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/ADT/SmallPtrSet.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/Analysis/AliasAnalysis.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineFunction.h"
26 #include "llvm/IR/DataLayout.h"
27 #include "llvm/IR/DerivedTypes.h"
28 #include "llvm/IR/Function.h"
29 #include "llvm/IR/LLVMContext.h"
30 #include "llvm/Support/CommandLine.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/ErrorHandling.h"
33 #include "llvm/Support/MathExtras.h"
34 #include "llvm/Support/raw_ostream.h"
35 #include "llvm/Target/TargetLowering.h"
36 #include "llvm/Target/TargetMachine.h"
37 #include "llvm/Target/TargetOptions.h"
38 #include "llvm/Target/TargetRegisterInfo.h"
39 #include "llvm/Target/TargetSubtargetInfo.h"
43 STATISTIC(NodesCombined , "Number of dag nodes combined");
44 STATISTIC(PreIndexedNodes , "Number of pre-indexed nodes created");
45 STATISTIC(PostIndexedNodes, "Number of post-indexed nodes created");
46 STATISTIC(OpsNarrowed , "Number of load/op/store narrowed");
47 STATISTIC(LdStFP2Int , "Number of fp load/store pairs transformed to int");
48 STATISTIC(SlicedLoads, "Number of load sliced");
52 CombinerAA("combiner-alias-analysis", cl::Hidden,
53 cl::desc("Enable DAG combiner alias-analysis heuristics"));
56 CombinerGlobalAA("combiner-global-alias-analysis", cl::Hidden,
57 cl::desc("Enable DAG combiner's use of IR alias analysis"));
59 // FIXME: Enable the use of TBAA. There are two known issues preventing this:
60 // 1. Stack coloring does not update TBAA when merging allocas
61 // 2. CGP inserts ptrtoint/inttoptr pairs when sinking address computations.
62 // Because BasicAA does not handle inttoptr, we'll often miss basic type
63 // punning idioms that we need to catch so we don't miscompile real-world
66 UseTBAA("combiner-use-tbaa", cl::Hidden, cl::init(false),
67 cl::desc("Enable DAG combiner's use of TBAA"));
70 static cl::opt<std::string>
71 CombinerAAOnlyFunc("combiner-aa-only-func", cl::Hidden,
72 cl::desc("Only use DAG-combiner alias analysis in this"
76 /// Hidden option to stress test load slicing, i.e., when this option
77 /// is enabled, load slicing bypasses most of its profitability guards.
79 StressLoadSlicing("combiner-stress-load-slicing", cl::Hidden,
80 cl::desc("Bypass the profitability model of load "
84 //------------------------------ DAGCombiner ---------------------------------//
88 const TargetLowering &TLI;
90 CodeGenOpt::Level OptLevel;
95 // Worklist of all of the nodes that need to be simplified.
97 // This has the semantics that when adding to the worklist,
98 // the item added must be next to be processed. It should
99 // also only appear once. The naive approach to this takes
102 // To reduce the insert/remove time to logarithmic, we use
103 // a set and a vector to maintain our worklist.
105 // The set contains the items on the worklist, but does not
106 // maintain the order they should be visited.
108 // The vector maintains the order nodes should be visited, but may
109 // contain duplicate or removed nodes. When choosing a node to
110 // visit, we pop off the order stack until we find an item that is
111 // also in the contents set. All operations are O(log N).
112 SmallPtrSet<SDNode*, 64> WorkListContents;
113 SmallVector<SDNode*, 64> WorkListOrder;
115 // AA - Used for DAG load/store alias analysis.
118 /// AddUsersToWorkList - When an instruction is simplified, add all users of
119 /// the instruction to the work lists because they might get more simplified
122 void AddUsersToWorkList(SDNode *N) {
123 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
128 /// visit - call the node-specific routine that knows how to fold each
129 /// particular type of node.
130 SDValue visit(SDNode *N);
133 /// AddToWorkList - Add to the work list making sure its instance is at the
134 /// back (next to be processed.)
135 void AddToWorkList(SDNode *N) {
136 WorkListContents.insert(N);
137 WorkListOrder.push_back(N);
140 /// removeFromWorkList - remove all instances of N from the worklist.
142 void removeFromWorkList(SDNode *N) {
143 WorkListContents.erase(N);
146 SDValue CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
149 SDValue CombineTo(SDNode *N, SDValue Res, bool AddTo = true) {
150 return CombineTo(N, &Res, 1, AddTo);
153 SDValue CombineTo(SDNode *N, SDValue Res0, SDValue Res1,
155 SDValue To[] = { Res0, Res1 };
156 return CombineTo(N, To, 2, AddTo);
159 void CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO);
163 /// SimplifyDemandedBits - Check the specified integer node value to see if
164 /// it can be simplified or if things it uses can be simplified by bit
165 /// propagation. If so, return true.
166 bool SimplifyDemandedBits(SDValue Op) {
167 unsigned BitWidth = Op.getValueType().getScalarType().getSizeInBits();
168 APInt Demanded = APInt::getAllOnesValue(BitWidth);
169 return SimplifyDemandedBits(Op, Demanded);
172 bool SimplifyDemandedBits(SDValue Op, const APInt &Demanded);
174 bool CombineToPreIndexedLoadStore(SDNode *N);
175 bool CombineToPostIndexedLoadStore(SDNode *N);
176 bool SliceUpLoad(SDNode *N);
178 void ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad);
179 SDValue PromoteOperand(SDValue Op, EVT PVT, bool &Replace);
180 SDValue SExtPromoteOperand(SDValue Op, EVT PVT);
181 SDValue ZExtPromoteOperand(SDValue Op, EVT PVT);
182 SDValue PromoteIntBinOp(SDValue Op);
183 SDValue PromoteIntShiftOp(SDValue Op);
184 SDValue PromoteExtend(SDValue Op);
185 bool PromoteLoad(SDValue Op);
187 void ExtendSetCCUses(const SmallVectorImpl<SDNode *> &SetCCs,
188 SDValue Trunc, SDValue ExtLoad, SDLoc DL,
189 ISD::NodeType ExtType);
191 /// combine - call the node-specific routine that knows how to fold each
192 /// particular type of node. If that doesn't do anything, try the
193 /// target-specific DAG combines.
194 SDValue combine(SDNode *N);
196 // Visitation implementation - Implement dag node combining for different
197 // node types. The semantics are as follows:
199 // SDValue.getNode() == 0 - No change was made
200 // SDValue.getNode() == N - N was replaced, is dead and has been handled.
201 // otherwise - N should be replaced by the returned Operand.
203 SDValue visitTokenFactor(SDNode *N);
204 SDValue visitMERGE_VALUES(SDNode *N);
205 SDValue visitADD(SDNode *N);
206 SDValue visitSUB(SDNode *N);
207 SDValue visitADDC(SDNode *N);
208 SDValue visitSUBC(SDNode *N);
209 SDValue visitADDE(SDNode *N);
210 SDValue visitSUBE(SDNode *N);
211 SDValue visitMUL(SDNode *N);
212 SDValue visitSDIV(SDNode *N);
213 SDValue visitUDIV(SDNode *N);
214 SDValue visitSREM(SDNode *N);
215 SDValue visitUREM(SDNode *N);
216 SDValue visitMULHU(SDNode *N);
217 SDValue visitMULHS(SDNode *N);
218 SDValue visitSMUL_LOHI(SDNode *N);
219 SDValue visitUMUL_LOHI(SDNode *N);
220 SDValue visitSMULO(SDNode *N);
221 SDValue visitUMULO(SDNode *N);
222 SDValue visitSDIVREM(SDNode *N);
223 SDValue visitUDIVREM(SDNode *N);
224 SDValue visitAND(SDNode *N);
225 SDValue visitOR(SDNode *N);
226 SDValue visitXOR(SDNode *N);
227 SDValue SimplifyVBinOp(SDNode *N);
228 SDValue SimplifyVUnaryOp(SDNode *N);
229 SDValue visitSHL(SDNode *N);
230 SDValue visitSRA(SDNode *N);
231 SDValue visitSRL(SDNode *N);
232 SDValue visitCTLZ(SDNode *N);
233 SDValue visitCTLZ_ZERO_UNDEF(SDNode *N);
234 SDValue visitCTTZ(SDNode *N);
235 SDValue visitCTTZ_ZERO_UNDEF(SDNode *N);
236 SDValue visitCTPOP(SDNode *N);
237 SDValue visitSELECT(SDNode *N);
238 SDValue visitVSELECT(SDNode *N);
239 SDValue visitSELECT_CC(SDNode *N);
240 SDValue visitSETCC(SDNode *N);
241 SDValue visitSIGN_EXTEND(SDNode *N);
242 SDValue visitZERO_EXTEND(SDNode *N);
243 SDValue visitANY_EXTEND(SDNode *N);
244 SDValue visitSIGN_EXTEND_INREG(SDNode *N);
245 SDValue visitTRUNCATE(SDNode *N);
246 SDValue visitBITCAST(SDNode *N);
247 SDValue visitBUILD_PAIR(SDNode *N);
248 SDValue visitFADD(SDNode *N);
249 SDValue visitFSUB(SDNode *N);
250 SDValue visitFMUL(SDNode *N);
251 SDValue visitFMA(SDNode *N);
252 SDValue visitFDIV(SDNode *N);
253 SDValue visitFREM(SDNode *N);
254 SDValue visitFCOPYSIGN(SDNode *N);
255 SDValue visitSINT_TO_FP(SDNode *N);
256 SDValue visitUINT_TO_FP(SDNode *N);
257 SDValue visitFP_TO_SINT(SDNode *N);
258 SDValue visitFP_TO_UINT(SDNode *N);
259 SDValue visitFP_ROUND(SDNode *N);
260 SDValue visitFP_ROUND_INREG(SDNode *N);
261 SDValue visitFP_EXTEND(SDNode *N);
262 SDValue visitFNEG(SDNode *N);
263 SDValue visitFABS(SDNode *N);
264 SDValue visitFCEIL(SDNode *N);
265 SDValue visitFTRUNC(SDNode *N);
266 SDValue visitFFLOOR(SDNode *N);
267 SDValue visitBRCOND(SDNode *N);
268 SDValue visitBR_CC(SDNode *N);
269 SDValue visitLOAD(SDNode *N);
270 SDValue visitSTORE(SDNode *N);
271 SDValue visitINSERT_VECTOR_ELT(SDNode *N);
272 SDValue visitEXTRACT_VECTOR_ELT(SDNode *N);
273 SDValue visitBUILD_VECTOR(SDNode *N);
274 SDValue visitCONCAT_VECTORS(SDNode *N);
275 SDValue visitEXTRACT_SUBVECTOR(SDNode *N);
276 SDValue visitVECTOR_SHUFFLE(SDNode *N);
277 SDValue visitINSERT_SUBVECTOR(SDNode *N);
279 SDValue XformToShuffleWithZero(SDNode *N);
280 SDValue ReassociateOps(unsigned Opc, SDLoc DL, SDValue LHS, SDValue RHS);
282 SDValue visitShiftByConstant(SDNode *N, unsigned Amt);
284 bool SimplifySelectOps(SDNode *SELECT, SDValue LHS, SDValue RHS);
285 SDValue SimplifyBinOpWithSameOpcodeHands(SDNode *N);
286 SDValue SimplifySelect(SDLoc DL, SDValue N0, SDValue N1, SDValue N2);
287 SDValue SimplifySelectCC(SDLoc DL, SDValue N0, SDValue N1, SDValue N2,
288 SDValue N3, ISD::CondCode CC,
289 bool NotExtCompare = false);
290 SDValue SimplifySetCC(EVT VT, SDValue N0, SDValue N1, ISD::CondCode Cond,
291 SDLoc DL, bool foldBooleans = true);
292 SDValue SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
294 SDValue CombineConsecutiveLoads(SDNode *N, EVT VT);
295 SDValue ConstantFoldBITCASTofBUILD_VECTOR(SDNode *, EVT);
296 SDValue BuildSDIV(SDNode *N);
297 SDValue BuildUDIV(SDNode *N);
298 SDValue MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
299 bool DemandHighBits = true);
300 SDValue MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1);
301 SDNode *MatchRotatePosNeg(SDValue Shifted, SDValue Pos, SDValue Neg,
302 SDValue InnerPos, SDValue InnerNeg,
303 unsigned PosOpcode, unsigned NegOpcode,
305 SDNode *MatchRotate(SDValue LHS, SDValue RHS, SDLoc DL);
306 SDValue ReduceLoadWidth(SDNode *N);
307 SDValue ReduceLoadOpStoreWidth(SDNode *N);
308 SDValue TransformFPLoadStorePair(SDNode *N);
309 SDValue reduceBuildVecExtToExtBuildVec(SDNode *N);
310 SDValue reduceBuildVecConvertToConvertBuildVec(SDNode *N);
312 SDValue GetDemandedBits(SDValue V, const APInt &Mask);
314 /// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes,
315 /// looking for aliasing nodes and adding them to the Aliases vector.
316 void GatherAllAliases(SDNode *N, SDValue OriginalChain,
317 SmallVectorImpl<SDValue> &Aliases);
319 /// isAlias - Return true if there is any possibility that the two addresses
321 bool isAlias(SDValue Ptr1, int64_t Size1, bool IsVolatile1,
322 const Value *SrcValue1, int SrcValueOffset1,
323 unsigned SrcValueAlign1,
324 const MDNode *TBAAInfo1,
325 SDValue Ptr2, int64_t Size2, bool IsVolatile2,
326 const Value *SrcValue2, int SrcValueOffset2,
327 unsigned SrcValueAlign2,
328 const MDNode *TBAAInfo2) const;
330 /// isAlias - Return true if there is any possibility that the two addresses
332 bool isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1);
334 /// FindAliasInfo - Extracts the relevant alias information from the memory
335 /// node. Returns true if the operand was a load.
336 bool FindAliasInfo(SDNode *N,
337 SDValue &Ptr, int64_t &Size, bool &IsVolatile,
338 const Value *&SrcValue, int &SrcValueOffset,
339 unsigned &SrcValueAlignment,
340 const MDNode *&TBAAInfo) const;
342 /// FindBetterChain - Walk up chain skipping non-aliasing memory nodes,
343 /// looking for a better chain (aliasing node.)
344 SDValue FindBetterChain(SDNode *N, SDValue Chain);
346 /// Merge consecutive store operations into a wide store.
347 /// This optimization uses wide integers or vectors when possible.
348 /// \return True if some memory operations were changed.
349 bool MergeConsecutiveStores(StoreSDNode *N);
352 DAGCombiner(SelectionDAG &D, AliasAnalysis &A, CodeGenOpt::Level OL)
353 : DAG(D), TLI(D.getTargetLoweringInfo()), Level(BeforeLegalizeTypes),
354 OptLevel(OL), LegalOperations(false), LegalTypes(false), AA(A) {
355 AttributeSet FnAttrs =
356 DAG.getMachineFunction().getFunction()->getAttributes();
358 FnAttrs.hasAttribute(AttributeSet::FunctionIndex,
359 Attribute::OptimizeForSize) ||
360 FnAttrs.hasAttribute(AttributeSet::FunctionIndex, Attribute::MinSize);
363 /// Run - runs the dag combiner on all nodes in the work list
364 void Run(CombineLevel AtLevel);
366 SelectionDAG &getDAG() const { return DAG; }
368 /// getShiftAmountTy - Returns a type large enough to hold any valid
369 /// shift amount - before type legalization these can be huge.
370 EVT getShiftAmountTy(EVT LHSTy) {
371 assert(LHSTy.isInteger() && "Shift amount is not an integer type!");
372 if (LHSTy.isVector())
374 return LegalTypes ? TLI.getScalarShiftAmountTy(LHSTy)
375 : TLI.getPointerTy();
378 /// isTypeLegal - This method returns true if we are running before type
379 /// legalization or if the specified VT is legal.
380 bool isTypeLegal(const EVT &VT) {
381 if (!LegalTypes) return true;
382 return TLI.isTypeLegal(VT);
385 /// getSetCCResultType - Convenience wrapper around
386 /// TargetLowering::getSetCCResultType
387 EVT getSetCCResultType(EVT VT) const {
388 return TLI.getSetCCResultType(*DAG.getContext(), VT);
395 /// WorkListRemover - This class is a DAGUpdateListener that removes any deleted
396 /// nodes from the worklist.
397 class WorkListRemover : public SelectionDAG::DAGUpdateListener {
400 explicit WorkListRemover(DAGCombiner &dc)
401 : SelectionDAG::DAGUpdateListener(dc.getDAG()), DC(dc) {}
403 virtual void NodeDeleted(SDNode *N, SDNode *E) {
404 DC.removeFromWorkList(N);
409 //===----------------------------------------------------------------------===//
410 // TargetLowering::DAGCombinerInfo implementation
411 //===----------------------------------------------------------------------===//
413 void TargetLowering::DAGCombinerInfo::AddToWorklist(SDNode *N) {
414 ((DAGCombiner*)DC)->AddToWorkList(N);
417 void TargetLowering::DAGCombinerInfo::RemoveFromWorklist(SDNode *N) {
418 ((DAGCombiner*)DC)->removeFromWorkList(N);
421 SDValue TargetLowering::DAGCombinerInfo::
422 CombineTo(SDNode *N, const std::vector<SDValue> &To, bool AddTo) {
423 return ((DAGCombiner*)DC)->CombineTo(N, &To[0], To.size(), AddTo);
426 SDValue TargetLowering::DAGCombinerInfo::
427 CombineTo(SDNode *N, SDValue Res, bool AddTo) {
428 return ((DAGCombiner*)DC)->CombineTo(N, Res, AddTo);
432 SDValue TargetLowering::DAGCombinerInfo::
433 CombineTo(SDNode *N, SDValue Res0, SDValue Res1, bool AddTo) {
434 return ((DAGCombiner*)DC)->CombineTo(N, Res0, Res1, AddTo);
437 void TargetLowering::DAGCombinerInfo::
438 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
439 return ((DAGCombiner*)DC)->CommitTargetLoweringOpt(TLO);
442 //===----------------------------------------------------------------------===//
444 //===----------------------------------------------------------------------===//
446 /// isNegatibleForFree - Return 1 if we can compute the negated form of the
447 /// specified expression for the same cost as the expression itself, or 2 if we
448 /// can compute the negated form more cheaply than the expression itself.
449 static char isNegatibleForFree(SDValue Op, bool LegalOperations,
450 const TargetLowering &TLI,
451 const TargetOptions *Options,
452 unsigned Depth = 0) {
453 // fneg is removable even if it has multiple uses.
454 if (Op.getOpcode() == ISD::FNEG) return 2;
456 // Don't allow anything with multiple uses.
457 if (!Op.hasOneUse()) return 0;
459 // Don't recurse exponentially.
460 if (Depth > 6) return 0;
462 switch (Op.getOpcode()) {
463 default: return false;
464 case ISD::ConstantFP:
465 // Don't invert constant FP values after legalize. The negated constant
466 // isn't necessarily legal.
467 return LegalOperations ? 0 : 1;
469 // FIXME: determine better conditions for this xform.
470 if (!Options->UnsafeFPMath) return 0;
472 // After operation legalization, it might not be legal to create new FSUBs.
473 if (LegalOperations &&
474 !TLI.isOperationLegalOrCustom(ISD::FSUB, Op.getValueType()))
477 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B)
478 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI,
481 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
482 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options,
485 // We can't turn -(A-B) into B-A when we honor signed zeros.
486 if (!Options->UnsafeFPMath) return 0;
488 // fold (fneg (fsub A, B)) -> (fsub B, A)
493 if (Options->HonorSignDependentRoundingFPMath()) return 0;
495 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y) or (fmul X, (fneg Y))
496 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI,
500 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options,
506 return isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI, Options,
511 /// GetNegatedExpression - If isNegatibleForFree returns true, this function
512 /// returns the newly negated expression.
513 static SDValue GetNegatedExpression(SDValue Op, SelectionDAG &DAG,
514 bool LegalOperations, unsigned Depth = 0) {
515 // fneg is removable even if it has multiple uses.
516 if (Op.getOpcode() == ISD::FNEG) return Op.getOperand(0);
518 // Don't allow anything with multiple uses.
519 assert(Op.hasOneUse() && "Unknown reuse!");
521 assert(Depth <= 6 && "GetNegatedExpression doesn't match isNegatibleForFree");
522 switch (Op.getOpcode()) {
523 default: llvm_unreachable("Unknown code");
524 case ISD::ConstantFP: {
525 APFloat V = cast<ConstantFPSDNode>(Op)->getValueAPF();
527 return DAG.getConstantFP(V, Op.getValueType());
530 // FIXME: determine better conditions for this xform.
531 assert(DAG.getTarget().Options.UnsafeFPMath);
533 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B)
534 if (isNegatibleForFree(Op.getOperand(0), LegalOperations,
535 DAG.getTargetLoweringInfo(),
536 &DAG.getTarget().Options, Depth+1))
537 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
538 GetNegatedExpression(Op.getOperand(0), DAG,
539 LegalOperations, Depth+1),
541 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
542 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
543 GetNegatedExpression(Op.getOperand(1), DAG,
544 LegalOperations, Depth+1),
547 // We can't turn -(A-B) into B-A when we honor signed zeros.
548 assert(DAG.getTarget().Options.UnsafeFPMath);
550 // fold (fneg (fsub 0, B)) -> B
551 if (ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(Op.getOperand(0)))
552 if (N0CFP->getValueAPF().isZero())
553 return Op.getOperand(1);
555 // fold (fneg (fsub A, B)) -> (fsub B, A)
556 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
557 Op.getOperand(1), Op.getOperand(0));
561 assert(!DAG.getTarget().Options.HonorSignDependentRoundingFPMath());
563 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y)
564 if (isNegatibleForFree(Op.getOperand(0), LegalOperations,
565 DAG.getTargetLoweringInfo(),
566 &DAG.getTarget().Options, Depth+1))
567 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
568 GetNegatedExpression(Op.getOperand(0), DAG,
569 LegalOperations, Depth+1),
572 // fold (fneg (fmul X, Y)) -> (fmul X, (fneg Y))
573 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
575 GetNegatedExpression(Op.getOperand(1), DAG,
576 LegalOperations, Depth+1));
580 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
581 GetNegatedExpression(Op.getOperand(0), DAG,
582 LegalOperations, Depth+1));
584 return DAG.getNode(ISD::FP_ROUND, SDLoc(Op), Op.getValueType(),
585 GetNegatedExpression(Op.getOperand(0), DAG,
586 LegalOperations, Depth+1),
592 // isSetCCEquivalent - Return true if this node is a setcc, or is a select_cc
593 // that selects between the values 1 and 0, making it equivalent to a setcc.
594 // Also, set the incoming LHS, RHS, and CC references to the appropriate
595 // nodes based on the type of node we are checking. This simplifies life a
596 // bit for the callers.
597 static bool isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS,
599 if (N.getOpcode() == ISD::SETCC) {
600 LHS = N.getOperand(0);
601 RHS = N.getOperand(1);
602 CC = N.getOperand(2);
605 if (N.getOpcode() == ISD::SELECT_CC &&
606 N.getOperand(2).getOpcode() == ISD::Constant &&
607 N.getOperand(3).getOpcode() == ISD::Constant &&
608 cast<ConstantSDNode>(N.getOperand(2))->getAPIntValue() == 1 &&
609 cast<ConstantSDNode>(N.getOperand(3))->isNullValue()) {
610 LHS = N.getOperand(0);
611 RHS = N.getOperand(1);
612 CC = N.getOperand(4);
618 // isOneUseSetCC - Return true if this is a SetCC-equivalent operation with only
619 // one use. If this is true, it allows the users to invert the operation for
620 // free when it is profitable to do so.
621 static bool isOneUseSetCC(SDValue N) {
623 if (isSetCCEquivalent(N, N0, N1, N2) && N.getNode()->hasOneUse())
628 // \brief Returns the SDNode if it is a constant BuildVector or constant int.
629 static SDNode *isConstantBuildVectorOrConstantInt(SDValue N) {
630 if (isa<ConstantSDNode>(N))
632 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N);
633 if(BV && BV->isConstant())
638 SDValue DAGCombiner::ReassociateOps(unsigned Opc, SDLoc DL,
639 SDValue N0, SDValue N1) {
640 EVT VT = N0.getValueType();
641 if (N0.getOpcode() == Opc) {
642 if (SDNode *L = isConstantBuildVectorOrConstantInt(N0.getOperand(1))) {
643 if (SDNode *R = isConstantBuildVectorOrConstantInt(N1)) {
644 // reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2))
645 SDValue OpNode = DAG.FoldConstantArithmetic(Opc, VT, L, R);
646 if (!OpNode.getNode())
648 return DAG.getNode(Opc, DL, VT, N0.getOperand(0), OpNode);
650 if (N0.hasOneUse()) {
651 // reassoc. (op (op x, c1), y) -> (op (op x, y), c1) iff x+c1 has one
653 SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N0.getOperand(0), N1);
654 if (!OpNode.getNode())
656 AddToWorkList(OpNode.getNode());
657 return DAG.getNode(Opc, DL, VT, OpNode, N0.getOperand(1));
662 if (N1.getOpcode() == Opc) {
663 if (SDNode *R = isConstantBuildVectorOrConstantInt(N1.getOperand(1))) {
664 if (SDNode *L = isConstantBuildVectorOrConstantInt(N0)) {
665 // reassoc. (op c2, (op x, c1)) -> (op x, (op c1, c2))
666 SDValue OpNode = DAG.FoldConstantArithmetic(Opc, VT, R, L);
667 if (!OpNode.getNode())
669 return DAG.getNode(Opc, DL, VT, N1.getOperand(0), OpNode);
671 if (N1.hasOneUse()) {
672 // reassoc. (op y, (op x, c1)) -> (op (op x, y), c1) iff x+c1 has one
674 SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N1.getOperand(0), N0);
675 if (!OpNode.getNode())
677 AddToWorkList(OpNode.getNode());
678 return DAG.getNode(Opc, DL, VT, OpNode, N1.getOperand(1));
686 SDValue DAGCombiner::CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
688 assert(N->getNumValues() == NumTo && "Broken CombineTo call!");
690 DEBUG(dbgs() << "\nReplacing.1 ";
692 dbgs() << "\nWith: ";
693 To[0].getNode()->dump(&DAG);
694 dbgs() << " and " << NumTo-1 << " other values\n";
695 for (unsigned i = 0, e = NumTo; i != e; ++i)
696 assert((!To[i].getNode() ||
697 N->getValueType(i) == To[i].getValueType()) &&
698 "Cannot combine value to value of different type!"));
699 WorkListRemover DeadNodes(*this);
700 DAG.ReplaceAllUsesWith(N, To);
702 // Push the new nodes and any users onto the worklist
703 for (unsigned i = 0, e = NumTo; i != e; ++i) {
704 if (To[i].getNode()) {
705 AddToWorkList(To[i].getNode());
706 AddUsersToWorkList(To[i].getNode());
711 // Finally, if the node is now dead, remove it from the graph. The node
712 // may not be dead if the replacement process recursively simplified to
713 // something else needing this node.
714 if (N->use_empty()) {
715 // Nodes can be reintroduced into the worklist. Make sure we do not
716 // process a node that has been replaced.
717 removeFromWorkList(N);
719 // Finally, since the node is now dead, remove it from the graph.
722 return SDValue(N, 0);
726 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
727 // Replace all uses. If any nodes become isomorphic to other nodes and
728 // are deleted, make sure to remove them from our worklist.
729 WorkListRemover DeadNodes(*this);
730 DAG.ReplaceAllUsesOfValueWith(TLO.Old, TLO.New);
732 // Push the new node and any (possibly new) users onto the worklist.
733 AddToWorkList(TLO.New.getNode());
734 AddUsersToWorkList(TLO.New.getNode());
736 // Finally, if the node is now dead, remove it from the graph. The node
737 // may not be dead if the replacement process recursively simplified to
738 // something else needing this node.
739 if (TLO.Old.getNode()->use_empty()) {
740 removeFromWorkList(TLO.Old.getNode());
742 // If the operands of this node are only used by the node, they will now
743 // be dead. Make sure to visit them first to delete dead nodes early.
744 for (unsigned i = 0, e = TLO.Old.getNode()->getNumOperands(); i != e; ++i)
745 if (TLO.Old.getNode()->getOperand(i).getNode()->hasOneUse())
746 AddToWorkList(TLO.Old.getNode()->getOperand(i).getNode());
748 DAG.DeleteNode(TLO.Old.getNode());
752 /// SimplifyDemandedBits - Check the specified integer node value to see if
753 /// it can be simplified or if things it uses can be simplified by bit
754 /// propagation. If so, return true.
755 bool DAGCombiner::SimplifyDemandedBits(SDValue Op, const APInt &Demanded) {
756 TargetLowering::TargetLoweringOpt TLO(DAG, LegalTypes, LegalOperations);
757 APInt KnownZero, KnownOne;
758 if (!TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
762 AddToWorkList(Op.getNode());
764 // Replace the old value with the new one.
766 DEBUG(dbgs() << "\nReplacing.2 ";
767 TLO.Old.getNode()->dump(&DAG);
768 dbgs() << "\nWith: ";
769 TLO.New.getNode()->dump(&DAG);
772 CommitTargetLoweringOpt(TLO);
776 void DAGCombiner::ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad) {
778 EVT VT = Load->getValueType(0);
779 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, VT, SDValue(ExtLoad, 0));
781 DEBUG(dbgs() << "\nReplacing.9 ";
783 dbgs() << "\nWith: ";
784 Trunc.getNode()->dump(&DAG);
786 WorkListRemover DeadNodes(*this);
787 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 0), Trunc);
788 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), SDValue(ExtLoad, 1));
789 removeFromWorkList(Load);
790 DAG.DeleteNode(Load);
791 AddToWorkList(Trunc.getNode());
794 SDValue DAGCombiner::PromoteOperand(SDValue Op, EVT PVT, bool &Replace) {
797 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Op)) {
798 EVT MemVT = LD->getMemoryVT();
799 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
800 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT) ? ISD::ZEXTLOAD
802 : LD->getExtensionType();
804 return DAG.getExtLoad(ExtType, dl, PVT,
805 LD->getChain(), LD->getBasePtr(),
806 MemVT, LD->getMemOperand());
809 unsigned Opc = Op.getOpcode();
812 case ISD::AssertSext:
813 return DAG.getNode(ISD::AssertSext, dl, PVT,
814 SExtPromoteOperand(Op.getOperand(0), PVT),
816 case ISD::AssertZext:
817 return DAG.getNode(ISD::AssertZext, dl, PVT,
818 ZExtPromoteOperand(Op.getOperand(0), PVT),
820 case ISD::Constant: {
822 Op.getValueType().isByteSized() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
823 return DAG.getNode(ExtOpc, dl, PVT, Op);
827 if (!TLI.isOperationLegal(ISD::ANY_EXTEND, PVT))
829 return DAG.getNode(ISD::ANY_EXTEND, dl, PVT, Op);
832 SDValue DAGCombiner::SExtPromoteOperand(SDValue Op, EVT PVT) {
833 if (!TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, PVT))
835 EVT OldVT = Op.getValueType();
837 bool Replace = false;
838 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
839 if (NewOp.getNode() == 0)
841 AddToWorkList(NewOp.getNode());
844 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
845 return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, NewOp.getValueType(), NewOp,
846 DAG.getValueType(OldVT));
849 SDValue DAGCombiner::ZExtPromoteOperand(SDValue Op, EVT PVT) {
850 EVT OldVT = Op.getValueType();
852 bool Replace = false;
853 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
854 if (NewOp.getNode() == 0)
856 AddToWorkList(NewOp.getNode());
859 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
860 return DAG.getZeroExtendInReg(NewOp, dl, OldVT);
863 /// PromoteIntBinOp - Promote the specified integer binary operation if the
864 /// target indicates it is beneficial. e.g. On x86, it's usually better to
865 /// promote i16 operations to i32 since i16 instructions are longer.
866 SDValue DAGCombiner::PromoteIntBinOp(SDValue Op) {
867 if (!LegalOperations)
870 EVT VT = Op.getValueType();
871 if (VT.isVector() || !VT.isInteger())
874 // If operation type is 'undesirable', e.g. i16 on x86, consider
876 unsigned Opc = Op.getOpcode();
877 if (TLI.isTypeDesirableForOp(Opc, VT))
881 // Consult target whether it is a good idea to promote this operation and
882 // what's the right type to promote it to.
883 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
884 assert(PVT != VT && "Don't know what type to promote to!");
886 bool Replace0 = false;
887 SDValue N0 = Op.getOperand(0);
888 SDValue NN0 = PromoteOperand(N0, PVT, Replace0);
889 if (NN0.getNode() == 0)
892 bool Replace1 = false;
893 SDValue N1 = Op.getOperand(1);
898 NN1 = PromoteOperand(N1, PVT, Replace1);
899 if (NN1.getNode() == 0)
903 AddToWorkList(NN0.getNode());
905 AddToWorkList(NN1.getNode());
908 ReplaceLoadWithPromotedLoad(N0.getNode(), NN0.getNode());
910 ReplaceLoadWithPromotedLoad(N1.getNode(), NN1.getNode());
912 DEBUG(dbgs() << "\nPromoting ";
913 Op.getNode()->dump(&DAG));
915 return DAG.getNode(ISD::TRUNCATE, dl, VT,
916 DAG.getNode(Opc, dl, PVT, NN0, NN1));
921 /// PromoteIntShiftOp - Promote the specified integer shift operation if the
922 /// target indicates it is beneficial. e.g. On x86, it's usually better to
923 /// promote i16 operations to i32 since i16 instructions are longer.
924 SDValue DAGCombiner::PromoteIntShiftOp(SDValue Op) {
925 if (!LegalOperations)
928 EVT VT = Op.getValueType();
929 if (VT.isVector() || !VT.isInteger())
932 // If operation type is 'undesirable', e.g. i16 on x86, consider
934 unsigned Opc = Op.getOpcode();
935 if (TLI.isTypeDesirableForOp(Opc, VT))
939 // Consult target whether it is a good idea to promote this operation and
940 // what's the right type to promote it to.
941 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
942 assert(PVT != VT && "Don't know what type to promote to!");
944 bool Replace = false;
945 SDValue N0 = Op.getOperand(0);
947 N0 = SExtPromoteOperand(Op.getOperand(0), PVT);
948 else if (Opc == ISD::SRL)
949 N0 = ZExtPromoteOperand(Op.getOperand(0), PVT);
951 N0 = PromoteOperand(N0, PVT, Replace);
952 if (N0.getNode() == 0)
955 AddToWorkList(N0.getNode());
957 ReplaceLoadWithPromotedLoad(Op.getOperand(0).getNode(), N0.getNode());
959 DEBUG(dbgs() << "\nPromoting ";
960 Op.getNode()->dump(&DAG));
962 return DAG.getNode(ISD::TRUNCATE, dl, VT,
963 DAG.getNode(Opc, dl, PVT, N0, Op.getOperand(1)));
968 SDValue DAGCombiner::PromoteExtend(SDValue Op) {
969 if (!LegalOperations)
972 EVT VT = Op.getValueType();
973 if (VT.isVector() || !VT.isInteger())
976 // If operation type is 'undesirable', e.g. i16 on x86, consider
978 unsigned Opc = Op.getOpcode();
979 if (TLI.isTypeDesirableForOp(Opc, VT))
983 // Consult target whether it is a good idea to promote this operation and
984 // what's the right type to promote it to.
985 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
986 assert(PVT != VT && "Don't know what type to promote to!");
987 // fold (aext (aext x)) -> (aext x)
988 // fold (aext (zext x)) -> (zext x)
989 // fold (aext (sext x)) -> (sext x)
990 DEBUG(dbgs() << "\nPromoting ";
991 Op.getNode()->dump(&DAG));
992 return DAG.getNode(Op.getOpcode(), SDLoc(Op), VT, Op.getOperand(0));
997 bool DAGCombiner::PromoteLoad(SDValue Op) {
998 if (!LegalOperations)
1001 EVT VT = Op.getValueType();
1002 if (VT.isVector() || !VT.isInteger())
1005 // If operation type is 'undesirable', e.g. i16 on x86, consider
1007 unsigned Opc = Op.getOpcode();
1008 if (TLI.isTypeDesirableForOp(Opc, VT))
1012 // Consult target whether it is a good idea to promote this operation and
1013 // what's the right type to promote it to.
1014 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1015 assert(PVT != VT && "Don't know what type to promote to!");
1018 SDNode *N = Op.getNode();
1019 LoadSDNode *LD = cast<LoadSDNode>(N);
1020 EVT MemVT = LD->getMemoryVT();
1021 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
1022 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT) ? ISD::ZEXTLOAD
1024 : LD->getExtensionType();
1025 SDValue NewLD = DAG.getExtLoad(ExtType, dl, PVT,
1026 LD->getChain(), LD->getBasePtr(),
1027 MemVT, LD->getMemOperand());
1028 SDValue Result = DAG.getNode(ISD::TRUNCATE, dl, VT, NewLD);
1030 DEBUG(dbgs() << "\nPromoting ";
1033 Result.getNode()->dump(&DAG);
1035 WorkListRemover DeadNodes(*this);
1036 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result);
1037 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), NewLD.getValue(1));
1038 removeFromWorkList(N);
1040 AddToWorkList(Result.getNode());
1047 //===----------------------------------------------------------------------===//
1048 // Main DAG Combiner implementation
1049 //===----------------------------------------------------------------------===//
1051 void DAGCombiner::Run(CombineLevel AtLevel) {
1052 // set the instance variables, so that the various visit routines may use it.
1054 LegalOperations = Level >= AfterLegalizeVectorOps;
1055 LegalTypes = Level >= AfterLegalizeTypes;
1057 // Add all the dag nodes to the worklist.
1058 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
1059 E = DAG.allnodes_end(); I != E; ++I)
1062 // Create a dummy node (which is not added to allnodes), that adds a reference
1063 // to the root node, preventing it from being deleted, and tracking any
1064 // changes of the root.
1065 HandleSDNode Dummy(DAG.getRoot());
1067 // The root of the dag may dangle to deleted nodes until the dag combiner is
1068 // done. Set it to null to avoid confusion.
1069 DAG.setRoot(SDValue());
1071 // while the worklist isn't empty, find a node and
1072 // try and combine it.
1073 while (!WorkListContents.empty()) {
1075 // The WorkListOrder holds the SDNodes in order, but it may contain
1077 // In order to avoid a linear scan, we use a set (O(log N)) to hold what the
1078 // worklist *should* contain, and check the node we want to visit is should
1079 // actually be visited.
1081 N = WorkListOrder.pop_back_val();
1082 } while (!WorkListContents.erase(N));
1084 // If N has no uses, it is dead. Make sure to revisit all N's operands once
1085 // N is deleted from the DAG, since they too may now be dead or may have a
1086 // reduced number of uses, allowing other xforms.
1087 if (N->use_empty() && N != &Dummy) {
1088 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1089 AddToWorkList(N->getOperand(i).getNode());
1095 SDValue RV = combine(N);
1097 if (RV.getNode() == 0)
1102 // If we get back the same node we passed in, rather than a new node or
1103 // zero, we know that the node must have defined multiple values and
1104 // CombineTo was used. Since CombineTo takes care of the worklist
1105 // mechanics for us, we have no work to do in this case.
1106 if (RV.getNode() == N)
1109 assert(N->getOpcode() != ISD::DELETED_NODE &&
1110 RV.getNode()->getOpcode() != ISD::DELETED_NODE &&
1111 "Node was deleted but visit returned new node!");
1113 DEBUG(dbgs() << "\nReplacing.3 ";
1115 dbgs() << "\nWith: ";
1116 RV.getNode()->dump(&DAG);
1119 // Transfer debug value.
1120 DAG.TransferDbgValues(SDValue(N, 0), RV);
1121 WorkListRemover DeadNodes(*this);
1122 if (N->getNumValues() == RV.getNode()->getNumValues())
1123 DAG.ReplaceAllUsesWith(N, RV.getNode());
1125 assert(N->getValueType(0) == RV.getValueType() &&
1126 N->getNumValues() == 1 && "Type mismatch");
1128 DAG.ReplaceAllUsesWith(N, &OpV);
1131 // Push the new node and any users onto the worklist
1132 AddToWorkList(RV.getNode());
1133 AddUsersToWorkList(RV.getNode());
1135 // Add any uses of the old node to the worklist in case this node is the
1136 // last one that uses them. They may become dead after this node is
1138 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1139 AddToWorkList(N->getOperand(i).getNode());
1141 // Finally, if the node is now dead, remove it from the graph. The node
1142 // may not be dead if the replacement process recursively simplified to
1143 // something else needing this node.
1144 if (N->use_empty()) {
1145 // Nodes can be reintroduced into the worklist. Make sure we do not
1146 // process a node that has been replaced.
1147 removeFromWorkList(N);
1149 // Finally, since the node is now dead, remove it from the graph.
1154 // If the root changed (e.g. it was a dead load, update the root).
1155 DAG.setRoot(Dummy.getValue());
1156 DAG.RemoveDeadNodes();
1159 SDValue DAGCombiner::visit(SDNode *N) {
1160 switch (N->getOpcode()) {
1162 case ISD::TokenFactor: return visitTokenFactor(N);
1163 case ISD::MERGE_VALUES: return visitMERGE_VALUES(N);
1164 case ISD::ADD: return visitADD(N);
1165 case ISD::SUB: return visitSUB(N);
1166 case ISD::ADDC: return visitADDC(N);
1167 case ISD::SUBC: return visitSUBC(N);
1168 case ISD::ADDE: return visitADDE(N);
1169 case ISD::SUBE: return visitSUBE(N);
1170 case ISD::MUL: return visitMUL(N);
1171 case ISD::SDIV: return visitSDIV(N);
1172 case ISD::UDIV: return visitUDIV(N);
1173 case ISD::SREM: return visitSREM(N);
1174 case ISD::UREM: return visitUREM(N);
1175 case ISD::MULHU: return visitMULHU(N);
1176 case ISD::MULHS: return visitMULHS(N);
1177 case ISD::SMUL_LOHI: return visitSMUL_LOHI(N);
1178 case ISD::UMUL_LOHI: return visitUMUL_LOHI(N);
1179 case ISD::SMULO: return visitSMULO(N);
1180 case ISD::UMULO: return visitUMULO(N);
1181 case ISD::SDIVREM: return visitSDIVREM(N);
1182 case ISD::UDIVREM: return visitUDIVREM(N);
1183 case ISD::AND: return visitAND(N);
1184 case ISD::OR: return visitOR(N);
1185 case ISD::XOR: return visitXOR(N);
1186 case ISD::SHL: return visitSHL(N);
1187 case ISD::SRA: return visitSRA(N);
1188 case ISD::SRL: return visitSRL(N);
1189 case ISD::CTLZ: return visitCTLZ(N);
1190 case ISD::CTLZ_ZERO_UNDEF: return visitCTLZ_ZERO_UNDEF(N);
1191 case ISD::CTTZ: return visitCTTZ(N);
1192 case ISD::CTTZ_ZERO_UNDEF: return visitCTTZ_ZERO_UNDEF(N);
1193 case ISD::CTPOP: return visitCTPOP(N);
1194 case ISD::SELECT: return visitSELECT(N);
1195 case ISD::VSELECT: return visitVSELECT(N);
1196 case ISD::SELECT_CC: return visitSELECT_CC(N);
1197 case ISD::SETCC: return visitSETCC(N);
1198 case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N);
1199 case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N);
1200 case ISD::ANY_EXTEND: return visitANY_EXTEND(N);
1201 case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N);
1202 case ISD::TRUNCATE: return visitTRUNCATE(N);
1203 case ISD::BITCAST: return visitBITCAST(N);
1204 case ISD::BUILD_PAIR: return visitBUILD_PAIR(N);
1205 case ISD::FADD: return visitFADD(N);
1206 case ISD::FSUB: return visitFSUB(N);
1207 case ISD::FMUL: return visitFMUL(N);
1208 case ISD::FMA: return visitFMA(N);
1209 case ISD::FDIV: return visitFDIV(N);
1210 case ISD::FREM: return visitFREM(N);
1211 case ISD::FCOPYSIGN: return visitFCOPYSIGN(N);
1212 case ISD::SINT_TO_FP: return visitSINT_TO_FP(N);
1213 case ISD::UINT_TO_FP: return visitUINT_TO_FP(N);
1214 case ISD::FP_TO_SINT: return visitFP_TO_SINT(N);
1215 case ISD::FP_TO_UINT: return visitFP_TO_UINT(N);
1216 case ISD::FP_ROUND: return visitFP_ROUND(N);
1217 case ISD::FP_ROUND_INREG: return visitFP_ROUND_INREG(N);
1218 case ISD::FP_EXTEND: return visitFP_EXTEND(N);
1219 case ISD::FNEG: return visitFNEG(N);
1220 case ISD::FABS: return visitFABS(N);
1221 case ISD::FFLOOR: return visitFFLOOR(N);
1222 case ISD::FCEIL: return visitFCEIL(N);
1223 case ISD::FTRUNC: return visitFTRUNC(N);
1224 case ISD::BRCOND: return visitBRCOND(N);
1225 case ISD::BR_CC: return visitBR_CC(N);
1226 case ISD::LOAD: return visitLOAD(N);
1227 case ISD::STORE: return visitSTORE(N);
1228 case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N);
1229 case ISD::EXTRACT_VECTOR_ELT: return visitEXTRACT_VECTOR_ELT(N);
1230 case ISD::BUILD_VECTOR: return visitBUILD_VECTOR(N);
1231 case ISD::CONCAT_VECTORS: return visitCONCAT_VECTORS(N);
1232 case ISD::EXTRACT_SUBVECTOR: return visitEXTRACT_SUBVECTOR(N);
1233 case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N);
1234 case ISD::INSERT_SUBVECTOR: return visitINSERT_SUBVECTOR(N);
1239 SDValue DAGCombiner::combine(SDNode *N) {
1240 SDValue RV = visit(N);
1242 // If nothing happened, try a target-specific DAG combine.
1243 if (RV.getNode() == 0) {
1244 assert(N->getOpcode() != ISD::DELETED_NODE &&
1245 "Node was deleted but visit returned NULL!");
1247 if (N->getOpcode() >= ISD::BUILTIN_OP_END ||
1248 TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode())) {
1250 // Expose the DAG combiner to the target combiner impls.
1251 TargetLowering::DAGCombinerInfo
1252 DagCombineInfo(DAG, Level, false, this);
1254 RV = TLI.PerformDAGCombine(N, DagCombineInfo);
1258 // If nothing happened still, try promoting the operation.
1259 if (RV.getNode() == 0) {
1260 switch (N->getOpcode()) {
1268 RV = PromoteIntBinOp(SDValue(N, 0));
1273 RV = PromoteIntShiftOp(SDValue(N, 0));
1275 case ISD::SIGN_EXTEND:
1276 case ISD::ZERO_EXTEND:
1277 case ISD::ANY_EXTEND:
1278 RV = PromoteExtend(SDValue(N, 0));
1281 if (PromoteLoad(SDValue(N, 0)))
1287 // If N is a commutative binary node, try commuting it to enable more
1289 if (RV.getNode() == 0 &&
1290 SelectionDAG::isCommutativeBinOp(N->getOpcode()) &&
1291 N->getNumValues() == 1) {
1292 SDValue N0 = N->getOperand(0);
1293 SDValue N1 = N->getOperand(1);
1295 // Constant operands are canonicalized to RHS.
1296 if (isa<ConstantSDNode>(N0) || !isa<ConstantSDNode>(N1)) {
1297 SDValue Ops[] = { N1, N0 };
1298 SDNode *CSENode = DAG.getNodeIfExists(N->getOpcode(), N->getVTList(),
1301 return SDValue(CSENode, 0);
1308 /// getInputChainForNode - Given a node, return its input chain if it has one,
1309 /// otherwise return a null sd operand.
1310 static SDValue getInputChainForNode(SDNode *N) {
1311 if (unsigned NumOps = N->getNumOperands()) {
1312 if (N->getOperand(0).getValueType() == MVT::Other)
1313 return N->getOperand(0);
1314 if (N->getOperand(NumOps-1).getValueType() == MVT::Other)
1315 return N->getOperand(NumOps-1);
1316 for (unsigned i = 1; i < NumOps-1; ++i)
1317 if (N->getOperand(i).getValueType() == MVT::Other)
1318 return N->getOperand(i);
1323 SDValue DAGCombiner::visitTokenFactor(SDNode *N) {
1324 // If N has two operands, where one has an input chain equal to the other,
1325 // the 'other' chain is redundant.
1326 if (N->getNumOperands() == 2) {
1327 if (getInputChainForNode(N->getOperand(0).getNode()) == N->getOperand(1))
1328 return N->getOperand(0);
1329 if (getInputChainForNode(N->getOperand(1).getNode()) == N->getOperand(0))
1330 return N->getOperand(1);
1333 SmallVector<SDNode *, 8> TFs; // List of token factors to visit.
1334 SmallVector<SDValue, 8> Ops; // Ops for replacing token factor.
1335 SmallPtrSet<SDNode*, 16> SeenOps;
1336 bool Changed = false; // If we should replace this token factor.
1338 // Start out with this token factor.
1341 // Iterate through token factors. The TFs grows when new token factors are
1343 for (unsigned i = 0; i < TFs.size(); ++i) {
1344 SDNode *TF = TFs[i];
1346 // Check each of the operands.
1347 for (unsigned i = 0, ie = TF->getNumOperands(); i != ie; ++i) {
1348 SDValue Op = TF->getOperand(i);
1350 switch (Op.getOpcode()) {
1351 case ISD::EntryToken:
1352 // Entry tokens don't need to be added to the list. They are
1357 case ISD::TokenFactor:
1358 if (Op.hasOneUse() &&
1359 std::find(TFs.begin(), TFs.end(), Op.getNode()) == TFs.end()) {
1360 // Queue up for processing.
1361 TFs.push_back(Op.getNode());
1362 // Clean up in case the token factor is removed.
1363 AddToWorkList(Op.getNode());
1370 // Only add if it isn't already in the list.
1371 if (SeenOps.insert(Op.getNode()))
1382 // If we've change things around then replace token factor.
1385 // The entry token is the only possible outcome.
1386 Result = DAG.getEntryNode();
1388 // New and improved token factor.
1389 Result = DAG.getNode(ISD::TokenFactor, SDLoc(N),
1390 MVT::Other, &Ops[0], Ops.size());
1393 // Don't add users to work list.
1394 return CombineTo(N, Result, false);
1400 /// MERGE_VALUES can always be eliminated.
1401 SDValue DAGCombiner::visitMERGE_VALUES(SDNode *N) {
1402 WorkListRemover DeadNodes(*this);
1403 // Replacing results may cause a different MERGE_VALUES to suddenly
1404 // be CSE'd with N, and carry its uses with it. Iterate until no
1405 // uses remain, to ensure that the node can be safely deleted.
1406 // First add the users of this node to the work list so that they
1407 // can be tried again once they have new operands.
1408 AddUsersToWorkList(N);
1410 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1411 DAG.ReplaceAllUsesOfValueWith(SDValue(N, i), N->getOperand(i));
1412 } while (!N->use_empty());
1413 removeFromWorkList(N);
1415 return SDValue(N, 0); // Return N so it doesn't get rechecked!
1419 SDValue combineShlAddConstant(SDLoc DL, SDValue N0, SDValue N1,
1420 SelectionDAG &DAG) {
1421 EVT VT = N0.getValueType();
1422 SDValue N00 = N0.getOperand(0);
1423 SDValue N01 = N0.getOperand(1);
1424 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N01);
1426 if (N01C && N00.getOpcode() == ISD::ADD && N00.getNode()->hasOneUse() &&
1427 isa<ConstantSDNode>(N00.getOperand(1))) {
1428 // fold (add (shl (add x, c1), c2), ) -> (add (add (shl x, c2), c1<<c2), )
1429 N0 = DAG.getNode(ISD::ADD, SDLoc(N0), VT,
1430 DAG.getNode(ISD::SHL, SDLoc(N00), VT,
1431 N00.getOperand(0), N01),
1432 DAG.getNode(ISD::SHL, SDLoc(N01), VT,
1433 N00.getOperand(1), N01));
1434 return DAG.getNode(ISD::ADD, DL, VT, N0, N1);
1440 SDValue DAGCombiner::visitADD(SDNode *N) {
1441 SDValue N0 = N->getOperand(0);
1442 SDValue N1 = N->getOperand(1);
1443 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1444 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1445 EVT VT = N0.getValueType();
1448 if (VT.isVector()) {
1449 SDValue FoldedVOp = SimplifyVBinOp(N);
1450 if (FoldedVOp.getNode()) return FoldedVOp;
1452 // fold (add x, 0) -> x, vector edition
1453 if (ISD::isBuildVectorAllZeros(N1.getNode()))
1455 if (ISD::isBuildVectorAllZeros(N0.getNode()))
1459 // fold (add x, undef) -> undef
1460 if (N0.getOpcode() == ISD::UNDEF)
1462 if (N1.getOpcode() == ISD::UNDEF)
1464 // fold (add c1, c2) -> c1+c2
1466 return DAG.FoldConstantArithmetic(ISD::ADD, VT, N0C, N1C);
1467 // canonicalize constant to RHS
1469 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N1, N0);
1470 // fold (add x, 0) -> x
1471 if (N1C && N1C->isNullValue())
1473 // fold (add Sym, c) -> Sym+c
1474 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1475 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA) && N1C &&
1476 GA->getOpcode() == ISD::GlobalAddress)
1477 return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT,
1479 (uint64_t)N1C->getSExtValue());
1480 // fold ((c1-A)+c2) -> (c1+c2)-A
1481 if (N1C && N0.getOpcode() == ISD::SUB)
1482 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getOperand(0)))
1483 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1484 DAG.getConstant(N1C->getAPIntValue()+
1485 N0C->getAPIntValue(), VT),
1488 SDValue RADD = ReassociateOps(ISD::ADD, SDLoc(N), N0, N1);
1489 if (RADD.getNode() != 0)
1491 // fold ((0-A) + B) -> B-A
1492 if (N0.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N0.getOperand(0)) &&
1493 cast<ConstantSDNode>(N0.getOperand(0))->isNullValue())
1494 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1, N0.getOperand(1));
1495 // fold (A + (0-B)) -> A-B
1496 if (N1.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N1.getOperand(0)) &&
1497 cast<ConstantSDNode>(N1.getOperand(0))->isNullValue())
1498 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, N1.getOperand(1));
1499 // fold (A+(B-A)) -> B
1500 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1))
1501 return N1.getOperand(0);
1502 // fold ((B-A)+A) -> B
1503 if (N0.getOpcode() == ISD::SUB && N1 == N0.getOperand(1))
1504 return N0.getOperand(0);
1505 // fold (A+(B-(A+C))) to (B-C)
1506 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1507 N0 == N1.getOperand(1).getOperand(0))
1508 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1.getOperand(0),
1509 N1.getOperand(1).getOperand(1));
1510 // fold (A+(B-(C+A))) to (B-C)
1511 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1512 N0 == N1.getOperand(1).getOperand(1))
1513 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1.getOperand(0),
1514 N1.getOperand(1).getOperand(0));
1515 // fold (A+((B-A)+or-C)) to (B+or-C)
1516 if ((N1.getOpcode() == ISD::SUB || N1.getOpcode() == ISD::ADD) &&
1517 N1.getOperand(0).getOpcode() == ISD::SUB &&
1518 N0 == N1.getOperand(0).getOperand(1))
1519 return DAG.getNode(N1.getOpcode(), SDLoc(N), VT,
1520 N1.getOperand(0).getOperand(0), N1.getOperand(1));
1522 // fold (A-B)+(C-D) to (A+C)-(B+D) when A or C is constant
1523 if (N0.getOpcode() == ISD::SUB && N1.getOpcode() == ISD::SUB) {
1524 SDValue N00 = N0.getOperand(0);
1525 SDValue N01 = N0.getOperand(1);
1526 SDValue N10 = N1.getOperand(0);
1527 SDValue N11 = N1.getOperand(1);
1529 if (isa<ConstantSDNode>(N00) || isa<ConstantSDNode>(N10))
1530 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1531 DAG.getNode(ISD::ADD, SDLoc(N0), VT, N00, N10),
1532 DAG.getNode(ISD::ADD, SDLoc(N1), VT, N01, N11));
1535 if (!VT.isVector() && SimplifyDemandedBits(SDValue(N, 0)))
1536 return SDValue(N, 0);
1538 // fold (a+b) -> (a|b) iff a and b share no bits.
1539 if (VT.isInteger() && !VT.isVector()) {
1540 APInt LHSZero, LHSOne;
1541 APInt RHSZero, RHSOne;
1542 DAG.ComputeMaskedBits(N0, LHSZero, LHSOne);
1544 if (LHSZero.getBoolValue()) {
1545 DAG.ComputeMaskedBits(N1, RHSZero, RHSOne);
1547 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1548 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1549 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero){
1550 if (!LegalOperations || TLI.isOperationLegal(ISD::OR, VT))
1551 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1);
1556 // fold (add (shl (add x, c1), c2), ) -> (add (add (shl x, c2), c1<<c2), )
1557 if (N0.getOpcode() == ISD::SHL && N0.getNode()->hasOneUse()) {
1558 SDValue Result = combineShlAddConstant(SDLoc(N), N0, N1, DAG);
1559 if (Result.getNode()) return Result;
1561 if (N1.getOpcode() == ISD::SHL && N1.getNode()->hasOneUse()) {
1562 SDValue Result = combineShlAddConstant(SDLoc(N), N1, N0, DAG);
1563 if (Result.getNode()) return Result;
1566 // fold (add x, shl(0 - y, n)) -> sub(x, shl(y, n))
1567 if (N1.getOpcode() == ISD::SHL &&
1568 N1.getOperand(0).getOpcode() == ISD::SUB)
1569 if (ConstantSDNode *C =
1570 dyn_cast<ConstantSDNode>(N1.getOperand(0).getOperand(0)))
1571 if (C->getAPIntValue() == 0)
1572 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N0,
1573 DAG.getNode(ISD::SHL, SDLoc(N), VT,
1574 N1.getOperand(0).getOperand(1),
1576 if (N0.getOpcode() == ISD::SHL &&
1577 N0.getOperand(0).getOpcode() == ISD::SUB)
1578 if (ConstantSDNode *C =
1579 dyn_cast<ConstantSDNode>(N0.getOperand(0).getOperand(0)))
1580 if (C->getAPIntValue() == 0)
1581 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1,
1582 DAG.getNode(ISD::SHL, SDLoc(N), VT,
1583 N0.getOperand(0).getOperand(1),
1586 if (N1.getOpcode() == ISD::AND) {
1587 SDValue AndOp0 = N1.getOperand(0);
1588 ConstantSDNode *AndOp1 = dyn_cast<ConstantSDNode>(N1->getOperand(1));
1589 unsigned NumSignBits = DAG.ComputeNumSignBits(AndOp0);
1590 unsigned DestBits = VT.getScalarType().getSizeInBits();
1592 // (add z, (and (sbbl x, x), 1)) -> (sub z, (sbbl x, x))
1593 // and similar xforms where the inner op is either ~0 or 0.
1594 if (NumSignBits == DestBits && AndOp1 && AndOp1->isOne()) {
1596 return DAG.getNode(ISD::SUB, DL, VT, N->getOperand(0), AndOp0);
1600 // add (sext i1), X -> sub X, (zext i1)
1601 if (N0.getOpcode() == ISD::SIGN_EXTEND &&
1602 N0.getOperand(0).getValueType() == MVT::i1 &&
1603 !TLI.isOperationLegal(ISD::SIGN_EXTEND, MVT::i1)) {
1605 SDValue ZExt = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0));
1606 return DAG.getNode(ISD::SUB, DL, VT, N1, ZExt);
1612 SDValue DAGCombiner::visitADDC(SDNode *N) {
1613 SDValue N0 = N->getOperand(0);
1614 SDValue N1 = N->getOperand(1);
1615 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1616 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1617 EVT VT = N0.getValueType();
1619 // If the flag result is dead, turn this into an ADD.
1620 if (!N->hasAnyUseOfValue(1))
1621 return CombineTo(N, DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N1),
1622 DAG.getNode(ISD::CARRY_FALSE,
1623 SDLoc(N), MVT::Glue));
1625 // canonicalize constant to RHS.
1627 return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N1, N0);
1629 // fold (addc x, 0) -> x + no carry out
1630 if (N1C && N1C->isNullValue())
1631 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE,
1632 SDLoc(N), MVT::Glue));
1634 // fold (addc a, b) -> (or a, b), CARRY_FALSE iff a and b share no bits.
1635 APInt LHSZero, LHSOne;
1636 APInt RHSZero, RHSOne;
1637 DAG.ComputeMaskedBits(N0, LHSZero, LHSOne);
1639 if (LHSZero.getBoolValue()) {
1640 DAG.ComputeMaskedBits(N1, RHSZero, RHSOne);
1642 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1643 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1644 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero)
1645 return CombineTo(N, DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1),
1646 DAG.getNode(ISD::CARRY_FALSE,
1647 SDLoc(N), MVT::Glue));
1653 SDValue DAGCombiner::visitADDE(SDNode *N) {
1654 SDValue N0 = N->getOperand(0);
1655 SDValue N1 = N->getOperand(1);
1656 SDValue CarryIn = N->getOperand(2);
1657 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1658 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1660 // canonicalize constant to RHS
1662 return DAG.getNode(ISD::ADDE, SDLoc(N), N->getVTList(),
1665 // fold (adde x, y, false) -> (addc x, y)
1666 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1667 return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N0, N1);
1672 // Since it may not be valid to emit a fold to zero for vector initializers
1673 // check if we can before folding.
1674 static SDValue tryFoldToZero(SDLoc DL, const TargetLowering &TLI, EVT VT,
1676 bool LegalOperations, bool LegalTypes) {
1678 return DAG.getConstant(0, VT);
1679 if (!LegalOperations || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
1680 return DAG.getConstant(0, VT);
1684 SDValue DAGCombiner::visitSUB(SDNode *N) {
1685 SDValue N0 = N->getOperand(0);
1686 SDValue N1 = N->getOperand(1);
1687 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1688 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1689 ConstantSDNode *N1C1 = N1.getOpcode() != ISD::ADD ? 0 :
1690 dyn_cast<ConstantSDNode>(N1.getOperand(1).getNode());
1691 EVT VT = N0.getValueType();
1694 if (VT.isVector()) {
1695 SDValue FoldedVOp = SimplifyVBinOp(N);
1696 if (FoldedVOp.getNode()) return FoldedVOp;
1698 // fold (sub x, 0) -> x, vector edition
1699 if (ISD::isBuildVectorAllZeros(N1.getNode()))
1703 // fold (sub x, x) -> 0
1704 // FIXME: Refactor this and xor and other similar operations together.
1706 return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations, LegalTypes);
1707 // fold (sub c1, c2) -> c1-c2
1709 return DAG.FoldConstantArithmetic(ISD::SUB, VT, N0C, N1C);
1710 // fold (sub x, c) -> (add x, -c)
1712 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0,
1713 DAG.getConstant(-N1C->getAPIntValue(), VT));
1714 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1)
1715 if (N0C && N0C->isAllOnesValue())
1716 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0);
1717 // fold A-(A-B) -> B
1718 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(0))
1719 return N1.getOperand(1);
1720 // fold (A+B)-A -> B
1721 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1)
1722 return N0.getOperand(1);
1723 // fold (A+B)-B -> A
1724 if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1)
1725 return N0.getOperand(0);
1726 // fold C2-(A+C1) -> (C2-C1)-A
1727 if (N1.getOpcode() == ISD::ADD && N0C && N1C1) {
1728 SDValue NewC = DAG.getConstant(N0C->getAPIntValue() - N1C1->getAPIntValue(),
1730 return DAG.getNode(ISD::SUB, SDLoc(N), VT, NewC,
1733 // fold ((A+(B+or-C))-B) -> A+or-C
1734 if (N0.getOpcode() == ISD::ADD &&
1735 (N0.getOperand(1).getOpcode() == ISD::SUB ||
1736 N0.getOperand(1).getOpcode() == ISD::ADD) &&
1737 N0.getOperand(1).getOperand(0) == N1)
1738 return DAG.getNode(N0.getOperand(1).getOpcode(), SDLoc(N), VT,
1739 N0.getOperand(0), N0.getOperand(1).getOperand(1));
1740 // fold ((A+(C+B))-B) -> A+C
1741 if (N0.getOpcode() == ISD::ADD &&
1742 N0.getOperand(1).getOpcode() == ISD::ADD &&
1743 N0.getOperand(1).getOperand(1) == N1)
1744 return DAG.getNode(ISD::ADD, SDLoc(N), VT,
1745 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1746 // fold ((A-(B-C))-C) -> A-B
1747 if (N0.getOpcode() == ISD::SUB &&
1748 N0.getOperand(1).getOpcode() == ISD::SUB &&
1749 N0.getOperand(1).getOperand(1) == N1)
1750 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1751 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1753 // If either operand of a sub is undef, the result is undef
1754 if (N0.getOpcode() == ISD::UNDEF)
1756 if (N1.getOpcode() == ISD::UNDEF)
1759 // If the relocation model supports it, consider symbol offsets.
1760 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1761 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA)) {
1762 // fold (sub Sym, c) -> Sym-c
1763 if (N1C && GA->getOpcode() == ISD::GlobalAddress)
1764 return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT,
1766 (uint64_t)N1C->getSExtValue());
1767 // fold (sub Sym+c1, Sym+c2) -> c1-c2
1768 if (GlobalAddressSDNode *GB = dyn_cast<GlobalAddressSDNode>(N1))
1769 if (GA->getGlobal() == GB->getGlobal())
1770 return DAG.getConstant((uint64_t)GA->getOffset() - GB->getOffset(),
1777 SDValue DAGCombiner::visitSUBC(SDNode *N) {
1778 SDValue N0 = N->getOperand(0);
1779 SDValue N1 = N->getOperand(1);
1780 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1781 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1782 EVT VT = N0.getValueType();
1784 // If the flag result is dead, turn this into an SUB.
1785 if (!N->hasAnyUseOfValue(1))
1786 return CombineTo(N, DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, N1),
1787 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1790 // fold (subc x, x) -> 0 + no borrow
1792 return CombineTo(N, DAG.getConstant(0, VT),
1793 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1796 // fold (subc x, 0) -> x + no borrow
1797 if (N1C && N1C->isNullValue())
1798 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1801 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1) + no borrow
1802 if (N0C && N0C->isAllOnesValue())
1803 return CombineTo(N, DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0),
1804 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1810 SDValue DAGCombiner::visitSUBE(SDNode *N) {
1811 SDValue N0 = N->getOperand(0);
1812 SDValue N1 = N->getOperand(1);
1813 SDValue CarryIn = N->getOperand(2);
1815 // fold (sube x, y, false) -> (subc x, y)
1816 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1817 return DAG.getNode(ISD::SUBC, SDLoc(N), N->getVTList(), N0, N1);
1822 /// isConstantSplatVector - Returns true if N is a BUILD_VECTOR node whose
1823 /// elements are all the same constant or undefined.
1824 static bool isConstantSplatVector(SDNode *N, APInt& SplatValue) {
1825 BuildVectorSDNode *C = dyn_cast<BuildVectorSDNode>(N);
1830 unsigned SplatBitSize;
1832 EVT EltVT = N->getValueType(0).getVectorElementType();
1833 return (C->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
1835 EltVT.getSizeInBits() >= SplatBitSize);
1838 SDValue DAGCombiner::visitMUL(SDNode *N) {
1839 SDValue N0 = N->getOperand(0);
1840 SDValue N1 = N->getOperand(1);
1841 EVT VT = N0.getValueType();
1843 // fold (mul x, undef) -> 0
1844 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
1845 return DAG.getConstant(0, VT);
1847 bool N0IsConst = false;
1848 bool N1IsConst = false;
1849 APInt ConstValue0, ConstValue1;
1851 if (VT.isVector()) {
1852 SDValue FoldedVOp = SimplifyVBinOp(N);
1853 if (FoldedVOp.getNode()) return FoldedVOp;
1855 N0IsConst = isConstantSplatVector(N0.getNode(), ConstValue0);
1856 N1IsConst = isConstantSplatVector(N1.getNode(), ConstValue1);
1858 N0IsConst = dyn_cast<ConstantSDNode>(N0) != 0;
1859 ConstValue0 = N0IsConst ? (dyn_cast<ConstantSDNode>(N0))->getAPIntValue()
1861 N1IsConst = dyn_cast<ConstantSDNode>(N1) != 0;
1862 ConstValue1 = N1IsConst ? (dyn_cast<ConstantSDNode>(N1))->getAPIntValue()
1866 // fold (mul c1, c2) -> c1*c2
1867 if (N0IsConst && N1IsConst)
1868 return DAG.FoldConstantArithmetic(ISD::MUL, VT, N0.getNode(), N1.getNode());
1870 // canonicalize constant to RHS
1871 if (N0IsConst && !N1IsConst)
1872 return DAG.getNode(ISD::MUL, SDLoc(N), VT, N1, N0);
1873 // fold (mul x, 0) -> 0
1874 if (N1IsConst && ConstValue1 == 0)
1876 // We require a splat of the entire scalar bit width for non-contiguous
1879 ConstValue1.getBitWidth() == VT.getScalarType().getSizeInBits();
1880 // fold (mul x, 1) -> x
1881 if (N1IsConst && ConstValue1 == 1 && IsFullSplat)
1883 // fold (mul x, -1) -> 0-x
1884 if (N1IsConst && ConstValue1.isAllOnesValue())
1885 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1886 DAG.getConstant(0, VT), N0);
1887 // fold (mul x, (1 << c)) -> x << c
1888 if (N1IsConst && ConstValue1.isPowerOf2() && IsFullSplat)
1889 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0,
1890 DAG.getConstant(ConstValue1.logBase2(),
1891 getShiftAmountTy(N0.getValueType())));
1892 // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c
1893 if (N1IsConst && (-ConstValue1).isPowerOf2() && IsFullSplat) {
1894 unsigned Log2Val = (-ConstValue1).logBase2();
1895 // FIXME: If the input is something that is easily negated (e.g. a
1896 // single-use add), we should put the negate there.
1897 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1898 DAG.getConstant(0, VT),
1899 DAG.getNode(ISD::SHL, SDLoc(N), VT, N0,
1900 DAG.getConstant(Log2Val,
1901 getShiftAmountTy(N0.getValueType()))));
1905 // (mul (shl X, c1), c2) -> (mul X, c2 << c1)
1906 if (N1IsConst && N0.getOpcode() == ISD::SHL &&
1907 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
1908 isa<ConstantSDNode>(N0.getOperand(1)))) {
1909 SDValue C3 = DAG.getNode(ISD::SHL, SDLoc(N), VT,
1910 N1, N0.getOperand(1));
1911 AddToWorkList(C3.getNode());
1912 return DAG.getNode(ISD::MUL, SDLoc(N), VT,
1913 N0.getOperand(0), C3);
1916 // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one
1919 SDValue Sh(0,0), Y(0,0);
1920 // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)).
1921 if (N0.getOpcode() == ISD::SHL &&
1922 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
1923 isa<ConstantSDNode>(N0.getOperand(1))) &&
1924 N0.getNode()->hasOneUse()) {
1926 } else if (N1.getOpcode() == ISD::SHL &&
1927 isa<ConstantSDNode>(N1.getOperand(1)) &&
1928 N1.getNode()->hasOneUse()) {
1933 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
1934 Sh.getOperand(0), Y);
1935 return DAG.getNode(ISD::SHL, SDLoc(N), VT,
1936 Mul, Sh.getOperand(1));
1940 // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2)
1941 if (N1IsConst && N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() &&
1942 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
1943 isa<ConstantSDNode>(N0.getOperand(1))))
1944 return DAG.getNode(ISD::ADD, SDLoc(N), VT,
1945 DAG.getNode(ISD::MUL, SDLoc(N0), VT,
1946 N0.getOperand(0), N1),
1947 DAG.getNode(ISD::MUL, SDLoc(N1), VT,
1948 N0.getOperand(1), N1));
1951 SDValue RMUL = ReassociateOps(ISD::MUL, SDLoc(N), N0, N1);
1952 if (RMUL.getNode() != 0)
1958 SDValue DAGCombiner::visitSDIV(SDNode *N) {
1959 SDValue N0 = N->getOperand(0);
1960 SDValue N1 = N->getOperand(1);
1961 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1962 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1963 EVT VT = N->getValueType(0);
1966 if (VT.isVector()) {
1967 SDValue FoldedVOp = SimplifyVBinOp(N);
1968 if (FoldedVOp.getNode()) return FoldedVOp;
1971 // fold (sdiv c1, c2) -> c1/c2
1972 if (N0C && N1C && !N1C->isNullValue())
1973 return DAG.FoldConstantArithmetic(ISD::SDIV, VT, N0C, N1C);
1974 // fold (sdiv X, 1) -> X
1975 if (N1C && N1C->getAPIntValue() == 1LL)
1977 // fold (sdiv X, -1) -> 0-X
1978 if (N1C && N1C->isAllOnesValue())
1979 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1980 DAG.getConstant(0, VT), N0);
1981 // If we know the sign bits of both operands are zero, strength reduce to a
1982 // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2
1983 if (!VT.isVector()) {
1984 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
1985 return DAG.getNode(ISD::UDIV, SDLoc(N), N1.getValueType(),
1988 // fold (sdiv X, pow2) -> simple ops after legalize
1989 if (N1C && !N1C->isNullValue() &&
1990 (N1C->getAPIntValue().isPowerOf2() ||
1991 (-N1C->getAPIntValue()).isPowerOf2())) {
1992 // If dividing by powers of two is cheap, then don't perform the following
1994 if (TLI.isPow2DivCheap())
1997 unsigned lg2 = N1C->getAPIntValue().countTrailingZeros();
1999 // Splat the sign bit into the register
2000 SDValue SGN = DAG.getNode(ISD::SRA, SDLoc(N), VT, N0,
2001 DAG.getConstant(VT.getSizeInBits()-1,
2002 getShiftAmountTy(N0.getValueType())));
2003 AddToWorkList(SGN.getNode());
2005 // Add (N0 < 0) ? abs2 - 1 : 0;
2006 SDValue SRL = DAG.getNode(ISD::SRL, SDLoc(N), VT, SGN,
2007 DAG.getConstant(VT.getSizeInBits() - lg2,
2008 getShiftAmountTy(SGN.getValueType())));
2009 SDValue ADD = DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, SRL);
2010 AddToWorkList(SRL.getNode());
2011 AddToWorkList(ADD.getNode()); // Divide by pow2
2012 SDValue SRA = DAG.getNode(ISD::SRA, SDLoc(N), VT, ADD,
2013 DAG.getConstant(lg2, getShiftAmountTy(ADD.getValueType())));
2015 // If we're dividing by a positive value, we're done. Otherwise, we must
2016 // negate the result.
2017 if (N1C->getAPIntValue().isNonNegative())
2020 AddToWorkList(SRA.getNode());
2021 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
2022 DAG.getConstant(0, VT), SRA);
2025 // if integer divide is expensive and we satisfy the requirements, emit an
2026 // alternate sequence.
2027 if (N1C && !N1C->isNullValue() && !TLI.isIntDivCheap()) {
2028 SDValue Op = BuildSDIV(N);
2029 if (Op.getNode()) return Op;
2033 if (N0.getOpcode() == ISD::UNDEF)
2034 return DAG.getConstant(0, VT);
2035 // X / undef -> undef
2036 if (N1.getOpcode() == ISD::UNDEF)
2042 SDValue DAGCombiner::visitUDIV(SDNode *N) {
2043 SDValue N0 = N->getOperand(0);
2044 SDValue N1 = N->getOperand(1);
2045 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
2046 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
2047 EVT VT = N->getValueType(0);
2050 if (VT.isVector()) {
2051 SDValue FoldedVOp = SimplifyVBinOp(N);
2052 if (FoldedVOp.getNode()) return FoldedVOp;
2055 // fold (udiv c1, c2) -> c1/c2
2056 if (N0C && N1C && !N1C->isNullValue())
2057 return DAG.FoldConstantArithmetic(ISD::UDIV, VT, N0C, N1C);
2058 // fold (udiv x, (1 << c)) -> x >>u c
2059 if (N1C && N1C->getAPIntValue().isPowerOf2())
2060 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0,
2061 DAG.getConstant(N1C->getAPIntValue().logBase2(),
2062 getShiftAmountTy(N0.getValueType())));
2063 // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2
2064 if (N1.getOpcode() == ISD::SHL) {
2065 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
2066 if (SHC->getAPIntValue().isPowerOf2()) {
2067 EVT ADDVT = N1.getOperand(1).getValueType();
2068 SDValue Add = DAG.getNode(ISD::ADD, SDLoc(N), ADDVT,
2070 DAG.getConstant(SHC->getAPIntValue()
2073 AddToWorkList(Add.getNode());
2074 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, Add);
2078 // fold (udiv x, c) -> alternate
2079 if (N1C && !N1C->isNullValue() && !TLI.isIntDivCheap()) {
2080 SDValue Op = BuildUDIV(N);
2081 if (Op.getNode()) return Op;
2085 if (N0.getOpcode() == ISD::UNDEF)
2086 return DAG.getConstant(0, VT);
2087 // X / undef -> undef
2088 if (N1.getOpcode() == ISD::UNDEF)
2094 SDValue DAGCombiner::visitSREM(SDNode *N) {
2095 SDValue N0 = N->getOperand(0);
2096 SDValue N1 = N->getOperand(1);
2097 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2098 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2099 EVT VT = N->getValueType(0);
2101 // fold (srem c1, c2) -> c1%c2
2102 if (N0C && N1C && !N1C->isNullValue())
2103 return DAG.FoldConstantArithmetic(ISD::SREM, VT, N0C, N1C);
2104 // If we know the sign bits of both operands are zero, strength reduce to a
2105 // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15
2106 if (!VT.isVector()) {
2107 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
2108 return DAG.getNode(ISD::UREM, SDLoc(N), VT, N0, N1);
2111 // If X/C can be simplified by the division-by-constant logic, lower
2112 // X%C to the equivalent of X-X/C*C.
2113 if (N1C && !N1C->isNullValue()) {
2114 SDValue Div = DAG.getNode(ISD::SDIV, SDLoc(N), VT, N0, N1);
2115 AddToWorkList(Div.getNode());
2116 SDValue OptimizedDiv = combine(Div.getNode());
2117 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2118 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2120 SDValue Sub = DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, Mul);
2121 AddToWorkList(Mul.getNode());
2127 if (N0.getOpcode() == ISD::UNDEF)
2128 return DAG.getConstant(0, VT);
2129 // X % undef -> undef
2130 if (N1.getOpcode() == ISD::UNDEF)
2136 SDValue DAGCombiner::visitUREM(SDNode *N) {
2137 SDValue N0 = N->getOperand(0);
2138 SDValue N1 = N->getOperand(1);
2139 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2140 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2141 EVT VT = N->getValueType(0);
2143 // fold (urem c1, c2) -> c1%c2
2144 if (N0C && N1C && !N1C->isNullValue())
2145 return DAG.FoldConstantArithmetic(ISD::UREM, VT, N0C, N1C);
2146 // fold (urem x, pow2) -> (and x, pow2-1)
2147 if (N1C && !N1C->isNullValue() && N1C->getAPIntValue().isPowerOf2())
2148 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0,
2149 DAG.getConstant(N1C->getAPIntValue()-1,VT));
2150 // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1))
2151 if (N1.getOpcode() == ISD::SHL) {
2152 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
2153 if (SHC->getAPIntValue().isPowerOf2()) {
2155 DAG.getNode(ISD::ADD, SDLoc(N), VT, N1,
2156 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()),
2158 AddToWorkList(Add.getNode());
2159 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, Add);
2164 // If X/C can be simplified by the division-by-constant logic, lower
2165 // X%C to the equivalent of X-X/C*C.
2166 if (N1C && !N1C->isNullValue()) {
2167 SDValue Div = DAG.getNode(ISD::UDIV, SDLoc(N), VT, N0, N1);
2168 AddToWorkList(Div.getNode());
2169 SDValue OptimizedDiv = combine(Div.getNode());
2170 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2171 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2173 SDValue Sub = DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, Mul);
2174 AddToWorkList(Mul.getNode());
2180 if (N0.getOpcode() == ISD::UNDEF)
2181 return DAG.getConstant(0, VT);
2182 // X % undef -> undef
2183 if (N1.getOpcode() == ISD::UNDEF)
2189 SDValue DAGCombiner::visitMULHS(SDNode *N) {
2190 SDValue N0 = N->getOperand(0);
2191 SDValue N1 = N->getOperand(1);
2192 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2193 EVT VT = N->getValueType(0);
2196 // fold (mulhs x, 0) -> 0
2197 if (N1C && N1C->isNullValue())
2199 // fold (mulhs x, 1) -> (sra x, size(x)-1)
2200 if (N1C && N1C->getAPIntValue() == 1)
2201 return DAG.getNode(ISD::SRA, SDLoc(N), N0.getValueType(), N0,
2202 DAG.getConstant(N0.getValueType().getSizeInBits() - 1,
2203 getShiftAmountTy(N0.getValueType())));
2204 // fold (mulhs x, undef) -> 0
2205 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2206 return DAG.getConstant(0, VT);
2208 // If the type twice as wide is legal, transform the mulhs to a wider multiply
2210 if (VT.isSimple() && !VT.isVector()) {
2211 MVT Simple = VT.getSimpleVT();
2212 unsigned SimpleSize = Simple.getSizeInBits();
2213 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2214 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2215 N0 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N0);
2216 N1 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N1);
2217 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2218 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2219 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2220 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2227 SDValue DAGCombiner::visitMULHU(SDNode *N) {
2228 SDValue N0 = N->getOperand(0);
2229 SDValue N1 = N->getOperand(1);
2230 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2231 EVT VT = N->getValueType(0);
2234 // fold (mulhu x, 0) -> 0
2235 if (N1C && N1C->isNullValue())
2237 // fold (mulhu x, 1) -> 0
2238 if (N1C && N1C->getAPIntValue() == 1)
2239 return DAG.getConstant(0, N0.getValueType());
2240 // fold (mulhu x, undef) -> 0
2241 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2242 return DAG.getConstant(0, VT);
2244 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2246 if (VT.isSimple() && !VT.isVector()) {
2247 MVT Simple = VT.getSimpleVT();
2248 unsigned SimpleSize = Simple.getSizeInBits();
2249 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2250 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2251 N0 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N0);
2252 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N1);
2253 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2254 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2255 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2256 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2263 /// SimplifyNodeWithTwoResults - Perform optimizations common to nodes that
2264 /// compute two values. LoOp and HiOp give the opcodes for the two computations
2265 /// that are being performed. Return true if a simplification was made.
2267 SDValue DAGCombiner::SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
2269 // If the high half is not needed, just compute the low half.
2270 bool HiExists = N->hasAnyUseOfValue(1);
2272 (!LegalOperations ||
2273 TLI.isOperationLegalOrCustom(LoOp, N->getValueType(0)))) {
2274 SDValue Res = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0),
2275 N->op_begin(), N->getNumOperands());
2276 return CombineTo(N, Res, Res);
2279 // If the low half is not needed, just compute the high half.
2280 bool LoExists = N->hasAnyUseOfValue(0);
2282 (!LegalOperations ||
2283 TLI.isOperationLegal(HiOp, N->getValueType(1)))) {
2284 SDValue Res = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1),
2285 N->op_begin(), N->getNumOperands());
2286 return CombineTo(N, Res, Res);
2289 // If both halves are used, return as it is.
2290 if (LoExists && HiExists)
2293 // If the two computed results can be simplified separately, separate them.
2295 SDValue Lo = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0),
2296 N->op_begin(), N->getNumOperands());
2297 AddToWorkList(Lo.getNode());
2298 SDValue LoOpt = combine(Lo.getNode());
2299 if (LoOpt.getNode() && LoOpt.getNode() != Lo.getNode() &&
2300 (!LegalOperations ||
2301 TLI.isOperationLegal(LoOpt.getOpcode(), LoOpt.getValueType())))
2302 return CombineTo(N, LoOpt, LoOpt);
2306 SDValue Hi = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1),
2307 N->op_begin(), N->getNumOperands());
2308 AddToWorkList(Hi.getNode());
2309 SDValue HiOpt = combine(Hi.getNode());
2310 if (HiOpt.getNode() && HiOpt != Hi &&
2311 (!LegalOperations ||
2312 TLI.isOperationLegal(HiOpt.getOpcode(), HiOpt.getValueType())))
2313 return CombineTo(N, HiOpt, HiOpt);
2319 SDValue DAGCombiner::visitSMUL_LOHI(SDNode *N) {
2320 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHS);
2321 if (Res.getNode()) return Res;
2323 EVT VT = N->getValueType(0);
2326 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2328 if (VT.isSimple() && !VT.isVector()) {
2329 MVT Simple = VT.getSimpleVT();
2330 unsigned SimpleSize = Simple.getSizeInBits();
2331 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2332 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2333 SDValue Lo = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(0));
2334 SDValue Hi = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(1));
2335 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2336 // Compute the high part as N1.
2337 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2338 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2339 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2340 // Compute the low part as N0.
2341 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2342 return CombineTo(N, Lo, Hi);
2349 SDValue DAGCombiner::visitUMUL_LOHI(SDNode *N) {
2350 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHU);
2351 if (Res.getNode()) return Res;
2353 EVT VT = N->getValueType(0);
2356 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2358 if (VT.isSimple() && !VT.isVector()) {
2359 MVT Simple = VT.getSimpleVT();
2360 unsigned SimpleSize = Simple.getSizeInBits();
2361 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2362 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2363 SDValue Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(0));
2364 SDValue Hi = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(1));
2365 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2366 // Compute the high part as N1.
2367 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2368 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2369 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2370 // Compute the low part as N0.
2371 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2372 return CombineTo(N, Lo, Hi);
2379 SDValue DAGCombiner::visitSMULO(SDNode *N) {
2380 // (smulo x, 2) -> (saddo x, x)
2381 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2382 if (C2->getAPIntValue() == 2)
2383 return DAG.getNode(ISD::SADDO, SDLoc(N), N->getVTList(),
2384 N->getOperand(0), N->getOperand(0));
2389 SDValue DAGCombiner::visitUMULO(SDNode *N) {
2390 // (umulo x, 2) -> (uaddo x, x)
2391 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2392 if (C2->getAPIntValue() == 2)
2393 return DAG.getNode(ISD::UADDO, SDLoc(N), N->getVTList(),
2394 N->getOperand(0), N->getOperand(0));
2399 SDValue DAGCombiner::visitSDIVREM(SDNode *N) {
2400 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::SDIV, ISD::SREM);
2401 if (Res.getNode()) return Res;
2406 SDValue DAGCombiner::visitUDIVREM(SDNode *N) {
2407 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::UDIV, ISD::UREM);
2408 if (Res.getNode()) return Res;
2413 /// SimplifyBinOpWithSameOpcodeHands - If this is a binary operator with
2414 /// two operands of the same opcode, try to simplify it.
2415 SDValue DAGCombiner::SimplifyBinOpWithSameOpcodeHands(SDNode *N) {
2416 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
2417 EVT VT = N0.getValueType();
2418 assert(N0.getOpcode() == N1.getOpcode() && "Bad input!");
2420 // Bail early if none of these transforms apply.
2421 if (N0.getNode()->getNumOperands() == 0) return SDValue();
2423 // For each of OP in AND/OR/XOR:
2424 // fold (OP (zext x), (zext y)) -> (zext (OP x, y))
2425 // fold (OP (sext x), (sext y)) -> (sext (OP x, y))
2426 // fold (OP (aext x), (aext y)) -> (aext (OP x, y))
2427 // fold (OP (trunc x), (trunc y)) -> (trunc (OP x, y)) (if trunc isn't free)
2429 // do not sink logical op inside of a vector extend, since it may combine
2431 EVT Op0VT = N0.getOperand(0).getValueType();
2432 if ((N0.getOpcode() == ISD::ZERO_EXTEND ||
2433 N0.getOpcode() == ISD::SIGN_EXTEND ||
2434 // Avoid infinite looping with PromoteIntBinOp.
2435 (N0.getOpcode() == ISD::ANY_EXTEND &&
2436 (!LegalTypes || TLI.isTypeDesirableForOp(N->getOpcode(), Op0VT))) ||
2437 (N0.getOpcode() == ISD::TRUNCATE &&
2438 (!TLI.isZExtFree(VT, Op0VT) ||
2439 !TLI.isTruncateFree(Op0VT, VT)) &&
2440 TLI.isTypeLegal(Op0VT))) &&
2442 Op0VT == N1.getOperand(0).getValueType() &&
2443 (!LegalOperations || TLI.isOperationLegal(N->getOpcode(), Op0VT))) {
2444 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0),
2445 N0.getOperand(0).getValueType(),
2446 N0.getOperand(0), N1.getOperand(0));
2447 AddToWorkList(ORNode.getNode());
2448 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, ORNode);
2451 // For each of OP in SHL/SRL/SRA/AND...
2452 // fold (and (OP x, z), (OP y, z)) -> (OP (and x, y), z)
2453 // fold (or (OP x, z), (OP y, z)) -> (OP (or x, y), z)
2454 // fold (xor (OP x, z), (OP y, z)) -> (OP (xor x, y), z)
2455 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL ||
2456 N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::AND) &&
2457 N0.getOperand(1) == N1.getOperand(1)) {
2458 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0),
2459 N0.getOperand(0).getValueType(),
2460 N0.getOperand(0), N1.getOperand(0));
2461 AddToWorkList(ORNode.getNode());
2462 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
2463 ORNode, N0.getOperand(1));
2466 // Simplify xor/and/or (bitcast(A), bitcast(B)) -> bitcast(op (A,B))
2467 // Only perform this optimization after type legalization and before
2468 // LegalizeVectorOprs. LegalizeVectorOprs promotes vector operations by
2469 // adding bitcasts. For example (xor v4i32) is promoted to (v2i64), and
2470 // we don't want to undo this promotion.
2471 // We also handle SCALAR_TO_VECTOR because xor/or/and operations are cheaper
2473 if ((N0.getOpcode() == ISD::BITCAST ||
2474 N0.getOpcode() == ISD::SCALAR_TO_VECTOR) &&
2475 Level == AfterLegalizeTypes) {
2476 SDValue In0 = N0.getOperand(0);
2477 SDValue In1 = N1.getOperand(0);
2478 EVT In0Ty = In0.getValueType();
2479 EVT In1Ty = In1.getValueType();
2481 // If both incoming values are integers, and the original types are the
2483 if (In0Ty.isInteger() && In1Ty.isInteger() && In0Ty == In1Ty) {
2484 SDValue Op = DAG.getNode(N->getOpcode(), DL, In0Ty, In0, In1);
2485 SDValue BC = DAG.getNode(N0.getOpcode(), DL, VT, Op);
2486 AddToWorkList(Op.getNode());
2491 // Xor/and/or are indifferent to the swizzle operation (shuffle of one value).
2492 // Simplify xor/and/or (shuff(A), shuff(B)) -> shuff(op (A,B))
2493 // If both shuffles use the same mask, and both shuffle within a single
2494 // vector, then it is worthwhile to move the swizzle after the operation.
2495 // The type-legalizer generates this pattern when loading illegal
2496 // vector types from memory. In many cases this allows additional shuffle
2498 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
2499 N0.getOperand(1).getOpcode() == ISD::UNDEF &&
2500 N1.getOperand(1).getOpcode() == ISD::UNDEF) {
2501 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(N0);
2502 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(N1);
2504 assert(N0.getOperand(0).getValueType() == N1.getOperand(1).getValueType() &&
2505 "Inputs to shuffles are not the same type");
2507 unsigned NumElts = VT.getVectorNumElements();
2509 // Check that both shuffles use the same mask. The masks are known to be of
2510 // the same length because the result vector type is the same.
2511 bool SameMask = true;
2512 for (unsigned i = 0; i != NumElts; ++i) {
2513 int Idx0 = SVN0->getMaskElt(i);
2514 int Idx1 = SVN1->getMaskElt(i);
2522 SDValue Op = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
2523 N0.getOperand(0), N1.getOperand(0));
2524 AddToWorkList(Op.getNode());
2525 return DAG.getVectorShuffle(VT, SDLoc(N), Op,
2526 DAG.getUNDEF(VT), &SVN0->getMask()[0]);
2533 SDValue DAGCombiner::visitAND(SDNode *N) {
2534 SDValue N0 = N->getOperand(0);
2535 SDValue N1 = N->getOperand(1);
2536 SDValue LL, LR, RL, RR, CC0, CC1;
2537 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2538 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2539 EVT VT = N1.getValueType();
2540 unsigned BitWidth = VT.getScalarType().getSizeInBits();
2543 if (VT.isVector()) {
2544 SDValue FoldedVOp = SimplifyVBinOp(N);
2545 if (FoldedVOp.getNode()) return FoldedVOp;
2547 // fold (and x, 0) -> 0, vector edition
2548 if (ISD::isBuildVectorAllZeros(N0.getNode()))
2550 if (ISD::isBuildVectorAllZeros(N1.getNode()))
2553 // fold (and x, -1) -> x, vector edition
2554 if (ISD::isBuildVectorAllOnes(N0.getNode()))
2556 if (ISD::isBuildVectorAllOnes(N1.getNode()))
2560 // fold (and x, undef) -> 0
2561 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2562 return DAG.getConstant(0, VT);
2563 // fold (and c1, c2) -> c1&c2
2565 return DAG.FoldConstantArithmetic(ISD::AND, VT, N0C, N1C);
2566 // canonicalize constant to RHS
2568 return DAG.getNode(ISD::AND, SDLoc(N), VT, N1, N0);
2569 // fold (and x, -1) -> x
2570 if (N1C && N1C->isAllOnesValue())
2572 // if (and x, c) is known to be zero, return 0
2573 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
2574 APInt::getAllOnesValue(BitWidth)))
2575 return DAG.getConstant(0, VT);
2577 SDValue RAND = ReassociateOps(ISD::AND, SDLoc(N), N0, N1);
2578 if (RAND.getNode() != 0)
2580 // fold (and (or x, C), D) -> D if (C & D) == D
2581 if (N1C && N0.getOpcode() == ISD::OR)
2582 if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
2583 if ((ORI->getAPIntValue() & N1C->getAPIntValue()) == N1C->getAPIntValue())
2585 // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits.
2586 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
2587 SDValue N0Op0 = N0.getOperand(0);
2588 APInt Mask = ~N1C->getAPIntValue();
2589 Mask = Mask.trunc(N0Op0.getValueSizeInBits());
2590 if (DAG.MaskedValueIsZero(N0Op0, Mask)) {
2591 SDValue Zext = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N),
2592 N0.getValueType(), N0Op0);
2594 // Replace uses of the AND with uses of the Zero extend node.
2597 // We actually want to replace all uses of the any_extend with the
2598 // zero_extend, to avoid duplicating things. This will later cause this
2599 // AND to be folded.
2600 CombineTo(N0.getNode(), Zext);
2601 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2604 // similarly fold (and (X (load ([non_ext|any_ext|zero_ext] V))), c) ->
2605 // (X (load ([non_ext|zero_ext] V))) if 'and' only clears top bits which must
2606 // already be zero by virtue of the width of the base type of the load.
2608 // the 'X' node here can either be nothing or an extract_vector_elt to catch
2610 if ((N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
2611 N0.getOperand(0).getOpcode() == ISD::LOAD) ||
2612 N0.getOpcode() == ISD::LOAD) {
2613 LoadSDNode *Load = cast<LoadSDNode>( (N0.getOpcode() == ISD::LOAD) ?
2614 N0 : N0.getOperand(0) );
2616 // Get the constant (if applicable) the zero'th operand is being ANDed with.
2617 // This can be a pure constant or a vector splat, in which case we treat the
2618 // vector as a scalar and use the splat value.
2619 APInt Constant = APInt::getNullValue(1);
2620 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
2621 Constant = C->getAPIntValue();
2622 } else if (BuildVectorSDNode *Vector = dyn_cast<BuildVectorSDNode>(N1)) {
2623 APInt SplatValue, SplatUndef;
2624 unsigned SplatBitSize;
2626 bool IsSplat = Vector->isConstantSplat(SplatValue, SplatUndef,
2627 SplatBitSize, HasAnyUndefs);
2629 // Undef bits can contribute to a possible optimisation if set, so
2631 SplatValue |= SplatUndef;
2633 // The splat value may be something like "0x00FFFFFF", which means 0 for
2634 // the first vector value and FF for the rest, repeating. We need a mask
2635 // that will apply equally to all members of the vector, so AND all the
2636 // lanes of the constant together.
2637 EVT VT = Vector->getValueType(0);
2638 unsigned BitWidth = VT.getVectorElementType().getSizeInBits();
2640 // If the splat value has been compressed to a bitlength lower
2641 // than the size of the vector lane, we need to re-expand it to
2643 if (BitWidth > SplatBitSize)
2644 for (SplatValue = SplatValue.zextOrTrunc(BitWidth);
2645 SplatBitSize < BitWidth;
2646 SplatBitSize = SplatBitSize * 2)
2647 SplatValue |= SplatValue.shl(SplatBitSize);
2649 Constant = APInt::getAllOnesValue(BitWidth);
2650 for (unsigned i = 0, n = SplatBitSize/BitWidth; i < n; ++i)
2651 Constant &= SplatValue.lshr(i*BitWidth).zextOrTrunc(BitWidth);
2655 // If we want to change an EXTLOAD to a ZEXTLOAD, ensure a ZEXTLOAD is
2656 // actually legal and isn't going to get expanded, else this is a false
2658 bool CanZextLoadProfitably = TLI.isLoadExtLegal(ISD::ZEXTLOAD,
2659 Load->getMemoryVT());
2661 // Resize the constant to the same size as the original memory access before
2662 // extension. If it is still the AllOnesValue then this AND is completely
2665 Constant.zextOrTrunc(Load->getMemoryVT().getScalarType().getSizeInBits());
2668 switch (Load->getExtensionType()) {
2669 default: B = false; break;
2670 case ISD::EXTLOAD: B = CanZextLoadProfitably; break;
2672 case ISD::NON_EXTLOAD: B = true; break;
2675 if (B && Constant.isAllOnesValue()) {
2676 // If the load type was an EXTLOAD, convert to ZEXTLOAD in order to
2677 // preserve semantics once we get rid of the AND.
2678 SDValue NewLoad(Load, 0);
2679 if (Load->getExtensionType() == ISD::EXTLOAD) {
2680 NewLoad = DAG.getLoad(Load->getAddressingMode(), ISD::ZEXTLOAD,
2681 Load->getValueType(0), SDLoc(Load),
2682 Load->getChain(), Load->getBasePtr(),
2683 Load->getOffset(), Load->getMemoryVT(),
2684 Load->getMemOperand());
2685 // Replace uses of the EXTLOAD with the new ZEXTLOAD.
2686 if (Load->getNumValues() == 3) {
2687 // PRE/POST_INC loads have 3 values.
2688 SDValue To[] = { NewLoad.getValue(0), NewLoad.getValue(1),
2689 NewLoad.getValue(2) };
2690 CombineTo(Load, To, 3, true);
2692 CombineTo(Load, NewLoad.getValue(0), NewLoad.getValue(1));
2696 // Fold the AND away, taking care not to fold to the old load node if we
2698 CombineTo(N, (N0.getNode() == Load) ? NewLoad : N0);
2700 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2703 // fold (and (setcc x), (setcc y)) -> (setcc (and x, y))
2704 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
2705 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
2706 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
2708 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
2709 LL.getValueType().isInteger()) {
2710 // fold (and (seteq X, 0), (seteq Y, 0)) -> (seteq (or X, Y), 0)
2711 if (cast<ConstantSDNode>(LR)->isNullValue() && Op1 == ISD::SETEQ) {
2712 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(N0),
2713 LR.getValueType(), LL, RL);
2714 AddToWorkList(ORNode.getNode());
2715 return DAG.getSetCC(SDLoc(N), VT, ORNode, LR, Op1);
2717 // fold (and (seteq X, -1), (seteq Y, -1)) -> (seteq (and X, Y), -1)
2718 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETEQ) {
2719 SDValue ANDNode = DAG.getNode(ISD::AND, SDLoc(N0),
2720 LR.getValueType(), LL, RL);
2721 AddToWorkList(ANDNode.getNode());
2722 return DAG.getSetCC(SDLoc(N), VT, ANDNode, LR, Op1);
2724 // fold (and (setgt X, -1), (setgt Y, -1)) -> (setgt (or X, Y), -1)
2725 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETGT) {
2726 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(N0),
2727 LR.getValueType(), LL, RL);
2728 AddToWorkList(ORNode.getNode());
2729 return DAG.getSetCC(SDLoc(N), VT, ORNode, LR, Op1);
2732 // Simplify (and (setne X, 0), (setne X, -1)) -> (setuge (add X, 1), 2)
2733 if (LL == RL && isa<ConstantSDNode>(LR) && isa<ConstantSDNode>(RR) &&
2734 Op0 == Op1 && LL.getValueType().isInteger() &&
2735 Op0 == ISD::SETNE && ((cast<ConstantSDNode>(LR)->isNullValue() &&
2736 cast<ConstantSDNode>(RR)->isAllOnesValue()) ||
2737 (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
2738 cast<ConstantSDNode>(RR)->isNullValue()))) {
2739 SDValue ADDNode = DAG.getNode(ISD::ADD, SDLoc(N0), LL.getValueType(),
2740 LL, DAG.getConstant(1, LL.getValueType()));
2741 AddToWorkList(ADDNode.getNode());
2742 return DAG.getSetCC(SDLoc(N), VT, ADDNode,
2743 DAG.getConstant(2, LL.getValueType()), ISD::SETUGE);
2745 // canonicalize equivalent to ll == rl
2746 if (LL == RR && LR == RL) {
2747 Op1 = ISD::getSetCCSwappedOperands(Op1);
2750 if (LL == RL && LR == RR) {
2751 bool isInteger = LL.getValueType().isInteger();
2752 ISD::CondCode Result = ISD::getSetCCAndOperation(Op0, Op1, isInteger);
2753 if (Result != ISD::SETCC_INVALID &&
2754 (!LegalOperations ||
2755 (TLI.isCondCodeLegal(Result, LL.getSimpleValueType()) &&
2756 TLI.isOperationLegal(ISD::SETCC,
2757 getSetCCResultType(N0.getSimpleValueType())))))
2758 return DAG.getSetCC(SDLoc(N), N0.getValueType(),
2763 // Simplify: (and (op x...), (op y...)) -> (op (and x, y))
2764 if (N0.getOpcode() == N1.getOpcode()) {
2765 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
2766 if (Tmp.getNode()) return Tmp;
2769 // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1)
2770 // fold (and (sra)) -> (and (srl)) when possible.
2771 if (!VT.isVector() &&
2772 SimplifyDemandedBits(SDValue(N, 0)))
2773 return SDValue(N, 0);
2775 // fold (zext_inreg (extload x)) -> (zextload x)
2776 if (ISD::isEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode())) {
2777 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2778 EVT MemVT = LN0->getMemoryVT();
2779 // If we zero all the possible extended bits, then we can turn this into
2780 // a zextload if we are running before legalize or the operation is legal.
2781 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
2782 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
2783 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
2784 ((!LegalOperations && !LN0->isVolatile()) ||
2785 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
2786 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT,
2787 LN0->getChain(), LN0->getBasePtr(),
2788 MemVT, LN0->getMemOperand());
2790 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
2791 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2794 // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use
2795 if (ISD::isSEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
2797 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2798 EVT MemVT = LN0->getMemoryVT();
2799 // If we zero all the possible extended bits, then we can turn this into
2800 // a zextload if we are running before legalize or the operation is legal.
2801 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
2802 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
2803 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
2804 ((!LegalOperations && !LN0->isVolatile()) ||
2805 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
2806 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT,
2807 LN0->getChain(), LN0->getBasePtr(),
2808 MemVT, LN0->getMemOperand());
2810 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
2811 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2815 // fold (and (load x), 255) -> (zextload x, i8)
2816 // fold (and (extload x, i16), 255) -> (zextload x, i8)
2817 // fold (and (any_ext (extload x, i16)), 255) -> (zextload x, i8)
2818 if (N1C && (N0.getOpcode() == ISD::LOAD ||
2819 (N0.getOpcode() == ISD::ANY_EXTEND &&
2820 N0.getOperand(0).getOpcode() == ISD::LOAD))) {
2821 bool HasAnyExt = N0.getOpcode() == ISD::ANY_EXTEND;
2822 LoadSDNode *LN0 = HasAnyExt
2823 ? cast<LoadSDNode>(N0.getOperand(0))
2824 : cast<LoadSDNode>(N0);
2825 if (LN0->getExtensionType() != ISD::SEXTLOAD &&
2826 LN0->isUnindexed() && N0.hasOneUse() && SDValue(LN0, 0).hasOneUse()) {
2827 uint32_t ActiveBits = N1C->getAPIntValue().getActiveBits();
2828 if (ActiveBits > 0 && APIntOps::isMask(ActiveBits, N1C->getAPIntValue())){
2829 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits);
2830 EVT LoadedVT = LN0->getMemoryVT();
2832 if (ExtVT == LoadedVT &&
2833 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) {
2834 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
2837 DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy,
2838 LN0->getChain(), LN0->getBasePtr(), ExtVT,
2839 LN0->getMemOperand());
2841 CombineTo(LN0, NewLoad, NewLoad.getValue(1));
2842 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2845 // Do not change the width of a volatile load.
2846 // Do not generate loads of non-round integer types since these can
2847 // be expensive (and would be wrong if the type is not byte sized).
2848 if (!LN0->isVolatile() && LoadedVT.bitsGT(ExtVT) && ExtVT.isRound() &&
2849 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) {
2850 EVT PtrType = LN0->getOperand(1).getValueType();
2852 unsigned Alignment = LN0->getAlignment();
2853 SDValue NewPtr = LN0->getBasePtr();
2855 // For big endian targets, we need to add an offset to the pointer
2856 // to load the correct bytes. For little endian systems, we merely
2857 // need to read fewer bytes from the same pointer.
2858 if (TLI.isBigEndian()) {
2859 unsigned LVTStoreBytes = LoadedVT.getStoreSize();
2860 unsigned EVTStoreBytes = ExtVT.getStoreSize();
2861 unsigned PtrOff = LVTStoreBytes - EVTStoreBytes;
2862 NewPtr = DAG.getNode(ISD::ADD, SDLoc(LN0), PtrType,
2863 NewPtr, DAG.getConstant(PtrOff, PtrType));
2864 Alignment = MinAlign(Alignment, PtrOff);
2867 AddToWorkList(NewPtr.getNode());
2869 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
2871 DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy,
2872 LN0->getChain(), NewPtr,
2873 LN0->getPointerInfo(),
2874 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
2875 Alignment, LN0->getTBAAInfo());
2877 CombineTo(LN0, Load, Load.getValue(1));
2878 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2884 if (N0.getOpcode() == ISD::ADD && N1.getOpcode() == ISD::SRL &&
2885 VT.getSizeInBits() <= 64) {
2886 if (ConstantSDNode *ADDI = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
2887 APInt ADDC = ADDI->getAPIntValue();
2888 if (!TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
2889 // Look for (and (add x, c1), (lshr y, c2)). If C1 wasn't a legal
2890 // immediate for an add, but it is legal if its top c2 bits are set,
2891 // transform the ADD so the immediate doesn't need to be materialized
2893 if (ConstantSDNode *SRLI = dyn_cast<ConstantSDNode>(N1.getOperand(1))) {
2894 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(),
2895 SRLI->getZExtValue());
2896 if (DAG.MaskedValueIsZero(N0.getOperand(1), Mask)) {
2898 if (TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
2900 DAG.getNode(ISD::ADD, SDLoc(N0), VT,
2901 N0.getOperand(0), DAG.getConstant(ADDC, VT));
2902 CombineTo(N0.getNode(), NewAdd);
2903 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2911 // fold (and (or (srl N, 8), (shl N, 8)), 0xffff) -> (srl (bswap N), const)
2912 if (N1C && N1C->getAPIntValue() == 0xffff && N0.getOpcode() == ISD::OR) {
2913 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
2914 N0.getOperand(1), false);
2915 if (BSwap.getNode())
2922 /// MatchBSwapHWord - Match (a >> 8) | (a << 8) as (bswap a) >> 16
2924 SDValue DAGCombiner::MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
2925 bool DemandHighBits) {
2926 if (!LegalOperations)
2929 EVT VT = N->getValueType(0);
2930 if (VT != MVT::i64 && VT != MVT::i32 && VT != MVT::i16)
2932 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
2935 // Recognize (and (shl a, 8), 0xff), (and (srl a, 8), 0xff00)
2936 bool LookPassAnd0 = false;
2937 bool LookPassAnd1 = false;
2938 if (N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode() == ISD::SRL)
2940 if (N1.getOpcode() == ISD::AND && N1.getOperand(0).getOpcode() == ISD::SHL)
2942 if (N0.getOpcode() == ISD::AND) {
2943 if (!N0.getNode()->hasOneUse())
2945 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
2946 if (!N01C || N01C->getZExtValue() != 0xFF00)
2948 N0 = N0.getOperand(0);
2949 LookPassAnd0 = true;
2952 if (N1.getOpcode() == ISD::AND) {
2953 if (!N1.getNode()->hasOneUse())
2955 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
2956 if (!N11C || N11C->getZExtValue() != 0xFF)
2958 N1 = N1.getOperand(0);
2959 LookPassAnd1 = true;
2962 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
2964 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
2966 if (!N0.getNode()->hasOneUse() ||
2967 !N1.getNode()->hasOneUse())
2970 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
2971 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
2974 if (N01C->getZExtValue() != 8 || N11C->getZExtValue() != 8)
2977 // Look for (shl (and a, 0xff), 8), (srl (and a, 0xff00), 8)
2978 SDValue N00 = N0->getOperand(0);
2979 if (!LookPassAnd0 && N00.getOpcode() == ISD::AND) {
2980 if (!N00.getNode()->hasOneUse())
2982 ConstantSDNode *N001C = dyn_cast<ConstantSDNode>(N00.getOperand(1));
2983 if (!N001C || N001C->getZExtValue() != 0xFF)
2985 N00 = N00.getOperand(0);
2986 LookPassAnd0 = true;
2989 SDValue N10 = N1->getOperand(0);
2990 if (!LookPassAnd1 && N10.getOpcode() == ISD::AND) {
2991 if (!N10.getNode()->hasOneUse())
2993 ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N10.getOperand(1));
2994 if (!N101C || N101C->getZExtValue() != 0xFF00)
2996 N10 = N10.getOperand(0);
2997 LookPassAnd1 = true;
3003 // Make sure everything beyond the low halfword gets set to zero since the SRL
3004 // 16 will clear the top bits.
3005 unsigned OpSizeInBits = VT.getSizeInBits();
3006 if (DemandHighBits && OpSizeInBits > 16) {
3007 // If the left-shift isn't masked out then the only way this is a bswap is
3008 // if all bits beyond the low 8 are 0. In that case the entire pattern
3009 // reduces to a left shift anyway: leave it for other parts of the combiner.
3013 // However, if the right shift isn't masked out then it might be because
3014 // it's not needed. See if we can spot that too.
3015 if (!LookPassAnd1 &&
3016 !DAG.MaskedValueIsZero(
3017 N10, APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - 16)))
3021 SDValue Res = DAG.getNode(ISD::BSWAP, SDLoc(N), VT, N00);
3022 if (OpSizeInBits > 16)
3023 Res = DAG.getNode(ISD::SRL, SDLoc(N), VT, Res,
3024 DAG.getConstant(OpSizeInBits-16, getShiftAmountTy(VT)));
3028 /// isBSwapHWordElement - Return true if the specified node is an element
3029 /// that makes up a 32-bit packed halfword byteswap. i.e.
3030 /// ((x&0xff)<<8)|((x&0xff00)>>8)|((x&0x00ff0000)<<8)|((x&0xff000000)>>8)
3031 static bool isBSwapHWordElement(SDValue N, SmallVectorImpl<SDNode *> &Parts) {
3032 if (!N.getNode()->hasOneUse())
3035 unsigned Opc = N.getOpcode();
3036 if (Opc != ISD::AND && Opc != ISD::SHL && Opc != ISD::SRL)
3039 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3044 switch (N1C->getZExtValue()) {
3047 case 0xFF: Num = 0; break;
3048 case 0xFF00: Num = 1; break;
3049 case 0xFF0000: Num = 2; break;
3050 case 0xFF000000: Num = 3; break;
3053 // Look for (x & 0xff) << 8 as well as ((x << 8) & 0xff00).
3054 SDValue N0 = N.getOperand(0);
3055 if (Opc == ISD::AND) {
3056 if (Num == 0 || Num == 2) {
3058 // (x >> 8) & 0xff0000
3059 if (N0.getOpcode() != ISD::SRL)
3061 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3062 if (!C || C->getZExtValue() != 8)
3065 // (x << 8) & 0xff00
3066 // (x << 8) & 0xff000000
3067 if (N0.getOpcode() != ISD::SHL)
3069 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3070 if (!C || C->getZExtValue() != 8)
3073 } else if (Opc == ISD::SHL) {
3075 // (x & 0xff0000) << 8
3076 if (Num != 0 && Num != 2)
3078 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3079 if (!C || C->getZExtValue() != 8)
3081 } else { // Opc == ISD::SRL
3082 // (x & 0xff00) >> 8
3083 // (x & 0xff000000) >> 8
3084 if (Num != 1 && Num != 3)
3086 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3087 if (!C || C->getZExtValue() != 8)
3094 Parts[Num] = N0.getOperand(0).getNode();
3098 /// MatchBSwapHWord - Match a 32-bit packed halfword bswap. That is
3099 /// ((x&0xff)<<8)|((x&0xff00)>>8)|((x&0x00ff0000)<<8)|((x&0xff000000)>>8)
3100 /// => (rotl (bswap x), 16)
3101 SDValue DAGCombiner::MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1) {
3102 if (!LegalOperations)
3105 EVT VT = N->getValueType(0);
3108 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
3111 SmallVector<SDNode*,4> Parts(4, (SDNode*)0);
3113 // (or (or (and), (and)), (or (and), (and)))
3114 // (or (or (or (and), (and)), (and)), (and))
3115 if (N0.getOpcode() != ISD::OR)
3117 SDValue N00 = N0.getOperand(0);
3118 SDValue N01 = N0.getOperand(1);
3120 if (N1.getOpcode() == ISD::OR &&
3121 N00.getNumOperands() == 2 && N01.getNumOperands() == 2) {
3122 // (or (or (and), (and)), (or (and), (and)))
3123 SDValue N000 = N00.getOperand(0);
3124 if (!isBSwapHWordElement(N000, Parts))
3127 SDValue N001 = N00.getOperand(1);
3128 if (!isBSwapHWordElement(N001, Parts))
3130 SDValue N010 = N01.getOperand(0);
3131 if (!isBSwapHWordElement(N010, Parts))
3133 SDValue N011 = N01.getOperand(1);
3134 if (!isBSwapHWordElement(N011, Parts))
3137 // (or (or (or (and), (and)), (and)), (and))
3138 if (!isBSwapHWordElement(N1, Parts))
3140 if (!isBSwapHWordElement(N01, Parts))
3142 if (N00.getOpcode() != ISD::OR)
3144 SDValue N000 = N00.getOperand(0);
3145 if (!isBSwapHWordElement(N000, Parts))
3147 SDValue N001 = N00.getOperand(1);
3148 if (!isBSwapHWordElement(N001, Parts))
3152 // Make sure the parts are all coming from the same node.
3153 if (Parts[0] != Parts[1] || Parts[0] != Parts[2] || Parts[0] != Parts[3])
3156 SDValue BSwap = DAG.getNode(ISD::BSWAP, SDLoc(N), VT,
3157 SDValue(Parts[0],0));
3159 // Result of the bswap should be rotated by 16. If it's not legal, then
3160 // do (x << 16) | (x >> 16).
3161 SDValue ShAmt = DAG.getConstant(16, getShiftAmountTy(VT));
3162 if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT))
3163 return DAG.getNode(ISD::ROTL, SDLoc(N), VT, BSwap, ShAmt);
3164 if (TLI.isOperationLegalOrCustom(ISD::ROTR, VT))
3165 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, BSwap, ShAmt);
3166 return DAG.getNode(ISD::OR, SDLoc(N), VT,
3167 DAG.getNode(ISD::SHL, SDLoc(N), VT, BSwap, ShAmt),
3168 DAG.getNode(ISD::SRL, SDLoc(N), VT, BSwap, ShAmt));
3171 SDValue DAGCombiner::visitOR(SDNode *N) {
3172 SDValue N0 = N->getOperand(0);
3173 SDValue N1 = N->getOperand(1);
3174 SDValue LL, LR, RL, RR, CC0, CC1;
3175 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3176 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3177 EVT VT = N1.getValueType();
3180 if (VT.isVector()) {
3181 SDValue FoldedVOp = SimplifyVBinOp(N);
3182 if (FoldedVOp.getNode()) return FoldedVOp;
3184 // fold (or x, 0) -> x, vector edition
3185 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3187 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3190 // fold (or x, -1) -> -1, vector edition
3191 if (ISD::isBuildVectorAllOnes(N0.getNode()))
3193 if (ISD::isBuildVectorAllOnes(N1.getNode()))
3197 // fold (or x, undef) -> -1
3198 if (!LegalOperations &&
3199 (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)) {
3200 EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT;
3201 return DAG.getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()), VT);
3203 // fold (or c1, c2) -> c1|c2
3205 return DAG.FoldConstantArithmetic(ISD::OR, VT, N0C, N1C);
3206 // canonicalize constant to RHS
3208 return DAG.getNode(ISD::OR, SDLoc(N), VT, N1, N0);
3209 // fold (or x, 0) -> x
3210 if (N1C && N1C->isNullValue())
3212 // fold (or x, -1) -> -1
3213 if (N1C && N1C->isAllOnesValue())
3215 // fold (or x, c) -> c iff (x & ~c) == 0
3216 if (N1C && DAG.MaskedValueIsZero(N0, ~N1C->getAPIntValue()))
3219 // Recognize halfword bswaps as (bswap + rotl 16) or (bswap + shl 16)
3220 SDValue BSwap = MatchBSwapHWord(N, N0, N1);
3221 if (BSwap.getNode() != 0)
3223 BSwap = MatchBSwapHWordLow(N, N0, N1);
3224 if (BSwap.getNode() != 0)
3228 SDValue ROR = ReassociateOps(ISD::OR, SDLoc(N), N0, N1);
3229 if (ROR.getNode() != 0)
3231 // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2)
3232 // iff (c1 & c2) == 0.
3233 if (N1C && N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
3234 isa<ConstantSDNode>(N0.getOperand(1))) {
3235 ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1));
3236 if ((C1->getAPIntValue() & N1C->getAPIntValue()) != 0) {
3237 SDValue COR = DAG.FoldConstantArithmetic(ISD::OR, VT, N1C, C1);
3240 return DAG.getNode(ISD::AND, SDLoc(N), VT,
3241 DAG.getNode(ISD::OR, SDLoc(N0), VT,
3242 N0.getOperand(0), N1), COR);
3245 // fold (or (setcc x), (setcc y)) -> (setcc (or x, y))
3246 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
3247 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
3248 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
3250 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
3251 LL.getValueType().isInteger()) {
3252 // fold (or (setne X, 0), (setne Y, 0)) -> (setne (or X, Y), 0)
3253 // fold (or (setlt X, 0), (setlt Y, 0)) -> (setne (or X, Y), 0)
3254 if (cast<ConstantSDNode>(LR)->isNullValue() &&
3255 (Op1 == ISD::SETNE || Op1 == ISD::SETLT)) {
3256 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(LR),
3257 LR.getValueType(), LL, RL);
3258 AddToWorkList(ORNode.getNode());
3259 return DAG.getSetCC(SDLoc(N), VT, ORNode, LR, Op1);
3261 // fold (or (setne X, -1), (setne Y, -1)) -> (setne (and X, Y), -1)
3262 // fold (or (setgt X, -1), (setgt Y -1)) -> (setgt (and X, Y), -1)
3263 if (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
3264 (Op1 == ISD::SETNE || Op1 == ISD::SETGT)) {
3265 SDValue ANDNode = DAG.getNode(ISD::AND, SDLoc(LR),
3266 LR.getValueType(), LL, RL);
3267 AddToWorkList(ANDNode.getNode());
3268 return DAG.getSetCC(SDLoc(N), VT, ANDNode, LR, Op1);
3271 // canonicalize equivalent to ll == rl
3272 if (LL == RR && LR == RL) {
3273 Op1 = ISD::getSetCCSwappedOperands(Op1);
3276 if (LL == RL && LR == RR) {
3277 bool isInteger = LL.getValueType().isInteger();
3278 ISD::CondCode Result = ISD::getSetCCOrOperation(Op0, Op1, isInteger);
3279 if (Result != ISD::SETCC_INVALID &&
3280 (!LegalOperations ||
3281 (TLI.isCondCodeLegal(Result, LL.getSimpleValueType()) &&
3282 TLI.isOperationLegal(ISD::SETCC,
3283 getSetCCResultType(N0.getValueType())))))
3284 return DAG.getSetCC(SDLoc(N), N0.getValueType(),
3289 // Simplify: (or (op x...), (op y...)) -> (op (or x, y))
3290 if (N0.getOpcode() == N1.getOpcode()) {
3291 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3292 if (Tmp.getNode()) return Tmp;
3295 // (or (and X, C1), (and Y, C2)) -> (and (or X, Y), C3) if possible.
3296 if (N0.getOpcode() == ISD::AND &&
3297 N1.getOpcode() == ISD::AND &&
3298 N0.getOperand(1).getOpcode() == ISD::Constant &&
3299 N1.getOperand(1).getOpcode() == ISD::Constant &&
3300 // Don't increase # computations.
3301 (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) {
3302 // We can only do this xform if we know that bits from X that are set in C2
3303 // but not in C1 are already zero. Likewise for Y.
3304 const APInt &LHSMask =
3305 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
3306 const APInt &RHSMask =
3307 cast<ConstantSDNode>(N1.getOperand(1))->getAPIntValue();
3309 if (DAG.MaskedValueIsZero(N0.getOperand(0), RHSMask&~LHSMask) &&
3310 DAG.MaskedValueIsZero(N1.getOperand(0), LHSMask&~RHSMask)) {
3311 SDValue X = DAG.getNode(ISD::OR, SDLoc(N0), VT,
3312 N0.getOperand(0), N1.getOperand(0));
3313 return DAG.getNode(ISD::AND, SDLoc(N), VT, X,
3314 DAG.getConstant(LHSMask | RHSMask, VT));
3318 // See if this is some rotate idiom.
3319 if (SDNode *Rot = MatchRotate(N0, N1, SDLoc(N)))
3320 return SDValue(Rot, 0);
3322 // Simplify the operands using demanded-bits information.
3323 if (!VT.isVector() &&
3324 SimplifyDemandedBits(SDValue(N, 0)))
3325 return SDValue(N, 0);
3330 /// MatchRotateHalf - Match "(X shl/srl V1) & V2" where V2 may not be present.
3331 static bool MatchRotateHalf(SDValue Op, SDValue &Shift, SDValue &Mask) {
3332 if (Op.getOpcode() == ISD::AND) {
3333 if (isa<ConstantSDNode>(Op.getOperand(1))) {
3334 Mask = Op.getOperand(1);
3335 Op = Op.getOperand(0);
3341 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SHL) {
3349 // Return true if we can prove that, whenever Neg and Pos are both in the
3350 // range [0, OpSize), Neg == (Pos == 0 ? 0 : OpSize - Pos). This means that
3351 // for two opposing shifts shift1 and shift2 and a value X with OpBits bits:
3353 // (or (shift1 X, Neg), (shift2 X, Pos))
3355 // reduces to a rotate in direction shift2 by Pos and a rotate in direction
3356 // shift1 by Neg. The range [0, OpSize) means that we only need to consider
3357 // shift amounts with defined behavior.
3358 static bool matchRotateSub(SDValue Pos, SDValue Neg, unsigned OpSize) {
3359 // If OpSize is a power of 2 then:
3361 // (a) (Pos == 0 ? 0 : OpSize - Pos) == (OpSize - Pos) & (OpSize - 1)
3362 // (b) Neg == Neg & (OpSize - 1) whenever Neg is in [0, OpSize).
3364 // So if OpSize is a power of 2 and Neg is (and Neg', OpSize-1), we check
3365 // for the stronger condition:
3367 // Neg & (OpSize - 1) == (OpSize - Pos) & (OpSize - 1) [A]
3369 // for all Neg and Pos. Since Neg & (OpSize - 1) == Neg' & (OpSize - 1)
3370 // we can just replace Neg with Neg' for the rest of the function.
3372 // In other cases we check for the even stronger condition:
3374 // Neg == OpSize - Pos [B]
3376 // for all Neg and Pos. Note that the (or ...) then invokes undefined
3377 // behavior if Pos == 0 (and consequently Neg == OpSize).
3379 // We could actually use [A] whenever OpSize is a power of 2, but the
3380 // only extra cases that it would match are those uninteresting ones
3381 // where Neg and Pos are never in range at the same time. E.g. for
3382 // OpSize == 32, using [A] would allow a Neg of the form (sub 64, Pos)
3383 // as well as (sub 32, Pos), but:
3385 // (or (shift1 X, (sub 64, Pos)), (shift2 X, Pos))
3387 // always invokes undefined behavior for 32-bit X.
3389 // Below, Mask == OpSize - 1 when using [A] and is all-ones otherwise.
3390 unsigned LoBits = 0;
3391 if (Neg.getOpcode() == ISD::AND &&
3392 isPowerOf2_64(OpSize) &&
3393 Neg.getOperand(1).getOpcode() == ISD::Constant &&
3394 cast<ConstantSDNode>(Neg.getOperand(1))->getAPIntValue() == OpSize - 1) {
3395 Neg = Neg.getOperand(0);
3396 LoBits = Log2_64(OpSize);
3399 // Check whether Neg has the form (sub NegC, NegOp1) for some NegC and NegOp1.
3400 if (Neg.getOpcode() != ISD::SUB)
3402 ConstantSDNode *NegC = dyn_cast<ConstantSDNode>(Neg.getOperand(0));
3405 SDValue NegOp1 = Neg.getOperand(1);
3407 // The condition we need is now:
3409 // (NegC - NegOp1) & Mask == (OpSize - Pos) & Mask
3411 // If NegOp1 == Pos then we need:
3413 // OpSize & Mask == NegC & Mask
3415 // (because "x & Mask" is a truncation and distributes through subtraction).
3418 Width = NegC->getAPIntValue();
3419 // Check for cases where Pos has the form (add NegOp1, PosC) for some PosC.
3420 // Then the condition we want to prove becomes:
3422 // (NegC - NegOp1) & Mask == (OpSize - (NegOp1 + PosC)) & Mask
3424 // which, again because "x & Mask" is a truncation, becomes:
3426 // NegC & Mask == (OpSize - PosC) & Mask
3427 // OpSize & Mask == (NegC + PosC) & Mask
3428 else if (Pos.getOpcode() == ISD::ADD &&
3429 Pos.getOperand(0) == NegOp1 &&
3430 Pos.getOperand(1).getOpcode() == ISD::Constant)
3431 Width = (cast<ConstantSDNode>(Pos.getOperand(1))->getAPIntValue() +
3432 NegC->getAPIntValue());
3436 // Now we just need to check that OpSize & Mask == Width & Mask.
3438 return Width.getLoBits(LoBits) == 0;
3439 return Width == OpSize;
3442 // A subroutine of MatchRotate used once we have found an OR of two opposite
3443 // shifts of Shifted. If Neg == <operand size> - Pos then the OR reduces
3444 // to both (PosOpcode Shifted, Pos) and (NegOpcode Shifted, Neg), with the
3445 // former being preferred if supported. InnerPos and InnerNeg are Pos and
3446 // Neg with outer conversions stripped away.
3447 SDNode *DAGCombiner::MatchRotatePosNeg(SDValue Shifted, SDValue Pos,
3448 SDValue Neg, SDValue InnerPos,
3449 SDValue InnerNeg, unsigned PosOpcode,
3450 unsigned NegOpcode, SDLoc DL) {
3451 // fold (or (shl x, (*ext y)),
3452 // (srl x, (*ext (sub 32, y)))) ->
3453 // (rotl x, y) or (rotr x, (sub 32, y))
3455 // fold (or (shl x, (*ext (sub 32, y))),
3456 // (srl x, (*ext y))) ->
3457 // (rotr x, y) or (rotl x, (sub 32, y))
3458 EVT VT = Shifted.getValueType();
3459 if (matchRotateSub(InnerPos, InnerNeg, VT.getSizeInBits())) {
3460 bool HasPos = TLI.isOperationLegalOrCustom(PosOpcode, VT);
3461 return DAG.getNode(HasPos ? PosOpcode : NegOpcode, DL, VT, Shifted,
3462 HasPos ? Pos : Neg).getNode();
3465 // fold (or (shl (*ext x), (*ext y)),
3466 // (srl (*ext x), (*ext (sub 32, y)))) ->
3467 // (*ext (rotl x, y)) or (*ext (rotr x, (sub 32, y)))
3469 // fold (or (shl (*ext x), (*ext (sub 32, y))),
3470 // (srl (*ext x), (*ext y))) ->
3471 // (*ext (rotr x, y)) or (*ext (rotl x, (sub 32, y)))
3472 if (Shifted.getOpcode() == ISD::ZERO_EXTEND ||
3473 Shifted.getOpcode() == ISD::ANY_EXTEND) {
3474 SDValue InnerShifted = Shifted.getOperand(0);
3475 EVT InnerVT = InnerShifted.getValueType();
3476 bool HasPosInner = TLI.isOperationLegalOrCustom(PosOpcode, InnerVT);
3477 if (HasPosInner || TLI.isOperationLegalOrCustom(NegOpcode, InnerVT)) {
3478 if (matchRotateSub(InnerPos, InnerNeg, InnerVT.getSizeInBits())) {
3479 SDValue V = DAG.getNode(HasPosInner ? PosOpcode : NegOpcode, DL,
3480 InnerVT, InnerShifted, HasPosInner ? Pos : Neg);
3481 return DAG.getNode(Shifted.getOpcode(), DL, VT, V).getNode();
3489 // MatchRotate - Handle an 'or' of two operands. If this is one of the many
3490 // idioms for rotate, and if the target supports rotation instructions, generate
3492 SDNode *DAGCombiner::MatchRotate(SDValue LHS, SDValue RHS, SDLoc DL) {
3493 // Must be a legal type. Expanded 'n promoted things won't work with rotates.
3494 EVT VT = LHS.getValueType();
3495 if (!TLI.isTypeLegal(VT)) return 0;
3497 // The target must have at least one rotate flavor.
3498 bool HasROTL = TLI.isOperationLegalOrCustom(ISD::ROTL, VT);
3499 bool HasROTR = TLI.isOperationLegalOrCustom(ISD::ROTR, VT);
3500 if (!HasROTL && !HasROTR) return 0;
3502 // Match "(X shl/srl V1) & V2" where V2 may not be present.
3503 SDValue LHSShift; // The shift.
3504 SDValue LHSMask; // AND value if any.
3505 if (!MatchRotateHalf(LHS, LHSShift, LHSMask))
3506 return 0; // Not part of a rotate.
3508 SDValue RHSShift; // The shift.
3509 SDValue RHSMask; // AND value if any.
3510 if (!MatchRotateHalf(RHS, RHSShift, RHSMask))
3511 return 0; // Not part of a rotate.
3513 if (LHSShift.getOperand(0) != RHSShift.getOperand(0))
3514 return 0; // Not shifting the same value.
3516 if (LHSShift.getOpcode() == RHSShift.getOpcode())
3517 return 0; // Shifts must disagree.
3519 // Canonicalize shl to left side in a shl/srl pair.
3520 if (RHSShift.getOpcode() == ISD::SHL) {
3521 std::swap(LHS, RHS);
3522 std::swap(LHSShift, RHSShift);
3523 std::swap(LHSMask , RHSMask );
3526 unsigned OpSizeInBits = VT.getSizeInBits();
3527 SDValue LHSShiftArg = LHSShift.getOperand(0);
3528 SDValue LHSShiftAmt = LHSShift.getOperand(1);
3529 SDValue RHSShiftArg = RHSShift.getOperand(0);
3530 SDValue RHSShiftAmt = RHSShift.getOperand(1);
3532 // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1)
3533 // fold (or (shl x, C1), (srl x, C2)) -> (rotr x, C2)
3534 if (LHSShiftAmt.getOpcode() == ISD::Constant &&
3535 RHSShiftAmt.getOpcode() == ISD::Constant) {
3536 uint64_t LShVal = cast<ConstantSDNode>(LHSShiftAmt)->getZExtValue();
3537 uint64_t RShVal = cast<ConstantSDNode>(RHSShiftAmt)->getZExtValue();
3538 if ((LShVal + RShVal) != OpSizeInBits)
3541 SDValue Rot = DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT,
3542 LHSShiftArg, HasROTL ? LHSShiftAmt : RHSShiftAmt);
3544 // If there is an AND of either shifted operand, apply it to the result.
3545 if (LHSMask.getNode() || RHSMask.getNode()) {
3546 APInt Mask = APInt::getAllOnesValue(OpSizeInBits);
3548 if (LHSMask.getNode()) {
3549 APInt RHSBits = APInt::getLowBitsSet(OpSizeInBits, LShVal);
3550 Mask &= cast<ConstantSDNode>(LHSMask)->getAPIntValue() | RHSBits;
3552 if (RHSMask.getNode()) {
3553 APInt LHSBits = APInt::getHighBitsSet(OpSizeInBits, RShVal);
3554 Mask &= cast<ConstantSDNode>(RHSMask)->getAPIntValue() | LHSBits;
3557 Rot = DAG.getNode(ISD::AND, DL, VT, Rot, DAG.getConstant(Mask, VT));
3560 return Rot.getNode();
3563 // If there is a mask here, and we have a variable shift, we can't be sure
3564 // that we're masking out the right stuff.
3565 if (LHSMask.getNode() || RHSMask.getNode())
3568 // If the shift amount is sign/zext/any-extended just peel it off.
3569 SDValue LExtOp0 = LHSShiftAmt;
3570 SDValue RExtOp0 = RHSShiftAmt;
3571 if ((LHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3572 LHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3573 LHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3574 LHSShiftAmt.getOpcode() == ISD::TRUNCATE) &&
3575 (RHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3576 RHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3577 RHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3578 RHSShiftAmt.getOpcode() == ISD::TRUNCATE)) {
3579 LExtOp0 = LHSShiftAmt.getOperand(0);
3580 RExtOp0 = RHSShiftAmt.getOperand(0);
3583 SDNode *TryL = MatchRotatePosNeg(LHSShiftArg, LHSShiftAmt, RHSShiftAmt,
3584 LExtOp0, RExtOp0, ISD::ROTL, ISD::ROTR, DL);
3588 SDNode *TryR = MatchRotatePosNeg(RHSShiftArg, RHSShiftAmt, LHSShiftAmt,
3589 RExtOp0, LExtOp0, ISD::ROTR, ISD::ROTL, DL);
3596 SDValue DAGCombiner::visitXOR(SDNode *N) {
3597 SDValue N0 = N->getOperand(0);
3598 SDValue N1 = N->getOperand(1);
3599 SDValue LHS, RHS, CC;
3600 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3601 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3602 EVT VT = N0.getValueType();
3605 if (VT.isVector()) {
3606 SDValue FoldedVOp = SimplifyVBinOp(N);
3607 if (FoldedVOp.getNode()) return FoldedVOp;
3609 // fold (xor x, 0) -> x, vector edition
3610 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3612 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3616 // fold (xor undef, undef) -> 0. This is a common idiom (misuse).
3617 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
3618 return DAG.getConstant(0, VT);
3619 // fold (xor x, undef) -> undef
3620 if (N0.getOpcode() == ISD::UNDEF)
3622 if (N1.getOpcode() == ISD::UNDEF)
3624 // fold (xor c1, c2) -> c1^c2
3626 return DAG.FoldConstantArithmetic(ISD::XOR, VT, N0C, N1C);
3627 // canonicalize constant to RHS
3629 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0);
3630 // fold (xor x, 0) -> x
3631 if (N1C && N1C->isNullValue())
3634 SDValue RXOR = ReassociateOps(ISD::XOR, SDLoc(N), N0, N1);
3635 if (RXOR.getNode() != 0)
3638 // fold !(x cc y) -> (x !cc y)
3639 if (N1C && N1C->getAPIntValue() == 1 && isSetCCEquivalent(N0, LHS, RHS, CC)) {
3640 bool isInt = LHS.getValueType().isInteger();
3641 ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
3644 if (!LegalOperations ||
3645 TLI.isCondCodeLegal(NotCC, LHS.getSimpleValueType())) {
3646 switch (N0.getOpcode()) {
3648 llvm_unreachable("Unhandled SetCC Equivalent!");
3650 return DAG.getSetCC(SDLoc(N), VT, LHS, RHS, NotCC);
3651 case ISD::SELECT_CC:
3652 return DAG.getSelectCC(SDLoc(N), LHS, RHS, N0.getOperand(2),
3653 N0.getOperand(3), NotCC);
3658 // fold (not (zext (setcc x, y))) -> (zext (not (setcc x, y)))
3659 if (N1C && N1C->getAPIntValue() == 1 && N0.getOpcode() == ISD::ZERO_EXTEND &&
3660 N0.getNode()->hasOneUse() &&
3661 isSetCCEquivalent(N0.getOperand(0), LHS, RHS, CC)){
3662 SDValue V = N0.getOperand(0);
3663 V = DAG.getNode(ISD::XOR, SDLoc(N0), V.getValueType(), V,
3664 DAG.getConstant(1, V.getValueType()));
3665 AddToWorkList(V.getNode());
3666 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, V);
3669 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are setcc
3670 if (N1C && N1C->getAPIntValue() == 1 && VT == MVT::i1 &&
3671 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3672 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3673 if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) {
3674 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3675 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS
3676 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS
3677 AddToWorkList(LHS.getNode()); AddToWorkList(RHS.getNode());
3678 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS);
3681 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are constants
3682 if (N1C && N1C->isAllOnesValue() &&
3683 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3684 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3685 if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) {
3686 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3687 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS
3688 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS
3689 AddToWorkList(LHS.getNode()); AddToWorkList(RHS.getNode());
3690 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS);
3693 // fold (xor (and x, y), y) -> (and (not x), y)
3694 if (N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
3695 N0->getOperand(1) == N1) {
3696 SDValue X = N0->getOperand(0);
3697 SDValue NotX = DAG.getNOT(SDLoc(X), X, VT);
3698 AddToWorkList(NotX.getNode());
3699 return DAG.getNode(ISD::AND, SDLoc(N), VT, NotX, N1);
3701 // fold (xor (xor x, c1), c2) -> (xor x, (xor c1, c2))
3702 if (N1C && N0.getOpcode() == ISD::XOR) {
3703 ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0));
3704 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3706 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N0.getOperand(1),
3707 DAG.getConstant(N1C->getAPIntValue() ^
3708 N00C->getAPIntValue(), VT));
3710 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N0.getOperand(0),
3711 DAG.getConstant(N1C->getAPIntValue() ^
3712 N01C->getAPIntValue(), VT));
3714 // fold (xor x, x) -> 0
3716 return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations, LegalTypes);
3718 // Simplify: xor (op x...), (op y...) -> (op (xor x, y))
3719 if (N0.getOpcode() == N1.getOpcode()) {
3720 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3721 if (Tmp.getNode()) return Tmp;
3724 // Simplify the expression using non-local knowledge.
3725 if (!VT.isVector() &&
3726 SimplifyDemandedBits(SDValue(N, 0)))
3727 return SDValue(N, 0);
3732 /// visitShiftByConstant - Handle transforms common to the three shifts, when
3733 /// the shift amount is a constant.
3734 SDValue DAGCombiner::visitShiftByConstant(SDNode *N, unsigned Amt) {
3735 SDNode *LHS = N->getOperand(0).getNode();
3736 if (!LHS->hasOneUse()) return SDValue();
3738 // We want to pull some binops through shifts, so that we have (and (shift))
3739 // instead of (shift (and)), likewise for add, or, xor, etc. This sort of
3740 // thing happens with address calculations, so it's important to canonicalize
3742 bool HighBitSet = false; // Can we transform this if the high bit is set?
3744 switch (LHS->getOpcode()) {
3745 default: return SDValue();
3748 HighBitSet = false; // We can only transform sra if the high bit is clear.
3751 HighBitSet = true; // We can only transform sra if the high bit is set.
3754 if (N->getOpcode() != ISD::SHL)
3755 return SDValue(); // only shl(add) not sr[al](add).
3756 HighBitSet = false; // We can only transform sra if the high bit is clear.
3760 // We require the RHS of the binop to be a constant as well.
3761 ConstantSDNode *BinOpCst = dyn_cast<ConstantSDNode>(LHS->getOperand(1));
3762 if (!BinOpCst) return SDValue();
3764 // FIXME: disable this unless the input to the binop is a shift by a constant.
3765 // If it is not a shift, it pessimizes some common cases like:
3767 // void foo(int *X, int i) { X[i & 1235] = 1; }
3768 // int bar(int *X, int i) { return X[i & 255]; }
3769 SDNode *BinOpLHSVal = LHS->getOperand(0).getNode();
3770 if ((BinOpLHSVal->getOpcode() != ISD::SHL &&
3771 BinOpLHSVal->getOpcode() != ISD::SRA &&
3772 BinOpLHSVal->getOpcode() != ISD::SRL) ||
3773 !isa<ConstantSDNode>(BinOpLHSVal->getOperand(1)))
3776 EVT VT = N->getValueType(0);
3778 // If this is a signed shift right, and the high bit is modified by the
3779 // logical operation, do not perform the transformation. The highBitSet
3780 // boolean indicates the value of the high bit of the constant which would
3781 // cause it to be modified for this operation.
3782 if (N->getOpcode() == ISD::SRA) {
3783 bool BinOpRHSSignSet = BinOpCst->getAPIntValue().isNegative();
3784 if (BinOpRHSSignSet != HighBitSet)
3788 // Fold the constants, shifting the binop RHS by the shift amount.
3789 SDValue NewRHS = DAG.getNode(N->getOpcode(), SDLoc(LHS->getOperand(1)),
3791 LHS->getOperand(1), N->getOperand(1));
3793 // Create the new shift.
3794 SDValue NewShift = DAG.getNode(N->getOpcode(),
3795 SDLoc(LHS->getOperand(0)),
3796 VT, LHS->getOperand(0), N->getOperand(1));
3798 // Create the new binop.
3799 return DAG.getNode(LHS->getOpcode(), SDLoc(N), VT, NewShift, NewRHS);
3802 SDValue DAGCombiner::visitSHL(SDNode *N) {
3803 SDValue N0 = N->getOperand(0);
3804 SDValue N1 = N->getOperand(1);
3805 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3806 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3807 EVT VT = N0.getValueType();
3808 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
3811 if (VT.isVector()) {
3812 SDValue FoldedVOp = SimplifyVBinOp(N);
3813 if (FoldedVOp.getNode()) return FoldedVOp;
3816 // fold (shl c1, c2) -> c1<<c2
3818 return DAG.FoldConstantArithmetic(ISD::SHL, VT, N0C, N1C);
3819 // fold (shl 0, x) -> 0
3820 if (N0C && N0C->isNullValue())
3822 // fold (shl x, c >= size(x)) -> undef
3823 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
3824 return DAG.getUNDEF(VT);
3825 // fold (shl x, 0) -> x
3826 if (N1C && N1C->isNullValue())
3828 // fold (shl undef, x) -> 0
3829 if (N0.getOpcode() == ISD::UNDEF)
3830 return DAG.getConstant(0, VT);
3831 // if (shl x, c) is known to be zero, return 0
3832 if (DAG.MaskedValueIsZero(SDValue(N, 0),
3833 APInt::getAllOnesValue(OpSizeInBits)))
3834 return DAG.getConstant(0, VT);
3835 // fold (shl x, (trunc (and y, c))) -> (shl x, (and (trunc y), (trunc c))).
3836 if (N1.getOpcode() == ISD::TRUNCATE &&
3837 N1.getOperand(0).getOpcode() == ISD::AND &&
3838 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
3839 SDValue N101 = N1.getOperand(0).getOperand(1);
3840 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
3841 EVT TruncVT = N1.getValueType();
3842 SDValue N100 = N1.getOperand(0).getOperand(0);
3843 APInt TruncC = N101C->getAPIntValue();
3844 TruncC = TruncC.trunc(TruncVT.getSizeInBits());
3845 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0,
3846 DAG.getNode(ISD::AND, SDLoc(N), TruncVT,
3847 DAG.getNode(ISD::TRUNCATE,
3850 DAG.getConstant(TruncC, TruncVT)));
3854 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
3855 return SDValue(N, 0);
3857 // fold (shl (shl x, c1), c2) -> 0 or (shl x, (add c1, c2))
3858 if (N1C && N0.getOpcode() == ISD::SHL &&
3859 N0.getOperand(1).getOpcode() == ISD::Constant) {
3860 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
3861 uint64_t c2 = N1C->getZExtValue();
3862 if (c1 + c2 >= OpSizeInBits)
3863 return DAG.getConstant(0, VT);
3864 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0),
3865 DAG.getConstant(c1 + c2, N1.getValueType()));
3868 // fold (shl (ext (shl x, c1)), c2) -> (ext (shl x, (add c1, c2)))
3869 // For this to be valid, the second form must not preserve any of the bits
3870 // that are shifted out by the inner shift in the first form. This means
3871 // the outer shift size must be >= the number of bits added by the ext.
3872 // As a corollary, we don't care what kind of ext it is.
3873 if (N1C && (N0.getOpcode() == ISD::ZERO_EXTEND ||
3874 N0.getOpcode() == ISD::ANY_EXTEND ||
3875 N0.getOpcode() == ISD::SIGN_EXTEND) &&
3876 N0.getOperand(0).getOpcode() == ISD::SHL &&
3877 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
3879 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
3880 uint64_t c2 = N1C->getZExtValue();
3881 EVT InnerShiftVT = N0.getOperand(0).getValueType();
3882 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
3883 if (c2 >= OpSizeInBits - InnerShiftSize) {
3884 if (c1 + c2 >= OpSizeInBits)
3885 return DAG.getConstant(0, VT);
3886 return DAG.getNode(ISD::SHL, SDLoc(N0), VT,
3887 DAG.getNode(N0.getOpcode(), SDLoc(N0), VT,
3888 N0.getOperand(0)->getOperand(0)),
3889 DAG.getConstant(c1 + c2, N1.getValueType()));
3893 // fold (shl (zext (srl x, C)), C) -> (zext (shl (srl x, C), C))
3894 // Only fold this if the inner zext has no other uses to avoid increasing
3895 // the total number of instructions.
3896 if (N1C && N0.getOpcode() == ISD::ZERO_EXTEND && N0.hasOneUse() &&
3897 N0.getOperand(0).getOpcode() == ISD::SRL &&
3898 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
3900 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
3901 if (c1 < VT.getSizeInBits()) {
3902 uint64_t c2 = N1C->getZExtValue();
3904 SDValue NewOp0 = N0.getOperand(0);
3905 EVT CountVT = NewOp0.getOperand(1).getValueType();
3906 SDValue NewSHL = DAG.getNode(ISD::SHL, SDLoc(N), NewOp0.getValueType(),
3907 NewOp0, DAG.getConstant(c2, CountVT));
3908 AddToWorkList(NewSHL.getNode());
3909 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N0), VT, NewSHL);
3914 // fold (shl (srl x, c1), c2) -> (and (shl x, (sub c2, c1), MASK) or
3915 // (and (srl x, (sub c1, c2), MASK)
3916 // Only fold this if the inner shift has no other uses -- if it does, folding
3917 // this will increase the total number of instructions.
3918 if (N1C && N0.getOpcode() == ISD::SRL && N0.hasOneUse() &&
3919 N0.getOperand(1).getOpcode() == ISD::Constant) {
3920 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
3921 if (c1 < VT.getSizeInBits()) {
3922 uint64_t c2 = N1C->getZExtValue();
3923 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(),
3924 VT.getSizeInBits() - c1);
3927 Mask = Mask.shl(c2-c1);
3928 Shift = DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0),
3929 DAG.getConstant(c2-c1, N1.getValueType()));
3931 Mask = Mask.lshr(c1-c2);
3932 Shift = DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0),
3933 DAG.getConstant(c1-c2, N1.getValueType()));
3935 return DAG.getNode(ISD::AND, SDLoc(N0), VT, Shift,
3936 DAG.getConstant(Mask, VT));
3939 // fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1))
3940 if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1)) {
3941 SDValue HiBitsMask =
3942 DAG.getConstant(APInt::getHighBitsSet(VT.getSizeInBits(),
3943 VT.getSizeInBits() -
3944 N1C->getZExtValue()),
3946 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0.getOperand(0),
3951 SDValue NewSHL = visitShiftByConstant(N, N1C->getZExtValue());
3952 if (NewSHL.getNode())
3959 SDValue DAGCombiner::visitSRA(SDNode *N) {
3960 SDValue N0 = N->getOperand(0);
3961 SDValue N1 = N->getOperand(1);
3962 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3963 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3964 EVT VT = N0.getValueType();
3965 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
3968 if (VT.isVector()) {
3969 SDValue FoldedVOp = SimplifyVBinOp(N);
3970 if (FoldedVOp.getNode()) return FoldedVOp;
3973 // fold (sra c1, c2) -> (sra c1, c2)
3975 return DAG.FoldConstantArithmetic(ISD::SRA, VT, N0C, N1C);
3976 // fold (sra 0, x) -> 0
3977 if (N0C && N0C->isNullValue())
3979 // fold (sra -1, x) -> -1
3980 if (N0C && N0C->isAllOnesValue())
3982 // fold (sra x, (setge c, size(x))) -> undef
3983 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
3984 return DAG.getUNDEF(VT);
3985 // fold (sra x, 0) -> x
3986 if (N1C && N1C->isNullValue())
3988 // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports
3990 if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) {
3991 unsigned LowBits = OpSizeInBits - (unsigned)N1C->getZExtValue();
3992 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), LowBits);
3994 ExtVT = EVT::getVectorVT(*DAG.getContext(),
3995 ExtVT, VT.getVectorNumElements());
3996 if ((!LegalOperations ||
3997 TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, ExtVT)))
3998 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
3999 N0.getOperand(0), DAG.getValueType(ExtVT));
4002 // fold (sra (sra x, c1), c2) -> (sra x, (add c1, c2))
4003 if (N1C && N0.getOpcode() == ISD::SRA) {
4004 if (ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
4005 unsigned Sum = N1C->getZExtValue() + C1->getZExtValue();
4006 if (Sum >= OpSizeInBits) Sum = OpSizeInBits-1;
4007 return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0.getOperand(0),
4008 DAG.getConstant(Sum, N1C->getValueType(0)));
4012 // fold (sra (shl X, m), (sub result_size, n))
4013 // -> (sign_extend (trunc (shl X, (sub (sub result_size, n), m)))) for
4014 // result_size - n != m.
4015 // If truncate is free for the target sext(shl) is likely to result in better
4017 if (N0.getOpcode() == ISD::SHL) {
4018 // Get the two constanst of the shifts, CN0 = m, CN = n.
4019 const ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
4021 // Determine what the truncate's result bitsize and type would be.
4023 EVT::getIntegerVT(*DAG.getContext(),
4024 OpSizeInBits - N1C->getZExtValue());
4025 // Determine the residual right-shift amount.
4026 signed ShiftAmt = N1C->getZExtValue() - N01C->getZExtValue();
4028 // If the shift is not a no-op (in which case this should be just a sign
4029 // extend already), the truncated to type is legal, sign_extend is legal
4030 // on that type, and the truncate to that type is both legal and free,
4031 // perform the transform.
4032 if ((ShiftAmt > 0) &&
4033 TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND, TruncVT) &&
4034 TLI.isOperationLegalOrCustom(ISD::TRUNCATE, VT) &&
4035 TLI.isTruncateFree(VT, TruncVT)) {
4037 SDValue Amt = DAG.getConstant(ShiftAmt,
4038 getShiftAmountTy(N0.getOperand(0).getValueType()));
4039 SDValue Shift = DAG.getNode(ISD::SRL, SDLoc(N0), VT,
4040 N0.getOperand(0), Amt);
4041 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), TruncVT,
4043 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N),
4044 N->getValueType(0), Trunc);
4049 // fold (sra x, (trunc (and y, c))) -> (sra x, (and (trunc y), (trunc c))).
4050 if (N1.getOpcode() == ISD::TRUNCATE &&
4051 N1.getOperand(0).getOpcode() == ISD::AND &&
4052 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
4053 SDValue N101 = N1.getOperand(0).getOperand(1);
4054 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
4055 EVT TruncVT = N1.getValueType();
4056 SDValue N100 = N1.getOperand(0).getOperand(0);
4057 APInt TruncC = N101C->getAPIntValue();
4058 TruncC = TruncC.trunc(TruncVT.getScalarType().getSizeInBits());
4059 return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0,
4060 DAG.getNode(ISD::AND, SDLoc(N),
4062 DAG.getNode(ISD::TRUNCATE,
4065 DAG.getConstant(TruncC, TruncVT)));
4069 // fold (sra (trunc (sr x, c1)), c2) -> (trunc (sra x, c1+c2))
4070 // if c1 is equal to the number of bits the trunc removes
4071 if (N0.getOpcode() == ISD::TRUNCATE &&
4072 (N0.getOperand(0).getOpcode() == ISD::SRL ||
4073 N0.getOperand(0).getOpcode() == ISD::SRA) &&
4074 N0.getOperand(0).hasOneUse() &&
4075 N0.getOperand(0).getOperand(1).hasOneUse() &&
4076 N1C && isa<ConstantSDNode>(N0.getOperand(0).getOperand(1))) {
4077 EVT LargeVT = N0.getOperand(0).getValueType();
4078 ConstantSDNode *LargeShiftAmt =
4079 cast<ConstantSDNode>(N0.getOperand(0).getOperand(1));
4081 if (LargeVT.getScalarType().getSizeInBits() - OpSizeInBits ==
4082 LargeShiftAmt->getZExtValue()) {
4084 DAG.getConstant(LargeShiftAmt->getZExtValue() + N1C->getZExtValue(),
4085 getShiftAmountTy(N0.getOperand(0).getOperand(0).getValueType()));
4086 SDValue SRA = DAG.getNode(ISD::SRA, SDLoc(N), LargeVT,
4087 N0.getOperand(0).getOperand(0), Amt);
4088 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, SRA);
4092 // Simplify, based on bits shifted out of the LHS.
4093 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4094 return SDValue(N, 0);
4097 // If the sign bit is known to be zero, switch this to a SRL.
4098 if (DAG.SignBitIsZero(N0))
4099 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, N1);
4102 SDValue NewSRA = visitShiftByConstant(N, N1C->getZExtValue());
4103 if (NewSRA.getNode())
4110 SDValue DAGCombiner::visitSRL(SDNode *N) {
4111 SDValue N0 = N->getOperand(0);
4112 SDValue N1 = N->getOperand(1);
4113 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4114 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4115 EVT VT = N0.getValueType();
4116 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
4119 if (VT.isVector()) {
4120 SDValue FoldedVOp = SimplifyVBinOp(N);
4121 if (FoldedVOp.getNode()) return FoldedVOp;
4124 // fold (srl c1, c2) -> c1 >>u c2
4126 return DAG.FoldConstantArithmetic(ISD::SRL, VT, N0C, N1C);
4127 // fold (srl 0, x) -> 0
4128 if (N0C && N0C->isNullValue())
4130 // fold (srl x, c >= size(x)) -> undef
4131 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4132 return DAG.getUNDEF(VT);
4133 // fold (srl x, 0) -> x
4134 if (N1C && N1C->isNullValue())
4136 // if (srl x, c) is known to be zero, return 0
4137 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
4138 APInt::getAllOnesValue(OpSizeInBits)))
4139 return DAG.getConstant(0, VT);
4141 // fold (srl (srl x, c1), c2) -> 0 or (srl x, (add c1, c2))
4142 if (N1C && N0.getOpcode() == ISD::SRL &&
4143 N0.getOperand(1).getOpcode() == ISD::Constant) {
4144 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
4145 uint64_t c2 = N1C->getZExtValue();
4146 if (c1 + c2 >= OpSizeInBits)
4147 return DAG.getConstant(0, VT);
4148 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0),
4149 DAG.getConstant(c1 + c2, N1.getValueType()));
4152 // fold (srl (trunc (srl x, c1)), c2) -> 0 or (trunc (srl x, (add c1, c2)))
4153 if (N1C && N0.getOpcode() == ISD::TRUNCATE &&
4154 N0.getOperand(0).getOpcode() == ISD::SRL &&
4155 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
4157 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
4158 uint64_t c2 = N1C->getZExtValue();
4159 EVT InnerShiftVT = N0.getOperand(0).getValueType();
4160 EVT ShiftCountVT = N0.getOperand(0)->getOperand(1).getValueType();
4161 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
4162 // This is only valid if the OpSizeInBits + c1 = size of inner shift.
4163 if (c1 + OpSizeInBits == InnerShiftSize) {
4164 if (c1 + c2 >= InnerShiftSize)
4165 return DAG.getConstant(0, VT);
4166 return DAG.getNode(ISD::TRUNCATE, SDLoc(N0), VT,
4167 DAG.getNode(ISD::SRL, SDLoc(N0), InnerShiftVT,
4168 N0.getOperand(0)->getOperand(0),
4169 DAG.getConstant(c1 + c2, ShiftCountVT)));
4173 // fold (srl (shl x, c), c) -> (and x, cst2)
4174 if (N1C && N0.getOpcode() == ISD::SHL && N0.getOperand(1) == N1 &&
4175 N0.getValueSizeInBits() <= 64) {
4176 uint64_t ShAmt = N1C->getZExtValue()+64-N0.getValueSizeInBits();
4177 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0.getOperand(0),
4178 DAG.getConstant(~0ULL >> ShAmt, VT));
4181 // fold (srl (anyextend x), c) -> (and (anyextend (srl x, c)), mask)
4182 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
4183 // Shifting in all undef bits?
4184 EVT SmallVT = N0.getOperand(0).getValueType();
4185 if (N1C->getZExtValue() >= SmallVT.getSizeInBits())
4186 return DAG.getUNDEF(VT);
4188 if (!LegalTypes || TLI.isTypeDesirableForOp(ISD::SRL, SmallVT)) {
4189 uint64_t ShiftAmt = N1C->getZExtValue();
4190 SDValue SmallShift = DAG.getNode(ISD::SRL, SDLoc(N0), SmallVT,
4192 DAG.getConstant(ShiftAmt, getShiftAmountTy(SmallVT)));
4193 AddToWorkList(SmallShift.getNode());
4194 APInt Mask = APInt::getAllOnesValue(VT.getSizeInBits()).lshr(ShiftAmt);
4195 return DAG.getNode(ISD::AND, SDLoc(N), VT,
4196 DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, SmallShift),
4197 DAG.getConstant(Mask, VT));
4201 // fold (srl (sra X, Y), 31) -> (srl X, 31). This srl only looks at the sign
4202 // bit, which is unmodified by sra.
4203 if (N1C && N1C->getZExtValue() + 1 == VT.getSizeInBits()) {
4204 if (N0.getOpcode() == ISD::SRA)
4205 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0), N1);
4208 // fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit).
4209 if (N1C && N0.getOpcode() == ISD::CTLZ &&
4210 N1C->getAPIntValue() == Log2_32(VT.getSizeInBits())) {
4211 APInt KnownZero, KnownOne;
4212 DAG.ComputeMaskedBits(N0.getOperand(0), KnownZero, KnownOne);
4214 // If any of the input bits are KnownOne, then the input couldn't be all
4215 // zeros, thus the result of the srl will always be zero.
4216 if (KnownOne.getBoolValue()) return DAG.getConstant(0, VT);
4218 // If all of the bits input the to ctlz node are known to be zero, then
4219 // the result of the ctlz is "32" and the result of the shift is one.
4220 APInt UnknownBits = ~KnownZero;
4221 if (UnknownBits == 0) return DAG.getConstant(1, VT);
4223 // Otherwise, check to see if there is exactly one bit input to the ctlz.
4224 if ((UnknownBits & (UnknownBits - 1)) == 0) {
4225 // Okay, we know that only that the single bit specified by UnknownBits
4226 // could be set on input to the CTLZ node. If this bit is set, the SRL
4227 // will return 0, if it is clear, it returns 1. Change the CTLZ/SRL pair
4228 // to an SRL/XOR pair, which is likely to simplify more.
4229 unsigned ShAmt = UnknownBits.countTrailingZeros();
4230 SDValue Op = N0.getOperand(0);
4233 Op = DAG.getNode(ISD::SRL, SDLoc(N0), VT, Op,
4234 DAG.getConstant(ShAmt, getShiftAmountTy(Op.getValueType())));
4235 AddToWorkList(Op.getNode());
4238 return DAG.getNode(ISD::XOR, SDLoc(N), VT,
4239 Op, DAG.getConstant(1, VT));
4243 // fold (srl x, (trunc (and y, c))) -> (srl x, (and (trunc y), (trunc c))).
4244 if (N1.getOpcode() == ISD::TRUNCATE &&
4245 N1.getOperand(0).getOpcode() == ISD::AND &&
4246 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
4247 SDValue N101 = N1.getOperand(0).getOperand(1);
4248 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
4249 EVT TruncVT = N1.getValueType();
4250 SDValue N100 = N1.getOperand(0).getOperand(0);
4251 APInt TruncC = N101C->getAPIntValue();
4252 TruncC = TruncC.trunc(TruncVT.getSizeInBits());
4253 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0,
4254 DAG.getNode(ISD::AND, SDLoc(N),
4256 DAG.getNode(ISD::TRUNCATE,
4259 DAG.getConstant(TruncC, TruncVT)));
4263 // fold operands of srl based on knowledge that the low bits are not
4265 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4266 return SDValue(N, 0);
4269 SDValue NewSRL = visitShiftByConstant(N, N1C->getZExtValue());
4270 if (NewSRL.getNode())
4274 // Attempt to convert a srl of a load into a narrower zero-extending load.
4275 SDValue NarrowLoad = ReduceLoadWidth(N);
4276 if (NarrowLoad.getNode())
4279 // Here is a common situation. We want to optimize:
4282 // %b = and i32 %a, 2
4283 // %c = srl i32 %b, 1
4284 // brcond i32 %c ...
4290 // %c = setcc eq %b, 0
4293 // However when after the source operand of SRL is optimized into AND, the SRL
4294 // itself may not be optimized further. Look for it and add the BRCOND into
4296 if (N->hasOneUse()) {
4297 SDNode *Use = *N->use_begin();
4298 if (Use->getOpcode() == ISD::BRCOND)
4300 else if (Use->getOpcode() == ISD::TRUNCATE && Use->hasOneUse()) {
4301 // Also look pass the truncate.
4302 Use = *Use->use_begin();
4303 if (Use->getOpcode() == ISD::BRCOND)
4311 SDValue DAGCombiner::visitCTLZ(SDNode *N) {
4312 SDValue N0 = N->getOperand(0);
4313 EVT VT = N->getValueType(0);
4315 // fold (ctlz c1) -> c2
4316 if (isa<ConstantSDNode>(N0))
4317 return DAG.getNode(ISD::CTLZ, SDLoc(N), VT, N0);
4321 SDValue DAGCombiner::visitCTLZ_ZERO_UNDEF(SDNode *N) {
4322 SDValue N0 = N->getOperand(0);
4323 EVT VT = N->getValueType(0);
4325 // fold (ctlz_zero_undef c1) -> c2
4326 if (isa<ConstantSDNode>(N0))
4327 return DAG.getNode(ISD::CTLZ_ZERO_UNDEF, SDLoc(N), VT, N0);
4331 SDValue DAGCombiner::visitCTTZ(SDNode *N) {
4332 SDValue N0 = N->getOperand(0);
4333 EVT VT = N->getValueType(0);
4335 // fold (cttz c1) -> c2
4336 if (isa<ConstantSDNode>(N0))
4337 return DAG.getNode(ISD::CTTZ, SDLoc(N), VT, N0);
4341 SDValue DAGCombiner::visitCTTZ_ZERO_UNDEF(SDNode *N) {
4342 SDValue N0 = N->getOperand(0);
4343 EVT VT = N->getValueType(0);
4345 // fold (cttz_zero_undef c1) -> c2
4346 if (isa<ConstantSDNode>(N0))
4347 return DAG.getNode(ISD::CTTZ_ZERO_UNDEF, SDLoc(N), VT, N0);
4351 SDValue DAGCombiner::visitCTPOP(SDNode *N) {
4352 SDValue N0 = N->getOperand(0);
4353 EVT VT = N->getValueType(0);
4355 // fold (ctpop c1) -> c2
4356 if (isa<ConstantSDNode>(N0))
4357 return DAG.getNode(ISD::CTPOP, SDLoc(N), VT, N0);
4361 SDValue DAGCombiner::visitSELECT(SDNode *N) {
4362 SDValue N0 = N->getOperand(0);
4363 SDValue N1 = N->getOperand(1);
4364 SDValue N2 = N->getOperand(2);
4365 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4366 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4367 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
4368 EVT VT = N->getValueType(0);
4369 EVT VT0 = N0.getValueType();
4371 // fold (select C, X, X) -> X
4374 // fold (select true, X, Y) -> X
4375 if (N0C && !N0C->isNullValue())
4377 // fold (select false, X, Y) -> Y
4378 if (N0C && N0C->isNullValue())
4380 // fold (select C, 1, X) -> (or C, X)
4381 if (VT == MVT::i1 && N1C && N1C->getAPIntValue() == 1)
4382 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N2);
4383 // fold (select C, 0, 1) -> (xor C, 1)
4384 if (VT.isInteger() &&
4387 TLI.getBooleanContents(false) ==
4388 TargetLowering::ZeroOrOneBooleanContent)) &&
4389 N1C && N2C && N1C->isNullValue() && N2C->getAPIntValue() == 1) {
4392 return DAG.getNode(ISD::XOR, SDLoc(N), VT0,
4393 N0, DAG.getConstant(1, VT0));
4394 XORNode = DAG.getNode(ISD::XOR, SDLoc(N0), VT0,
4395 N0, DAG.getConstant(1, VT0));
4396 AddToWorkList(XORNode.getNode());
4398 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, XORNode);
4399 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, XORNode);
4401 // fold (select C, 0, X) -> (and (not C), X)
4402 if (VT == VT0 && VT == MVT::i1 && N1C && N1C->isNullValue()) {
4403 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT);
4404 AddToWorkList(NOTNode.getNode());
4405 return DAG.getNode(ISD::AND, SDLoc(N), VT, NOTNode, N2);
4407 // fold (select C, X, 1) -> (or (not C), X)
4408 if (VT == VT0 && VT == MVT::i1 && N2C && N2C->getAPIntValue() == 1) {
4409 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT);
4410 AddToWorkList(NOTNode.getNode());
4411 return DAG.getNode(ISD::OR, SDLoc(N), VT, NOTNode, N1);
4413 // fold (select C, X, 0) -> (and C, X)
4414 if (VT == MVT::i1 && N2C && N2C->isNullValue())
4415 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, N1);
4416 // fold (select X, X, Y) -> (or X, Y)
4417 // fold (select X, 1, Y) -> (or X, Y)
4418 if (VT == MVT::i1 && (N0 == N1 || (N1C && N1C->getAPIntValue() == 1)))
4419 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N2);
4420 // fold (select X, Y, X) -> (and X, Y)
4421 // fold (select X, Y, 0) -> (and X, Y)
4422 if (VT == MVT::i1 && (N0 == N2 || (N2C && N2C->getAPIntValue() == 0)))
4423 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, N1);
4425 // If we can fold this based on the true/false value, do so.
4426 if (SimplifySelectOps(N, N1, N2))
4427 return SDValue(N, 0); // Don't revisit N.
4429 // fold selects based on a setcc into other things, such as min/max/abs
4430 if (N0.getOpcode() == ISD::SETCC) {
4432 // Check against MVT::Other for SELECT_CC, which is a workaround for targets
4433 // having to say they don't support SELECT_CC on every type the DAG knows
4434 // about, since there is no way to mark an opcode illegal at all value types
4435 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, MVT::Other) &&
4436 TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT))
4437 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT,
4438 N0.getOperand(0), N0.getOperand(1),
4439 N1, N2, N0.getOperand(2));
4440 return SimplifySelect(SDLoc(N), N0, N1, N2);
4447 std::pair<SDValue, SDValue> SplitVSETCC(const SDNode *N, SelectionDAG &DAG) {
4450 llvm::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0));
4452 // Split the inputs.
4453 SDValue Lo, Hi, LL, LH, RL, RH;
4454 llvm::tie(LL, LH) = DAG.SplitVectorOperand(N, 0);
4455 llvm::tie(RL, RH) = DAG.SplitVectorOperand(N, 1);
4457 Lo = DAG.getNode(N->getOpcode(), DL, LoVT, LL, RL, N->getOperand(2));
4458 Hi = DAG.getNode(N->getOpcode(), DL, HiVT, LH, RH, N->getOperand(2));
4460 return std::make_pair(Lo, Hi);
4463 SDValue DAGCombiner::visitVSELECT(SDNode *N) {
4464 SDValue N0 = N->getOperand(0);
4465 SDValue N1 = N->getOperand(1);
4466 SDValue N2 = N->getOperand(2);
4469 // Canonicalize integer abs.
4470 // vselect (setg[te] X, 0), X, -X ->
4471 // vselect (setgt X, -1), X, -X ->
4472 // vselect (setl[te] X, 0), -X, X ->
4473 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
4474 if (N0.getOpcode() == ISD::SETCC) {
4475 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
4476 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
4478 bool RHSIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
4480 if (((RHSIsAllZeros && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
4481 (ISD::isBuildVectorAllOnes(RHS.getNode()) && CC == ISD::SETGT)) &&
4482 N1 == LHS && N2.getOpcode() == ISD::SUB && N1 == N2.getOperand(1))
4483 isAbs = ISD::isBuildVectorAllZeros(N2.getOperand(0).getNode());
4484 else if ((RHSIsAllZeros && (CC == ISD::SETLT || CC == ISD::SETLE)) &&
4485 N2 == LHS && N1.getOpcode() == ISD::SUB && N2 == N1.getOperand(1))
4486 isAbs = ISD::isBuildVectorAllZeros(N1.getOperand(0).getNode());
4489 EVT VT = LHS.getValueType();
4490 SDValue Shift = DAG.getNode(
4491 ISD::SRA, DL, VT, LHS,
4492 DAG.getConstant(VT.getScalarType().getSizeInBits() - 1, VT));
4493 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, LHS, Shift);
4494 AddToWorkList(Shift.getNode());
4495 AddToWorkList(Add.getNode());
4496 return DAG.getNode(ISD::XOR, DL, VT, Add, Shift);
4500 // If the VSELECT result requires splitting and the mask is provided by a
4501 // SETCC, then split both nodes and its operands before legalization. This
4502 // prevents the type legalizer from unrolling SETCC into scalar comparisons
4503 // and enables future optimizations (e.g. min/max pattern matching on X86).
4504 if (N0.getOpcode() == ISD::SETCC) {
4505 EVT VT = N->getValueType(0);
4507 // Check if any splitting is required.
4508 if (TLI.getTypeAction(*DAG.getContext(), VT) !=
4509 TargetLowering::TypeSplitVector)
4512 SDValue Lo, Hi, CCLo, CCHi, LL, LH, RL, RH;
4513 llvm::tie(CCLo, CCHi) = SplitVSETCC(N0.getNode(), DAG);
4514 llvm::tie(LL, LH) = DAG.SplitVectorOperand(N, 1);
4515 llvm::tie(RL, RH) = DAG.SplitVectorOperand(N, 2);
4517 Lo = DAG.getNode(N->getOpcode(), DL, LL.getValueType(), CCLo, LL, RL);
4518 Hi = DAG.getNode(N->getOpcode(), DL, LH.getValueType(), CCHi, LH, RH);
4520 // Add the new VSELECT nodes to the work list in case they need to be split
4522 AddToWorkList(Lo.getNode());
4523 AddToWorkList(Hi.getNode());
4525 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
4528 // Fold (vselect (build_vector all_ones), N1, N2) -> N1
4529 if (ISD::isBuildVectorAllOnes(N0.getNode()))
4531 // Fold (vselect (build_vector all_zeros), N1, N2) -> N2
4532 if (ISD::isBuildVectorAllZeros(N0.getNode()))
4538 SDValue DAGCombiner::visitSELECT_CC(SDNode *N) {
4539 SDValue N0 = N->getOperand(0);
4540 SDValue N1 = N->getOperand(1);
4541 SDValue N2 = N->getOperand(2);
4542 SDValue N3 = N->getOperand(3);
4543 SDValue N4 = N->getOperand(4);
4544 ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get();
4546 // fold select_cc lhs, rhs, x, x, cc -> x
4550 // Determine if the condition we're dealing with is constant
4551 SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()),
4552 N0, N1, CC, SDLoc(N), false);
4553 if (SCC.getNode()) {
4554 AddToWorkList(SCC.getNode());
4556 if (ConstantSDNode *SCCC = dyn_cast<ConstantSDNode>(SCC.getNode())) {
4557 if (!SCCC->isNullValue())
4558 return N2; // cond always true -> true val
4560 return N3; // cond always false -> false val
4563 // Fold to a simpler select_cc
4564 if (SCC.getOpcode() == ISD::SETCC)
4565 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), N2.getValueType(),
4566 SCC.getOperand(0), SCC.getOperand(1), N2, N3,
4570 // If we can fold this based on the true/false value, do so.
4571 if (SimplifySelectOps(N, N2, N3))
4572 return SDValue(N, 0); // Don't revisit N.
4574 // fold select_cc into other things, such as min/max/abs
4575 return SimplifySelectCC(SDLoc(N), N0, N1, N2, N3, CC);
4578 SDValue DAGCombiner::visitSETCC(SDNode *N) {
4579 return SimplifySetCC(N->getValueType(0), N->getOperand(0), N->getOperand(1),
4580 cast<CondCodeSDNode>(N->getOperand(2))->get(),
4584 // tryToFoldExtendOfConstant - Try to fold a sext/zext/aext
4585 // dag node into a ConstantSDNode or a build_vector of constants.
4586 // This function is called by the DAGCombiner when visiting sext/zext/aext
4587 // dag nodes (see for example method DAGCombiner::visitSIGN_EXTEND).
4588 // Vector extends are not folded if operations are legal; this is to
4589 // avoid introducing illegal build_vector dag nodes.
4590 static SDNode *tryToFoldExtendOfConstant(SDNode *N, const TargetLowering &TLI,
4591 SelectionDAG &DAG, bool LegalTypes,
4592 bool LegalOperations) {
4593 unsigned Opcode = N->getOpcode();
4594 SDValue N0 = N->getOperand(0);
4595 EVT VT = N->getValueType(0);
4597 assert((Opcode == ISD::SIGN_EXTEND || Opcode == ISD::ZERO_EXTEND ||
4598 Opcode == ISD::ANY_EXTEND) && "Expected EXTEND dag node in input!");
4600 // fold (sext c1) -> c1
4601 // fold (zext c1) -> c1
4602 // fold (aext c1) -> c1
4603 if (isa<ConstantSDNode>(N0))
4604 return DAG.getNode(Opcode, SDLoc(N), VT, N0).getNode();
4606 // fold (sext (build_vector AllConstants) -> (build_vector AllConstants)
4607 // fold (zext (build_vector AllConstants) -> (build_vector AllConstants)
4608 // fold (aext (build_vector AllConstants) -> (build_vector AllConstants)
4609 EVT SVT = VT.getScalarType();
4610 if (!(VT.isVector() &&
4611 (!LegalTypes || (!LegalOperations && TLI.isTypeLegal(SVT))) &&
4612 ISD::isBuildVectorOfConstantSDNodes(N0.getNode())))
4615 // We can fold this node into a build_vector.
4616 unsigned VTBits = SVT.getSizeInBits();
4617 unsigned EVTBits = N0->getValueType(0).getScalarType().getSizeInBits();
4618 unsigned ShAmt = VTBits - EVTBits;
4619 SmallVector<SDValue, 8> Elts;
4620 unsigned NumElts = N0->getNumOperands();
4623 for (unsigned i=0; i != NumElts; ++i) {
4624 SDValue Op = N0->getOperand(i);
4625 if (Op->getOpcode() == ISD::UNDEF) {
4626 Elts.push_back(DAG.getUNDEF(SVT));
4630 ConstantSDNode *CurrentND = cast<ConstantSDNode>(Op);
4631 const APInt &C = APInt(VTBits, CurrentND->getAPIntValue().getZExtValue());
4632 if (Opcode == ISD::SIGN_EXTEND)
4633 Elts.push_back(DAG.getConstant(C.shl(ShAmt).ashr(ShAmt).getZExtValue(),
4636 Elts.push_back(DAG.getConstant(C.shl(ShAmt).lshr(ShAmt).getZExtValue(),
4640 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, &Elts[0], NumElts).getNode();
4643 // ExtendUsesToFormExtLoad - Trying to extend uses of a load to enable this:
4644 // "fold ({s|z|a}ext (load x)) -> ({s|z|a}ext (truncate ({s|z|a}extload x)))"
4645 // transformation. Returns true if extension are possible and the above
4646 // mentioned transformation is profitable.
4647 static bool ExtendUsesToFormExtLoad(SDNode *N, SDValue N0,
4649 SmallVectorImpl<SDNode *> &ExtendNodes,
4650 const TargetLowering &TLI) {
4651 bool HasCopyToRegUses = false;
4652 bool isTruncFree = TLI.isTruncateFree(N->getValueType(0), N0.getValueType());
4653 for (SDNode::use_iterator UI = N0.getNode()->use_begin(),
4654 UE = N0.getNode()->use_end();
4659 if (UI.getUse().getResNo() != N0.getResNo())
4661 // FIXME: Only extend SETCC N, N and SETCC N, c for now.
4662 if (ExtOpc != ISD::ANY_EXTEND && User->getOpcode() == ISD::SETCC) {
4663 ISD::CondCode CC = cast<CondCodeSDNode>(User->getOperand(2))->get();
4664 if (ExtOpc == ISD::ZERO_EXTEND && ISD::isSignedIntSetCC(CC))
4665 // Sign bits will be lost after a zext.
4668 for (unsigned i = 0; i != 2; ++i) {
4669 SDValue UseOp = User->getOperand(i);
4672 if (!isa<ConstantSDNode>(UseOp))
4677 ExtendNodes.push_back(User);
4680 // If truncates aren't free and there are users we can't
4681 // extend, it isn't worthwhile.
4684 // Remember if this value is live-out.
4685 if (User->getOpcode() == ISD::CopyToReg)
4686 HasCopyToRegUses = true;
4689 if (HasCopyToRegUses) {
4690 bool BothLiveOut = false;
4691 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
4693 SDUse &Use = UI.getUse();
4694 if (Use.getResNo() == 0 && Use.getUser()->getOpcode() == ISD::CopyToReg) {
4700 // Both unextended and extended values are live out. There had better be
4701 // a good reason for the transformation.
4702 return ExtendNodes.size();
4707 void DAGCombiner::ExtendSetCCUses(const SmallVectorImpl<SDNode *> &SetCCs,
4708 SDValue Trunc, SDValue ExtLoad, SDLoc DL,
4709 ISD::NodeType ExtType) {
4710 // Extend SetCC uses if necessary.
4711 for (unsigned i = 0, e = SetCCs.size(); i != e; ++i) {
4712 SDNode *SetCC = SetCCs[i];
4713 SmallVector<SDValue, 4> Ops;
4715 for (unsigned j = 0; j != 2; ++j) {
4716 SDValue SOp = SetCC->getOperand(j);
4718 Ops.push_back(ExtLoad);
4720 Ops.push_back(DAG.getNode(ExtType, DL, ExtLoad->getValueType(0), SOp));
4723 Ops.push_back(SetCC->getOperand(2));
4724 CombineTo(SetCC, DAG.getNode(ISD::SETCC, DL, SetCC->getValueType(0),
4725 &Ops[0], Ops.size()));
4729 SDValue DAGCombiner::visitSIGN_EXTEND(SDNode *N) {
4730 SDValue N0 = N->getOperand(0);
4731 EVT VT = N->getValueType(0);
4733 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
4735 return SDValue(Res, 0);
4737 // fold (sext (sext x)) -> (sext x)
4738 // fold (sext (aext x)) -> (sext x)
4739 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
4740 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT,
4743 if (N0.getOpcode() == ISD::TRUNCATE) {
4744 // fold (sext (truncate (load x))) -> (sext (smaller load x))
4745 // fold (sext (truncate (srl (load x), c))) -> (sext (smaller load (x+c/n)))
4746 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
4747 if (NarrowLoad.getNode()) {
4748 SDNode* oye = N0.getNode()->getOperand(0).getNode();
4749 if (NarrowLoad.getNode() != N0.getNode()) {
4750 CombineTo(N0.getNode(), NarrowLoad);
4751 // CombineTo deleted the truncate, if needed, but not what's under it.
4754 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4757 // See if the value being truncated is already sign extended. If so, just
4758 // eliminate the trunc/sext pair.
4759 SDValue Op = N0.getOperand(0);
4760 unsigned OpBits = Op.getValueType().getScalarType().getSizeInBits();
4761 unsigned MidBits = N0.getValueType().getScalarType().getSizeInBits();
4762 unsigned DestBits = VT.getScalarType().getSizeInBits();
4763 unsigned NumSignBits = DAG.ComputeNumSignBits(Op);
4765 if (OpBits == DestBits) {
4766 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
4767 // bits, it is already ready.
4768 if (NumSignBits > DestBits-MidBits)
4770 } else if (OpBits < DestBits) {
4771 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
4772 // bits, just sext from i32.
4773 if (NumSignBits > OpBits-MidBits)
4774 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, Op);
4776 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
4777 // bits, just truncate to i32.
4778 if (NumSignBits > OpBits-MidBits)
4779 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
4782 // fold (sext (truncate x)) -> (sextinreg x).
4783 if (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG,
4784 N0.getValueType())) {
4785 if (OpBits < DestBits)
4786 Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N0), VT, Op);
4787 else if (OpBits > DestBits)
4788 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), VT, Op);
4789 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, Op,
4790 DAG.getValueType(N0.getValueType()));
4794 // fold (sext (load x)) -> (sext (truncate (sextload x)))
4795 // None of the supported targets knows how to perform load and sign extend
4796 // on vectors in one instruction. We only perform this transformation on
4798 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
4799 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
4800 TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()))) {
4801 bool DoXform = true;
4802 SmallVector<SDNode*, 4> SetCCs;
4803 if (!N0.hasOneUse())
4804 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::SIGN_EXTEND, SetCCs, TLI);
4806 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4807 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
4809 LN0->getBasePtr(), N0.getValueType(),
4810 LN0->getMemOperand());
4811 CombineTo(N, ExtLoad);
4812 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
4813 N0.getValueType(), ExtLoad);
4814 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
4815 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
4817 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4821 // fold (sext (sextload x)) -> (sext (truncate (sextload x)))
4822 // fold (sext ( extload x)) -> (sext (truncate (sextload x)))
4823 if ((ISD::isSEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
4824 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
4825 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4826 EVT MemVT = LN0->getMemoryVT();
4827 if ((!LegalOperations && !LN0->isVolatile()) ||
4828 TLI.isLoadExtLegal(ISD::SEXTLOAD, MemVT)) {
4829 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
4831 LN0->getBasePtr(), MemVT,
4832 LN0->getMemOperand());
4833 CombineTo(N, ExtLoad);
4834 CombineTo(N0.getNode(),
4835 DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
4836 N0.getValueType(), ExtLoad),
4837 ExtLoad.getValue(1));
4838 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4842 // fold (sext (and/or/xor (load x), cst)) ->
4843 // (and/or/xor (sextload x), (sext cst))
4844 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
4845 N0.getOpcode() == ISD::XOR) &&
4846 isa<LoadSDNode>(N0.getOperand(0)) &&
4847 N0.getOperand(1).getOpcode() == ISD::Constant &&
4848 TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()) &&
4849 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
4850 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
4851 if (LN0->getExtensionType() != ISD::ZEXTLOAD) {
4852 bool DoXform = true;
4853 SmallVector<SDNode*, 4> SetCCs;
4854 if (!N0.hasOneUse())
4855 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::SIGN_EXTEND,
4858 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(LN0), VT,
4859 LN0->getChain(), LN0->getBasePtr(),
4861 LN0->getMemOperand());
4862 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
4863 Mask = Mask.sext(VT.getSizeInBits());
4864 SDValue And = DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
4865 ExtLoad, DAG.getConstant(Mask, VT));
4866 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
4867 SDLoc(N0.getOperand(0)),
4868 N0.getOperand(0).getValueType(), ExtLoad);
4870 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
4871 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
4873 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4878 if (N0.getOpcode() == ISD::SETCC) {
4879 // sext(setcc) -> sext_in_reg(vsetcc) for vectors.
4880 // Only do this before legalize for now.
4881 if (VT.isVector() && !LegalOperations &&
4882 TLI.getBooleanContents(true) ==
4883 TargetLowering::ZeroOrNegativeOneBooleanContent) {
4884 EVT N0VT = N0.getOperand(0).getValueType();
4885 // On some architectures (such as SSE/NEON/etc) the SETCC result type is
4886 // of the same size as the compared operands. Only optimize sext(setcc())
4887 // if this is the case.
4888 EVT SVT = getSetCCResultType(N0VT);
4890 // We know that the # elements of the results is the same as the
4891 // # elements of the compare (and the # elements of the compare result
4892 // for that matter). Check to see that they are the same size. If so,
4893 // we know that the element size of the sext'd result matches the
4894 // element size of the compare operands.
4895 if (VT.getSizeInBits() == SVT.getSizeInBits())
4896 return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
4898 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4900 // If the desired elements are smaller or larger than the source
4901 // elements we can use a matching integer vector type and then
4902 // truncate/sign extend
4903 EVT MatchingVectorType = N0VT.changeVectorElementTypeToInteger();
4904 if (SVT == MatchingVectorType) {
4905 SDValue VsetCC = DAG.getSetCC(SDLoc(N), MatchingVectorType,
4906 N0.getOperand(0), N0.getOperand(1),
4907 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4908 return DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT);
4912 // sext(setcc x, y, cc) -> (select (setcc x, y, cc), -1, 0)
4913 unsigned ElementWidth = VT.getScalarType().getSizeInBits();
4915 DAG.getConstant(APInt::getAllOnesValue(ElementWidth), VT);
4917 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
4918 NegOne, DAG.getConstant(0, VT),
4919 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
4920 if (SCC.getNode()) return SCC;
4922 if (!VT.isVector()) {
4923 EVT SetCCVT = getSetCCResultType(N0.getOperand(0).getValueType());
4924 if (!LegalOperations || TLI.isOperationLegal(ISD::SETCC, SetCCVT)) {
4926 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
4927 SDValue SetCC = DAG.getSetCC(DL,
4929 N0.getOperand(0), N0.getOperand(1), CC);
4930 EVT SelectVT = getSetCCResultType(VT);
4931 return DAG.getSelect(DL, VT,
4932 DAG.getSExtOrTrunc(SetCC, DL, SelectVT),
4933 NegOne, DAG.getConstant(0, VT));
4939 // fold (sext x) -> (zext x) if the sign bit is known zero.
4940 if ((!LegalOperations || TLI.isOperationLegal(ISD::ZERO_EXTEND, VT)) &&
4941 DAG.SignBitIsZero(N0))
4942 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, N0);
4947 // isTruncateOf - If N is a truncate of some other value, return true, record
4948 // the value being truncated in Op and which of Op's bits are zero in KnownZero.
4949 // This function computes KnownZero to avoid a duplicated call to
4950 // ComputeMaskedBits in the caller.
4951 static bool isTruncateOf(SelectionDAG &DAG, SDValue N, SDValue &Op,
4954 if (N->getOpcode() == ISD::TRUNCATE) {
4955 Op = N->getOperand(0);
4956 DAG.ComputeMaskedBits(Op, KnownZero, KnownOne);
4960 if (N->getOpcode() != ISD::SETCC || N->getValueType(0) != MVT::i1 ||
4961 cast<CondCodeSDNode>(N->getOperand(2))->get() != ISD::SETNE)
4964 SDValue Op0 = N->getOperand(0);
4965 SDValue Op1 = N->getOperand(1);
4966 assert(Op0.getValueType() == Op1.getValueType());
4968 ConstantSDNode *COp0 = dyn_cast<ConstantSDNode>(Op0);
4969 ConstantSDNode *COp1 = dyn_cast<ConstantSDNode>(Op1);
4970 if (COp0 && COp0->isNullValue())
4972 else if (COp1 && COp1->isNullValue())
4977 DAG.ComputeMaskedBits(Op, KnownZero, KnownOne);
4979 if (!(KnownZero | APInt(Op.getValueSizeInBits(), 1)).isAllOnesValue())
4985 SDValue DAGCombiner::visitZERO_EXTEND(SDNode *N) {
4986 SDValue N0 = N->getOperand(0);
4987 EVT VT = N->getValueType(0);
4989 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
4991 return SDValue(Res, 0);
4993 // fold (zext (zext x)) -> (zext x)
4994 // fold (zext (aext x)) -> (zext x)
4995 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
4996 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT,
4999 // fold (zext (truncate x)) -> (zext x) or
5000 // (zext (truncate x)) -> (truncate x)
5001 // This is valid when the truncated bits of x are already zero.
5002 // FIXME: We should extend this to work for vectors too.
5005 if (!VT.isVector() && isTruncateOf(DAG, N0, Op, KnownZero)) {
5006 APInt TruncatedBits =
5007 (Op.getValueSizeInBits() == N0.getValueSizeInBits()) ?
5008 APInt(Op.getValueSizeInBits(), 0) :
5009 APInt::getBitsSet(Op.getValueSizeInBits(),
5010 N0.getValueSizeInBits(),
5011 std::min(Op.getValueSizeInBits(),
5012 VT.getSizeInBits()));
5013 if (TruncatedBits == (KnownZero & TruncatedBits)) {
5014 if (VT.bitsGT(Op.getValueType()))
5015 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, Op);
5016 if (VT.bitsLT(Op.getValueType()))
5017 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
5023 // fold (zext (truncate (load x))) -> (zext (smaller load x))
5024 // fold (zext (truncate (srl (load x), c))) -> (zext (small load (x+c/n)))
5025 if (N0.getOpcode() == ISD::TRUNCATE) {
5026 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5027 if (NarrowLoad.getNode()) {
5028 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5029 if (NarrowLoad.getNode() != N0.getNode()) {
5030 CombineTo(N0.getNode(), NarrowLoad);
5031 // CombineTo deleted the truncate, if needed, but not what's under it.
5034 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5038 // fold (zext (truncate x)) -> (and x, mask)
5039 if (N0.getOpcode() == ISD::TRUNCATE &&
5040 (!LegalOperations || TLI.isOperationLegal(ISD::AND, VT))) {
5042 // fold (zext (truncate (load x))) -> (zext (smaller load x))
5043 // fold (zext (truncate (srl (load x), c))) -> (zext (smaller load (x+c/n)))
5044 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5045 if (NarrowLoad.getNode()) {
5046 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5047 if (NarrowLoad.getNode() != N0.getNode()) {
5048 CombineTo(N0.getNode(), NarrowLoad);
5049 // CombineTo deleted the truncate, if needed, but not what's under it.
5052 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5055 SDValue Op = N0.getOperand(0);
5056 if (Op.getValueType().bitsLT(VT)) {
5057 Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, Op);
5058 AddToWorkList(Op.getNode());
5059 } else if (Op.getValueType().bitsGT(VT)) {
5060 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
5061 AddToWorkList(Op.getNode());
5063 return DAG.getZeroExtendInReg(Op, SDLoc(N),
5064 N0.getValueType().getScalarType());
5067 // Fold (zext (and (trunc x), cst)) -> (and x, cst),
5068 // if either of the casts is not free.
5069 if (N0.getOpcode() == ISD::AND &&
5070 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
5071 N0.getOperand(1).getOpcode() == ISD::Constant &&
5072 (!TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
5073 N0.getValueType()) ||
5074 !TLI.isZExtFree(N0.getValueType(), VT))) {
5075 SDValue X = N0.getOperand(0).getOperand(0);
5076 if (X.getValueType().bitsLT(VT)) {
5077 X = DAG.getNode(ISD::ANY_EXTEND, SDLoc(X), VT, X);
5078 } else if (X.getValueType().bitsGT(VT)) {
5079 X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X);
5081 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5082 Mask = Mask.zext(VT.getSizeInBits());
5083 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5084 X, DAG.getConstant(Mask, VT));
5087 // fold (zext (load x)) -> (zext (truncate (zextload x)))
5088 // None of the supported targets knows how to perform load and vector_zext
5089 // on vectors in one instruction. We only perform this transformation on
5091 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
5092 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5093 TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()))) {
5094 bool DoXform = true;
5095 SmallVector<SDNode*, 4> SetCCs;
5096 if (!N0.hasOneUse())
5097 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ZERO_EXTEND, SetCCs, TLI);
5099 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5100 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT,
5102 LN0->getBasePtr(), N0.getValueType(),
5103 LN0->getMemOperand());
5104 CombineTo(N, ExtLoad);
5105 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5106 N0.getValueType(), ExtLoad);
5107 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
5109 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5111 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5115 // fold (zext (and/or/xor (load x), cst)) ->
5116 // (and/or/xor (zextload x), (zext cst))
5117 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
5118 N0.getOpcode() == ISD::XOR) &&
5119 isa<LoadSDNode>(N0.getOperand(0)) &&
5120 N0.getOperand(1).getOpcode() == ISD::Constant &&
5121 TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()) &&
5122 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
5123 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
5124 if (LN0->getExtensionType() != ISD::SEXTLOAD) {
5125 bool DoXform = true;
5126 SmallVector<SDNode*, 4> SetCCs;
5127 if (!N0.hasOneUse())
5128 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::ZERO_EXTEND,
5131 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), VT,
5132 LN0->getChain(), LN0->getBasePtr(),
5134 LN0->getMemOperand());
5135 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5136 Mask = Mask.zext(VT.getSizeInBits());
5137 SDValue And = DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
5138 ExtLoad, DAG.getConstant(Mask, VT));
5139 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
5140 SDLoc(N0.getOperand(0)),
5141 N0.getOperand(0).getValueType(), ExtLoad);
5143 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
5144 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5146 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5151 // fold (zext (zextload x)) -> (zext (truncate (zextload x)))
5152 // fold (zext ( extload x)) -> (zext (truncate (zextload x)))
5153 if ((ISD::isZEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
5154 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
5155 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5156 EVT MemVT = LN0->getMemoryVT();
5157 if ((!LegalOperations && !LN0->isVolatile()) ||
5158 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT)) {
5159 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT,
5161 LN0->getBasePtr(), MemVT,
5162 LN0->getMemOperand());
5163 CombineTo(N, ExtLoad);
5164 CombineTo(N0.getNode(),
5165 DAG.getNode(ISD::TRUNCATE, SDLoc(N0), N0.getValueType(),
5167 ExtLoad.getValue(1));
5168 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5172 if (N0.getOpcode() == ISD::SETCC) {
5173 if (!LegalOperations && VT.isVector() &&
5174 N0.getValueType().getVectorElementType() == MVT::i1) {
5175 EVT N0VT = N0.getOperand(0).getValueType();
5176 if (getSetCCResultType(N0VT) == N0.getValueType())
5179 // zext(setcc) -> (and (vsetcc), (1, 1, ...) for vectors.
5180 // Only do this before legalize for now.
5181 EVT EltVT = VT.getVectorElementType();
5182 SmallVector<SDValue,8> OneOps(VT.getVectorNumElements(),
5183 DAG.getConstant(1, EltVT));
5184 if (VT.getSizeInBits() == N0VT.getSizeInBits())
5185 // We know that the # elements of the results is the same as the
5186 // # elements of the compare (and the # elements of the compare result
5187 // for that matter). Check to see that they are the same size. If so,
5188 // we know that the element size of the sext'd result matches the
5189 // element size of the compare operands.
5190 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5191 DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
5193 cast<CondCodeSDNode>(N0.getOperand(2))->get()),
5194 DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT,
5195 &OneOps[0], OneOps.size()));
5197 // If the desired elements are smaller or larger than the source
5198 // elements we can use a matching integer vector type and then
5199 // truncate/sign extend
5200 EVT MatchingElementType =
5201 EVT::getIntegerVT(*DAG.getContext(),
5202 N0VT.getScalarType().getSizeInBits());
5203 EVT MatchingVectorType =
5204 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
5205 N0VT.getVectorNumElements());
5207 DAG.getSetCC(SDLoc(N), MatchingVectorType, N0.getOperand(0),
5209 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5210 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5211 DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT),
5212 DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT,
5213 &OneOps[0], OneOps.size()));
5216 // zext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
5218 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
5219 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
5220 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
5221 if (SCC.getNode()) return SCC;
5224 // (zext (shl (zext x), cst)) -> (shl (zext x), cst)
5225 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL) &&
5226 isa<ConstantSDNode>(N0.getOperand(1)) &&
5227 N0.getOperand(0).getOpcode() == ISD::ZERO_EXTEND &&
5229 SDValue ShAmt = N0.getOperand(1);
5230 unsigned ShAmtVal = cast<ConstantSDNode>(ShAmt)->getZExtValue();
5231 if (N0.getOpcode() == ISD::SHL) {
5232 SDValue InnerZExt = N0.getOperand(0);
5233 // If the original shl may be shifting out bits, do not perform this
5235 unsigned KnownZeroBits = InnerZExt.getValueType().getSizeInBits() -
5236 InnerZExt.getOperand(0).getValueType().getSizeInBits();
5237 if (ShAmtVal > KnownZeroBits)
5243 // Ensure that the shift amount is wide enough for the shifted value.
5244 if (VT.getSizeInBits() >= 256)
5245 ShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, ShAmt);
5247 return DAG.getNode(N0.getOpcode(), DL, VT,
5248 DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0)),
5255 SDValue DAGCombiner::visitANY_EXTEND(SDNode *N) {
5256 SDValue N0 = N->getOperand(0);
5257 EVT VT = N->getValueType(0);
5259 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
5261 return SDValue(Res, 0);
5263 // fold (aext (aext x)) -> (aext x)
5264 // fold (aext (zext x)) -> (zext x)
5265 // fold (aext (sext x)) -> (sext x)
5266 if (N0.getOpcode() == ISD::ANY_EXTEND ||
5267 N0.getOpcode() == ISD::ZERO_EXTEND ||
5268 N0.getOpcode() == ISD::SIGN_EXTEND)
5269 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, N0.getOperand(0));
5271 // fold (aext (truncate (load x))) -> (aext (smaller load x))
5272 // fold (aext (truncate (srl (load x), c))) -> (aext (small load (x+c/n)))
5273 if (N0.getOpcode() == ISD::TRUNCATE) {
5274 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5275 if (NarrowLoad.getNode()) {
5276 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5277 if (NarrowLoad.getNode() != N0.getNode()) {
5278 CombineTo(N0.getNode(), NarrowLoad);
5279 // CombineTo deleted the truncate, if needed, but not what's under it.
5282 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5286 // fold (aext (truncate x))
5287 if (N0.getOpcode() == ISD::TRUNCATE) {
5288 SDValue TruncOp = N0.getOperand(0);
5289 if (TruncOp.getValueType() == VT)
5290 return TruncOp; // x iff x size == zext size.
5291 if (TruncOp.getValueType().bitsGT(VT))
5292 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, TruncOp);
5293 return DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, TruncOp);
5296 // Fold (aext (and (trunc x), cst)) -> (and x, cst)
5297 // if the trunc is not free.
5298 if (N0.getOpcode() == ISD::AND &&
5299 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
5300 N0.getOperand(1).getOpcode() == ISD::Constant &&
5301 !TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
5302 N0.getValueType())) {
5303 SDValue X = N0.getOperand(0).getOperand(0);
5304 if (X.getValueType().bitsLT(VT)) {
5305 X = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, X);
5306 } else if (X.getValueType().bitsGT(VT)) {
5307 X = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, X);
5309 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5310 Mask = Mask.zext(VT.getSizeInBits());
5311 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5312 X, DAG.getConstant(Mask, VT));
5315 // fold (aext (load x)) -> (aext (truncate (extload x)))
5316 // None of the supported targets knows how to perform load and any_ext
5317 // on vectors in one instruction. We only perform this transformation on
5319 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
5320 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5321 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType()))) {
5322 bool DoXform = true;
5323 SmallVector<SDNode*, 4> SetCCs;
5324 if (!N0.hasOneUse())
5325 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ANY_EXTEND, SetCCs, TLI);
5327 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5328 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT,
5330 LN0->getBasePtr(), N0.getValueType(),
5331 LN0->getMemOperand());
5332 CombineTo(N, ExtLoad);
5333 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5334 N0.getValueType(), ExtLoad);
5335 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
5336 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5338 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5342 // fold (aext (zextload x)) -> (aext (truncate (zextload x)))
5343 // fold (aext (sextload x)) -> (aext (truncate (sextload x)))
5344 // fold (aext ( extload x)) -> (aext (truncate (extload x)))
5345 if (N0.getOpcode() == ISD::LOAD &&
5346 !ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
5348 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5349 EVT MemVT = LN0->getMemoryVT();
5350 SDValue ExtLoad = DAG.getExtLoad(LN0->getExtensionType(), SDLoc(N),
5351 VT, LN0->getChain(), LN0->getBasePtr(),
5352 MemVT, LN0->getMemOperand());
5353 CombineTo(N, ExtLoad);
5354 CombineTo(N0.getNode(),
5355 DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5356 N0.getValueType(), ExtLoad),
5357 ExtLoad.getValue(1));
5358 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5361 if (N0.getOpcode() == ISD::SETCC) {
5362 // aext(setcc) -> sext_in_reg(vsetcc) for vectors.
5363 // Only do this before legalize for now.
5364 if (VT.isVector() && !LegalOperations) {
5365 EVT N0VT = N0.getOperand(0).getValueType();
5366 // We know that the # elements of the results is the same as the
5367 // # elements of the compare (and the # elements of the compare result
5368 // for that matter). Check to see that they are the same size. If so,
5369 // we know that the element size of the sext'd result matches the
5370 // element size of the compare operands.
5371 if (VT.getSizeInBits() == N0VT.getSizeInBits())
5372 return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
5374 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5375 // If the desired elements are smaller or larger than the source
5376 // elements we can use a matching integer vector type and then
5377 // truncate/sign extend
5379 EVT MatchingElementType =
5380 EVT::getIntegerVT(*DAG.getContext(),
5381 N0VT.getScalarType().getSizeInBits());
5382 EVT MatchingVectorType =
5383 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
5384 N0VT.getVectorNumElements());
5386 DAG.getSetCC(SDLoc(N), MatchingVectorType, N0.getOperand(0),
5388 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5389 return DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT);
5393 // aext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
5395 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
5396 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
5397 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
5405 /// GetDemandedBits - See if the specified operand can be simplified with the
5406 /// knowledge that only the bits specified by Mask are used. If so, return the
5407 /// simpler operand, otherwise return a null SDValue.
5408 SDValue DAGCombiner::GetDemandedBits(SDValue V, const APInt &Mask) {
5409 switch (V.getOpcode()) {
5411 case ISD::Constant: {
5412 const ConstantSDNode *CV = cast<ConstantSDNode>(V.getNode());
5413 assert(CV != 0 && "Const value should be ConstSDNode.");
5414 const APInt &CVal = CV->getAPIntValue();
5415 APInt NewVal = CVal & Mask;
5417 return DAG.getConstant(NewVal, V.getValueType());
5422 // If the LHS or RHS don't contribute bits to the or, drop them.
5423 if (DAG.MaskedValueIsZero(V.getOperand(0), Mask))
5424 return V.getOperand(1);
5425 if (DAG.MaskedValueIsZero(V.getOperand(1), Mask))
5426 return V.getOperand(0);
5429 // Only look at single-use SRLs.
5430 if (!V.getNode()->hasOneUse())
5432 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(V.getOperand(1))) {
5433 // See if we can recursively simplify the LHS.
5434 unsigned Amt = RHSC->getZExtValue();
5436 // Watch out for shift count overflow though.
5437 if (Amt >= Mask.getBitWidth()) break;
5438 APInt NewMask = Mask << Amt;
5439 SDValue SimplifyLHS = GetDemandedBits(V.getOperand(0), NewMask);
5440 if (SimplifyLHS.getNode())
5441 return DAG.getNode(ISD::SRL, SDLoc(V), V.getValueType(),
5442 SimplifyLHS, V.getOperand(1));
5448 /// ReduceLoadWidth - If the result of a wider load is shifted to right of N
5449 /// bits and then truncated to a narrower type and where N is a multiple
5450 /// of number of bits of the narrower type, transform it to a narrower load
5451 /// from address + N / num of bits of new type. If the result is to be
5452 /// extended, also fold the extension to form a extending load.
5453 SDValue DAGCombiner::ReduceLoadWidth(SDNode *N) {
5454 unsigned Opc = N->getOpcode();
5456 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
5457 SDValue N0 = N->getOperand(0);
5458 EVT VT = N->getValueType(0);
5461 // This transformation isn't valid for vector loads.
5465 // Special case: SIGN_EXTEND_INREG is basically truncating to ExtVT then
5467 if (Opc == ISD::SIGN_EXTEND_INREG) {
5468 ExtType = ISD::SEXTLOAD;
5469 ExtVT = cast<VTSDNode>(N->getOperand(1))->getVT();
5470 } else if (Opc == ISD::SRL) {
5471 // Another special-case: SRL is basically zero-extending a narrower value.
5472 ExtType = ISD::ZEXTLOAD;
5474 ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1));
5475 if (!N01) return SDValue();
5476 ExtVT = EVT::getIntegerVT(*DAG.getContext(),
5477 VT.getSizeInBits() - N01->getZExtValue());
5479 if (LegalOperations && !TLI.isLoadExtLegal(ExtType, ExtVT))
5482 unsigned EVTBits = ExtVT.getSizeInBits();
5484 // Do not generate loads of non-round integer types since these can
5485 // be expensive (and would be wrong if the type is not byte sized).
5486 if (!ExtVT.isRound())
5490 if (N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
5491 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
5492 ShAmt = N01->getZExtValue();
5493 // Is the shift amount a multiple of size of VT?
5494 if ((ShAmt & (EVTBits-1)) == 0) {
5495 N0 = N0.getOperand(0);
5496 // Is the load width a multiple of size of VT?
5497 if ((N0.getValueType().getSizeInBits() & (EVTBits-1)) != 0)
5501 // At this point, we must have a load or else we can't do the transform.
5502 if (!isa<LoadSDNode>(N0)) return SDValue();
5504 // Because a SRL must be assumed to *need* to zero-extend the high bits
5505 // (as opposed to anyext the high bits), we can't combine the zextload
5506 // lowering of SRL and an sextload.
5507 if (cast<LoadSDNode>(N0)->getExtensionType() == ISD::SEXTLOAD)
5510 // If the shift amount is larger than the input type then we're not
5511 // accessing any of the loaded bytes. If the load was a zextload/extload
5512 // then the result of the shift+trunc is zero/undef (handled elsewhere).
5513 if (ShAmt >= cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits())
5518 // If the load is shifted left (and the result isn't shifted back right),
5519 // we can fold the truncate through the shift.
5520 unsigned ShLeftAmt = 0;
5521 if (ShAmt == 0 && N0.getOpcode() == ISD::SHL && N0.hasOneUse() &&
5522 ExtVT == VT && TLI.isNarrowingProfitable(N0.getValueType(), VT)) {
5523 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
5524 ShLeftAmt = N01->getZExtValue();
5525 N0 = N0.getOperand(0);
5529 // If we haven't found a load, we can't narrow it. Don't transform one with
5530 // multiple uses, this would require adding a new load.
5531 if (!isa<LoadSDNode>(N0) || !N0.hasOneUse())
5534 // Don't change the width of a volatile load.
5535 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5536 if (LN0->isVolatile())
5539 // Verify that we are actually reducing a load width here.
5540 if (LN0->getMemoryVT().getSizeInBits() < EVTBits)
5543 // For the transform to be legal, the load must produce only two values
5544 // (the value loaded and the chain). Don't transform a pre-increment
5545 // load, for example, which produces an extra value. Otherwise the
5546 // transformation is not equivalent, and the downstream logic to replace
5547 // uses gets things wrong.
5548 if (LN0->getNumValues() > 2)
5551 // If the load that we're shrinking is an extload and we're not just
5552 // discarding the extension we can't simply shrink the load. Bail.
5553 // TODO: It would be possible to merge the extensions in some cases.
5554 if (LN0->getExtensionType() != ISD::NON_EXTLOAD &&
5555 LN0->getMemoryVT().getSizeInBits() < ExtVT.getSizeInBits() + ShAmt)
5558 EVT PtrType = N0.getOperand(1).getValueType();
5560 if (PtrType == MVT::Untyped || PtrType.isExtended())
5561 // It's not possible to generate a constant of extended or untyped type.
5564 // For big endian targets, we need to adjust the offset to the pointer to
5565 // load the correct bytes.
5566 if (TLI.isBigEndian()) {
5567 unsigned LVTStoreBits = LN0->getMemoryVT().getStoreSizeInBits();
5568 unsigned EVTStoreBits = ExtVT.getStoreSizeInBits();
5569 ShAmt = LVTStoreBits - EVTStoreBits - ShAmt;
5572 uint64_t PtrOff = ShAmt / 8;
5573 unsigned NewAlign = MinAlign(LN0->getAlignment(), PtrOff);
5574 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(LN0),
5575 PtrType, LN0->getBasePtr(),
5576 DAG.getConstant(PtrOff, PtrType));
5577 AddToWorkList(NewPtr.getNode());
5580 if (ExtType == ISD::NON_EXTLOAD)
5581 Load = DAG.getLoad(VT, SDLoc(N0), LN0->getChain(), NewPtr,
5582 LN0->getPointerInfo().getWithOffset(PtrOff),
5583 LN0->isVolatile(), LN0->isNonTemporal(),
5584 LN0->isInvariant(), NewAlign, LN0->getTBAAInfo());
5586 Load = DAG.getExtLoad(ExtType, SDLoc(N0), VT, LN0->getChain(),NewPtr,
5587 LN0->getPointerInfo().getWithOffset(PtrOff),
5588 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
5589 NewAlign, LN0->getTBAAInfo());
5591 // Replace the old load's chain with the new load's chain.
5592 WorkListRemover DeadNodes(*this);
5593 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1));
5595 // Shift the result left, if we've swallowed a left shift.
5596 SDValue Result = Load;
5597 if (ShLeftAmt != 0) {
5598 EVT ShImmTy = getShiftAmountTy(Result.getValueType());
5599 if (!isUIntN(ShImmTy.getSizeInBits(), ShLeftAmt))
5601 // If the shift amount is as large as the result size (but, presumably,
5602 // no larger than the source) then the useful bits of the result are
5603 // zero; we can't simply return the shortened shift, because the result
5604 // of that operation is undefined.
5605 if (ShLeftAmt >= VT.getSizeInBits())
5606 Result = DAG.getConstant(0, VT);
5608 Result = DAG.getNode(ISD::SHL, SDLoc(N0), VT,
5609 Result, DAG.getConstant(ShLeftAmt, ShImmTy));
5612 // Return the new loaded value.
5616 SDValue DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
5617 SDValue N0 = N->getOperand(0);
5618 SDValue N1 = N->getOperand(1);
5619 EVT VT = N->getValueType(0);
5620 EVT EVT = cast<VTSDNode>(N1)->getVT();
5621 unsigned VTBits = VT.getScalarType().getSizeInBits();
5622 unsigned EVTBits = EVT.getScalarType().getSizeInBits();
5624 // fold (sext_in_reg c1) -> c1
5625 if (isa<ConstantSDNode>(N0) || N0.getOpcode() == ISD::UNDEF)
5626 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, N0, N1);
5628 // If the input is already sign extended, just drop the extension.
5629 if (DAG.ComputeNumSignBits(N0) >= VTBits-EVTBits+1)
5632 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2
5633 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
5634 EVT.bitsLT(cast<VTSDNode>(N0.getOperand(1))->getVT()))
5635 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
5636 N0.getOperand(0), N1);
5638 // fold (sext_in_reg (sext x)) -> (sext x)
5639 // fold (sext_in_reg (aext x)) -> (sext x)
5640 // if x is small enough.
5641 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) {
5642 SDValue N00 = N0.getOperand(0);
5643 if (N00.getValueType().getScalarType().getSizeInBits() <= EVTBits &&
5644 (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND, VT)))
5645 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, N00, N1);
5648 // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is known zero.
5649 if (DAG.MaskedValueIsZero(N0, APInt::getBitsSet(VTBits, EVTBits-1, EVTBits)))
5650 return DAG.getZeroExtendInReg(N0, SDLoc(N), EVT);
5652 // fold operands of sext_in_reg based on knowledge that the top bits are not
5654 if (SimplifyDemandedBits(SDValue(N, 0)))
5655 return SDValue(N, 0);
5657 // fold (sext_in_reg (load x)) -> (smaller sextload x)
5658 // fold (sext_in_reg (srl (load x), c)) -> (smaller sextload (x+c/evtbits))
5659 SDValue NarrowLoad = ReduceLoadWidth(N);
5660 if (NarrowLoad.getNode())
5663 // fold (sext_in_reg (srl X, 24), i8) -> (sra X, 24)
5664 // fold (sext_in_reg (srl X, 23), i8) -> (sra X, 23) iff possible.
5665 // We already fold "(sext_in_reg (srl X, 25), i8) -> srl X, 25" above.
5666 if (N0.getOpcode() == ISD::SRL) {
5667 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
5668 if (ShAmt->getZExtValue()+EVTBits <= VTBits) {
5669 // We can turn this into an SRA iff the input to the SRL is already sign
5671 unsigned InSignBits = DAG.ComputeNumSignBits(N0.getOperand(0));
5672 if (VTBits-(ShAmt->getZExtValue()+EVTBits) < InSignBits)
5673 return DAG.getNode(ISD::SRA, SDLoc(N), VT,
5674 N0.getOperand(0), N0.getOperand(1));
5678 // fold (sext_inreg (extload x)) -> (sextload x)
5679 if (ISD::isEXTLoad(N0.getNode()) &&
5680 ISD::isUNINDEXEDLoad(N0.getNode()) &&
5681 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
5682 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5683 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) {
5684 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5685 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5687 LN0->getBasePtr(), EVT,
5688 LN0->getMemOperand());
5689 CombineTo(N, ExtLoad);
5690 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
5691 AddToWorkList(ExtLoad.getNode());
5692 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5694 // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use
5695 if (ISD::isZEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
5697 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
5698 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5699 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) {
5700 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5701 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5703 LN0->getBasePtr(), EVT,
5704 LN0->getMemOperand());
5705 CombineTo(N, ExtLoad);
5706 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
5707 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5710 // Form (sext_inreg (bswap >> 16)) or (sext_inreg (rotl (bswap) 16))
5711 if (EVTBits <= 16 && N0.getOpcode() == ISD::OR) {
5712 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
5713 N0.getOperand(1), false);
5714 if (BSwap.getNode() != 0)
5715 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
5719 // Fold a sext_inreg of a build_vector of ConstantSDNodes or undefs
5720 // into a build_vector.
5721 if (ISD::isBuildVectorOfConstantSDNodes(N0.getNode())) {
5722 SmallVector<SDValue, 8> Elts;
5723 unsigned NumElts = N0->getNumOperands();
5724 unsigned ShAmt = VTBits - EVTBits;
5726 for (unsigned i = 0; i != NumElts; ++i) {
5727 SDValue Op = N0->getOperand(i);
5728 if (Op->getOpcode() == ISD::UNDEF) {
5733 ConstantSDNode *CurrentND = cast<ConstantSDNode>(Op);
5734 const APInt &C = APInt(VTBits, CurrentND->getAPIntValue().getZExtValue());
5735 Elts.push_back(DAG.getConstant(C.shl(ShAmt).ashr(ShAmt).getZExtValue(),
5736 Op.getValueType()));
5739 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, &Elts[0], NumElts);
5745 SDValue DAGCombiner::visitTRUNCATE(SDNode *N) {
5746 SDValue N0 = N->getOperand(0);
5747 EVT VT = N->getValueType(0);
5748 bool isLE = TLI.isLittleEndian();
5751 if (N0.getValueType() == N->getValueType(0))
5753 // fold (truncate c1) -> c1
5754 if (isa<ConstantSDNode>(N0))
5755 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0);
5756 // fold (truncate (truncate x)) -> (truncate x)
5757 if (N0.getOpcode() == ISD::TRUNCATE)
5758 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0));
5759 // fold (truncate (ext x)) -> (ext x) or (truncate x) or x
5760 if (N0.getOpcode() == ISD::ZERO_EXTEND ||
5761 N0.getOpcode() == ISD::SIGN_EXTEND ||
5762 N0.getOpcode() == ISD::ANY_EXTEND) {
5763 if (N0.getOperand(0).getValueType().bitsLT(VT))
5764 // if the source is smaller than the dest, we still need an extend
5765 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
5767 if (N0.getOperand(0).getValueType().bitsGT(VT))
5768 // if the source is larger than the dest, than we just need the truncate
5769 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0));
5770 // if the source and dest are the same type, we can drop both the extend
5771 // and the truncate.
5772 return N0.getOperand(0);
5775 // Fold extract-and-trunc into a narrow extract. For example:
5776 // i64 x = EXTRACT_VECTOR_ELT(v2i64 val, i32 1)
5777 // i32 y = TRUNCATE(i64 x)
5779 // v16i8 b = BITCAST (v2i64 val)
5780 // i8 x = EXTRACT_VECTOR_ELT(v16i8 b, i32 8)
5782 // Note: We only run this optimization after type legalization (which often
5783 // creates this pattern) and before operation legalization after which
5784 // we need to be more careful about the vector instructions that we generate.
5785 if (N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5786 LegalTypes && !LegalOperations && N0->hasOneUse()) {
5788 EVT VecTy = N0.getOperand(0).getValueType();
5789 EVT ExTy = N0.getValueType();
5790 EVT TrTy = N->getValueType(0);
5792 unsigned NumElem = VecTy.getVectorNumElements();
5793 unsigned SizeRatio = ExTy.getSizeInBits()/TrTy.getSizeInBits();
5795 EVT NVT = EVT::getVectorVT(*DAG.getContext(), TrTy, SizeRatio * NumElem);
5796 assert(NVT.getSizeInBits() == VecTy.getSizeInBits() && "Invalid Size");
5798 SDValue EltNo = N0->getOperand(1);
5799 if (isa<ConstantSDNode>(EltNo) && isTypeLegal(NVT)) {
5800 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
5801 EVT IndexTy = TLI.getVectorIdxTy();
5802 int Index = isLE ? (Elt*SizeRatio) : (Elt*SizeRatio + (SizeRatio-1));
5804 SDValue V = DAG.getNode(ISD::BITCAST, SDLoc(N),
5805 NVT, N0.getOperand(0));
5807 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
5809 DAG.getConstant(Index, IndexTy));
5813 // Fold a series of buildvector, bitcast, and truncate if possible.
5815 // (2xi32 trunc (bitcast ((4xi32)buildvector x, x, y, y) 2xi64)) to
5816 // (2xi32 (buildvector x, y)).
5817 if (Level == AfterLegalizeVectorOps && VT.isVector() &&
5818 N0.getOpcode() == ISD::BITCAST && N0.hasOneUse() &&
5819 N0.getOperand(0).getOpcode() == ISD::BUILD_VECTOR &&
5820 N0.getOperand(0).hasOneUse()) {
5822 SDValue BuildVect = N0.getOperand(0);
5823 EVT BuildVectEltTy = BuildVect.getValueType().getVectorElementType();
5824 EVT TruncVecEltTy = VT.getVectorElementType();
5826 // Check that the element types match.
5827 if (BuildVectEltTy == TruncVecEltTy) {
5828 // Now we only need to compute the offset of the truncated elements.
5829 unsigned BuildVecNumElts = BuildVect.getNumOperands();
5830 unsigned TruncVecNumElts = VT.getVectorNumElements();
5831 unsigned TruncEltOffset = BuildVecNumElts / TruncVecNumElts;
5833 assert((BuildVecNumElts % TruncVecNumElts) == 0 &&
5834 "Invalid number of elements");
5836 SmallVector<SDValue, 8> Opnds;
5837 for (unsigned i = 0, e = BuildVecNumElts; i != e; i += TruncEltOffset)
5838 Opnds.push_back(BuildVect.getOperand(i));
5840 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, &Opnds[0],
5845 // See if we can simplify the input to this truncate through knowledge that
5846 // only the low bits are being used.
5847 // For example "trunc (or (shl x, 8), y)" // -> trunc y
5848 // Currently we only perform this optimization on scalars because vectors
5849 // may have different active low bits.
5850 if (!VT.isVector()) {
5852 GetDemandedBits(N0, APInt::getLowBitsSet(N0.getValueSizeInBits(),
5853 VT.getSizeInBits()));
5854 if (Shorter.getNode())
5855 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Shorter);
5857 // fold (truncate (load x)) -> (smaller load x)
5858 // fold (truncate (srl (load x), c)) -> (smaller load (x+c/evtbits))
5859 if (!LegalTypes || TLI.isTypeDesirableForOp(N0.getOpcode(), VT)) {
5860 SDValue Reduced = ReduceLoadWidth(N);
5861 if (Reduced.getNode())
5863 // Handle the case where the load remains an extending load even
5864 // after truncation.
5865 if (N0.hasOneUse() && ISD::isUNINDEXEDLoad(N0.getNode())) {
5866 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5867 if (!LN0->isVolatile() &&
5868 LN0->getMemoryVT().getStoreSizeInBits() < VT.getSizeInBits()) {
5869 SDValue NewLoad = DAG.getExtLoad(LN0->getExtensionType(), SDLoc(LN0),
5870 VT, LN0->getChain(), LN0->getBasePtr(),
5872 LN0->getMemOperand());
5873 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLoad.getValue(1));
5878 // fold (trunc (concat ... x ...)) -> (concat ..., (trunc x), ...)),
5879 // where ... are all 'undef'.
5880 if (N0.getOpcode() == ISD::CONCAT_VECTORS && !LegalTypes) {
5881 SmallVector<EVT, 8> VTs;
5884 unsigned NumDefs = 0;
5886 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) {
5887 SDValue X = N0.getOperand(i);
5888 if (X.getOpcode() != ISD::UNDEF) {
5893 // Stop if more than one members are non-undef.
5896 VTs.push_back(EVT::getVectorVT(*DAG.getContext(),
5897 VT.getVectorElementType(),
5898 X.getValueType().getVectorNumElements()));
5902 return DAG.getUNDEF(VT);
5905 assert(V.getNode() && "The single defined operand is empty!");
5906 SmallVector<SDValue, 8> Opnds;
5907 for (unsigned i = 0, e = VTs.size(); i != e; ++i) {
5909 Opnds.push_back(DAG.getUNDEF(VTs[i]));
5912 SDValue NV = DAG.getNode(ISD::TRUNCATE, SDLoc(V), VTs[i], V);
5913 AddToWorkList(NV.getNode());
5914 Opnds.push_back(NV);
5916 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
5917 &Opnds[0], Opnds.size());
5921 // Simplify the operands using demanded-bits information.
5922 if (!VT.isVector() &&
5923 SimplifyDemandedBits(SDValue(N, 0)))
5924 return SDValue(N, 0);
5929 static SDNode *getBuildPairElt(SDNode *N, unsigned i) {
5930 SDValue Elt = N->getOperand(i);
5931 if (Elt.getOpcode() != ISD::MERGE_VALUES)
5932 return Elt.getNode();
5933 return Elt.getOperand(Elt.getResNo()).getNode();
5936 /// CombineConsecutiveLoads - build_pair (load, load) -> load
5937 /// if load locations are consecutive.
5938 SDValue DAGCombiner::CombineConsecutiveLoads(SDNode *N, EVT VT) {
5939 assert(N->getOpcode() == ISD::BUILD_PAIR);
5941 LoadSDNode *LD1 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 0));
5942 LoadSDNode *LD2 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 1));
5943 if (!LD1 || !LD2 || !ISD::isNON_EXTLoad(LD1) || !LD1->hasOneUse() ||
5944 LD1->getPointerInfo().getAddrSpace() !=
5945 LD2->getPointerInfo().getAddrSpace())
5947 EVT LD1VT = LD1->getValueType(0);
5949 if (ISD::isNON_EXTLoad(LD2) &&
5951 // If both are volatile this would reduce the number of volatile loads.
5952 // If one is volatile it might be ok, but play conservative and bail out.
5953 !LD1->isVolatile() &&
5954 !LD2->isVolatile() &&
5955 DAG.isConsecutiveLoad(LD2, LD1, LD1VT.getSizeInBits()/8, 1)) {
5956 unsigned Align = LD1->getAlignment();
5957 unsigned NewAlign = TLI.getDataLayout()->
5958 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
5960 if (NewAlign <= Align &&
5961 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)))
5962 return DAG.getLoad(VT, SDLoc(N), LD1->getChain(),
5963 LD1->getBasePtr(), LD1->getPointerInfo(),
5964 false, false, false, Align);
5970 SDValue DAGCombiner::visitBITCAST(SDNode *N) {
5971 SDValue N0 = N->getOperand(0);
5972 EVT VT = N->getValueType(0);
5974 // If the input is a BUILD_VECTOR with all constant elements, fold this now.
5975 // Only do this before legalize, since afterward the target may be depending
5976 // on the bitconvert.
5977 // First check to see if this is all constant.
5979 N0.getOpcode() == ISD::BUILD_VECTOR && N0.getNode()->hasOneUse() &&
5981 bool isSimple = cast<BuildVectorSDNode>(N0)->isConstant();
5983 EVT DestEltVT = N->getValueType(0).getVectorElementType();
5984 assert(!DestEltVT.isVector() &&
5985 "Element type of vector ValueType must not be vector!");
5987 return ConstantFoldBITCASTofBUILD_VECTOR(N0.getNode(), DestEltVT);
5990 // If the input is a constant, let getNode fold it.
5991 if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) {
5992 SDValue Res = DAG.getNode(ISD::BITCAST, SDLoc(N), VT, N0);
5993 if (Res.getNode() != N) {
5994 if (!LegalOperations ||
5995 TLI.isOperationLegal(Res.getNode()->getOpcode(), VT))
5998 // Folding it resulted in an illegal node, and it's too late to
5999 // do that. Clean up the old node and forego the transformation.
6000 // Ideally this won't happen very often, because instcombine
6001 // and the earlier dagcombine runs (where illegal nodes are
6002 // permitted) should have folded most of them already.
6003 DAG.DeleteNode(Res.getNode());
6007 // (conv (conv x, t1), t2) -> (conv x, t2)
6008 if (N0.getOpcode() == ISD::BITCAST)
6009 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT,
6012 // fold (conv (load x)) -> (load (conv*)x)
6013 // If the resultant load doesn't need a higher alignment than the original!
6014 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
6015 // Do not change the width of a volatile load.
6016 !cast<LoadSDNode>(N0)->isVolatile() &&
6017 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)) &&
6018 TLI.isLoadBitCastBeneficial(N0.getValueType(), VT)) {
6019 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6020 unsigned Align = TLI.getDataLayout()->
6021 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
6022 unsigned OrigAlign = LN0->getAlignment();
6024 if (Align <= OrigAlign) {
6025 SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(),
6026 LN0->getBasePtr(), LN0->getPointerInfo(),
6027 LN0->isVolatile(), LN0->isNonTemporal(),
6028 LN0->isInvariant(), OrigAlign,
6029 LN0->getTBAAInfo());
6031 CombineTo(N0.getNode(),
6032 DAG.getNode(ISD::BITCAST, SDLoc(N0),
6033 N0.getValueType(), Load),
6039 // fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit)
6040 // fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit))
6041 // This often reduces constant pool loads.
6042 if (((N0.getOpcode() == ISD::FNEG && !TLI.isFNegFree(N0.getValueType())) ||
6043 (N0.getOpcode() == ISD::FABS && !TLI.isFAbsFree(N0.getValueType()))) &&
6044 N0.getNode()->hasOneUse() && VT.isInteger() &&
6045 !VT.isVector() && !N0.getValueType().isVector()) {
6046 SDValue NewConv = DAG.getNode(ISD::BITCAST, SDLoc(N0), VT,
6048 AddToWorkList(NewConv.getNode());
6050 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
6051 if (N0.getOpcode() == ISD::FNEG)
6052 return DAG.getNode(ISD::XOR, SDLoc(N), VT,
6053 NewConv, DAG.getConstant(SignBit, VT));
6054 assert(N0.getOpcode() == ISD::FABS);
6055 return DAG.getNode(ISD::AND, SDLoc(N), VT,
6056 NewConv, DAG.getConstant(~SignBit, VT));
6059 // fold (bitconvert (fcopysign cst, x)) ->
6060 // (or (and (bitconvert x), sign), (and cst, (not sign)))
6061 // Note that we don't handle (copysign x, cst) because this can always be
6062 // folded to an fneg or fabs.
6063 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse() &&
6064 isa<ConstantFPSDNode>(N0.getOperand(0)) &&
6065 VT.isInteger() && !VT.isVector()) {
6066 unsigned OrigXWidth = N0.getOperand(1).getValueType().getSizeInBits();
6067 EVT IntXVT = EVT::getIntegerVT(*DAG.getContext(), OrigXWidth);
6068 if (isTypeLegal(IntXVT)) {
6069 SDValue X = DAG.getNode(ISD::BITCAST, SDLoc(N0),
6070 IntXVT, N0.getOperand(1));
6071 AddToWorkList(X.getNode());
6073 // If X has a different width than the result/lhs, sext it or truncate it.
6074 unsigned VTWidth = VT.getSizeInBits();
6075 if (OrigXWidth < VTWidth) {
6076 X = DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, X);
6077 AddToWorkList(X.getNode());
6078 } else if (OrigXWidth > VTWidth) {
6079 // To get the sign bit in the right place, we have to shift it right
6080 // before truncating.
6081 X = DAG.getNode(ISD::SRL, SDLoc(X),
6082 X.getValueType(), X,
6083 DAG.getConstant(OrigXWidth-VTWidth, X.getValueType()));
6084 AddToWorkList(X.getNode());
6085 X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X);
6086 AddToWorkList(X.getNode());
6089 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
6090 X = DAG.getNode(ISD::AND, SDLoc(X), VT,
6091 X, DAG.getConstant(SignBit, VT));
6092 AddToWorkList(X.getNode());
6094 SDValue Cst = DAG.getNode(ISD::BITCAST, SDLoc(N0),
6095 VT, N0.getOperand(0));
6096 Cst = DAG.getNode(ISD::AND, SDLoc(Cst), VT,
6097 Cst, DAG.getConstant(~SignBit, VT));
6098 AddToWorkList(Cst.getNode());
6100 return DAG.getNode(ISD::OR, SDLoc(N), VT, X, Cst);
6104 // bitconvert(build_pair(ld, ld)) -> ld iff load locations are consecutive.
6105 if (N0.getOpcode() == ISD::BUILD_PAIR) {
6106 SDValue CombineLD = CombineConsecutiveLoads(N0.getNode(), VT);
6107 if (CombineLD.getNode())
6114 SDValue DAGCombiner::visitBUILD_PAIR(SDNode *N) {
6115 EVT VT = N->getValueType(0);
6116 return CombineConsecutiveLoads(N, VT);
6119 /// ConstantFoldBITCASTofBUILD_VECTOR - We know that BV is a build_vector
6120 /// node with Constant, ConstantFP or Undef operands. DstEltVT indicates the
6121 /// destination element value type.
6122 SDValue DAGCombiner::
6123 ConstantFoldBITCASTofBUILD_VECTOR(SDNode *BV, EVT DstEltVT) {
6124 EVT SrcEltVT = BV->getValueType(0).getVectorElementType();
6126 // If this is already the right type, we're done.
6127 if (SrcEltVT == DstEltVT) return SDValue(BV, 0);
6129 unsigned SrcBitSize = SrcEltVT.getSizeInBits();
6130 unsigned DstBitSize = DstEltVT.getSizeInBits();
6132 // If this is a conversion of N elements of one type to N elements of another
6133 // type, convert each element. This handles FP<->INT cases.
6134 if (SrcBitSize == DstBitSize) {
6135 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
6136 BV->getValueType(0).getVectorNumElements());
6138 // Due to the FP element handling below calling this routine recursively,
6139 // we can end up with a scalar-to-vector node here.
6140 if (BV->getOpcode() == ISD::SCALAR_TO_VECTOR)
6141 return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(BV), VT,
6142 DAG.getNode(ISD::BITCAST, SDLoc(BV),
6143 DstEltVT, BV->getOperand(0)));
6145 SmallVector<SDValue, 8> Ops;
6146 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
6147 SDValue Op = BV->getOperand(i);
6148 // If the vector element type is not legal, the BUILD_VECTOR operands
6149 // are promoted and implicitly truncated. Make that explicit here.
6150 if (Op.getValueType() != SrcEltVT)
6151 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(BV), SrcEltVT, Op);
6152 Ops.push_back(DAG.getNode(ISD::BITCAST, SDLoc(BV),
6154 AddToWorkList(Ops.back().getNode());
6156 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT,
6157 &Ops[0], Ops.size());
6160 // Otherwise, we're growing or shrinking the elements. To avoid having to
6161 // handle annoying details of growing/shrinking FP values, we convert them to
6163 if (SrcEltVT.isFloatingPoint()) {
6164 // Convert the input float vector to a int vector where the elements are the
6166 assert((SrcEltVT == MVT::f32 || SrcEltVT == MVT::f64) && "Unknown FP VT!");
6167 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), SrcEltVT.getSizeInBits());
6168 BV = ConstantFoldBITCASTofBUILD_VECTOR(BV, IntVT).getNode();
6172 // Now we know the input is an integer vector. If the output is a FP type,
6173 // convert to integer first, then to FP of the right size.
6174 if (DstEltVT.isFloatingPoint()) {
6175 assert((DstEltVT == MVT::f32 || DstEltVT == MVT::f64) && "Unknown FP VT!");
6176 EVT TmpVT = EVT::getIntegerVT(*DAG.getContext(), DstEltVT.getSizeInBits());
6177 SDNode *Tmp = ConstantFoldBITCASTofBUILD_VECTOR(BV, TmpVT).getNode();
6179 // Next, convert to FP elements of the same size.
6180 return ConstantFoldBITCASTofBUILD_VECTOR(Tmp, DstEltVT);
6183 // Okay, we know the src/dst types are both integers of differing types.
6184 // Handling growing first.
6185 assert(SrcEltVT.isInteger() && DstEltVT.isInteger());
6186 if (SrcBitSize < DstBitSize) {
6187 unsigned NumInputsPerOutput = DstBitSize/SrcBitSize;
6189 SmallVector<SDValue, 8> Ops;
6190 for (unsigned i = 0, e = BV->getNumOperands(); i != e;
6191 i += NumInputsPerOutput) {
6192 bool isLE = TLI.isLittleEndian();
6193 APInt NewBits = APInt(DstBitSize, 0);
6194 bool EltIsUndef = true;
6195 for (unsigned j = 0; j != NumInputsPerOutput; ++j) {
6196 // Shift the previously computed bits over.
6197 NewBits <<= SrcBitSize;
6198 SDValue Op = BV->getOperand(i+ (isLE ? (NumInputsPerOutput-j-1) : j));
6199 if (Op.getOpcode() == ISD::UNDEF) continue;
6202 NewBits |= cast<ConstantSDNode>(Op)->getAPIntValue().
6203 zextOrTrunc(SrcBitSize).zext(DstBitSize);
6207 Ops.push_back(DAG.getUNDEF(DstEltVT));
6209 Ops.push_back(DAG.getConstant(NewBits, DstEltVT));
6212 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, Ops.size());
6213 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT,
6214 &Ops[0], Ops.size());
6217 // Finally, this must be the case where we are shrinking elements: each input
6218 // turns into multiple outputs.
6219 bool isS2V = ISD::isScalarToVector(BV);
6220 unsigned NumOutputsPerInput = SrcBitSize/DstBitSize;
6221 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
6222 NumOutputsPerInput*BV->getNumOperands());
6223 SmallVector<SDValue, 8> Ops;
6225 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
6226 if (BV->getOperand(i).getOpcode() == ISD::UNDEF) {
6227 for (unsigned j = 0; j != NumOutputsPerInput; ++j)
6228 Ops.push_back(DAG.getUNDEF(DstEltVT));
6232 APInt OpVal = cast<ConstantSDNode>(BV->getOperand(i))->
6233 getAPIntValue().zextOrTrunc(SrcBitSize);
6235 for (unsigned j = 0; j != NumOutputsPerInput; ++j) {
6236 APInt ThisVal = OpVal.trunc(DstBitSize);
6237 Ops.push_back(DAG.getConstant(ThisVal, DstEltVT));
6238 if (isS2V && i == 0 && j == 0 && ThisVal.zext(SrcBitSize) == OpVal)
6239 // Simply turn this into a SCALAR_TO_VECTOR of the new type.
6240 return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(BV), VT,
6242 OpVal = OpVal.lshr(DstBitSize);
6245 // For big endian targets, swap the order of the pieces of each element.
6246 if (TLI.isBigEndian())
6247 std::reverse(Ops.end()-NumOutputsPerInput, Ops.end());
6250 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT,
6251 &Ops[0], Ops.size());
6254 SDValue DAGCombiner::visitFADD(SDNode *N) {
6255 SDValue N0 = N->getOperand(0);
6256 SDValue N1 = N->getOperand(1);
6257 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6258 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6259 EVT VT = N->getValueType(0);
6262 if (VT.isVector()) {
6263 SDValue FoldedVOp = SimplifyVBinOp(N);
6264 if (FoldedVOp.getNode()) return FoldedVOp;
6267 // fold (fadd c1, c2) -> c1 + c2
6269 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N1);
6270 // canonicalize constant to RHS
6271 if (N0CFP && !N1CFP)
6272 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N1, N0);
6273 // fold (fadd A, 0) -> A
6274 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
6275 N1CFP->getValueAPF().isZero())
6277 // fold (fadd A, (fneg B)) -> (fsub A, B)
6278 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
6279 isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options) == 2)
6280 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N0,
6281 GetNegatedExpression(N1, DAG, LegalOperations));
6282 // fold (fadd (fneg A), B) -> (fsub B, A)
6283 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
6284 isNegatibleForFree(N0, LegalOperations, TLI, &DAG.getTarget().Options) == 2)
6285 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N1,
6286 GetNegatedExpression(N0, DAG, LegalOperations));
6288 // If allowed, fold (fadd (fadd x, c1), c2) -> (fadd x, (fadd c1, c2))
6289 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
6290 N0.getOpcode() == ISD::FADD && N0.getNode()->hasOneUse() &&
6291 isa<ConstantFPSDNode>(N0.getOperand(1)))
6292 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0.getOperand(0),
6293 DAG.getNode(ISD::FADD, SDLoc(N), VT,
6294 N0.getOperand(1), N1));
6296 // No FP constant should be created after legalization as Instruction
6297 // Selection pass has hard time in dealing with FP constant.
6299 // We don't need test this condition for transformation like following, as
6300 // the DAG being transformed implies it is legal to take FP constant as
6303 // (fadd (fmul c, x), x) -> (fmul c+1, x)
6305 bool AllowNewFpConst = (Level < AfterLegalizeDAG);
6307 // If allow, fold (fadd (fneg x), x) -> 0.0
6308 if (AllowNewFpConst && DAG.getTarget().Options.UnsafeFPMath &&
6309 N0.getOpcode() == ISD::FNEG && N0.getOperand(0) == N1)
6310 return DAG.getConstantFP(0.0, VT);
6312 // If allow, fold (fadd x, (fneg x)) -> 0.0
6313 if (AllowNewFpConst && DAG.getTarget().Options.UnsafeFPMath &&
6314 N1.getOpcode() == ISD::FNEG && N1.getOperand(0) == N0)
6315 return DAG.getConstantFP(0.0, VT);
6317 // In unsafe math mode, we can fold chains of FADD's of the same value
6318 // into multiplications. This transform is not safe in general because
6319 // we are reducing the number of rounding steps.
6320 if (DAG.getTarget().Options.UnsafeFPMath &&
6321 TLI.isOperationLegalOrCustom(ISD::FMUL, VT) &&
6323 if (N0.getOpcode() == ISD::FMUL) {
6324 ConstantFPSDNode *CFP00 = dyn_cast<ConstantFPSDNode>(N0.getOperand(0));
6325 ConstantFPSDNode *CFP01 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
6327 // (fadd (fmul c, x), x) -> (fmul x, c+1)
6328 if (CFP00 && !CFP01 && N0.getOperand(1) == N1) {
6329 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6331 DAG.getConstantFP(1.0, VT));
6332 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6336 // (fadd (fmul x, c), x) -> (fmul x, c+1)
6337 if (CFP01 && !CFP00 && N0.getOperand(0) == N1) {
6338 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6340 DAG.getConstantFP(1.0, VT));
6341 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6345 // (fadd (fmul c, x), (fadd x, x)) -> (fmul x, c+2)
6346 if (CFP00 && !CFP01 && N1.getOpcode() == ISD::FADD &&
6347 N1.getOperand(0) == N1.getOperand(1) &&
6348 N0.getOperand(1) == N1.getOperand(0)) {
6349 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6351 DAG.getConstantFP(2.0, VT));
6352 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6353 N0.getOperand(1), NewCFP);
6356 // (fadd (fmul x, c), (fadd x, x)) -> (fmul x, c+2)
6357 if (CFP01 && !CFP00 && N1.getOpcode() == ISD::FADD &&
6358 N1.getOperand(0) == N1.getOperand(1) &&
6359 N0.getOperand(0) == N1.getOperand(0)) {
6360 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6362 DAG.getConstantFP(2.0, VT));
6363 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6364 N0.getOperand(0), NewCFP);
6368 if (N1.getOpcode() == ISD::FMUL) {
6369 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0));
6370 ConstantFPSDNode *CFP11 = dyn_cast<ConstantFPSDNode>(N1.getOperand(1));
6372 // (fadd x, (fmul c, x)) -> (fmul x, c+1)
6373 if (CFP10 && !CFP11 && N1.getOperand(1) == N0) {
6374 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6376 DAG.getConstantFP(1.0, VT));
6377 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6381 // (fadd x, (fmul x, c)) -> (fmul x, c+1)
6382 if (CFP11 && !CFP10 && N1.getOperand(0) == N0) {
6383 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6385 DAG.getConstantFP(1.0, VT));
6386 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6391 // (fadd (fadd x, x), (fmul c, x)) -> (fmul x, c+2)
6392 if (CFP10 && !CFP11 && N0.getOpcode() == ISD::FADD &&
6393 N0.getOperand(0) == N0.getOperand(1) &&
6394 N1.getOperand(1) == N0.getOperand(0)) {
6395 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6397 DAG.getConstantFP(2.0, VT));
6398 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6399 N1.getOperand(1), NewCFP);
6402 // (fadd (fadd x, x), (fmul x, c)) -> (fmul x, c+2)
6403 if (CFP11 && !CFP10 && N0.getOpcode() == ISD::FADD &&
6404 N0.getOperand(0) == N0.getOperand(1) &&
6405 N1.getOperand(0) == N0.getOperand(0)) {
6406 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6408 DAG.getConstantFP(2.0, VT));
6409 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6410 N1.getOperand(0), NewCFP);
6414 if (N0.getOpcode() == ISD::FADD && AllowNewFpConst) {
6415 ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N0.getOperand(0));
6416 // (fadd (fadd x, x), x) -> (fmul x, 3.0)
6417 if (!CFP && N0.getOperand(0) == N0.getOperand(1) &&
6418 (N0.getOperand(0) == N1))
6419 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6420 N1, DAG.getConstantFP(3.0, VT));
6423 if (N1.getOpcode() == ISD::FADD && AllowNewFpConst) {
6424 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0));
6425 // (fadd x, (fadd x, x)) -> (fmul x, 3.0)
6426 if (!CFP10 && N1.getOperand(0) == N1.getOperand(1) &&
6427 N1.getOperand(0) == N0)
6428 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6429 N0, DAG.getConstantFP(3.0, VT));
6432 // (fadd (fadd x, x), (fadd x, x)) -> (fmul x, 4.0)
6433 if (AllowNewFpConst &&
6434 N0.getOpcode() == ISD::FADD && N1.getOpcode() == ISD::FADD &&
6435 N0.getOperand(0) == N0.getOperand(1) &&
6436 N1.getOperand(0) == N1.getOperand(1) &&
6437 N0.getOperand(0) == N1.getOperand(0))
6438 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6440 DAG.getConstantFP(4.0, VT));
6443 // FADD -> FMA combines:
6444 if ((DAG.getTarget().Options.AllowFPOpFusion == FPOpFusion::Fast ||
6445 DAG.getTarget().Options.UnsafeFPMath) &&
6446 DAG.getTarget().getTargetLowering()->isFMAFasterThanFMulAndFAdd(VT) &&
6447 (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FMA, VT))) {
6449 // fold (fadd (fmul x, y), z) -> (fma x, y, z)
6450 if (N0.getOpcode() == ISD::FMUL && N0->hasOneUse())
6451 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
6452 N0.getOperand(0), N0.getOperand(1), N1);
6454 // fold (fadd x, (fmul y, z)) -> (fma y, z, x)
6455 // Note: Commutes FADD operands.
6456 if (N1.getOpcode() == ISD::FMUL && N1->hasOneUse())
6457 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
6458 N1.getOperand(0), N1.getOperand(1), N0);
6464 SDValue DAGCombiner::visitFSUB(SDNode *N) {
6465 SDValue N0 = N->getOperand(0);
6466 SDValue N1 = N->getOperand(1);
6467 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6468 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6469 EVT VT = N->getValueType(0);
6473 if (VT.isVector()) {
6474 SDValue FoldedVOp = SimplifyVBinOp(N);
6475 if (FoldedVOp.getNode()) return FoldedVOp;
6478 // fold (fsub c1, c2) -> c1-c2
6480 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N0, N1);
6481 // fold (fsub A, 0) -> A
6482 if (DAG.getTarget().Options.UnsafeFPMath &&
6483 N1CFP && N1CFP->getValueAPF().isZero())
6485 // fold (fsub 0, B) -> -B
6486 if (DAG.getTarget().Options.UnsafeFPMath &&
6487 N0CFP && N0CFP->getValueAPF().isZero()) {
6488 if (isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options))
6489 return GetNegatedExpression(N1, DAG, LegalOperations);
6490 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
6491 return DAG.getNode(ISD::FNEG, dl, VT, N1);
6493 // fold (fsub A, (fneg B)) -> (fadd A, B)
6494 if (isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options))
6495 return DAG.getNode(ISD::FADD, dl, VT, N0,
6496 GetNegatedExpression(N1, DAG, LegalOperations));
6498 // If 'unsafe math' is enabled, fold
6499 // (fsub x, x) -> 0.0 &
6500 // (fsub x, (fadd x, y)) -> (fneg y) &
6501 // (fsub x, (fadd y, x)) -> (fneg y)
6502 if (DAG.getTarget().Options.UnsafeFPMath) {
6504 return DAG.getConstantFP(0.0f, VT);
6506 if (N1.getOpcode() == ISD::FADD) {
6507 SDValue N10 = N1->getOperand(0);
6508 SDValue N11 = N1->getOperand(1);
6510 if (N10 == N0 && isNegatibleForFree(N11, LegalOperations, TLI,
6511 &DAG.getTarget().Options))
6512 return GetNegatedExpression(N11, DAG, LegalOperations);
6514 if (N11 == N0 && isNegatibleForFree(N10, LegalOperations, TLI,
6515 &DAG.getTarget().Options))
6516 return GetNegatedExpression(N10, DAG, LegalOperations);
6520 // FSUB -> FMA combines:
6521 if ((DAG.getTarget().Options.AllowFPOpFusion == FPOpFusion::Fast ||
6522 DAG.getTarget().Options.UnsafeFPMath) &&
6523 DAG.getTarget().getTargetLowering()->isFMAFasterThanFMulAndFAdd(VT) &&
6524 (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FMA, VT))) {
6526 // fold (fsub (fmul x, y), z) -> (fma x, y, (fneg z))
6527 if (N0.getOpcode() == ISD::FMUL && N0->hasOneUse())
6528 return DAG.getNode(ISD::FMA, dl, VT,
6529 N0.getOperand(0), N0.getOperand(1),
6530 DAG.getNode(ISD::FNEG, dl, VT, N1));
6532 // fold (fsub x, (fmul y, z)) -> (fma (fneg y), z, x)
6533 // Note: Commutes FSUB operands.
6534 if (N1.getOpcode() == ISD::FMUL && N1->hasOneUse())
6535 return DAG.getNode(ISD::FMA, dl, VT,
6536 DAG.getNode(ISD::FNEG, dl, VT,
6538 N1.getOperand(1), N0);
6540 // fold (fsub (fneg (fmul, x, y)), z) -> (fma (fneg x), y, (fneg z))
6541 if (N0.getOpcode() == ISD::FNEG &&
6542 N0.getOperand(0).getOpcode() == ISD::FMUL &&
6543 N0->hasOneUse() && N0.getOperand(0).hasOneUse()) {
6544 SDValue N00 = N0.getOperand(0).getOperand(0);
6545 SDValue N01 = N0.getOperand(0).getOperand(1);
6546 return DAG.getNode(ISD::FMA, dl, VT,
6547 DAG.getNode(ISD::FNEG, dl, VT, N00), N01,
6548 DAG.getNode(ISD::FNEG, dl, VT, N1));
6555 SDValue DAGCombiner::visitFMUL(SDNode *N) {
6556 SDValue N0 = N->getOperand(0);
6557 SDValue N1 = N->getOperand(1);
6558 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6559 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6560 EVT VT = N->getValueType(0);
6561 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6564 if (VT.isVector()) {
6565 SDValue FoldedVOp = SimplifyVBinOp(N);
6566 if (FoldedVOp.getNode()) return FoldedVOp;
6569 // fold (fmul c1, c2) -> c1*c2
6571 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0, N1);
6572 // canonicalize constant to RHS
6573 if (N0CFP && !N1CFP)
6574 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N1, N0);
6575 // fold (fmul A, 0) -> 0
6576 if (DAG.getTarget().Options.UnsafeFPMath &&
6577 N1CFP && N1CFP->getValueAPF().isZero())
6579 // fold (fmul A, 0) -> 0, vector edition.
6580 if (DAG.getTarget().Options.UnsafeFPMath &&
6581 ISD::isBuildVectorAllZeros(N1.getNode()))
6583 // fold (fmul A, 1.0) -> A
6584 if (N1CFP && N1CFP->isExactlyValue(1.0))
6586 // fold (fmul X, 2.0) -> (fadd X, X)
6587 if (N1CFP && N1CFP->isExactlyValue(+2.0))
6588 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N0);
6589 // fold (fmul X, -1.0) -> (fneg X)
6590 if (N1CFP && N1CFP->isExactlyValue(-1.0))
6591 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
6592 return DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0);
6594 // fold (fmul (fneg X), (fneg Y)) -> (fmul X, Y)
6595 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI,
6596 &DAG.getTarget().Options)) {
6597 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI,
6598 &DAG.getTarget().Options)) {
6599 // Both can be negated for free, check to see if at least one is cheaper
6601 if (LHSNeg == 2 || RHSNeg == 2)
6602 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6603 GetNegatedExpression(N0, DAG, LegalOperations),
6604 GetNegatedExpression(N1, DAG, LegalOperations));
6608 // If allowed, fold (fmul (fmul x, c1), c2) -> (fmul x, (fmul c1, c2))
6609 if (DAG.getTarget().Options.UnsafeFPMath &&
6610 N1CFP && N0.getOpcode() == ISD::FMUL &&
6611 N0.getNode()->hasOneUse() && isa<ConstantFPSDNode>(N0.getOperand(1)))
6612 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0.getOperand(0),
6613 DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6614 N0.getOperand(1), N1));
6619 SDValue DAGCombiner::visitFMA(SDNode *N) {
6620 SDValue N0 = N->getOperand(0);
6621 SDValue N1 = N->getOperand(1);
6622 SDValue N2 = N->getOperand(2);
6623 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6624 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6625 EVT VT = N->getValueType(0);
6628 if (DAG.getTarget().Options.UnsafeFPMath) {
6629 if (N0CFP && N0CFP->isZero())
6631 if (N1CFP && N1CFP->isZero())
6634 if (N0CFP && N0CFP->isExactlyValue(1.0))
6635 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N1, N2);
6636 if (N1CFP && N1CFP->isExactlyValue(1.0))
6637 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N2);
6639 // Canonicalize (fma c, x, y) -> (fma x, c, y)
6640 if (N0CFP && !N1CFP)
6641 return DAG.getNode(ISD::FMA, SDLoc(N), VT, N1, N0, N2);
6643 // (fma x, c1, (fmul x, c2)) -> (fmul x, c1+c2)
6644 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
6645 N2.getOpcode() == ISD::FMUL &&
6646 N0 == N2.getOperand(0) &&
6647 N2.getOperand(1).getOpcode() == ISD::ConstantFP) {
6648 return DAG.getNode(ISD::FMUL, dl, VT, N0,
6649 DAG.getNode(ISD::FADD, dl, VT, N1, N2.getOperand(1)));
6653 // (fma (fmul x, c1), c2, y) -> (fma x, c1*c2, y)
6654 if (DAG.getTarget().Options.UnsafeFPMath &&
6655 N0.getOpcode() == ISD::FMUL && N1CFP &&
6656 N0.getOperand(1).getOpcode() == ISD::ConstantFP) {
6657 return DAG.getNode(ISD::FMA, dl, VT,
6659 DAG.getNode(ISD::FMUL, dl, VT, N1, N0.getOperand(1)),
6663 // (fma x, 1, y) -> (fadd x, y)
6664 // (fma x, -1, y) -> (fadd (fneg x), y)
6666 if (N1CFP->isExactlyValue(1.0))
6667 return DAG.getNode(ISD::FADD, dl, VT, N0, N2);
6669 if (N1CFP->isExactlyValue(-1.0) &&
6670 (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))) {
6671 SDValue RHSNeg = DAG.getNode(ISD::FNEG, dl, VT, N0);
6672 AddToWorkList(RHSNeg.getNode());
6673 return DAG.getNode(ISD::FADD, dl, VT, N2, RHSNeg);
6677 // (fma x, c, x) -> (fmul x, (c+1))
6678 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP && N0 == N2)
6679 return DAG.getNode(ISD::FMUL, dl, VT, N0,
6680 DAG.getNode(ISD::FADD, dl, VT,
6681 N1, DAG.getConstantFP(1.0, VT)));
6683 // (fma x, c, (fneg x)) -> (fmul x, (c-1))
6684 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
6685 N2.getOpcode() == ISD::FNEG && N2.getOperand(0) == N0)
6686 return DAG.getNode(ISD::FMUL, dl, VT, N0,
6687 DAG.getNode(ISD::FADD, dl, VT,
6688 N1, DAG.getConstantFP(-1.0, VT)));
6694 SDValue DAGCombiner::visitFDIV(SDNode *N) {
6695 SDValue N0 = N->getOperand(0);
6696 SDValue N1 = N->getOperand(1);
6697 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6698 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6699 EVT VT = N->getValueType(0);
6700 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6703 if (VT.isVector()) {
6704 SDValue FoldedVOp = SimplifyVBinOp(N);
6705 if (FoldedVOp.getNode()) return FoldedVOp;
6708 // fold (fdiv c1, c2) -> c1/c2
6710 return DAG.getNode(ISD::FDIV, SDLoc(N), VT, N0, N1);
6712 // fold (fdiv X, c2) -> fmul X, 1/c2 if losing precision is acceptable.
6713 if (N1CFP && DAG.getTarget().Options.UnsafeFPMath) {
6714 // Compute the reciprocal 1.0 / c2.
6715 APFloat N1APF = N1CFP->getValueAPF();
6716 APFloat Recip(N1APF.getSemantics(), 1); // 1.0
6717 APFloat::opStatus st = Recip.divide(N1APF, APFloat::rmNearestTiesToEven);
6718 // Only do the transform if the reciprocal is a legal fp immediate that
6719 // isn't too nasty (eg NaN, denormal, ...).
6720 if ((st == APFloat::opOK || st == APFloat::opInexact) && // Not too nasty
6721 (!LegalOperations ||
6722 // FIXME: custom lowering of ConstantFP might fail (see e.g. ARM
6723 // backend)... we should handle this gracefully after Legalize.
6724 // TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT) ||
6725 TLI.isOperationLegal(llvm::ISD::ConstantFP, VT) ||
6726 TLI.isFPImmLegal(Recip, VT)))
6727 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0,
6728 DAG.getConstantFP(Recip, VT));
6731 // (fdiv (fneg X), (fneg Y)) -> (fdiv X, Y)
6732 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI,
6733 &DAG.getTarget().Options)) {
6734 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI,
6735 &DAG.getTarget().Options)) {
6736 // Both can be negated for free, check to see if at least one is cheaper
6738 if (LHSNeg == 2 || RHSNeg == 2)
6739 return DAG.getNode(ISD::FDIV, SDLoc(N), VT,
6740 GetNegatedExpression(N0, DAG, LegalOperations),
6741 GetNegatedExpression(N1, DAG, LegalOperations));
6748 SDValue DAGCombiner::visitFREM(SDNode *N) {
6749 SDValue N0 = N->getOperand(0);
6750 SDValue N1 = N->getOperand(1);
6751 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6752 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6753 EVT VT = N->getValueType(0);
6755 // fold (frem c1, c2) -> fmod(c1,c2)
6757 return DAG.getNode(ISD::FREM, SDLoc(N), VT, N0, N1);
6762 SDValue DAGCombiner::visitFCOPYSIGN(SDNode *N) {
6763 SDValue N0 = N->getOperand(0);
6764 SDValue N1 = N->getOperand(1);
6765 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6766 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6767 EVT VT = N->getValueType(0);
6769 if (N0CFP && N1CFP) // Constant fold
6770 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT, N0, N1);
6773 const APFloat& V = N1CFP->getValueAPF();
6774 // copysign(x, c1) -> fabs(x) iff ispos(c1)
6775 // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1)
6776 if (!V.isNegative()) {
6777 if (!LegalOperations || TLI.isOperationLegal(ISD::FABS, VT))
6778 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
6780 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
6781 return DAG.getNode(ISD::FNEG, SDLoc(N), VT,
6782 DAG.getNode(ISD::FABS, SDLoc(N0), VT, N0));
6786 // copysign(fabs(x), y) -> copysign(x, y)
6787 // copysign(fneg(x), y) -> copysign(x, y)
6788 // copysign(copysign(x,z), y) -> copysign(x, y)
6789 if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG ||
6790 N0.getOpcode() == ISD::FCOPYSIGN)
6791 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
6792 N0.getOperand(0), N1);
6794 // copysign(x, abs(y)) -> abs(x)
6795 if (N1.getOpcode() == ISD::FABS)
6796 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
6798 // copysign(x, copysign(y,z)) -> copysign(x, z)
6799 if (N1.getOpcode() == ISD::FCOPYSIGN)
6800 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
6801 N0, N1.getOperand(1));
6803 // copysign(x, fp_extend(y)) -> copysign(x, y)
6804 // copysign(x, fp_round(y)) -> copysign(x, y)
6805 if (N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND)
6806 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
6807 N0, N1.getOperand(0));
6812 SDValue DAGCombiner::visitSINT_TO_FP(SDNode *N) {
6813 SDValue N0 = N->getOperand(0);
6814 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
6815 EVT VT = N->getValueType(0);
6816 EVT OpVT = N0.getValueType();
6818 // fold (sint_to_fp c1) -> c1fp
6820 // ...but only if the target supports immediate floating-point values
6821 (!LegalOperations ||
6822 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
6823 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0);
6825 // If the input is a legal type, and SINT_TO_FP is not legal on this target,
6826 // but UINT_TO_FP is legal on this target, try to convert.
6827 if (!TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT) &&
6828 TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT)) {
6829 // If the sign bit is known to be zero, we can change this to UINT_TO_FP.
6830 if (DAG.SignBitIsZero(N0))
6831 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0);
6834 // The next optimizations are desirable only if SELECT_CC can be lowered.
6835 // Check against MVT::Other for SELECT_CC, which is a workaround for targets
6836 // having to say they don't support SELECT_CC on every type the DAG knows
6837 // about, since there is no way to mark an opcode illegal at all value types
6838 // (See also visitSELECT)
6839 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, MVT::Other)) {
6840 // fold (sint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
6841 if (N0.getOpcode() == ISD::SETCC && N0.getValueType() == MVT::i1 &&
6843 (!LegalOperations ||
6844 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
6846 { N0.getOperand(0), N0.getOperand(1),
6847 DAG.getConstantFP(-1.0, VT) , DAG.getConstantFP(0.0, VT),
6849 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops, 5);
6852 // fold (sint_to_fp (zext (setcc x, y, cc))) ->
6853 // (select_cc x, y, 1.0, 0.0,, cc)
6854 if (N0.getOpcode() == ISD::ZERO_EXTEND &&
6855 N0.getOperand(0).getOpcode() == ISD::SETCC &&!VT.isVector() &&
6856 (!LegalOperations ||
6857 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
6859 { N0.getOperand(0).getOperand(0), N0.getOperand(0).getOperand(1),
6860 DAG.getConstantFP(1.0, VT) , DAG.getConstantFP(0.0, VT),
6861 N0.getOperand(0).getOperand(2) };
6862 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops, 5);
6869 SDValue DAGCombiner::visitUINT_TO_FP(SDNode *N) {
6870 SDValue N0 = N->getOperand(0);
6871 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
6872 EVT VT = N->getValueType(0);
6873 EVT OpVT = N0.getValueType();
6875 // fold (uint_to_fp c1) -> c1fp
6877 // ...but only if the target supports immediate floating-point values
6878 (!LegalOperations ||
6879 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
6880 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0);
6882 // If the input is a legal type, and UINT_TO_FP is not legal on this target,
6883 // but SINT_TO_FP is legal on this target, try to convert.
6884 if (!TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT) &&
6885 TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT)) {
6886 // If the sign bit is known to be zero, we can change this to SINT_TO_FP.
6887 if (DAG.SignBitIsZero(N0))
6888 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0);
6891 // The next optimizations are desirable only if SELECT_CC can be lowered.
6892 // Check against MVT::Other for SELECT_CC, which is a workaround for targets
6893 // having to say they don't support SELECT_CC on every type the DAG knows
6894 // about, since there is no way to mark an opcode illegal at all value types
6895 // (See also visitSELECT)
6896 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, MVT::Other)) {
6897 // fold (uint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
6899 if (N0.getOpcode() == ISD::SETCC && !VT.isVector() &&
6900 (!LegalOperations ||
6901 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
6903 { N0.getOperand(0), N0.getOperand(1),
6904 DAG.getConstantFP(1.0, VT), DAG.getConstantFP(0.0, VT),
6906 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops, 5);
6913 SDValue DAGCombiner::visitFP_TO_SINT(SDNode *N) {
6914 SDValue N0 = N->getOperand(0);
6915 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6916 EVT VT = N->getValueType(0);
6918 // fold (fp_to_sint c1fp) -> c1
6920 return DAG.getNode(ISD::FP_TO_SINT, SDLoc(N), VT, N0);
6925 SDValue DAGCombiner::visitFP_TO_UINT(SDNode *N) {
6926 SDValue N0 = N->getOperand(0);
6927 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6928 EVT VT = N->getValueType(0);
6930 // fold (fp_to_uint c1fp) -> c1
6932 return DAG.getNode(ISD::FP_TO_UINT, SDLoc(N), VT, N0);
6937 SDValue DAGCombiner::visitFP_ROUND(SDNode *N) {
6938 SDValue N0 = N->getOperand(0);
6939 SDValue N1 = N->getOperand(1);
6940 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6941 EVT VT = N->getValueType(0);
6943 // fold (fp_round c1fp) -> c1fp
6945 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, N0, N1);
6947 // fold (fp_round (fp_extend x)) -> x
6948 if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType())
6949 return N0.getOperand(0);
6951 // fold (fp_round (fp_round x)) -> (fp_round x)
6952 if (N0.getOpcode() == ISD::FP_ROUND) {
6953 // This is a value preserving truncation if both round's are.
6954 bool IsTrunc = N->getConstantOperandVal(1) == 1 &&
6955 N0.getNode()->getConstantOperandVal(1) == 1;
6956 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, N0.getOperand(0),
6957 DAG.getIntPtrConstant(IsTrunc));
6960 // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y)
6961 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse()) {
6962 SDValue Tmp = DAG.getNode(ISD::FP_ROUND, SDLoc(N0), VT,
6963 N0.getOperand(0), N1);
6964 AddToWorkList(Tmp.getNode());
6965 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
6966 Tmp, N0.getOperand(1));
6972 SDValue DAGCombiner::visitFP_ROUND_INREG(SDNode *N) {
6973 SDValue N0 = N->getOperand(0);
6974 EVT VT = N->getValueType(0);
6975 EVT EVT = cast<VTSDNode>(N->getOperand(1))->getVT();
6976 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6978 // fold (fp_round_inreg c1fp) -> c1fp
6979 if (N0CFP && isTypeLegal(EVT)) {
6980 SDValue Round = DAG.getConstantFP(*N0CFP->getConstantFPValue(), EVT);
6981 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, Round);
6987 SDValue DAGCombiner::visitFP_EXTEND(SDNode *N) {
6988 SDValue N0 = N->getOperand(0);
6989 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6990 EVT VT = N->getValueType(0);
6992 // If this is fp_round(fpextend), don't fold it, allow ourselves to be folded.
6993 if (N->hasOneUse() &&
6994 N->use_begin()->getOpcode() == ISD::FP_ROUND)
6997 // fold (fp_extend c1fp) -> c1fp
6999 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, N0);
7001 // Turn fp_extend(fp_round(X, 1)) -> x since the fp_round doesn't affect the
7003 if (N0.getOpcode() == ISD::FP_ROUND
7004 && N0.getNode()->getConstantOperandVal(1) == 1) {
7005 SDValue In = N0.getOperand(0);
7006 if (In.getValueType() == VT) return In;
7007 if (VT.bitsLT(In.getValueType()))
7008 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT,
7009 In, N0.getOperand(1));
7010 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, In);
7013 // fold (fpext (load x)) -> (fpext (fptrunc (extload x)))
7014 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
7015 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
7016 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType()))) {
7017 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
7018 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT,
7020 LN0->getBasePtr(), N0.getValueType(),
7021 LN0->getMemOperand());
7022 CombineTo(N, ExtLoad);
7023 CombineTo(N0.getNode(),
7024 DAG.getNode(ISD::FP_ROUND, SDLoc(N0),
7025 N0.getValueType(), ExtLoad, DAG.getIntPtrConstant(1)),
7026 ExtLoad.getValue(1));
7027 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7033 SDValue DAGCombiner::visitFNEG(SDNode *N) {
7034 SDValue N0 = N->getOperand(0);
7035 EVT VT = N->getValueType(0);
7037 if (VT.isVector()) {
7038 SDValue FoldedVOp = SimplifyVUnaryOp(N);
7039 if (FoldedVOp.getNode()) return FoldedVOp;
7042 if (isNegatibleForFree(N0, LegalOperations, DAG.getTargetLoweringInfo(),
7043 &DAG.getTarget().Options))
7044 return GetNegatedExpression(N0, DAG, LegalOperations);
7046 // Transform fneg(bitconvert(x)) -> bitconvert(x^sign) to avoid loading
7047 // constant pool values.
7048 if (!TLI.isFNegFree(VT) && N0.getOpcode() == ISD::BITCAST &&
7050 N0.getNode()->hasOneUse() &&
7051 N0.getOperand(0).getValueType().isInteger()) {
7052 SDValue Int = N0.getOperand(0);
7053 EVT IntVT = Int.getValueType();
7054 if (IntVT.isInteger() && !IntVT.isVector()) {
7055 Int = DAG.getNode(ISD::XOR, SDLoc(N0), IntVT, Int,
7056 DAG.getConstant(APInt::getSignBit(IntVT.getSizeInBits()), IntVT));
7057 AddToWorkList(Int.getNode());
7058 return DAG.getNode(ISD::BITCAST, SDLoc(N),
7063 // (fneg (fmul c, x)) -> (fmul -c, x)
7064 if (N0.getOpcode() == ISD::FMUL) {
7065 ConstantFPSDNode *CFP1 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
7067 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
7069 DAG.getNode(ISD::FNEG, SDLoc(N), VT,
7076 SDValue DAGCombiner::visitFCEIL(SDNode *N) {
7077 SDValue N0 = N->getOperand(0);
7078 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7079 EVT VT = N->getValueType(0);
7081 // fold (fceil c1) -> fceil(c1)
7083 return DAG.getNode(ISD::FCEIL, SDLoc(N), VT, N0);
7088 SDValue DAGCombiner::visitFTRUNC(SDNode *N) {
7089 SDValue N0 = N->getOperand(0);
7090 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7091 EVT VT = N->getValueType(0);
7093 // fold (ftrunc c1) -> ftrunc(c1)
7095 return DAG.getNode(ISD::FTRUNC, SDLoc(N), VT, N0);
7100 SDValue DAGCombiner::visitFFLOOR(SDNode *N) {
7101 SDValue N0 = N->getOperand(0);
7102 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7103 EVT VT = N->getValueType(0);
7105 // fold (ffloor c1) -> ffloor(c1)
7107 return DAG.getNode(ISD::FFLOOR, SDLoc(N), VT, N0);
7112 SDValue DAGCombiner::visitFABS(SDNode *N) {
7113 SDValue N0 = N->getOperand(0);
7114 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7115 EVT VT = N->getValueType(0);
7117 if (VT.isVector()) {
7118 SDValue FoldedVOp = SimplifyVUnaryOp(N);
7119 if (FoldedVOp.getNode()) return FoldedVOp;
7122 // fold (fabs c1) -> fabs(c1)
7124 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
7125 // fold (fabs (fabs x)) -> (fabs x)
7126 if (N0.getOpcode() == ISD::FABS)
7127 return N->getOperand(0);
7128 // fold (fabs (fneg x)) -> (fabs x)
7129 // fold (fabs (fcopysign x, y)) -> (fabs x)
7130 if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN)
7131 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0.getOperand(0));
7133 // Transform fabs(bitconvert(x)) -> bitconvert(x&~sign) to avoid loading
7134 // constant pool values.
7135 if (!TLI.isFAbsFree(VT) &&
7136 N0.getOpcode() == ISD::BITCAST && N0.getNode()->hasOneUse() &&
7137 N0.getOperand(0).getValueType().isInteger() &&
7138 !N0.getOperand(0).getValueType().isVector()) {
7139 SDValue Int = N0.getOperand(0);
7140 EVT IntVT = Int.getValueType();
7141 if (IntVT.isInteger() && !IntVT.isVector()) {
7142 Int = DAG.getNode(ISD::AND, SDLoc(N0), IntVT, Int,
7143 DAG.getConstant(~APInt::getSignBit(IntVT.getSizeInBits()), IntVT));
7144 AddToWorkList(Int.getNode());
7145 return DAG.getNode(ISD::BITCAST, SDLoc(N),
7146 N->getValueType(0), Int);
7153 SDValue DAGCombiner::visitBRCOND(SDNode *N) {
7154 SDValue Chain = N->getOperand(0);
7155 SDValue N1 = N->getOperand(1);
7156 SDValue N2 = N->getOperand(2);
7158 // If N is a constant we could fold this into a fallthrough or unconditional
7159 // branch. However that doesn't happen very often in normal code, because
7160 // Instcombine/SimplifyCFG should have handled the available opportunities.
7161 // If we did this folding here, it would be necessary to update the
7162 // MachineBasicBlock CFG, which is awkward.
7164 // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal
7166 if (N1.getOpcode() == ISD::SETCC &&
7167 TLI.isOperationLegalOrCustom(ISD::BR_CC,
7168 N1.getOperand(0).getValueType())) {
7169 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other,
7170 Chain, N1.getOperand(2),
7171 N1.getOperand(0), N1.getOperand(1), N2);
7174 if ((N1.hasOneUse() && N1.getOpcode() == ISD::SRL) ||
7175 ((N1.getOpcode() == ISD::TRUNCATE && N1.hasOneUse()) &&
7176 (N1.getOperand(0).hasOneUse() &&
7177 N1.getOperand(0).getOpcode() == ISD::SRL))) {
7179 if (N1.getOpcode() == ISD::TRUNCATE) {
7180 // Look pass the truncate.
7181 Trunc = N1.getNode();
7182 N1 = N1.getOperand(0);
7185 // Match this pattern so that we can generate simpler code:
7188 // %b = and i32 %a, 2
7189 // %c = srl i32 %b, 1
7190 // brcond i32 %c ...
7195 // %b = and i32 %a, 2
7196 // %c = setcc eq %b, 0
7199 // This applies only when the AND constant value has one bit set and the
7200 // SRL constant is equal to the log2 of the AND constant. The back-end is
7201 // smart enough to convert the result into a TEST/JMP sequence.
7202 SDValue Op0 = N1.getOperand(0);
7203 SDValue Op1 = N1.getOperand(1);
7205 if (Op0.getOpcode() == ISD::AND &&
7206 Op1.getOpcode() == ISD::Constant) {
7207 SDValue AndOp1 = Op0.getOperand(1);
7209 if (AndOp1.getOpcode() == ISD::Constant) {
7210 const APInt &AndConst = cast<ConstantSDNode>(AndOp1)->getAPIntValue();
7212 if (AndConst.isPowerOf2() &&
7213 cast<ConstantSDNode>(Op1)->getAPIntValue()==AndConst.logBase2()) {
7215 DAG.getSetCC(SDLoc(N),
7216 getSetCCResultType(Op0.getValueType()),
7217 Op0, DAG.getConstant(0, Op0.getValueType()),
7220 SDValue NewBRCond = DAG.getNode(ISD::BRCOND, SDLoc(N),
7221 MVT::Other, Chain, SetCC, N2);
7222 // Don't add the new BRCond into the worklist or else SimplifySelectCC
7223 // will convert it back to (X & C1) >> C2.
7224 CombineTo(N, NewBRCond, false);
7225 // Truncate is dead.
7227 removeFromWorkList(Trunc);
7228 DAG.DeleteNode(Trunc);
7230 // Replace the uses of SRL with SETCC
7231 WorkListRemover DeadNodes(*this);
7232 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
7233 removeFromWorkList(N1.getNode());
7234 DAG.DeleteNode(N1.getNode());
7235 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7241 // Restore N1 if the above transformation doesn't match.
7242 N1 = N->getOperand(1);
7245 // Transform br(xor(x, y)) -> br(x != y)
7246 // Transform br(xor(xor(x,y), 1)) -> br (x == y)
7247 if (N1.hasOneUse() && N1.getOpcode() == ISD::XOR) {
7248 SDNode *TheXor = N1.getNode();
7249 SDValue Op0 = TheXor->getOperand(0);
7250 SDValue Op1 = TheXor->getOperand(1);
7251 if (Op0.getOpcode() == Op1.getOpcode()) {
7252 // Avoid missing important xor optimizations.
7253 SDValue Tmp = visitXOR(TheXor);
7254 if (Tmp.getNode()) {
7255 if (Tmp.getNode() != TheXor) {
7256 DEBUG(dbgs() << "\nReplacing.8 ";
7258 dbgs() << "\nWith: ";
7259 Tmp.getNode()->dump(&DAG);
7261 WorkListRemover DeadNodes(*this);
7262 DAG.ReplaceAllUsesOfValueWith(N1, Tmp);
7263 removeFromWorkList(TheXor);
7264 DAG.DeleteNode(TheXor);
7265 return DAG.getNode(ISD::BRCOND, SDLoc(N),
7266 MVT::Other, Chain, Tmp, N2);
7269 // visitXOR has changed XOR's operands or replaced the XOR completely,
7271 return SDValue(N, 0);
7275 if (Op0.getOpcode() != ISD::SETCC && Op1.getOpcode() != ISD::SETCC) {
7277 if (ConstantSDNode *RHSCI = dyn_cast<ConstantSDNode>(Op0))
7278 if (RHSCI->getAPIntValue() == 1 && Op0.hasOneUse() &&
7279 Op0.getOpcode() == ISD::XOR) {
7280 TheXor = Op0.getNode();
7284 EVT SetCCVT = N1.getValueType();
7286 SetCCVT = getSetCCResultType(SetCCVT);
7287 SDValue SetCC = DAG.getSetCC(SDLoc(TheXor),
7290 Equal ? ISD::SETEQ : ISD::SETNE);
7291 // Replace the uses of XOR with SETCC
7292 WorkListRemover DeadNodes(*this);
7293 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
7294 removeFromWorkList(N1.getNode());
7295 DAG.DeleteNode(N1.getNode());
7296 return DAG.getNode(ISD::BRCOND, SDLoc(N),
7297 MVT::Other, Chain, SetCC, N2);
7304 // Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB.
7306 SDValue DAGCombiner::visitBR_CC(SDNode *N) {
7307 CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1));
7308 SDValue CondLHS = N->getOperand(2), CondRHS = N->getOperand(3);
7310 // If N is a constant we could fold this into a fallthrough or unconditional
7311 // branch. However that doesn't happen very often in normal code, because
7312 // Instcombine/SimplifyCFG should have handled the available opportunities.
7313 // If we did this folding here, it would be necessary to update the
7314 // MachineBasicBlock CFG, which is awkward.
7316 // Use SimplifySetCC to simplify SETCC's.
7317 SDValue Simp = SimplifySetCC(getSetCCResultType(CondLHS.getValueType()),
7318 CondLHS, CondRHS, CC->get(), SDLoc(N),
7320 if (Simp.getNode()) AddToWorkList(Simp.getNode());
7322 // fold to a simpler setcc
7323 if (Simp.getNode() && Simp.getOpcode() == ISD::SETCC)
7324 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other,
7325 N->getOperand(0), Simp.getOperand(2),
7326 Simp.getOperand(0), Simp.getOperand(1),
7332 /// canFoldInAddressingMode - Return true if 'Use' is a load or a store that
7333 /// uses N as its base pointer and that N may be folded in the load / store
7334 /// addressing mode.
7335 static bool canFoldInAddressingMode(SDNode *N, SDNode *Use,
7337 const TargetLowering &TLI) {
7339 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Use)) {
7340 if (LD->isIndexed() || LD->getBasePtr().getNode() != N)
7342 VT = Use->getValueType(0);
7343 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(Use)) {
7344 if (ST->isIndexed() || ST->getBasePtr().getNode() != N)
7346 VT = ST->getValue().getValueType();
7350 TargetLowering::AddrMode AM;
7351 if (N->getOpcode() == ISD::ADD) {
7352 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
7355 AM.BaseOffs = Offset->getSExtValue();
7359 } else if (N->getOpcode() == ISD::SUB) {
7360 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
7363 AM.BaseOffs = -Offset->getSExtValue();
7370 return TLI.isLegalAddressingMode(AM, VT.getTypeForEVT(*DAG.getContext()));
7373 /// CombineToPreIndexedLoadStore - Try turning a load / store into a
7374 /// pre-indexed load / store when the base pointer is an add or subtract
7375 /// and it has other uses besides the load / store. After the
7376 /// transformation, the new indexed load / store has effectively folded
7377 /// the add / subtract in and all of its other uses are redirected to the
7378 /// new load / store.
7379 bool DAGCombiner::CombineToPreIndexedLoadStore(SDNode *N) {
7380 if (Level < AfterLegalizeDAG)
7386 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7387 if (LD->isIndexed())
7389 VT = LD->getMemoryVT();
7390 if (!TLI.isIndexedLoadLegal(ISD::PRE_INC, VT) &&
7391 !TLI.isIndexedLoadLegal(ISD::PRE_DEC, VT))
7393 Ptr = LD->getBasePtr();
7394 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7395 if (ST->isIndexed())
7397 VT = ST->getMemoryVT();
7398 if (!TLI.isIndexedStoreLegal(ISD::PRE_INC, VT) &&
7399 !TLI.isIndexedStoreLegal(ISD::PRE_DEC, VT))
7401 Ptr = ST->getBasePtr();
7407 // If the pointer is not an add/sub, or if it doesn't have multiple uses, bail
7408 // out. There is no reason to make this a preinc/predec.
7409 if ((Ptr.getOpcode() != ISD::ADD && Ptr.getOpcode() != ISD::SUB) ||
7410 Ptr.getNode()->hasOneUse())
7413 // Ask the target to do addressing mode selection.
7416 ISD::MemIndexedMode AM = ISD::UNINDEXED;
7417 if (!TLI.getPreIndexedAddressParts(N, BasePtr, Offset, AM, DAG))
7420 // Backends without true r+i pre-indexed forms may need to pass a
7421 // constant base with a variable offset so that constant coercion
7422 // will work with the patterns in canonical form.
7423 bool Swapped = false;
7424 if (isa<ConstantSDNode>(BasePtr)) {
7425 std::swap(BasePtr, Offset);
7429 // Don't create a indexed load / store with zero offset.
7430 if (isa<ConstantSDNode>(Offset) &&
7431 cast<ConstantSDNode>(Offset)->isNullValue())
7434 // Try turning it into a pre-indexed load / store except when:
7435 // 1) The new base ptr is a frame index.
7436 // 2) If N is a store and the new base ptr is either the same as or is a
7437 // predecessor of the value being stored.
7438 // 3) Another use of old base ptr is a predecessor of N. If ptr is folded
7439 // that would create a cycle.
7440 // 4) All uses are load / store ops that use it as old base ptr.
7442 // Check #1. Preinc'ing a frame index would require copying the stack pointer
7443 // (plus the implicit offset) to a register to preinc anyway.
7444 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
7449 SDValue Val = cast<StoreSDNode>(N)->getValue();
7450 if (Val == BasePtr || BasePtr.getNode()->isPredecessorOf(Val.getNode()))
7454 // If the offset is a constant, there may be other adds of constants that
7455 // can be folded with this one. We should do this to avoid having to keep
7456 // a copy of the original base pointer.
7457 SmallVector<SDNode *, 16> OtherUses;
7458 if (isa<ConstantSDNode>(Offset))
7459 for (SDNode::use_iterator I = BasePtr.getNode()->use_begin(),
7460 E = BasePtr.getNode()->use_end(); I != E; ++I) {
7462 if (Use == Ptr.getNode())
7465 if (Use->isPredecessorOf(N))
7468 if (Use->getOpcode() != ISD::ADD && Use->getOpcode() != ISD::SUB) {
7473 SDValue Op0 = Use->getOperand(0), Op1 = Use->getOperand(1);
7474 if (Op1.getNode() == BasePtr.getNode())
7475 std::swap(Op0, Op1);
7476 assert(Op0.getNode() == BasePtr.getNode() &&
7477 "Use of ADD/SUB but not an operand");
7479 if (!isa<ConstantSDNode>(Op1)) {
7484 // FIXME: In some cases, we can be smarter about this.
7485 if (Op1.getValueType() != Offset.getValueType()) {
7490 OtherUses.push_back(Use);
7494 std::swap(BasePtr, Offset);
7496 // Now check for #3 and #4.
7497 bool RealUse = false;
7499 // Caches for hasPredecessorHelper
7500 SmallPtrSet<const SDNode *, 32> Visited;
7501 SmallVector<const SDNode *, 16> Worklist;
7503 for (SDNode::use_iterator I = Ptr.getNode()->use_begin(),
7504 E = Ptr.getNode()->use_end(); I != E; ++I) {
7508 if (N->hasPredecessorHelper(Use, Visited, Worklist))
7511 // If Ptr may be folded in addressing mode of other use, then it's
7512 // not profitable to do this transformation.
7513 if (!canFoldInAddressingMode(Ptr.getNode(), Use, DAG, TLI))
7522 Result = DAG.getIndexedLoad(SDValue(N,0), SDLoc(N),
7523 BasePtr, Offset, AM);
7525 Result = DAG.getIndexedStore(SDValue(N,0), SDLoc(N),
7526 BasePtr, Offset, AM);
7529 DEBUG(dbgs() << "\nReplacing.4 ";
7531 dbgs() << "\nWith: ";
7532 Result.getNode()->dump(&DAG);
7534 WorkListRemover DeadNodes(*this);
7536 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
7537 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
7539 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
7542 // Finally, since the node is now dead, remove it from the graph.
7546 std::swap(BasePtr, Offset);
7548 // Replace other uses of BasePtr that can be updated to use Ptr
7549 for (unsigned i = 0, e = OtherUses.size(); i != e; ++i) {
7550 unsigned OffsetIdx = 1;
7551 if (OtherUses[i]->getOperand(OffsetIdx).getNode() == BasePtr.getNode())
7553 assert(OtherUses[i]->getOperand(!OffsetIdx).getNode() ==
7554 BasePtr.getNode() && "Expected BasePtr operand");
7556 // We need to replace ptr0 in the following expression:
7557 // x0 * offset0 + y0 * ptr0 = t0
7559 // x1 * offset1 + y1 * ptr0 = t1 (the indexed load/store)
7561 // where x0, x1, y0 and y1 in {-1, 1} are given by the types of the
7562 // indexed load/store and the expresion that needs to be re-written.
7564 // Therefore, we have:
7565 // t0 = (x0 * offset0 - x1 * y0 * y1 *offset1) + (y0 * y1) * t1
7567 ConstantSDNode *CN =
7568 cast<ConstantSDNode>(OtherUses[i]->getOperand(OffsetIdx));
7570 APInt Offset0 = CN->getAPIntValue();
7571 APInt Offset1 = cast<ConstantSDNode>(Offset)->getAPIntValue();
7573 X0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 1) ? -1 : 1;
7574 Y0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 0) ? -1 : 1;
7575 X1 = (AM == ISD::PRE_DEC && !Swapped) ? -1 : 1;
7576 Y1 = (AM == ISD::PRE_DEC && Swapped) ? -1 : 1;
7578 unsigned Opcode = (Y0 * Y1 < 0) ? ISD::SUB : ISD::ADD;
7580 APInt CNV = Offset0;
7581 if (X0 < 0) CNV = -CNV;
7582 if (X1 * Y0 * Y1 < 0) CNV = CNV + Offset1;
7583 else CNV = CNV - Offset1;
7585 // We can now generate the new expression.
7586 SDValue NewOp1 = DAG.getConstant(CNV, CN->getValueType(0));
7587 SDValue NewOp2 = Result.getValue(isLoad ? 1 : 0);
7589 SDValue NewUse = DAG.getNode(Opcode,
7590 SDLoc(OtherUses[i]),
7591 OtherUses[i]->getValueType(0), NewOp1, NewOp2);
7592 DAG.ReplaceAllUsesOfValueWith(SDValue(OtherUses[i], 0), NewUse);
7593 removeFromWorkList(OtherUses[i]);
7594 DAG.DeleteNode(OtherUses[i]);
7597 // Replace the uses of Ptr with uses of the updated base value.
7598 DAG.ReplaceAllUsesOfValueWith(Ptr, Result.getValue(isLoad ? 1 : 0));
7599 removeFromWorkList(Ptr.getNode());
7600 DAG.DeleteNode(Ptr.getNode());
7605 /// CombineToPostIndexedLoadStore - Try to combine a load / store with a
7606 /// add / sub of the base pointer node into a post-indexed load / store.
7607 /// The transformation folded the add / subtract into the new indexed
7608 /// load / store effectively and all of its uses are redirected to the
7609 /// new load / store.
7610 bool DAGCombiner::CombineToPostIndexedLoadStore(SDNode *N) {
7611 if (Level < AfterLegalizeDAG)
7617 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7618 if (LD->isIndexed())
7620 VT = LD->getMemoryVT();
7621 if (!TLI.isIndexedLoadLegal(ISD::POST_INC, VT) &&
7622 !TLI.isIndexedLoadLegal(ISD::POST_DEC, VT))
7624 Ptr = LD->getBasePtr();
7625 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7626 if (ST->isIndexed())
7628 VT = ST->getMemoryVT();
7629 if (!TLI.isIndexedStoreLegal(ISD::POST_INC, VT) &&
7630 !TLI.isIndexedStoreLegal(ISD::POST_DEC, VT))
7632 Ptr = ST->getBasePtr();
7638 if (Ptr.getNode()->hasOneUse())
7641 for (SDNode::use_iterator I = Ptr.getNode()->use_begin(),
7642 E = Ptr.getNode()->use_end(); I != E; ++I) {
7645 (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB))
7650 ISD::MemIndexedMode AM = ISD::UNINDEXED;
7651 if (TLI.getPostIndexedAddressParts(N, Op, BasePtr, Offset, AM, DAG)) {
7652 // Don't create a indexed load / store with zero offset.
7653 if (isa<ConstantSDNode>(Offset) &&
7654 cast<ConstantSDNode>(Offset)->isNullValue())
7657 // Try turning it into a post-indexed load / store except when
7658 // 1) All uses are load / store ops that use it as base ptr (and
7659 // it may be folded as addressing mmode).
7660 // 2) Op must be independent of N, i.e. Op is neither a predecessor
7661 // nor a successor of N. Otherwise, if Op is folded that would
7664 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
7668 bool TryNext = false;
7669 for (SDNode::use_iterator II = BasePtr.getNode()->use_begin(),
7670 EE = BasePtr.getNode()->use_end(); II != EE; ++II) {
7672 if (Use == Ptr.getNode())
7675 // If all the uses are load / store addresses, then don't do the
7677 if (Use->getOpcode() == ISD::ADD || Use->getOpcode() == ISD::SUB){
7678 bool RealUse = false;
7679 for (SDNode::use_iterator III = Use->use_begin(),
7680 EEE = Use->use_end(); III != EEE; ++III) {
7681 SDNode *UseUse = *III;
7682 if (!canFoldInAddressingMode(Use, UseUse, DAG, TLI))
7697 if (!Op->isPredecessorOf(N) && !N->isPredecessorOf(Op)) {
7698 SDValue Result = isLoad
7699 ? DAG.getIndexedLoad(SDValue(N,0), SDLoc(N),
7700 BasePtr, Offset, AM)
7701 : DAG.getIndexedStore(SDValue(N,0), SDLoc(N),
7702 BasePtr, Offset, AM);
7705 DEBUG(dbgs() << "\nReplacing.5 ";
7707 dbgs() << "\nWith: ";
7708 Result.getNode()->dump(&DAG);
7710 WorkListRemover DeadNodes(*this);
7712 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
7713 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
7715 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
7718 // Finally, since the node is now dead, remove it from the graph.
7721 // Replace the uses of Use with uses of the updated base value.
7722 DAG.ReplaceAllUsesOfValueWith(SDValue(Op, 0),
7723 Result.getValue(isLoad ? 1 : 0));
7724 removeFromWorkList(Op);
7734 SDValue DAGCombiner::visitLOAD(SDNode *N) {
7735 LoadSDNode *LD = cast<LoadSDNode>(N);
7736 SDValue Chain = LD->getChain();
7737 SDValue Ptr = LD->getBasePtr();
7739 // If load is not volatile and there are no uses of the loaded value (and
7740 // the updated indexed value in case of indexed loads), change uses of the
7741 // chain value into uses of the chain input (i.e. delete the dead load).
7742 if (!LD->isVolatile()) {
7743 if (N->getValueType(1) == MVT::Other) {
7745 if (!N->hasAnyUseOfValue(0)) {
7746 // It's not safe to use the two value CombineTo variant here. e.g.
7747 // v1, chain2 = load chain1, loc
7748 // v2, chain3 = load chain2, loc
7750 // Now we replace use of chain2 with chain1. This makes the second load
7751 // isomorphic to the one we are deleting, and thus makes this load live.
7752 DEBUG(dbgs() << "\nReplacing.6 ";
7754 dbgs() << "\nWith chain: ";
7755 Chain.getNode()->dump(&DAG);
7757 WorkListRemover DeadNodes(*this);
7758 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
7760 if (N->use_empty()) {
7761 removeFromWorkList(N);
7765 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7769 assert(N->getValueType(2) == MVT::Other && "Malformed indexed loads?");
7770 if (!N->hasAnyUseOfValue(0) && !N->hasAnyUseOfValue(1)) {
7771 SDValue Undef = DAG.getUNDEF(N->getValueType(0));
7772 DEBUG(dbgs() << "\nReplacing.7 ";
7774 dbgs() << "\nWith: ";
7775 Undef.getNode()->dump(&DAG);
7776 dbgs() << " and 2 other values\n");
7777 WorkListRemover DeadNodes(*this);
7778 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Undef);
7779 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1),
7780 DAG.getUNDEF(N->getValueType(1)));
7781 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 2), Chain);
7782 removeFromWorkList(N);
7784 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7789 // If this load is directly stored, replace the load value with the stored
7791 // TODO: Handle store large -> read small portion.
7792 // TODO: Handle TRUNCSTORE/LOADEXT
7793 if (ISD::isNormalLoad(N) && !LD->isVolatile()) {
7794 if (ISD::isNON_TRUNCStore(Chain.getNode())) {
7795 StoreSDNode *PrevST = cast<StoreSDNode>(Chain);
7796 if (PrevST->getBasePtr() == Ptr &&
7797 PrevST->getValue().getValueType() == N->getValueType(0))
7798 return CombineTo(N, Chain.getOperand(1), Chain);
7802 // Try to infer better alignment information than the load already has.
7803 if (OptLevel != CodeGenOpt::None && LD->isUnindexed()) {
7804 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
7805 if (Align > LD->getMemOperand()->getBaseAlignment()) {
7807 DAG.getExtLoad(LD->getExtensionType(), SDLoc(N),
7808 LD->getValueType(0),
7809 Chain, Ptr, LD->getPointerInfo(),
7811 LD->isVolatile(), LD->isNonTemporal(), Align,
7813 return CombineTo(N, NewLoad, SDValue(NewLoad.getNode(), 1), true);
7818 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA :
7819 TLI.getTargetMachine().getSubtarget<TargetSubtargetInfo>().useAA();
7821 if (CombinerAAOnlyFunc.getNumOccurrences() &&
7822 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
7825 if (UseAA && LD->isUnindexed()) {
7826 // Walk up chain skipping non-aliasing memory nodes.
7827 SDValue BetterChain = FindBetterChain(N, Chain);
7829 // If there is a better chain.
7830 if (Chain != BetterChain) {
7833 // Replace the chain to void dependency.
7834 if (LD->getExtensionType() == ISD::NON_EXTLOAD) {
7835 ReplLoad = DAG.getLoad(N->getValueType(0), SDLoc(LD),
7836 BetterChain, Ptr, LD->getMemOperand());
7838 ReplLoad = DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD),
7839 LD->getValueType(0),
7840 BetterChain, Ptr, LD->getMemoryVT(),
7841 LD->getMemOperand());
7844 // Create token factor to keep old chain connected.
7845 SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N),
7846 MVT::Other, Chain, ReplLoad.getValue(1));
7848 // Make sure the new and old chains are cleaned up.
7849 AddToWorkList(Token.getNode());
7851 // Replace uses with load result and token factor. Don't add users
7853 return CombineTo(N, ReplLoad.getValue(0), Token, false);
7857 // Try transforming N to an indexed load.
7858 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
7859 return SDValue(N, 0);
7861 // Try to slice up N to more direct loads if the slices are mapped to
7862 // different register banks or pairing can take place.
7864 return SDValue(N, 0);
7870 /// \brief Helper structure used to slice a load in smaller loads.
7871 /// Basically a slice is obtained from the following sequence:
7872 /// Origin = load Ty1, Base
7873 /// Shift = srl Ty1 Origin, CstTy Amount
7874 /// Inst = trunc Shift to Ty2
7876 /// Then, it will be rewriten into:
7877 /// Slice = load SliceTy, Base + SliceOffset
7878 /// [Inst = zext Slice to Ty2], only if SliceTy <> Ty2
7880 /// SliceTy is deduced from the number of bits that are actually used to
7882 struct LoadedSlice {
7883 /// \brief Helper structure used to compute the cost of a slice.
7885 /// Are we optimizing for code size.
7890 unsigned CrossRegisterBanksCopies;
7894 Cost(bool ForCodeSize = false)
7895 : ForCodeSize(ForCodeSize), Loads(0), Truncates(0),
7896 CrossRegisterBanksCopies(0), ZExts(0), Shift(0) {}
7898 /// \brief Get the cost of one isolated slice.
7899 Cost(const LoadedSlice &LS, bool ForCodeSize = false)
7900 : ForCodeSize(ForCodeSize), Loads(1), Truncates(0),
7901 CrossRegisterBanksCopies(0), ZExts(0), Shift(0) {
7902 EVT TruncType = LS.Inst->getValueType(0);
7903 EVT LoadedType = LS.getLoadedType();
7904 if (TruncType != LoadedType &&
7905 !LS.DAG->getTargetLoweringInfo().isZExtFree(LoadedType, TruncType))
7909 /// \brief Account for slicing gain in the current cost.
7910 /// Slicing provide a few gains like removing a shift or a
7911 /// truncate. This method allows to grow the cost of the original
7912 /// load with the gain from this slice.
7913 void addSliceGain(const LoadedSlice &LS) {
7914 // Each slice saves a truncate.
7915 const TargetLowering &TLI = LS.DAG->getTargetLoweringInfo();
7916 if (!TLI.isTruncateFree(LS.Inst->getValueType(0),
7917 LS.Inst->getOperand(0).getValueType()))
7919 // If there is a shift amount, this slice gets rid of it.
7922 // If this slice can merge a cross register bank copy, account for it.
7923 if (LS.canMergeExpensiveCrossRegisterBankCopy())
7924 ++CrossRegisterBanksCopies;
7927 Cost &operator+=(const Cost &RHS) {
7929 Truncates += RHS.Truncates;
7930 CrossRegisterBanksCopies += RHS.CrossRegisterBanksCopies;
7936 bool operator==(const Cost &RHS) const {
7937 return Loads == RHS.Loads && Truncates == RHS.Truncates &&
7938 CrossRegisterBanksCopies == RHS.CrossRegisterBanksCopies &&
7939 ZExts == RHS.ZExts && Shift == RHS.Shift;
7942 bool operator!=(const Cost &RHS) const { return !(*this == RHS); }
7944 bool operator<(const Cost &RHS) const {
7945 // Assume cross register banks copies are as expensive as loads.
7946 // FIXME: Do we want some more target hooks?
7947 unsigned ExpensiveOpsLHS = Loads + CrossRegisterBanksCopies;
7948 unsigned ExpensiveOpsRHS = RHS.Loads + RHS.CrossRegisterBanksCopies;
7949 // Unless we are optimizing for code size, consider the
7950 // expensive operation first.
7951 if (!ForCodeSize && ExpensiveOpsLHS != ExpensiveOpsRHS)
7952 return ExpensiveOpsLHS < ExpensiveOpsRHS;
7953 return (Truncates + ZExts + Shift + ExpensiveOpsLHS) <
7954 (RHS.Truncates + RHS.ZExts + RHS.Shift + ExpensiveOpsRHS);
7957 bool operator>(const Cost &RHS) const { return RHS < *this; }
7959 bool operator<=(const Cost &RHS) const { return !(RHS < *this); }
7961 bool operator>=(const Cost &RHS) const { return !(*this < RHS); }
7963 // The last instruction that represent the slice. This should be a
7964 // truncate instruction.
7966 // The original load instruction.
7968 // The right shift amount in bits from the original load.
7970 // The DAG from which Origin came from.
7971 // This is used to get some contextual information about legal types, etc.
7974 LoadedSlice(SDNode *Inst = NULL, LoadSDNode *Origin = NULL,
7975 unsigned Shift = 0, SelectionDAG *DAG = NULL)
7976 : Inst(Inst), Origin(Origin), Shift(Shift), DAG(DAG) {}
7978 LoadedSlice(const LoadedSlice &LS)
7979 : Inst(LS.Inst), Origin(LS.Origin), Shift(LS.Shift), DAG(LS.DAG) {}
7981 /// \brief Get the bits used in a chunk of bits \p BitWidth large.
7982 /// \return Result is \p BitWidth and has used bits set to 1 and
7983 /// not used bits set to 0.
7984 APInt getUsedBits() const {
7985 // Reproduce the trunc(lshr) sequence:
7986 // - Start from the truncated value.
7987 // - Zero extend to the desired bit width.
7989 assert(Origin && "No original load to compare against.");
7990 unsigned BitWidth = Origin->getValueSizeInBits(0);
7991 assert(Inst && "This slice is not bound to an instruction");
7992 assert(Inst->getValueSizeInBits(0) <= BitWidth &&
7993 "Extracted slice is bigger than the whole type!");
7994 APInt UsedBits(Inst->getValueSizeInBits(0), 0);
7995 UsedBits.setAllBits();
7996 UsedBits = UsedBits.zext(BitWidth);
8001 /// \brief Get the size of the slice to be loaded in bytes.
8002 unsigned getLoadedSize() const {
8003 unsigned SliceSize = getUsedBits().countPopulation();
8004 assert(!(SliceSize & 0x7) && "Size is not a multiple of a byte.");
8005 return SliceSize / 8;
8008 /// \brief Get the type that will be loaded for this slice.
8009 /// Note: This may not be the final type for the slice.
8010 EVT getLoadedType() const {
8011 assert(DAG && "Missing context");
8012 LLVMContext &Ctxt = *DAG->getContext();
8013 return EVT::getIntegerVT(Ctxt, getLoadedSize() * 8);
8016 /// \brief Get the alignment of the load used for this slice.
8017 unsigned getAlignment() const {
8018 unsigned Alignment = Origin->getAlignment();
8019 unsigned Offset = getOffsetFromBase();
8021 Alignment = MinAlign(Alignment, Alignment + Offset);
8025 /// \brief Check if this slice can be rewritten with legal operations.
8026 bool isLegal() const {
8027 // An invalid slice is not legal.
8028 if (!Origin || !Inst || !DAG)
8031 // Offsets are for indexed load only, we do not handle that.
8032 if (Origin->getOffset().getOpcode() != ISD::UNDEF)
8035 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
8037 // Check that the type is legal.
8038 EVT SliceType = getLoadedType();
8039 if (!TLI.isTypeLegal(SliceType))
8042 // Check that the load is legal for this type.
8043 if (!TLI.isOperationLegal(ISD::LOAD, SliceType))
8046 // Check that the offset can be computed.
8047 // 1. Check its type.
8048 EVT PtrType = Origin->getBasePtr().getValueType();
8049 if (PtrType == MVT::Untyped || PtrType.isExtended())
8052 // 2. Check that it fits in the immediate.
8053 if (!TLI.isLegalAddImmediate(getOffsetFromBase()))
8056 // 3. Check that the computation is legal.
8057 if (!TLI.isOperationLegal(ISD::ADD, PtrType))
8060 // Check that the zext is legal if it needs one.
8061 EVT TruncateType = Inst->getValueType(0);
8062 if (TruncateType != SliceType &&
8063 !TLI.isOperationLegal(ISD::ZERO_EXTEND, TruncateType))
8069 /// \brief Get the offset in bytes of this slice in the original chunk of
8071 /// \pre DAG != NULL.
8072 uint64_t getOffsetFromBase() const {
8073 assert(DAG && "Missing context.");
8075 DAG->getTargetLoweringInfo().getDataLayout()->isBigEndian();
8076 assert(!(Shift & 0x7) && "Shifts not aligned on Bytes are not supported.");
8077 uint64_t Offset = Shift / 8;
8078 unsigned TySizeInBytes = Origin->getValueSizeInBits(0) / 8;
8079 assert(!(Origin->getValueSizeInBits(0) & 0x7) &&
8080 "The size of the original loaded type is not a multiple of a"
8082 // If Offset is bigger than TySizeInBytes, it means we are loading all
8083 // zeros. This should have been optimized before in the process.
8084 assert(TySizeInBytes > Offset &&
8085 "Invalid shift amount for given loaded size");
8087 Offset = TySizeInBytes - Offset - getLoadedSize();
8091 /// \brief Generate the sequence of instructions to load the slice
8092 /// represented by this object and redirect the uses of this slice to
8093 /// this new sequence of instructions.
8094 /// \pre this->Inst && this->Origin are valid Instructions and this
8095 /// object passed the legal check: LoadedSlice::isLegal returned true.
8096 /// \return The last instruction of the sequence used to load the slice.
8097 SDValue loadSlice() const {
8098 assert(Inst && Origin && "Unable to replace a non-existing slice.");
8099 const SDValue &OldBaseAddr = Origin->getBasePtr();
8100 SDValue BaseAddr = OldBaseAddr;
8101 // Get the offset in that chunk of bytes w.r.t. the endianess.
8102 int64_t Offset = static_cast<int64_t>(getOffsetFromBase());
8103 assert(Offset >= 0 && "Offset too big to fit in int64_t!");
8105 // BaseAddr = BaseAddr + Offset.
8106 EVT ArithType = BaseAddr.getValueType();
8107 BaseAddr = DAG->getNode(ISD::ADD, SDLoc(Origin), ArithType, BaseAddr,
8108 DAG->getConstant(Offset, ArithType));
8111 // Create the type of the loaded slice according to its size.
8112 EVT SliceType = getLoadedType();
8114 // Create the load for the slice.
8115 SDValue LastInst = DAG->getLoad(
8116 SliceType, SDLoc(Origin), Origin->getChain(), BaseAddr,
8117 Origin->getPointerInfo().getWithOffset(Offset), Origin->isVolatile(),
8118 Origin->isNonTemporal(), Origin->isInvariant(), getAlignment());
8119 // If the final type is not the same as the loaded type, this means that
8120 // we have to pad with zero. Create a zero extend for that.
8121 EVT FinalType = Inst->getValueType(0);
8122 if (SliceType != FinalType)
8124 DAG->getNode(ISD::ZERO_EXTEND, SDLoc(LastInst), FinalType, LastInst);
8128 /// \brief Check if this slice can be merged with an expensive cross register
8129 /// bank copy. E.g.,
8131 /// f = bitcast i32 i to float
8132 bool canMergeExpensiveCrossRegisterBankCopy() const {
8133 if (!Inst || !Inst->hasOneUse())
8135 SDNode *Use = *Inst->use_begin();
8136 if (Use->getOpcode() != ISD::BITCAST)
8138 assert(DAG && "Missing context");
8139 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
8140 EVT ResVT = Use->getValueType(0);
8141 const TargetRegisterClass *ResRC = TLI.getRegClassFor(ResVT.getSimpleVT());
8142 const TargetRegisterClass *ArgRC =
8143 TLI.getRegClassFor(Use->getOperand(0).getValueType().getSimpleVT());
8144 if (ArgRC == ResRC || !TLI.isOperationLegal(ISD::LOAD, ResVT))
8147 // At this point, we know that we perform a cross-register-bank copy.
8148 // Check if it is expensive.
8149 const TargetRegisterInfo *TRI = TLI.getTargetMachine().getRegisterInfo();
8150 // Assume bitcasts are cheap, unless both register classes do not
8151 // explicitly share a common sub class.
8152 if (!TRI || TRI->getCommonSubClass(ArgRC, ResRC))
8155 // Check if it will be merged with the load.
8156 // 1. Check the alignment constraint.
8157 unsigned RequiredAlignment = TLI.getDataLayout()->getABITypeAlignment(
8158 ResVT.getTypeForEVT(*DAG->getContext()));
8160 if (RequiredAlignment > getAlignment())
8163 // 2. Check that the load is a legal operation for that type.
8164 if (!TLI.isOperationLegal(ISD::LOAD, ResVT))
8167 // 3. Check that we do not have a zext in the way.
8168 if (Inst->getValueType(0) != getLoadedType())
8176 /// \brief Sorts LoadedSlice according to their offset.
8177 struct LoadedSliceSorter {
8178 bool operator()(const LoadedSlice &LHS, const LoadedSlice &RHS) {
8179 assert(LHS.Origin == RHS.Origin && "Different bases not implemented.");
8180 return LHS.getOffsetFromBase() < RHS.getOffsetFromBase();
8184 /// \brief Check that all bits set in \p UsedBits form a dense region, i.e.,
8185 /// \p UsedBits looks like 0..0 1..1 0..0.
8186 static bool areUsedBitsDense(const APInt &UsedBits) {
8187 // If all the bits are one, this is dense!
8188 if (UsedBits.isAllOnesValue())
8191 // Get rid of the unused bits on the right.
8192 APInt NarrowedUsedBits = UsedBits.lshr(UsedBits.countTrailingZeros());
8193 // Get rid of the unused bits on the left.
8194 if (NarrowedUsedBits.countLeadingZeros())
8195 NarrowedUsedBits = NarrowedUsedBits.trunc(NarrowedUsedBits.getActiveBits());
8196 // Check that the chunk of bits is completely used.
8197 return NarrowedUsedBits.isAllOnesValue();
8200 /// \brief Check whether or not \p First and \p Second are next to each other
8201 /// in memory. This means that there is no hole between the bits loaded
8202 /// by \p First and the bits loaded by \p Second.
8203 static bool areSlicesNextToEachOther(const LoadedSlice &First,
8204 const LoadedSlice &Second) {
8205 assert(First.Origin == Second.Origin && First.Origin &&
8206 "Unable to match different memory origins.");
8207 APInt UsedBits = First.getUsedBits();
8208 assert((UsedBits & Second.getUsedBits()) == 0 &&
8209 "Slices are not supposed to overlap.");
8210 UsedBits |= Second.getUsedBits();
8211 return areUsedBitsDense(UsedBits);
8214 /// \brief Adjust the \p GlobalLSCost according to the target
8215 /// paring capabilities and the layout of the slices.
8216 /// \pre \p GlobalLSCost should account for at least as many loads as
8217 /// there is in the slices in \p LoadedSlices.
8218 static void adjustCostForPairing(SmallVectorImpl<LoadedSlice> &LoadedSlices,
8219 LoadedSlice::Cost &GlobalLSCost) {
8220 unsigned NumberOfSlices = LoadedSlices.size();
8221 // If there is less than 2 elements, no pairing is possible.
8222 if (NumberOfSlices < 2)
8225 // Sort the slices so that elements that are likely to be next to each
8226 // other in memory are next to each other in the list.
8227 std::sort(LoadedSlices.begin(), LoadedSlices.end(), LoadedSliceSorter());
8228 const TargetLowering &TLI = LoadedSlices[0].DAG->getTargetLoweringInfo();
8229 // First (resp. Second) is the first (resp. Second) potentially candidate
8230 // to be placed in a paired load.
8231 const LoadedSlice *First = NULL;
8232 const LoadedSlice *Second = NULL;
8233 for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice,
8234 // Set the beginning of the pair.
8237 Second = &LoadedSlices[CurrSlice];
8239 // If First is NULL, it means we start a new pair.
8240 // Get to the next slice.
8244 EVT LoadedType = First->getLoadedType();
8246 // If the types of the slices are different, we cannot pair them.
8247 if (LoadedType != Second->getLoadedType())
8250 // Check if the target supplies paired loads for this type.
8251 unsigned RequiredAlignment = 0;
8252 if (!TLI.hasPairedLoad(LoadedType, RequiredAlignment)) {
8253 // move to the next pair, this type is hopeless.
8257 // Check if we meet the alignment requirement.
8258 if (RequiredAlignment > First->getAlignment())
8261 // Check that both loads are next to each other in memory.
8262 if (!areSlicesNextToEachOther(*First, *Second))
8265 assert(GlobalLSCost.Loads > 0 && "We save more loads than we created!");
8266 --GlobalLSCost.Loads;
8267 // Move to the next pair.
8272 /// \brief Check the profitability of all involved LoadedSlice.
8273 /// Currently, it is considered profitable if there is exactly two
8274 /// involved slices (1) which are (2) next to each other in memory, and
8275 /// whose cost (\see LoadedSlice::Cost) is smaller than the original load (3).
8277 /// Note: The order of the elements in \p LoadedSlices may be modified, but not
8278 /// the elements themselves.
8280 /// FIXME: When the cost model will be mature enough, we can relax
8281 /// constraints (1) and (2).
8282 static bool isSlicingProfitable(SmallVectorImpl<LoadedSlice> &LoadedSlices,
8283 const APInt &UsedBits, bool ForCodeSize) {
8284 unsigned NumberOfSlices = LoadedSlices.size();
8285 if (StressLoadSlicing)
8286 return NumberOfSlices > 1;
8289 if (NumberOfSlices != 2)
8293 if (!areUsedBitsDense(UsedBits))
8297 LoadedSlice::Cost OrigCost(ForCodeSize), GlobalSlicingCost(ForCodeSize);
8298 // The original code has one big load.
8300 for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice) {
8301 const LoadedSlice &LS = LoadedSlices[CurrSlice];
8302 // Accumulate the cost of all the slices.
8303 LoadedSlice::Cost SliceCost(LS, ForCodeSize);
8304 GlobalSlicingCost += SliceCost;
8306 // Account as cost in the original configuration the gain obtained
8307 // with the current slices.
8308 OrigCost.addSliceGain(LS);
8311 // If the target supports paired load, adjust the cost accordingly.
8312 adjustCostForPairing(LoadedSlices, GlobalSlicingCost);
8313 return OrigCost > GlobalSlicingCost;
8316 /// \brief If the given load, \p LI, is used only by trunc or trunc(lshr)
8317 /// operations, split it in the various pieces being extracted.
8319 /// This sort of thing is introduced by SROA.
8320 /// This slicing takes care not to insert overlapping loads.
8321 /// \pre LI is a simple load (i.e., not an atomic or volatile load).
8322 bool DAGCombiner::SliceUpLoad(SDNode *N) {
8323 if (Level < AfterLegalizeDAG)
8326 LoadSDNode *LD = cast<LoadSDNode>(N);
8327 if (LD->isVolatile() || !ISD::isNormalLoad(LD) ||
8328 !LD->getValueType(0).isInteger())
8331 // Keep track of already used bits to detect overlapping values.
8332 // In that case, we will just abort the transformation.
8333 APInt UsedBits(LD->getValueSizeInBits(0), 0);
8335 SmallVector<LoadedSlice, 4> LoadedSlices;
8337 // Check if this load is used as several smaller chunks of bits.
8338 // Basically, look for uses in trunc or trunc(lshr) and record a new chain
8339 // of computation for each trunc.
8340 for (SDNode::use_iterator UI = LD->use_begin(), UIEnd = LD->use_end();
8341 UI != UIEnd; ++UI) {
8342 // Skip the uses of the chain.
8343 if (UI.getUse().getResNo() != 0)
8349 // Check if this is a trunc(lshr).
8350 if (User->getOpcode() == ISD::SRL && User->hasOneUse() &&
8351 isa<ConstantSDNode>(User->getOperand(1))) {
8352 Shift = cast<ConstantSDNode>(User->getOperand(1))->getZExtValue();
8353 User = *User->use_begin();
8356 // At this point, User is a Truncate, iff we encountered, trunc or
8358 if (User->getOpcode() != ISD::TRUNCATE)
8361 // The width of the type must be a power of 2 and greater than 8-bits.
8362 // Otherwise the load cannot be represented in LLVM IR.
8363 // Moreover, if we shifted with a non-8-bits multiple, the slice
8364 // will be across several bytes. We do not support that.
8365 unsigned Width = User->getValueSizeInBits(0);
8366 if (Width < 8 || !isPowerOf2_32(Width) || (Shift & 0x7))
8369 // Build the slice for this chain of computations.
8370 LoadedSlice LS(User, LD, Shift, &DAG);
8371 APInt CurrentUsedBits = LS.getUsedBits();
8373 // Check if this slice overlaps with another.
8374 if ((CurrentUsedBits & UsedBits) != 0)
8376 // Update the bits used globally.
8377 UsedBits |= CurrentUsedBits;
8379 // Check if the new slice would be legal.
8383 // Record the slice.
8384 LoadedSlices.push_back(LS);
8387 // Abort slicing if it does not seem to be profitable.
8388 if (!isSlicingProfitable(LoadedSlices, UsedBits, ForCodeSize))
8393 // Rewrite each chain to use an independent load.
8394 // By construction, each chain can be represented by a unique load.
8396 // Prepare the argument for the new token factor for all the slices.
8397 SmallVector<SDValue, 8> ArgChains;
8398 for (SmallVectorImpl<LoadedSlice>::const_iterator
8399 LSIt = LoadedSlices.begin(),
8400 LSItEnd = LoadedSlices.end();
8401 LSIt != LSItEnd; ++LSIt) {
8402 SDValue SliceInst = LSIt->loadSlice();
8403 CombineTo(LSIt->Inst, SliceInst, true);
8404 if (SliceInst.getNode()->getOpcode() != ISD::LOAD)
8405 SliceInst = SliceInst.getOperand(0);
8406 assert(SliceInst->getOpcode() == ISD::LOAD &&
8407 "It takes more than a zext to get to the loaded slice!!");
8408 ArgChains.push_back(SliceInst.getValue(1));
8411 SDValue Chain = DAG.getNode(ISD::TokenFactor, SDLoc(LD), MVT::Other,
8412 &ArgChains[0], ArgChains.size());
8413 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
8417 /// CheckForMaskedLoad - Check to see if V is (and load (ptr), imm), where the
8418 /// load is having specific bytes cleared out. If so, return the byte size
8419 /// being masked out and the shift amount.
8420 static std::pair<unsigned, unsigned>
8421 CheckForMaskedLoad(SDValue V, SDValue Ptr, SDValue Chain) {
8422 std::pair<unsigned, unsigned> Result(0, 0);
8424 // Check for the structure we're looking for.
8425 if (V->getOpcode() != ISD::AND ||
8426 !isa<ConstantSDNode>(V->getOperand(1)) ||
8427 !ISD::isNormalLoad(V->getOperand(0).getNode()))
8430 // Check the chain and pointer.
8431 LoadSDNode *LD = cast<LoadSDNode>(V->getOperand(0));
8432 if (LD->getBasePtr() != Ptr) return Result; // Not from same pointer.
8434 // The store should be chained directly to the load or be an operand of a
8436 if (LD == Chain.getNode())
8438 else if (Chain->getOpcode() != ISD::TokenFactor)
8439 return Result; // Fail.
8442 for (unsigned i = 0, e = Chain->getNumOperands(); i != e; ++i)
8443 if (Chain->getOperand(i).getNode() == LD) {
8447 if (!isOk) return Result;
8450 // This only handles simple types.
8451 if (V.getValueType() != MVT::i16 &&
8452 V.getValueType() != MVT::i32 &&
8453 V.getValueType() != MVT::i64)
8456 // Check the constant mask. Invert it so that the bits being masked out are
8457 // 0 and the bits being kept are 1. Use getSExtValue so that leading bits
8458 // follow the sign bit for uniformity.
8459 uint64_t NotMask = ~cast<ConstantSDNode>(V->getOperand(1))->getSExtValue();
8460 unsigned NotMaskLZ = countLeadingZeros(NotMask);
8461 if (NotMaskLZ & 7) return Result; // Must be multiple of a byte.
8462 unsigned NotMaskTZ = countTrailingZeros(NotMask);
8463 if (NotMaskTZ & 7) return Result; // Must be multiple of a byte.
8464 if (NotMaskLZ == 64) return Result; // All zero mask.
8466 // See if we have a continuous run of bits. If so, we have 0*1+0*
8467 if (CountTrailingOnes_64(NotMask >> NotMaskTZ)+NotMaskTZ+NotMaskLZ != 64)
8470 // Adjust NotMaskLZ down to be from the actual size of the int instead of i64.
8471 if (V.getValueType() != MVT::i64 && NotMaskLZ)
8472 NotMaskLZ -= 64-V.getValueSizeInBits();
8474 unsigned MaskedBytes = (V.getValueSizeInBits()-NotMaskLZ-NotMaskTZ)/8;
8475 switch (MaskedBytes) {
8479 default: return Result; // All one mask, or 5-byte mask.
8482 // Verify that the first bit starts at a multiple of mask so that the access
8483 // is aligned the same as the access width.
8484 if (NotMaskTZ && NotMaskTZ/8 % MaskedBytes) return Result;
8486 Result.first = MaskedBytes;
8487 Result.second = NotMaskTZ/8;
8492 /// ShrinkLoadReplaceStoreWithStore - Check to see if IVal is something that
8493 /// provides a value as specified by MaskInfo. If so, replace the specified
8494 /// store with a narrower store of truncated IVal.
8496 ShrinkLoadReplaceStoreWithStore(const std::pair<unsigned, unsigned> &MaskInfo,
8497 SDValue IVal, StoreSDNode *St,
8499 unsigned NumBytes = MaskInfo.first;
8500 unsigned ByteShift = MaskInfo.second;
8501 SelectionDAG &DAG = DC->getDAG();
8503 // Check to see if IVal is all zeros in the part being masked in by the 'or'
8504 // that uses this. If not, this is not a replacement.
8505 APInt Mask = ~APInt::getBitsSet(IVal.getValueSizeInBits(),
8506 ByteShift*8, (ByteShift+NumBytes)*8);
8507 if (!DAG.MaskedValueIsZero(IVal, Mask)) return 0;
8509 // Check that it is legal on the target to do this. It is legal if the new
8510 // VT we're shrinking to (i8/i16/i32) is legal or we're still before type
8512 MVT VT = MVT::getIntegerVT(NumBytes*8);
8513 if (!DC->isTypeLegal(VT))
8516 // Okay, we can do this! Replace the 'St' store with a store of IVal that is
8517 // shifted by ByteShift and truncated down to NumBytes.
8519 IVal = DAG.getNode(ISD::SRL, SDLoc(IVal), IVal.getValueType(), IVal,
8520 DAG.getConstant(ByteShift*8,
8521 DC->getShiftAmountTy(IVal.getValueType())));
8523 // Figure out the offset for the store and the alignment of the access.
8525 unsigned NewAlign = St->getAlignment();
8527 if (DAG.getTargetLoweringInfo().isLittleEndian())
8528 StOffset = ByteShift;
8530 StOffset = IVal.getValueType().getStoreSize() - ByteShift - NumBytes;
8532 SDValue Ptr = St->getBasePtr();
8534 Ptr = DAG.getNode(ISD::ADD, SDLoc(IVal), Ptr.getValueType(),
8535 Ptr, DAG.getConstant(StOffset, Ptr.getValueType()));
8536 NewAlign = MinAlign(NewAlign, StOffset);
8539 // Truncate down to the new size.
8540 IVal = DAG.getNode(ISD::TRUNCATE, SDLoc(IVal), VT, IVal);
8543 return DAG.getStore(St->getChain(), SDLoc(St), IVal, Ptr,
8544 St->getPointerInfo().getWithOffset(StOffset),
8545 false, false, NewAlign).getNode();
8549 /// ReduceLoadOpStoreWidth - Look for sequence of load / op / store where op is
8550 /// one of 'or', 'xor', and 'and' of immediates. If 'op' is only touching some
8551 /// of the loaded bits, try narrowing the load and store if it would end up
8552 /// being a win for performance or code size.
8553 SDValue DAGCombiner::ReduceLoadOpStoreWidth(SDNode *N) {
8554 StoreSDNode *ST = cast<StoreSDNode>(N);
8555 if (ST->isVolatile())
8558 SDValue Chain = ST->getChain();
8559 SDValue Value = ST->getValue();
8560 SDValue Ptr = ST->getBasePtr();
8561 EVT VT = Value.getValueType();
8563 if (ST->isTruncatingStore() || VT.isVector() || !Value.hasOneUse())
8566 unsigned Opc = Value.getOpcode();
8568 // If this is "store (or X, Y), P" and X is "(and (load P), cst)", where cst
8569 // is a byte mask indicating a consecutive number of bytes, check to see if
8570 // Y is known to provide just those bytes. If so, we try to replace the
8571 // load + replace + store sequence with a single (narrower) store, which makes
8573 if (Opc == ISD::OR) {
8574 std::pair<unsigned, unsigned> MaskedLoad;
8575 MaskedLoad = CheckForMaskedLoad(Value.getOperand(0), Ptr, Chain);
8576 if (MaskedLoad.first)
8577 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
8578 Value.getOperand(1), ST,this))
8579 return SDValue(NewST, 0);
8581 // Or is commutative, so try swapping X and Y.
8582 MaskedLoad = CheckForMaskedLoad(Value.getOperand(1), Ptr, Chain);
8583 if (MaskedLoad.first)
8584 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
8585 Value.getOperand(0), ST,this))
8586 return SDValue(NewST, 0);
8589 if ((Opc != ISD::OR && Opc != ISD::XOR && Opc != ISD::AND) ||
8590 Value.getOperand(1).getOpcode() != ISD::Constant)
8593 SDValue N0 = Value.getOperand(0);
8594 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
8595 Chain == SDValue(N0.getNode(), 1)) {
8596 LoadSDNode *LD = cast<LoadSDNode>(N0);
8597 if (LD->getBasePtr() != Ptr ||
8598 LD->getPointerInfo().getAddrSpace() !=
8599 ST->getPointerInfo().getAddrSpace())
8602 // Find the type to narrow it the load / op / store to.
8603 SDValue N1 = Value.getOperand(1);
8604 unsigned BitWidth = N1.getValueSizeInBits();
8605 APInt Imm = cast<ConstantSDNode>(N1)->getAPIntValue();
8606 if (Opc == ISD::AND)
8607 Imm ^= APInt::getAllOnesValue(BitWidth);
8608 if (Imm == 0 || Imm.isAllOnesValue())
8610 unsigned ShAmt = Imm.countTrailingZeros();
8611 unsigned MSB = BitWidth - Imm.countLeadingZeros() - 1;
8612 unsigned NewBW = NextPowerOf2(MSB - ShAmt);
8613 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
8614 while (NewBW < BitWidth &&
8615 !(TLI.isOperationLegalOrCustom(Opc, NewVT) &&
8616 TLI.isNarrowingProfitable(VT, NewVT))) {
8617 NewBW = NextPowerOf2(NewBW);
8618 NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
8620 if (NewBW >= BitWidth)
8623 // If the lsb changed does not start at the type bitwidth boundary,
8624 // start at the previous one.
8626 ShAmt = (((ShAmt + NewBW - 1) / NewBW) * NewBW) - NewBW;
8627 APInt Mask = APInt::getBitsSet(BitWidth, ShAmt,
8628 std::min(BitWidth, ShAmt + NewBW));
8629 if ((Imm & Mask) == Imm) {
8630 APInt NewImm = (Imm & Mask).lshr(ShAmt).trunc(NewBW);
8631 if (Opc == ISD::AND)
8632 NewImm ^= APInt::getAllOnesValue(NewBW);
8633 uint64_t PtrOff = ShAmt / 8;
8634 // For big endian targets, we need to adjust the offset to the pointer to
8635 // load the correct bytes.
8636 if (TLI.isBigEndian())
8637 PtrOff = (BitWidth + 7 - NewBW) / 8 - PtrOff;
8639 unsigned NewAlign = MinAlign(LD->getAlignment(), PtrOff);
8640 Type *NewVTTy = NewVT.getTypeForEVT(*DAG.getContext());
8641 if (NewAlign < TLI.getDataLayout()->getABITypeAlignment(NewVTTy))
8644 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(LD),
8645 Ptr.getValueType(), Ptr,
8646 DAG.getConstant(PtrOff, Ptr.getValueType()));
8647 SDValue NewLD = DAG.getLoad(NewVT, SDLoc(N0),
8648 LD->getChain(), NewPtr,
8649 LD->getPointerInfo().getWithOffset(PtrOff),
8650 LD->isVolatile(), LD->isNonTemporal(),
8651 LD->isInvariant(), NewAlign,
8653 SDValue NewVal = DAG.getNode(Opc, SDLoc(Value), NewVT, NewLD,
8654 DAG.getConstant(NewImm, NewVT));
8655 SDValue NewST = DAG.getStore(Chain, SDLoc(N),
8657 ST->getPointerInfo().getWithOffset(PtrOff),
8658 false, false, NewAlign);
8660 AddToWorkList(NewPtr.getNode());
8661 AddToWorkList(NewLD.getNode());
8662 AddToWorkList(NewVal.getNode());
8663 WorkListRemover DeadNodes(*this);
8664 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLD.getValue(1));
8673 /// TransformFPLoadStorePair - For a given floating point load / store pair,
8674 /// if the load value isn't used by any other operations, then consider
8675 /// transforming the pair to integer load / store operations if the target
8676 /// deems the transformation profitable.
8677 SDValue DAGCombiner::TransformFPLoadStorePair(SDNode *N) {
8678 StoreSDNode *ST = cast<StoreSDNode>(N);
8679 SDValue Chain = ST->getChain();
8680 SDValue Value = ST->getValue();
8681 if (ISD::isNormalStore(ST) && ISD::isNormalLoad(Value.getNode()) &&
8682 Value.hasOneUse() &&
8683 Chain == SDValue(Value.getNode(), 1)) {
8684 LoadSDNode *LD = cast<LoadSDNode>(Value);
8685 EVT VT = LD->getMemoryVT();
8686 if (!VT.isFloatingPoint() ||
8687 VT != ST->getMemoryVT() ||
8688 LD->isNonTemporal() ||
8689 ST->isNonTemporal() ||
8690 LD->getPointerInfo().getAddrSpace() != 0 ||
8691 ST->getPointerInfo().getAddrSpace() != 0)
8694 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
8695 if (!TLI.isOperationLegal(ISD::LOAD, IntVT) ||
8696 !TLI.isOperationLegal(ISD::STORE, IntVT) ||
8697 !TLI.isDesirableToTransformToIntegerOp(ISD::LOAD, VT) ||
8698 !TLI.isDesirableToTransformToIntegerOp(ISD::STORE, VT))
8701 unsigned LDAlign = LD->getAlignment();
8702 unsigned STAlign = ST->getAlignment();
8703 Type *IntVTTy = IntVT.getTypeForEVT(*DAG.getContext());
8704 unsigned ABIAlign = TLI.getDataLayout()->getABITypeAlignment(IntVTTy);
8705 if (LDAlign < ABIAlign || STAlign < ABIAlign)
8708 SDValue NewLD = DAG.getLoad(IntVT, SDLoc(Value),
8709 LD->getChain(), LD->getBasePtr(),
8710 LD->getPointerInfo(),
8711 false, false, false, LDAlign);
8713 SDValue NewST = DAG.getStore(NewLD.getValue(1), SDLoc(N),
8714 NewLD, ST->getBasePtr(),
8715 ST->getPointerInfo(),
8716 false, false, STAlign);
8718 AddToWorkList(NewLD.getNode());
8719 AddToWorkList(NewST.getNode());
8720 WorkListRemover DeadNodes(*this);
8721 DAG.ReplaceAllUsesOfValueWith(Value.getValue(1), NewLD.getValue(1));
8729 /// Helper struct to parse and store a memory address as base + index + offset.
8730 /// We ignore sign extensions when it is safe to do so.
8731 /// The following two expressions are not equivalent. To differentiate we need
8732 /// to store whether there was a sign extension involved in the index
8734 /// (load (i64 add (i64 copyfromreg %c)
8735 /// (i64 signextend (add (i8 load %index)
8739 /// (load (i64 add (i64 copyfromreg %c)
8740 /// (i64 signextend (i32 add (i32 signextend (i8 load %index))
8742 struct BaseIndexOffset {
8746 bool IsIndexSignExt;
8748 BaseIndexOffset() : Offset(0), IsIndexSignExt(false) {}
8750 BaseIndexOffset(SDValue Base, SDValue Index, int64_t Offset,
8751 bool IsIndexSignExt) :
8752 Base(Base), Index(Index), Offset(Offset), IsIndexSignExt(IsIndexSignExt) {}
8754 bool equalBaseIndex(const BaseIndexOffset &Other) {
8755 return Other.Base == Base && Other.Index == Index &&
8756 Other.IsIndexSignExt == IsIndexSignExt;
8759 /// Parses tree in Ptr for base, index, offset addresses.
8760 static BaseIndexOffset match(SDValue Ptr) {
8761 bool IsIndexSignExt = false;
8763 // We only can pattern match BASE + INDEX + OFFSET. If Ptr is not an ADD
8764 // instruction, then it could be just the BASE or everything else we don't
8765 // know how to handle. Just use Ptr as BASE and give up.
8766 if (Ptr->getOpcode() != ISD::ADD)
8767 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
8769 // We know that we have at least an ADD instruction. Try to pattern match
8770 // the simple case of BASE + OFFSET.
8771 if (isa<ConstantSDNode>(Ptr->getOperand(1))) {
8772 int64_t Offset = cast<ConstantSDNode>(Ptr->getOperand(1))->getSExtValue();
8773 return BaseIndexOffset(Ptr->getOperand(0), SDValue(), Offset,
8777 // Inside a loop the current BASE pointer is calculated using an ADD and a
8778 // MUL instruction. In this case Ptr is the actual BASE pointer.
8779 // (i64 add (i64 %array_ptr)
8780 // (i64 mul (i64 %induction_var)
8781 // (i64 %element_size)))
8782 if (Ptr->getOperand(1)->getOpcode() == ISD::MUL)
8783 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
8785 // Look at Base + Index + Offset cases.
8786 SDValue Base = Ptr->getOperand(0);
8787 SDValue IndexOffset = Ptr->getOperand(1);
8789 // Skip signextends.
8790 if (IndexOffset->getOpcode() == ISD::SIGN_EXTEND) {
8791 IndexOffset = IndexOffset->getOperand(0);
8792 IsIndexSignExt = true;
8795 // Either the case of Base + Index (no offset) or something else.
8796 if (IndexOffset->getOpcode() != ISD::ADD)
8797 return BaseIndexOffset(Base, IndexOffset, 0, IsIndexSignExt);
8799 // Now we have the case of Base + Index + offset.
8800 SDValue Index = IndexOffset->getOperand(0);
8801 SDValue Offset = IndexOffset->getOperand(1);
8803 if (!isa<ConstantSDNode>(Offset))
8804 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
8806 // Ignore signextends.
8807 if (Index->getOpcode() == ISD::SIGN_EXTEND) {
8808 Index = Index->getOperand(0);
8809 IsIndexSignExt = true;
8810 } else IsIndexSignExt = false;
8812 int64_t Off = cast<ConstantSDNode>(Offset)->getSExtValue();
8813 return BaseIndexOffset(Base, Index, Off, IsIndexSignExt);
8817 /// Holds a pointer to an LSBaseSDNode as well as information on where it
8818 /// is located in a sequence of memory operations connected by a chain.
8820 MemOpLink (LSBaseSDNode *N, int64_t Offset, unsigned Seq):
8821 MemNode(N), OffsetFromBase(Offset), SequenceNum(Seq) { }
8822 // Ptr to the mem node.
8823 LSBaseSDNode *MemNode;
8824 // Offset from the base ptr.
8825 int64_t OffsetFromBase;
8826 // What is the sequence number of this mem node.
8827 // Lowest mem operand in the DAG starts at zero.
8828 unsigned SequenceNum;
8831 /// Sorts store nodes in a link according to their offset from a shared
8833 struct ConsecutiveMemoryChainSorter {
8834 bool operator()(MemOpLink LHS, MemOpLink RHS) {
8836 LHS.OffsetFromBase < RHS.OffsetFromBase ||
8837 (LHS.OffsetFromBase == RHS.OffsetFromBase &&
8838 LHS.SequenceNum > RHS.SequenceNum);
8842 bool DAGCombiner::MergeConsecutiveStores(StoreSDNode* St) {
8843 EVT MemVT = St->getMemoryVT();
8844 int64_t ElementSizeBytes = MemVT.getSizeInBits()/8;
8845 bool NoVectors = DAG.getMachineFunction().getFunction()->getAttributes().
8846 hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat);
8848 // Don't merge vectors into wider inputs.
8849 if (MemVT.isVector() || !MemVT.isSimple())
8852 // Perform an early exit check. Do not bother looking at stored values that
8853 // are not constants or loads.
8854 SDValue StoredVal = St->getValue();
8855 bool IsLoadSrc = isa<LoadSDNode>(StoredVal);
8856 if (!isa<ConstantSDNode>(StoredVal) && !isa<ConstantFPSDNode>(StoredVal) &&
8860 // Only look at ends of store sequences.
8861 SDValue Chain = SDValue(St, 1);
8862 if (Chain->hasOneUse() && Chain->use_begin()->getOpcode() == ISD::STORE)
8865 // This holds the base pointer, index, and the offset in bytes from the base
8867 BaseIndexOffset BasePtr = BaseIndexOffset::match(St->getBasePtr());
8869 // We must have a base and an offset.
8870 if (!BasePtr.Base.getNode())
8873 // Do not handle stores to undef base pointers.
8874 if (BasePtr.Base.getOpcode() == ISD::UNDEF)
8877 // Save the LoadSDNodes that we find in the chain.
8878 // We need to make sure that these nodes do not interfere with
8879 // any of the store nodes.
8880 SmallVector<LSBaseSDNode*, 8> AliasLoadNodes;
8882 // Save the StoreSDNodes that we find in the chain.
8883 SmallVector<MemOpLink, 8> StoreNodes;
8885 // Walk up the chain and look for nodes with offsets from the same
8886 // base pointer. Stop when reaching an instruction with a different kind
8887 // or instruction which has a different base pointer.
8889 StoreSDNode *Index = St;
8891 // If the chain has more than one use, then we can't reorder the mem ops.
8892 if (Index != St && !SDValue(Index, 1)->hasOneUse())
8895 // Find the base pointer and offset for this memory node.
8896 BaseIndexOffset Ptr = BaseIndexOffset::match(Index->getBasePtr());
8898 // Check that the base pointer is the same as the original one.
8899 if (!Ptr.equalBaseIndex(BasePtr))
8902 // Check that the alignment is the same.
8903 if (Index->getAlignment() != St->getAlignment())
8906 // The memory operands must not be volatile.
8907 if (Index->isVolatile() || Index->isIndexed())
8911 if (StoreSDNode *St = dyn_cast<StoreSDNode>(Index))
8912 if (St->isTruncatingStore())
8915 // The stored memory type must be the same.
8916 if (Index->getMemoryVT() != MemVT)
8919 // We do not allow unaligned stores because we want to prevent overriding
8921 if (Index->getAlignment()*8 != MemVT.getSizeInBits())
8924 // We found a potential memory operand to merge.
8925 StoreNodes.push_back(MemOpLink(Index, Ptr.Offset, Seq++));
8927 // Find the next memory operand in the chain. If the next operand in the
8928 // chain is a store then move up and continue the scan with the next
8929 // memory operand. If the next operand is a load save it and use alias
8930 // information to check if it interferes with anything.
8931 SDNode *NextInChain = Index->getChain().getNode();
8933 if (StoreSDNode *STn = dyn_cast<StoreSDNode>(NextInChain)) {
8934 // We found a store node. Use it for the next iteration.
8937 } else if (LoadSDNode *Ldn = dyn_cast<LoadSDNode>(NextInChain)) {
8938 if (Ldn->isVolatile()) {
8943 // Save the load node for later. Continue the scan.
8944 AliasLoadNodes.push_back(Ldn);
8945 NextInChain = Ldn->getChain().getNode();
8954 // Check if there is anything to merge.
8955 if (StoreNodes.size() < 2)
8958 // Sort the memory operands according to their distance from the base pointer.
8959 std::sort(StoreNodes.begin(), StoreNodes.end(),
8960 ConsecutiveMemoryChainSorter());
8962 // Scan the memory operations on the chain and find the first non-consecutive
8963 // store memory address.
8964 unsigned LastConsecutiveStore = 0;
8965 int64_t StartAddress = StoreNodes[0].OffsetFromBase;
8966 for (unsigned i = 0, e = StoreNodes.size(); i < e; ++i) {
8968 // Check that the addresses are consecutive starting from the second
8969 // element in the list of stores.
8971 int64_t CurrAddress = StoreNodes[i].OffsetFromBase;
8972 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
8977 // Check if this store interferes with any of the loads that we found.
8978 for (unsigned ld = 0, lde = AliasLoadNodes.size(); ld < lde; ++ld)
8979 if (isAlias(AliasLoadNodes[ld], StoreNodes[i].MemNode)) {
8983 // We found a load that alias with this store. Stop the sequence.
8987 // Mark this node as useful.
8988 LastConsecutiveStore = i;
8991 // The node with the lowest store address.
8992 LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode;
8994 // Store the constants into memory as one consecutive store.
8996 unsigned LastLegalType = 0;
8997 unsigned LastLegalVectorType = 0;
8998 bool NonZero = false;
8999 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
9000 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
9001 SDValue StoredVal = St->getValue();
9003 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(StoredVal)) {
9004 NonZero |= !C->isNullValue();
9005 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(StoredVal)) {
9006 NonZero |= !C->getConstantFPValue()->isNullValue();
9012 // Find a legal type for the constant store.
9013 unsigned StoreBW = (i+1) * ElementSizeBytes * 8;
9014 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
9015 if (TLI.isTypeLegal(StoreTy))
9016 LastLegalType = i+1;
9017 // Or check whether a truncstore is legal.
9018 else if (TLI.getTypeAction(*DAG.getContext(), StoreTy) ==
9019 TargetLowering::TypePromoteInteger) {
9020 EVT LegalizedStoredValueTy =
9021 TLI.getTypeToTransformTo(*DAG.getContext(), StoredVal.getValueType());
9022 if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy))
9023 LastLegalType = i+1;
9026 // Find a legal type for the vector store.
9027 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
9028 if (TLI.isTypeLegal(Ty))
9029 LastLegalVectorType = i + 1;
9032 // We only use vectors if the constant is known to be zero and the
9033 // function is not marked with the noimplicitfloat attribute.
9034 if (NonZero || NoVectors)
9035 LastLegalVectorType = 0;
9037 // Check if we found a legal integer type to store.
9038 if (LastLegalType == 0 && LastLegalVectorType == 0)
9041 bool UseVector = (LastLegalVectorType > LastLegalType) && !NoVectors;
9042 unsigned NumElem = UseVector ? LastLegalVectorType : LastLegalType;
9044 // Make sure we have something to merge.
9048 unsigned EarliestNodeUsed = 0;
9049 for (unsigned i=0; i < NumElem; ++i) {
9050 // Find a chain for the new wide-store operand. Notice that some
9051 // of the store nodes that we found may not be selected for inclusion
9052 // in the wide store. The chain we use needs to be the chain of the
9053 // earliest store node which is *used* and replaced by the wide store.
9054 if (StoreNodes[i].SequenceNum > StoreNodes[EarliestNodeUsed].SequenceNum)
9055 EarliestNodeUsed = i;
9058 // The earliest Node in the DAG.
9059 LSBaseSDNode *EarliestOp = StoreNodes[EarliestNodeUsed].MemNode;
9060 SDLoc DL(StoreNodes[0].MemNode);
9064 // Find a legal type for the vector store.
9065 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
9066 assert(TLI.isTypeLegal(Ty) && "Illegal vector store");
9067 StoredVal = DAG.getConstant(0, Ty);
9069 unsigned StoreBW = NumElem * ElementSizeBytes * 8;
9070 APInt StoreInt(StoreBW, 0);
9072 // Construct a single integer constant which is made of the smaller
9074 bool IsLE = TLI.isLittleEndian();
9075 for (unsigned i = 0; i < NumElem ; ++i) {
9076 unsigned Idx = IsLE ?(NumElem - 1 - i) : i;
9077 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[Idx].MemNode);
9078 SDValue Val = St->getValue();
9079 StoreInt<<=ElementSizeBytes*8;
9080 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val)) {
9081 StoreInt|=C->getAPIntValue().zext(StoreBW);
9082 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Val)) {
9083 StoreInt|= C->getValueAPF().bitcastToAPInt().zext(StoreBW);
9085 assert(false && "Invalid constant element type");
9089 // Create the new Load and Store operations.
9090 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
9091 StoredVal = DAG.getConstant(StoreInt, StoreTy);
9094 SDValue NewStore = DAG.getStore(EarliestOp->getChain(), DL, StoredVal,
9095 FirstInChain->getBasePtr(),
9096 FirstInChain->getPointerInfo(),
9098 FirstInChain->getAlignment());
9100 // Replace the first store with the new store
9101 CombineTo(EarliestOp, NewStore);
9102 // Erase all other stores.
9103 for (unsigned i = 0; i < NumElem ; ++i) {
9104 if (StoreNodes[i].MemNode == EarliestOp)
9106 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
9107 // ReplaceAllUsesWith will replace all uses that existed when it was
9108 // called, but graph optimizations may cause new ones to appear. For
9109 // example, the case in pr14333 looks like
9111 // St's chain -> St -> another store -> X
9113 // And the only difference from St to the other store is the chain.
9114 // When we change it's chain to be St's chain they become identical,
9115 // get CSEed and the net result is that X is now a use of St.
9116 // Since we know that St is redundant, just iterate.
9117 while (!St->use_empty())
9118 DAG.ReplaceAllUsesWith(SDValue(St, 0), St->getChain());
9119 removeFromWorkList(St);
9126 // Below we handle the case of multiple consecutive stores that
9127 // come from multiple consecutive loads. We merge them into a single
9128 // wide load and a single wide store.
9130 // Look for load nodes which are used by the stored values.
9131 SmallVector<MemOpLink, 8> LoadNodes;
9133 // Find acceptable loads. Loads need to have the same chain (token factor),
9134 // must not be zext, volatile, indexed, and they must be consecutive.
9135 BaseIndexOffset LdBasePtr;
9136 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
9137 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
9138 LoadSDNode *Ld = dyn_cast<LoadSDNode>(St->getValue());
9141 // Loads must only have one use.
9142 if (!Ld->hasNUsesOfValue(1, 0))
9145 // Check that the alignment is the same as the stores.
9146 if (Ld->getAlignment() != St->getAlignment())
9149 // The memory operands must not be volatile.
9150 if (Ld->isVolatile() || Ld->isIndexed())
9153 // We do not accept ext loads.
9154 if (Ld->getExtensionType() != ISD::NON_EXTLOAD)
9157 // The stored memory type must be the same.
9158 if (Ld->getMemoryVT() != MemVT)
9161 BaseIndexOffset LdPtr = BaseIndexOffset::match(Ld->getBasePtr());
9162 // If this is not the first ptr that we check.
9163 if (LdBasePtr.Base.getNode()) {
9164 // The base ptr must be the same.
9165 if (!LdPtr.equalBaseIndex(LdBasePtr))
9168 // Check that all other base pointers are the same as this one.
9172 // We found a potential memory operand to merge.
9173 LoadNodes.push_back(MemOpLink(Ld, LdPtr.Offset, 0));
9176 if (LoadNodes.size() < 2)
9179 // Scan the memory operations on the chain and find the first non-consecutive
9180 // load memory address. These variables hold the index in the store node
9182 unsigned LastConsecutiveLoad = 0;
9183 // This variable refers to the size and not index in the array.
9184 unsigned LastLegalVectorType = 0;
9185 unsigned LastLegalIntegerType = 0;
9186 StartAddress = LoadNodes[0].OffsetFromBase;
9187 SDValue FirstChain = LoadNodes[0].MemNode->getChain();
9188 for (unsigned i = 1; i < LoadNodes.size(); ++i) {
9189 // All loads much share the same chain.
9190 if (LoadNodes[i].MemNode->getChain() != FirstChain)
9193 int64_t CurrAddress = LoadNodes[i].OffsetFromBase;
9194 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
9196 LastConsecutiveLoad = i;
9198 // Find a legal type for the vector store.
9199 EVT StoreTy = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
9200 if (TLI.isTypeLegal(StoreTy))
9201 LastLegalVectorType = i + 1;
9203 // Find a legal type for the integer store.
9204 unsigned StoreBW = (i+1) * ElementSizeBytes * 8;
9205 StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
9206 if (TLI.isTypeLegal(StoreTy))
9207 LastLegalIntegerType = i + 1;
9208 // Or check whether a truncstore and extload is legal.
9209 else if (TLI.getTypeAction(*DAG.getContext(), StoreTy) ==
9210 TargetLowering::TypePromoteInteger) {
9211 EVT LegalizedStoredValueTy =
9212 TLI.getTypeToTransformTo(*DAG.getContext(), StoreTy);
9213 if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy) &&
9214 TLI.isLoadExtLegal(ISD::ZEXTLOAD, StoreTy) &&
9215 TLI.isLoadExtLegal(ISD::SEXTLOAD, StoreTy) &&
9216 TLI.isLoadExtLegal(ISD::EXTLOAD, StoreTy))
9217 LastLegalIntegerType = i+1;
9221 // Only use vector types if the vector type is larger than the integer type.
9222 // If they are the same, use integers.
9223 bool UseVectorTy = LastLegalVectorType > LastLegalIntegerType && !NoVectors;
9224 unsigned LastLegalType = std::max(LastLegalVectorType, LastLegalIntegerType);
9226 // We add +1 here because the LastXXX variables refer to location while
9227 // the NumElem refers to array/index size.
9228 unsigned NumElem = std::min(LastConsecutiveStore, LastConsecutiveLoad) + 1;
9229 NumElem = std::min(LastLegalType, NumElem);
9234 // The earliest Node in the DAG.
9235 unsigned EarliestNodeUsed = 0;
9236 LSBaseSDNode *EarliestOp = StoreNodes[EarliestNodeUsed].MemNode;
9237 for (unsigned i=1; i<NumElem; ++i) {
9238 // Find a chain for the new wide-store operand. Notice that some
9239 // of the store nodes that we found may not be selected for inclusion
9240 // in the wide store. The chain we use needs to be the chain of the
9241 // earliest store node which is *used* and replaced by the wide store.
9242 if (StoreNodes[i].SequenceNum > StoreNodes[EarliestNodeUsed].SequenceNum)
9243 EarliestNodeUsed = i;
9246 // Find if it is better to use vectors or integers to load and store
9250 JointMemOpVT = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
9252 unsigned StoreBW = NumElem * ElementSizeBytes * 8;
9253 JointMemOpVT = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
9256 SDLoc LoadDL(LoadNodes[0].MemNode);
9257 SDLoc StoreDL(StoreNodes[0].MemNode);
9259 LoadSDNode *FirstLoad = cast<LoadSDNode>(LoadNodes[0].MemNode);
9260 SDValue NewLoad = DAG.getLoad(JointMemOpVT, LoadDL,
9261 FirstLoad->getChain(),
9262 FirstLoad->getBasePtr(),
9263 FirstLoad->getPointerInfo(),
9264 false, false, false,
9265 FirstLoad->getAlignment());
9267 SDValue NewStore = DAG.getStore(EarliestOp->getChain(), StoreDL, NewLoad,
9268 FirstInChain->getBasePtr(),
9269 FirstInChain->getPointerInfo(), false, false,
9270 FirstInChain->getAlignment());
9272 // Replace one of the loads with the new load.
9273 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[0].MemNode);
9274 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1),
9275 SDValue(NewLoad.getNode(), 1));
9277 // Remove the rest of the load chains.
9278 for (unsigned i = 1; i < NumElem ; ++i) {
9279 // Replace all chain users of the old load nodes with the chain of the new
9281 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[i].MemNode);
9282 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Ld->getChain());
9285 // Replace the first store with the new store.
9286 CombineTo(EarliestOp, NewStore);
9287 // Erase all other stores.
9288 for (unsigned i = 0; i < NumElem ; ++i) {
9289 // Remove all Store nodes.
9290 if (StoreNodes[i].MemNode == EarliestOp)
9292 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
9293 DAG.ReplaceAllUsesOfValueWith(SDValue(St, 0), St->getChain());
9294 removeFromWorkList(St);
9301 SDValue DAGCombiner::visitSTORE(SDNode *N) {
9302 StoreSDNode *ST = cast<StoreSDNode>(N);
9303 SDValue Chain = ST->getChain();
9304 SDValue Value = ST->getValue();
9305 SDValue Ptr = ST->getBasePtr();
9307 // If this is a store of a bit convert, store the input value if the
9308 // resultant store does not need a higher alignment than the original.
9309 if (Value.getOpcode() == ISD::BITCAST && !ST->isTruncatingStore() &&
9310 ST->isUnindexed()) {
9311 unsigned OrigAlign = ST->getAlignment();
9312 EVT SVT = Value.getOperand(0).getValueType();
9313 unsigned Align = TLI.getDataLayout()->
9314 getABITypeAlignment(SVT.getTypeForEVT(*DAG.getContext()));
9315 if (Align <= OrigAlign &&
9316 ((!LegalOperations && !ST->isVolatile()) ||
9317 TLI.isOperationLegalOrCustom(ISD::STORE, SVT)))
9318 return DAG.getStore(Chain, SDLoc(N), Value.getOperand(0),
9319 Ptr, ST->getPointerInfo(), ST->isVolatile(),
9320 ST->isNonTemporal(), OrigAlign,
9324 // Turn 'store undef, Ptr' -> nothing.
9325 if (Value.getOpcode() == ISD::UNDEF && ST->isUnindexed())
9328 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
9329 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Value)) {
9330 // NOTE: If the original store is volatile, this transform must not increase
9331 // the number of stores. For example, on x86-32 an f64 can be stored in one
9332 // processor operation but an i64 (which is not legal) requires two. So the
9333 // transform should not be done in this case.
9334 if (Value.getOpcode() != ISD::TargetConstantFP) {
9336 switch (CFP->getSimpleValueType(0).SimpleTy) {
9337 default: llvm_unreachable("Unknown FP type");
9338 case MVT::f16: // We don't do this for these yet.
9344 if ((isTypeLegal(MVT::i32) && !LegalOperations && !ST->isVolatile()) ||
9345 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
9346 Tmp = DAG.getConstant((uint32_t)CFP->getValueAPF().
9347 bitcastToAPInt().getZExtValue(), MVT::i32);
9348 return DAG.getStore(Chain, SDLoc(N), Tmp,
9349 Ptr, ST->getMemOperand());
9353 if ((TLI.isTypeLegal(MVT::i64) && !LegalOperations &&
9354 !ST->isVolatile()) ||
9355 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i64)) {
9356 Tmp = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
9357 getZExtValue(), MVT::i64);
9358 return DAG.getStore(Chain, SDLoc(N), Tmp,
9359 Ptr, ST->getMemOperand());
9362 if (!ST->isVolatile() &&
9363 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
9364 // Many FP stores are not made apparent until after legalize, e.g. for
9365 // argument passing. Since this is so common, custom legalize the
9366 // 64-bit integer store into two 32-bit stores.
9367 uint64_t Val = CFP->getValueAPF().bitcastToAPInt().getZExtValue();
9368 SDValue Lo = DAG.getConstant(Val & 0xFFFFFFFF, MVT::i32);
9369 SDValue Hi = DAG.getConstant(Val >> 32, MVT::i32);
9370 if (TLI.isBigEndian()) std::swap(Lo, Hi);
9372 unsigned Alignment = ST->getAlignment();
9373 bool isVolatile = ST->isVolatile();
9374 bool isNonTemporal = ST->isNonTemporal();
9375 const MDNode *TBAAInfo = ST->getTBAAInfo();
9377 SDValue St0 = DAG.getStore(Chain, SDLoc(ST), Lo,
9378 Ptr, ST->getPointerInfo(),
9379 isVolatile, isNonTemporal,
9380 ST->getAlignment(), TBAAInfo);
9381 Ptr = DAG.getNode(ISD::ADD, SDLoc(N), Ptr.getValueType(), Ptr,
9382 DAG.getConstant(4, Ptr.getValueType()));
9383 Alignment = MinAlign(Alignment, 4U);
9384 SDValue St1 = DAG.getStore(Chain, SDLoc(ST), Hi,
9385 Ptr, ST->getPointerInfo().getWithOffset(4),
9386 isVolatile, isNonTemporal,
9387 Alignment, TBAAInfo);
9388 return DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other,
9397 // Try to infer better alignment information than the store already has.
9398 if (OptLevel != CodeGenOpt::None && ST->isUnindexed()) {
9399 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
9400 if (Align > ST->getAlignment())
9401 return DAG.getTruncStore(Chain, SDLoc(N), Value,
9402 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
9403 ST->isVolatile(), ST->isNonTemporal(), Align,
9408 // Try transforming a pair floating point load / store ops to integer
9409 // load / store ops.
9410 SDValue NewST = TransformFPLoadStorePair(N);
9411 if (NewST.getNode())
9414 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA :
9415 TLI.getTargetMachine().getSubtarget<TargetSubtargetInfo>().useAA();
9417 if (CombinerAAOnlyFunc.getNumOccurrences() &&
9418 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
9421 if (UseAA && ST->isUnindexed()) {
9422 // Walk up chain skipping non-aliasing memory nodes.
9423 SDValue BetterChain = FindBetterChain(N, Chain);
9425 // If there is a better chain.
9426 if (Chain != BetterChain) {
9429 // Replace the chain to avoid dependency.
9430 if (ST->isTruncatingStore()) {
9431 ReplStore = DAG.getTruncStore(BetterChain, SDLoc(N), Value, Ptr,
9432 ST->getMemoryVT(), ST->getMemOperand());
9434 ReplStore = DAG.getStore(BetterChain, SDLoc(N), Value, Ptr,
9435 ST->getMemOperand());
9438 // Create token to keep both nodes around.
9439 SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N),
9440 MVT::Other, Chain, ReplStore);
9442 // Make sure the new and old chains are cleaned up.
9443 AddToWorkList(Token.getNode());
9445 // Don't add users to work list.
9446 return CombineTo(N, Token, false);
9450 // Try transforming N to an indexed store.
9451 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
9452 return SDValue(N, 0);
9454 // FIXME: is there such a thing as a truncating indexed store?
9455 if (ST->isTruncatingStore() && ST->isUnindexed() &&
9456 Value.getValueType().isInteger()) {
9457 // See if we can simplify the input to this truncstore with knowledge that
9458 // only the low bits are being used. For example:
9459 // "truncstore (or (shl x, 8), y), i8" -> "truncstore y, i8"
9461 GetDemandedBits(Value,
9462 APInt::getLowBitsSet(
9463 Value.getValueType().getScalarType().getSizeInBits(),
9464 ST->getMemoryVT().getScalarType().getSizeInBits()));
9465 AddToWorkList(Value.getNode());
9466 if (Shorter.getNode())
9467 return DAG.getTruncStore(Chain, SDLoc(N), Shorter,
9468 Ptr, ST->getMemoryVT(), ST->getMemOperand());
9470 // Otherwise, see if we can simplify the operation with
9471 // SimplifyDemandedBits, which only works if the value has a single use.
9472 if (SimplifyDemandedBits(Value,
9473 APInt::getLowBitsSet(
9474 Value.getValueType().getScalarType().getSizeInBits(),
9475 ST->getMemoryVT().getScalarType().getSizeInBits())))
9476 return SDValue(N, 0);
9479 // If this is a load followed by a store to the same location, then the store
9481 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Value)) {
9482 if (Ld->getBasePtr() == Ptr && ST->getMemoryVT() == Ld->getMemoryVT() &&
9483 ST->isUnindexed() && !ST->isVolatile() &&
9484 // There can't be any side effects between the load and store, such as
9486 Chain.reachesChainWithoutSideEffects(SDValue(Ld, 1))) {
9487 // The store is dead, remove it.
9492 // If this is an FP_ROUND or TRUNC followed by a store, fold this into a
9493 // truncating store. We can do this even if this is already a truncstore.
9494 if ((Value.getOpcode() == ISD::FP_ROUND || Value.getOpcode() == ISD::TRUNCATE)
9495 && Value.getNode()->hasOneUse() && ST->isUnindexed() &&
9496 TLI.isTruncStoreLegal(Value.getOperand(0).getValueType(),
9497 ST->getMemoryVT())) {
9498 return DAG.getTruncStore(Chain, SDLoc(N), Value.getOperand(0),
9499 Ptr, ST->getMemoryVT(), ST->getMemOperand());
9502 // Only perform this optimization before the types are legal, because we
9503 // don't want to perform this optimization on every DAGCombine invocation.
9505 bool EverChanged = false;
9508 // There can be multiple store sequences on the same chain.
9509 // Keep trying to merge store sequences until we are unable to do so
9510 // or until we merge the last store on the chain.
9511 bool Changed = MergeConsecutiveStores(ST);
9512 EverChanged |= Changed;
9513 if (!Changed) break;
9514 } while (ST->getOpcode() != ISD::DELETED_NODE);
9517 return SDValue(N, 0);
9520 return ReduceLoadOpStoreWidth(N);
9523 SDValue DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) {
9524 SDValue InVec = N->getOperand(0);
9525 SDValue InVal = N->getOperand(1);
9526 SDValue EltNo = N->getOperand(2);
9529 // If the inserted element is an UNDEF, just use the input vector.
9530 if (InVal.getOpcode() == ISD::UNDEF)
9533 EVT VT = InVec.getValueType();
9535 // If we can't generate a legal BUILD_VECTOR, exit
9536 if (LegalOperations && !TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
9539 // Check that we know which element is being inserted
9540 if (!isa<ConstantSDNode>(EltNo))
9542 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
9544 // Check that the operand is a BUILD_VECTOR (or UNDEF, which can essentially
9545 // be converted to a BUILD_VECTOR). Fill in the Ops vector with the
9547 SmallVector<SDValue, 8> Ops;
9548 // Do not combine these two vectors if the output vector will not replace
9549 // the input vector.
9550 if (InVec.getOpcode() == ISD::BUILD_VECTOR && InVec.hasOneUse()) {
9551 Ops.append(InVec.getNode()->op_begin(),
9552 InVec.getNode()->op_end());
9553 } else if (InVec.getOpcode() == ISD::UNDEF) {
9554 unsigned NElts = VT.getVectorNumElements();
9555 Ops.append(NElts, DAG.getUNDEF(InVal.getValueType()));
9560 // Insert the element
9561 if (Elt < Ops.size()) {
9562 // All the operands of BUILD_VECTOR must have the same type;
9563 // we enforce that here.
9564 EVT OpVT = Ops[0].getValueType();
9565 if (InVal.getValueType() != OpVT)
9566 InVal = OpVT.bitsGT(InVal.getValueType()) ?
9567 DAG.getNode(ISD::ANY_EXTEND, dl, OpVT, InVal) :
9568 DAG.getNode(ISD::TRUNCATE, dl, OpVT, InVal);
9572 // Return the new vector
9573 return DAG.getNode(ISD::BUILD_VECTOR, dl,
9574 VT, &Ops[0], Ops.size());
9577 SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) {
9578 // (vextract (scalar_to_vector val, 0) -> val
9579 SDValue InVec = N->getOperand(0);
9580 EVT VT = InVec.getValueType();
9581 EVT NVT = N->getValueType(0);
9583 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
9584 // Check if the result type doesn't match the inserted element type. A
9585 // SCALAR_TO_VECTOR may truncate the inserted element and the
9586 // EXTRACT_VECTOR_ELT may widen the extracted vector.
9587 SDValue InOp = InVec.getOperand(0);
9588 if (InOp.getValueType() != NVT) {
9589 assert(InOp.getValueType().isInteger() && NVT.isInteger());
9590 return DAG.getSExtOrTrunc(InOp, SDLoc(InVec), NVT);
9595 SDValue EltNo = N->getOperand(1);
9596 bool ConstEltNo = isa<ConstantSDNode>(EltNo);
9598 // Transform: (EXTRACT_VECTOR_ELT( VECTOR_SHUFFLE )) -> EXTRACT_VECTOR_ELT.
9599 // We only perform this optimization before the op legalization phase because
9600 // we may introduce new vector instructions which are not backed by TD
9601 // patterns. For example on AVX, extracting elements from a wide vector
9602 // without using extract_subvector.
9603 if (InVec.getOpcode() == ISD::VECTOR_SHUFFLE
9604 && ConstEltNo && !LegalOperations) {
9605 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
9606 int NumElem = VT.getVectorNumElements();
9607 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(InVec);
9608 // Find the new index to extract from.
9609 int OrigElt = SVOp->getMaskElt(Elt);
9611 // Extracting an undef index is undef.
9613 return DAG.getUNDEF(NVT);
9615 // Select the right vector half to extract from.
9616 if (OrigElt < NumElem) {
9617 InVec = InVec->getOperand(0);
9619 InVec = InVec->getOperand(1);
9623 EVT IndexTy = TLI.getVectorIdxTy();
9624 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(N), NVT,
9625 InVec, DAG.getConstant(OrigElt, IndexTy));
9628 // Perform only after legalization to ensure build_vector / vector_shuffle
9629 // optimizations have already been done.
9630 if (!LegalOperations) return SDValue();
9632 // (vextract (v4f32 load $addr), c) -> (f32 load $addr+c*size)
9633 // (vextract (v4f32 s2v (f32 load $addr)), c) -> (f32 load $addr+c*size)
9634 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), 0) -> (f32 load $addr)
9637 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
9638 bool NewLoad = false;
9639 bool BCNumEltsChanged = false;
9640 EVT ExtVT = VT.getVectorElementType();
9643 // If the result of load has to be truncated, then it's not necessarily
9645 if (NVT.bitsLT(LVT) && !TLI.isTruncateFree(LVT, NVT))
9648 if (InVec.getOpcode() == ISD::BITCAST) {
9649 // Don't duplicate a load with other uses.
9650 if (!InVec.hasOneUse())
9653 EVT BCVT = InVec.getOperand(0).getValueType();
9654 if (!BCVT.isVector() || ExtVT.bitsGT(BCVT.getVectorElementType()))
9656 if (VT.getVectorNumElements() != BCVT.getVectorNumElements())
9657 BCNumEltsChanged = true;
9658 InVec = InVec.getOperand(0);
9659 ExtVT = BCVT.getVectorElementType();
9663 LoadSDNode *LN0 = NULL;
9664 const ShuffleVectorSDNode *SVN = NULL;
9665 if (ISD::isNormalLoad(InVec.getNode())) {
9666 LN0 = cast<LoadSDNode>(InVec);
9667 } else if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR &&
9668 InVec.getOperand(0).getValueType() == ExtVT &&
9669 ISD::isNormalLoad(InVec.getOperand(0).getNode())) {
9670 // Don't duplicate a load with other uses.
9671 if (!InVec.hasOneUse())
9674 LN0 = cast<LoadSDNode>(InVec.getOperand(0));
9675 } else if ((SVN = dyn_cast<ShuffleVectorSDNode>(InVec))) {
9676 // (vextract (vector_shuffle (load $addr), v2, <1, u, u, u>), 1)
9678 // (load $addr+1*size)
9680 // Don't duplicate a load with other uses.
9681 if (!InVec.hasOneUse())
9684 // If the bit convert changed the number of elements, it is unsafe
9685 // to examine the mask.
9686 if (BCNumEltsChanged)
9689 // Select the input vector, guarding against out of range extract vector.
9690 unsigned NumElems = VT.getVectorNumElements();
9691 int Idx = (Elt > (int)NumElems) ? -1 : SVN->getMaskElt(Elt);
9692 InVec = (Idx < (int)NumElems) ? InVec.getOperand(0) : InVec.getOperand(1);
9694 if (InVec.getOpcode() == ISD::BITCAST) {
9695 // Don't duplicate a load with other uses.
9696 if (!InVec.hasOneUse())
9699 InVec = InVec.getOperand(0);
9701 if (ISD::isNormalLoad(InVec.getNode())) {
9702 LN0 = cast<LoadSDNode>(InVec);
9703 Elt = (Idx < (int)NumElems) ? Idx : Idx - (int)NumElems;
9707 // Make sure we found a non-volatile load and the extractelement is
9709 if (!LN0 || !LN0->hasNUsesOfValue(1,0) || LN0->isVolatile())
9712 // If Idx was -1 above, Elt is going to be -1, so just return undef.
9714 return DAG.getUNDEF(LVT);
9716 unsigned Align = LN0->getAlignment();
9718 // Check the resultant load doesn't need a higher alignment than the
9722 ->getABITypeAlignment(LVT.getTypeForEVT(*DAG.getContext()));
9724 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, LVT))
9730 SDValue NewPtr = LN0->getBasePtr();
9731 unsigned PtrOff = 0;
9734 PtrOff = LVT.getSizeInBits() * Elt / 8;
9735 EVT PtrType = NewPtr.getValueType();
9736 if (TLI.isBigEndian())
9737 PtrOff = VT.getSizeInBits() / 8 - PtrOff;
9738 NewPtr = DAG.getNode(ISD::ADD, SDLoc(N), PtrType, NewPtr,
9739 DAG.getConstant(PtrOff, PtrType));
9742 // The replacement we need to do here is a little tricky: we need to
9743 // replace an extractelement of a load with a load.
9744 // Use ReplaceAllUsesOfValuesWith to do the replacement.
9745 // Note that this replacement assumes that the extractvalue is the only
9746 // use of the load; that's okay because we don't want to perform this
9747 // transformation in other cases anyway.
9750 if (NVT.bitsGT(LVT)) {
9751 // If the result type of vextract is wider than the load, then issue an
9752 // extending load instead.
9753 ISD::LoadExtType ExtType = TLI.isLoadExtLegal(ISD::ZEXTLOAD, LVT)
9754 ? ISD::ZEXTLOAD : ISD::EXTLOAD;
9755 Load = DAG.getExtLoad(ExtType, SDLoc(N), NVT, LN0->getChain(),
9756 NewPtr, LN0->getPointerInfo().getWithOffset(PtrOff),
9757 LVT, LN0->isVolatile(), LN0->isNonTemporal(),
9758 Align, LN0->getTBAAInfo());
9759 Chain = Load.getValue(1);
9761 Load = DAG.getLoad(LVT, SDLoc(N), LN0->getChain(), NewPtr,
9762 LN0->getPointerInfo().getWithOffset(PtrOff),
9763 LN0->isVolatile(), LN0->isNonTemporal(),
9764 LN0->isInvariant(), Align, LN0->getTBAAInfo());
9765 Chain = Load.getValue(1);
9766 if (NVT.bitsLT(LVT))
9767 Load = DAG.getNode(ISD::TRUNCATE, SDLoc(N), NVT, Load);
9769 Load = DAG.getNode(ISD::BITCAST, SDLoc(N), NVT, Load);
9771 WorkListRemover DeadNodes(*this);
9772 SDValue From[] = { SDValue(N, 0), SDValue(LN0,1) };
9773 SDValue To[] = { Load, Chain };
9774 DAG.ReplaceAllUsesOfValuesWith(From, To, 2);
9775 // Since we're explcitly calling ReplaceAllUses, add the new node to the
9776 // worklist explicitly as well.
9777 AddToWorkList(Load.getNode());
9778 AddUsersToWorkList(Load.getNode()); // Add users too
9779 // Make sure to revisit this node to clean it up; it will usually be dead.
9781 return SDValue(N, 0);
9787 // Simplify (build_vec (ext )) to (bitcast (build_vec ))
9788 SDValue DAGCombiner::reduceBuildVecExtToExtBuildVec(SDNode *N) {
9789 // We perform this optimization post type-legalization because
9790 // the type-legalizer often scalarizes integer-promoted vectors.
9791 // Performing this optimization before may create bit-casts which
9792 // will be type-legalized to complex code sequences.
9793 // We perform this optimization only before the operation legalizer because we
9794 // may introduce illegal operations.
9795 if (Level != AfterLegalizeVectorOps && Level != AfterLegalizeTypes)
9798 unsigned NumInScalars = N->getNumOperands();
9800 EVT VT = N->getValueType(0);
9802 // Check to see if this is a BUILD_VECTOR of a bunch of values
9803 // which come from any_extend or zero_extend nodes. If so, we can create
9804 // a new BUILD_VECTOR using bit-casts which may enable other BUILD_VECTOR
9805 // optimizations. We do not handle sign-extend because we can't fill the sign
9807 EVT SourceType = MVT::Other;
9808 bool AllAnyExt = true;
9810 for (unsigned i = 0; i != NumInScalars; ++i) {
9811 SDValue In = N->getOperand(i);
9812 // Ignore undef inputs.
9813 if (In.getOpcode() == ISD::UNDEF) continue;
9815 bool AnyExt = In.getOpcode() == ISD::ANY_EXTEND;
9816 bool ZeroExt = In.getOpcode() == ISD::ZERO_EXTEND;
9818 // Abort if the element is not an extension.
9819 if (!ZeroExt && !AnyExt) {
9820 SourceType = MVT::Other;
9824 // The input is a ZeroExt or AnyExt. Check the original type.
9825 EVT InTy = In.getOperand(0).getValueType();
9827 // Check that all of the widened source types are the same.
9828 if (SourceType == MVT::Other)
9831 else if (InTy != SourceType) {
9832 // Multiple income types. Abort.
9833 SourceType = MVT::Other;
9837 // Check if all of the extends are ANY_EXTENDs.
9838 AllAnyExt &= AnyExt;
9841 // In order to have valid types, all of the inputs must be extended from the
9842 // same source type and all of the inputs must be any or zero extend.
9843 // Scalar sizes must be a power of two.
9844 EVT OutScalarTy = VT.getScalarType();
9845 bool ValidTypes = SourceType != MVT::Other &&
9846 isPowerOf2_32(OutScalarTy.getSizeInBits()) &&
9847 isPowerOf2_32(SourceType.getSizeInBits());
9849 // Create a new simpler BUILD_VECTOR sequence which other optimizations can
9850 // turn into a single shuffle instruction.
9854 bool isLE = TLI.isLittleEndian();
9855 unsigned ElemRatio = OutScalarTy.getSizeInBits()/SourceType.getSizeInBits();
9856 assert(ElemRatio > 1 && "Invalid element size ratio");
9857 SDValue Filler = AllAnyExt ? DAG.getUNDEF(SourceType):
9858 DAG.getConstant(0, SourceType);
9860 unsigned NewBVElems = ElemRatio * VT.getVectorNumElements();
9861 SmallVector<SDValue, 8> Ops(NewBVElems, Filler);
9863 // Populate the new build_vector
9864 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
9865 SDValue Cast = N->getOperand(i);
9866 assert((Cast.getOpcode() == ISD::ANY_EXTEND ||
9867 Cast.getOpcode() == ISD::ZERO_EXTEND ||
9868 Cast.getOpcode() == ISD::UNDEF) && "Invalid cast opcode");
9870 if (Cast.getOpcode() == ISD::UNDEF)
9871 In = DAG.getUNDEF(SourceType);
9873 In = Cast->getOperand(0);
9874 unsigned Index = isLE ? (i * ElemRatio) :
9875 (i * ElemRatio + (ElemRatio - 1));
9877 assert(Index < Ops.size() && "Invalid index");
9881 // The type of the new BUILD_VECTOR node.
9882 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), SourceType, NewBVElems);
9883 assert(VecVT.getSizeInBits() == VT.getSizeInBits() &&
9884 "Invalid vector size");
9885 // Check if the new vector type is legal.
9886 if (!isTypeLegal(VecVT)) return SDValue();
9888 // Make the new BUILD_VECTOR.
9889 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], Ops.size());
9891 // The new BUILD_VECTOR node has the potential to be further optimized.
9892 AddToWorkList(BV.getNode());
9893 // Bitcast to the desired type.
9894 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
9897 SDValue DAGCombiner::reduceBuildVecConvertToConvertBuildVec(SDNode *N) {
9898 EVT VT = N->getValueType(0);
9900 unsigned NumInScalars = N->getNumOperands();
9903 EVT SrcVT = MVT::Other;
9904 unsigned Opcode = ISD::DELETED_NODE;
9905 unsigned NumDefs = 0;
9907 for (unsigned i = 0; i != NumInScalars; ++i) {
9908 SDValue In = N->getOperand(i);
9909 unsigned Opc = In.getOpcode();
9911 if (Opc == ISD::UNDEF)
9914 // If all scalar values are floats and converted from integers.
9915 if (Opcode == ISD::DELETED_NODE &&
9916 (Opc == ISD::UINT_TO_FP || Opc == ISD::SINT_TO_FP)) {
9923 EVT InVT = In.getOperand(0).getValueType();
9925 // If all scalar values are typed differently, bail out. It's chosen to
9926 // simplify BUILD_VECTOR of integer types.
9927 if (SrcVT == MVT::Other)
9934 // If the vector has just one element defined, it's not worth to fold it into
9935 // a vectorized one.
9939 assert((Opcode == ISD::UINT_TO_FP || Opcode == ISD::SINT_TO_FP)
9940 && "Should only handle conversion from integer to float.");
9941 assert(SrcVT != MVT::Other && "Cannot determine source type!");
9943 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SrcVT, NumInScalars);
9945 if (!TLI.isOperationLegalOrCustom(Opcode, NVT))
9948 SmallVector<SDValue, 8> Opnds;
9949 for (unsigned i = 0; i != NumInScalars; ++i) {
9950 SDValue In = N->getOperand(i);
9952 if (In.getOpcode() == ISD::UNDEF)
9953 Opnds.push_back(DAG.getUNDEF(SrcVT));
9955 Opnds.push_back(In.getOperand(0));
9957 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NVT,
9958 &Opnds[0], Opnds.size());
9959 AddToWorkList(BV.getNode());
9961 return DAG.getNode(Opcode, dl, VT, BV);
9964 SDValue DAGCombiner::visitBUILD_VECTOR(SDNode *N) {
9965 unsigned NumInScalars = N->getNumOperands();
9967 EVT VT = N->getValueType(0);
9969 // A vector built entirely of undefs is undef.
9970 if (ISD::allOperandsUndef(N))
9971 return DAG.getUNDEF(VT);
9973 SDValue V = reduceBuildVecExtToExtBuildVec(N);
9977 V = reduceBuildVecConvertToConvertBuildVec(N);
9981 // Check to see if this is a BUILD_VECTOR of a bunch of EXTRACT_VECTOR_ELT
9982 // operations. If so, and if the EXTRACT_VECTOR_ELT vector inputs come from
9983 // at most two distinct vectors, turn this into a shuffle node.
9985 // May only combine to shuffle after legalize if shuffle is legal.
9986 if (LegalOperations &&
9987 !TLI.isOperationLegalOrCustom(ISD::VECTOR_SHUFFLE, VT))
9990 SDValue VecIn1, VecIn2;
9991 for (unsigned i = 0; i != NumInScalars; ++i) {
9992 // Ignore undef inputs.
9993 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
9995 // If this input is something other than a EXTRACT_VECTOR_ELT with a
9996 // constant index, bail out.
9997 if (N->getOperand(i).getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
9998 !isa<ConstantSDNode>(N->getOperand(i).getOperand(1))) {
9999 VecIn1 = VecIn2 = SDValue(0, 0);
10003 // We allow up to two distinct input vectors.
10004 SDValue ExtractedFromVec = N->getOperand(i).getOperand(0);
10005 if (ExtractedFromVec == VecIn1 || ExtractedFromVec == VecIn2)
10008 if (VecIn1.getNode() == 0) {
10009 VecIn1 = ExtractedFromVec;
10010 } else if (VecIn2.getNode() == 0) {
10011 VecIn2 = ExtractedFromVec;
10013 // Too many inputs.
10014 VecIn1 = VecIn2 = SDValue(0, 0);
10019 // If everything is good, we can make a shuffle operation.
10020 if (VecIn1.getNode()) {
10021 SmallVector<int, 8> Mask;
10022 for (unsigned i = 0; i != NumInScalars; ++i) {
10023 if (N->getOperand(i).getOpcode() == ISD::UNDEF) {
10024 Mask.push_back(-1);
10028 // If extracting from the first vector, just use the index directly.
10029 SDValue Extract = N->getOperand(i);
10030 SDValue ExtVal = Extract.getOperand(1);
10031 if (Extract.getOperand(0) == VecIn1) {
10032 unsigned ExtIndex = cast<ConstantSDNode>(ExtVal)->getZExtValue();
10033 if (ExtIndex > VT.getVectorNumElements())
10036 Mask.push_back(ExtIndex);
10040 // Otherwise, use InIdx + VecSize
10041 unsigned Idx = cast<ConstantSDNode>(ExtVal)->getZExtValue();
10042 Mask.push_back(Idx+NumInScalars);
10045 // We can't generate a shuffle node with mismatched input and output types.
10046 // Attempt to transform a single input vector to the correct type.
10047 if ((VT != VecIn1.getValueType())) {
10048 // We don't support shuffeling between TWO values of different types.
10049 if (VecIn2.getNode() != 0)
10052 // We only support widening of vectors which are half the size of the
10053 // output registers. For example XMM->YMM widening on X86 with AVX.
10054 if (VecIn1.getValueType().getSizeInBits()*2 != VT.getSizeInBits())
10057 // If the input vector type has a different base type to the output
10058 // vector type, bail out.
10059 if (VecIn1.getValueType().getVectorElementType() !=
10060 VT.getVectorElementType())
10063 // Widen the input vector by adding undef values.
10064 VecIn1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
10065 VecIn1, DAG.getUNDEF(VecIn1.getValueType()));
10068 // If VecIn2 is unused then change it to undef.
10069 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
10071 // Check that we were able to transform all incoming values to the same
10073 if (VecIn2.getValueType() != VecIn1.getValueType() ||
10074 VecIn1.getValueType() != VT)
10077 // Only type-legal BUILD_VECTOR nodes are converted to shuffle nodes.
10078 if (!isTypeLegal(VT))
10081 // Return the new VECTOR_SHUFFLE node.
10085 return DAG.getVectorShuffle(VT, dl, Ops[0], Ops[1], &Mask[0]);
10091 SDValue DAGCombiner::visitCONCAT_VECTORS(SDNode *N) {
10092 // TODO: Check to see if this is a CONCAT_VECTORS of a bunch of
10093 // EXTRACT_SUBVECTOR operations. If so, and if the EXTRACT_SUBVECTOR vector
10094 // inputs come from at most two distinct vectors, turn this into a shuffle
10097 // If we only have one input vector, we don't need to do any concatenation.
10098 if (N->getNumOperands() == 1)
10099 return N->getOperand(0);
10101 // Check if all of the operands are undefs.
10102 EVT VT = N->getValueType(0);
10103 if (ISD::allOperandsUndef(N))
10104 return DAG.getUNDEF(VT);
10106 // Optimize concat_vectors where one of the vectors is undef.
10107 if (N->getNumOperands() == 2 &&
10108 N->getOperand(1)->getOpcode() == ISD::UNDEF) {
10109 SDValue In = N->getOperand(0);
10110 assert(In.getValueType().isVector() && "Must concat vectors");
10112 // Transform: concat_vectors(scalar, undef) -> scalar_to_vector(sclr).
10113 if (In->getOpcode() == ISD::BITCAST &&
10114 !In->getOperand(0)->getValueType(0).isVector()) {
10115 SDValue Scalar = In->getOperand(0);
10116 EVT SclTy = Scalar->getValueType(0);
10118 if (!SclTy.isFloatingPoint() && !SclTy.isInteger())
10121 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SclTy,
10122 VT.getSizeInBits() / SclTy.getSizeInBits());
10123 if (!TLI.isTypeLegal(NVT) || !TLI.isTypeLegal(Scalar.getValueType()))
10126 SDLoc dl = SDLoc(N);
10127 SDValue Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, NVT, Scalar);
10128 return DAG.getNode(ISD::BITCAST, dl, VT, Res);
10132 // Type legalization of vectors and DAG canonicalization of SHUFFLE_VECTOR
10133 // nodes often generate nop CONCAT_VECTOR nodes.
10134 // Scan the CONCAT_VECTOR operands and look for a CONCAT operations that
10135 // place the incoming vectors at the exact same location.
10136 SDValue SingleSource = SDValue();
10137 unsigned PartNumElem = N->getOperand(0).getValueType().getVectorNumElements();
10139 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
10140 SDValue Op = N->getOperand(i);
10142 if (Op.getOpcode() == ISD::UNDEF)
10145 // Check if this is the identity extract:
10146 if (Op.getOpcode() != ISD::EXTRACT_SUBVECTOR)
10149 // Find the single incoming vector for the extract_subvector.
10150 if (SingleSource.getNode()) {
10151 if (Op.getOperand(0) != SingleSource)
10154 SingleSource = Op.getOperand(0);
10156 // Check the source type is the same as the type of the result.
10157 // If not, this concat may extend the vector, so we can not
10158 // optimize it away.
10159 if (SingleSource.getValueType() != N->getValueType(0))
10163 unsigned IdentityIndex = i * PartNumElem;
10164 ConstantSDNode *CS = dyn_cast<ConstantSDNode>(Op.getOperand(1));
10165 // The extract index must be constant.
10169 // Check that we are reading from the identity index.
10170 if (CS->getZExtValue() != IdentityIndex)
10174 if (SingleSource.getNode())
10175 return SingleSource;
10180 SDValue DAGCombiner::visitEXTRACT_SUBVECTOR(SDNode* N) {
10181 EVT NVT = N->getValueType(0);
10182 SDValue V = N->getOperand(0);
10184 if (V->getOpcode() == ISD::CONCAT_VECTORS) {
10186 // (extract_subvec (concat V1, V2, ...), i)
10189 // Only operand 0 is checked as 'concat' assumes all inputs of the same
10191 if (V->getOperand(0).getValueType() != NVT)
10193 unsigned Idx = dyn_cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
10194 unsigned NumElems = NVT.getVectorNumElements();
10195 assert((Idx % NumElems) == 0 &&
10196 "IDX in concat is not a multiple of the result vector length.");
10197 return V->getOperand(Idx / NumElems);
10201 if (V->getOpcode() == ISD::BITCAST)
10202 V = V.getOperand(0);
10204 if (V->getOpcode() == ISD::INSERT_SUBVECTOR) {
10206 // Handle only simple case where vector being inserted and vector
10207 // being extracted are of same type, and are half size of larger vectors.
10208 EVT BigVT = V->getOperand(0).getValueType();
10209 EVT SmallVT = V->getOperand(1).getValueType();
10210 if (!NVT.bitsEq(SmallVT) || NVT.getSizeInBits()*2 != BigVT.getSizeInBits())
10213 // Only handle cases where both indexes are constants with the same type.
10214 ConstantSDNode *ExtIdx = dyn_cast<ConstantSDNode>(N->getOperand(1));
10215 ConstantSDNode *InsIdx = dyn_cast<ConstantSDNode>(V->getOperand(2));
10217 if (InsIdx && ExtIdx &&
10218 InsIdx->getValueType(0).getSizeInBits() <= 64 &&
10219 ExtIdx->getValueType(0).getSizeInBits() <= 64) {
10221 // (extract_subvec (insert_subvec V1, V2, InsIdx), ExtIdx)
10223 // indices are equal or bit offsets are equal => V1
10224 // otherwise => (extract_subvec V1, ExtIdx)
10225 if (InsIdx->getZExtValue() * SmallVT.getScalarType().getSizeInBits() ==
10226 ExtIdx->getZExtValue() * NVT.getScalarType().getSizeInBits())
10227 return DAG.getNode(ISD::BITCAST, dl, NVT, V->getOperand(1));
10228 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, NVT,
10229 DAG.getNode(ISD::BITCAST, dl,
10230 N->getOperand(0).getValueType(),
10231 V->getOperand(0)), N->getOperand(1));
10238 // Tries to turn a shuffle of two CONCAT_VECTORS into a single concat.
10239 static SDValue partitionShuffleOfConcats(SDNode *N, SelectionDAG &DAG) {
10240 EVT VT = N->getValueType(0);
10241 unsigned NumElts = VT.getVectorNumElements();
10243 SDValue N0 = N->getOperand(0);
10244 SDValue N1 = N->getOperand(1);
10245 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
10247 SmallVector<SDValue, 4> Ops;
10248 EVT ConcatVT = N0.getOperand(0).getValueType();
10249 unsigned NumElemsPerConcat = ConcatVT.getVectorNumElements();
10250 unsigned NumConcats = NumElts / NumElemsPerConcat;
10252 // Look at every vector that's inserted. We're looking for exact
10253 // subvector-sized copies from a concatenated vector
10254 for (unsigned I = 0; I != NumConcats; ++I) {
10255 // Make sure we're dealing with a copy.
10256 unsigned Begin = I * NumElemsPerConcat;
10257 bool AllUndef = true, NoUndef = true;
10258 for (unsigned J = Begin; J != Begin + NumElemsPerConcat; ++J) {
10259 if (SVN->getMaskElt(J) >= 0)
10266 if (SVN->getMaskElt(Begin) % NumElemsPerConcat != 0)
10269 for (unsigned J = 1; J != NumElemsPerConcat; ++J)
10270 if (SVN->getMaskElt(Begin + J - 1) + 1 != SVN->getMaskElt(Begin + J))
10273 unsigned FirstElt = SVN->getMaskElt(Begin) / NumElemsPerConcat;
10274 if (FirstElt < N0.getNumOperands())
10275 Ops.push_back(N0.getOperand(FirstElt));
10277 Ops.push_back(N1.getOperand(FirstElt - N0.getNumOperands()));
10279 } else if (AllUndef) {
10280 Ops.push_back(DAG.getUNDEF(N0.getOperand(0).getValueType()));
10281 } else { // Mixed with general masks and undefs, can't do optimization.
10286 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Ops.data(),
10290 SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
10291 EVT VT = N->getValueType(0);
10292 unsigned NumElts = VT.getVectorNumElements();
10294 SDValue N0 = N->getOperand(0);
10295 SDValue N1 = N->getOperand(1);
10297 assert(N0.getValueType() == VT && "Vector shuffle must be normalized in DAG");
10299 // Canonicalize shuffle undef, undef -> undef
10300 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
10301 return DAG.getUNDEF(VT);
10303 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
10305 // Canonicalize shuffle v, v -> v, undef
10307 SmallVector<int, 8> NewMask;
10308 for (unsigned i = 0; i != NumElts; ++i) {
10309 int Idx = SVN->getMaskElt(i);
10310 if (Idx >= (int)NumElts) Idx -= NumElts;
10311 NewMask.push_back(Idx);
10313 return DAG.getVectorShuffle(VT, SDLoc(N), N0, DAG.getUNDEF(VT),
10317 // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask.
10318 if (N0.getOpcode() == ISD::UNDEF) {
10319 SmallVector<int, 8> NewMask;
10320 for (unsigned i = 0; i != NumElts; ++i) {
10321 int Idx = SVN->getMaskElt(i);
10323 if (Idx >= (int)NumElts)
10326 Idx = -1; // remove reference to lhs
10328 NewMask.push_back(Idx);
10330 return DAG.getVectorShuffle(VT, SDLoc(N), N1, DAG.getUNDEF(VT),
10334 // Remove references to rhs if it is undef
10335 if (N1.getOpcode() == ISD::UNDEF) {
10336 bool Changed = false;
10337 SmallVector<int, 8> NewMask;
10338 for (unsigned i = 0; i != NumElts; ++i) {
10339 int Idx = SVN->getMaskElt(i);
10340 if (Idx >= (int)NumElts) {
10344 NewMask.push_back(Idx);
10347 return DAG.getVectorShuffle(VT, SDLoc(N), N0, N1, &NewMask[0]);
10350 // If it is a splat, check if the argument vector is another splat or a
10351 // build_vector with all scalar elements the same.
10352 if (SVN->isSplat() && SVN->getSplatIndex() < (int)NumElts) {
10353 SDNode *V = N0.getNode();
10355 // If this is a bit convert that changes the element type of the vector but
10356 // not the number of vector elements, look through it. Be careful not to
10357 // look though conversions that change things like v4f32 to v2f64.
10358 if (V->getOpcode() == ISD::BITCAST) {
10359 SDValue ConvInput = V->getOperand(0);
10360 if (ConvInput.getValueType().isVector() &&
10361 ConvInput.getValueType().getVectorNumElements() == NumElts)
10362 V = ConvInput.getNode();
10365 if (V->getOpcode() == ISD::BUILD_VECTOR) {
10366 assert(V->getNumOperands() == NumElts &&
10367 "BUILD_VECTOR has wrong number of operands");
10369 bool AllSame = true;
10370 for (unsigned i = 0; i != NumElts; ++i) {
10371 if (V->getOperand(i).getOpcode() != ISD::UNDEF) {
10372 Base = V->getOperand(i);
10376 // Splat of <u, u, u, u>, return <u, u, u, u>
10377 if (!Base.getNode())
10379 for (unsigned i = 0; i != NumElts; ++i) {
10380 if (V->getOperand(i) != Base) {
10385 // Splat of <x, x, x, x>, return <x, x, x, x>
10391 if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
10392 Level < AfterLegalizeVectorOps &&
10393 (N1.getOpcode() == ISD::UNDEF ||
10394 (N1.getOpcode() == ISD::CONCAT_VECTORS &&
10395 N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()))) {
10396 SDValue V = partitionShuffleOfConcats(N, DAG);
10402 // If this shuffle node is simply a swizzle of another shuffle node,
10403 // and it reverses the swizzle of the previous shuffle then we can
10404 // optimize shuffle(shuffle(x, undef), undef) -> x.
10405 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
10406 N1.getOpcode() == ISD::UNDEF) {
10408 ShuffleVectorSDNode *OtherSV = cast<ShuffleVectorSDNode>(N0);
10410 // Shuffle nodes can only reverse shuffles with a single non-undef value.
10411 if (N0.getOperand(1).getOpcode() != ISD::UNDEF)
10414 // The incoming shuffle must be of the same type as the result of the
10415 // current shuffle.
10416 assert(OtherSV->getOperand(0).getValueType() == VT &&
10417 "Shuffle types don't match");
10419 for (unsigned i = 0; i != NumElts; ++i) {
10420 int Idx = SVN->getMaskElt(i);
10421 assert(Idx < (int)NumElts && "Index references undef operand");
10422 // Next, this index comes from the first value, which is the incoming
10423 // shuffle. Adopt the incoming index.
10425 Idx = OtherSV->getMaskElt(Idx);
10427 // The combined shuffle must map each index to itself.
10428 if (Idx >= 0 && (unsigned)Idx != i)
10432 return OtherSV->getOperand(0);
10438 SDValue DAGCombiner::visitINSERT_SUBVECTOR(SDNode *N) {
10439 SDValue N0 = N->getOperand(0);
10440 SDValue N2 = N->getOperand(2);
10442 // If the input vector is a concatenation, and the insert replaces
10443 // one of the halves, we can optimize into a single concat_vectors.
10444 if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
10445 N0->getNumOperands() == 2 && N2.getOpcode() == ISD::Constant) {
10446 APInt InsIdx = cast<ConstantSDNode>(N2)->getAPIntValue();
10447 EVT VT = N->getValueType(0);
10449 // Lower half: fold (insert_subvector (concat_vectors X, Y), Z) ->
10450 // (concat_vectors Z, Y)
10452 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
10453 N->getOperand(1), N0.getOperand(1));
10455 // Upper half: fold (insert_subvector (concat_vectors X, Y), Z) ->
10456 // (concat_vectors X, Z)
10457 if (InsIdx == VT.getVectorNumElements()/2)
10458 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
10459 N0.getOperand(0), N->getOperand(1));
10465 /// XformToShuffleWithZero - Returns a vector_shuffle if it able to transform
10466 /// an AND to a vector_shuffle with the destination vector and a zero vector.
10467 /// e.g. AND V, <0xffffffff, 0, 0xffffffff, 0>. ==>
10468 /// vector_shuffle V, Zero, <0, 4, 2, 4>
10469 SDValue DAGCombiner::XformToShuffleWithZero(SDNode *N) {
10470 EVT VT = N->getValueType(0);
10472 SDValue LHS = N->getOperand(0);
10473 SDValue RHS = N->getOperand(1);
10474 if (N->getOpcode() == ISD::AND) {
10475 if (RHS.getOpcode() == ISD::BITCAST)
10476 RHS = RHS.getOperand(0);
10477 if (RHS.getOpcode() == ISD::BUILD_VECTOR) {
10478 SmallVector<int, 8> Indices;
10479 unsigned NumElts = RHS.getNumOperands();
10480 for (unsigned i = 0; i != NumElts; ++i) {
10481 SDValue Elt = RHS.getOperand(i);
10482 if (!isa<ConstantSDNode>(Elt))
10485 if (cast<ConstantSDNode>(Elt)->isAllOnesValue())
10486 Indices.push_back(i);
10487 else if (cast<ConstantSDNode>(Elt)->isNullValue())
10488 Indices.push_back(NumElts);
10493 // Let's see if the target supports this vector_shuffle.
10494 EVT RVT = RHS.getValueType();
10495 if (!TLI.isVectorClearMaskLegal(Indices, RVT))
10498 // Return the new VECTOR_SHUFFLE node.
10499 EVT EltVT = RVT.getVectorElementType();
10500 SmallVector<SDValue,8> ZeroOps(RVT.getVectorNumElements(),
10501 DAG.getConstant(0, EltVT));
10502 SDValue Zero = DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N),
10503 RVT, &ZeroOps[0], ZeroOps.size());
10504 LHS = DAG.getNode(ISD::BITCAST, dl, RVT, LHS);
10505 SDValue Shuf = DAG.getVectorShuffle(RVT, dl, LHS, Zero, &Indices[0]);
10506 return DAG.getNode(ISD::BITCAST, dl, VT, Shuf);
10513 /// SimplifyVBinOp - Visit a binary vector operation, like ADD.
10514 SDValue DAGCombiner::SimplifyVBinOp(SDNode *N) {
10515 assert(N->getValueType(0).isVector() &&
10516 "SimplifyVBinOp only works on vectors!");
10518 SDValue LHS = N->getOperand(0);
10519 SDValue RHS = N->getOperand(1);
10520 SDValue Shuffle = XformToShuffleWithZero(N);
10521 if (Shuffle.getNode()) return Shuffle;
10523 // If the LHS and RHS are BUILD_VECTOR nodes, see if we can constant fold
10525 if (LHS.getOpcode() == ISD::BUILD_VECTOR &&
10526 RHS.getOpcode() == ISD::BUILD_VECTOR) {
10527 // Check if both vectors are constants. If not bail out.
10528 if (!(cast<BuildVectorSDNode>(LHS)->isConstant() &&
10529 cast<BuildVectorSDNode>(RHS)->isConstant()))
10532 SmallVector<SDValue, 8> Ops;
10533 for (unsigned i = 0, e = LHS.getNumOperands(); i != e; ++i) {
10534 SDValue LHSOp = LHS.getOperand(i);
10535 SDValue RHSOp = RHS.getOperand(i);
10537 // Can't fold divide by zero.
10538 if (N->getOpcode() == ISD::SDIV || N->getOpcode() == ISD::UDIV ||
10539 N->getOpcode() == ISD::FDIV) {
10540 if ((RHSOp.getOpcode() == ISD::Constant &&
10541 cast<ConstantSDNode>(RHSOp.getNode())->isNullValue()) ||
10542 (RHSOp.getOpcode() == ISD::ConstantFP &&
10543 cast<ConstantFPSDNode>(RHSOp.getNode())->getValueAPF().isZero()))
10547 EVT VT = LHSOp.getValueType();
10548 EVT RVT = RHSOp.getValueType();
10550 // Integer BUILD_VECTOR operands may have types larger than the element
10551 // size (e.g., when the element type is not legal). Prior to type
10552 // legalization, the types may not match between the two BUILD_VECTORS.
10553 // Truncate one of the operands to make them match.
10554 if (RVT.getSizeInBits() > VT.getSizeInBits()) {
10555 RHSOp = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, RHSOp);
10557 LHSOp = DAG.getNode(ISD::TRUNCATE, SDLoc(N), RVT, LHSOp);
10561 SDValue FoldOp = DAG.getNode(N->getOpcode(), SDLoc(LHS), VT,
10563 if (FoldOp.getOpcode() != ISD::UNDEF &&
10564 FoldOp.getOpcode() != ISD::Constant &&
10565 FoldOp.getOpcode() != ISD::ConstantFP)
10567 Ops.push_back(FoldOp);
10568 AddToWorkList(FoldOp.getNode());
10571 if (Ops.size() == LHS.getNumOperands())
10572 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N),
10573 LHS.getValueType(), &Ops[0], Ops.size());
10579 /// SimplifyVUnaryOp - Visit a binary vector operation, like FABS/FNEG.
10580 SDValue DAGCombiner::SimplifyVUnaryOp(SDNode *N) {
10581 assert(N->getValueType(0).isVector() &&
10582 "SimplifyVUnaryOp only works on vectors!");
10584 SDValue N0 = N->getOperand(0);
10586 if (N0.getOpcode() != ISD::BUILD_VECTOR)
10589 // Operand is a BUILD_VECTOR node, see if we can constant fold it.
10590 SmallVector<SDValue, 8> Ops;
10591 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) {
10592 SDValue Op = N0.getOperand(i);
10593 if (Op.getOpcode() != ISD::UNDEF &&
10594 Op.getOpcode() != ISD::ConstantFP)
10596 EVT EltVT = Op.getValueType();
10597 SDValue FoldOp = DAG.getNode(N->getOpcode(), SDLoc(N0), EltVT, Op);
10598 if (FoldOp.getOpcode() != ISD::UNDEF &&
10599 FoldOp.getOpcode() != ISD::ConstantFP)
10601 Ops.push_back(FoldOp);
10602 AddToWorkList(FoldOp.getNode());
10605 if (Ops.size() != N0.getNumOperands())
10608 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N),
10609 N0.getValueType(), &Ops[0], Ops.size());
10612 SDValue DAGCombiner::SimplifySelect(SDLoc DL, SDValue N0,
10613 SDValue N1, SDValue N2){
10614 assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!");
10616 SDValue SCC = SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1), N1, N2,
10617 cast<CondCodeSDNode>(N0.getOperand(2))->get());
10619 // If we got a simplified select_cc node back from SimplifySelectCC, then
10620 // break it down into a new SETCC node, and a new SELECT node, and then return
10621 // the SELECT node, since we were called with a SELECT node.
10622 if (SCC.getNode()) {
10623 // Check to see if we got a select_cc back (to turn into setcc/select).
10624 // Otherwise, just return whatever node we got back, like fabs.
10625 if (SCC.getOpcode() == ISD::SELECT_CC) {
10626 SDValue SETCC = DAG.getNode(ISD::SETCC, SDLoc(N0),
10628 SCC.getOperand(0), SCC.getOperand(1),
10629 SCC.getOperand(4));
10630 AddToWorkList(SETCC.getNode());
10631 return DAG.getSelect(SDLoc(SCC), SCC.getValueType(),
10632 SCC.getOperand(2), SCC.getOperand(3), SETCC);
10640 /// SimplifySelectOps - Given a SELECT or a SELECT_CC node, where LHS and RHS
10641 /// are the two values being selected between, see if we can simplify the
10642 /// select. Callers of this should assume that TheSelect is deleted if this
10643 /// returns true. As such, they should return the appropriate thing (e.g. the
10644 /// node) back to the top-level of the DAG combiner loop to avoid it being
10646 bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDValue LHS,
10649 // Cannot simplify select with vector condition
10650 if (TheSelect->getOperand(0).getValueType().isVector()) return false;
10652 // If this is a select from two identical things, try to pull the operation
10653 // through the select.
10654 if (LHS.getOpcode() != RHS.getOpcode() ||
10655 !LHS.hasOneUse() || !RHS.hasOneUse())
10658 // If this is a load and the token chain is identical, replace the select
10659 // of two loads with a load through a select of the address to load from.
10660 // This triggers in things like "select bool X, 10.0, 123.0" after the FP
10661 // constants have been dropped into the constant pool.
10662 if (LHS.getOpcode() == ISD::LOAD) {
10663 LoadSDNode *LLD = cast<LoadSDNode>(LHS);
10664 LoadSDNode *RLD = cast<LoadSDNode>(RHS);
10666 // Token chains must be identical.
10667 if (LHS.getOperand(0) != RHS.getOperand(0) ||
10668 // Do not let this transformation reduce the number of volatile loads.
10669 LLD->isVolatile() || RLD->isVolatile() ||
10670 // If this is an EXTLOAD, the VT's must match.
10671 LLD->getMemoryVT() != RLD->getMemoryVT() ||
10672 // If this is an EXTLOAD, the kind of extension must match.
10673 (LLD->getExtensionType() != RLD->getExtensionType() &&
10674 // The only exception is if one of the extensions is anyext.
10675 LLD->getExtensionType() != ISD::EXTLOAD &&
10676 RLD->getExtensionType() != ISD::EXTLOAD) ||
10677 // FIXME: this discards src value information. This is
10678 // over-conservative. It would be beneficial to be able to remember
10679 // both potential memory locations. Since we are discarding
10680 // src value info, don't do the transformation if the memory
10681 // locations are not in the default address space.
10682 LLD->getPointerInfo().getAddrSpace() != 0 ||
10683 RLD->getPointerInfo().getAddrSpace() != 0 ||
10684 !TLI.isOperationLegalOrCustom(TheSelect->getOpcode(),
10685 LLD->getBasePtr().getValueType()))
10688 // Check that the select condition doesn't reach either load. If so,
10689 // folding this will induce a cycle into the DAG. If not, this is safe to
10690 // xform, so create a select of the addresses.
10692 if (TheSelect->getOpcode() == ISD::SELECT) {
10693 SDNode *CondNode = TheSelect->getOperand(0).getNode();
10694 if ((LLD->hasAnyUseOfValue(1) && LLD->isPredecessorOf(CondNode)) ||
10695 (RLD->hasAnyUseOfValue(1) && RLD->isPredecessorOf(CondNode)))
10697 // The loads must not depend on one another.
10698 if (LLD->isPredecessorOf(RLD) ||
10699 RLD->isPredecessorOf(LLD))
10701 Addr = DAG.getSelect(SDLoc(TheSelect),
10702 LLD->getBasePtr().getValueType(),
10703 TheSelect->getOperand(0), LLD->getBasePtr(),
10704 RLD->getBasePtr());
10705 } else { // Otherwise SELECT_CC
10706 SDNode *CondLHS = TheSelect->getOperand(0).getNode();
10707 SDNode *CondRHS = TheSelect->getOperand(1).getNode();
10709 if ((LLD->hasAnyUseOfValue(1) &&
10710 (LLD->isPredecessorOf(CondLHS) || LLD->isPredecessorOf(CondRHS))) ||
10711 (RLD->hasAnyUseOfValue(1) &&
10712 (RLD->isPredecessorOf(CondLHS) || RLD->isPredecessorOf(CondRHS))))
10715 Addr = DAG.getNode(ISD::SELECT_CC, SDLoc(TheSelect),
10716 LLD->getBasePtr().getValueType(),
10717 TheSelect->getOperand(0),
10718 TheSelect->getOperand(1),
10719 LLD->getBasePtr(), RLD->getBasePtr(),
10720 TheSelect->getOperand(4));
10724 if (LLD->getExtensionType() == ISD::NON_EXTLOAD) {
10725 Load = DAG.getLoad(TheSelect->getValueType(0),
10727 // FIXME: Discards pointer and TBAA info.
10728 LLD->getChain(), Addr, MachinePointerInfo(),
10729 LLD->isVolatile(), LLD->isNonTemporal(),
10730 LLD->isInvariant(), LLD->getAlignment());
10732 Load = DAG.getExtLoad(LLD->getExtensionType() == ISD::EXTLOAD ?
10733 RLD->getExtensionType() : LLD->getExtensionType(),
10735 TheSelect->getValueType(0),
10736 // FIXME: Discards pointer and TBAA info.
10737 LLD->getChain(), Addr, MachinePointerInfo(),
10738 LLD->getMemoryVT(), LLD->isVolatile(),
10739 LLD->isNonTemporal(), LLD->getAlignment());
10742 // Users of the select now use the result of the load.
10743 CombineTo(TheSelect, Load);
10745 // Users of the old loads now use the new load's chain. We know the
10746 // old-load value is dead now.
10747 CombineTo(LHS.getNode(), Load.getValue(0), Load.getValue(1));
10748 CombineTo(RHS.getNode(), Load.getValue(0), Load.getValue(1));
10755 /// SimplifySelectCC - Simplify an expression of the form (N0 cond N1) ? N2 : N3
10756 /// where 'cond' is the comparison specified by CC.
10757 SDValue DAGCombiner::SimplifySelectCC(SDLoc DL, SDValue N0, SDValue N1,
10758 SDValue N2, SDValue N3,
10759 ISD::CondCode CC, bool NotExtCompare) {
10760 // (x ? y : y) -> y.
10761 if (N2 == N3) return N2;
10763 EVT VT = N2.getValueType();
10764 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
10765 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
10766 ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N3.getNode());
10768 // Determine if the condition we're dealing with is constant
10769 SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()),
10770 N0, N1, CC, DL, false);
10771 if (SCC.getNode()) AddToWorkList(SCC.getNode());
10772 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode());
10774 // fold select_cc true, x, y -> x
10775 if (SCCC && !SCCC->isNullValue())
10777 // fold select_cc false, x, y -> y
10778 if (SCCC && SCCC->isNullValue())
10781 // Check to see if we can simplify the select into an fabs node
10782 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) {
10783 // Allow either -0.0 or 0.0
10784 if (CFP->getValueAPF().isZero()) {
10785 // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs
10786 if ((CC == ISD::SETGE || CC == ISD::SETGT) &&
10787 N0 == N2 && N3.getOpcode() == ISD::FNEG &&
10788 N2 == N3.getOperand(0))
10789 return DAG.getNode(ISD::FABS, DL, VT, N0);
10791 // select (setl[te] X, +/-0.0), fneg(X), X -> fabs
10792 if ((CC == ISD::SETLT || CC == ISD::SETLE) &&
10793 N0 == N3 && N2.getOpcode() == ISD::FNEG &&
10794 N2.getOperand(0) == N3)
10795 return DAG.getNode(ISD::FABS, DL, VT, N3);
10799 // Turn "(a cond b) ? 1.0f : 2.0f" into "load (tmp + ((a cond b) ? 0 : 4)"
10800 // where "tmp" is a constant pool entry containing an array with 1.0 and 2.0
10801 // in it. This is a win when the constant is not otherwise available because
10802 // it replaces two constant pool loads with one. We only do this if the FP
10803 // type is known to be legal, because if it isn't, then we are before legalize
10804 // types an we want the other legalization to happen first (e.g. to avoid
10805 // messing with soft float) and if the ConstantFP is not legal, because if
10806 // it is legal, we may not need to store the FP constant in a constant pool.
10807 if (ConstantFPSDNode *TV = dyn_cast<ConstantFPSDNode>(N2))
10808 if (ConstantFPSDNode *FV = dyn_cast<ConstantFPSDNode>(N3)) {
10809 if (TLI.isTypeLegal(N2.getValueType()) &&
10810 (TLI.getOperationAction(ISD::ConstantFP, N2.getValueType()) !=
10811 TargetLowering::Legal) &&
10812 // If both constants have multiple uses, then we won't need to do an
10813 // extra load, they are likely around in registers for other users.
10814 (TV->hasOneUse() || FV->hasOneUse())) {
10815 Constant *Elts[] = {
10816 const_cast<ConstantFP*>(FV->getConstantFPValue()),
10817 const_cast<ConstantFP*>(TV->getConstantFPValue())
10819 Type *FPTy = Elts[0]->getType();
10820 const DataLayout &TD = *TLI.getDataLayout();
10822 // Create a ConstantArray of the two constants.
10823 Constant *CA = ConstantArray::get(ArrayType::get(FPTy, 2), Elts);
10824 SDValue CPIdx = DAG.getConstantPool(CA, TLI.getPointerTy(),
10825 TD.getPrefTypeAlignment(FPTy));
10826 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
10828 // Get the offsets to the 0 and 1 element of the array so that we can
10829 // select between them.
10830 SDValue Zero = DAG.getIntPtrConstant(0);
10831 unsigned EltSize = (unsigned)TD.getTypeAllocSize(Elts[0]->getType());
10832 SDValue One = DAG.getIntPtrConstant(EltSize);
10834 SDValue Cond = DAG.getSetCC(DL,
10835 getSetCCResultType(N0.getValueType()),
10837 AddToWorkList(Cond.getNode());
10838 SDValue CstOffset = DAG.getSelect(DL, Zero.getValueType(),
10840 AddToWorkList(CstOffset.getNode());
10841 CPIdx = DAG.getNode(ISD::ADD, DL, CPIdx.getValueType(), CPIdx,
10843 AddToWorkList(CPIdx.getNode());
10844 return DAG.getLoad(TV->getValueType(0), DL, DAG.getEntryNode(), CPIdx,
10845 MachinePointerInfo::getConstantPool(), false,
10846 false, false, Alignment);
10851 // Check to see if we can perform the "gzip trick", transforming
10852 // (select_cc setlt X, 0, A, 0) -> (and (sra X, (sub size(X), 1), A)
10853 if (N1C && N3C && N3C->isNullValue() && CC == ISD::SETLT &&
10854 (N1C->isNullValue() || // (a < 0) ? b : 0
10855 (N1C->getAPIntValue() == 1 && N0 == N2))) { // (a < 1) ? a : 0
10856 EVT XType = N0.getValueType();
10857 EVT AType = N2.getValueType();
10858 if (XType.bitsGE(AType)) {
10859 // and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a
10860 // single-bit constant.
10861 if (N2C && ((N2C->getAPIntValue() & (N2C->getAPIntValue()-1)) == 0)) {
10862 unsigned ShCtV = N2C->getAPIntValue().logBase2();
10863 ShCtV = XType.getSizeInBits()-ShCtV-1;
10864 SDValue ShCt = DAG.getConstant(ShCtV,
10865 getShiftAmountTy(N0.getValueType()));
10866 SDValue Shift = DAG.getNode(ISD::SRL, SDLoc(N0),
10868 AddToWorkList(Shift.getNode());
10870 if (XType.bitsGT(AType)) {
10871 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
10872 AddToWorkList(Shift.getNode());
10875 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
10878 SDValue Shift = DAG.getNode(ISD::SRA, SDLoc(N0),
10880 DAG.getConstant(XType.getSizeInBits()-1,
10881 getShiftAmountTy(N0.getValueType())));
10882 AddToWorkList(Shift.getNode());
10884 if (XType.bitsGT(AType)) {
10885 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
10886 AddToWorkList(Shift.getNode());
10889 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
10893 // fold (select_cc seteq (and x, y), 0, 0, A) -> (and (shr (shl x)) A)
10894 // where y is has a single bit set.
10895 // A plaintext description would be, we can turn the SELECT_CC into an AND
10896 // when the condition can be materialized as an all-ones register. Any
10897 // single bit-test can be materialized as an all-ones register with
10898 // shift-left and shift-right-arith.
10899 if (CC == ISD::SETEQ && N0->getOpcode() == ISD::AND &&
10900 N0->getValueType(0) == VT &&
10901 N1C && N1C->isNullValue() &&
10902 N2C && N2C->isNullValue()) {
10903 SDValue AndLHS = N0->getOperand(0);
10904 ConstantSDNode *ConstAndRHS = dyn_cast<ConstantSDNode>(N0->getOperand(1));
10905 if (ConstAndRHS && ConstAndRHS->getAPIntValue().countPopulation() == 1) {
10906 // Shift the tested bit over the sign bit.
10907 APInt AndMask = ConstAndRHS->getAPIntValue();
10909 DAG.getConstant(AndMask.countLeadingZeros(),
10910 getShiftAmountTy(AndLHS.getValueType()));
10911 SDValue Shl = DAG.getNode(ISD::SHL, SDLoc(N0), VT, AndLHS, ShlAmt);
10913 // Now arithmetic right shift it all the way over, so the result is either
10914 // all-ones, or zero.
10916 DAG.getConstant(AndMask.getBitWidth()-1,
10917 getShiftAmountTy(Shl.getValueType()));
10918 SDValue Shr = DAG.getNode(ISD::SRA, SDLoc(N0), VT, Shl, ShrAmt);
10920 return DAG.getNode(ISD::AND, DL, VT, Shr, N3);
10924 // fold select C, 16, 0 -> shl C, 4
10925 if (N2C && N3C && N3C->isNullValue() && N2C->getAPIntValue().isPowerOf2() &&
10926 TLI.getBooleanContents(N0.getValueType().isVector()) ==
10927 TargetLowering::ZeroOrOneBooleanContent) {
10929 // If the caller doesn't want us to simplify this into a zext of a compare,
10931 if (NotExtCompare && N2C->getAPIntValue() == 1)
10934 // Get a SetCC of the condition
10935 // NOTE: Don't create a SETCC if it's not legal on this target.
10936 if (!LegalOperations ||
10937 TLI.isOperationLegal(ISD::SETCC,
10938 LegalTypes ? getSetCCResultType(N0.getValueType()) : MVT::i1)) {
10940 // cast from setcc result type to select result type
10942 SCC = DAG.getSetCC(DL, getSetCCResultType(N0.getValueType()),
10944 if (N2.getValueType().bitsLT(SCC.getValueType()))
10945 Temp = DAG.getZeroExtendInReg(SCC, SDLoc(N2),
10946 N2.getValueType());
10948 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2),
10949 N2.getValueType(), SCC);
10951 SCC = DAG.getSetCC(SDLoc(N0), MVT::i1, N0, N1, CC);
10952 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2),
10953 N2.getValueType(), SCC);
10956 AddToWorkList(SCC.getNode());
10957 AddToWorkList(Temp.getNode());
10959 if (N2C->getAPIntValue() == 1)
10962 // shl setcc result by log2 n2c
10963 return DAG.getNode(
10964 ISD::SHL, DL, N2.getValueType(), Temp,
10965 DAG.getConstant(N2C->getAPIntValue().logBase2(),
10966 getShiftAmountTy(Temp.getValueType())));
10970 // Check to see if this is the equivalent of setcc
10971 // FIXME: Turn all of these into setcc if setcc if setcc is legal
10972 // otherwise, go ahead with the folds.
10973 if (0 && N3C && N3C->isNullValue() && N2C && (N2C->getAPIntValue() == 1ULL)) {
10974 EVT XType = N0.getValueType();
10975 if (!LegalOperations ||
10976 TLI.isOperationLegal(ISD::SETCC, getSetCCResultType(XType))) {
10977 SDValue Res = DAG.getSetCC(DL, getSetCCResultType(XType), N0, N1, CC);
10978 if (Res.getValueType() != VT)
10979 Res = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Res);
10983 // fold (seteq X, 0) -> (srl (ctlz X, log2(size(X))))
10984 if (N1C && N1C->isNullValue() && CC == ISD::SETEQ &&
10985 (!LegalOperations ||
10986 TLI.isOperationLegal(ISD::CTLZ, XType))) {
10987 SDValue Ctlz = DAG.getNode(ISD::CTLZ, SDLoc(N0), XType, N0);
10988 return DAG.getNode(ISD::SRL, DL, XType, Ctlz,
10989 DAG.getConstant(Log2_32(XType.getSizeInBits()),
10990 getShiftAmountTy(Ctlz.getValueType())));
10992 // fold (setgt X, 0) -> (srl (and (-X, ~X), size(X)-1))
10993 if (N1C && N1C->isNullValue() && CC == ISD::SETGT) {
10994 SDValue NegN0 = DAG.getNode(ISD::SUB, SDLoc(N0),
10995 XType, DAG.getConstant(0, XType), N0);
10996 SDValue NotN0 = DAG.getNOT(SDLoc(N0), N0, XType);
10997 return DAG.getNode(ISD::SRL, DL, XType,
10998 DAG.getNode(ISD::AND, DL, XType, NegN0, NotN0),
10999 DAG.getConstant(XType.getSizeInBits()-1,
11000 getShiftAmountTy(XType)));
11002 // fold (setgt X, -1) -> (xor (srl (X, size(X)-1), 1))
11003 if (N1C && N1C->isAllOnesValue() && CC == ISD::SETGT) {
11004 SDValue Sign = DAG.getNode(ISD::SRL, SDLoc(N0), XType, N0,
11005 DAG.getConstant(XType.getSizeInBits()-1,
11006 getShiftAmountTy(N0.getValueType())));
11007 return DAG.getNode(ISD::XOR, DL, XType, Sign, DAG.getConstant(1, XType));
11011 // Check to see if this is an integer abs.
11012 // select_cc setg[te] X, 0, X, -X ->
11013 // select_cc setgt X, -1, X, -X ->
11014 // select_cc setl[te] X, 0, -X, X ->
11015 // select_cc setlt X, 1, -X, X ->
11016 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
11018 ConstantSDNode *SubC = NULL;
11019 if (((N1C->isNullValue() && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
11020 (N1C->isAllOnesValue() && CC == ISD::SETGT)) &&
11021 N0 == N2 && N3.getOpcode() == ISD::SUB && N0 == N3.getOperand(1))
11022 SubC = dyn_cast<ConstantSDNode>(N3.getOperand(0));
11023 else if (((N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE)) ||
11024 (N1C->isOne() && CC == ISD::SETLT)) &&
11025 N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1))
11026 SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0));
11028 EVT XType = N0.getValueType();
11029 if (SubC && SubC->isNullValue() && XType.isInteger()) {
11030 SDValue Shift = DAG.getNode(ISD::SRA, SDLoc(N0), XType,
11032 DAG.getConstant(XType.getSizeInBits()-1,
11033 getShiftAmountTy(N0.getValueType())));
11034 SDValue Add = DAG.getNode(ISD::ADD, SDLoc(N0),
11036 AddToWorkList(Shift.getNode());
11037 AddToWorkList(Add.getNode());
11038 return DAG.getNode(ISD::XOR, DL, XType, Add, Shift);
11045 /// SimplifySetCC - This is a stub for TargetLowering::SimplifySetCC.
11046 SDValue DAGCombiner::SimplifySetCC(EVT VT, SDValue N0,
11047 SDValue N1, ISD::CondCode Cond,
11048 SDLoc DL, bool foldBooleans) {
11049 TargetLowering::DAGCombinerInfo
11050 DagCombineInfo(DAG, Level, false, this);
11051 return TLI.SimplifySetCC(VT, N0, N1, Cond, foldBooleans, DagCombineInfo, DL);
11054 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
11055 /// return a DAG expression to select that will generate the same value by
11056 /// multiplying by a magic number. See:
11057 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
11058 SDValue DAGCombiner::BuildSDIV(SDNode *N) {
11059 std::vector<SDNode*> Built;
11060 SDValue S = TLI.BuildSDIV(N, DAG, LegalOperations, &Built);
11062 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end();
11064 AddToWorkList(*ii);
11068 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
11069 /// return a DAG expression to select that will generate the same value by
11070 /// multiplying by a magic number. See:
11071 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
11072 SDValue DAGCombiner::BuildUDIV(SDNode *N) {
11073 std::vector<SDNode*> Built;
11074 SDValue S = TLI.BuildUDIV(N, DAG, LegalOperations, &Built);
11076 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end();
11078 AddToWorkList(*ii);
11082 /// FindBaseOffset - Return true if base is a frame index, which is known not
11083 // to alias with anything but itself. Provides base object and offset as
11085 static bool FindBaseOffset(SDValue Ptr, SDValue &Base, int64_t &Offset,
11086 const GlobalValue *&GV, const void *&CV) {
11087 // Assume it is a primitive operation.
11088 Base = Ptr; Offset = 0; GV = 0; CV = 0;
11090 // If it's an adding a simple constant then integrate the offset.
11091 if (Base.getOpcode() == ISD::ADD) {
11092 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Base.getOperand(1))) {
11093 Base = Base.getOperand(0);
11094 Offset += C->getZExtValue();
11098 // Return the underlying GlobalValue, and update the Offset. Return false
11099 // for GlobalAddressSDNode since the same GlobalAddress may be represented
11100 // by multiple nodes with different offsets.
11101 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Base)) {
11102 GV = G->getGlobal();
11103 Offset += G->getOffset();
11107 // Return the underlying Constant value, and update the Offset. Return false
11108 // for ConstantSDNodes since the same constant pool entry may be represented
11109 // by multiple nodes with different offsets.
11110 if (ConstantPoolSDNode *C = dyn_cast<ConstantPoolSDNode>(Base)) {
11111 CV = C->isMachineConstantPoolEntry() ? (const void *)C->getMachineCPVal()
11112 : (const void *)C->getConstVal();
11113 Offset += C->getOffset();
11116 // If it's any of the following then it can't alias with anything but itself.
11117 return isa<FrameIndexSDNode>(Base);
11120 /// isAlias - Return true if there is any possibility that the two addresses
11122 bool DAGCombiner::isAlias(SDValue Ptr1, int64_t Size1, bool IsVolatile1,
11123 const Value *SrcValue1, int SrcValueOffset1,
11124 unsigned SrcValueAlign1,
11125 const MDNode *TBAAInfo1,
11126 SDValue Ptr2, int64_t Size2, bool IsVolatile2,
11127 const Value *SrcValue2, int SrcValueOffset2,
11128 unsigned SrcValueAlign2,
11129 const MDNode *TBAAInfo2) const {
11130 // If they are the same then they must be aliases.
11131 if (Ptr1 == Ptr2) return true;
11133 // If they are both volatile then they cannot be reordered.
11134 if (IsVolatile1 && IsVolatile2) return true;
11136 // Gather base node and offset information.
11137 SDValue Base1, Base2;
11138 int64_t Offset1, Offset2;
11139 const GlobalValue *GV1, *GV2;
11140 const void *CV1, *CV2;
11141 bool isFrameIndex1 = FindBaseOffset(Ptr1, Base1, Offset1, GV1, CV1);
11142 bool isFrameIndex2 = FindBaseOffset(Ptr2, Base2, Offset2, GV2, CV2);
11144 // If they have a same base address then check to see if they overlap.
11145 if (Base1 == Base2 || (GV1 && (GV1 == GV2)) || (CV1 && (CV1 == CV2)))
11146 return !((Offset1 + Size1) <= Offset2 || (Offset2 + Size2) <= Offset1);
11148 // It is possible for different frame indices to alias each other, mostly
11149 // when tail call optimization reuses return address slots for arguments.
11150 // To catch this case, look up the actual index of frame indices to compute
11151 // the real alias relationship.
11152 if (isFrameIndex1 && isFrameIndex2) {
11153 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
11154 Offset1 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base1)->getIndex());
11155 Offset2 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base2)->getIndex());
11156 return !((Offset1 + Size1) <= Offset2 || (Offset2 + Size2) <= Offset1);
11159 // Otherwise, if we know what the bases are, and they aren't identical, then
11160 // we know they cannot alias.
11161 if ((isFrameIndex1 || CV1 || GV1) && (isFrameIndex2 || CV2 || GV2))
11164 // If we know required SrcValue1 and SrcValue2 have relatively large alignment
11165 // compared to the size and offset of the access, we may be able to prove they
11166 // do not alias. This check is conservative for now to catch cases created by
11167 // splitting vector types.
11168 if ((SrcValueAlign1 == SrcValueAlign2) &&
11169 (SrcValueOffset1 != SrcValueOffset2) &&
11170 (Size1 == Size2) && (SrcValueAlign1 > Size1)) {
11171 int64_t OffAlign1 = SrcValueOffset1 % SrcValueAlign1;
11172 int64_t OffAlign2 = SrcValueOffset2 % SrcValueAlign1;
11174 // There is no overlap between these relatively aligned accesses of similar
11175 // size, return no alias.
11176 if ((OffAlign1 + Size1) <= OffAlign2 || (OffAlign2 + Size2) <= OffAlign1)
11180 bool UseAA = CombinerGlobalAA.getNumOccurrences() > 0 ? CombinerGlobalAA :
11181 TLI.getTargetMachine().getSubtarget<TargetSubtargetInfo>().useAA();
11183 if (CombinerAAOnlyFunc.getNumOccurrences() &&
11184 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
11187 if (UseAA && SrcValue1 && SrcValue2) {
11188 // Use alias analysis information.
11189 int64_t MinOffset = std::min(SrcValueOffset1, SrcValueOffset2);
11190 int64_t Overlap1 = Size1 + SrcValueOffset1 - MinOffset;
11191 int64_t Overlap2 = Size2 + SrcValueOffset2 - MinOffset;
11192 AliasAnalysis::AliasResult AAResult =
11193 AA.alias(AliasAnalysis::Location(SrcValue1, Overlap1,
11194 UseTBAA ? TBAAInfo1 : 0),
11195 AliasAnalysis::Location(SrcValue2, Overlap2,
11196 UseTBAA ? TBAAInfo2 : 0));
11197 if (AAResult == AliasAnalysis::NoAlias)
11201 // Otherwise we have to assume they alias.
11205 bool DAGCombiner::isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) {
11206 SDValue Ptr0, Ptr1;
11207 int64_t Size0, Size1;
11208 bool IsVolatile0, IsVolatile1;
11209 const Value *SrcValue0, *SrcValue1;
11210 int SrcValueOffset0, SrcValueOffset1;
11211 unsigned SrcValueAlign0, SrcValueAlign1;
11212 const MDNode *SrcTBAAInfo0, *SrcTBAAInfo1;
11213 FindAliasInfo(Op0, Ptr0, Size0, IsVolatile0, SrcValue0, SrcValueOffset0,
11214 SrcValueAlign0, SrcTBAAInfo0);
11215 FindAliasInfo(Op1, Ptr1, Size1, IsVolatile1, SrcValue1, SrcValueOffset1,
11216 SrcValueAlign1, SrcTBAAInfo1);
11217 return isAlias(Ptr0, Size0, IsVolatile0, SrcValue0, SrcValueOffset0,
11218 SrcValueAlign0, SrcTBAAInfo0,
11219 Ptr1, Size1, IsVolatile1, SrcValue1, SrcValueOffset1,
11220 SrcValueAlign1, SrcTBAAInfo1);
11223 /// FindAliasInfo - Extracts the relevant alias information from the memory
11224 /// node. Returns true if the operand was a nonvolatile load.
11225 bool DAGCombiner::FindAliasInfo(SDNode *N,
11226 SDValue &Ptr, int64_t &Size, bool &IsVolatile,
11227 const Value *&SrcValue,
11228 int &SrcValueOffset,
11229 unsigned &SrcValueAlign,
11230 const MDNode *&TBAAInfo) const {
11231 LSBaseSDNode *LS = cast<LSBaseSDNode>(N);
11233 Ptr = LS->getBasePtr();
11234 Size = LS->getMemoryVT().getSizeInBits() >> 3;
11235 IsVolatile = LS->isVolatile();
11236 SrcValue = LS->getSrcValue();
11237 SrcValueOffset = LS->getSrcValueOffset();
11238 SrcValueAlign = LS->getOriginalAlignment();
11239 TBAAInfo = LS->getTBAAInfo();
11240 return isa<LoadSDNode>(LS) && !IsVolatile;
11243 /// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes,
11244 /// looking for aliasing nodes and adding them to the Aliases vector.
11245 void DAGCombiner::GatherAllAliases(SDNode *N, SDValue OriginalChain,
11246 SmallVectorImpl<SDValue> &Aliases) {
11247 SmallVector<SDValue, 8> Chains; // List of chains to visit.
11248 SmallPtrSet<SDNode *, 16> Visited; // Visited node set.
11250 // Get alias information for node.
11254 const Value *SrcValue;
11255 int SrcValueOffset;
11256 unsigned SrcValueAlign;
11257 const MDNode *SrcTBAAInfo;
11258 bool IsLoad = FindAliasInfo(N, Ptr, Size, IsVolatile, SrcValue,
11259 SrcValueOffset, SrcValueAlign, SrcTBAAInfo);
11262 Chains.push_back(OriginalChain);
11263 unsigned Depth = 0;
11265 // Look at each chain and determine if it is an alias. If so, add it to the
11266 // aliases list. If not, then continue up the chain looking for the next
11268 while (!Chains.empty()) {
11269 SDValue Chain = Chains.back();
11272 // For TokenFactor nodes, look at each operand and only continue up the
11273 // chain until we find two aliases. If we've seen two aliases, assume we'll
11274 // find more and revert to original chain since the xform is unlikely to be
11277 // FIXME: The depth check could be made to return the last non-aliasing
11278 // chain we found before we hit a tokenfactor rather than the original
11280 if (Depth > 6 || Aliases.size() == 2) {
11282 Aliases.push_back(OriginalChain);
11286 // Don't bother if we've been before.
11287 if (!Visited.insert(Chain.getNode()))
11290 switch (Chain.getOpcode()) {
11291 case ISD::EntryToken:
11292 // Entry token is ideal chain operand, but handled in FindBetterChain.
11297 // Get alias information for Chain.
11301 const Value *OpSrcValue;
11302 int OpSrcValueOffset;
11303 unsigned OpSrcValueAlign;
11304 const MDNode *OpSrcTBAAInfo;
11305 bool IsOpLoad = FindAliasInfo(Chain.getNode(), OpPtr, OpSize,
11306 OpIsVolatile, OpSrcValue, OpSrcValueOffset,
11310 // If chain is alias then stop here.
11311 if (!(IsLoad && IsOpLoad) &&
11312 isAlias(Ptr, Size, IsVolatile, SrcValue, SrcValueOffset,
11313 SrcValueAlign, SrcTBAAInfo,
11314 OpPtr, OpSize, OpIsVolatile, OpSrcValue, OpSrcValueOffset,
11315 OpSrcValueAlign, OpSrcTBAAInfo)) {
11316 Aliases.push_back(Chain);
11318 // Look further up the chain.
11319 Chains.push_back(Chain.getOperand(0));
11325 case ISD::TokenFactor:
11326 // We have to check each of the operands of the token factor for "small"
11327 // token factors, so we queue them up. Adding the operands to the queue
11328 // (stack) in reverse order maintains the original order and increases the
11329 // likelihood that getNode will find a matching token factor (CSE.)
11330 if (Chain.getNumOperands() > 16) {
11331 Aliases.push_back(Chain);
11334 for (unsigned n = Chain.getNumOperands(); n;)
11335 Chains.push_back(Chain.getOperand(--n));
11340 // For all other instructions we will just have to take what we can get.
11341 Aliases.push_back(Chain);
11346 // We need to be careful here to also search for aliases through the
11347 // value operand of a store, etc. Consider the following situation:
11349 // L1 = load Token1, %52
11350 // S1 = store Token1, L1, %51
11351 // L2 = load Token1, %52+8
11352 // S2 = store Token1, L2, %51+8
11353 // Token2 = Token(S1, S2)
11354 // L3 = load Token2, %53
11355 // S3 = store Token2, L3, %52
11356 // L4 = load Token2, %53+8
11357 // S4 = store Token2, L4, %52+8
11358 // If we search for aliases of S3 (which loads address %52), and we look
11359 // only through the chain, then we'll miss the trivial dependence on L1
11360 // (which also loads from %52). We then might change all loads and
11361 // stores to use Token1 as their chain operand, which could result in
11362 // copying %53 into %52 before copying %52 into %51 (which should
11365 // The problem is, however, that searching for such data dependencies
11366 // can become expensive, and the cost is not directly related to the
11367 // chain depth. Instead, we'll rule out such configurations here by
11368 // insisting that we've visited all chain users (except for users
11369 // of the original chain, which is not necessary). When doing this,
11370 // we need to look through nodes we don't care about (otherwise, things
11371 // like register copies will interfere with trivial cases).
11373 SmallVector<const SDNode *, 16> Worklist;
11374 for (SmallPtrSet<SDNode *, 16>::iterator I = Visited.begin(),
11375 IE = Visited.end(); I != IE; ++I)
11376 if (*I != OriginalChain.getNode())
11377 Worklist.push_back(*I);
11379 while (!Worklist.empty()) {
11380 const SDNode *M = Worklist.pop_back_val();
11382 // We have already visited M, and want to make sure we've visited any uses
11383 // of M that we care about. For uses that we've not visisted, and don't
11384 // care about, queue them to the worklist.
11386 for (SDNode::use_iterator UI = M->use_begin(),
11387 UIE = M->use_end(); UI != UIE; ++UI)
11388 if (UI.getUse().getValueType() == MVT::Other && Visited.insert(*UI)) {
11389 if (isa<MemIntrinsicSDNode>(*UI) || isa<MemSDNode>(*UI)) {
11390 // We've not visited this use, and we care about it (it could have an
11391 // ordering dependency with the original node).
11393 Aliases.push_back(OriginalChain);
11397 // We've not visited this use, but we don't care about it. Mark it as
11398 // visited and enqueue it to the worklist.
11399 Worklist.push_back(*UI);
11404 /// FindBetterChain - Walk up chain skipping non-aliasing memory nodes, looking
11405 /// for a better chain (aliasing node.)
11406 SDValue DAGCombiner::FindBetterChain(SDNode *N, SDValue OldChain) {
11407 SmallVector<SDValue, 8> Aliases; // Ops for replacing token factor.
11409 // Accumulate all the aliases to this node.
11410 GatherAllAliases(N, OldChain, Aliases);
11412 // If no operands then chain to entry token.
11413 if (Aliases.size() == 0)
11414 return DAG.getEntryNode();
11416 // If a single operand then chain to it. We don't need to revisit it.
11417 if (Aliases.size() == 1)
11420 // Construct a custom tailored token factor.
11421 return DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other,
11422 &Aliases[0], Aliases.size());
11425 // SelectionDAG::Combine - This is the entry point for the file.
11427 void SelectionDAG::Combine(CombineLevel Level, AliasAnalysis &AA,
11428 CodeGenOpt::Level OptLevel) {
11429 /// run - This is the main entry point to this class.
11431 DAGCombiner(*this, AA, OptLevel).Run(Level);