), VT, ORNode, LR, Op1);
2586 // fold (and (seteq X, -1), (seteq Y, -1)) -> (seteq (and X, Y), -1)
2587 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETEQ) {
2588 SDValue ANDNode = DAG.getNode(ISD::AND, N0.getDebugLoc(),
2589 LR.getValueType(), LL, RL);
2590 AddToWorkList(ANDNode.getNode());
2591 return DAG.getSetCC(N->getDebugLoc(), VT, ANDNode, LR, Op1);
2593 // fold (and (setgt X, -1), (setgt Y, -1)) -> (setgt (or X, Y), -1)
2594 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETGT) {
2595 SDValue ORNode = DAG.getNode(ISD::OR, N0.getDebugLoc(),
2596 LR.getValueType(), LL, RL);
2597 AddToWorkList(ORNode.getNode());
2598 return DAG.getSetCC(N->getDebugLoc(), VT, ORNode, LR, Op1);
2601 // canonicalize equivalent to ll == rl
2602 if (LL == RR && LR == RL) {
2603 Op1 = ISD::getSetCCSwappedOperands(Op1);
2606 if (LL == RL && LR == RR) {
2607 bool isInteger = LL.getValueType().isInteger();
2608 ISD::CondCode Result = ISD::getSetCCAndOperation(Op0, Op1, isInteger);
2609 if (Result != ISD::SETCC_INVALID &&
2610 (!LegalOperations || TLI.isCondCodeLegal(Result, LL.getValueType())))
2611 return DAG.getSetCC(N->getDebugLoc(), N0.getValueType(),
2616 // Simplify: (and (op x...), (op y...)) -> (op (and x, y))
2617 if (N0.getOpcode() == N1.getOpcode()) {
2618 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
2619 if (Tmp.getNode()) return Tmp;
2622 // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1)
2623 // fold (and (sra)) -> (and (srl)) when possible.
2624 if (!VT.isVector() &&
2625 SimplifyDemandedBits(SDValue(N, 0)))
2626 return SDValue(N, 0);
2628 // fold (zext_inreg (extload x)) -> (zextload x)
2629 if (ISD::isEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode())) {
2630 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2631 EVT MemVT = LN0->getMemoryVT();
2632 // If we zero all the possible extended bits, then we can turn this into
2633 // a zextload if we are running before legalize or the operation is legal.
2634 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
2635 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
2636 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
2637 ((!LegalOperations && !LN0->isVolatile()) ||
2638 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
2639 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N0.getDebugLoc(), VT,
2640 LN0->getChain(), LN0->getBasePtr(),
2641 LN0->getPointerInfo(), MemVT,
2642 LN0->isVolatile(), LN0->isNonTemporal(),
2643 LN0->getAlignment());
2645 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
2646 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2649 // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use
2650 if (ISD::isSEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
2652 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2653 EVT MemVT = LN0->getMemoryVT();
2654 // If we zero all the possible extended bits, then we can turn this into
2655 // a zextload if we are running before legalize or the operation is legal.
2656 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
2657 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
2658 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
2659 ((!LegalOperations && !LN0->isVolatile()) ||
2660 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
2661 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N0.getDebugLoc(), VT,
2663 LN0->getBasePtr(), LN0->getPointerInfo(),
2665 LN0->isVolatile(), LN0->isNonTemporal(),
2666 LN0->getAlignment());
2668 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
2669 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2673 // fold (and (load x), 255) -> (zextload x, i8)
2674 // fold (and (extload x, i16), 255) -> (zextload x, i8)
2675 // fold (and (any_ext (extload x, i16)), 255) -> (zextload x, i8)
2676 if (N1C && (N0.getOpcode() == ISD::LOAD ||
2677 (N0.getOpcode() == ISD::ANY_EXTEND &&
2678 N0.getOperand(0).getOpcode() == ISD::LOAD))) {
2679 bool HasAnyExt = N0.getOpcode() == ISD::ANY_EXTEND;
2680 LoadSDNode *LN0 = HasAnyExt
2681 ? cast<LoadSDNode>(N0.getOperand(0))
2682 : cast<LoadSDNode>(N0);
2683 if (LN0->getExtensionType() != ISD::SEXTLOAD &&
2684 LN0->isUnindexed() && N0.hasOneUse() && LN0->hasOneUse()) {
2685 uint32_t ActiveBits = N1C->getAPIntValue().getActiveBits();
2686 if (ActiveBits > 0 && APIntOps::isMask(ActiveBits, N1C->getAPIntValue())){
2687 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits);
2688 EVT LoadedVT = LN0->getMemoryVT();
2690 if (ExtVT == LoadedVT &&
2691 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) {
2692 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
2695 DAG.getExtLoad(ISD::ZEXTLOAD, LN0->getDebugLoc(), LoadResultTy,
2696 LN0->getChain(), LN0->getBasePtr(),
2697 LN0->getPointerInfo(),
2698 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
2699 LN0->getAlignment());
2701 CombineTo(LN0, NewLoad, NewLoad.getValue(1));
2702 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2705 // Do not change the width of a volatile load.
2706 // Do not generate loads of non-round integer types since these can
2707 // be expensive (and would be wrong if the type is not byte sized).
2708 if (!LN0->isVolatile() && LoadedVT.bitsGT(ExtVT) && ExtVT.isRound() &&
2709 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) {
2710 EVT PtrType = LN0->getOperand(1).getValueType();
2712 unsigned Alignment = LN0->getAlignment();
2713 SDValue NewPtr = LN0->getBasePtr();
2715 // For big endian targets, we need to add an offset to the pointer
2716 // to load the correct bytes. For little endian systems, we merely
2717 // need to read fewer bytes from the same pointer.
2718 if (TLI.isBigEndian()) {
2719 unsigned LVTStoreBytes = LoadedVT.getStoreSize();
2720 unsigned EVTStoreBytes = ExtVT.getStoreSize();
2721 unsigned PtrOff = LVTStoreBytes - EVTStoreBytes;
2722 NewPtr = DAG.getNode(ISD::ADD, LN0->getDebugLoc(), PtrType,
2723 NewPtr, DAG.getConstant(PtrOff, PtrType));
2724 Alignment = MinAlign(Alignment, PtrOff);
2727 AddToWorkList(NewPtr.getNode());
2729 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
2731 DAG.getExtLoad(ISD::ZEXTLOAD, LN0->getDebugLoc(), LoadResultTy,
2732 LN0->getChain(), NewPtr,
2733 LN0->getPointerInfo(),
2734 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
2737 CombineTo(LN0, Load, Load.getValue(1));
2738 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2744 if (N0.getOpcode() == ISD::ADD && N1.getOpcode() == ISD::SRL &&
2745 VT.getSizeInBits() <= 64) {
2746 if (ConstantSDNode *ADDI = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
2747 APInt ADDC = ADDI->getAPIntValue();
2748 if (!TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
2749 // Look for (and (add x, c1), (lshr y, c2)). If C1 wasn't a legal
2750 // immediate for an add, but it is legal if its top c2 bits are set,
2751 // transform the ADD so the immediate doesn't need to be materialized
2753 if (ConstantSDNode *SRLI = dyn_cast<ConstantSDNode>(N1.getOperand(1))) {
2754 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(),
2755 SRLI->getZExtValue());
2756 if (DAG.MaskedValueIsZero(N0.getOperand(1), Mask)) {
2758 if (TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
2760 DAG.getNode(ISD::ADD, N0.getDebugLoc(), VT,
2761 N0.getOperand(0), DAG.getConstant(ADDC, VT));
2762 CombineTo(N0.getNode(), NewAdd);
2763 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2775 /// MatchBSwapHWord - Match (a >> 8) | (a << 8) as (bswap a) >> 16
2777 SDValue DAGCombiner::MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
2778 bool DemandHighBits) {
2779 if (!LegalOperations)
2782 EVT VT = N->getValueType(0);
2783 if (VT != MVT::i64 && VT != MVT::i32 && VT != MVT::i16)
2785 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
2788 // Recognize (and (shl a, 8), 0xff), (and (srl a, 8), 0xff00)
2789 bool LookPassAnd0 = false;
2790 bool LookPassAnd1 = false;
2791 if (N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode() == ISD::SRL)
2793 if (N1.getOpcode() == ISD::AND && N1.getOperand(0).getOpcode() == ISD::SHL)
2795 if (N0.getOpcode() == ISD::AND) {
2796 if (!N0.getNode()->hasOneUse())
2798 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
2799 if (!N01C || N01C->getZExtValue() != 0xFF00)
2801 N0 = N0.getOperand(0);
2802 LookPassAnd0 = true;
2805 if (N1.getOpcode() == ISD::AND) {
2806 if (!N1.getNode()->hasOneUse())
2808 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
2809 if (!N11C || N11C->getZExtValue() != 0xFF)
2811 N1 = N1.getOperand(0);
2812 LookPassAnd1 = true;
2815 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
2817 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
2819 if (!N0.getNode()->hasOneUse() ||
2820 !N1.getNode()->hasOneUse())
2823 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
2824 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
2827 if (N01C->getZExtValue() != 8 || N11C->getZExtValue() != 8)
2830 // Look for (shl (and a, 0xff), 8), (srl (and a, 0xff00), 8)
2831 SDValue N00 = N0->getOperand(0);
2832 if (!LookPassAnd0 && N00.getOpcode() == ISD::AND) {
2833 if (!N00.getNode()->hasOneUse())
2835 ConstantSDNode *N001C = dyn_cast<ConstantSDNode>(N00.getOperand(1));
2836 if (!N001C || N001C->getZExtValue() != 0xFF)
2838 N00 = N00.getOperand(0);
2839 LookPassAnd0 = true;
2842 SDValue N10 = N1->getOperand(0);
2843 if (!LookPassAnd1 && N10.getOpcode() == ISD::AND) {
2844 if (!N10.getNode()->hasOneUse())
2846 ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N10.getOperand(1));
2847 if (!N101C || N101C->getZExtValue() != 0xFF00)
2849 N10 = N10.getOperand(0);
2850 LookPassAnd1 = true;
2856 // Make sure everything beyond the low halfword is zero since the SRL 16
2857 // will clear the top bits.
2858 unsigned OpSizeInBits = VT.getSizeInBits();
2859 if (DemandHighBits && OpSizeInBits > 16 &&
2860 (!LookPassAnd0 || !LookPassAnd1) &&
2861 !DAG.MaskedValueIsZero(N10, APInt::getHighBitsSet(OpSizeInBits, 16)))
2864 SDValue Res = DAG.getNode(ISD::BSWAP, N->getDebugLoc(), VT, N00);
2865 if (OpSizeInBits > 16)
2866 Res = DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, Res,
2867 DAG.getConstant(OpSizeInBits-16, getShiftAmountTy(VT)));
2871 /// isBSwapHWordElement - Return true if the specified node is an element
2872 /// that makes up a 32-bit packed halfword byteswap. i.e.
2873 /// ((x&0xff)<<8)|((x&0xff00)>>8)|((x&0x00ff0000)<<8)|((x&0xff000000)>>8)
2874 static bool isBSwapHWordElement(SDValue N, SmallVector<SDNode*,4> &Parts) {
2875 if (!N.getNode()->hasOneUse())
2878 unsigned Opc = N.getOpcode();
2879 if (Opc != ISD::AND && Opc != ISD::SHL && Opc != ISD::SRL)
2882 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N.getOperand(1));
2887 switch (N1C->getZExtValue()) {
2890 case 0xFF: Num = 0; break;
2891 case 0xFF00: Num = 1; break;
2892 case 0xFF0000: Num = 2; break;
2893 case 0xFF000000: Num = 3; break;
2896 // Look for (x & 0xff) << 8 as well as ((x << 8) & 0xff00).
2897 SDValue N0 = N.getOperand(0);
2898 if (Opc == ISD::AND) {
2899 if (Num == 0 || Num == 2) {
2901 // (x >> 8) & 0xff0000
2902 if (N0.getOpcode() != ISD::SRL)
2904 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
2905 if (!C || C->getZExtValue() != 8)
2908 // (x << 8) & 0xff00
2909 // (x << 8) & 0xff000000
2910 if (N0.getOpcode() != ISD::SHL)
2912 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
2913 if (!C || C->getZExtValue() != 8)
2916 } else if (Opc == ISD::SHL) {
2918 // (x & 0xff0000) << 8
2919 if (Num != 0 && Num != 2)
2921 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
2922 if (!C || C->getZExtValue() != 8)
2924 } else { // Opc == ISD::SRL
2925 // (x & 0xff00) >> 8
2926 // (x & 0xff000000) >> 8
2927 if (Num != 1 && Num != 3)
2929 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
2930 if (!C || C->getZExtValue() != 8)
2937 Parts[Num] = N0.getOperand(0).getNode();
2941 /// MatchBSwapHWord - Match a 32-bit packed halfword bswap. That is
2942 /// ((x&0xff)<<8)|((x&0xff00)>>8)|((x&0x00ff0000)<<8)|((x&0xff000000)>>8)
2943 /// => (rotl (bswap x), 16)
2944 SDValue DAGCombiner::MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1) {
2945 if (!LegalOperations)
2948 EVT VT = N->getValueType(0);
2951 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
2954 SmallVector<SDNode*,4> Parts(4, (SDNode*)0);
2956 // (or (or (and), (and)), (or (and), (and)))
2957 // (or (or (or (and), (and)), (and)), (and))
2958 if (N0.getOpcode() != ISD::OR)
2960 SDValue N00 = N0.getOperand(0);
2961 SDValue N01 = N0.getOperand(1);
2963 if (N1.getOpcode() == ISD::OR) {
2964 // (or (or (and), (and)), (or (and), (and)))
2965 SDValue N000 = N00.getOperand(0);
2966 if (!isBSwapHWordElement(N000, Parts))
2969 SDValue N001 = N00.getOperand(1);
2970 if (!isBSwapHWordElement(N001, Parts))
2972 SDValue N010 = N01.getOperand(0);
2973 if (!isBSwapHWordElement(N010, Parts))
2975 SDValue N011 = N01.getOperand(1);
2976 if (!isBSwapHWordElement(N011, Parts))
2979 // (or (or (or (and), (and)), (and)), (and))
2980 if (!isBSwapHWordElement(N1, Parts))
2982 if (!isBSwapHWordElement(N01, Parts))
2984 if (N00.getOpcode() != ISD::OR)
2986 SDValue N000 = N00.getOperand(0);
2987 if (!isBSwapHWordElement(N000, Parts))
2989 SDValue N001 = N00.getOperand(1);
2990 if (!isBSwapHWordElement(N001, Parts))
2994 // Make sure the parts are all coming from the same node.
2995 if (Parts[0] != Parts[1] || Parts[0] != Parts[2] || Parts[0] != Parts[3])
2998 SDValue BSwap = DAG.getNode(ISD::BSWAP, N->getDebugLoc(), VT,
2999 SDValue(Parts[0],0));
3001 // Result of the bswap should be rotated by 16. If it's not legal, than
3002 // do (x << 16) | (x >> 16).
3003 SDValue ShAmt = DAG.getConstant(16, getShiftAmountTy(VT));
3004 if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT))
3005 return DAG.getNode(ISD::ROTL, N->getDebugLoc(), VT, BSwap, ShAmt);
3006 else if (TLI.isOperationLegalOrCustom(ISD::ROTR, VT))
3007 return DAG.getNode(ISD::ROTR, N->getDebugLoc(), VT, BSwap, ShAmt);
3008 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT,
3009 DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, BSwap, ShAmt),
3010 DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, BSwap, ShAmt));
3013 SDValue DAGCombiner::visitOR(SDNode *N) {
3014 SDValue N0 = N->getOperand(0);
3015 SDValue N1 = N->getOperand(1);
3016 SDValue LL, LR, RL, RR, CC0, CC1;
3017 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3018 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3019 EVT VT = N1.getValueType();
3022 if (VT.isVector()) {
3023 SDValue FoldedVOp = SimplifyVBinOp(N);
3024 if (FoldedVOp.getNode()) return FoldedVOp;
3027 // fold (or x, undef) -> -1
3028 if (!LegalOperations &&
3029 (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)) {
3030 EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT;
3031 return DAG.getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()), VT);
3033 // fold (or c1, c2) -> c1|c2
3035 return DAG.FoldConstantArithmetic(ISD::OR, VT, N0C, N1C);
3036 // canonicalize constant to RHS
3038 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N1, N0);
3039 // fold (or x, 0) -> x
3040 if (N1C && N1C->isNullValue())
3042 // fold (or x, -1) -> -1
3043 if (N1C && N1C->isAllOnesValue())
3045 // fold (or x, c) -> c iff (x & ~c) == 0
3046 if (N1C && DAG.MaskedValueIsZero(N0, ~N1C->getAPIntValue()))
3049 // Recognize halfword bswaps as (bswap + rotl 16) or (bswap + shl 16)
3050 SDValue BSwap = MatchBSwapHWord(N, N0, N1);
3051 if (BSwap.getNode() != 0)
3053 BSwap = MatchBSwapHWordLow(N, N0, N1);
3054 if (BSwap.getNode() != 0)
3058 SDValue ROR = ReassociateOps(ISD::OR, N->getDebugLoc(), N0, N1);
3059 if (ROR.getNode() != 0)
3061 // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2)
3062 // iff (c1 & c2) == 0.
3063 if (N1C && N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
3064 isa<ConstantSDNode>(N0.getOperand(1))) {
3065 ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1));
3066 if ((C1->getAPIntValue() & N1C->getAPIntValue()) != 0)
3067 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
3068 DAG.getNode(ISD::OR, N0.getDebugLoc(), VT,
3069 N0.getOperand(0), N1),
3070 DAG.FoldConstantArithmetic(ISD::OR, VT, N1C, C1));
3072 // fold (or (setcc x), (setcc y)) -> (setcc (or x, y))
3073 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
3074 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
3075 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
3077 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
3078 LL.getValueType().isInteger()) {
3079 // fold (or (setne X, 0), (setne Y, 0)) -> (setne (or X, Y), 0)
3080 // fold (or (setlt X, 0), (setlt Y, 0)) -> (setne (or X, Y), 0)
3081 if (cast<ConstantSDNode>(LR)->isNullValue() &&
3082 (Op1 == ISD::SETNE || Op1 == ISD::SETLT)) {
3083 SDValue ORNode = DAG.getNode(ISD::OR, LR.getDebugLoc(),
3084 LR.getValueType(), LL, RL);
3085 AddToWorkList(ORNode.getNode());
3086 return DAG.getSetCC(N->getDebugLoc(), VT, ORNode, LR, Op1);
3088 // fold (or (setne X, -1), (setne Y, -1)) -> (setne (and X, Y), -1)
3089 // fold (or (setgt X, -1), (setgt Y -1)) -> (setgt (and X, Y), -1)
3090 if (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
3091 (Op1 == ISD::SETNE || Op1 == ISD::SETGT)) {
3092 SDValue ANDNode = DAG.getNode(ISD::AND, LR.getDebugLoc(),
3093 LR.getValueType(), LL, RL);
3094 AddToWorkList(ANDNode.getNode());
3095 return DAG.getSetCC(N->getDebugLoc(), VT, ANDNode, LR, Op1);
3098 // canonicalize equivalent to ll == rl
3099 if (LL == RR && LR == RL) {
3100 Op1 = ISD::getSetCCSwappedOperands(Op1);
3103 if (LL == RL && LR == RR) {
3104 bool isInteger = LL.getValueType().isInteger();
3105 ISD::CondCode Result = ISD::getSetCCOrOperation(Op0, Op1, isInteger);
3106 if (Result != ISD::SETCC_INVALID &&
3107 (!LegalOperations || TLI.isCondCodeLegal(Result, LL.getValueType())))
3108 return DAG.getSetCC(N->getDebugLoc(), N0.getValueType(),
3113 // Simplify: (or (op x...), (op y...)) -> (op (or x, y))
3114 if (N0.getOpcode() == N1.getOpcode()) {
3115 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3116 if (Tmp.getNode()) return Tmp;
3119 // (or (and X, C1), (and Y, C2)) -> (and (or X, Y), C3) if possible.
3120 if (N0.getOpcode() == ISD::AND &&
3121 N1.getOpcode() == ISD::AND &&
3122 N0.getOperand(1).getOpcode() == ISD::Constant &&
3123 N1.getOperand(1).getOpcode() == ISD::Constant &&
3124 // Don't increase # computations.
3125 (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) {
3126 // We can only do this xform if we know that bits from X that are set in C2
3127 // but not in C1 are already zero. Likewise for Y.
3128 const APInt &LHSMask =
3129 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
3130 const APInt &RHSMask =
3131 cast<ConstantSDNode>(N1.getOperand(1))->getAPIntValue();
3133 if (DAG.MaskedValueIsZero(N0.getOperand(0), RHSMask&~LHSMask) &&
3134 DAG.MaskedValueIsZero(N1.getOperand(0), LHSMask&~RHSMask)) {
3135 SDValue X = DAG.getNode(ISD::OR, N0.getDebugLoc(), VT,
3136 N0.getOperand(0), N1.getOperand(0));
3137 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, X,
3138 DAG.getConstant(LHSMask | RHSMask, VT));
3142 // See if this is some rotate idiom.
3143 if (SDNode *Rot = MatchRotate(N0, N1, N->getDebugLoc()))
3144 return SDValue(Rot, 0);
3146 // Simplify the operands using demanded-bits information.
3147 if (!VT.isVector() &&
3148 SimplifyDemandedBits(SDValue(N, 0)))
3149 return SDValue(N, 0);
3154 /// MatchRotateHalf - Match "(X shl/srl V1) & V2" where V2 may not be present.
3155 static bool MatchRotateHalf(SDValue Op, SDValue &Shift, SDValue &Mask) {
3156 if (Op.getOpcode() == ISD::AND) {
3157 if (isa<ConstantSDNode>(Op.getOperand(1))) {
3158 Mask = Op.getOperand(1);
3159 Op = Op.getOperand(0);
3165 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SHL) {
3173 // MatchRotate - Handle an 'or' of two operands. If this is one of the many
3174 // idioms for rotate, and if the target supports rotation instructions, generate
3176 SDNode *DAGCombiner::MatchRotate(SDValue LHS, SDValue RHS, DebugLoc DL) {
3177 // Must be a legal type. Expanded 'n promoted things won't work with rotates.
3178 EVT VT = LHS.getValueType();
3179 if (!TLI.isTypeLegal(VT)) return 0;
3181 // The target must have at least one rotate flavor.
3182 bool HasROTL = TLI.isOperationLegalOrCustom(ISD::ROTL, VT);
3183 bool HasROTR = TLI.isOperationLegalOrCustom(ISD::ROTR, VT);
3184 if (!HasROTL && !HasROTR) return 0;
3186 // Match "(X shl/srl V1) & V2" where V2 may not be present.
3187 SDValue LHSShift; // The shift.
3188 SDValue LHSMask; // AND value if any.
3189 if (!MatchRotateHalf(LHS, LHSShift, LHSMask))
3190 return 0; // Not part of a rotate.
3192 SDValue RHSShift; // The shift.
3193 SDValue RHSMask; // AND value if any.
3194 if (!MatchRotateHalf(RHS, RHSShift, RHSMask))
3195 return 0; // Not part of a rotate.
3197 if (LHSShift.getOperand(0) != RHSShift.getOperand(0))
3198 return 0; // Not shifting the same value.
3200 if (LHSShift.getOpcode() == RHSShift.getOpcode())
3201 return 0; // Shifts must disagree.
3203 // Canonicalize shl to left side in a shl/srl pair.
3204 if (RHSShift.getOpcode() == ISD::SHL) {
3205 std::swap(LHS, RHS);
3206 std::swap(LHSShift, RHSShift);
3207 std::swap(LHSMask , RHSMask );
3210 unsigned OpSizeInBits = VT.getSizeInBits();
3211 SDValue LHSShiftArg = LHSShift.getOperand(0);
3212 SDValue LHSShiftAmt = LHSShift.getOperand(1);
3213 SDValue RHSShiftAmt = RHSShift.getOperand(1);
3215 // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1)
3216 // fold (or (shl x, C1), (srl x, C2)) -> (rotr x, C2)
3217 if (LHSShiftAmt.getOpcode() == ISD::Constant &&
3218 RHSShiftAmt.getOpcode() == ISD::Constant) {
3219 uint64_t LShVal = cast<ConstantSDNode>(LHSShiftAmt)->getZExtValue();
3220 uint64_t RShVal = cast<ConstantSDNode>(RHSShiftAmt)->getZExtValue();
3221 if ((LShVal + RShVal) != OpSizeInBits)
3226 Rot = DAG.getNode(ISD::ROTL, DL, VT, LHSShiftArg, LHSShiftAmt);
3228 Rot = DAG.getNode(ISD::ROTR, DL, VT, LHSShiftArg, RHSShiftAmt);
3230 // If there is an AND of either shifted operand, apply it to the result.
3231 if (LHSMask.getNode() || RHSMask.getNode()) {
3232 APInt Mask = APInt::getAllOnesValue(OpSizeInBits);
3234 if (LHSMask.getNode()) {
3235 APInt RHSBits = APInt::getLowBitsSet(OpSizeInBits, LShVal);
3236 Mask &= cast<ConstantSDNode>(LHSMask)->getAPIntValue() | RHSBits;
3238 if (RHSMask.getNode()) {
3239 APInt LHSBits = APInt::getHighBitsSet(OpSizeInBits, RShVal);
3240 Mask &= cast<ConstantSDNode>(RHSMask)->getAPIntValue() | LHSBits;
3243 Rot = DAG.getNode(ISD::AND, DL, VT, Rot, DAG.getConstant(Mask, VT));
3246 return Rot.getNode();
3249 // If there is a mask here, and we have a variable shift, we can't be sure
3250 // that we're masking out the right stuff.
3251 if (LHSMask.getNode() || RHSMask.getNode())
3254 // fold (or (shl x, y), (srl x, (sub 32, y))) -> (rotl x, y)
3255 // fold (or (shl x, y), (srl x, (sub 32, y))) -> (rotr x, (sub 32, y))
3256 if (RHSShiftAmt.getOpcode() == ISD::SUB &&
3257 LHSShiftAmt == RHSShiftAmt.getOperand(1)) {
3258 if (ConstantSDNode *SUBC =
3259 dyn_cast<ConstantSDNode>(RHSShiftAmt.getOperand(0))) {
3260 if (SUBC->getAPIntValue() == OpSizeInBits) {
3262 return DAG.getNode(ISD::ROTL, DL, VT,
3263 LHSShiftArg, LHSShiftAmt).getNode();
3265 return DAG.getNode(ISD::ROTR, DL, VT,
3266 LHSShiftArg, RHSShiftAmt).getNode();
3271 // fold (or (shl x, (sub 32, y)), (srl x, r)) -> (rotr x, y)
3272 // fold (or (shl x, (sub 32, y)), (srl x, r)) -> (rotl x, (sub 32, y))
3273 if (LHSShiftAmt.getOpcode() == ISD::SUB &&
3274 RHSShiftAmt == LHSShiftAmt.getOperand(1)) {
3275 if (ConstantSDNode *SUBC =
3276 dyn_cast<ConstantSDNode>(LHSShiftAmt.getOperand(0))) {
3277 if (SUBC->getAPIntValue() == OpSizeInBits) {
3279 return DAG.getNode(ISD::ROTR, DL, VT,
3280 LHSShiftArg, RHSShiftAmt).getNode();
3282 return DAG.getNode(ISD::ROTL, DL, VT,
3283 LHSShiftArg, LHSShiftAmt).getNode();
3288 // Look for sign/zext/any-extended or truncate cases:
3289 if ((LHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND
3290 || LHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND
3291 || LHSShiftAmt.getOpcode() == ISD::ANY_EXTEND
3292 || LHSShiftAmt.getOpcode() == ISD::TRUNCATE) &&
3293 (RHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND
3294 || RHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND
3295 || RHSShiftAmt.getOpcode() == ISD::ANY_EXTEND
3296 || RHSShiftAmt.getOpcode() == ISD::TRUNCATE)) {
3297 SDValue LExtOp0 = LHSShiftAmt.getOperand(0);
3298 SDValue RExtOp0 = RHSShiftAmt.getOperand(0);
3299 if (RExtOp0.getOpcode() == ISD::SUB &&
3300 RExtOp0.getOperand(1) == LExtOp0) {
3301 // fold (or (shl x, (*ext y)), (srl x, (*ext (sub 32, y)))) ->
3303 // fold (or (shl x, (*ext y)), (srl x, (*ext (sub 32, y)))) ->
3304 // (rotr x, (sub 32, y))
3305 if (ConstantSDNode *SUBC =
3306 dyn_cast<ConstantSDNode>(RExtOp0.getOperand(0))) {
3307 if (SUBC->getAPIntValue() == OpSizeInBits) {
3308 return DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT,
3310 HasROTL ? LHSShiftAmt : RHSShiftAmt).getNode();
3313 } else if (LExtOp0.getOpcode() == ISD::SUB &&
3314 RExtOp0 == LExtOp0.getOperand(1)) {
3315 // fold (or (shl x, (*ext (sub 32, y))), (srl x, (*ext y))) ->
3317 // fold (or (shl x, (*ext (sub 32, y))), (srl x, (*ext y))) ->
3318 // (rotl x, (sub 32, y))
3319 if (ConstantSDNode *SUBC =
3320 dyn_cast<ConstantSDNode>(LExtOp0.getOperand(0))) {
3321 if (SUBC->getAPIntValue() == OpSizeInBits) {
3322 return DAG.getNode(HasROTR ? ISD::ROTR : ISD::ROTL, DL, VT,
3324 HasROTR ? RHSShiftAmt : LHSShiftAmt).getNode();
3333 SDValue DAGCombiner::visitXOR(SDNode *N) {
3334 SDValue N0 = N->getOperand(0);
3335 SDValue N1 = N->getOperand(1);
3336 SDValue LHS, RHS, CC;
3337 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3338 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3339 EVT VT = N0.getValueType();
3342 if (VT.isVector()) {
3343 SDValue FoldedVOp = SimplifyVBinOp(N);
3344 if (FoldedVOp.getNode()) return FoldedVOp;
3347 // fold (xor undef, undef) -> 0. This is a common idiom (misuse).
3348 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
3349 return DAG.getConstant(0, VT);
3350 // fold (xor x, undef) -> undef
3351 if (N0.getOpcode() == ISD::UNDEF)
3353 if (N1.getOpcode() == ISD::UNDEF)
3355 // fold (xor c1, c2) -> c1^c2
3357 return DAG.FoldConstantArithmetic(ISD::XOR, VT, N0C, N1C);
3358 // canonicalize constant to RHS
3360 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N1, N0);
3361 // fold (xor x, 0) -> x
3362 if (N1C && N1C->isNullValue())
3365 SDValue RXOR = ReassociateOps(ISD::XOR, N->getDebugLoc(), N0, N1);
3366 if (RXOR.getNode() != 0)
3369 // fold !(x cc y) -> (x !cc y)
3370 if (N1C && N1C->getAPIntValue() == 1 && isSetCCEquivalent(N0, LHS, RHS, CC)) {
3371 bool isInt = LHS.getValueType().isInteger();
3372 ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
3375 if (!LegalOperations || TLI.isCondCodeLegal(NotCC, LHS.getValueType())) {
3376 switch (N0.getOpcode()) {
3378 llvm_unreachable("Unhandled SetCC Equivalent!");
3380 return DAG.getSetCC(N->getDebugLoc(), VT, LHS, RHS, NotCC);
3381 case ISD::SELECT_CC:
3382 return DAG.getSelectCC(N->getDebugLoc(), LHS, RHS, N0.getOperand(2),
3383 N0.getOperand(3), NotCC);
3388 // fold (not (zext (setcc x, y))) -> (zext (not (setcc x, y)))
3389 if (N1C && N1C->getAPIntValue() == 1 && N0.getOpcode() == ISD::ZERO_EXTEND &&
3390 N0.getNode()->hasOneUse() &&
3391 isSetCCEquivalent(N0.getOperand(0), LHS, RHS, CC)){
3392 SDValue V = N0.getOperand(0);
3393 V = DAG.getNode(ISD::XOR, N0.getDebugLoc(), V.getValueType(), V,
3394 DAG.getConstant(1, V.getValueType()));
3395 AddToWorkList(V.getNode());
3396 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, V);
3399 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are setcc
3400 if (N1C && N1C->getAPIntValue() == 1 && VT == MVT::i1 &&
3401 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3402 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3403 if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) {
3404 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3405 LHS = DAG.getNode(ISD::XOR, LHS.getDebugLoc(), VT, LHS, N1); // LHS = ~LHS
3406 RHS = DAG.getNode(ISD::XOR, RHS.getDebugLoc(), VT, RHS, N1); // RHS = ~RHS
3407 AddToWorkList(LHS.getNode()); AddToWorkList(RHS.getNode());
3408 return DAG.getNode(NewOpcode, N->getDebugLoc(), VT, LHS, RHS);
3411 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are constants
3412 if (N1C && N1C->isAllOnesValue() &&
3413 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3414 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3415 if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) {
3416 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3417 LHS = DAG.getNode(ISD::XOR, LHS.getDebugLoc(), VT, LHS, N1); // LHS = ~LHS
3418 RHS = DAG.getNode(ISD::XOR, RHS.getDebugLoc(), VT, RHS, N1); // RHS = ~RHS
3419 AddToWorkList(LHS.getNode()); AddToWorkList(RHS.getNode());
3420 return DAG.getNode(NewOpcode, N->getDebugLoc(), VT, LHS, RHS);
3423 // fold (xor (xor x, c1), c2) -> (xor x, (xor c1, c2))
3424 if (N1C && N0.getOpcode() == ISD::XOR) {
3425 ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0));
3426 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3428 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N0.getOperand(1),
3429 DAG.getConstant(N1C->getAPIntValue() ^
3430 N00C->getAPIntValue(), VT));
3432 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N0.getOperand(0),
3433 DAG.getConstant(N1C->getAPIntValue() ^
3434 N01C->getAPIntValue(), VT));
3436 // fold (xor x, x) -> 0
3438 return tryFoldToZero(N->getDebugLoc(), TLI, VT, DAG, LegalOperations);
3440 // Simplify: xor (op x...), (op y...) -> (op (xor x, y))
3441 if (N0.getOpcode() == N1.getOpcode()) {
3442 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3443 if (Tmp.getNode()) return Tmp;
3446 // Simplify the expression using non-local knowledge.
3447 if (!VT.isVector() &&
3448 SimplifyDemandedBits(SDValue(N, 0)))
3449 return SDValue(N, 0);
3454 /// visitShiftByConstant - Handle transforms common to the three shifts, when
3455 /// the shift amount is a constant.
3456 SDValue DAGCombiner::visitShiftByConstant(SDNode *N, unsigned Amt) {
3457 SDNode *LHS = N->getOperand(0).getNode();
3458 if (!LHS->hasOneUse()) return SDValue();
3460 // We want to pull some binops through shifts, so that we have (and (shift))
3461 // instead of (shift (and)), likewise for add, or, xor, etc. This sort of
3462 // thing happens with address calculations, so it's important to canonicalize
3464 bool HighBitSet = false; // Can we transform this if the high bit is set?
3466 switch (LHS->getOpcode()) {
3467 default: return SDValue();
3470 HighBitSet = false; // We can only transform sra if the high bit is clear.
3473 HighBitSet = true; // We can only transform sra if the high bit is set.
3476 if (N->getOpcode() != ISD::SHL)
3477 return SDValue(); // only shl(add) not sr[al](add).
3478 HighBitSet = false; // We can only transform sra if the high bit is clear.
3482 // We require the RHS of the binop to be a constant as well.
3483 ConstantSDNode *BinOpCst = dyn_cast<ConstantSDNode>(LHS->getOperand(1));
3484 if (!BinOpCst) return SDValue();
3486 // FIXME: disable this unless the input to the binop is a shift by a constant.
3487 // If it is not a shift, it pessimizes some common cases like:
3489 // void foo(int *X, int i) { X[i & 1235] = 1; }
3490 // int bar(int *X, int i) { return X[i & 255]; }
3491 SDNode *BinOpLHSVal = LHS->getOperand(0).getNode();
3492 if ((BinOpLHSVal->getOpcode() != ISD::SHL &&
3493 BinOpLHSVal->getOpcode() != ISD::SRA &&
3494 BinOpLHSVal->getOpcode() != ISD::SRL) ||
3495 !isa<ConstantSDNode>(BinOpLHSVal->getOperand(1)))
3498 EVT VT = N->getValueType(0);
3500 // If this is a signed shift right, and the high bit is modified by the
3501 // logical operation, do not perform the transformation. The highBitSet
3502 // boolean indicates the value of the high bit of the constant which would
3503 // cause it to be modified for this operation.
3504 if (N->getOpcode() == ISD::SRA) {
3505 bool BinOpRHSSignSet = BinOpCst->getAPIntValue().isNegative();
3506 if (BinOpRHSSignSet != HighBitSet)
3510 // Fold the constants, shifting the binop RHS by the shift amount.
3511 SDValue NewRHS = DAG.getNode(N->getOpcode(), LHS->getOperand(1).getDebugLoc(),
3513 LHS->getOperand(1), N->getOperand(1));
3515 // Create the new shift.
3516 SDValue NewShift = DAG.getNode(N->getOpcode(),
3517 LHS->getOperand(0).getDebugLoc(),
3518 VT, LHS->getOperand(0), N->getOperand(1));
3520 // Create the new binop.
3521 return DAG.getNode(LHS->getOpcode(), N->getDebugLoc(), VT, NewShift, NewRHS);
3524 SDValue DAGCombiner::visitSHL(SDNode *N) {
3525 SDValue N0 = N->getOperand(0);
3526 SDValue N1 = N->getOperand(1);
3527 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3528 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3529 EVT VT = N0.getValueType();
3530 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
3532 // fold (shl c1, c2) -> c1<<c2
3534 return DAG.FoldConstantArithmetic(ISD::SHL, VT, N0C, N1C);
3535 // fold (shl 0, x) -> 0
3536 if (N0C && N0C->isNullValue())
3538 // fold (shl x, c >= size(x)) -> undef
3539 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
3540 return DAG.getUNDEF(VT);
3541 // fold (shl x, 0) -> x
3542 if (N1C && N1C->isNullValue())
3544 // fold (shl undef, x) -> 0
3545 if (N0.getOpcode() == ISD::UNDEF)
3546 return DAG.getConstant(0, VT);
3547 // if (shl x, c) is known to be zero, return 0
3548 if (DAG.MaskedValueIsZero(SDValue(N, 0),
3549 APInt::getAllOnesValue(OpSizeInBits)))
3550 return DAG.getConstant(0, VT);
3551 // fold (shl x, (trunc (and y, c))) -> (shl x, (and (trunc y), (trunc c))).
3552 if (N1.getOpcode() == ISD::TRUNCATE &&
3553 N1.getOperand(0).getOpcode() == ISD::AND &&
3554 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
3555 SDValue N101 = N1.getOperand(0).getOperand(1);
3556 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
3557 EVT TruncVT = N1.getValueType();
3558 SDValue N100 = N1.getOperand(0).getOperand(0);
3559 APInt TruncC = N101C->getAPIntValue();
3560 TruncC = TruncC.trunc(TruncVT.getSizeInBits());
3561 return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0,
3562 DAG.getNode(ISD::AND, N->getDebugLoc(), TruncVT,
3563 DAG.getNode(ISD::TRUNCATE,
3566 DAG.getConstant(TruncC, TruncVT)));
3570 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
3571 return SDValue(N, 0);
3573 // fold (shl (shl x, c1), c2) -> 0 or (shl x, (add c1, c2))
3574 if (N1C && N0.getOpcode() == ISD::SHL &&
3575 N0.getOperand(1).getOpcode() == ISD::Constant) {
3576 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
3577 uint64_t c2 = N1C->getZExtValue();
3578 if (c1 + c2 >= OpSizeInBits)
3579 return DAG.getConstant(0, VT);
3580 return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0.getOperand(0),
3581 DAG.getConstant(c1 + c2, N1.getValueType()));
3584 // fold (shl (ext (shl x, c1)), c2) -> (ext (shl x, (add c1, c2)))
3585 // For this to be valid, the second form must not preserve any of the bits
3586 // that are shifted out by the inner shift in the first form. This means
3587 // the outer shift size must be >= the number of bits added by the ext.
3588 // As a corollary, we don't care what kind of ext it is.
3589 if (N1C && (N0.getOpcode() == ISD::ZERO_EXTEND ||
3590 N0.getOpcode() == ISD::ANY_EXTEND ||
3591 N0.getOpcode() == ISD::SIGN_EXTEND) &&
3592 N0.getOperand(0).getOpcode() == ISD::SHL &&
3593 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
3595 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
3596 uint64_t c2 = N1C->getZExtValue();
3597 EVT InnerShiftVT = N0.getOperand(0).getValueType();
3598 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
3599 if (c2 >= OpSizeInBits - InnerShiftSize) {
3600 if (c1 + c2 >= OpSizeInBits)
3601 return DAG.getConstant(0, VT);
3602 return DAG.getNode(ISD::SHL, N0->getDebugLoc(), VT,
3603 DAG.getNode(N0.getOpcode(), N0->getDebugLoc(), VT,
3604 N0.getOperand(0)->getOperand(0)),
3605 DAG.getConstant(c1 + c2, N1.getValueType()));
3609 // fold (shl (srl x, c1), c2) -> (and (shl x, (sub c2, c1), MASK) or
3610 // (and (srl x, (sub c1, c2), MASK)
3611 // Only fold this if the inner shift has no other uses -- if it does, folding
3612 // this will increase the total number of instructions.
3613 if (N1C && N0.getOpcode() == ISD::SRL && N0.hasOneUse() &&
3614 N0.getOperand(1).getOpcode() == ISD::Constant) {
3615 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
3616 if (c1 < VT.getSizeInBits()) {
3617 uint64_t c2 = N1C->getZExtValue();
3618 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(),
3619 VT.getSizeInBits() - c1);
3622 Mask = Mask.shl(c2-c1);
3623 Shift = DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0.getOperand(0),
3624 DAG.getConstant(c2-c1, N1.getValueType()));
3626 Mask = Mask.lshr(c1-c2);
3627 Shift = DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0.getOperand(0),
3628 DAG.getConstant(c1-c2, N1.getValueType()));
3630 return DAG.getNode(ISD::AND, N0.getDebugLoc(), VT, Shift,
3631 DAG.getConstant(Mask, VT));
3634 // fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1))
3635 if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1)) {
3636 SDValue HiBitsMask =
3637 DAG.getConstant(APInt::getHighBitsSet(VT.getSizeInBits(),
3638 VT.getSizeInBits() -
3639 N1C->getZExtValue()),
3641 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0.getOperand(0),
3646 SDValue NewSHL = visitShiftByConstant(N, N1C->getZExtValue());
3647 if (NewSHL.getNode())
3654 SDValue DAGCombiner::visitSRA(SDNode *N) {
3655 SDValue N0 = N->getOperand(0);
3656 SDValue N1 = N->getOperand(1);
3657 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3658 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3659 EVT VT = N0.getValueType();
3660 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
3662 // fold (sra c1, c2) -> (sra c1, c2)
3664 return DAG.FoldConstantArithmetic(ISD::SRA, VT, N0C, N1C);
3665 // fold (sra 0, x) -> 0
3666 if (N0C && N0C->isNullValue())
3668 // fold (sra -1, x) -> -1
3669 if (N0C && N0C->isAllOnesValue())
3671 // fold (sra x, (setge c, size(x))) -> undef
3672 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
3673 return DAG.getUNDEF(VT);
3674 // fold (sra x, 0) -> x
3675 if (N1C && N1C->isNullValue())
3677 // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports
3679 if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) {
3680 unsigned LowBits = OpSizeInBits - (unsigned)N1C->getZExtValue();
3681 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), LowBits);
3683 ExtVT = EVT::getVectorVT(*DAG.getContext(),
3684 ExtVT, VT.getVectorNumElements());
3685 if ((!LegalOperations ||
3686 TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, ExtVT)))
3687 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT,
3688 N0.getOperand(0), DAG.getValueType(ExtVT));
3691 // fold (sra (sra x, c1), c2) -> (sra x, (add c1, c2))
3692 if (N1C && N0.getOpcode() == ISD::SRA) {
3693 if (ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
3694 unsigned Sum = N1C->getZExtValue() + C1->getZExtValue();
3695 if (Sum >= OpSizeInBits) Sum = OpSizeInBits-1;
3696 return DAG.getNode(ISD::SRA, N->getDebugLoc(), VT, N0.getOperand(0),
3697 DAG.getConstant(Sum, N1C->getValueType(0)));
3701 // fold (sra (shl X, m), (sub result_size, n))
3702 // -> (sign_extend (trunc (shl X, (sub (sub result_size, n), m)))) for
3703 // result_size - n != m.
3704 // If truncate is free for the target sext(shl) is likely to result in better
3706 if (N0.getOpcode() == ISD::SHL) {
3707 // Get the two constanst of the shifts, CN0 = m, CN = n.
3708 const ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3710 // Determine what the truncate's result bitsize and type would be.
3712 EVT::getIntegerVT(*DAG.getContext(),
3713 OpSizeInBits - N1C->getZExtValue());
3714 // Determine the residual right-shift amount.
3715 signed ShiftAmt = N1C->getZExtValue() - N01C->getZExtValue();
3717 // If the shift is not a no-op (in which case this should be just a sign
3718 // extend already), the truncated to type is legal, sign_extend is legal
3719 // on that type, and the truncate to that type is both legal and free,
3720 // perform the transform.
3721 if ((ShiftAmt > 0) &&
3722 TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND, TruncVT) &&
3723 TLI.isOperationLegalOrCustom(ISD::TRUNCATE, VT) &&
3724 TLI.isTruncateFree(VT, TruncVT)) {
3726 SDValue Amt = DAG.getConstant(ShiftAmt,
3727 getShiftAmountTy(N0.getOperand(0).getValueType()));
3728 SDValue Shift = DAG.getNode(ISD::SRL, N0.getDebugLoc(), VT,
3729 N0.getOperand(0), Amt);
3730 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(), TruncVT,
3732 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(),
3733 N->getValueType(0), Trunc);
3738 // fold (sra x, (trunc (and y, c))) -> (sra x, (and (trunc y), (trunc c))).
3739 if (N1.getOpcode() == ISD::TRUNCATE &&
3740 N1.getOperand(0).getOpcode() == ISD::AND &&
3741 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
3742 SDValue N101 = N1.getOperand(0).getOperand(1);
3743 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
3744 EVT TruncVT = N1.getValueType();
3745 SDValue N100 = N1.getOperand(0).getOperand(0);
3746 APInt TruncC = N101C->getAPIntValue();
3747 TruncC = TruncC.trunc(TruncVT.getScalarType().getSizeInBits());
3748 return DAG.getNode(ISD::SRA, N->getDebugLoc(), VT, N0,
3749 DAG.getNode(ISD::AND, N->getDebugLoc(),
3751 DAG.getNode(ISD::TRUNCATE,
3754 DAG.getConstant(TruncC, TruncVT)));
3758 // fold (sra (trunc (sr x, c1)), c2) -> (trunc (sra x, c1+c2))
3759 // if c1 is equal to the number of bits the trunc removes
3760 if (N0.getOpcode() == ISD::TRUNCATE &&
3761 (N0.getOperand(0).getOpcode() == ISD::SRL ||
3762 N0.getOperand(0).getOpcode() == ISD::SRA) &&
3763 N0.getOperand(0).hasOneUse() &&
3764 N0.getOperand(0).getOperand(1).hasOneUse() &&
3765 N1C && isa<ConstantSDNode>(N0.getOperand(0).getOperand(1))) {
3766 EVT LargeVT = N0.getOperand(0).getValueType();
3767 ConstantSDNode *LargeShiftAmt =
3768 cast<ConstantSDNode>(N0.getOperand(0).getOperand(1));
3770 if (LargeVT.getScalarType().getSizeInBits() - OpSizeInBits ==
3771 LargeShiftAmt->getZExtValue()) {
3773 DAG.getConstant(LargeShiftAmt->getZExtValue() + N1C->getZExtValue(),
3774 getShiftAmountTy(N0.getOperand(0).getOperand(0).getValueType()));
3775 SDValue SRA = DAG.getNode(ISD::SRA, N->getDebugLoc(), LargeVT,
3776 N0.getOperand(0).getOperand(0), Amt);
3777 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, SRA);
3781 // Simplify, based on bits shifted out of the LHS.
3782 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
3783 return SDValue(N, 0);
3786 // If the sign bit is known to be zero, switch this to a SRL.
3787 if (DAG.SignBitIsZero(N0))
3788 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0, N1);
3791 SDValue NewSRA = visitShiftByConstant(N, N1C->getZExtValue());
3792 if (NewSRA.getNode())
3799 SDValue DAGCombiner::visitSRL(SDNode *N) {
3800 SDValue N0 = N->getOperand(0);
3801 SDValue N1 = N->getOperand(1);
3802 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3803 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3804 EVT VT = N0.getValueType();
3805 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
3807 // fold (srl c1, c2) -> c1 >>u c2
3809 return DAG.FoldConstantArithmetic(ISD::SRL, VT, N0C, N1C);
3810 // fold (srl 0, x) -> 0
3811 if (N0C && N0C->isNullValue())
3813 // fold (srl x, c >= size(x)) -> undef
3814 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
3815 return DAG.getUNDEF(VT);
3816 // fold (srl x, 0) -> x
3817 if (N1C && N1C->isNullValue())
3819 // if (srl x, c) is known to be zero, return 0
3820 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
3821 APInt::getAllOnesValue(OpSizeInBits)))
3822 return DAG.getConstant(0, VT);
3824 // fold (srl (srl x, c1), c2) -> 0 or (srl x, (add c1, c2))
3825 if (N1C && N0.getOpcode() == ISD::SRL &&
3826 N0.getOperand(1).getOpcode() == ISD::Constant) {
3827 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
3828 uint64_t c2 = N1C->getZExtValue();
3829 if (c1 + c2 >= OpSizeInBits)
3830 return DAG.getConstant(0, VT);
3831 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0.getOperand(0),
3832 DAG.getConstant(c1 + c2, N1.getValueType()));
3835 // fold (srl (trunc (srl x, c1)), c2) -> 0 or (trunc (srl x, (add c1, c2)))
3836 if (N1C && N0.getOpcode() == ISD::TRUNCATE &&
3837 N0.getOperand(0).getOpcode() == ISD::SRL &&
3838 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
3840 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
3841 uint64_t c2 = N1C->getZExtValue();
3842 EVT InnerShiftVT = N0.getOperand(0).getValueType();
3843 EVT ShiftCountVT = N0.getOperand(0)->getOperand(1).getValueType();
3844 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
3845 // This is only valid if the OpSizeInBits + c1 = size of inner shift.
3846 if (c1 + OpSizeInBits == InnerShiftSize) {
3847 if (c1 + c2 >= InnerShiftSize)
3848 return DAG.getConstant(0, VT);
3849 return DAG.getNode(ISD::TRUNCATE, N0->getDebugLoc(), VT,
3850 DAG.getNode(ISD::SRL, N0->getDebugLoc(), InnerShiftVT,
3851 N0.getOperand(0)->getOperand(0),
3852 DAG.getConstant(c1 + c2, ShiftCountVT)));
3856 // fold (srl (shl x, c), c) -> (and x, cst2)
3857 if (N1C && N0.getOpcode() == ISD::SHL && N0.getOperand(1) == N1 &&
3858 N0.getValueSizeInBits() <= 64) {
3859 uint64_t ShAmt = N1C->getZExtValue()+64-N0.getValueSizeInBits();
3860 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0.getOperand(0),
3861 DAG.getConstant(~0ULL >> ShAmt, VT));
3865 // fold (srl (anyextend x), c) -> (anyextend (srl x, c))
3866 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
3867 // Shifting in all undef bits?
3868 EVT SmallVT = N0.getOperand(0).getValueType();
3869 if (N1C->getZExtValue() >= SmallVT.getSizeInBits())
3870 return DAG.getUNDEF(VT);
3872 if (!LegalTypes || TLI.isTypeDesirableForOp(ISD::SRL, SmallVT)) {
3873 uint64_t ShiftAmt = N1C->getZExtValue();
3874 SDValue SmallShift = DAG.getNode(ISD::SRL, N0.getDebugLoc(), SmallVT,
3876 DAG.getConstant(ShiftAmt, getShiftAmountTy(SmallVT)));
3877 AddToWorkList(SmallShift.getNode());
3878 return DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, SmallShift);
3882 // fold (srl (sra X, Y), 31) -> (srl X, 31). This srl only looks at the sign
3883 // bit, which is unmodified by sra.
3884 if (N1C && N1C->getZExtValue() + 1 == VT.getSizeInBits()) {
3885 if (N0.getOpcode() == ISD::SRA)
3886 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0.getOperand(0), N1);
3889 // fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit).
3890 if (N1C && N0.getOpcode() == ISD::CTLZ &&
3891 N1C->getAPIntValue() == Log2_32(VT.getSizeInBits())) {
3892 APInt KnownZero, KnownOne;
3893 DAG.ComputeMaskedBits(N0.getOperand(0), KnownZero, KnownOne);
3895 // If any of the input bits are KnownOne, then the input couldn't be all
3896 // zeros, thus the result of the srl will always be zero.
3897 if (KnownOne.getBoolValue()) return DAG.getConstant(0, VT);
3899 // If all of the bits input the to ctlz node are known to be zero, then
3900 // the result of the ctlz is "32" and the result of the shift is one.
3901 APInt UnknownBits = ~KnownZero;
3902 if (UnknownBits == 0) return DAG.getConstant(1, VT);
3904 // Otherwise, check to see if there is exactly one bit input to the ctlz.
3905 if ((UnknownBits & (UnknownBits - 1)) == 0) {
3906 // Okay, we know that only that the single bit specified by UnknownBits
3907 // could be set on input to the CTLZ node. If this bit is set, the SRL
3908 // will return 0, if it is clear, it returns 1. Change the CTLZ/SRL pair
3909 // to an SRL/XOR pair, which is likely to simplify more.
3910 unsigned ShAmt = UnknownBits.countTrailingZeros();
3911 SDValue Op = N0.getOperand(0);
3914 Op = DAG.getNode(ISD::SRL, N0.getDebugLoc(), VT, Op,
3915 DAG.getConstant(ShAmt, getShiftAmountTy(Op.getValueType())));
3916 AddToWorkList(Op.getNode());
3919 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT,
3920 Op, DAG.getConstant(1, VT));
3924 // fold (srl x, (trunc (and y, c))) -> (srl x, (and (trunc y), (trunc c))).
3925 if (N1.getOpcode() == ISD::TRUNCATE &&
3926 N1.getOperand(0).getOpcode() == ISD::AND &&
3927 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
3928 SDValue N101 = N1.getOperand(0).getOperand(1);
3929 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
3930 EVT TruncVT = N1.getValueType();
3931 SDValue N100 = N1.getOperand(0).getOperand(0);
3932 APInt TruncC = N101C->getAPIntValue();
3933 TruncC = TruncC.trunc(TruncVT.getSizeInBits());
3934 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0,
3935 DAG.getNode(ISD::AND, N->getDebugLoc(),
3937 DAG.getNode(ISD::TRUNCATE,
3940 DAG.getConstant(TruncC, TruncVT)));
3944 // fold operands of srl based on knowledge that the low bits are not
3946 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
3947 return SDValue(N, 0);
3950 SDValue NewSRL = visitShiftByConstant(N, N1C->getZExtValue());
3951 if (NewSRL.getNode())
3955 // Attempt to convert a srl of a load into a narrower zero-extending load.
3956 SDValue NarrowLoad = ReduceLoadWidth(N);
3957 if (NarrowLoad.getNode())
3960 // Here is a common situation. We want to optimize:
3963 // %b = and i32 %a, 2
3964 // %c = srl i32 %b, 1
3965 // brcond i32 %c ...
3971 // %c = setcc eq %b, 0
3974 // However when after the source operand of SRL is optimized into AND, the SRL
3975 // itself may not be optimized further. Look for it and add the BRCOND into
3977 if (N->hasOneUse()) {
3978 SDNode *Use = *N->use_begin();
3979 if (Use->getOpcode() == ISD::BRCOND)
3981 else if (Use->getOpcode() == ISD::TRUNCATE && Use->hasOneUse()) {
3982 // Also look pass the truncate.
3983 Use = *Use->use_begin();
3984 if (Use->getOpcode() == ISD::BRCOND)
3992 SDValue DAGCombiner::visitCTLZ(SDNode *N) {
3993 SDValue N0 = N->getOperand(0);
3994 EVT VT = N->getValueType(0);
3996 // fold (ctlz c1) -> c2
3997 if (isa<ConstantSDNode>(N0))
3998 return DAG.getNode(ISD::CTLZ, N->getDebugLoc(), VT, N0);
4002 SDValue DAGCombiner::visitCTLZ_ZERO_UNDEF(SDNode *N) {
4003 SDValue N0 = N->getOperand(0);
4004 EVT VT = N->getValueType(0);
4006 // fold (ctlz_zero_undef c1) -> c2
4007 if (isa<ConstantSDNode>(N0))
4008 return DAG.getNode(ISD::CTLZ_ZERO_UNDEF, N->getDebugLoc(), VT, N0);
4012 SDValue DAGCombiner::visitCTTZ(SDNode *N) {
4013 SDValue N0 = N->getOperand(0);
4014 EVT VT = N->getValueType(0);
4016 // fold (cttz c1) -> c2
4017 if (isa<ConstantSDNode>(N0))
4018 return DAG.getNode(ISD::CTTZ, N->getDebugLoc(), VT, N0);
4022 SDValue DAGCombiner::visitCTTZ_ZERO_UNDEF(SDNode *N) {
4023 SDValue N0 = N->getOperand(0);
4024 EVT VT = N->getValueType(0);
4026 // fold (cttz_zero_undef c1) -> c2
4027 if (isa<ConstantSDNode>(N0))
4028 return DAG.getNode(ISD::CTTZ_ZERO_UNDEF, N->getDebugLoc(), VT, N0);
4032 SDValue DAGCombiner::visitCTPOP(SDNode *N) {
4033 SDValue N0 = N->getOperand(0);
4034 EVT VT = N->getValueType(0);
4036 // fold (ctpop c1) -> c2
4037 if (isa<ConstantSDNode>(N0))
4038 return DAG.getNode(ISD::CTPOP, N->getDebugLoc(), VT, N0);
4042 SDValue DAGCombiner::visitSELECT(SDNode *N) {
4043 SDValue N0 = N->getOperand(0);
4044 SDValue N1 = N->getOperand(1);
4045 SDValue N2 = N->getOperand(2);
4046 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4047 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4048 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
4049 EVT VT = N->getValueType(0);
4050 EVT VT0 = N0.getValueType();
4052 // fold (select C, X, X) -> X
4055 // fold (select true, X, Y) -> X
4056 if (N0C && !N0C->isNullValue())
4058 // fold (select false, X, Y) -> Y
4059 if (N0C && N0C->isNullValue())
4061 // fold (select C, 1, X) -> (or C, X)
4062 if (VT == MVT::i1 && N1C && N1C->getAPIntValue() == 1)
4063 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N0, N2);
4064 // fold (select C, 0, 1) -> (xor C, 1)
4065 if (VT.isInteger() &&
4068 TLI.getBooleanContents(false) ==
4069 TargetLowering::ZeroOrOneBooleanContent)) &&
4070 N1C && N2C && N1C->isNullValue() && N2C->getAPIntValue() == 1) {
4073 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT0,
4074 N0, DAG.getConstant(1, VT0));
4075 XORNode = DAG.getNode(ISD::XOR, N0.getDebugLoc(), VT0,
4076 N0, DAG.getConstant(1, VT0));
4077 AddToWorkList(XORNode.getNode());
4079 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, XORNode);
4080 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, XORNode);
4082 // fold (select C, 0, X) -> (and (not C), X)
4083 if (VT == VT0 && VT == MVT::i1 && N1C && N1C->isNullValue()) {
4084 SDValue NOTNode = DAG.getNOT(N0.getDebugLoc(), N0, VT);
4085 AddToWorkList(NOTNode.getNode());
4086 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, NOTNode, N2);
4088 // fold (select C, X, 1) -> (or (not C), X)
4089 if (VT == VT0 && VT == MVT::i1 && N2C && N2C->getAPIntValue() == 1) {
4090 SDValue NOTNode = DAG.getNOT(N0.getDebugLoc(), N0, VT);
4091 AddToWorkList(NOTNode.getNode());
4092 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, NOTNode, N1);
4094 // fold (select C, X, 0) -> (and C, X)
4095 if (VT == MVT::i1 && N2C && N2C->isNullValue())
4096 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0, N1);
4097 // fold (select X, X, Y) -> (or X, Y)
4098 // fold (select X, 1, Y) -> (or X, Y)
4099 if (VT == MVT::i1 && (N0 == N1 || (N1C && N1C->getAPIntValue() == 1)))
4100 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N0, N2);
4101 // fold (select X, Y, X) -> (and X, Y)
4102 // fold (select X, Y, 0) -> (and X, Y)
4103 if (VT == MVT::i1 && (N0 == N2 || (N2C && N2C->getAPIntValue() == 0)))
4104 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0, N1);
4106 // If we can fold this based on the true/false value, do so.
4107 if (SimplifySelectOps(N, N1, N2))
4108 return SDValue(N, 0); // Don't revisit N.
4110 // fold selects based on a setcc into other things, such as min/max/abs
4111 if (N0.getOpcode() == ISD::SETCC) {
4113 // Check against MVT::Other for SELECT_CC, which is a workaround for targets
4114 // having to say they don't support SELECT_CC on every type the DAG knows
4115 // about, since there is no way to mark an opcode illegal at all value types
4116 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, MVT::Other) &&
4117 TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT))
4118 return DAG.getNode(ISD::SELECT_CC, N->getDebugLoc(), VT,
4119 N0.getOperand(0), N0.getOperand(1),
4120 N1, N2, N0.getOperand(2));
4121 return SimplifySelect(N->getDebugLoc(), N0, N1, N2);
4127 SDValue DAGCombiner::visitSELECT_CC(SDNode *N) {
4128 SDValue N0 = N->getOperand(0);
4129 SDValue N1 = N->getOperand(1);
4130 SDValue N2 = N->getOperand(2);
4131 SDValue N3 = N->getOperand(3);
4132 SDValue N4 = N->getOperand(4);
4133 ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get();
4135 // fold select_cc lhs, rhs, x, x, cc -> x
4139 // Determine if the condition we're dealing with is constant
4140 SDValue SCC = SimplifySetCC(TLI.getSetCCResultType(N0.getValueType()),
4141 N0, N1, CC, N->getDebugLoc(), false);
4142 if (SCC.getNode()) AddToWorkList(SCC.getNode());
4144 if (ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode())) {
4145 if (!SCCC->isNullValue())
4146 return N2; // cond always true -> true val
4148 return N3; // cond always false -> false val
4151 // Fold to a simpler select_cc
4152 if (SCC.getNode() && SCC.getOpcode() == ISD::SETCC)
4153 return DAG.getNode(ISD::SELECT_CC, N->getDebugLoc(), N2.getValueType(),
4154 SCC.getOperand(0), SCC.getOperand(1), N2, N3,
4157 // If we can fold this based on the true/false value, do so.
4158 if (SimplifySelectOps(N, N2, N3))
4159 return SDValue(N, 0); // Don't revisit N.
4161 // fold select_cc into other things, such as min/max/abs
4162 return SimplifySelectCC(N->getDebugLoc(), N0, N1, N2, N3, CC);
4165 SDValue DAGCombiner::visitSETCC(SDNode *N) {
4166 return SimplifySetCC(N->getValueType(0), N->getOperand(0), N->getOperand(1),
4167 cast<CondCodeSDNode>(N->getOperand(2))->get(),
4171 // ExtendUsesToFormExtLoad - Trying to extend uses of a load to enable this:
4172 // "fold ({s|z|a}ext (load x)) -> ({s|z|a}ext (truncate ({s|z|a}extload x)))"
4173 // transformation. Returns true if extension are possible and the above
4174 // mentioned transformation is profitable.
4175 static bool ExtendUsesToFormExtLoad(SDNode *N, SDValue N0,
4177 SmallVector<SDNode*, 4> &ExtendNodes,
4178 const TargetLowering &TLI) {
4179 bool HasCopyToRegUses = false;
4180 bool isTruncFree = TLI.isTruncateFree(N->getValueType(0), N0.getValueType());
4181 for (SDNode::use_iterator UI = N0.getNode()->use_begin(),
4182 UE = N0.getNode()->use_end();
4187 if (UI.getUse().getResNo() != N0.getResNo())
4189 // FIXME: Only extend SETCC N, N and SETCC N, c for now.
4190 if (ExtOpc != ISD::ANY_EXTEND && User->getOpcode() == ISD::SETCC) {
4191 ISD::CondCode CC = cast<CondCodeSDNode>(User->getOperand(2))->get();
4192 if (ExtOpc == ISD::ZERO_EXTEND && ISD::isSignedIntSetCC(CC))
4193 // Sign bits will be lost after a zext.
4196 for (unsigned i = 0; i != 2; ++i) {
4197 SDValue UseOp = User->getOperand(i);
4200 if (!isa<ConstantSDNode>(UseOp))
4205 ExtendNodes.push_back(User);
4208 // If truncates aren't free and there are users we can't
4209 // extend, it isn't worthwhile.
4212 // Remember if this value is live-out.
4213 if (User->getOpcode() == ISD::CopyToReg)
4214 HasCopyToRegUses = true;
4217 if (HasCopyToRegUses) {
4218 bool BothLiveOut = false;
4219 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
4221 SDUse &Use = UI.getUse();
4222 if (Use.getResNo() == 0 && Use.getUser()->getOpcode() == ISD::CopyToReg) {
4228 // Both unextended and extended values are live out. There had better be
4229 // a good reason for the transformation.
4230 return ExtendNodes.size();
4235 void DAGCombiner::ExtendSetCCUses(SmallVector<SDNode*, 4> SetCCs,
4236 SDValue Trunc, SDValue ExtLoad, DebugLoc DL,
4237 ISD::NodeType ExtType) {
4238 // Extend SetCC uses if necessary.
4239 for (unsigned i = 0, e = SetCCs.size(); i != e; ++i) {
4240 SDNode *SetCC = SetCCs[i];
4241 SmallVector<SDValue, 4> Ops;
4243 for (unsigned j = 0; j != 2; ++j) {
4244 SDValue SOp = SetCC->getOperand(j);
4246 Ops.push_back(ExtLoad);
4248 Ops.push_back(DAG.getNode(ExtType, DL, ExtLoad->getValueType(0), SOp));
4251 Ops.push_back(SetCC->getOperand(2));
4252 CombineTo(SetCC, DAG.getNode(ISD::SETCC, DL, SetCC->getValueType(0),
4253 &Ops[0], Ops.size()));
4257 SDValue DAGCombiner::visitSIGN_EXTEND(SDNode *N) {
4258 SDValue N0 = N->getOperand(0);
4259 EVT VT = N->getValueType(0);
4261 // fold (sext c1) -> c1
4262 if (isa<ConstantSDNode>(N0))
4263 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, N0);
4265 // fold (sext (sext x)) -> (sext x)
4266 // fold (sext (aext x)) -> (sext x)
4267 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
4268 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT,
4271 if (N0.getOpcode() == ISD::TRUNCATE) {
4272 // fold (sext (truncate (load x))) -> (sext (smaller load x))
4273 // fold (sext (truncate (srl (load x), c))) -> (sext (smaller load (x+c/n)))
4274 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
4275 if (NarrowLoad.getNode()) {
4276 SDNode* oye = N0.getNode()->getOperand(0).getNode();
4277 if (NarrowLoad.getNode() != N0.getNode()) {
4278 CombineTo(N0.getNode(), NarrowLoad);
4279 // CombineTo deleted the truncate, if needed, but not what's under it.
4282 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4285 // See if the value being truncated is already sign extended. If so, just
4286 // eliminate the trunc/sext pair.
4287 SDValue Op = N0.getOperand(0);
4288 unsigned OpBits = Op.getValueType().getScalarType().getSizeInBits();
4289 unsigned MidBits = N0.getValueType().getScalarType().getSizeInBits();
4290 unsigned DestBits = VT.getScalarType().getSizeInBits();
4291 unsigned NumSignBits = DAG.ComputeNumSignBits(Op);
4293 if (OpBits == DestBits) {
4294 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
4295 // bits, it is already ready.
4296 if (NumSignBits > DestBits-MidBits)
4298 } else if (OpBits < DestBits) {
4299 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
4300 // bits, just sext from i32.
4301 if (NumSignBits > OpBits-MidBits)
4302 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, Op);
4304 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
4305 // bits, just truncate to i32.
4306 if (NumSignBits > OpBits-MidBits)
4307 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, Op);
4310 // fold (sext (truncate x)) -> (sextinreg x).
4311 if (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG,
4312 N0.getValueType())) {
4313 if (OpBits < DestBits)
4314 Op = DAG.getNode(ISD::ANY_EXTEND, N0.getDebugLoc(), VT, Op);
4315 else if (OpBits > DestBits)
4316 Op = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(), VT, Op);
4317 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT, Op,
4318 DAG.getValueType(N0.getValueType()));
4322 // fold (sext (load x)) -> (sext (truncate (sextload x)))
4323 // None of the supported targets knows how to perform load and sign extend
4324 // on vectors in one instruction. We only perform this transformation on
4326 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
4327 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
4328 TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()))) {
4329 bool DoXform = true;
4330 SmallVector<SDNode*, 4> SetCCs;
4331 if (!N0.hasOneUse())
4332 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::SIGN_EXTEND, SetCCs, TLI);
4334 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4335 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
4337 LN0->getBasePtr(), LN0->getPointerInfo(),
4339 LN0->isVolatile(), LN0->isNonTemporal(),
4340 LN0->getAlignment());
4341 CombineTo(N, ExtLoad);
4342 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4343 N0.getValueType(), ExtLoad);
4344 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
4345 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, N->getDebugLoc(),
4347 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4351 // fold (sext (sextload x)) -> (sext (truncate (sextload x)))
4352 // fold (sext ( extload x)) -> (sext (truncate (sextload x)))
4353 if ((ISD::isSEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
4354 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
4355 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4356 EVT MemVT = LN0->getMemoryVT();
4357 if ((!LegalOperations && !LN0->isVolatile()) ||
4358 TLI.isLoadExtLegal(ISD::SEXTLOAD, MemVT)) {
4359 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
4361 LN0->getBasePtr(), LN0->getPointerInfo(),
4363 LN0->isVolatile(), LN0->isNonTemporal(),
4364 LN0->getAlignment());
4365 CombineTo(N, ExtLoad);
4366 CombineTo(N0.getNode(),
4367 DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4368 N0.getValueType(), ExtLoad),
4369 ExtLoad.getValue(1));
4370 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4374 // fold (sext (and/or/xor (load x), cst)) ->
4375 // (and/or/xor (sextload x), (sext cst))
4376 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
4377 N0.getOpcode() == ISD::XOR) &&
4378 isa<LoadSDNode>(N0.getOperand(0)) &&
4379 N0.getOperand(1).getOpcode() == ISD::Constant &&
4380 TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()) &&
4381 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
4382 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
4383 if (LN0->getExtensionType() != ISD::ZEXTLOAD) {
4384 bool DoXform = true;
4385 SmallVector<SDNode*, 4> SetCCs;
4386 if (!N0.hasOneUse())
4387 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::SIGN_EXTEND,
4390 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, LN0->getDebugLoc(), VT,
4391 LN0->getChain(), LN0->getBasePtr(),
4392 LN0->getPointerInfo(),
4395 LN0->isNonTemporal(),
4396 LN0->getAlignment());
4397 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
4398 Mask = Mask.sext(VT.getSizeInBits());
4399 SDValue And = DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT,
4400 ExtLoad, DAG.getConstant(Mask, VT));
4401 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
4402 N0.getOperand(0).getDebugLoc(),
4403 N0.getOperand(0).getValueType(), ExtLoad);
4405 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
4406 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, N->getDebugLoc(),
4408 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4413 if (N0.getOpcode() == ISD::SETCC) {
4414 // sext(setcc) -> sext_in_reg(vsetcc) for vectors.
4415 // Only do this before legalize for now.
4416 if (VT.isVector() && !LegalOperations) {
4417 EVT N0VT = N0.getOperand(0).getValueType();
4418 // On some architectures (such as SSE/NEON/etc) the SETCC result type is
4419 // of the same size as the compared operands. Only optimize sext(setcc())
4420 // if this is the case.
4421 EVT SVT = TLI.getSetCCResultType(N0VT);
4423 // We know that the # elements of the results is the same as the
4424 // # elements of the compare (and the # elements of the compare result
4425 // for that matter). Check to see that they are the same size. If so,
4426 // we know that the element size of the sext'd result matches the
4427 // element size of the compare operands.
4428 if (VT.getSizeInBits() == SVT.getSizeInBits())
4429 return DAG.getSetCC(N->getDebugLoc(), VT, N0.getOperand(0),
4431 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4432 // If the desired elements are smaller or larger than the source
4433 // elements we can use a matching integer vector type and then
4434 // truncate/sign extend
4436 EVT MatchingElementType =
4437 EVT::getIntegerVT(*DAG.getContext(),
4438 N0VT.getScalarType().getSizeInBits());
4439 EVT MatchingVectorType =
4440 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
4441 N0VT.getVectorNumElements());
4443 if (SVT == MatchingVectorType) {
4444 SDValue VsetCC = DAG.getSetCC(N->getDebugLoc(), MatchingVectorType,
4445 N0.getOperand(0), N0.getOperand(1),
4446 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4447 return DAG.getSExtOrTrunc(VsetCC, N->getDebugLoc(), VT);
4452 // sext(setcc x, y, cc) -> (select_cc x, y, -1, 0, cc)
4453 unsigned ElementWidth = VT.getScalarType().getSizeInBits();
4455 DAG.getConstant(APInt::getAllOnesValue(ElementWidth), VT);
4457 SimplifySelectCC(N->getDebugLoc(), N0.getOperand(0), N0.getOperand(1),
4458 NegOne, DAG.getConstant(0, VT),
4459 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
4460 if (SCC.getNode()) return SCC;
4461 if (!LegalOperations ||
4462 TLI.isOperationLegal(ISD::SETCC, TLI.getSetCCResultType(VT)))
4463 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
4464 DAG.getSetCC(N->getDebugLoc(),
4465 TLI.getSetCCResultType(VT),
4466 N0.getOperand(0), N0.getOperand(1),
4467 cast<CondCodeSDNode>(N0.getOperand(2))->get()),
4468 NegOne, DAG.getConstant(0, VT));
4471 // fold (sext x) -> (zext x) if the sign bit is known zero.
4472 if ((!LegalOperations || TLI.isOperationLegal(ISD::ZERO_EXTEND, VT)) &&
4473 DAG.SignBitIsZero(N0))
4474 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, N0);
4479 // isTruncateOf - If N is a truncate of some other value, return true, record
4480 // the value being truncated in Op and which of Op's bits are zero in KnownZero.
4481 // This function computes KnownZero to avoid a duplicated call to
4482 // ComputeMaskedBits in the caller.
4483 static bool isTruncateOf(SelectionDAG &DAG, SDValue N, SDValue &Op,
4486 if (N->getOpcode() == ISD::TRUNCATE) {
4487 Op = N->getOperand(0);
4488 DAG.ComputeMaskedBits(Op, KnownZero, KnownOne);
4492 if (N->getOpcode() != ISD::SETCC || N->getValueType(0) != MVT::i1 ||
4493 cast<CondCodeSDNode>(N->getOperand(2))->get() != ISD::SETNE)
4496 SDValue Op0 = N->getOperand(0);
4497 SDValue Op1 = N->getOperand(1);
4498 assert(Op0.getValueType() == Op1.getValueType());
4500 ConstantSDNode *COp0 = dyn_cast<ConstantSDNode>(Op0);
4501 ConstantSDNode *COp1 = dyn_cast<ConstantSDNode>(Op1);
4502 if (COp0 && COp0->isNullValue())
4504 else if (COp1 && COp1->isNullValue())
4509 DAG.ComputeMaskedBits(Op, KnownZero, KnownOne);
4511 if (!(KnownZero | APInt(Op.getValueSizeInBits(), 1)).isAllOnesValue())
4517 SDValue DAGCombiner::visitZERO_EXTEND(SDNode *N) {
4518 SDValue N0 = N->getOperand(0);
4519 EVT VT = N->getValueType(0);
4521 // fold (zext c1) -> c1
4522 if (isa<ConstantSDNode>(N0))
4523 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, N0);
4524 // fold (zext (zext x)) -> (zext x)
4525 // fold (zext (aext x)) -> (zext x)
4526 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
4527 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT,
4530 // fold (zext (truncate x)) -> (zext x) or
4531 // (zext (truncate x)) -> (truncate x)
4532 // This is valid when the truncated bits of x are already zero.
4533 // FIXME: We should extend this to work for vectors too.
4536 if (!VT.isVector() && isTruncateOf(DAG, N0, Op, KnownZero)) {
4537 APInt TruncatedBits =
4538 (Op.getValueSizeInBits() == N0.getValueSizeInBits()) ?
4539 APInt(Op.getValueSizeInBits(), 0) :
4540 APInt::getBitsSet(Op.getValueSizeInBits(),
4541 N0.getValueSizeInBits(),
4542 std::min(Op.getValueSizeInBits(),
4543 VT.getSizeInBits()));
4544 if (TruncatedBits == (KnownZero & TruncatedBits)) {
4545 if (VT.bitsGT(Op.getValueType()))
4546 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, Op);
4547 if (VT.bitsLT(Op.getValueType()))
4548 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, Op);
4554 // fold (zext (truncate (load x))) -> (zext (smaller load x))
4555 // fold (zext (truncate (srl (load x), c))) -> (zext (small load (x+c/n)))
4556 if (N0.getOpcode() == ISD::TRUNCATE) {
4557 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
4558 if (NarrowLoad.getNode()) {
4559 SDNode* oye = N0.getNode()->getOperand(0).getNode();
4560 if (NarrowLoad.getNode() != N0.getNode()) {
4561 CombineTo(N0.getNode(), NarrowLoad);
4562 // CombineTo deleted the truncate, if needed, but not what's under it.
4565 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4569 // fold (zext (truncate x)) -> (and x, mask)
4570 if (N0.getOpcode() == ISD::TRUNCATE &&
4571 (!LegalOperations || TLI.isOperationLegal(ISD::AND, VT))) {
4573 // fold (zext (truncate (load x))) -> (zext (smaller load x))
4574 // fold (zext (truncate (srl (load x), c))) -> (zext (smaller load (x+c/n)))
4575 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
4576 if (NarrowLoad.getNode()) {
4577 SDNode* oye = N0.getNode()->getOperand(0).getNode();
4578 if (NarrowLoad.getNode() != N0.getNode()) {
4579 CombineTo(N0.getNode(), NarrowLoad);
4580 // CombineTo deleted the truncate, if needed, but not what's under it.
4583 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4586 SDValue Op = N0.getOperand(0);
4587 if (Op.getValueType().bitsLT(VT)) {
4588 Op = DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, Op);
4589 AddToWorkList(Op.getNode());
4590 } else if (Op.getValueType().bitsGT(VT)) {
4591 Op = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, Op);
4592 AddToWorkList(Op.getNode());
4594 return DAG.getZeroExtendInReg(Op, N->getDebugLoc(),
4595 N0.getValueType().getScalarType());
4598 // Fold (zext (and (trunc x), cst)) -> (and x, cst),
4599 // if either of the casts is not free.
4600 if (N0.getOpcode() == ISD::AND &&
4601 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
4602 N0.getOperand(1).getOpcode() == ISD::Constant &&
4603 (!TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
4604 N0.getValueType()) ||
4605 !TLI.isZExtFree(N0.getValueType(), VT))) {
4606 SDValue X = N0.getOperand(0).getOperand(0);
4607 if (X.getValueType().bitsLT(VT)) {
4608 X = DAG.getNode(ISD::ANY_EXTEND, X.getDebugLoc(), VT, X);
4609 } else if (X.getValueType().bitsGT(VT)) {
4610 X = DAG.getNode(ISD::TRUNCATE, X.getDebugLoc(), VT, X);
4612 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
4613 Mask = Mask.zext(VT.getSizeInBits());
4614 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
4615 X, DAG.getConstant(Mask, VT));
4618 // fold (zext (load x)) -> (zext (truncate (zextload x)))
4619 // None of the supported targets knows how to perform load and vector_zext
4620 // on vectors in one instruction. We only perform this transformation on
4622 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
4623 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
4624 TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()))) {
4625 bool DoXform = true;
4626 SmallVector<SDNode*, 4> SetCCs;
4627 if (!N0.hasOneUse())
4628 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ZERO_EXTEND, SetCCs, TLI);
4630 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4631 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N->getDebugLoc(), VT,
4633 LN0->getBasePtr(), LN0->getPointerInfo(),
4635 LN0->isVolatile(), LN0->isNonTemporal(),
4636 LN0->getAlignment());
4637 CombineTo(N, ExtLoad);
4638 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4639 N0.getValueType(), ExtLoad);
4640 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
4642 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, N->getDebugLoc(),
4644 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4648 // fold (zext (and/or/xor (load x), cst)) ->
4649 // (and/or/xor (zextload x), (zext cst))
4650 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
4651 N0.getOpcode() == ISD::XOR) &&
4652 isa<LoadSDNode>(N0.getOperand(0)) &&
4653 N0.getOperand(1).getOpcode() == ISD::Constant &&
4654 TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()) &&
4655 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
4656 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
4657 if (LN0->getExtensionType() != ISD::SEXTLOAD) {
4658 bool DoXform = true;
4659 SmallVector<SDNode*, 4> SetCCs;
4660 if (!N0.hasOneUse())
4661 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::ZERO_EXTEND,
4664 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, LN0->getDebugLoc(), VT,
4665 LN0->getChain(), LN0->getBasePtr(),
4666 LN0->getPointerInfo(),
4669 LN0->isNonTemporal(),
4670 LN0->getAlignment());
4671 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
4672 Mask = Mask.zext(VT.getSizeInBits());
4673 SDValue And = DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT,
4674 ExtLoad, DAG.getConstant(Mask, VT));
4675 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
4676 N0.getOperand(0).getDebugLoc(),
4677 N0.getOperand(0).getValueType(), ExtLoad);
4679 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
4680 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, N->getDebugLoc(),
4682 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4687 // fold (zext (zextload x)) -> (zext (truncate (zextload x)))
4688 // fold (zext ( extload x)) -> (zext (truncate (zextload x)))
4689 if ((ISD::isZEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
4690 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
4691 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4692 EVT MemVT = LN0->getMemoryVT();
4693 if ((!LegalOperations && !LN0->isVolatile()) ||
4694 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT)) {
4695 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N->getDebugLoc(), VT,
4697 LN0->getBasePtr(), LN0->getPointerInfo(),
4699 LN0->isVolatile(), LN0->isNonTemporal(),
4700 LN0->getAlignment());
4701 CombineTo(N, ExtLoad);
4702 CombineTo(N0.getNode(),
4703 DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(), N0.getValueType(),
4705 ExtLoad.getValue(1));
4706 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4710 if (N0.getOpcode() == ISD::SETCC) {
4711 if (!LegalOperations && VT.isVector()) {
4712 // zext(setcc) -> (and (vsetcc), (1, 1, ...) for vectors.
4713 // Only do this before legalize for now.
4714 EVT N0VT = N0.getOperand(0).getValueType();
4715 EVT EltVT = VT.getVectorElementType();
4716 SmallVector<SDValue,8> OneOps(VT.getVectorNumElements(),
4717 DAG.getConstant(1, EltVT));
4718 if (VT.getSizeInBits() == N0VT.getSizeInBits())
4719 // We know that the # elements of the results is the same as the
4720 // # elements of the compare (and the # elements of the compare result
4721 // for that matter). Check to see that they are the same size. If so,
4722 // we know that the element size of the sext'd result matches the
4723 // element size of the compare operands.
4724 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
4725 DAG.getSetCC(N->getDebugLoc(), VT, N0.getOperand(0),
4727 cast<CondCodeSDNode>(N0.getOperand(2))->get()),
4728 DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), VT,
4729 &OneOps[0], OneOps.size()));
4731 // If the desired elements are smaller or larger than the source
4732 // elements we can use a matching integer vector type and then
4733 // truncate/sign extend
4734 EVT MatchingElementType =
4735 EVT::getIntegerVT(*DAG.getContext(),
4736 N0VT.getScalarType().getSizeInBits());
4737 EVT MatchingVectorType =
4738 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
4739 N0VT.getVectorNumElements());
4741 DAG.getSetCC(N->getDebugLoc(), MatchingVectorType, N0.getOperand(0),
4743 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4744 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
4745 DAG.getSExtOrTrunc(VsetCC, N->getDebugLoc(), VT),
4746 DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), VT,
4747 &OneOps[0], OneOps.size()));
4750 // zext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
4752 SimplifySelectCC(N->getDebugLoc(), N0.getOperand(0), N0.getOperand(1),
4753 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
4754 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
4755 if (SCC.getNode()) return SCC;
4758 // (zext (shl (zext x), cst)) -> (shl (zext x), cst)
4759 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL) &&
4760 isa<ConstantSDNode>(N0.getOperand(1)) &&
4761 N0.getOperand(0).getOpcode() == ISD::ZERO_EXTEND &&
4763 SDValue ShAmt = N0.getOperand(1);
4764 unsigned ShAmtVal = cast<ConstantSDNode>(ShAmt)->getZExtValue();
4765 if (N0.getOpcode() == ISD::SHL) {
4766 SDValue InnerZExt = N0.getOperand(0);
4767 // If the original shl may be shifting out bits, do not perform this
4769 unsigned KnownZeroBits = InnerZExt.getValueType().getSizeInBits() -
4770 InnerZExt.getOperand(0).getValueType().getSizeInBits();
4771 if (ShAmtVal > KnownZeroBits)
4775 DebugLoc DL = N->getDebugLoc();
4777 // Ensure that the shift amount is wide enough for the shifted value.
4778 if (VT.getSizeInBits() >= 256)
4779 ShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, ShAmt);
4781 return DAG.getNode(N0.getOpcode(), DL, VT,
4782 DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0)),
4789 SDValue DAGCombiner::visitANY_EXTEND(SDNode *N) {
4790 SDValue N0 = N->getOperand(0);
4791 EVT VT = N->getValueType(0);
4793 // fold (aext c1) -> c1
4794 if (isa<ConstantSDNode>(N0))
4795 return DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, N0);
4796 // fold (aext (aext x)) -> (aext x)
4797 // fold (aext (zext x)) -> (zext x)
4798 // fold (aext (sext x)) -> (sext x)
4799 if (N0.getOpcode() == ISD::ANY_EXTEND ||
4800 N0.getOpcode() == ISD::ZERO_EXTEND ||
4801 N0.getOpcode() == ISD::SIGN_EXTEND)
4802 return DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT, N0.getOperand(0));
4804 // fold (aext (truncate (load x))) -> (aext (smaller load x))
4805 // fold (aext (truncate (srl (load x), c))) -> (aext (small load (x+c/n)))
4806 if (N0.getOpcode() == ISD::TRUNCATE) {
4807 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
4808 if (NarrowLoad.getNode()) {
4809 SDNode* oye = N0.getNode()->getOperand(0).getNode();
4810 if (NarrowLoad.getNode() != N0.getNode()) {
4811 CombineTo(N0.getNode(), NarrowLoad);
4812 // CombineTo deleted the truncate, if needed, but not what's under it.
4815 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4819 // fold (aext (truncate x))
4820 if (N0.getOpcode() == ISD::TRUNCATE) {
4821 SDValue TruncOp = N0.getOperand(0);
4822 if (TruncOp.getValueType() == VT)
4823 return TruncOp; // x iff x size == zext size.
4824 if (TruncOp.getValueType().bitsGT(VT))
4825 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, TruncOp);
4826 return DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, TruncOp);
4829 // Fold (aext (and (trunc x), cst)) -> (and x, cst)
4830 // if the trunc is not free.
4831 if (N0.getOpcode() == ISD::AND &&
4832 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
4833 N0.getOperand(1).getOpcode() == ISD::Constant &&
4834 !TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
4835 N0.getValueType())) {
4836 SDValue X = N0.getOperand(0).getOperand(0);
4837 if (X.getValueType().bitsLT(VT)) {
4838 X = DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, X);
4839 } else if (X.getValueType().bitsGT(VT)) {
4840 X = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, X);
4842 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
4843 Mask = Mask.zext(VT.getSizeInBits());
4844 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
4845 X, DAG.getConstant(Mask, VT));
4848 // fold (aext (load x)) -> (aext (truncate (extload x)))
4849 // None of the supported targets knows how to perform load and any_ext
4850 // on vectors in one instruction. We only perform this transformation on
4852 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
4853 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
4854 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType()))) {
4855 bool DoXform = true;
4856 SmallVector<SDNode*, 4> SetCCs;
4857 if (!N0.hasOneUse())
4858 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ANY_EXTEND, SetCCs, TLI);
4860 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4861 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, N->getDebugLoc(), VT,
4863 LN0->getBasePtr(), LN0->getPointerInfo(),
4865 LN0->isVolatile(), LN0->isNonTemporal(),
4866 LN0->getAlignment());
4867 CombineTo(N, ExtLoad);
4868 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4869 N0.getValueType(), ExtLoad);
4870 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
4871 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, N->getDebugLoc(),
4873 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4877 // fold (aext (zextload x)) -> (aext (truncate (zextload x)))
4878 // fold (aext (sextload x)) -> (aext (truncate (sextload x)))
4879 // fold (aext ( extload x)) -> (aext (truncate (extload x)))
4880 if (N0.getOpcode() == ISD::LOAD &&
4881 !ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
4883 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4884 EVT MemVT = LN0->getMemoryVT();
4885 SDValue ExtLoad = DAG.getExtLoad(LN0->getExtensionType(), N->getDebugLoc(),
4886 VT, LN0->getChain(), LN0->getBasePtr(),
4887 LN0->getPointerInfo(), MemVT,
4888 LN0->isVolatile(), LN0->isNonTemporal(),
4889 LN0->getAlignment());
4890 CombineTo(N, ExtLoad);
4891 CombineTo(N0.getNode(),
4892 DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4893 N0.getValueType(), ExtLoad),
4894 ExtLoad.getValue(1));
4895 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4898 if (N0.getOpcode() == ISD::SETCC) {
4899 // aext(setcc) -> sext_in_reg(vsetcc) for vectors.
4900 // Only do this before legalize for now.
4901 if (VT.isVector() && !LegalOperations) {
4902 EVT N0VT = N0.getOperand(0).getValueType();
4903 // We know that the # elements of the results is the same as the
4904 // # elements of the compare (and the # elements of the compare result
4905 // for that matter). Check to see that they are the same size. If so,
4906 // we know that the element size of the sext'd result matches the
4907 // element size of the compare operands.
4908 if (VT.getSizeInBits() == N0VT.getSizeInBits())
4909 return DAG.getSetCC(N->getDebugLoc(), VT, N0.getOperand(0),
4911 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4912 // If the desired elements are smaller or larger than the source
4913 // elements we can use a matching integer vector type and then
4914 // truncate/sign extend
4916 EVT MatchingElementType =
4917 EVT::getIntegerVT(*DAG.getContext(),
4918 N0VT.getScalarType().getSizeInBits());
4919 EVT MatchingVectorType =
4920 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
4921 N0VT.getVectorNumElements());
4923 DAG.getSetCC(N->getDebugLoc(), MatchingVectorType, N0.getOperand(0),
4925 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4926 return DAG.getSExtOrTrunc(VsetCC, N->getDebugLoc(), VT);
4930 // aext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
4932 SimplifySelectCC(N->getDebugLoc(), N0.getOperand(0), N0.getOperand(1),
4933 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
4934 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
4942 /// GetDemandedBits - See if the specified operand can be simplified with the
4943 /// knowledge that only the bits specified by Mask are used. If so, return the
4944 /// simpler operand, otherwise return a null SDValue.
4945 SDValue DAGCombiner::GetDemandedBits(SDValue V, const APInt &Mask) {
4946 switch (V.getOpcode()) {
4948 case ISD::Constant: {
4949 const ConstantSDNode *CV = cast<ConstantSDNode>(V.getNode());
4950 assert(CV != 0 && "Const value should be ConstSDNode.");
4951 const APInt &CVal = CV->getAPIntValue();
4952 APInt NewVal = CVal & Mask;
4953 if (NewVal != CVal) {
4954 return DAG.getConstant(NewVal, V.getValueType());
4960 // If the LHS or RHS don't contribute bits to the or, drop them.
4961 if (DAG.MaskedValueIsZero(V.getOperand(0), Mask))
4962 return V.getOperand(1);
4963 if (DAG.MaskedValueIsZero(V.getOperand(1), Mask))
4964 return V.getOperand(0);
4967 // Only look at single-use SRLs.
4968 if (!V.getNode()->hasOneUse())
4970 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(V.getOperand(1))) {
4971 // See if we can recursively simplify the LHS.
4972 unsigned Amt = RHSC->getZExtValue();
4974 // Watch out for shift count overflow though.
4975 if (Amt >= Mask.getBitWidth()) break;
4976 APInt NewMask = Mask << Amt;
4977 SDValue SimplifyLHS = GetDemandedBits(V.getOperand(0), NewMask);
4978 if (SimplifyLHS.getNode())
4979 return DAG.getNode(ISD::SRL, V.getDebugLoc(), V.getValueType(),
4980 SimplifyLHS, V.getOperand(1));
4986 /// ReduceLoadWidth - If the result of a wider load is shifted to right of N
4987 /// bits and then truncated to a narrower type and where N is a multiple
4988 /// of number of bits of the narrower type, transform it to a narrower load
4989 /// from address + N / num of bits of new type. If the result is to be
4990 /// extended, also fold the extension to form a extending load.
4991 SDValue DAGCombiner::ReduceLoadWidth(SDNode *N) {
4992 unsigned Opc = N->getOpcode();
4994 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
4995 SDValue N0 = N->getOperand(0);
4996 EVT VT = N->getValueType(0);
4999 // This transformation isn't valid for vector loads.
5003 // Special case: SIGN_EXTEND_INREG is basically truncating to ExtVT then
5005 if (Opc == ISD::SIGN_EXTEND_INREG) {
5006 ExtType = ISD::SEXTLOAD;
5007 ExtVT = cast<VTSDNode>(N->getOperand(1))->getVT();
5008 } else if (Opc == ISD::SRL) {
5009 // Another special-case: SRL is basically zero-extending a narrower value.
5010 ExtType = ISD::ZEXTLOAD;
5012 ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1));
5013 if (!N01) return SDValue();
5014 ExtVT = EVT::getIntegerVT(*DAG.getContext(),
5015 VT.getSizeInBits() - N01->getZExtValue());
5017 if (LegalOperations && !TLI.isLoadExtLegal(ExtType, ExtVT))
5020 unsigned EVTBits = ExtVT.getSizeInBits();
5022 // Do not generate loads of non-round integer types since these can
5023 // be expensive (and would be wrong if the type is not byte sized).
5024 if (!ExtVT.isRound())
5028 if (N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
5029 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
5030 ShAmt = N01->getZExtValue();
5031 // Is the shift amount a multiple of size of VT?
5032 if ((ShAmt & (EVTBits-1)) == 0) {
5033 N0 = N0.getOperand(0);
5034 // Is the load width a multiple of size of VT?
5035 if ((N0.getValueType().getSizeInBits() & (EVTBits-1)) != 0)
5039 // At this point, we must have a load or else we can't do the transform.
5040 if (!isa<LoadSDNode>(N0)) return SDValue();
5042 // If the shift amount is larger than the input type then we're not
5043 // accessing any of the loaded bytes. If the load was a zextload/extload
5044 // then the result of the shift+trunc is zero/undef (handled elsewhere).
5045 // If the load was a sextload then the result is a splat of the sign bit
5046 // of the extended byte. This is not worth optimizing for.
5047 if (ShAmt >= cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits())
5052 // If the load is shifted left (and the result isn't shifted back right),
5053 // we can fold the truncate through the shift.
5054 unsigned ShLeftAmt = 0;
5055 if (ShAmt == 0 && N0.getOpcode() == ISD::SHL && N0.hasOneUse() &&
5056 ExtVT == VT && TLI.isNarrowingProfitable(N0.getValueType(), VT)) {
5057 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
5058 ShLeftAmt = N01->getZExtValue();
5059 N0 = N0.getOperand(0);
5063 // If we haven't found a load, we can't narrow it. Don't transform one with
5064 // multiple uses, this would require adding a new load.
5065 if (!isa<LoadSDNode>(N0) || !N0.hasOneUse() ||
5066 // Don't change the width of a volatile load.
5067 cast<LoadSDNode>(N0)->isVolatile())
5070 // Verify that we are actually reducing a load width here.
5071 if (cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits() < EVTBits)
5074 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5075 EVT PtrType = N0.getOperand(1).getValueType();
5077 if (PtrType == MVT::Untyped || PtrType.isExtended())
5078 // It's not possible to generate a constant of extended or untyped type.
5081 // For big endian targets, we need to adjust the offset to the pointer to
5082 // load the correct bytes.
5083 if (TLI.isBigEndian()) {
5084 unsigned LVTStoreBits = LN0->getMemoryVT().getStoreSizeInBits();
5085 unsigned EVTStoreBits = ExtVT.getStoreSizeInBits();
5086 ShAmt = LVTStoreBits - EVTStoreBits - ShAmt;
5089 uint64_t PtrOff = ShAmt / 8;
5090 unsigned NewAlign = MinAlign(LN0->getAlignment(), PtrOff);
5091 SDValue NewPtr = DAG.getNode(ISD::ADD, LN0->getDebugLoc(),
5092 PtrType, LN0->getBasePtr(),
5093 DAG.getConstant(PtrOff, PtrType));
5094 AddToWorkList(NewPtr.getNode());
5097 if (ExtType == ISD::NON_EXTLOAD)
5098 Load = DAG.getLoad(VT, N0.getDebugLoc(), LN0->getChain(), NewPtr,
5099 LN0->getPointerInfo().getWithOffset(PtrOff),
5100 LN0->isVolatile(), LN0->isNonTemporal(),
5101 LN0->isInvariant(), NewAlign);
5103 Load = DAG.getExtLoad(ExtType, N0.getDebugLoc(), VT, LN0->getChain(),NewPtr,
5104 LN0->getPointerInfo().getWithOffset(PtrOff),
5105 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
5108 // Replace the old load's chain with the new load's chain.
5109 WorkListRemover DeadNodes(*this);
5110 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1));
5112 // Shift the result left, if we've swallowed a left shift.
5113 SDValue Result = Load;
5114 if (ShLeftAmt != 0) {
5115 EVT ShImmTy = getShiftAmountTy(Result.getValueType());
5116 if (!isUIntN(ShImmTy.getSizeInBits(), ShLeftAmt))
5118 Result = DAG.getNode(ISD::SHL, N0.getDebugLoc(), VT,
5119 Result, DAG.getConstant(ShLeftAmt, ShImmTy));
5122 // Return the new loaded value.
5126 SDValue DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
5127 SDValue N0 = N->getOperand(0);
5128 SDValue N1 = N->getOperand(1);
5129 EVT VT = N->getValueType(0);
5130 EVT EVT = cast<VTSDNode>(N1)->getVT();
5131 unsigned VTBits = VT.getScalarType().getSizeInBits();
5132 unsigned EVTBits = EVT.getScalarType().getSizeInBits();
5134 // fold (sext_in_reg c1) -> c1
5135 if (isa<ConstantSDNode>(N0) || N0.getOpcode() == ISD::UNDEF)
5136 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT, N0, N1);
5138 // If the input is already sign extended, just drop the extension.
5139 if (DAG.ComputeNumSignBits(N0) >= VTBits-EVTBits+1)
5142 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2
5143 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
5144 EVT.bitsLT(cast<VTSDNode>(N0.getOperand(1))->getVT())) {
5145 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT,
5146 N0.getOperand(0), N1);
5149 // fold (sext_in_reg (sext x)) -> (sext x)
5150 // fold (sext_in_reg (aext x)) -> (sext x)
5151 // if x is small enough.
5152 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) {
5153 SDValue N00 = N0.getOperand(0);
5154 if (N00.getValueType().getScalarType().getSizeInBits() <= EVTBits &&
5155 (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND, VT)))
5156 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, N00, N1);
5159 // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is known zero.
5160 if (DAG.MaskedValueIsZero(N0, APInt::getBitsSet(VTBits, EVTBits-1, EVTBits)))
5161 return DAG.getZeroExtendInReg(N0, N->getDebugLoc(), EVT);
5163 // fold operands of sext_in_reg based on knowledge that the top bits are not
5165 if (SimplifyDemandedBits(SDValue(N, 0)))
5166 return SDValue(N, 0);
5168 // fold (sext_in_reg (load x)) -> (smaller sextload x)
5169 // fold (sext_in_reg (srl (load x), c)) -> (smaller sextload (x+c/evtbits))
5170 SDValue NarrowLoad = ReduceLoadWidth(N);
5171 if (NarrowLoad.getNode())
5174 // fold (sext_in_reg (srl X, 24), i8) -> (sra X, 24)
5175 // fold (sext_in_reg (srl X, 23), i8) -> (sra X, 23) iff possible.
5176 // We already fold "(sext_in_reg (srl X, 25), i8) -> srl X, 25" above.
5177 if (N0.getOpcode() == ISD::SRL) {
5178 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
5179 if (ShAmt->getZExtValue()+EVTBits <= VTBits) {
5180 // We can turn this into an SRA iff the input to the SRL is already sign
5182 unsigned InSignBits = DAG.ComputeNumSignBits(N0.getOperand(0));
5183 if (VTBits-(ShAmt->getZExtValue()+EVTBits) < InSignBits)
5184 return DAG.getNode(ISD::SRA, N->getDebugLoc(), VT,
5185 N0.getOperand(0), N0.getOperand(1));
5189 // fold (sext_inreg (extload x)) -> (sextload x)
5190 if (ISD::isEXTLoad(N0.getNode()) &&
5191 ISD::isUNINDEXEDLoad(N0.getNode()) &&
5192 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
5193 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5194 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) {
5195 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5196 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
5198 LN0->getBasePtr(), LN0->getPointerInfo(),
5200 LN0->isVolatile(), LN0->isNonTemporal(),
5201 LN0->getAlignment());
5202 CombineTo(N, ExtLoad);
5203 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
5204 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5206 // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use
5207 if (ISD::isZEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
5209 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
5210 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5211 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) {
5212 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5213 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
5215 LN0->getBasePtr(), LN0->getPointerInfo(),
5217 LN0->isVolatile(), LN0->isNonTemporal(),
5218 LN0->getAlignment());
5219 CombineTo(N, ExtLoad);
5220 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
5221 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5224 // Form (sext_inreg (bswap >> 16)) or (sext_inreg (rotl (bswap) 16))
5225 if (EVTBits <= 16 && N0.getOpcode() == ISD::OR) {
5226 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
5227 N0.getOperand(1), false);
5228 if (BSwap.getNode() != 0)
5229 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT,
5236 SDValue DAGCombiner::visitTRUNCATE(SDNode *N) {
5237 SDValue N0 = N->getOperand(0);
5238 EVT VT = N->getValueType(0);
5239 bool isLE = TLI.isLittleEndian();
5242 if (N0.getValueType() == N->getValueType(0))
5244 // fold (truncate c1) -> c1
5245 if (isa<ConstantSDNode>(N0))
5246 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, N0);
5247 // fold (truncate (truncate x)) -> (truncate x)
5248 if (N0.getOpcode() == ISD::TRUNCATE)
5249 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, N0.getOperand(0));
5250 // fold (truncate (ext x)) -> (ext x) or (truncate x) or x
5251 if (N0.getOpcode() == ISD::ZERO_EXTEND ||
5252 N0.getOpcode() == ISD::SIGN_EXTEND ||
5253 N0.getOpcode() == ISD::ANY_EXTEND) {
5254 if (N0.getOperand(0).getValueType().bitsLT(VT))
5255 // if the source is smaller than the dest, we still need an extend
5256 return DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT,
5258 else if (N0.getOperand(0).getValueType().bitsGT(VT))
5259 // if the source is larger than the dest, than we just need the truncate
5260 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, N0.getOperand(0));
5262 // if the source and dest are the same type, we can drop both the extend
5263 // and the truncate.
5264 return N0.getOperand(0);
5267 // Fold extract-and-trunc into a narrow extract. For example:
5268 // i64 x = EXTRACT_VECTOR_ELT(v2i64 val, i32 1)
5269 // i32 y = TRUNCATE(i64 x)
5271 // v16i8 b = BITCAST (v2i64 val)
5272 // i8 x = EXTRACT_VECTOR_ELT(v16i8 b, i32 8)
5274 // Note: We only run this optimization after type legalization (which often
5275 // creates this pattern) and before operation legalization after which
5276 // we need to be more careful about the vector instructions that we generate.
5277 if (N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5278 LegalTypes && !LegalOperations && N0->hasOneUse()) {
5280 EVT VecTy = N0.getOperand(0).getValueType();
5281 EVT ExTy = N0.getValueType();
5282 EVT TrTy = N->getValueType(0);
5284 unsigned NumElem = VecTy.getVectorNumElements();
5285 unsigned SizeRatio = ExTy.getSizeInBits()/TrTy.getSizeInBits();
5287 EVT NVT = EVT::getVectorVT(*DAG.getContext(), TrTy, SizeRatio * NumElem);
5288 assert(NVT.getSizeInBits() == VecTy.getSizeInBits() && "Invalid Size");
5290 SDValue EltNo = N0->getOperand(1);
5291 if (isa<ConstantSDNode>(EltNo) && isTypeLegal(NVT)) {
5292 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
5293 EVT IndexTy = N0->getOperand(1).getValueType();
5294 int Index = isLE ? (Elt*SizeRatio) : (Elt*SizeRatio + (SizeRatio-1));
5296 SDValue V = DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
5297 NVT, N0.getOperand(0));
5299 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
5300 N->getDebugLoc(), TrTy, V,
5301 DAG.getConstant(Index, IndexTy));
5305 // See if we can simplify the input to this truncate through knowledge that
5306 // only the low bits are being used.
5307 // For example "trunc (or (shl x, 8), y)" // -> trunc y
5308 // Currently we only perform this optimization on scalars because vectors
5309 // may have different active low bits.
5310 if (!VT.isVector()) {
5312 GetDemandedBits(N0, APInt::getLowBitsSet(N0.getValueSizeInBits(),
5313 VT.getSizeInBits()));
5314 if (Shorter.getNode())
5315 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, Shorter);
5317 // fold (truncate (load x)) -> (smaller load x)
5318 // fold (truncate (srl (load x), c)) -> (smaller load (x+c/evtbits))
5319 if (!LegalTypes || TLI.isTypeDesirableForOp(N0.getOpcode(), VT)) {
5320 SDValue Reduced = ReduceLoadWidth(N);
5321 if (Reduced.getNode())
5325 // Simplify the operands using demanded-bits information.
5326 if (!VT.isVector() &&
5327 SimplifyDemandedBits(SDValue(N, 0)))
5328 return SDValue(N, 0);
5333 static SDNode *getBuildPairElt(SDNode *N, unsigned i) {
5334 SDValue Elt = N->getOperand(i);
5335 if (Elt.getOpcode() != ISD::MERGE_VALUES)
5336 return Elt.getNode();
5337 return Elt.getOperand(Elt.getResNo()).getNode();
5340 /// CombineConsecutiveLoads - build_pair (load, load) -> load
5341 /// if load locations are consecutive.
5342 SDValue DAGCombiner::CombineConsecutiveLoads(SDNode *N, EVT VT) {
5343 assert(N->getOpcode() == ISD::BUILD_PAIR);
5345 LoadSDNode *LD1 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 0));
5346 LoadSDNode *LD2 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 1));
5347 if (!LD1 || !LD2 || !ISD::isNON_EXTLoad(LD1) || !LD1->hasOneUse() ||
5348 LD1->getPointerInfo().getAddrSpace() !=
5349 LD2->getPointerInfo().getAddrSpace())
5351 EVT LD1VT = LD1->getValueType(0);
5353 if (ISD::isNON_EXTLoad(LD2) &&
5355 // If both are volatile this would reduce the number of volatile loads.
5356 // If one is volatile it might be ok, but play conservative and bail out.
5357 !LD1->isVolatile() &&
5358 !LD2->isVolatile() &&
5359 DAG.isConsecutiveLoad(LD2, LD1, LD1VT.getSizeInBits()/8, 1)) {
5360 unsigned Align = LD1->getAlignment();
5361 unsigned NewAlign = TLI.getTargetData()->
5362 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
5364 if (NewAlign <= Align &&
5365 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)))
5366 return DAG.getLoad(VT, N->getDebugLoc(), LD1->getChain(),
5367 LD1->getBasePtr(), LD1->getPointerInfo(),
5368 false, false, false, Align);
5374 SDValue DAGCombiner::visitBITCAST(SDNode *N) {
5375 SDValue N0 = N->getOperand(0);
5376 EVT VT = N->getValueType(0);
5378 // If the input is a BUILD_VECTOR with all constant elements, fold this now.
5379 // Only do this before legalize, since afterward the target may be depending
5380 // on the bitconvert.
5381 // First check to see if this is all constant.
5383 N0.getOpcode() == ISD::BUILD_VECTOR && N0.getNode()->hasOneUse() &&
5385 bool isSimple = true;
5386 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i)
5387 if (N0.getOperand(i).getOpcode() != ISD::UNDEF &&
5388 N0.getOperand(i).getOpcode() != ISD::Constant &&
5389 N0.getOperand(i).getOpcode() != ISD::ConstantFP) {
5394 EVT DestEltVT = N->getValueType(0).getVectorElementType();
5395 assert(!DestEltVT.isVector() &&
5396 "Element type of vector ValueType must not be vector!");
5398 return ConstantFoldBITCASTofBUILD_VECTOR(N0.getNode(), DestEltVT);
5401 // If the input is a constant, let getNode fold it.
5402 if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) {
5403 SDValue Res = DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, N0);
5404 if (Res.getNode() != N) {
5405 if (!LegalOperations ||
5406 TLI.isOperationLegal(Res.getNode()->getOpcode(), VT))
5409 // Folding it resulted in an illegal node, and it's too late to
5410 // do that. Clean up the old node and forego the transformation.
5411 // Ideally this won't happen very often, because instcombine
5412 // and the earlier dagcombine runs (where illegal nodes are
5413 // permitted) should have folded most of them already.
5414 DAG.DeleteNode(Res.getNode());
5418 // (conv (conv x, t1), t2) -> (conv x, t2)
5419 if (N0.getOpcode() == ISD::BITCAST)
5420 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT,
5423 // fold (conv (load x)) -> (load (conv*)x)
5424 // If the resultant load doesn't need a higher alignment than the original!
5425 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
5426 // Do not change the width of a volatile load.
5427 !cast<LoadSDNode>(N0)->isVolatile() &&
5428 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT))) {
5429 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5430 unsigned Align = TLI.getTargetData()->
5431 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
5432 unsigned OrigAlign = LN0->getAlignment();
5434 if (Align <= OrigAlign) {
5435 SDValue Load = DAG.getLoad(VT, N->getDebugLoc(), LN0->getChain(),
5436 LN0->getBasePtr(), LN0->getPointerInfo(),
5437 LN0->isVolatile(), LN0->isNonTemporal(),
5438 LN0->isInvariant(), OrigAlign);
5440 CombineTo(N0.getNode(),
5441 DAG.getNode(ISD::BITCAST, N0.getDebugLoc(),
5442 N0.getValueType(), Load),
5448 // fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit)
5449 // fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit))
5450 // This often reduces constant pool loads.
5451 if (((N0.getOpcode() == ISD::FNEG && !TLI.isFNegFree(VT)) ||
5452 (N0.getOpcode() == ISD::FABS && !TLI.isFAbsFree(VT))) &&
5453 N0.getNode()->hasOneUse() && VT.isInteger() &&
5454 !VT.isVector() && !N0.getValueType().isVector()) {
5455 SDValue NewConv = DAG.getNode(ISD::BITCAST, N0.getDebugLoc(), VT,
5457 AddToWorkList(NewConv.getNode());
5459 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
5460 if (N0.getOpcode() == ISD::FNEG)
5461 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT,
5462 NewConv, DAG.getConstant(SignBit, VT));
5463 assert(N0.getOpcode() == ISD::FABS);
5464 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
5465 NewConv, DAG.getConstant(~SignBit, VT));
5468 // fold (bitconvert (fcopysign cst, x)) ->
5469 // (or (and (bitconvert x), sign), (and cst, (not sign)))
5470 // Note that we don't handle (copysign x, cst) because this can always be
5471 // folded to an fneg or fabs.
5472 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse() &&
5473 isa<ConstantFPSDNode>(N0.getOperand(0)) &&
5474 VT.isInteger() && !VT.isVector()) {
5475 unsigned OrigXWidth = N0.getOperand(1).getValueType().getSizeInBits();
5476 EVT IntXVT = EVT::getIntegerVT(*DAG.getContext(), OrigXWidth);
5477 if (isTypeLegal(IntXVT)) {
5478 SDValue X = DAG.getNode(ISD::BITCAST, N0.getDebugLoc(),
5479 IntXVT, N0.getOperand(1));
5480 AddToWorkList(X.getNode());
5482 // If X has a different width than the result/lhs, sext it or truncate it.
5483 unsigned VTWidth = VT.getSizeInBits();
5484 if (OrigXWidth < VTWidth) {
5485 X = DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, X);
5486 AddToWorkList(X.getNode());
5487 } else if (OrigXWidth > VTWidth) {
5488 // To get the sign bit in the right place, we have to shift it right
5489 // before truncating.
5490 X = DAG.getNode(ISD::SRL, X.getDebugLoc(),
5491 X.getValueType(), X,
5492 DAG.getConstant(OrigXWidth-VTWidth, X.getValueType()));
5493 AddToWorkList(X.getNode());
5494 X = DAG.getNode(ISD::TRUNCATE, X.getDebugLoc(), VT, X);
5495 AddToWorkList(X.getNode());
5498 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
5499 X = DAG.getNode(ISD::AND, X.getDebugLoc(), VT,
5500 X, DAG.getConstant(SignBit, VT));
5501 AddToWorkList(X.getNode());
5503 SDValue Cst = DAG.getNode(ISD::BITCAST, N0.getDebugLoc(),
5504 VT, N0.getOperand(0));
5505 Cst = DAG.getNode(ISD::AND, Cst.getDebugLoc(), VT,
5506 Cst, DAG.getConstant(~SignBit, VT));
5507 AddToWorkList(Cst.getNode());
5509 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, X, Cst);
5513 // bitconvert(build_pair(ld, ld)) -> ld iff load locations are consecutive.
5514 if (N0.getOpcode() == ISD::BUILD_PAIR) {
5515 SDValue CombineLD = CombineConsecutiveLoads(N0.getNode(), VT);
5516 if (CombineLD.getNode())
5523 SDValue DAGCombiner::visitBUILD_PAIR(SDNode *N) {
5524 EVT VT = N->getValueType(0);
5525 return CombineConsecutiveLoads(N, VT);
5528 /// ConstantFoldBITCASTofBUILD_VECTOR - We know that BV is a build_vector
5529 /// node with Constant, ConstantFP or Undef operands. DstEltVT indicates the
5530 /// destination element value type.
5531 SDValue DAGCombiner::
5532 ConstantFoldBITCASTofBUILD_VECTOR(SDNode *BV, EVT DstEltVT) {
5533 EVT SrcEltVT = BV->getValueType(0).getVectorElementType();
5535 // If this is already the right type, we're done.
5536 if (SrcEltVT == DstEltVT) return SDValue(BV, 0);
5538 unsigned SrcBitSize = SrcEltVT.getSizeInBits();
5539 unsigned DstBitSize = DstEltVT.getSizeInBits();
5541 // If this is a conversion of N elements of one type to N elements of another
5542 // type, convert each element. This handles FP<->INT cases.
5543 if (SrcBitSize == DstBitSize) {
5544 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
5545 BV->getValueType(0).getVectorNumElements());
5547 // Due to the FP element handling below calling this routine recursively,
5548 // we can end up with a scalar-to-vector node here.
5549 if (BV->getOpcode() == ISD::SCALAR_TO_VECTOR)
5550 return DAG.getNode(ISD::SCALAR_TO_VECTOR, BV->getDebugLoc(), VT,
5551 DAG.getNode(ISD::BITCAST, BV->getDebugLoc(),
5552 DstEltVT, BV->getOperand(0)));
5554 SmallVector<SDValue, 8> Ops;
5555 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
5556 SDValue Op = BV->getOperand(i);
5557 // If the vector element type is not legal, the BUILD_VECTOR operands
5558 // are promoted and implicitly truncated. Make that explicit here.
5559 if (Op.getValueType() != SrcEltVT)
5560 Op = DAG.getNode(ISD::TRUNCATE, BV->getDebugLoc(), SrcEltVT, Op);
5561 Ops.push_back(DAG.getNode(ISD::BITCAST, BV->getDebugLoc(),
5563 AddToWorkList(Ops.back().getNode());
5565 return DAG.getNode(ISD::BUILD_VECTOR, BV->getDebugLoc(), VT,
5566 &Ops[0], Ops.size());
5569 // Otherwise, we're growing or shrinking the elements. To avoid having to
5570 // handle annoying details of growing/shrinking FP values, we convert them to
5572 if (SrcEltVT.isFloatingPoint()) {
5573 // Convert the input float vector to a int vector where the elements are the
5575 assert((SrcEltVT == MVT::f32 || SrcEltVT == MVT::f64) && "Unknown FP VT!");
5576 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), SrcEltVT.getSizeInBits());
5577 BV = ConstantFoldBITCASTofBUILD_VECTOR(BV, IntVT).getNode();
5581 // Now we know the input is an integer vector. If the output is a FP type,
5582 // convert to integer first, then to FP of the right size.
5583 if (DstEltVT.isFloatingPoint()) {
5584 assert((DstEltVT == MVT::f32 || DstEltVT == MVT::f64) && "Unknown FP VT!");
5585 EVT TmpVT = EVT::getIntegerVT(*DAG.getContext(), DstEltVT.getSizeInBits());
5586 SDNode *Tmp = ConstantFoldBITCASTofBUILD_VECTOR(BV, TmpVT).getNode();
5588 // Next, convert to FP elements of the same size.
5589 return ConstantFoldBITCASTofBUILD_VECTOR(Tmp, DstEltVT);
5592 // Okay, we know the src/dst types are both integers of differing types.
5593 // Handling growing first.
5594 assert(SrcEltVT.isInteger() && DstEltVT.isInteger());
5595 if (SrcBitSize < DstBitSize) {
5596 unsigned NumInputsPerOutput = DstBitSize/SrcBitSize;
5598 SmallVector<SDValue, 8> Ops;
5599 for (unsigned i = 0, e = BV->getNumOperands(); i != e;
5600 i += NumInputsPerOutput) {
5601 bool isLE = TLI.isLittleEndian();
5602 APInt NewBits = APInt(DstBitSize, 0);
5603 bool EltIsUndef = true;
5604 for (unsigned j = 0; j != NumInputsPerOutput; ++j) {
5605 // Shift the previously computed bits over.
5606 NewBits <<= SrcBitSize;
5607 SDValue Op = BV->getOperand(i+ (isLE ? (NumInputsPerOutput-j-1) : j));
5608 if (Op.getOpcode() == ISD::UNDEF) continue;
5611 NewBits |= cast<ConstantSDNode>(Op)->getAPIntValue().
5612 zextOrTrunc(SrcBitSize).zext(DstBitSize);
5616 Ops.push_back(DAG.getUNDEF(DstEltVT));
5618 Ops.push_back(DAG.getConstant(NewBits, DstEltVT));
5621 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, Ops.size());
5622 return DAG.getNode(ISD::BUILD_VECTOR, BV->getDebugLoc(), VT,
5623 &Ops[0], Ops.size());
5626 // Finally, this must be the case where we are shrinking elements: each input
5627 // turns into multiple outputs.
5628 bool isS2V = ISD::isScalarToVector(BV);
5629 unsigned NumOutputsPerInput = SrcBitSize/DstBitSize;
5630 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
5631 NumOutputsPerInput*BV->getNumOperands());
5632 SmallVector<SDValue, 8> Ops;
5634 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
5635 if (BV->getOperand(i).getOpcode() == ISD::UNDEF) {
5636 for (unsigned j = 0; j != NumOutputsPerInput; ++j)
5637 Ops.push_back(DAG.getUNDEF(DstEltVT));
5641 APInt OpVal = cast<ConstantSDNode>(BV->getOperand(i))->
5642 getAPIntValue().zextOrTrunc(SrcBitSize);
5644 for (unsigned j = 0; j != NumOutputsPerInput; ++j) {
5645 APInt ThisVal = OpVal.trunc(DstBitSize);
5646 Ops.push_back(DAG.getConstant(ThisVal, DstEltVT));
5647 if (isS2V && i == 0 && j == 0 && ThisVal.zext(SrcBitSize) == OpVal)
5648 // Simply turn this into a SCALAR_TO_VECTOR of the new type.
5649 return DAG.getNode(ISD::SCALAR_TO_VECTOR, BV->getDebugLoc(), VT,
5651 OpVal = OpVal.lshr(DstBitSize);
5654 // For big endian targets, swap the order of the pieces of each element.
5655 if (TLI.isBigEndian())
5656 std::reverse(Ops.end()-NumOutputsPerInput, Ops.end());
5659 return DAG.getNode(ISD::BUILD_VECTOR, BV->getDebugLoc(), VT,
5660 &Ops[0], Ops.size());
5663 SDValue DAGCombiner::visitFADD(SDNode *N) {
5664 SDValue N0 = N->getOperand(0);
5665 SDValue N1 = N->getOperand(1);
5666 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5667 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
5668 EVT VT = N->getValueType(0);
5671 if (VT.isVector()) {
5672 SDValue FoldedVOp = SimplifyVBinOp(N);
5673 if (FoldedVOp.getNode()) return FoldedVOp;
5676 // fold (fadd c1, c2) -> c1 + c2
5677 if (N0CFP && N1CFP && VT != MVT::ppcf128)
5678 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0, N1);
5679 // canonicalize constant to RHS
5680 if (N0CFP && !N1CFP)
5681 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N1, N0);
5682 // fold (fadd A, 0) -> A
5683 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
5684 N1CFP->getValueAPF().isZero())
5686 // fold (fadd A, (fneg B)) -> (fsub A, B)
5687 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
5688 isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options) == 2)
5689 return DAG.getNode(ISD::FSUB, N->getDebugLoc(), VT, N0,
5690 GetNegatedExpression(N1, DAG, LegalOperations));
5691 // fold (fadd (fneg A), B) -> (fsub B, A)
5692 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
5693 isNegatibleForFree(N0, LegalOperations, TLI, &DAG.getTarget().Options) == 2)
5694 return DAG.getNode(ISD::FSUB, N->getDebugLoc(), VT, N1,
5695 GetNegatedExpression(N0, DAG, LegalOperations));
5697 // If allowed, fold (fadd (fadd x, c1), c2) -> (fadd x, (fadd c1, c2))
5698 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
5699 N0.getOpcode() == ISD::FADD && N0.getNode()->hasOneUse() &&
5700 isa<ConstantFPSDNode>(N0.getOperand(1)))
5701 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0.getOperand(0),
5702 DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
5703 N0.getOperand(1), N1));
5705 // In unsafe math mode, we can fold chains of FADD's of the same value
5706 // into multiplications. This transform is not safe in general because
5707 // we are reducing the number of rounding steps.
5708 if (DAG.getTarget().Options.UnsafeFPMath &&
5709 TLI.isOperationLegalOrCustom(ISD::FMUL, VT) &&
5711 if (N0.getOpcode() == ISD::FMUL) {
5712 ConstantFPSDNode *CFP00 = dyn_cast<ConstantFPSDNode>(N0.getOperand(0));
5713 ConstantFPSDNode *CFP01 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
5715 // (fadd (fmul c, x), x) -> (fmul c+1, x)
5716 if (CFP00 && !CFP01 && N0.getOperand(1) == N1) {
5717 SDValue NewCFP = DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
5719 DAG.getConstantFP(1.0, VT));
5720 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5724 // (fadd (fmul x, c), x) -> (fmul c+1, x)
5725 if (CFP01 && !CFP00 && N0.getOperand(0) == N1) {
5726 SDValue NewCFP = DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
5728 DAG.getConstantFP(1.0, VT));
5729 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5733 // (fadd (fadd x, x), x) -> (fmul 3.0, x)
5734 if (!CFP00 && !CFP01 && N0.getOperand(0) == N0.getOperand(1) &&
5735 N0.getOperand(0) == N1) {
5736 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5737 N1, DAG.getConstantFP(3.0, VT));
5740 // (fadd (fmul c, x), (fadd x, x)) -> (fmul c+2, x)
5741 if (CFP00 && !CFP01 && N1.getOpcode() == ISD::FADD &&
5742 N1.getOperand(0) == N1.getOperand(1) &&
5743 N0.getOperand(1) == N1.getOperand(0)) {
5744 SDValue NewCFP = DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
5746 DAG.getConstantFP(2.0, VT));
5747 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5748 N0.getOperand(1), NewCFP);
5751 // (fadd (fmul x, c), (fadd x, x)) -> (fmul c+2, x)
5752 if (CFP01 && !CFP00 && N1.getOpcode() == ISD::FADD &&
5753 N1.getOperand(0) == N1.getOperand(1) &&
5754 N0.getOperand(0) == N1.getOperand(0)) {
5755 SDValue NewCFP = DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
5757 DAG.getConstantFP(2.0, VT));
5758 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5759 N0.getOperand(0), NewCFP);
5763 if (N1.getOpcode() == ISD::FMUL) {
5764 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0));
5765 ConstantFPSDNode *CFP11 = dyn_cast<ConstantFPSDNode>(N1.getOperand(1));
5767 // (fadd x, (fmul c, x)) -> (fmul c+1, x)
5768 if (CFP10 && !CFP11 && N1.getOperand(1) == N0) {
5769 SDValue NewCFP = DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
5771 DAG.getConstantFP(1.0, VT));
5772 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5776 // (fadd x, (fmul x, c)) -> (fmul c+1, x)
5777 if (CFP11 && !CFP10 && N1.getOperand(0) == N0) {
5778 SDValue NewCFP = DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
5780 DAG.getConstantFP(1.0, VT));
5781 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5785 // (fadd x, (fadd x, x)) -> (fmul 3.0, x)
5786 if (!CFP10 && !CFP11 && N1.getOperand(0) == N1.getOperand(1) &&
5787 N1.getOperand(0) == N0) {
5788 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5789 N0, DAG.getConstantFP(3.0, VT));
5792 // (fadd (fadd x, x), (fmul c, x)) -> (fmul c+2, x)
5793 if (CFP10 && !CFP11 && N1.getOpcode() == ISD::FADD &&
5794 N1.getOperand(0) == N1.getOperand(1) &&
5795 N0.getOperand(1) == N1.getOperand(0)) {
5796 SDValue NewCFP = DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
5798 DAG.getConstantFP(2.0, VT));
5799 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5800 N0.getOperand(1), NewCFP);
5803 // (fadd (fadd x, x), (fmul x, c)) -> (fmul c+2, x)
5804 if (CFP11 && !CFP10 && N1.getOpcode() == ISD::FADD &&
5805 N1.getOperand(0) == N1.getOperand(1) &&
5806 N0.getOperand(0) == N1.getOperand(0)) {
5807 SDValue NewCFP = DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
5809 DAG.getConstantFP(2.0, VT));
5810 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5811 N0.getOperand(0), NewCFP);
5815 // (fadd (fadd x, x), (fadd x, x)) -> (fmul 4.0, x)
5816 if (N0.getOpcode() == ISD::FADD && N1.getOpcode() == ISD::FADD &&
5817 N0.getOperand(0) == N0.getOperand(1) &&
5818 N1.getOperand(0) == N1.getOperand(1) &&
5819 N0.getOperand(0) == N1.getOperand(0)) {
5820 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5822 DAG.getConstantFP(4.0, VT));
5826 // FADD -> FMA combines:
5827 if ((DAG.getTarget().Options.AllowFPOpFusion == FPOpFusion::Fast ||
5828 DAG.getTarget().Options.UnsafeFPMath) &&
5829 DAG.getTarget().getTargetLowering()->isFMAFasterThanMulAndAdd(VT) &&
5830 TLI.isOperationLegalOrCustom(ISD::FMA, VT)) {
5832 // fold (fadd (fmul x, y), z) -> (fma x, y, z)
5833 if (N0.getOpcode() == ISD::FMUL && N0->hasOneUse()) {
5834 return DAG.getNode(ISD::FMA, N->getDebugLoc(), VT,
5835 N0.getOperand(0), N0.getOperand(1), N1);
5838 // fold (fadd x, (fmul y, z)) -> (fma y, z, x)
5839 // Note: Commutes FADD operands.
5840 if (N1.getOpcode() == ISD::FMUL && N1->hasOneUse()) {
5841 return DAG.getNode(ISD::FMA, N->getDebugLoc(), VT,
5842 N1.getOperand(0), N1.getOperand(1), N0);
5849 SDValue DAGCombiner::visitFSUB(SDNode *N) {
5850 SDValue N0 = N->getOperand(0);
5851 SDValue N1 = N->getOperand(1);
5852 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5853 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
5854 EVT VT = N->getValueType(0);
5855 DebugLoc dl = N->getDebugLoc();
5858 if (VT.isVector()) {
5859 SDValue FoldedVOp = SimplifyVBinOp(N);
5860 if (FoldedVOp.getNode()) return FoldedVOp;
5863 // fold (fsub c1, c2) -> c1-c2
5864 if (N0CFP && N1CFP && VT != MVT::ppcf128)
5865 return DAG.getNode(ISD::FSUB, N->getDebugLoc(), VT, N0, N1);
5866 // fold (fsub A, 0) -> A
5867 if (DAG.getTarget().Options.UnsafeFPMath &&
5868 N1CFP && N1CFP->getValueAPF().isZero())
5870 // fold (fsub 0, B) -> -B
5871 if (DAG.getTarget().Options.UnsafeFPMath &&
5872 N0CFP && N0CFP->getValueAPF().isZero()) {
5873 if (isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options))
5874 return GetNegatedExpression(N1, DAG, LegalOperations);
5875 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
5876 return DAG.getNode(ISD::FNEG, dl, VT, N1);
5878 // fold (fsub A, (fneg B)) -> (fadd A, B)
5879 if (isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options))
5880 return DAG.getNode(ISD::FADD, dl, VT, N0,
5881 GetNegatedExpression(N1, DAG, LegalOperations));
5883 // If 'unsafe math' is enabled, fold
5884 // (fsub x, x) -> 0.0 &
5885 // (fsub x, (fadd x, y)) -> (fneg y) &
5886 // (fsub x, (fadd y, x)) -> (fneg y)
5887 if (DAG.getTarget().Options.UnsafeFPMath) {
5889 return DAG.getConstantFP(0.0f, VT);
5891 if (N1.getOpcode() == ISD::FADD) {
5892 SDValue N10 = N1->getOperand(0);
5893 SDValue N11 = N1->getOperand(1);
5895 if (N10 == N0 && isNegatibleForFree(N11, LegalOperations, TLI,
5896 &DAG.getTarget().Options))
5897 return GetNegatedExpression(N11, DAG, LegalOperations);
5898 else if (N11 == N0 && isNegatibleForFree(N10, LegalOperations, TLI,
5899 &DAG.getTarget().Options))
5900 return GetNegatedExpression(N10, DAG, LegalOperations);
5904 // FSUB -> FMA combines:
5905 if ((DAG.getTarget().Options.AllowFPOpFusion == FPOpFusion::Fast ||
5906 DAG.getTarget().Options.UnsafeFPMath) &&
5907 DAG.getTarget().getTargetLowering()->isFMAFasterThanMulAndAdd(VT) &&
5908 TLI.isOperationLegalOrCustom(ISD::FMA, VT)) {
5910 // fold (fsub (fmul x, y), z) -> (fma x, y, (fneg z))
5911 if (N0.getOpcode() == ISD::FMUL && N0->hasOneUse()) {
5912 return DAG.getNode(ISD::FMA, dl, VT,
5913 N0.getOperand(0), N0.getOperand(1),
5914 DAG.getNode(ISD::FNEG, dl, VT, N1));
5917 // fold (fsub x, (fmul y, z)) -> (fma (fneg y), z, x)
5918 // Note: Commutes FSUB operands.
5919 if (N1.getOpcode() == ISD::FMUL && N1->hasOneUse()) {
5920 return DAG.getNode(ISD::FMA, dl, VT,
5921 DAG.getNode(ISD::FNEG, dl, VT,
5923 N1.getOperand(1), N0);
5926 // fold (fsub (-(fmul, x, y)), z) -> (fma (fneg x), y, (fneg z))
5927 if (N0.getOpcode() == ISD::FNEG &&
5928 N0.getOperand(0).getOpcode() == ISD::FMUL &&
5929 N0->hasOneUse() && N0.getOperand(0).hasOneUse()) {
5930 SDValue N00 = N0.getOperand(0).getOperand(0);
5931 SDValue N01 = N0.getOperand(0).getOperand(1);
5932 return DAG.getNode(ISD::FMA, dl, VT,
5933 DAG.getNode(ISD::FNEG, dl, VT, N00), N01,
5934 DAG.getNode(ISD::FNEG, dl, VT, N1));
5941 SDValue DAGCombiner::visitFMUL(SDNode *N) {
5942 SDValue N0 = N->getOperand(0);
5943 SDValue N1 = N->getOperand(1);
5944 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5945 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
5946 EVT VT = N->getValueType(0);
5947 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5950 if (VT.isVector()) {
5951 SDValue FoldedVOp = SimplifyVBinOp(N);
5952 if (FoldedVOp.getNode()) return FoldedVOp;
5955 // fold (fmul c1, c2) -> c1*c2
5956 if (N0CFP && N1CFP && VT != MVT::ppcf128)
5957 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT, N0, N1);
5958 // canonicalize constant to RHS
5959 if (N0CFP && !N1CFP)
5960 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT, N1, N0);
5961 // fold (fmul A, 0) -> 0
5962 if (DAG.getTarget().Options.UnsafeFPMath &&
5963 N1CFP && N1CFP->getValueAPF().isZero())
5965 // fold (fmul A, 0) -> 0, vector edition.
5966 if (DAG.getTarget().Options.UnsafeFPMath &&
5967 ISD::isBuildVectorAllZeros(N1.getNode()))
5969 // fold (fmul A, 1.0) -> A
5970 if (N1CFP && N1CFP->isExactlyValue(1.0))
5972 // fold (fmul X, 2.0) -> (fadd X, X)
5973 if (N1CFP && N1CFP->isExactlyValue(+2.0))
5974 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0, N0);
5975 // fold (fmul X, -1.0) -> (fneg X)
5976 if (N1CFP && N1CFP->isExactlyValue(-1.0))
5977 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
5978 return DAG.getNode(ISD::FNEG, N->getDebugLoc(), VT, N0);
5980 // fold (fmul (fneg X), (fneg Y)) -> (fmul X, Y)
5981 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI,
5982 &DAG.getTarget().Options)) {
5983 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI,
5984 &DAG.getTarget().Options)) {
5985 // Both can be negated for free, check to see if at least one is cheaper
5987 if (LHSNeg == 2 || RHSNeg == 2)
5988 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5989 GetNegatedExpression(N0, DAG, LegalOperations),
5990 GetNegatedExpression(N1, DAG, LegalOperations));
5994 // If allowed, fold (fmul (fmul x, c1), c2) -> (fmul x, (fmul c1, c2))
5995 if (DAG.getTarget().Options.UnsafeFPMath &&
5996 N1CFP && N0.getOpcode() == ISD::FMUL &&
5997 N0.getNode()->hasOneUse() && isa<ConstantFPSDNode>(N0.getOperand(1)))
5998 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT, N0.getOperand(0),
5999 DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
6000 N0.getOperand(1), N1));
6005 SDValue DAGCombiner::visitFMA(SDNode *N) {
6006 SDValue N0 = N->getOperand(0);
6007 SDValue N1 = N->getOperand(1);
6008 SDValue N2 = N->getOperand(2);
6009 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6010 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6011 EVT VT = N->getValueType(0);
6012 DebugLoc dl = N->getDebugLoc();
6014 if (N0CFP && N0CFP->isExactlyValue(1.0))
6015 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N1, N2);
6016 if (N1CFP && N1CFP->isExactlyValue(1.0))
6017 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0, N2);
6019 // Canonicalize (fma c, x, y) -> (fma x, c, y)
6020 if (N0CFP && !N1CFP)
6021 return DAG.getNode(ISD::FMA, N->getDebugLoc(), VT, N1, N0, N2);
6023 // (fma x, c1, (fmul x, c2)) -> (fmul x, c1+c2)
6024 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
6025 N2.getOpcode() == ISD::FMUL &&
6026 N0 == N2.getOperand(0) &&
6027 N2.getOperand(1).getOpcode() == ISD::ConstantFP) {
6028 return DAG.getNode(ISD::FMUL, dl, VT, N0,
6029 DAG.getNode(ISD::FADD, dl, VT, N1, N2.getOperand(1)));
6033 // (fma (fmul x, c1), c2, y) -> (fma x, c1*c2, y)
6034 if (DAG.getTarget().Options.UnsafeFPMath &&
6035 N0.getOpcode() == ISD::FMUL && N1CFP &&
6036 N0.getOperand(1).getOpcode() == ISD::ConstantFP) {
6037 return DAG.getNode(ISD::FMA, dl, VT,
6039 DAG.getNode(ISD::FMUL, dl, VT, N1, N0.getOperand(1)),
6043 // (fma x, 1, y) -> (fadd x, y)
6044 // (fma x, -1, y) -> (fadd (fneg x), y)
6046 if (N1CFP->isExactlyValue(1.0))
6047 return DAG.getNode(ISD::FADD, dl, VT, N0, N2);
6049 if (N1CFP->isExactlyValue(-1.0) &&
6050 (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))) {
6051 SDValue RHSNeg = DAG.getNode(ISD::FNEG, dl, VT, N0);
6052 AddToWorkList(RHSNeg.getNode());
6053 return DAG.getNode(ISD::FADD, dl, VT, N2, RHSNeg);
6057 // (fma x, c, x) -> (fmul x, (c+1))
6058 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP && N0 == N2) {
6059 return DAG.getNode(ISD::FMUL, dl, VT,
6061 DAG.getNode(ISD::FADD, dl, VT,
6062 N1, DAG.getConstantFP(1.0, VT)));
6065 // (fma x, c, (fneg x)) -> (fmul x, (c-1))
6066 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
6067 N2.getOpcode() == ISD::FNEG && N2.getOperand(0) == N0) {
6068 return DAG.getNode(ISD::FMUL, dl, VT,
6070 DAG.getNode(ISD::FADD, dl, VT,
6071 N1, DAG.getConstantFP(-1.0, VT)));
6078 SDValue DAGCombiner::visitFDIV(SDNode *N) {
6079 SDValue N0 = N->getOperand(0);
6080 SDValue N1 = N->getOperand(1);
6081 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6082 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6083 EVT VT = N->getValueType(0);
6084 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6087 if (VT.isVector()) {
6088 SDValue FoldedVOp = SimplifyVBinOp(N);
6089 if (FoldedVOp.getNode()) return FoldedVOp;
6092 // fold (fdiv c1, c2) -> c1/c2
6093 if (N0CFP && N1CFP && VT != MVT::ppcf128)
6094 return DAG.getNode(ISD::FDIV, N->getDebugLoc(), VT, N0, N1);
6096 // fold (fdiv X, c2) -> fmul X, 1/c2 if losing precision is acceptable.
6097 if (N1CFP && VT != MVT::ppcf128 && DAG.getTarget().Options.UnsafeFPMath) {
6098 // Compute the reciprocal 1.0 / c2.
6099 APFloat N1APF = N1CFP->getValueAPF();
6100 APFloat Recip(N1APF.getSemantics(), 1); // 1.0
6101 APFloat::opStatus st = Recip.divide(N1APF, APFloat::rmNearestTiesToEven);
6102 // Only do the transform if the reciprocal is a legal fp immediate that
6103 // isn't too nasty (eg NaN, denormal, ...).
6104 if ((st == APFloat::opOK || st == APFloat::opInexact) && // Not too nasty
6105 (!LegalOperations ||
6106 // FIXME: custom lowering of ConstantFP might fail (see e.g. ARM
6107 // backend)... we should handle this gracefully after Legalize.
6108 // TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT) ||
6109 TLI.isOperationLegal(llvm::ISD::ConstantFP, VT) ||
6110 TLI.isFPImmLegal(Recip, VT)))
6111 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT, N0,
6112 DAG.getConstantFP(Recip, VT));
6115 // (fdiv (fneg X), (fneg Y)) -> (fdiv X, Y)
6116 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI,
6117 &DAG.getTarget().Options)) {
6118 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI,
6119 &DAG.getTarget().Options)) {
6120 // Both can be negated for free, check to see if at least one is cheaper
6122 if (LHSNeg == 2 || RHSNeg == 2)
6123 return DAG.getNode(ISD::FDIV, N->getDebugLoc(), VT,
6124 GetNegatedExpression(N0, DAG, LegalOperations),
6125 GetNegatedExpression(N1, DAG, LegalOperations));
6132 SDValue DAGCombiner::visitFREM(SDNode *N) {
6133 SDValue N0 = N->getOperand(0);
6134 SDValue N1 = N->getOperand(1);
6135 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6136 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6137 EVT VT = N->getValueType(0);
6139 // fold (frem c1, c2) -> fmod(c1,c2)
6140 if (N0CFP && N1CFP && VT != MVT::ppcf128)
6141 return DAG.getNode(ISD::FREM, N->getDebugLoc(), VT, N0, N1);
6146 SDValue DAGCombiner::visitFCOPYSIGN(SDNode *N) {
6147 SDValue N0 = N->getOperand(0);
6148 SDValue N1 = N->getOperand(1);
6149 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6150 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6151 EVT VT = N->getValueType(0);
6153 if (N0CFP && N1CFP && VT != MVT::ppcf128) // Constant fold
6154 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT, N0, N1);
6157 const APFloat& V = N1CFP->getValueAPF();
6158 // copysign(x, c1) -> fabs(x) iff ispos(c1)
6159 // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1)
6160 if (!V.isNegative()) {
6161 if (!LegalOperations || TLI.isOperationLegal(ISD::FABS, VT))
6162 return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0);
6164 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
6165 return DAG.getNode(ISD::FNEG, N->getDebugLoc(), VT,
6166 DAG.getNode(ISD::FABS, N0.getDebugLoc(), VT, N0));
6170 // copysign(fabs(x), y) -> copysign(x, y)
6171 // copysign(fneg(x), y) -> copysign(x, y)
6172 // copysign(copysign(x,z), y) -> copysign(x, y)
6173 if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG ||
6174 N0.getOpcode() == ISD::FCOPYSIGN)
6175 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
6176 N0.getOperand(0), N1);
6178 // copysign(x, abs(y)) -> abs(x)
6179 if (N1.getOpcode() == ISD::FABS)
6180 return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0);
6182 // copysign(x, copysign(y,z)) -> copysign(x, z)
6183 if (N1.getOpcode() == ISD::FCOPYSIGN)
6184 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
6185 N0, N1.getOperand(1));
6187 // copysign(x, fp_extend(y)) -> copysign(x, y)
6188 // copysign(x, fp_round(y)) -> copysign(x, y)
6189 if (N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND)
6190 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
6191 N0, N1.getOperand(0));
6196 SDValue DAGCombiner::visitSINT_TO_FP(SDNode *N) {
6197 SDValue N0 = N->getOperand(0);
6198 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
6199 EVT VT = N->getValueType(0);
6200 EVT OpVT = N0.getValueType();
6202 // fold (sint_to_fp c1) -> c1fp
6203 if (N0C && OpVT != MVT::ppcf128 &&
6204 // ...but only if the target supports immediate floating-point values
6205 (!LegalOperations ||
6206 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
6207 return DAG.getNode(ISD::SINT_TO_FP, N->getDebugLoc(), VT, N0);
6209 // If the input is a legal type, and SINT_TO_FP is not legal on this target,
6210 // but UINT_TO_FP is legal on this target, try to convert.
6211 if (!TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT) &&
6212 TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT)) {
6213 // If the sign bit is known to be zero, we can change this to UINT_TO_FP.
6214 if (DAG.SignBitIsZero(N0))
6215 return DAG.getNode(ISD::UINT_TO_FP, N->getDebugLoc(), VT, N0);
6218 // The next optimizations are desireable only if SELECT_CC can be lowered.
6219 // Check against MVT::Other for SELECT_CC, which is a workaround for targets
6220 // having to say they don't support SELECT_CC on every type the DAG knows
6221 // about, since there is no way to mark an opcode illegal at all value types
6222 // (See also visitSELECT)
6223 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, MVT::Other)) {
6224 // fold (sint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
6225 if (N0.getOpcode() == ISD::SETCC && N0.getValueType() == MVT::i1 &&
6227 (!LegalOperations ||
6228 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
6230 { N0.getOperand(0), N0.getOperand(1),
6231 DAG.getConstantFP(-1.0, VT) , DAG.getConstantFP(0.0, VT),
6233 return DAG.getNode(ISD::SELECT_CC, N->getDebugLoc(), VT, Ops, 5);
6236 // fold (sint_to_fp (zext (setcc x, y, cc))) ->
6237 // (select_cc x, y, 1.0, 0.0,, cc)
6238 if (N0.getOpcode() == ISD::ZERO_EXTEND &&
6239 N0.getOperand(0).getOpcode() == ISD::SETCC &&!VT.isVector() &&
6240 (!LegalOperations ||
6241 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
6243 { N0.getOperand(0).getOperand(0), N0.getOperand(0).getOperand(1),
6244 DAG.getConstantFP(1.0, VT) , DAG.getConstantFP(0.0, VT),
6245 N0.getOperand(0).getOperand(2) };
6246 return DAG.getNode(ISD::SELECT_CC, N->getDebugLoc(), VT, Ops, 5);
6253 SDValue DAGCombiner::visitUINT_TO_FP(SDNode *N) {
6254 SDValue N0 = N->getOperand(0);
6255 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
6256 EVT VT = N->getValueType(0);
6257 EVT OpVT = N0.getValueType();
6259 // fold (uint_to_fp c1) -> c1fp
6260 if (N0C && OpVT != MVT::ppcf128 &&
6261 // ...but only if the target supports immediate floating-point values
6262 (!LegalOperations ||
6263 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
6264 return DAG.getNode(ISD::UINT_TO_FP, N->getDebugLoc(), VT, N0);
6266 // If the input is a legal type, and UINT_TO_FP is not legal on this target,
6267 // but SINT_TO_FP is legal on this target, try to convert.
6268 if (!TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT) &&
6269 TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT)) {
6270 // If the sign bit is known to be zero, we can change this to SINT_TO_FP.
6271 if (DAG.SignBitIsZero(N0))
6272 return DAG.getNode(ISD::SINT_TO_FP, N->getDebugLoc(), VT, N0);
6275 // The next optimizations are desireable only if SELECT_CC can be lowered.
6276 // Check against MVT::Other for SELECT_CC, which is a workaround for targets
6277 // having to say they don't support SELECT_CC on every type the DAG knows
6278 // about, since there is no way to mark an opcode illegal at all value types
6279 // (See also visitSELECT)
6280 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, MVT::Other)) {
6281 // fold (uint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
6283 if (N0.getOpcode() == ISD::SETCC && !VT.isVector() &&
6284 (!LegalOperations ||
6285 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
6287 { N0.getOperand(0), N0.getOperand(1),
6288 DAG.getConstantFP(1.0, VT), DAG.getConstantFP(0.0, VT),
6290 return DAG.getNode(ISD::SELECT_CC, N->getDebugLoc(), VT, Ops, 5);
6297 SDValue DAGCombiner::visitFP_TO_SINT(SDNode *N) {
6298 SDValue N0 = N->getOperand(0);
6299 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6300 EVT VT = N->getValueType(0);
6302 // fold (fp_to_sint c1fp) -> c1
6304 return DAG.getNode(ISD::FP_TO_SINT, N->getDebugLoc(), VT, N0);
6309 SDValue DAGCombiner::visitFP_TO_UINT(SDNode *N) {
6310 SDValue N0 = N->getOperand(0);
6311 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6312 EVT VT = N->getValueType(0);
6314 // fold (fp_to_uint c1fp) -> c1
6315 if (N0CFP && VT != MVT::ppcf128)
6316 return DAG.getNode(ISD::FP_TO_UINT, N->getDebugLoc(), VT, N0);
6321 SDValue DAGCombiner::visitFP_ROUND(SDNode *N) {
6322 SDValue N0 = N->getOperand(0);
6323 SDValue N1 = N->getOperand(1);
6324 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6325 EVT VT = N->getValueType(0);
6327 // fold (fp_round c1fp) -> c1fp
6328 if (N0CFP && N0.getValueType() != MVT::ppcf128)
6329 return DAG.getNode(ISD::FP_ROUND, N->getDebugLoc(), VT, N0, N1);
6331 // fold (fp_round (fp_extend x)) -> x
6332 if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType())
6333 return N0.getOperand(0);
6335 // fold (fp_round (fp_round x)) -> (fp_round x)
6336 if (N0.getOpcode() == ISD::FP_ROUND) {
6337 // This is a value preserving truncation if both round's are.
6338 bool IsTrunc = N->getConstantOperandVal(1) == 1 &&
6339 N0.getNode()->getConstantOperandVal(1) == 1;
6340 return DAG.getNode(ISD::FP_ROUND, N->getDebugLoc(), VT, N0.getOperand(0),
6341 DAG.getIntPtrConstant(IsTrunc));
6344 // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y)
6345 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse()) {
6346 SDValue Tmp = DAG.getNode(ISD::FP_ROUND, N0.getDebugLoc(), VT,
6347 N0.getOperand(0), N1);
6348 AddToWorkList(Tmp.getNode());
6349 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
6350 Tmp, N0.getOperand(1));
6356 SDValue DAGCombiner::visitFP_ROUND_INREG(SDNode *N) {
6357 SDValue N0 = N->getOperand(0);
6358 EVT VT = N->getValueType(0);
6359 EVT EVT = cast<VTSDNode>(N->getOperand(1))->getVT();
6360 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6362 // fold (fp_round_inreg c1fp) -> c1fp
6363 if (N0CFP && isTypeLegal(EVT)) {
6364 SDValue Round = DAG.getConstantFP(*N0CFP->getConstantFPValue(), EVT);
6365 return DAG.getNode(ISD::FP_EXTEND, N->getDebugLoc(), VT, Round);
6371 SDValue DAGCombiner::visitFP_EXTEND(SDNode *N) {
6372 SDValue N0 = N->getOperand(0);
6373 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6374 EVT VT = N->getValueType(0);
6376 // If this is fp_round(fpextend), don't fold it, allow ourselves to be folded.
6377 if (N->hasOneUse() &&
6378 N->use_begin()->getOpcode() == ISD::FP_ROUND)
6381 // fold (fp_extend c1fp) -> c1fp
6382 if (N0CFP && VT != MVT::ppcf128)
6383 return DAG.getNode(ISD::FP_EXTEND, N->getDebugLoc(), VT, N0);
6385 // Turn fp_extend(fp_round(X, 1)) -> x since the fp_round doesn't affect the
6387 if (N0.getOpcode() == ISD::FP_ROUND
6388 && N0.getNode()->getConstantOperandVal(1) == 1) {
6389 SDValue In = N0.getOperand(0);
6390 if (In.getValueType() == VT) return In;
6391 if (VT.bitsLT(In.getValueType()))
6392 return DAG.getNode(ISD::FP_ROUND, N->getDebugLoc(), VT,
6393 In, N0.getOperand(1));
6394 return DAG.getNode(ISD::FP_EXTEND, N->getDebugLoc(), VT, In);
6397 // fold (fpext (load x)) -> (fpext (fptrunc (extload x)))
6398 if (ISD::isNON_EXTLoad(N0.getNode()) && N0.hasOneUse() &&
6399 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
6400 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType()))) {
6401 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6402 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, N->getDebugLoc(), VT,
6404 LN0->getBasePtr(), LN0->getPointerInfo(),
6406 LN0->isVolatile(), LN0->isNonTemporal(),
6407 LN0->getAlignment());
6408 CombineTo(N, ExtLoad);
6409 CombineTo(N0.getNode(),
6410 DAG.getNode(ISD::FP_ROUND, N0.getDebugLoc(),
6411 N0.getValueType(), ExtLoad, DAG.getIntPtrConstant(1)),
6412 ExtLoad.getValue(1));
6413 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6419 SDValue DAGCombiner::visitFNEG(SDNode *N) {
6420 SDValue N0 = N->getOperand(0);
6421 EVT VT = N->getValueType(0);
6423 if (VT.isVector()) {
6424 SDValue FoldedVOp = SimplifyVUnaryOp(N);
6425 if (FoldedVOp.getNode()) return FoldedVOp;
6428 if (isNegatibleForFree(N0, LegalOperations, DAG.getTargetLoweringInfo(),
6429 &DAG.getTarget().Options))
6430 return GetNegatedExpression(N0, DAG, LegalOperations);
6432 // Transform fneg(bitconvert(x)) -> bitconvert(x^sign) to avoid loading
6433 // constant pool values.
6434 if (!TLI.isFNegFree(VT) && N0.getOpcode() == ISD::BITCAST &&
6436 N0.getNode()->hasOneUse() &&
6437 N0.getOperand(0).getValueType().isInteger()) {
6438 SDValue Int = N0.getOperand(0);
6439 EVT IntVT = Int.getValueType();
6440 if (IntVT.isInteger() && !IntVT.isVector()) {
6441 Int = DAG.getNode(ISD::XOR, N0.getDebugLoc(), IntVT, Int,
6442 DAG.getConstant(APInt::getSignBit(IntVT.getSizeInBits()), IntVT));
6443 AddToWorkList(Int.getNode());
6444 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
6449 // (fneg (fmul c, x)) -> (fmul -c, x)
6450 if (N0.getOpcode() == ISD::FMUL) {
6451 ConstantFPSDNode *CFP1 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
6453 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
6455 DAG.getNode(ISD::FNEG, N->getDebugLoc(), VT,
6463 SDValue DAGCombiner::visitFCEIL(SDNode *N) {
6464 SDValue N0 = N->getOperand(0);
6465 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6466 EVT VT = N->getValueType(0);
6468 // fold (fceil c1) -> fceil(c1)
6469 if (N0CFP && VT != MVT::ppcf128)
6470 return DAG.getNode(ISD::FCEIL, N->getDebugLoc(), VT, N0);
6475 SDValue DAGCombiner::visitFTRUNC(SDNode *N) {
6476 SDValue N0 = N->getOperand(0);
6477 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6478 EVT VT = N->getValueType(0);
6480 // fold (ftrunc c1) -> ftrunc(c1)
6481 if (N0CFP && VT != MVT::ppcf128)
6482 return DAG.getNode(ISD::FTRUNC, N->getDebugLoc(), VT, N0);
6487 SDValue DAGCombiner::visitFFLOOR(SDNode *N) {
6488 SDValue N0 = N->getOperand(0);
6489 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6490 EVT VT = N->getValueType(0);
6492 // fold (ffloor c1) -> ffloor(c1)
6493 if (N0CFP && VT != MVT::ppcf128)
6494 return DAG.getNode(ISD::FFLOOR, N->getDebugLoc(), VT, N0);
6499 SDValue DAGCombiner::visitFABS(SDNode *N) {
6500 SDValue N0 = N->getOperand(0);
6501 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6502 EVT VT = N->getValueType(0);
6504 if (VT.isVector()) {
6505 SDValue FoldedVOp = SimplifyVUnaryOp(N);
6506 if (FoldedVOp.getNode()) return FoldedVOp;
6509 // fold (fabs c1) -> fabs(c1)
6510 if (N0CFP && VT != MVT::ppcf128)
6511 return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0);
6512 // fold (fabs (fabs x)) -> (fabs x)
6513 if (N0.getOpcode() == ISD::FABS)
6514 return N->getOperand(0);
6515 // fold (fabs (fneg x)) -> (fabs x)
6516 // fold (fabs (fcopysign x, y)) -> (fabs x)
6517 if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN)
6518 return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0.getOperand(0));
6520 // Transform fabs(bitconvert(x)) -> bitconvert(x&~sign) to avoid loading
6521 // constant pool values.
6522 if (!TLI.isFAbsFree(VT) &&
6523 N0.getOpcode() == ISD::BITCAST && N0.getNode()->hasOneUse() &&
6524 N0.getOperand(0).getValueType().isInteger() &&
6525 !N0.getOperand(0).getValueType().isVector()) {
6526 SDValue Int = N0.getOperand(0);
6527 EVT IntVT = Int.getValueType();
6528 if (IntVT.isInteger() && !IntVT.isVector()) {
6529 Int = DAG.getNode(ISD::AND, N0.getDebugLoc(), IntVT, Int,
6530 DAG.getConstant(~APInt::getSignBit(IntVT.getSizeInBits()), IntVT));
6531 AddToWorkList(Int.getNode());
6532 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
6533 N->getValueType(0), Int);
6540 SDValue DAGCombiner::visitBRCOND(SDNode *N) {
6541 SDValue Chain = N->getOperand(0);
6542 SDValue N1 = N->getOperand(1);
6543 SDValue N2 = N->getOperand(2);
6545 // If N is a constant we could fold this into a fallthrough or unconditional
6546 // branch. However that doesn't happen very often in normal code, because
6547 // Instcombine/SimplifyCFG should have handled the available opportunities.
6548 // If we did this folding here, it would be necessary to update the
6549 // MachineBasicBlock CFG, which is awkward.
6551 // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal
6553 if (N1.getOpcode() == ISD::SETCC &&
6554 TLI.isOperationLegalOrCustom(ISD::BR_CC, MVT::Other)) {
6555 return DAG.getNode(ISD::BR_CC, N->getDebugLoc(), MVT::Other,
6556 Chain, N1.getOperand(2),
6557 N1.getOperand(0), N1.getOperand(1), N2);
6560 if ((N1.hasOneUse() && N1.getOpcode() == ISD::SRL) ||
6561 ((N1.getOpcode() == ISD::TRUNCATE && N1.hasOneUse()) &&
6562 (N1.getOperand(0).hasOneUse() &&
6563 N1.getOperand(0).getOpcode() == ISD::SRL))) {
6565 if (N1.getOpcode() == ISD::TRUNCATE) {
6566 // Look pass the truncate.
6567 Trunc = N1.getNode();
6568 N1 = N1.getOperand(0);
6571 // Match this pattern so that we can generate simpler code:
6574 // %b = and i32 %a, 2
6575 // %c = srl i32 %b, 1
6576 // brcond i32 %c ...
6581 // %b = and i32 %a, 2
6582 // %c = setcc eq %b, 0
6585 // This applies only when the AND constant value has one bit set and the
6586 // SRL constant is equal to the log2 of the AND constant. The back-end is
6587 // smart enough to convert the result into a TEST/JMP sequence.
6588 SDValue Op0 = N1.getOperand(0);
6589 SDValue Op1 = N1.getOperand(1);
6591 if (Op0.getOpcode() == ISD::AND &&
6592 Op1.getOpcode() == ISD::Constant) {
6593 SDValue AndOp1 = Op0.getOperand(1);
6595 if (AndOp1.getOpcode() == ISD::Constant) {
6596 const APInt &AndConst = cast<ConstantSDNode>(AndOp1)->getAPIntValue();
6598 if (AndConst.isPowerOf2() &&
6599 cast<ConstantSDNode>(Op1)->getAPIntValue()==AndConst.logBase2()) {
6601 DAG.getSetCC(N->getDebugLoc(),
6602 TLI.getSetCCResultType(Op0.getValueType()),
6603 Op0, DAG.getConstant(0, Op0.getValueType()),
6606 SDValue NewBRCond = DAG.getNode(ISD::BRCOND, N->getDebugLoc(),
6607 MVT::Other, Chain, SetCC, N2);
6608 // Don't add the new BRCond into the worklist or else SimplifySelectCC
6609 // will convert it back to (X & C1) >> C2.
6610 CombineTo(N, NewBRCond, false);
6611 // Truncate is dead.
6613 removeFromWorkList(Trunc);
6614 DAG.DeleteNode(Trunc);
6616 // Replace the uses of SRL with SETCC
6617 WorkListRemover DeadNodes(*this);
6618 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
6619 removeFromWorkList(N1.getNode());
6620 DAG.DeleteNode(N1.getNode());
6621 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6627 // Restore N1 if the above transformation doesn't match.
6628 N1 = N->getOperand(1);
6631 // Transform br(xor(x, y)) -> br(x != y)
6632 // Transform br(xor(xor(x,y), 1)) -> br (x == y)
6633 if (N1.hasOneUse() && N1.getOpcode() == ISD::XOR) {
6634 SDNode *TheXor = N1.getNode();
6635 SDValue Op0 = TheXor->getOperand(0);
6636 SDValue Op1 = TheXor->getOperand(1);
6637 if (Op0.getOpcode() == Op1.getOpcode()) {
6638 // Avoid missing important xor optimizations.
6639 SDValue Tmp = visitXOR(TheXor);
6640 if (Tmp.getNode() && Tmp.getNode() != TheXor) {
6641 DEBUG(dbgs() << "\nReplacing.8 ";
6643 dbgs() << "\nWith: ";
6644 Tmp.getNode()->dump(&DAG);
6646 WorkListRemover DeadNodes(*this);
6647 DAG.ReplaceAllUsesOfValueWith(N1, Tmp);
6648 removeFromWorkList(TheXor);
6649 DAG.DeleteNode(TheXor);
6650 return DAG.getNode(ISD::BRCOND, N->getDebugLoc(),
6651 MVT::Other, Chain, Tmp, N2);
6655 if (Op0.getOpcode() != ISD::SETCC && Op1.getOpcode() != ISD::SETCC) {
6657 if (ConstantSDNode *RHSCI = dyn_cast<ConstantSDNode>(Op0))
6658 if (RHSCI->getAPIntValue() == 1 && Op0.hasOneUse() &&
6659 Op0.getOpcode() == ISD::XOR) {
6660 TheXor = Op0.getNode();
6664 EVT SetCCVT = N1.getValueType();
6666 SetCCVT = TLI.getSetCCResultType(SetCCVT);
6667 SDValue SetCC = DAG.getSetCC(TheXor->getDebugLoc(),
6670 Equal ? ISD::SETEQ : ISD::SETNE);
6671 // Replace the uses of XOR with SETCC
6672 WorkListRemover DeadNodes(*this);
6673 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
6674 removeFromWorkList(N1.getNode());
6675 DAG.DeleteNode(N1.getNode());
6676 return DAG.getNode(ISD::BRCOND, N->getDebugLoc(),
6677 MVT::Other, Chain, SetCC, N2);
6684 // Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB.
6686 SDValue DAGCombiner::visitBR_CC(SDNode *N) {
6687 CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1));
6688 SDValue CondLHS = N->getOperand(2), CondRHS = N->getOperand(3);
6690 // If N is a constant we could fold this into a fallthrough or unconditional
6691 // branch. However that doesn't happen very often in normal code, because
6692 // Instcombine/SimplifyCFG should have handled the available opportunities.
6693 // If we did this folding here, it would be necessary to update the
6694 // MachineBasicBlock CFG, which is awkward.
6696 // Use SimplifySetCC to simplify SETCC's.
6697 SDValue Simp = SimplifySetCC(TLI.getSetCCResultType(CondLHS.getValueType()),
6698 CondLHS, CondRHS, CC->get(), N->getDebugLoc(),
6700 if (Simp.getNode()) AddToWorkList(Simp.getNode());
6702 // fold to a simpler setcc
6703 if (Simp.getNode() && Simp.getOpcode() == ISD::SETCC)
6704 return DAG.getNode(ISD::BR_CC, N->getDebugLoc(), MVT::Other,
6705 N->getOperand(0), Simp.getOperand(2),
6706 Simp.getOperand(0), Simp.getOperand(1),
6712 /// canFoldInAddressingMode - Return true if 'Use' is a load or a store that
6713 /// uses N as its base pointer and that N may be folded in the load / store
6714 /// addressing mode.
6715 static bool canFoldInAddressingMode(SDNode *N, SDNode *Use,
6717 const TargetLowering &TLI) {
6719 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Use)) {
6720 if (LD->isIndexed() || LD->getBasePtr().getNode() != N)
6722 VT = Use->getValueType(0);
6723 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(Use)) {
6724 if (ST->isIndexed() || ST->getBasePtr().getNode() != N)
6726 VT = ST->getValue().getValueType();
6730 TargetLowering::AddrMode AM;
6731 if (N->getOpcode() == ISD::ADD) {
6732 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
6735 AM.BaseOffs = Offset->getSExtValue();
6739 } else if (N->getOpcode() == ISD::SUB) {
6740 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
6743 AM.BaseOffs = -Offset->getSExtValue();
6750 return TLI.isLegalAddressingMode(AM, VT.getTypeForEVT(*DAG.getContext()));
6753 /// CombineToPreIndexedLoadStore - Try turning a load / store into a
6754 /// pre-indexed load / store when the base pointer is an add or subtract
6755 /// and it has other uses besides the load / store. After the
6756 /// transformation, the new indexed load / store has effectively folded
6757 /// the add / subtract in and all of its other uses are redirected to the
6758 /// new load / store.
6759 bool DAGCombiner::CombineToPreIndexedLoadStore(SDNode *N) {
6760 if (Level < AfterLegalizeDAG)
6766 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
6767 if (LD->isIndexed())
6769 VT = LD->getMemoryVT();
6770 if (!TLI.isIndexedLoadLegal(ISD::PRE_INC, VT) &&
6771 !TLI.isIndexedLoadLegal(ISD::PRE_DEC, VT))
6773 Ptr = LD->getBasePtr();
6774 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
6775 if (ST->isIndexed())
6777 VT = ST->getMemoryVT();
6778 if (!TLI.isIndexedStoreLegal(ISD::PRE_INC, VT) &&
6779 !TLI.isIndexedStoreLegal(ISD::PRE_DEC, VT))
6781 Ptr = ST->getBasePtr();
6787 // If the pointer is not an add/sub, or if it doesn't have multiple uses, bail
6788 // out. There is no reason to make this a preinc/predec.
6789 if ((Ptr.getOpcode() != ISD::ADD && Ptr.getOpcode() != ISD::SUB) ||
6790 Ptr.getNode()->hasOneUse())
6793 // Ask the target to do addressing mode selection.
6796 ISD::MemIndexedMode AM = ISD::UNINDEXED;
6797 if (!TLI.getPreIndexedAddressParts(N, BasePtr, Offset, AM, DAG))
6799 // Don't create a indexed load / store with zero offset.
6800 if (isa<ConstantSDNode>(Offset) &&
6801 cast<ConstantSDNode>(Offset)->isNullValue())
6804 // Try turning it into a pre-indexed load / store except when:
6805 // 1) The new base ptr is a frame index.
6806 // 2) If N is a store and the new base ptr is either the same as or is a
6807 // predecessor of the value being stored.
6808 // 3) Another use of old base ptr is a predecessor of N. If ptr is folded
6809 // that would create a cycle.
6810 // 4) All uses are load / store ops that use it as old base ptr.
6812 // Check #1. Preinc'ing a frame index would require copying the stack pointer
6813 // (plus the implicit offset) to a register to preinc anyway.
6814 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
6819 SDValue Val = cast<StoreSDNode>(N)->getValue();
6820 if (Val == BasePtr || BasePtr.getNode()->isPredecessorOf(Val.getNode()))
6824 // Now check for #3 and #4.
6825 bool RealUse = false;
6827 // Caches for hasPredecessorHelper
6828 SmallPtrSet<const SDNode *, 32> Visited;
6829 SmallVector<const SDNode *, 16> Worklist;
6831 for (SDNode::use_iterator I = Ptr.getNode()->use_begin(),
6832 E = Ptr.getNode()->use_end(); I != E; ++I) {
6836 if (N->hasPredecessorHelper(Use, Visited, Worklist))
6839 // If Ptr may be folded in addressing mode of other use, then it's
6840 // not profitable to do this transformation.
6841 if (!canFoldInAddressingMode(Ptr.getNode(), Use, DAG, TLI))
6850 Result = DAG.getIndexedLoad(SDValue(N,0), N->getDebugLoc(),
6851 BasePtr, Offset, AM);
6853 Result = DAG.getIndexedStore(SDValue(N,0), N->getDebugLoc(),
6854 BasePtr, Offset, AM);
6857 DEBUG(dbgs() << "\nReplacing.4 ";
6859 dbgs() << "\nWith: ";
6860 Result.getNode()->dump(&DAG);
6862 WorkListRemover DeadNodes(*this);
6864 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
6865 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
6867 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
6870 // Finally, since the node is now dead, remove it from the graph.
6873 // Replace the uses of Ptr with uses of the updated base value.
6874 DAG.ReplaceAllUsesOfValueWith(Ptr, Result.getValue(isLoad ? 1 : 0));
6875 removeFromWorkList(Ptr.getNode());
6876 DAG.DeleteNode(Ptr.getNode());
6881 /// CombineToPostIndexedLoadStore - Try to combine a load / store with a
6882 /// add / sub of the base pointer node into a post-indexed load / store.
6883 /// The transformation folded the add / subtract into the new indexed
6884 /// load / store effectively and all of its uses are redirected to the
6885 /// new load / store.
6886 bool DAGCombiner::CombineToPostIndexedLoadStore(SDNode *N) {
6887 if (Level < AfterLegalizeDAG)
6893 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
6894 if (LD->isIndexed())
6896 VT = LD->getMemoryVT();
6897 if (!TLI.isIndexedLoadLegal(ISD::POST_INC, VT) &&
6898 !TLI.isIndexedLoadLegal(ISD::POST_DEC, VT))
6900 Ptr = LD->getBasePtr();
6901 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
6902 if (ST->isIndexed())
6904 VT = ST->getMemoryVT();
6905 if (!TLI.isIndexedStoreLegal(ISD::POST_INC, VT) &&
6906 !TLI.isIndexedStoreLegal(ISD::POST_DEC, VT))
6908 Ptr = ST->getBasePtr();
6914 if (Ptr.getNode()->hasOneUse())
6917 for (SDNode::use_iterator I = Ptr.getNode()->use_begin(),
6918 E = Ptr.getNode()->use_end(); I != E; ++I) {
6921 (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB))
6926 ISD::MemIndexedMode AM = ISD::UNINDEXED;
6927 if (TLI.getPostIndexedAddressParts(N, Op, BasePtr, Offset, AM, DAG)) {
6928 // Don't create a indexed load / store with zero offset.
6929 if (isa<ConstantSDNode>(Offset) &&
6930 cast<ConstantSDNode>(Offset)->isNullValue())
6933 // Try turning it into a post-indexed load / store except when
6934 // 1) All uses are load / store ops that use it as base ptr (and
6935 // it may be folded as addressing mmode).
6936 // 2) Op must be independent of N, i.e. Op is neither a predecessor
6937 // nor a successor of N. Otherwise, if Op is folded that would
6940 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
6944 bool TryNext = false;
6945 for (SDNode::use_iterator II = BasePtr.getNode()->use_begin(),
6946 EE = BasePtr.getNode()->use_end(); II != EE; ++II) {
6948 if (Use == Ptr.getNode())
6951 // If all the uses are load / store addresses, then don't do the
6953 if (Use->getOpcode() == ISD::ADD || Use->getOpcode() == ISD::SUB){
6954 bool RealUse = false;
6955 for (SDNode::use_iterator III = Use->use_begin(),
6956 EEE = Use->use_end(); III != EEE; ++III) {
6957 SDNode *UseUse = *III;
6958 if (!canFoldInAddressingMode(Use, UseUse, DAG, TLI))
6973 if (!Op->isPredecessorOf(N) && !N->isPredecessorOf(Op)) {
6974 SDValue Result = isLoad
6975 ? DAG.getIndexedLoad(SDValue(N,0), N->getDebugLoc(),
6976 BasePtr, Offset, AM)
6977 : DAG.getIndexedStore(SDValue(N,0), N->getDebugLoc(),
6978 BasePtr, Offset, AM);
6981 DEBUG(dbgs() << "\nReplacing.5 ";
6983 dbgs() << "\nWith: ";
6984 Result.getNode()->dump(&DAG);
6986 WorkListRemover DeadNodes(*this);
6988 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
6989 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
6991 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
6994 // Finally, since the node is now dead, remove it from the graph.
6997 // Replace the uses of Use with uses of the updated base value.
6998 DAG.ReplaceAllUsesOfValueWith(SDValue(Op, 0),
6999 Result.getValue(isLoad ? 1 : 0));
7000 removeFromWorkList(Op);
7010 SDValue DAGCombiner::visitLOAD(SDNode *N) {
7011 LoadSDNode *LD = cast<LoadSDNode>(N);
7012 SDValue Chain = LD->getChain();
7013 SDValue Ptr = LD->getBasePtr();
7015 // If load is not volatile and there are no uses of the loaded value (and
7016 // the updated indexed value in case of indexed loads), change uses of the
7017 // chain value into uses of the chain input (i.e. delete the dead load).
7018 if (!LD->isVolatile()) {
7019 if (N->getValueType(1) == MVT::Other) {
7021 if (!N->hasAnyUseOfValue(0)) {
7022 // It's not safe to use the two value CombineTo variant here. e.g.
7023 // v1, chain2 = load chain1, loc
7024 // v2, chain3 = load chain2, loc
7026 // Now we replace use of chain2 with chain1. This makes the second load
7027 // isomorphic to the one we are deleting, and thus makes this load live.
7028 DEBUG(dbgs() << "\nReplacing.6 ";
7030 dbgs() << "\nWith chain: ";
7031 Chain.getNode()->dump(&DAG);
7033 WorkListRemover DeadNodes(*this);
7034 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
7036 if (N->use_empty()) {
7037 removeFromWorkList(N);
7041 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7045 assert(N->getValueType(2) == MVT::Other && "Malformed indexed loads?");
7046 if (!N->hasAnyUseOfValue(0) && !N->hasAnyUseOfValue(1)) {
7047 SDValue Undef = DAG.getUNDEF(N->getValueType(0));
7048 DEBUG(dbgs() << "\nReplacing.7 ";
7050 dbgs() << "\nWith: ";
7051 Undef.getNode()->dump(&DAG);
7052 dbgs() << " and 2 other values\n");
7053 WorkListRemover DeadNodes(*this);
7054 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Undef);
7055 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1),
7056 DAG.getUNDEF(N->getValueType(1)));
7057 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 2), Chain);
7058 removeFromWorkList(N);
7060 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7065 // If this load is directly stored, replace the load value with the stored
7067 // TODO: Handle store large -> read small portion.
7068 // TODO: Handle TRUNCSTORE/LOADEXT
7069 if (ISD::isNormalLoad(N) && !LD->isVolatile()) {
7070 if (ISD::isNON_TRUNCStore(Chain.getNode())) {
7071 StoreSDNode *PrevST = cast<StoreSDNode>(Chain);
7072 if (PrevST->getBasePtr() == Ptr &&
7073 PrevST->getValue().getValueType() == N->getValueType(0))
7074 return CombineTo(N, Chain.getOperand(1), Chain);
7078 // Try to infer better alignment information than the load already has.
7079 if (OptLevel != CodeGenOpt::None && LD->isUnindexed()) {
7080 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
7081 if (Align > LD->getAlignment())
7082 return DAG.getExtLoad(LD->getExtensionType(), N->getDebugLoc(),
7083 LD->getValueType(0),
7084 Chain, Ptr, LD->getPointerInfo(),
7086 LD->isVolatile(), LD->isNonTemporal(), Align);
7091 // Walk up chain skipping non-aliasing memory nodes.
7092 SDValue BetterChain = FindBetterChain(N, Chain);
7094 // If there is a better chain.
7095 if (Chain != BetterChain) {
7098 // Replace the chain to void dependency.
7099 if (LD->getExtensionType() == ISD::NON_EXTLOAD) {
7100 ReplLoad = DAG.getLoad(N->getValueType(0), LD->getDebugLoc(),
7101 BetterChain, Ptr, LD->getPointerInfo(),
7102 LD->isVolatile(), LD->isNonTemporal(),
7103 LD->isInvariant(), LD->getAlignment());
7105 ReplLoad = DAG.getExtLoad(LD->getExtensionType(), LD->getDebugLoc(),
7106 LD->getValueType(0),
7107 BetterChain, Ptr, LD->getPointerInfo(),
7110 LD->isNonTemporal(),
7111 LD->getAlignment());
7114 // Create token factor to keep old chain connected.
7115 SDValue Token = DAG.getNode(ISD::TokenFactor, N->getDebugLoc(),
7116 MVT::Other, Chain, ReplLoad.getValue(1));
7118 // Make sure the new and old chains are cleaned up.
7119 AddToWorkList(Token.getNode());
7121 // Replace uses with load result and token factor. Don't add users
7123 return CombineTo(N, ReplLoad.getValue(0), Token, false);
7127 // Try transforming N to an indexed load.
7128 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
7129 return SDValue(N, 0);
7134 /// CheckForMaskedLoad - Check to see if V is (and load (ptr), imm), where the
7135 /// load is having specific bytes cleared out. If so, return the byte size
7136 /// being masked out and the shift amount.
7137 static std::pair<unsigned, unsigned>
7138 CheckForMaskedLoad(SDValue V, SDValue Ptr, SDValue Chain) {
7139 std::pair<unsigned, unsigned> Result(0, 0);
7141 // Check for the structure we're looking for.
7142 if (V->getOpcode() != ISD::AND ||
7143 !isa<ConstantSDNode>(V->getOperand(1)) ||
7144 !ISD::isNormalLoad(V->getOperand(0).getNode()))
7147 // Check the chain and pointer.
7148 LoadSDNode *LD = cast<LoadSDNode>(V->getOperand(0));
7149 if (LD->getBasePtr() != Ptr) return Result; // Not from same pointer.
7151 // The store should be chained directly to the load or be an operand of a
7153 if (LD == Chain.getNode())
7155 else if (Chain->getOpcode() != ISD::TokenFactor)
7156 return Result; // Fail.
7159 for (unsigned i = 0, e = Chain->getNumOperands(); i != e; ++i)
7160 if (Chain->getOperand(i).getNode() == LD) {
7164 if (!isOk) return Result;
7167 // This only handles simple types.
7168 if (V.getValueType() != MVT::i16 &&
7169 V.getValueType() != MVT::i32 &&
7170 V.getValueType() != MVT::i64)
7173 // Check the constant mask. Invert it so that the bits being masked out are
7174 // 0 and the bits being kept are 1. Use getSExtValue so that leading bits
7175 // follow the sign bit for uniformity.
7176 uint64_t NotMask = ~cast<ConstantSDNode>(V->getOperand(1))->getSExtValue();
7177 unsigned NotMaskLZ = CountLeadingZeros_64(NotMask);
7178 if (NotMaskLZ & 7) return Result; // Must be multiple of a byte.
7179 unsigned NotMaskTZ = CountTrailingZeros_64(NotMask);
7180 if (NotMaskTZ & 7) return Result; // Must be multiple of a byte.
7181 if (NotMaskLZ == 64) return Result; // All zero mask.
7183 // See if we have a continuous run of bits. If so, we have 0*1+0*
7184 if (CountTrailingOnes_64(NotMask >> NotMaskTZ)+NotMaskTZ+NotMaskLZ != 64)
7187 // Adjust NotMaskLZ down to be from the actual size of the int instead of i64.
7188 if (V.getValueType() != MVT::i64 && NotMaskLZ)
7189 NotMaskLZ -= 64-V.getValueSizeInBits();
7191 unsigned MaskedBytes = (V.getValueSizeInBits()-NotMaskLZ-NotMaskTZ)/8;
7192 switch (MaskedBytes) {
7196 default: return Result; // All one mask, or 5-byte mask.
7199 // Verify that the first bit starts at a multiple of mask so that the access
7200 // is aligned the same as the access width.
7201 if (NotMaskTZ && NotMaskTZ/8 % MaskedBytes) return Result;
7203 Result.first = MaskedBytes;
7204 Result.second = NotMaskTZ/8;
7209 /// ShrinkLoadReplaceStoreWithStore - Check to see if IVal is something that
7210 /// provides a value as specified by MaskInfo. If so, replace the specified
7211 /// store with a narrower store of truncated IVal.
7213 ShrinkLoadReplaceStoreWithStore(const std::pair<unsigned, unsigned> &MaskInfo,
7214 SDValue IVal, StoreSDNode *St,
7216 unsigned NumBytes = MaskInfo.first;
7217 unsigned ByteShift = MaskInfo.second;
7218 SelectionDAG &DAG = DC->getDAG();
7220 // Check to see if IVal is all zeros in the part being masked in by the 'or'
7221 // that uses this. If not, this is not a replacement.
7222 APInt Mask = ~APInt::getBitsSet(IVal.getValueSizeInBits(),
7223 ByteShift*8, (ByteShift+NumBytes)*8);
7224 if (!DAG.MaskedValueIsZero(IVal, Mask)) return 0;
7226 // Check that it is legal on the target to do this. It is legal if the new
7227 // VT we're shrinking to (i8/i16/i32) is legal or we're still before type
7229 MVT VT = MVT::getIntegerVT(NumBytes*8);
7230 if (!DC->isTypeLegal(VT))
7233 // Okay, we can do this! Replace the 'St' store with a store of IVal that is
7234 // shifted by ByteShift and truncated down to NumBytes.
7236 IVal = DAG.getNode(ISD::SRL, IVal->getDebugLoc(), IVal.getValueType(), IVal,
7237 DAG.getConstant(ByteShift*8,
7238 DC->getShiftAmountTy(IVal.getValueType())));
7240 // Figure out the offset for the store and the alignment of the access.
7242 unsigned NewAlign = St->getAlignment();
7244 if (DAG.getTargetLoweringInfo().isLittleEndian())
7245 StOffset = ByteShift;
7247 StOffset = IVal.getValueType().getStoreSize() - ByteShift - NumBytes;
7249 SDValue Ptr = St->getBasePtr();
7251 Ptr = DAG.getNode(ISD::ADD, IVal->getDebugLoc(), Ptr.getValueType(),
7252 Ptr, DAG.getConstant(StOffset, Ptr.getValueType()));
7253 NewAlign = MinAlign(NewAlign, StOffset);
7256 // Truncate down to the new size.
7257 IVal = DAG.getNode(ISD::TRUNCATE, IVal->getDebugLoc(), VT, IVal);
7260 return DAG.getStore(St->getChain(), St->getDebugLoc(), IVal, Ptr,
7261 St->getPointerInfo().getWithOffset(StOffset),
7262 false, false, NewAlign).getNode();
7266 /// ReduceLoadOpStoreWidth - Look for sequence of load / op / store where op is
7267 /// one of 'or', 'xor', and 'and' of immediates. If 'op' is only touching some
7268 /// of the loaded bits, try narrowing the load and store if it would end up
7269 /// being a win for performance or code size.
7270 SDValue DAGCombiner::ReduceLoadOpStoreWidth(SDNode *N) {
7271 StoreSDNode *ST = cast<StoreSDNode>(N);
7272 if (ST->isVolatile())
7275 SDValue Chain = ST->getChain();
7276 SDValue Value = ST->getValue();
7277 SDValue Ptr = ST->getBasePtr();
7278 EVT VT = Value.getValueType();
7280 if (ST->isTruncatingStore() || VT.isVector() || !Value.hasOneUse())
7283 unsigned Opc = Value.getOpcode();
7285 // If this is "store (or X, Y), P" and X is "(and (load P), cst)", where cst
7286 // is a byte mask indicating a consecutive number of bytes, check to see if
7287 // Y is known to provide just those bytes. If so, we try to replace the
7288 // load + replace + store sequence with a single (narrower) store, which makes
7290 if (Opc == ISD::OR) {
7291 std::pair<unsigned, unsigned> MaskedLoad;
7292 MaskedLoad = CheckForMaskedLoad(Value.getOperand(0), Ptr, Chain);
7293 if (MaskedLoad.first)
7294 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
7295 Value.getOperand(1), ST,this))
7296 return SDValue(NewST, 0);
7298 // Or is commutative, so try swapping X and Y.
7299 MaskedLoad = CheckForMaskedLoad(Value.getOperand(1), Ptr, Chain);
7300 if (MaskedLoad.first)
7301 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
7302 Value.getOperand(0), ST,this))
7303 return SDValue(NewST, 0);
7306 if ((Opc != ISD::OR && Opc != ISD::XOR && Opc != ISD::AND) ||
7307 Value.getOperand(1).getOpcode() != ISD::Constant)
7310 SDValue N0 = Value.getOperand(0);
7311 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
7312 Chain == SDValue(N0.getNode(), 1)) {
7313 LoadSDNode *LD = cast<LoadSDNode>(N0);
7314 if (LD->getBasePtr() != Ptr ||
7315 LD->getPointerInfo().getAddrSpace() !=
7316 ST->getPointerInfo().getAddrSpace())
7319 // Find the type to narrow it the load / op / store to.
7320 SDValue N1 = Value.getOperand(1);
7321 unsigned BitWidth = N1.getValueSizeInBits();
7322 APInt Imm = cast<ConstantSDNode>(N1)->getAPIntValue();
7323 if (Opc == ISD::AND)
7324 Imm ^= APInt::getAllOnesValue(BitWidth);
7325 if (Imm == 0 || Imm.isAllOnesValue())
7327 unsigned ShAmt = Imm.countTrailingZeros();
7328 unsigned MSB = BitWidth - Imm.countLeadingZeros() - 1;
7329 unsigned NewBW = NextPowerOf2(MSB - ShAmt);
7330 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
7331 while (NewBW < BitWidth &&
7332 !(TLI.isOperationLegalOrCustom(Opc, NewVT) &&
7333 TLI.isNarrowingProfitable(VT, NewVT))) {
7334 NewBW = NextPowerOf2(NewBW);
7335 NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
7337 if (NewBW >= BitWidth)
7340 // If the lsb changed does not start at the type bitwidth boundary,
7341 // start at the previous one.
7343 ShAmt = (((ShAmt + NewBW - 1) / NewBW) * NewBW) - NewBW;
7344 APInt Mask = APInt::getBitsSet(BitWidth, ShAmt, ShAmt + NewBW);
7345 if ((Imm & Mask) == Imm) {
7346 APInt NewImm = (Imm & Mask).lshr(ShAmt).trunc(NewBW);
7347 if (Opc == ISD::AND)
7348 NewImm ^= APInt::getAllOnesValue(NewBW);
7349 uint64_t PtrOff = ShAmt / 8;
7350 // For big endian targets, we need to adjust the offset to the pointer to
7351 // load the correct bytes.
7352 if (TLI.isBigEndian())
7353 PtrOff = (BitWidth + 7 - NewBW) / 8 - PtrOff;
7355 unsigned NewAlign = MinAlign(LD->getAlignment(), PtrOff);
7356 Type *NewVTTy = NewVT.getTypeForEVT(*DAG.getContext());
7357 if (NewAlign < TLI.getTargetData()->getABITypeAlignment(NewVTTy))
7360 SDValue NewPtr = DAG.getNode(ISD::ADD, LD->getDebugLoc(),
7361 Ptr.getValueType(), Ptr,
7362 DAG.getConstant(PtrOff, Ptr.getValueType()));
7363 SDValue NewLD = DAG.getLoad(NewVT, N0.getDebugLoc(),
7364 LD->getChain(), NewPtr,
7365 LD->getPointerInfo().getWithOffset(PtrOff),
7366 LD->isVolatile(), LD->isNonTemporal(),
7367 LD->isInvariant(), NewAlign);
7368 SDValue NewVal = DAG.getNode(Opc, Value.getDebugLoc(), NewVT, NewLD,
7369 DAG.getConstant(NewImm, NewVT));
7370 SDValue NewST = DAG.getStore(Chain, N->getDebugLoc(),
7372 ST->getPointerInfo().getWithOffset(PtrOff),
7373 false, false, NewAlign);
7375 AddToWorkList(NewPtr.getNode());
7376 AddToWorkList(NewLD.getNode());
7377 AddToWorkList(NewVal.getNode());
7378 WorkListRemover DeadNodes(*this);
7379 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLD.getValue(1));
7388 /// TransformFPLoadStorePair - For a given floating point load / store pair,
7389 /// if the load value isn't used by any other operations, then consider
7390 /// transforming the pair to integer load / store operations if the target
7391 /// deems the transformation profitable.
7392 SDValue DAGCombiner::TransformFPLoadStorePair(SDNode *N) {
7393 StoreSDNode *ST = cast<StoreSDNode>(N);
7394 SDValue Chain = ST->getChain();
7395 SDValue Value = ST->getValue();
7396 if (ISD::isNormalStore(ST) && ISD::isNormalLoad(Value.getNode()) &&
7397 Value.hasOneUse() &&
7398 Chain == SDValue(Value.getNode(), 1)) {
7399 LoadSDNode *LD = cast<LoadSDNode>(Value);
7400 EVT VT = LD->getMemoryVT();
7401 if (!VT.isFloatingPoint() ||
7402 VT != ST->getMemoryVT() ||
7403 LD->isNonTemporal() ||
7404 ST->isNonTemporal() ||
7405 LD->getPointerInfo().getAddrSpace() != 0 ||
7406 ST->getPointerInfo().getAddrSpace() != 0)
7409 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
7410 if (!TLI.isOperationLegal(ISD::LOAD, IntVT) ||
7411 !TLI.isOperationLegal(ISD::STORE, IntVT) ||
7412 !TLI.isDesirableToTransformToIntegerOp(ISD::LOAD, VT) ||
7413 !TLI.isDesirableToTransformToIntegerOp(ISD::STORE, VT))
7416 unsigned LDAlign = LD->getAlignment();
7417 unsigned STAlign = ST->getAlignment();
7418 Type *IntVTTy = IntVT.getTypeForEVT(*DAG.getContext());
7419 unsigned ABIAlign = TLI.getTargetData()->getABITypeAlignment(IntVTTy);
7420 if (LDAlign < ABIAlign || STAlign < ABIAlign)
7423 SDValue NewLD = DAG.getLoad(IntVT, Value.getDebugLoc(),
7424 LD->getChain(), LD->getBasePtr(),
7425 LD->getPointerInfo(),
7426 false, false, false, LDAlign);
7428 SDValue NewST = DAG.getStore(NewLD.getValue(1), N->getDebugLoc(),
7429 NewLD, ST->getBasePtr(),
7430 ST->getPointerInfo(),
7431 false, false, STAlign);
7433 AddToWorkList(NewLD.getNode());
7434 AddToWorkList(NewST.getNode());
7435 WorkListRemover DeadNodes(*this);
7436 DAG.ReplaceAllUsesOfValueWith(Value.getValue(1), NewLD.getValue(1));
7444 /// Returns the base pointer and an integer offset from that object.
7445 static std::pair<SDValue, int64_t> GetPointerBaseAndOffset(SDValue Ptr) {
7446 if (Ptr->getOpcode() == ISD::ADD && isa<ConstantSDNode>(Ptr->getOperand(1))) {
7447 int64_t Offset = cast<ConstantSDNode>(Ptr->getOperand(1))->getSExtValue();
7448 SDValue Base = Ptr->getOperand(0);
7449 return std::make_pair(Base, Offset);
7452 return std::make_pair(Ptr, 0);
7455 struct ConsecutiveMemoryChainSorter {
7456 typedef std::pair<LSBaseSDNode*, int64_t> MemLink;
7457 bool operator()(MemLink LHS, MemLink RHS) {
7458 return LHS.second < RHS.second;
7462 bool DAGCombiner::MergeConsecutiveStores(StoreSDNode* St) {
7463 EVT MemVT = St->getMemoryVT();
7464 int64_t ElementSizeBytes = MemVT.getSizeInBits()/8;
7466 // Don't handle vectors.
7467 if (MemVT.isVector() || !MemVT.isSimple())
7470 // Perform an early exit check. Do not bother looking at stored values that
7471 // are not constants or loads.
7472 SDValue StoredVal = St->getValue();
7473 if (!isa<ConstantSDNode>(StoredVal) && !isa<ConstantFPSDNode>(StoredVal) &&
7474 !isa<LoadSDNode>(StoredVal))
7477 // Is this a load-to-store or a const-store.
7478 bool IsLoadSrc = isa<LoadSDNode>(StoredVal);
7480 // Only look at ends of store chains.
7481 SDValue Chain = SDValue(St, 1);
7482 if (Chain->hasOneUse() && Chain->use_begin()->getOpcode() == ISD::STORE)
7485 // This holds the base pointer and the offset in bytes from the base pointer.
7486 std::pair<SDValue, int64_t> BasePtr =
7487 GetPointerBaseAndOffset(St->getBasePtr());
7489 // We must have a base and an offset.
7490 if (!BasePtr.first.getNode())
7493 // Do not handle stores to undef base pointers.
7494 if (BasePtr.first.getOpcode() == ISD::UNDEF)
7497 SmallVector<std::pair<StoreSDNode*, int64_t>, 8> StoreNodes;
7498 // Walk up the chain and look for nodes with offsets from the same
7499 // base pointer. Stop when reaching an instruction with a different kind
7500 // or instruction which has a different base pointer.
7501 StoreSDNode *Index = St;
7503 // If the chain has more than one use, then we can't reorder the mem ops.
7504 if (Index != St && !SDValue(Index, 1)->hasOneUse())
7507 // Find the base pointer and offset for this memory node.
7508 std::pair<SDValue, int64_t> Ptr =
7509 GetPointerBaseAndOffset(Index->getBasePtr());
7511 // Check that the base pointer is the same as the original one.
7512 if (Ptr.first.getNode() != BasePtr.first.getNode())
7515 // Check that the alignment is the same.
7516 if (Index->getAlignment() != St->getAlignment())
7519 // The memory operands must not be volatile.
7520 if (Index->isVolatile() || Index->isIndexed())
7524 if (StoreSDNode *St = dyn_cast<StoreSDNode>(Index))
7525 if (St->isTruncatingStore())
7528 // The stored memory type must be the same.
7529 if (Index->getMemoryVT() != MemVT)
7532 // We found a potential memory operand to merge.
7533 StoreNodes.push_back(std::make_pair(Index,Ptr.second));
7535 // Move up the chain to the next memory operation.
7536 Index = dyn_cast<StoreSDNode>(Index->getChain().getNode());
7539 // Check if there is anything to merge.
7540 if (StoreNodes.size() < 2)
7543 // Remember which node is the earliest node in the chain.
7544 LSBaseSDNode *EarliestOp = StoreNodes.back().first;
7546 // Sort the memory operands according to their distance from the base pointer.
7547 std::sort(StoreNodes.begin(), StoreNodes.end(),
7548 ConsecutiveMemoryChainSorter());
7550 // Scan the memory operations on the chain and find the first non-consecutive
7551 // store memory address.
7552 unsigned LastConsecutiveStore = 0;
7553 int64_t StartAddress = StoreNodes[0].second;
7554 for (unsigned i=1; i<StoreNodes.size(); ++i) {
7555 int64_t CurrAddress = StoreNodes[i].second;
7556 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
7558 LastConsecutiveStore = i;
7561 // Store the constants into memory as one consecutive store.
7563 unsigned LastConst = 0;
7564 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
7565 SDValue StoredVal = StoreNodes[i].first->getValue();
7566 bool IsConst = (isa<ConstantSDNode>(StoredVal) || isa<ConstantFPSDNode>(StoredVal));
7571 unsigned NumElem = std::min(LastConsecutiveStore + 1, LastConst + 1);
7575 EVT JointMemOpVT = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
7576 DebugLoc DL = StoreNodes[0].first->getDebugLoc();
7577 SmallVector<SDValue, 8> Ops;
7579 for (unsigned i = 0; i < NumElem ; ++i) {
7580 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].first);
7581 Ops.push_back(St->getValue());
7584 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, DL,
7585 JointMemOpVT, &Ops[0], Ops.size());
7587 SDValue NewStore = DAG.getStore(EarliestOp->getChain(), DL, BV,
7588 EarliestOp->getBasePtr(),
7589 EarliestOp->getPointerInfo(), false, false,
7590 EarliestOp->getAlignment());
7592 for (unsigned i = 0; i < NumElem ; ++i) {
7593 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].first);
7594 CombineTo(St, NewStore);
7599 // Look for load nodes wich are used by the stored values.
7600 SmallVector<std::pair<LoadSDNode*, int64_t>, 8> LoadNodes;
7602 // Find acceptible loads. Loads need to have the same chain (token factor),
7603 // must not be zext, volatile, indexed, and they must be consecutive.
7605 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
7606 LoadSDNode *Ld = dyn_cast<LoadSDNode>(StoreNodes[i].first->getValue());
7609 // Loads must only have one use.
7610 if (!Ld->hasNUsesOfValue(1, 0))
7613 // Check that the alignment is the same as the stores.
7614 if (Ld->getAlignment() != St->getAlignment())
7617 // The memory operands must not be volatile.
7618 if (Ld->isVolatile() || Ld->isIndexed())
7621 if (Ld->getExtensionType() != ISD::NON_EXTLOAD)
7624 // The stored memory type must be the same.
7625 if (Ld->getMemoryVT() != MemVT)
7628 std::pair<SDValue, int64_t> LdPtr =
7629 GetPointerBaseAndOffset(Ld->getBasePtr());
7631 // If this is not the first ptr that we check.
7632 if (LdBasePtr.getNode()) {
7633 // The base ptr must be the same,
7634 if (LdPtr.first != LdBasePtr)
7637 LdBasePtr = LdPtr.first;
7640 // We found a potential memory operand to merge.
7641 LoadNodes.push_back(std::make_pair(Ld, LdPtr.second));
7644 if (LoadNodes.size() < 2)
7647 // Scan the memory operations on the chain and find the first non-consecutive
7648 // load memory address.
7649 unsigned LastConsecutiveLoad = 0;
7650 StartAddress = LoadNodes[0].second;
7651 for (unsigned i=1; i<LoadNodes.size(); ++i) {
7652 int64_t CurrAddress = LoadNodes[i].second;
7653 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
7655 LastConsecutiveLoad = i;
7659 std::min(LastConsecutiveStore + 1, LastConsecutiveLoad + 1);
7661 EVT JointMemOpVT = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
7662 DebugLoc LoadDL = LoadNodes[0].first->getDebugLoc();
7663 DebugLoc StoreDL = StoreNodes[0].first->getDebugLoc();
7665 LoadSDNode *FirstLoad = LoadNodes[0].first;
7666 SDValue NewLoad = DAG.getLoad(JointMemOpVT, LoadDL,
7667 FirstLoad->getChain(),
7668 FirstLoad->getBasePtr(),
7669 FirstLoad->getPointerInfo(),
7670 false, false, false,
7671 FirstLoad->getAlignment());
7673 SDValue NewStore = DAG.getStore(EarliestOp->getChain(), StoreDL, NewLoad,
7674 EarliestOp->getBasePtr(),
7675 EarliestOp->getPointerInfo(), false, false,
7676 EarliestOp->getAlignment());
7678 for (unsigned i = 0; i < NumElem ; ++i) {
7679 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].first);
7680 CombineTo(St, NewStore);
7686 SDValue DAGCombiner::visitSTORE(SDNode *N) {
7687 StoreSDNode *ST = cast<StoreSDNode>(N);
7688 SDValue Chain = ST->getChain();
7689 SDValue Value = ST->getValue();
7690 SDValue Ptr = ST->getBasePtr();
7692 // If this is a store of a bit convert, store the input value if the
7693 // resultant store does not need a higher alignment than the original.
7694 if (Value.getOpcode() == ISD::BITCAST && !ST->isTruncatingStore() &&
7695 ST->isUnindexed()) {
7696 unsigned OrigAlign = ST->getAlignment();
7697 EVT SVT = Value.getOperand(0).getValueType();
7698 unsigned Align = TLI.getTargetData()->
7699 getABITypeAlignment(SVT.getTypeForEVT(*DAG.getContext()));
7700 if (Align <= OrigAlign &&
7701 ((!LegalOperations && !ST->isVolatile()) ||
7702 TLI.isOperationLegalOrCustom(ISD::STORE, SVT)))
7703 return DAG.getStore(Chain, N->getDebugLoc(), Value.getOperand(0),
7704 Ptr, ST->getPointerInfo(), ST->isVolatile(),
7705 ST->isNonTemporal(), OrigAlign);
7708 // Turn 'store undef, Ptr' -> nothing.
7709 if (Value.getOpcode() == ISD::UNDEF && ST->isUnindexed())
7712 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
7713 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Value)) {
7714 // NOTE: If the original store is volatile, this transform must not increase
7715 // the number of stores. For example, on x86-32 an f64 can be stored in one
7716 // processor operation but an i64 (which is not legal) requires two. So the
7717 // transform should not be done in this case.
7718 if (Value.getOpcode() != ISD::TargetConstantFP) {
7720 switch (CFP->getValueType(0).getSimpleVT().SimpleTy) {
7721 default: llvm_unreachable("Unknown FP type");
7722 case MVT::f16: // We don't do this for these yet.
7728 if ((isTypeLegal(MVT::i32) && !LegalOperations && !ST->isVolatile()) ||
7729 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
7730 Tmp = DAG.getConstant((uint32_t)CFP->getValueAPF().
7731 bitcastToAPInt().getZExtValue(), MVT::i32);
7732 return DAG.getStore(Chain, N->getDebugLoc(), Tmp,
7733 Ptr, ST->getPointerInfo(), ST->isVolatile(),
7734 ST->isNonTemporal(), ST->getAlignment());
7738 if ((TLI.isTypeLegal(MVT::i64) && !LegalOperations &&
7739 !ST->isVolatile()) ||
7740 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i64)) {
7741 Tmp = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
7742 getZExtValue(), MVT::i64);
7743 return DAG.getStore(Chain, N->getDebugLoc(), Tmp,
7744 Ptr, ST->getPointerInfo(), ST->isVolatile(),
7745 ST->isNonTemporal(), ST->getAlignment());
7748 if (!ST->isVolatile() &&
7749 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
7750 // Many FP stores are not made apparent until after legalize, e.g. for
7751 // argument passing. Since this is so common, custom legalize the
7752 // 64-bit integer store into two 32-bit stores.
7753 uint64_t Val = CFP->getValueAPF().bitcastToAPInt().getZExtValue();
7754 SDValue Lo = DAG.getConstant(Val & 0xFFFFFFFF, MVT::i32);
7755 SDValue Hi = DAG.getConstant(Val >> 32, MVT::i32);
7756 if (TLI.isBigEndian()) std::swap(Lo, Hi);
7758 unsigned Alignment = ST->getAlignment();
7759 bool isVolatile = ST->isVolatile();
7760 bool isNonTemporal = ST->isNonTemporal();
7762 SDValue St0 = DAG.getStore(Chain, ST->getDebugLoc(), Lo,
7763 Ptr, ST->getPointerInfo(),
7764 isVolatile, isNonTemporal,
7765 ST->getAlignment());
7766 Ptr = DAG.getNode(ISD::ADD, N->getDebugLoc(), Ptr.getValueType(), Ptr,
7767 DAG.getConstant(4, Ptr.getValueType()));
7768 Alignment = MinAlign(Alignment, 4U);
7769 SDValue St1 = DAG.getStore(Chain, ST->getDebugLoc(), Hi,
7770 Ptr, ST->getPointerInfo().getWithOffset(4),
7771 isVolatile, isNonTemporal,
7773 return DAG.getNode(ISD::TokenFactor, N->getDebugLoc(), MVT::Other,
7782 // Try to infer better alignment information than the store already has.
7783 if (OptLevel != CodeGenOpt::None && ST->isUnindexed()) {
7784 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
7785 if (Align > ST->getAlignment())
7786 return DAG.getTruncStore(Chain, N->getDebugLoc(), Value,
7787 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
7788 ST->isVolatile(), ST->isNonTemporal(), Align);
7792 // Try transforming a pair floating point load / store ops to integer
7793 // load / store ops.
7794 SDValue NewST = TransformFPLoadStorePair(N);
7795 if (NewST.getNode())
7799 // Walk up chain skipping non-aliasing memory nodes.
7800 SDValue BetterChain = FindBetterChain(N, Chain);
7802 // If there is a better chain.
7803 if (Chain != BetterChain) {
7806 // Replace the chain to avoid dependency.
7807 if (ST->isTruncatingStore()) {
7808 ReplStore = DAG.getTruncStore(BetterChain, N->getDebugLoc(), Value, Ptr,
7809 ST->getPointerInfo(),
7810 ST->getMemoryVT(), ST->isVolatile(),
7811 ST->isNonTemporal(), ST->getAlignment());
7813 ReplStore = DAG.getStore(BetterChain, N->getDebugLoc(), Value, Ptr,
7814 ST->getPointerInfo(),
7815 ST->isVolatile(), ST->isNonTemporal(),
7816 ST->getAlignment());
7819 // Create token to keep both nodes around.
7820 SDValue Token = DAG.getNode(ISD::TokenFactor, N->getDebugLoc(),
7821 MVT::Other, Chain, ReplStore);
7823 // Make sure the new and old chains are cleaned up.
7824 AddToWorkList(Token.getNode());
7826 // Don't add users to work list.
7827 return CombineTo(N, Token, false);
7831 // Try transforming N to an indexed store.
7832 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
7833 return SDValue(N, 0);
7835 // FIXME: is there such a thing as a truncating indexed store?
7836 if (ST->isTruncatingStore() && ST->isUnindexed() &&
7837 Value.getValueType().isInteger()) {
7838 // See if we can simplify the input to this truncstore with knowledge that
7839 // only the low bits are being used. For example:
7840 // "truncstore (or (shl x, 8), y), i8" -> "truncstore y, i8"
7842 GetDemandedBits(Value,
7843 APInt::getLowBitsSet(
7844 Value.getValueType().getScalarType().getSizeInBits(),
7845 ST->getMemoryVT().getScalarType().getSizeInBits()));
7846 AddToWorkList(Value.getNode());
7847 if (Shorter.getNode())
7848 return DAG.getTruncStore(Chain, N->getDebugLoc(), Shorter,
7849 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
7850 ST->isVolatile(), ST->isNonTemporal(),
7851 ST->getAlignment());
7853 // Otherwise, see if we can simplify the operation with
7854 // SimplifyDemandedBits, which only works if the value has a single use.
7855 if (SimplifyDemandedBits(Value,
7856 APInt::getLowBitsSet(
7857 Value.getValueType().getScalarType().getSizeInBits(),
7858 ST->getMemoryVT().getScalarType().getSizeInBits())))
7859 return SDValue(N, 0);
7862 // If this is a load followed by a store to the same location, then the store
7864 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Value)) {
7865 if (Ld->getBasePtr() == Ptr && ST->getMemoryVT() == Ld->getMemoryVT() &&
7866 ST->isUnindexed() && !ST->isVolatile() &&
7867 // There can't be any side effects between the load and store, such as
7869 Chain.reachesChainWithoutSideEffects(SDValue(Ld, 1))) {
7870 // The store is dead, remove it.
7875 // If this is an FP_ROUND or TRUNC followed by a store, fold this into a
7876 // truncating store. We can do this even if this is already a truncstore.
7877 if ((Value.getOpcode() == ISD::FP_ROUND || Value.getOpcode() == ISD::TRUNCATE)
7878 && Value.getNode()->hasOneUse() && ST->isUnindexed() &&
7879 TLI.isTruncStoreLegal(Value.getOperand(0).getValueType(),
7880 ST->getMemoryVT())) {
7881 return DAG.getTruncStore(Chain, N->getDebugLoc(), Value.getOperand(0),
7882 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
7883 ST->isVolatile(), ST->isNonTemporal(),
7884 ST->getAlignment());
7888 // Only perform this optimization before the types are legal, because we
7889 // don't want to generate illegal types in this optimization.
7890 if (!LegalTypes && MergeConsecutiveStores(ST))
7891 return SDValue(N, 0);
7893 return ReduceLoadOpStoreWidth(N);
7896 SDValue DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) {
7897 SDValue InVec = N->getOperand(0);
7898 SDValue InVal = N->getOperand(1);
7899 SDValue EltNo = N->getOperand(2);
7900 DebugLoc dl = N->getDebugLoc();
7902 // If the inserted element is an UNDEF, just use the input vector.
7903 if (InVal.getOpcode() == ISD::UNDEF)
7906 EVT VT = InVec.getValueType();
7908 // If we can't generate a legal BUILD_VECTOR, exit
7909 if (LegalOperations && !TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
7912 // Check that we know which element is being inserted
7913 if (!isa<ConstantSDNode>(EltNo))
7915 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
7917 // Check that the operand is a BUILD_VECTOR (or UNDEF, which can essentially
7918 // be converted to a BUILD_VECTOR). Fill in the Ops vector with the
7920 SmallVector<SDValue, 8> Ops;
7921 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
7922 Ops.append(InVec.getNode()->op_begin(),
7923 InVec.getNode()->op_end());
7924 } else if (InVec.getOpcode() == ISD::UNDEF) {
7925 unsigned NElts = VT.getVectorNumElements();
7926 Ops.append(NElts, DAG.getUNDEF(InVal.getValueType()));
7931 // Insert the element
7932 if (Elt < Ops.size()) {
7933 // All the operands of BUILD_VECTOR must have the same type;
7934 // we enforce that here.
7935 EVT OpVT = Ops[0].getValueType();
7936 if (InVal.getValueType() != OpVT)
7937 InVal = OpVT.bitsGT(InVal.getValueType()) ?
7938 DAG.getNode(ISD::ANY_EXTEND, dl, OpVT, InVal) :
7939 DAG.getNode(ISD::TRUNCATE, dl, OpVT, InVal);
7943 // Return the new vector
7944 return DAG.getNode(ISD::BUILD_VECTOR, dl,
7945 VT, &Ops[0], Ops.size());
7948 SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) {
7949 // (vextract (scalar_to_vector val, 0) -> val
7950 SDValue InVec = N->getOperand(0);
7951 EVT VT = InVec.getValueType();
7952 EVT NVT = N->getValueType(0);
7954 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
7955 // Check if the result type doesn't match the inserted element type. A
7956 // SCALAR_TO_VECTOR may truncate the inserted element and the
7957 // EXTRACT_VECTOR_ELT may widen the extracted vector.
7958 SDValue InOp = InVec.getOperand(0);
7959 if (InOp.getValueType() != NVT) {
7960 assert(InOp.getValueType().isInteger() && NVT.isInteger());
7961 return DAG.getSExtOrTrunc(InOp, InVec.getDebugLoc(), NVT);
7966 SDValue EltNo = N->getOperand(1);
7967 bool ConstEltNo = isa<ConstantSDNode>(EltNo);
7969 // Transform: (EXTRACT_VECTOR_ELT( VECTOR_SHUFFLE )) -> EXTRACT_VECTOR_ELT.
7970 // We only perform this optimization before the op legalization phase because
7971 // we may introduce new vector instructions which are not backed by TD
7972 // patterns. For example on AVX, extracting elements from a wide vector
7973 // without using extract_subvector.
7974 if (InVec.getOpcode() == ISD::VECTOR_SHUFFLE
7975 && ConstEltNo && !LegalOperations) {
7976 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
7977 int NumElem = VT.getVectorNumElements();
7978 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(InVec);
7979 // Find the new index to extract from.
7980 int OrigElt = SVOp->getMaskElt(Elt);
7982 // Extracting an undef index is undef.
7984 return DAG.getUNDEF(NVT);
7986 // Select the right vector half to extract from.
7987 if (OrigElt < NumElem) {
7988 InVec = InVec->getOperand(0);
7990 InVec = InVec->getOperand(1);
7994 EVT IndexTy = N->getOperand(1).getValueType();
7995 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, N->getDebugLoc(), NVT,
7996 InVec, DAG.getConstant(OrigElt, IndexTy));
7999 // Perform only after legalization to ensure build_vector / vector_shuffle
8000 // optimizations have already been done.
8001 if (!LegalOperations) return SDValue();
8003 // (vextract (v4f32 load $addr), c) -> (f32 load $addr+c*size)
8004 // (vextract (v4f32 s2v (f32 load $addr)), c) -> (f32 load $addr+c*size)
8005 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), 0) -> (f32 load $addr)
8008 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
8009 bool NewLoad = false;
8010 bool BCNumEltsChanged = false;
8011 EVT ExtVT = VT.getVectorElementType();
8014 // If the result of load has to be truncated, then it's not necessarily
8016 if (NVT.bitsLT(LVT) && !TLI.isTruncateFree(LVT, NVT))
8019 if (InVec.getOpcode() == ISD::BITCAST) {
8020 // Don't duplicate a load with other uses.
8021 if (!InVec.hasOneUse())
8024 EVT BCVT = InVec.getOperand(0).getValueType();
8025 if (!BCVT.isVector() || ExtVT.bitsGT(BCVT.getVectorElementType()))
8027 if (VT.getVectorNumElements() != BCVT.getVectorNumElements())
8028 BCNumEltsChanged = true;
8029 InVec = InVec.getOperand(0);
8030 ExtVT = BCVT.getVectorElementType();
8034 LoadSDNode *LN0 = NULL;
8035 const ShuffleVectorSDNode *SVN = NULL;
8036 if (ISD::isNormalLoad(InVec.getNode())) {
8037 LN0 = cast<LoadSDNode>(InVec);
8038 } else if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR &&
8039 InVec.getOperand(0).getValueType() == ExtVT &&
8040 ISD::isNormalLoad(InVec.getOperand(0).getNode())) {
8041 // Don't duplicate a load with other uses.
8042 if (!InVec.hasOneUse())
8045 LN0 = cast<LoadSDNode>(InVec.getOperand(0));
8046 } else if ((SVN = dyn_cast<ShuffleVectorSDNode>(InVec))) {
8047 // (vextract (vector_shuffle (load $addr), v2, <1, u, u, u>), 1)
8049 // (load $addr+1*size)
8051 // Don't duplicate a load with other uses.
8052 if (!InVec.hasOneUse())
8055 // If the bit convert changed the number of elements, it is unsafe
8056 // to examine the mask.
8057 if (BCNumEltsChanged)
8060 // Select the input vector, guarding against out of range extract vector.
8061 unsigned NumElems = VT.getVectorNumElements();
8062 int Idx = (Elt > (int)NumElems) ? -1 : SVN->getMaskElt(Elt);
8063 InVec = (Idx < (int)NumElems) ? InVec.getOperand(0) : InVec.getOperand(1);
8065 if (InVec.getOpcode() == ISD::BITCAST) {
8066 // Don't duplicate a load with other uses.
8067 if (!InVec.hasOneUse())
8070 InVec = InVec.getOperand(0);
8072 if (ISD::isNormalLoad(InVec.getNode())) {
8073 LN0 = cast<LoadSDNode>(InVec);
8074 Elt = (Idx < (int)NumElems) ? Idx : Idx - (int)NumElems;
8078 // Make sure we found a non-volatile load and the extractelement is
8080 if (!LN0 || !LN0->hasNUsesOfValue(1,0) || LN0->isVolatile())
8083 // If Idx was -1 above, Elt is going to be -1, so just return undef.
8085 return DAG.getUNDEF(LVT);
8087 unsigned Align = LN0->getAlignment();
8089 // Check the resultant load doesn't need a higher alignment than the
8093 ->getABITypeAlignment(LVT.getTypeForEVT(*DAG.getContext()));
8095 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, LVT))
8101 SDValue NewPtr = LN0->getBasePtr();
8102 unsigned PtrOff = 0;
8105 PtrOff = LVT.getSizeInBits() * Elt / 8;
8106 EVT PtrType = NewPtr.getValueType();
8107 if (TLI.isBigEndian())
8108 PtrOff = VT.getSizeInBits() / 8 - PtrOff;
8109 NewPtr = DAG.getNode(ISD::ADD, N->getDebugLoc(), PtrType, NewPtr,
8110 DAG.getConstant(PtrOff, PtrType));
8113 // The replacement we need to do here is a little tricky: we need to
8114 // replace an extractelement of a load with a load.
8115 // Use ReplaceAllUsesOfValuesWith to do the replacement.
8116 // Note that this replacement assumes that the extractvalue is the only
8117 // use of the load; that's okay because we don't want to perform this
8118 // transformation in other cases anyway.
8121 if (NVT.bitsGT(LVT)) {
8122 // If the result type of vextract is wider than the load, then issue an
8123 // extending load instead.
8124 ISD::LoadExtType ExtType = TLI.isLoadExtLegal(ISD::ZEXTLOAD, LVT)
8125 ? ISD::ZEXTLOAD : ISD::EXTLOAD;
8126 Load = DAG.getExtLoad(ExtType, N->getDebugLoc(), NVT, LN0->getChain(),
8127 NewPtr, LN0->getPointerInfo().getWithOffset(PtrOff),
8128 LVT, LN0->isVolatile(), LN0->isNonTemporal(),Align);
8129 Chain = Load.getValue(1);
8131 Load = DAG.getLoad(LVT, N->getDebugLoc(), LN0->getChain(), NewPtr,
8132 LN0->getPointerInfo().getWithOffset(PtrOff),
8133 LN0->isVolatile(), LN0->isNonTemporal(),
8134 LN0->isInvariant(), Align);
8135 Chain = Load.getValue(1);
8136 if (NVT.bitsLT(LVT))
8137 Load = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), NVT, Load);
8139 Load = DAG.getNode(ISD::BITCAST, N->getDebugLoc(), NVT, Load);
8141 WorkListRemover DeadNodes(*this);
8142 SDValue From[] = { SDValue(N, 0), SDValue(LN0,1) };
8143 SDValue To[] = { Load, Chain };
8144 DAG.ReplaceAllUsesOfValuesWith(From, To, 2);
8145 // Since we're explcitly calling ReplaceAllUses, add the new node to the
8146 // worklist explicitly as well.
8147 AddToWorkList(Load.getNode());
8148 AddUsersToWorkList(Load.getNode()); // Add users too
8149 // Make sure to revisit this node to clean it up; it will usually be dead.
8151 return SDValue(N, 0);
8157 SDValue DAGCombiner::visitBUILD_VECTOR(SDNode *N) {
8158 unsigned NumInScalars = N->getNumOperands();
8159 DebugLoc dl = N->getDebugLoc();
8160 EVT VT = N->getValueType(0);
8162 // A vector built entirely of undefs is undef.
8163 if (ISD::allOperandsUndef(N))
8164 return DAG.getUNDEF(VT);
8166 // Check to see if this is a BUILD_VECTOR of a bunch of values
8167 // which come from any_extend or zero_extend nodes. If so, we can create
8168 // a new BUILD_VECTOR using bit-casts which may enable other BUILD_VECTOR
8169 // optimizations. We do not handle sign-extend because we can't fill the sign
8171 EVT SourceType = MVT::Other;
8172 bool AllAnyExt = true;
8174 for (unsigned i = 0; i != NumInScalars; ++i) {
8175 SDValue In = N->getOperand(i);
8176 // Ignore undef inputs.
8177 if (In.getOpcode() == ISD::UNDEF) continue;
8179 bool AnyExt = In.getOpcode() == ISD::ANY_EXTEND;
8180 bool ZeroExt = In.getOpcode() == ISD::ZERO_EXTEND;
8182 // Abort if the element is not an extension.
8183 if (!ZeroExt && !AnyExt) {
8184 SourceType = MVT::Other;
8188 // The input is a ZeroExt or AnyExt. Check the original type.
8189 EVT InTy = In.getOperand(0).getValueType();
8191 // Check that all of the widened source types are the same.
8192 if (SourceType == MVT::Other)
8195 else if (InTy != SourceType) {
8196 // Multiple income types. Abort.
8197 SourceType = MVT::Other;
8201 // Check if all of the extends are ANY_EXTENDs.
8202 AllAnyExt &= AnyExt;
8205 // In order to have valid types, all of the inputs must be extended from the
8206 // same source type and all of the inputs must be any or zero extend.
8207 // Scalar sizes must be a power of two.
8208 EVT OutScalarTy = N->getValueType(0).getScalarType();
8209 bool ValidTypes = SourceType != MVT::Other &&
8210 isPowerOf2_32(OutScalarTy.getSizeInBits()) &&
8211 isPowerOf2_32(SourceType.getSizeInBits());
8213 // We perform this optimization post type-legalization because
8214 // the type-legalizer often scalarizes integer-promoted vectors.
8215 // Performing this optimization before may create bit-casts which
8216 // will be type-legalized to complex code sequences.
8217 // We perform this optimization only before the operation legalizer because we
8218 // may introduce illegal operations.
8219 // Create a new simpler BUILD_VECTOR sequence which other optimizations can
8220 // turn into a single shuffle instruction.
8221 if ((Level == AfterLegalizeVectorOps || Level == AfterLegalizeTypes) &&
8223 bool isLE = TLI.isLittleEndian();
8224 unsigned ElemRatio = OutScalarTy.getSizeInBits()/SourceType.getSizeInBits();
8225 assert(ElemRatio > 1 && "Invalid element size ratio");
8226 SDValue Filler = AllAnyExt ? DAG.getUNDEF(SourceType):
8227 DAG.getConstant(0, SourceType);
8229 unsigned NewBVElems = ElemRatio * N->getValueType(0).getVectorNumElements();
8230 SmallVector<SDValue, 8> Ops(NewBVElems, Filler);
8232 // Populate the new build_vector
8233 for (unsigned i=0; i < N->getNumOperands(); ++i) {
8234 SDValue Cast = N->getOperand(i);
8235 assert((Cast.getOpcode() == ISD::ANY_EXTEND ||
8236 Cast.getOpcode() == ISD::ZERO_EXTEND ||
8237 Cast.getOpcode() == ISD::UNDEF) && "Invalid cast opcode");
8239 if (Cast.getOpcode() == ISD::UNDEF)
8240 In = DAG.getUNDEF(SourceType);
8242 In = Cast->getOperand(0);
8243 unsigned Index = isLE ? (i * ElemRatio) :
8244 (i * ElemRatio + (ElemRatio - 1));
8246 assert(Index < Ops.size() && "Invalid index");
8250 // The type of the new BUILD_VECTOR node.
8251 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), SourceType, NewBVElems);
8252 assert(VecVT.getSizeInBits() == N->getValueType(0).getSizeInBits() &&
8253 "Invalid vector size");
8254 // Check if the new vector type is legal.
8255 if (!isTypeLegal(VecVT)) return SDValue();
8257 // Make the new BUILD_VECTOR.
8258 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
8259 VecVT, &Ops[0], Ops.size());
8261 // The new BUILD_VECTOR node has the potential to be further optimized.
8262 AddToWorkList(BV.getNode());
8263 // Bitcast to the desired type.
8264 return DAG.getNode(ISD::BITCAST, dl, N->getValueType(0), BV);
8267 // Check to see if this is a BUILD_VECTOR of a bunch of EXTRACT_VECTOR_ELT
8268 // operations. If so, and if the EXTRACT_VECTOR_ELT vector inputs come from
8269 // at most two distinct vectors, turn this into a shuffle node.
8271 // May only combine to shuffle after legalize if shuffle is legal.
8272 if (LegalOperations &&
8273 !TLI.isOperationLegalOrCustom(ISD::VECTOR_SHUFFLE, VT))
8276 SDValue VecIn1, VecIn2;
8277 for (unsigned i = 0; i != NumInScalars; ++i) {
8278 // Ignore undef inputs.
8279 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
8281 // If this input is something other than a EXTRACT_VECTOR_ELT with a
8282 // constant index, bail out.
8283 if (N->getOperand(i).getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
8284 !isa<ConstantSDNode>(N->getOperand(i).getOperand(1))) {
8285 VecIn1 = VecIn2 = SDValue(0, 0);
8289 // We allow up to two distinct input vectors.
8290 SDValue ExtractedFromVec = N->getOperand(i).getOperand(0);
8291 if (ExtractedFromVec == VecIn1 || ExtractedFromVec == VecIn2)
8294 if (VecIn1.getNode() == 0) {
8295 VecIn1 = ExtractedFromVec;
8296 } else if (VecIn2.getNode() == 0) {
8297 VecIn2 = ExtractedFromVec;
8300 VecIn1 = VecIn2 = SDValue(0, 0);
8305 // If everything is good, we can make a shuffle operation.
8306 if (VecIn1.getNode()) {
8307 SmallVector<int, 8> Mask;
8308 for (unsigned i = 0; i != NumInScalars; ++i) {
8309 if (N->getOperand(i).getOpcode() == ISD::UNDEF) {
8314 // If extracting from the first vector, just use the index directly.
8315 SDValue Extract = N->getOperand(i);
8316 SDValue ExtVal = Extract.getOperand(1);
8317 if (Extract.getOperand(0) == VecIn1) {
8318 unsigned ExtIndex = cast<ConstantSDNode>(ExtVal)->getZExtValue();
8319 if (ExtIndex > VT.getVectorNumElements())
8322 Mask.push_back(ExtIndex);
8326 // Otherwise, use InIdx + VecSize
8327 unsigned Idx = cast<ConstantSDNode>(ExtVal)->getZExtValue();
8328 Mask.push_back(Idx+NumInScalars);
8331 // We can't generate a shuffle node with mismatched input and output types.
8332 // Attempt to transform a single input vector to the correct type.
8333 if ((VT != VecIn1.getValueType())) {
8334 // We don't support shuffeling between TWO values of different types.
8335 if (VecIn2.getNode() != 0)
8338 // We only support widening of vectors which are half the size of the
8339 // output registers. For example XMM->YMM widening on X86 with AVX.
8340 if (VecIn1.getValueType().getSizeInBits()*2 != VT.getSizeInBits())
8343 // If the input vector type has a different base type to the output
8344 // vector type, bail out.
8345 if (VecIn1.getValueType().getVectorElementType() !=
8346 VT.getVectorElementType())
8349 // Widen the input vector by adding undef values.
8350 VecIn1 = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
8351 VecIn1, DAG.getUNDEF(VecIn1.getValueType()));
8354 // If VecIn2 is unused then change it to undef.
8355 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
8357 // Check that we were able to transform all incoming values to the same
8359 if (VecIn2.getValueType() != VecIn1.getValueType() ||
8360 VecIn1.getValueType() != VT)
8363 // Only type-legal BUILD_VECTOR nodes are converted to shuffle nodes.
8364 if (!isTypeLegal(VT))
8367 // Return the new VECTOR_SHUFFLE node.
8371 return DAG.getVectorShuffle(VT, N->getDebugLoc(), Ops[0], Ops[1], &Mask[0]);
8377 SDValue DAGCombiner::visitCONCAT_VECTORS(SDNode *N) {
8378 // TODO: Check to see if this is a CONCAT_VECTORS of a bunch of
8379 // EXTRACT_SUBVECTOR operations. If so, and if the EXTRACT_SUBVECTOR vector
8380 // inputs come from at most two distinct vectors, turn this into a shuffle
8383 // If we only have one input vector, we don't need to do any concatenation.
8384 if (N->getNumOperands() == 1)
8385 return N->getOperand(0);
8387 // Check if all of the operands are undefs.
8388 if (ISD::allOperandsUndef(N))
8389 return DAG.getUNDEF(N->getValueType(0));
8394 SDValue DAGCombiner::visitEXTRACT_SUBVECTOR(SDNode* N) {
8395 EVT NVT = N->getValueType(0);
8396 SDValue V = N->getOperand(0);
8398 if (V->getOpcode() == ISD::INSERT_SUBVECTOR) {
8399 // Handle only simple case where vector being inserted and vector
8400 // being extracted are of same type, and are half size of larger vectors.
8401 EVT BigVT = V->getOperand(0).getValueType();
8402 EVT SmallVT = V->getOperand(1).getValueType();
8403 if (NVT != SmallVT || NVT.getSizeInBits()*2 != BigVT.getSizeInBits())
8406 // Only handle cases where both indexes are constants with the same type.
8407 ConstantSDNode *InsIdx = dyn_cast<ConstantSDNode>(N->getOperand(1));
8408 ConstantSDNode *ExtIdx = dyn_cast<ConstantSDNode>(V->getOperand(2));
8410 if (InsIdx && ExtIdx &&
8411 InsIdx->getValueType(0).getSizeInBits() <= 64 &&
8412 ExtIdx->getValueType(0).getSizeInBits() <= 64) {
8414 // (extract_subvec (insert_subvec V1, V2, InsIdx), ExtIdx)
8416 // indices are equal => V1
8417 // otherwise => (extract_subvec V1, ExtIdx)
8418 if (InsIdx->getZExtValue() == ExtIdx->getZExtValue())
8419 return V->getOperand(1);
8420 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, N->getDebugLoc(), NVT,
8421 V->getOperand(0), N->getOperand(1));
8428 SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
8429 EVT VT = N->getValueType(0);
8430 unsigned NumElts = VT.getVectorNumElements();
8432 SDValue N0 = N->getOperand(0);
8433 SDValue N1 = N->getOperand(1);
8435 assert(N0.getValueType() == VT && "Vector shuffle must be normalized in DAG");
8437 // Canonicalize shuffle undef, undef -> undef
8438 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
8439 return DAG.getUNDEF(VT);
8441 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8443 // Canonicalize shuffle v, v -> v, undef
8445 SmallVector<int, 8> NewMask;
8446 for (unsigned i = 0; i != NumElts; ++i) {
8447 int Idx = SVN->getMaskElt(i);
8448 if (Idx >= (int)NumElts) Idx -= NumElts;
8449 NewMask.push_back(Idx);
8451 return DAG.getVectorShuffle(VT, N->getDebugLoc(), N0, DAG.getUNDEF(VT),
8455 // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask.
8456 if (N0.getOpcode() == ISD::UNDEF) {
8457 SmallVector<int, 8> NewMask;
8458 for (unsigned i = 0; i != NumElts; ++i) {
8459 int Idx = SVN->getMaskElt(i);
8461 if (Idx < (int)NumElts)
8466 NewMask.push_back(Idx);
8468 return DAG.getVectorShuffle(VT, N->getDebugLoc(), N1, DAG.getUNDEF(VT),
8472 // Remove references to rhs if it is undef
8473 if (N1.getOpcode() == ISD::UNDEF) {
8474 bool Changed = false;
8475 SmallVector<int, 8> NewMask;
8476 for (unsigned i = 0; i != NumElts; ++i) {
8477 int Idx = SVN->getMaskElt(i);
8478 if (Idx >= (int)NumElts) {
8482 NewMask.push_back(Idx);
8485 return DAG.getVectorShuffle(VT, N->getDebugLoc(), N0, N1, &NewMask[0]);
8488 // If it is a splat, check if the argument vector is another splat or a
8489 // build_vector with all scalar elements the same.
8490 if (SVN->isSplat() && SVN->getSplatIndex() < (int)NumElts) {
8491 SDNode *V = N0.getNode();
8493 // If this is a bit convert that changes the element type of the vector but
8494 // not the number of vector elements, look through it. Be careful not to
8495 // look though conversions that change things like v4f32 to v2f64.
8496 if (V->getOpcode() == ISD::BITCAST) {
8497 SDValue ConvInput = V->getOperand(0);
8498 if (ConvInput.getValueType().isVector() &&
8499 ConvInput.getValueType().getVectorNumElements() == NumElts)
8500 V = ConvInput.getNode();
8503 if (V->getOpcode() == ISD::BUILD_VECTOR) {
8504 assert(V->getNumOperands() == NumElts &&
8505 "BUILD_VECTOR has wrong number of operands");
8507 bool AllSame = true;
8508 for (unsigned i = 0; i != NumElts; ++i) {
8509 if (V->getOperand(i).getOpcode() != ISD::UNDEF) {
8510 Base = V->getOperand(i);
8514 // Splat of <u, u, u, u>, return <u, u, u, u>
8515 if (!Base.getNode())
8517 for (unsigned i = 0; i != NumElts; ++i) {
8518 if (V->getOperand(i) != Base) {
8523 // Splat of <x, x, x, x>, return <x, x, x, x>
8529 // If this shuffle node is simply a swizzle of another shuffle node,
8530 // and it reverses the swizzle of the previous shuffle then we can
8531 // optimize shuffle(shuffle(x, undef), undef) -> x.
8532 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
8533 N1.getOpcode() == ISD::UNDEF) {
8535 ShuffleVectorSDNode *OtherSV = cast<ShuffleVectorSDNode>(N0);
8537 // Shuffle nodes can only reverse shuffles with a single non-undef value.
8538 if (N0.getOperand(1).getOpcode() != ISD::UNDEF)
8541 // The incoming shuffle must be of the same type as the result of the
8543 assert(OtherSV->getOperand(0).getValueType() == VT &&
8544 "Shuffle types don't match");
8546 for (unsigned i = 0; i != NumElts; ++i) {
8547 int Idx = SVN->getMaskElt(i);
8548 assert(Idx < (int)NumElts && "Index references undef operand");
8549 // Next, this index comes from the first value, which is the incoming
8550 // shuffle. Adopt the incoming index.
8552 Idx = OtherSV->getMaskElt(Idx);
8554 // The combined shuffle must map each index to itself.
8555 if (Idx >= 0 && (unsigned)Idx != i)
8559 return OtherSV->getOperand(0);
8565 SDValue DAGCombiner::visitMEMBARRIER(SDNode* N) {
8566 if (!TLI.getShouldFoldAtomicFences())
8569 SDValue atomic = N->getOperand(0);
8570 switch (atomic.getOpcode()) {
8571 case ISD::ATOMIC_CMP_SWAP:
8572 case ISD::ATOMIC_SWAP:
8573 case ISD::ATOMIC_LOAD_ADD:
8574 case ISD::ATOMIC_LOAD_SUB:
8575 case ISD::ATOMIC_LOAD_AND:
8576 case ISD::ATOMIC_LOAD_OR:
8577 case ISD::ATOMIC_LOAD_XOR:
8578 case ISD::ATOMIC_LOAD_NAND:
8579 case ISD::ATOMIC_LOAD_MIN:
8580 case ISD::ATOMIC_LOAD_MAX:
8581 case ISD::ATOMIC_LOAD_UMIN:
8582 case ISD::ATOMIC_LOAD_UMAX:
8588 SDValue fence = atomic.getOperand(0);
8589 if (fence.getOpcode() != ISD::MEMBARRIER)
8592 switch (atomic.getOpcode()) {
8593 case ISD::ATOMIC_CMP_SWAP:
8594 return SDValue(DAG.UpdateNodeOperands(atomic.getNode(),
8595 fence.getOperand(0),
8596 atomic.getOperand(1), atomic.getOperand(2),
8597 atomic.getOperand(3)), atomic.getResNo());
8598 case ISD::ATOMIC_SWAP:
8599 case ISD::ATOMIC_LOAD_ADD:
8600 case ISD::ATOMIC_LOAD_SUB:
8601 case ISD::ATOMIC_LOAD_AND:
8602 case ISD::ATOMIC_LOAD_OR:
8603 case ISD::ATOMIC_LOAD_XOR:
8604 case ISD::ATOMIC_LOAD_NAND:
8605 case ISD::ATOMIC_LOAD_MIN:
8606 case ISD::ATOMIC_LOAD_MAX:
8607 case ISD::ATOMIC_LOAD_UMIN:
8608 case ISD::ATOMIC_LOAD_UMAX:
8609 return SDValue(DAG.UpdateNodeOperands(atomic.getNode(),
8610 fence.getOperand(0),
8611 atomic.getOperand(1), atomic.getOperand(2)),
8618 /// XformToShuffleWithZero - Returns a vector_shuffle if it able to transform
8619 /// an AND to a vector_shuffle with the destination vector and a zero vector.
8620 /// e.g. AND V, <0xffffffff, 0, 0xffffffff, 0>. ==>
8621 /// vector_shuffle V, Zero, <0, 4, 2, 4>
8622 SDValue DAGCombiner::XformToShuffleWithZero(SDNode *N) {
8623 EVT VT = N->getValueType(0);
8624 DebugLoc dl = N->getDebugLoc();
8625 SDValue LHS = N->getOperand(0);
8626 SDValue RHS = N->getOperand(1);
8627 if (N->getOpcode() == ISD::AND) {
8628 if (RHS.getOpcode() == ISD::BITCAST)
8629 RHS = RHS.getOperand(0);
8630 if (RHS.getOpcode() == ISD::BUILD_VECTOR) {
8631 SmallVector<int, 8> Indices;
8632 unsigned NumElts = RHS.getNumOperands();
8633 for (unsigned i = 0; i != NumElts; ++i) {
8634 SDValue Elt = RHS.getOperand(i);
8635 if (!isa<ConstantSDNode>(Elt))
8638 if (cast<ConstantSDNode>(Elt)->isAllOnesValue())
8639 Indices.push_back(i);
8640 else if (cast<ConstantSDNode>(Elt)->isNullValue())
8641 Indices.push_back(NumElts);
8646 // Let's see if the target supports this vector_shuffle.
8647 EVT RVT = RHS.getValueType();
8648 if (!TLI.isVectorClearMaskLegal(Indices, RVT))
8651 // Return the new VECTOR_SHUFFLE node.
8652 EVT EltVT = RVT.getVectorElementType();
8653 SmallVector<SDValue,8> ZeroOps(RVT.getVectorNumElements(),
8654 DAG.getConstant(0, EltVT));
8655 SDValue Zero = DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
8656 RVT, &ZeroOps[0], ZeroOps.size());
8657 LHS = DAG.getNode(ISD::BITCAST, dl, RVT, LHS);
8658 SDValue Shuf = DAG.getVectorShuffle(RVT, dl, LHS, Zero, &Indices[0]);
8659 return DAG.getNode(ISD::BITCAST, dl, VT, Shuf);
8666 /// SimplifyVBinOp - Visit a binary vector operation, like ADD.
8667 SDValue DAGCombiner::SimplifyVBinOp(SDNode *N) {
8668 // After legalize, the target may be depending on adds and other
8669 // binary ops to provide legal ways to construct constants or other
8670 // things. Simplifying them may result in a loss of legality.
8671 if (LegalOperations) return SDValue();
8673 assert(N->getValueType(0).isVector() &&
8674 "SimplifyVBinOp only works on vectors!");
8676 SDValue LHS = N->getOperand(0);
8677 SDValue RHS = N->getOperand(1);
8678 SDValue Shuffle = XformToShuffleWithZero(N);
8679 if (Shuffle.getNode()) return Shuffle;
8681 // If the LHS and RHS are BUILD_VECTOR nodes, see if we can constant fold
8683 if (LHS.getOpcode() == ISD::BUILD_VECTOR &&
8684 RHS.getOpcode() == ISD::BUILD_VECTOR) {
8685 SmallVector<SDValue, 8> Ops;
8686 for (unsigned i = 0, e = LHS.getNumOperands(); i != e; ++i) {
8687 SDValue LHSOp = LHS.getOperand(i);
8688 SDValue RHSOp = RHS.getOperand(i);
8689 // If these two elements can't be folded, bail out.
8690 if ((LHSOp.getOpcode() != ISD::UNDEF &&
8691 LHSOp.getOpcode() != ISD::Constant &&
8692 LHSOp.getOpcode() != ISD::ConstantFP) ||
8693 (RHSOp.getOpcode() != ISD::UNDEF &&
8694 RHSOp.getOpcode() != ISD::Constant &&
8695 RHSOp.getOpcode() != ISD::ConstantFP))
8698 // Can't fold divide by zero.
8699 if (N->getOpcode() == ISD::SDIV || N->getOpcode() == ISD::UDIV ||
8700 N->getOpcode() == ISD::FDIV) {
8701 if ((RHSOp.getOpcode() == ISD::Constant &&
8702 cast<ConstantSDNode>(RHSOp.getNode())->isNullValue()) ||
8703 (RHSOp.getOpcode() == ISD::ConstantFP &&
8704 cast<ConstantFPSDNode>(RHSOp.getNode())->getValueAPF().isZero()))
8708 EVT VT = LHSOp.getValueType();
8709 EVT RVT = RHSOp.getValueType();
8711 // Integer BUILD_VECTOR operands may have types larger than the element
8712 // size (e.g., when the element type is not legal). Prior to type
8713 // legalization, the types may not match between the two BUILD_VECTORS.
8714 // Truncate one of the operands to make them match.
8715 if (RVT.getSizeInBits() > VT.getSizeInBits()) {
8716 RHSOp = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, RHSOp);
8718 LHSOp = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), RVT, LHSOp);
8722 SDValue FoldOp = DAG.getNode(N->getOpcode(), LHS.getDebugLoc(), VT,
8724 if (FoldOp.getOpcode() != ISD::UNDEF &&
8725 FoldOp.getOpcode() != ISD::Constant &&
8726 FoldOp.getOpcode() != ISD::ConstantFP)
8728 Ops.push_back(FoldOp);
8729 AddToWorkList(FoldOp.getNode());
8732 if (Ops.size() == LHS.getNumOperands())
8733 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
8734 LHS.getValueType(), &Ops[0], Ops.size());
8740 /// SimplifyVUnaryOp - Visit a binary vector operation, like FABS/FNEG.
8741 SDValue DAGCombiner::SimplifyVUnaryOp(SDNode *N) {
8742 // After legalize, the target may be depending on adds and other
8743 // binary ops to provide legal ways to construct constants or other
8744 // things. Simplifying them may result in a loss of legality.
8745 if (LegalOperations) return SDValue();
8747 assert(N->getValueType(0).isVector() &&
8748 "SimplifyVUnaryOp only works on vectors!");
8750 SDValue N0 = N->getOperand(0);
8752 if (N0.getOpcode() != ISD::BUILD_VECTOR)
8755 // Operand is a BUILD_VECTOR node, see if we can constant fold it.
8756 SmallVector<SDValue, 8> Ops;
8757 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) {
8758 SDValue Op = N0.getOperand(i);
8759 if (Op.getOpcode() != ISD::UNDEF &&
8760 Op.getOpcode() != ISD::ConstantFP)
8762 EVT EltVT = Op.getValueType();
8763 SDValue FoldOp = DAG.getNode(N->getOpcode(), N0.getDebugLoc(), EltVT, Op);
8764 if (FoldOp.getOpcode() != ISD::UNDEF &&
8765 FoldOp.getOpcode() != ISD::ConstantFP)
8767 Ops.push_back(FoldOp);
8768 AddToWorkList(FoldOp.getNode());
8771 if (Ops.size() != N0.getNumOperands())
8774 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
8775 N0.getValueType(), &Ops[0], Ops.size());
8778 SDValue DAGCombiner::SimplifySelect(DebugLoc DL, SDValue N0,
8779 SDValue N1, SDValue N2){
8780 assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!");
8782 SDValue SCC = SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1), N1, N2,
8783 cast<CondCodeSDNode>(N0.getOperand(2))->get());
8785 // If we got a simplified select_cc node back from SimplifySelectCC, then
8786 // break it down into a new SETCC node, and a new SELECT node, and then return
8787 // the SELECT node, since we were called with a SELECT node.
8788 if (SCC.getNode()) {
8789 // Check to see if we got a select_cc back (to turn into setcc/select).
8790 // Otherwise, just return whatever node we got back, like fabs.
8791 if (SCC.getOpcode() == ISD::SELECT_CC) {
8792 SDValue SETCC = DAG.getNode(ISD::SETCC, N0.getDebugLoc(),
8794 SCC.getOperand(0), SCC.getOperand(1),
8796 AddToWorkList(SETCC.getNode());
8797 return DAG.getNode(ISD::SELECT, SCC.getDebugLoc(), SCC.getValueType(),
8798 SCC.getOperand(2), SCC.getOperand(3), SETCC);
8806 /// SimplifySelectOps - Given a SELECT or a SELECT_CC node, where LHS and RHS
8807 /// are the two values being selected between, see if we can simplify the
8808 /// select. Callers of this should assume that TheSelect is deleted if this
8809 /// returns true. As such, they should return the appropriate thing (e.g. the
8810 /// node) back to the top-level of the DAG combiner loop to avoid it being
8812 bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDValue LHS,
8815 // Cannot simplify select with vector condition
8816 if (TheSelect->getOperand(0).getValueType().isVector()) return false;
8818 // If this is a select from two identical things, try to pull the operation
8819 // through the select.
8820 if (LHS.getOpcode() != RHS.getOpcode() ||
8821 !LHS.hasOneUse() || !RHS.hasOneUse())
8824 // If this is a load and the token chain is identical, replace the select
8825 // of two loads with a load through a select of the address to load from.
8826 // This triggers in things like "select bool X, 10.0, 123.0" after the FP
8827 // constants have been dropped into the constant pool.
8828 if (LHS.getOpcode() == ISD::LOAD) {
8829 LoadSDNode *LLD = cast<LoadSDNode>(LHS);
8830 LoadSDNode *RLD = cast<LoadSDNode>(RHS);
8832 // Token chains must be identical.
8833 if (LHS.getOperand(0) != RHS.getOperand(0) ||
8834 // Do not let this transformation reduce the number of volatile loads.
8835 LLD->isVolatile() || RLD->isVolatile() ||
8836 // If this is an EXTLOAD, the VT's must match.
8837 LLD->getMemoryVT() != RLD->getMemoryVT() ||
8838 // If this is an EXTLOAD, the kind of extension must match.
8839 (LLD->getExtensionType() != RLD->getExtensionType() &&
8840 // The only exception is if one of the extensions is anyext.
8841 LLD->getExtensionType() != ISD::EXTLOAD &&
8842 RLD->getExtensionType() != ISD::EXTLOAD) ||
8843 // FIXME: this discards src value information. This is
8844 // over-conservative. It would be beneficial to be able to remember
8845 // both potential memory locations. Since we are discarding
8846 // src value info, don't do the transformation if the memory
8847 // locations are not in the default address space.
8848 LLD->getPointerInfo().getAddrSpace() != 0 ||
8849 RLD->getPointerInfo().getAddrSpace() != 0)
8852 // Check that the select condition doesn't reach either load. If so,
8853 // folding this will induce a cycle into the DAG. If not, this is safe to
8854 // xform, so create a select of the addresses.
8856 if (TheSelect->getOpcode() == ISD::SELECT) {
8857 SDNode *CondNode = TheSelect->getOperand(0).getNode();
8858 if ((LLD->hasAnyUseOfValue(1) && LLD->isPredecessorOf(CondNode)) ||
8859 (RLD->hasAnyUseOfValue(1) && RLD->isPredecessorOf(CondNode)))
8861 Addr = DAG.getNode(ISD::SELECT, TheSelect->getDebugLoc(),
8862 LLD->getBasePtr().getValueType(),
8863 TheSelect->getOperand(0), LLD->getBasePtr(),
8865 } else { // Otherwise SELECT_CC
8866 SDNode *CondLHS = TheSelect->getOperand(0).getNode();
8867 SDNode *CondRHS = TheSelect->getOperand(1).getNode();
8869 if ((LLD->hasAnyUseOfValue(1) &&
8870 (LLD->isPredecessorOf(CondLHS) || LLD->isPredecessorOf(CondRHS))) ||
8871 (RLD->hasAnyUseOfValue(1) &&
8872 (RLD->isPredecessorOf(CondLHS) || RLD->isPredecessorOf(CondRHS))))
8875 Addr = DAG.getNode(ISD::SELECT_CC, TheSelect->getDebugLoc(),
8876 LLD->getBasePtr().getValueType(),
8877 TheSelect->getOperand(0),
8878 TheSelect->getOperand(1),
8879 LLD->getBasePtr(), RLD->getBasePtr(),
8880 TheSelect->getOperand(4));
8884 if (LLD->getExtensionType() == ISD::NON_EXTLOAD) {
8885 Load = DAG.getLoad(TheSelect->getValueType(0),
8886 TheSelect->getDebugLoc(),
8887 // FIXME: Discards pointer info.
8888 LLD->getChain(), Addr, MachinePointerInfo(),
8889 LLD->isVolatile(), LLD->isNonTemporal(),
8890 LLD->isInvariant(), LLD->getAlignment());
8892 Load = DAG.getExtLoad(LLD->getExtensionType() == ISD::EXTLOAD ?
8893 RLD->getExtensionType() : LLD->getExtensionType(),
8894 TheSelect->getDebugLoc(),
8895 TheSelect->getValueType(0),
8896 // FIXME: Discards pointer info.
8897 LLD->getChain(), Addr, MachinePointerInfo(),
8898 LLD->getMemoryVT(), LLD->isVolatile(),
8899 LLD->isNonTemporal(), LLD->getAlignment());
8902 // Users of the select now use the result of the load.
8903 CombineTo(TheSelect, Load);
8905 // Users of the old loads now use the new load's chain. We know the
8906 // old-load value is dead now.
8907 CombineTo(LHS.getNode(), Load.getValue(0), Load.getValue(1));
8908 CombineTo(RHS.getNode(), Load.getValue(0), Load.getValue(1));
8915 /// SimplifySelectCC - Simplify an expression of the form (N0 cond N1) ? N2 : N3
8916 /// where 'cond' is the comparison specified by CC.
8917 SDValue DAGCombiner::SimplifySelectCC(DebugLoc DL, SDValue N0, SDValue N1,
8918 SDValue N2, SDValue N3,
8919 ISD::CondCode CC, bool NotExtCompare) {
8920 // (x ? y : y) -> y.
8921 if (N2 == N3) return N2;
8923 EVT VT = N2.getValueType();
8924 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
8925 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
8926 ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N3.getNode());
8928 // Determine if the condition we're dealing with is constant
8929 SDValue SCC = SimplifySetCC(TLI.getSetCCResultType(N0.getValueType()),
8930 N0, N1, CC, DL, false);
8931 if (SCC.getNode()) AddToWorkList(SCC.getNode());
8932 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode());
8934 // fold select_cc true, x, y -> x
8935 if (SCCC && !SCCC->isNullValue())
8937 // fold select_cc false, x, y -> y
8938 if (SCCC && SCCC->isNullValue())
8941 // Check to see if we can simplify the select into an fabs node
8942 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) {
8943 // Allow either -0.0 or 0.0
8944 if (CFP->getValueAPF().isZero()) {
8945 // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs
8946 if ((CC == ISD::SETGE || CC == ISD::SETGT) &&
8947 N0 == N2 && N3.getOpcode() == ISD::FNEG &&
8948 N2 == N3.getOperand(0))
8949 return DAG.getNode(ISD::FABS, DL, VT, N0);
8951 // select (setl[te] X, +/-0.0), fneg(X), X -> fabs
8952 if ((CC == ISD::SETLT || CC == ISD::SETLE) &&
8953 N0 == N3 && N2.getOpcode() == ISD::FNEG &&
8954 N2.getOperand(0) == N3)
8955 return DAG.getNode(ISD::FABS, DL, VT, N3);
8959 // Turn "(a cond b) ? 1.0f : 2.0f" into "load (tmp + ((a cond b) ? 0 : 4)"
8960 // where "tmp" is a constant pool entry containing an array with 1.0 and 2.0
8961 // in it. This is a win when the constant is not otherwise available because
8962 // it replaces two constant pool loads with one. We only do this if the FP
8963 // type is known to be legal, because if it isn't, then we are before legalize
8964 // types an we want the other legalization to happen first (e.g. to avoid
8965 // messing with soft float) and if the ConstantFP is not legal, because if
8966 // it is legal, we may not need to store the FP constant in a constant pool.
8967 if (ConstantFPSDNode *TV = dyn_cast<ConstantFPSDNode>(N2))
8968 if (ConstantFPSDNode *FV = dyn_cast<ConstantFPSDNode>(N3)) {
8969 if (TLI.isTypeLegal(N2.getValueType()) &&
8970 (TLI.getOperationAction(ISD::ConstantFP, N2.getValueType()) !=
8971 TargetLowering::Legal) &&
8972 // If both constants have multiple uses, then we won't need to do an
8973 // extra load, they are likely around in registers for other users.
8974 (TV->hasOneUse() || FV->hasOneUse())) {
8975 Constant *Elts[] = {
8976 const_cast<ConstantFP*>(FV->getConstantFPValue()),
8977 const_cast<ConstantFP*>(TV->getConstantFPValue())
8979 Type *FPTy = Elts[0]->getType();
8980 const TargetData &TD = *TLI.getTargetData();
8982 // Create a ConstantArray of the two constants.
8983 Constant *CA = ConstantArray::get(ArrayType::get(FPTy, 2), Elts);
8984 SDValue CPIdx = DAG.getConstantPool(CA, TLI.getPointerTy(),
8985 TD.getPrefTypeAlignment(FPTy));
8986 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
8988 // Get the offsets to the 0 and 1 element of the array so that we can
8989 // select between them.
8990 SDValue Zero = DAG.getIntPtrConstant(0);
8991 unsigned EltSize = (unsigned)TD.getTypeAllocSize(Elts[0]->getType());
8992 SDValue One = DAG.getIntPtrConstant(EltSize);
8994 SDValue Cond = DAG.getSetCC(DL,
8995 TLI.getSetCCResultType(N0.getValueType()),
8997 AddToWorkList(Cond.getNode());
8998 SDValue CstOffset = DAG.getNode(ISD::SELECT, DL, Zero.getValueType(),
9000 AddToWorkList(CstOffset.getNode());
9001 CPIdx = DAG.getNode(ISD::ADD, DL, TLI.getPointerTy(), CPIdx,
9003 AddToWorkList(CPIdx.getNode());
9004 return DAG.getLoad(TV->getValueType(0), DL, DAG.getEntryNode(), CPIdx,
9005 MachinePointerInfo::getConstantPool(), false,
9006 false, false, Alignment);
9011 // Check to see if we can perform the "gzip trick", transforming
9012 // (select_cc setlt X, 0, A, 0) -> (and (sra X, (sub size(X), 1), A)
9013 if (N1C && N3C && N3C->isNullValue() && CC == ISD::SETLT &&
9014 (N1C->isNullValue() || // (a < 0) ? b : 0
9015 (N1C->getAPIntValue() == 1 && N0 == N2))) { // (a < 1) ? a : 0
9016 EVT XType = N0.getValueType();
9017 EVT AType = N2.getValueType();
9018 if (XType.bitsGE(AType)) {
9019 // and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a
9020 // single-bit constant.
9021 if (N2C && ((N2C->getAPIntValue() & (N2C->getAPIntValue()-1)) == 0)) {
9022 unsigned ShCtV = N2C->getAPIntValue().logBase2();
9023 ShCtV = XType.getSizeInBits()-ShCtV-1;
9024 SDValue ShCt = DAG.getConstant(ShCtV,
9025 getShiftAmountTy(N0.getValueType()));
9026 SDValue Shift = DAG.getNode(ISD::SRL, N0.getDebugLoc(),
9028 AddToWorkList(Shift.getNode());
9030 if (XType.bitsGT(AType)) {
9031 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
9032 AddToWorkList(Shift.getNode());
9035 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
9038 SDValue Shift = DAG.getNode(ISD::SRA, N0.getDebugLoc(),
9040 DAG.getConstant(XType.getSizeInBits()-1,
9041 getShiftAmountTy(N0.getValueType())));
9042 AddToWorkList(Shift.getNode());
9044 if (XType.bitsGT(AType)) {
9045 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
9046 AddToWorkList(Shift.getNode());
9049 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
9053 // fold (select_cc seteq (and x, y), 0, 0, A) -> (and (shr (shl x)) A)
9054 // where y is has a single bit set.
9055 // A plaintext description would be, we can turn the SELECT_CC into an AND
9056 // when the condition can be materialized as an all-ones register. Any
9057 // single bit-test can be materialized as an all-ones register with
9058 // shift-left and shift-right-arith.
9059 if (CC == ISD::SETEQ && N0->getOpcode() == ISD::AND &&
9060 N0->getValueType(0) == VT &&
9061 N1C && N1C->isNullValue() &&
9062 N2C && N2C->isNullValue()) {
9063 SDValue AndLHS = N0->getOperand(0);
9064 ConstantSDNode *ConstAndRHS = dyn_cast<ConstantSDNode>(N0->getOperand(1));
9065 if (ConstAndRHS && ConstAndRHS->getAPIntValue().countPopulation() == 1) {
9066 // Shift the tested bit over the sign bit.
9067 APInt AndMask = ConstAndRHS->getAPIntValue();
9069 DAG.getConstant(AndMask.countLeadingZeros(),
9070 getShiftAmountTy(AndLHS.getValueType()));
9071 SDValue Shl = DAG.getNode(ISD::SHL, N0.getDebugLoc(), VT, AndLHS, ShlAmt);
9073 // Now arithmetic right shift it all the way over, so the result is either
9074 // all-ones, or zero.
9076 DAG.getConstant(AndMask.getBitWidth()-1,
9077 getShiftAmountTy(Shl.getValueType()));
9078 SDValue Shr = DAG.getNode(ISD::SRA, N0.getDebugLoc(), VT, Shl, ShrAmt);
9080 return DAG.getNode(ISD::AND, DL, VT, Shr, N3);
9084 // fold select C, 16, 0 -> shl C, 4
9085 if (N2C && N3C && N3C->isNullValue() && N2C->getAPIntValue().isPowerOf2() &&
9086 TLI.getBooleanContents(N0.getValueType().isVector()) ==
9087 TargetLowering::ZeroOrOneBooleanContent) {
9089 // If the caller doesn't want us to simplify this into a zext of a compare,
9091 if (NotExtCompare && N2C->getAPIntValue() == 1)
9094 // Get a SetCC of the condition
9095 // FIXME: Should probably make sure that setcc is legal if we ever have a
9096 // target where it isn't.
9098 // cast from setcc result type to select result type
9100 SCC = DAG.getSetCC(DL, TLI.getSetCCResultType(N0.getValueType()),
9102 if (N2.getValueType().bitsLT(SCC.getValueType()))
9103 Temp = DAG.getZeroExtendInReg(SCC, N2.getDebugLoc(), N2.getValueType());
9105 Temp = DAG.getNode(ISD::ZERO_EXTEND, N2.getDebugLoc(),
9106 N2.getValueType(), SCC);
9108 SCC = DAG.getSetCC(N0.getDebugLoc(), MVT::i1, N0, N1, CC);
9109 Temp = DAG.getNode(ISD::ZERO_EXTEND, N2.getDebugLoc(),
9110 N2.getValueType(), SCC);
9113 AddToWorkList(SCC.getNode());
9114 AddToWorkList(Temp.getNode());
9116 if (N2C->getAPIntValue() == 1)
9119 // shl setcc result by log2 n2c
9120 return DAG.getNode(ISD::SHL, DL, N2.getValueType(), Temp,
9121 DAG.getConstant(N2C->getAPIntValue().logBase2(),
9122 getShiftAmountTy(Temp.getValueType())));
9125 // Check to see if this is the equivalent of setcc
9126 // FIXME: Turn all of these into setcc if setcc if setcc is legal
9127 // otherwise, go ahead with the folds.
9128 if (0 && N3C && N3C->isNullValue() && N2C && (N2C->getAPIntValue() == 1ULL)) {
9129 EVT XType = N0.getValueType();
9130 if (!LegalOperations ||
9131 TLI.isOperationLegal(ISD::SETCC, TLI.getSetCCResultType(XType))) {
9132 SDValue Res = DAG.getSetCC(DL, TLI.getSetCCResultType(XType), N0, N1, CC);
9133 if (Res.getValueType() != VT)
9134 Res = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Res);
9138 // fold (seteq X, 0) -> (srl (ctlz X, log2(size(X))))
9139 if (N1C && N1C->isNullValue() && CC == ISD::SETEQ &&
9140 (!LegalOperations ||
9141 TLI.isOperationLegal(ISD::CTLZ, XType))) {
9142 SDValue Ctlz = DAG.getNode(ISD::CTLZ, N0.getDebugLoc(), XType, N0);
9143 return DAG.getNode(ISD::SRL, DL, XType, Ctlz,
9144 DAG.getConstant(Log2_32(XType.getSizeInBits()),
9145 getShiftAmountTy(Ctlz.getValueType())));
9147 // fold (setgt X, 0) -> (srl (and (-X, ~X), size(X)-1))
9148 if (N1C && N1C->isNullValue() && CC == ISD::SETGT) {
9149 SDValue NegN0 = DAG.getNode(ISD::SUB, N0.getDebugLoc(),
9150 XType, DAG.getConstant(0, XType), N0);
9151 SDValue NotN0 = DAG.getNOT(N0.getDebugLoc(), N0, XType);
9152 return DAG.getNode(ISD::SRL, DL, XType,
9153 DAG.getNode(ISD::AND, DL, XType, NegN0, NotN0),
9154 DAG.getConstant(XType.getSizeInBits()-1,
9155 getShiftAmountTy(XType)));
9157 // fold (setgt X, -1) -> (xor (srl (X, size(X)-1), 1))
9158 if (N1C && N1C->isAllOnesValue() && CC == ISD::SETGT) {
9159 SDValue Sign = DAG.getNode(ISD::SRL, N0.getDebugLoc(), XType, N0,
9160 DAG.getConstant(XType.getSizeInBits()-1,
9161 getShiftAmountTy(N0.getValueType())));
9162 return DAG.getNode(ISD::XOR, DL, XType, Sign, DAG.getConstant(1, XType));
9166 // Check to see if this is an integer abs.
9167 // select_cc setg[te] X, 0, X, -X ->
9168 // select_cc setgt X, -1, X, -X ->
9169 // select_cc setl[te] X, 0, -X, X ->
9170 // select_cc setlt X, 1, -X, X ->
9171 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
9173 ConstantSDNode *SubC = NULL;
9174 if (((N1C->isNullValue() && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
9175 (N1C->isAllOnesValue() && CC == ISD::SETGT)) &&
9176 N0 == N2 && N3.getOpcode() == ISD::SUB && N0 == N3.getOperand(1))
9177 SubC = dyn_cast<ConstantSDNode>(N3.getOperand(0));
9178 else if (((N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE)) ||
9179 (N1C->isOne() && CC == ISD::SETLT)) &&
9180 N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1))
9181 SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0));
9183 EVT XType = N0.getValueType();
9184 if (SubC && SubC->isNullValue() && XType.isInteger()) {
9185 SDValue Shift = DAG.getNode(ISD::SRA, N0.getDebugLoc(), XType,
9187 DAG.getConstant(XType.getSizeInBits()-1,
9188 getShiftAmountTy(N0.getValueType())));
9189 SDValue Add = DAG.getNode(ISD::ADD, N0.getDebugLoc(),
9191 AddToWorkList(Shift.getNode());
9192 AddToWorkList(Add.getNode());
9193 return DAG.getNode(ISD::XOR, DL, XType, Add, Shift);
9200 /// SimplifySetCC - This is a stub for TargetLowering::SimplifySetCC.
9201 SDValue DAGCombiner::SimplifySetCC(EVT VT, SDValue N0,
9202 SDValue N1, ISD::CondCode Cond,
9203 DebugLoc DL, bool foldBooleans) {
9204 TargetLowering::DAGCombinerInfo
9205 DagCombineInfo(DAG, !LegalTypes, !LegalOperations, false, this);
9206 return TLI.SimplifySetCC(VT, N0, N1, Cond, foldBooleans, DagCombineInfo, DL);
9209 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
9210 /// return a DAG expression to select that will generate the same value by
9211 /// multiplying by a magic number. See:
9212 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
9213 SDValue DAGCombiner::BuildSDIV(SDNode *N) {
9214 std::vector<SDNode*> Built;
9215 SDValue S = TLI.BuildSDIV(N, DAG, LegalOperations, &Built);
9217 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end();
9223 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
9224 /// return a DAG expression to select that will generate the same value by
9225 /// multiplying by a magic number. See:
9226 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
9227 SDValue DAGCombiner::BuildUDIV(SDNode *N) {
9228 std::vector<SDNode*> Built;
9229 SDValue S = TLI.BuildUDIV(N, DAG, LegalOperations, &Built);
9231 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end();
9237 /// FindBaseOffset - Return true if base is a frame index, which is known not
9238 // to alias with anything but itself. Provides base object and offset as
9240 static bool FindBaseOffset(SDValue Ptr, SDValue &Base, int64_t &Offset,
9241 const GlobalValue *&GV, const void *&CV) {
9242 // Assume it is a primitive operation.
9243 Base = Ptr; Offset = 0; GV = 0; CV = 0;
9245 // If it's an adding a simple constant then integrate the offset.
9246 if (Base.getOpcode() == ISD::ADD) {
9247 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Base.getOperand(1))) {
9248 Base = Base.getOperand(0);
9249 Offset += C->getZExtValue();
9253 // Return the underlying GlobalValue, and update the Offset. Return false
9254 // for GlobalAddressSDNode since the same GlobalAddress may be represented
9255 // by multiple nodes with different offsets.
9256 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Base)) {
9257 GV = G->getGlobal();
9258 Offset += G->getOffset();
9262 // Return the underlying Constant value, and update the Offset. Return false
9263 // for ConstantSDNodes since the same constant pool entry may be represented
9264 // by multiple nodes with different offsets.
9265 if (ConstantPoolSDNode *C = dyn_cast<ConstantPoolSDNode>(Base)) {
9266 CV = C->isMachineConstantPoolEntry() ? (const void *)C->getMachineCPVal()
9267 : (const void *)C->getConstVal();
9268 Offset += C->getOffset();
9271 // If it's any of the following then it can't alias with anything but itself.
9272 return isa<FrameIndexSDNode>(Base);
9275 /// isAlias - Return true if there is any possibility that the two addresses
9277 bool DAGCombiner::isAlias(SDValue Ptr1, int64_t Size1,
9278 const Value *SrcValue1, int SrcValueOffset1,
9279 unsigned SrcValueAlign1,
9280 const MDNode *TBAAInfo1,
9281 SDValue Ptr2, int64_t Size2,
9282 const Value *SrcValue2, int SrcValueOffset2,
9283 unsigned SrcValueAlign2,
9284 const MDNode *TBAAInfo2) const {
9285 // If they are the same then they must be aliases.
9286 if (Ptr1 == Ptr2) return true;
9288 // Gather base node and offset information.
9289 SDValue Base1, Base2;
9290 int64_t Offset1, Offset2;
9291 const GlobalValue *GV1, *GV2;
9292 const void *CV1, *CV2;
9293 bool isFrameIndex1 = FindBaseOffset(Ptr1, Base1, Offset1, GV1, CV1);
9294 bool isFrameIndex2 = FindBaseOffset(Ptr2, Base2, Offset2, GV2, CV2);
9296 // If they have a same base address then check to see if they overlap.
9297 if (Base1 == Base2 || (GV1 && (GV1 == GV2)) || (CV1 && (CV1 == CV2)))
9298 return !((Offset1 + Size1) <= Offset2 || (Offset2 + Size2) <= Offset1);
9300 // It is possible for different frame indices to alias each other, mostly
9301 // when tail call optimization reuses return address slots for arguments.
9302 // To catch this case, look up the actual index of frame indices to compute
9303 // the real alias relationship.
9304 if (isFrameIndex1 && isFrameIndex2) {
9305 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
9306 Offset1 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base1)->getIndex());
9307 Offset2 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base2)->getIndex());
9308 return !((Offset1 + Size1) <= Offset2 || (Offset2 + Size2) <= Offset1);
9311 // Otherwise, if we know what the bases are, and they aren't identical, then
9312 // we know they cannot alias.
9313 if ((isFrameIndex1 || CV1 || GV1) && (isFrameIndex2 || CV2 || GV2))
9316 // If we know required SrcValue1 and SrcValue2 have relatively large alignment
9317 // compared to the size and offset of the access, we may be able to prove they
9318 // do not alias. This check is conservative for now to catch cases created by
9319 // splitting vector types.
9320 if ((SrcValueAlign1 == SrcValueAlign2) &&
9321 (SrcValueOffset1 != SrcValueOffset2) &&
9322 (Size1 == Size2) && (SrcValueAlign1 > Size1)) {
9323 int64_t OffAlign1 = SrcValueOffset1 % SrcValueAlign1;
9324 int64_t OffAlign2 = SrcValueOffset2 % SrcValueAlign1;
9326 // There is no overlap between these relatively aligned accesses of similar
9327 // size, return no alias.
9328 if ((OffAlign1 + Size1) <= OffAlign2 || (OffAlign2 + Size2) <= OffAlign1)
9332 if (CombinerGlobalAA) {
9333 // Use alias analysis information.
9334 int64_t MinOffset = std::min(SrcValueOffset1, SrcValueOffset2);
9335 int64_t Overlap1 = Size1 + SrcValueOffset1 - MinOffset;
9336 int64_t Overlap2 = Size2 + SrcValueOffset2 - MinOffset;
9337 AliasAnalysis::AliasResult AAResult =
9338 AA.alias(AliasAnalysis::Location(SrcValue1, Overlap1, TBAAInfo1),
9339 AliasAnalysis::Location(SrcValue2, Overlap2, TBAAInfo2));
9340 if (AAResult == AliasAnalysis::NoAlias)
9344 // Otherwise we have to assume they alias.
9348 /// FindAliasInfo - Extracts the relevant alias information from the memory
9349 /// node. Returns true if the operand was a load.
9350 bool DAGCombiner::FindAliasInfo(SDNode *N,
9351 SDValue &Ptr, int64_t &Size,
9352 const Value *&SrcValue,
9353 int &SrcValueOffset,
9354 unsigned &SrcValueAlign,
9355 const MDNode *&TBAAInfo) const {
9356 LSBaseSDNode *LS = cast<LSBaseSDNode>(N);
9358 Ptr = LS->getBasePtr();
9359 Size = LS->getMemoryVT().getSizeInBits() >> 3;
9360 SrcValue = LS->getSrcValue();
9361 SrcValueOffset = LS->getSrcValueOffset();
9362 SrcValueAlign = LS->getOriginalAlignment();
9363 TBAAInfo = LS->getTBAAInfo();
9364 return isa<LoadSDNode>(LS);
9367 /// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes,
9368 /// looking for aliasing nodes and adding them to the Aliases vector.
9369 void DAGCombiner::GatherAllAliases(SDNode *N, SDValue OriginalChain,
9370 SmallVector<SDValue, 8> &Aliases) {
9371 SmallVector<SDValue, 8> Chains; // List of chains to visit.
9372 SmallPtrSet<SDNode *, 16> Visited; // Visited node set.
9374 // Get alias information for node.
9377 const Value *SrcValue;
9379 unsigned SrcValueAlign;
9380 const MDNode *SrcTBAAInfo;
9381 bool IsLoad = FindAliasInfo(N, Ptr, Size, SrcValue, SrcValueOffset,
9382 SrcValueAlign, SrcTBAAInfo);
9385 Chains.push_back(OriginalChain);
9388 // Look at each chain and determine if it is an alias. If so, add it to the
9389 // aliases list. If not, then continue up the chain looking for the next
9391 while (!Chains.empty()) {
9392 SDValue Chain = Chains.back();
9395 // For TokenFactor nodes, look at each operand and only continue up the
9396 // chain until we find two aliases. If we've seen two aliases, assume we'll
9397 // find more and revert to original chain since the xform is unlikely to be
9400 // FIXME: The depth check could be made to return the last non-aliasing
9401 // chain we found before we hit a tokenfactor rather than the original
9403 if (Depth > 6 || Aliases.size() == 2) {
9405 Aliases.push_back(OriginalChain);
9409 // Don't bother if we've been before.
9410 if (!Visited.insert(Chain.getNode()))
9413 switch (Chain.getOpcode()) {
9414 case ISD::EntryToken:
9415 // Entry token is ideal chain operand, but handled in FindBetterChain.
9420 // Get alias information for Chain.
9423 const Value *OpSrcValue;
9424 int OpSrcValueOffset;
9425 unsigned OpSrcValueAlign;
9426 const MDNode *OpSrcTBAAInfo;
9427 bool IsOpLoad = FindAliasInfo(Chain.getNode(), OpPtr, OpSize,
9428 OpSrcValue, OpSrcValueOffset,
9432 // If chain is alias then stop here.
9433 if (!(IsLoad && IsOpLoad) &&
9434 isAlias(Ptr, Size, SrcValue, SrcValueOffset, SrcValueAlign,
9436 OpPtr, OpSize, OpSrcValue, OpSrcValueOffset,
9437 OpSrcValueAlign, OpSrcTBAAInfo)) {
9438 Aliases.push_back(Chain);
9440 // Look further up the chain.
9441 Chains.push_back(Chain.getOperand(0));
9447 case ISD::TokenFactor:
9448 // We have to check each of the operands of the token factor for "small"
9449 // token factors, so we queue them up. Adding the operands to the queue
9450 // (stack) in reverse order maintains the original order and increases the
9451 // likelihood that getNode will find a matching token factor (CSE.)
9452 if (Chain.getNumOperands() > 16) {
9453 Aliases.push_back(Chain);
9456 for (unsigned n = Chain.getNumOperands(); n;)
9457 Chains.push_back(Chain.getOperand(--n));
9462 // For all other instructions we will just have to take what we can get.
9463 Aliases.push_back(Chain);
9469 /// FindBetterChain - Walk up chain skipping non-aliasing memory nodes, looking
9470 /// for a better chain (aliasing node.)
9471 SDValue DAGCombiner::FindBetterChain(SDNode *N, SDValue OldChain) {
9472 SmallVector<SDValue, 8> Aliases; // Ops for replacing token factor.
9474 // Accumulate all the aliases to this node.
9475 GatherAllAliases(N, OldChain, Aliases);
9477 // If no operands then chain to entry token.
9478 if (Aliases.size() == 0)
9479 return DAG.getEntryNode();
9481 // If a single operand then chain to it. We don't need to revisit it.
9482 if (Aliases.size() == 1)
9485 // Construct a custom tailored token factor.
9486 return DAG.getNode(ISD::TokenFactor, N->getDebugLoc(), MVT::Other,
9487 &Aliases[0], Aliases.size());
9490 // SelectionDAG::Combine - This is the entry point for the file.
9492 void SelectionDAG::Combine(CombineLevel Level, AliasAnalysis &AA,
9493 CodeGenOpt::Level OptLevel) {
9494 /// run - This is the main entry point to this class.
9496 DAGCombiner(*this, AA, OptLevel).Run(Level);