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 #include "llvm/CodeGen/SelectionDAG.h"
20 #include "llvm/ADT/SmallPtrSet.h"
21 #include "llvm/ADT/Statistic.h"
22 #include "llvm/Analysis/AliasAnalysis.h"
23 #include "llvm/CodeGen/MachineFrameInfo.h"
24 #include "llvm/CodeGen/MachineFunction.h"
25 #include "llvm/IR/DataLayout.h"
26 #include "llvm/IR/DerivedTypes.h"
27 #include "llvm/IR/Function.h"
28 #include "llvm/IR/LLVMContext.h"
29 #include "llvm/Support/CommandLine.h"
30 #include "llvm/Support/Debug.h"
31 #include "llvm/Support/ErrorHandling.h"
32 #include "llvm/Support/MathExtras.h"
33 #include "llvm/Support/raw_ostream.h"
34 #include "llvm/Target/TargetLowering.h"
35 #include "llvm/Target/TargetMachine.h"
36 #include "llvm/Target/TargetOptions.h"
37 #include "llvm/Target/TargetRegisterInfo.h"
38 #include "llvm/Target/TargetSubtargetInfo.h"
42 #define DEBUG_TYPE "dagcombine"
44 STATISTIC(NodesCombined , "Number of dag nodes combined");
45 STATISTIC(PreIndexedNodes , "Number of pre-indexed nodes created");
46 STATISTIC(PostIndexedNodes, "Number of post-indexed nodes created");
47 STATISTIC(OpsNarrowed , "Number of load/op/store narrowed");
48 STATISTIC(LdStFP2Int , "Number of fp load/store pairs transformed to int");
49 STATISTIC(SlicedLoads, "Number of load sliced");
53 CombinerAA("combiner-alias-analysis", cl::Hidden,
54 cl::desc("Enable DAG combiner alias-analysis heuristics"));
57 CombinerGlobalAA("combiner-global-alias-analysis", cl::Hidden,
58 cl::desc("Enable DAG combiner's use of IR alias analysis"));
61 UseTBAA("combiner-use-tbaa", cl::Hidden, cl::init(true),
62 cl::desc("Enable DAG combiner's use of TBAA"));
65 static cl::opt<std::string>
66 CombinerAAOnlyFunc("combiner-aa-only-func", cl::Hidden,
67 cl::desc("Only use DAG-combiner alias analysis in this"
71 /// Hidden option to stress test load slicing, i.e., when this option
72 /// is enabled, load slicing bypasses most of its profitability guards.
74 StressLoadSlicing("combiner-stress-load-slicing", cl::Hidden,
75 cl::desc("Bypass the profitability model of load "
79 //------------------------------ DAGCombiner ---------------------------------//
83 const TargetLowering &TLI;
85 CodeGenOpt::Level OptLevel;
90 // Worklist of all of the nodes that need to be simplified.
92 // This has the semantics that when adding to the worklist,
93 // the item added must be next to be processed. It should
94 // also only appear once. The naive approach to this takes
97 // To reduce the insert/remove time to logarithmic, we use
98 // a set and a vector to maintain our worklist.
100 // The set contains the items on the worklist, but does not
101 // maintain the order they should be visited.
103 // The vector maintains the order nodes should be visited, but may
104 // contain duplicate or removed nodes. When choosing a node to
105 // visit, we pop off the order stack until we find an item that is
106 // also in the contents set. All operations are O(log N).
107 SmallPtrSet<SDNode*, 64> WorkListContents;
108 SmallVector<SDNode*, 64> WorkListOrder;
110 // AA - Used for DAG load/store alias analysis.
113 /// AddUsersToWorkList - When an instruction is simplified, add all users of
114 /// the instruction to the work lists because they might get more simplified
117 void AddUsersToWorkList(SDNode *N) {
118 for (SDNode *Node : N->uses())
122 /// visit - call the node-specific routine that knows how to fold each
123 /// particular type of node.
124 SDValue visit(SDNode *N);
127 /// AddToWorkList - Add to the work list making sure its instance is at the
128 /// back (next to be processed.)
129 void AddToWorkList(SDNode *N) {
130 WorkListContents.insert(N);
131 WorkListOrder.push_back(N);
134 /// removeFromWorkList - remove all instances of N from the worklist.
136 void removeFromWorkList(SDNode *N) {
137 WorkListContents.erase(N);
140 SDValue CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
143 SDValue CombineTo(SDNode *N, SDValue Res, bool AddTo = true) {
144 return CombineTo(N, &Res, 1, AddTo);
147 SDValue CombineTo(SDNode *N, SDValue Res0, SDValue Res1,
149 SDValue To[] = { Res0, Res1 };
150 return CombineTo(N, To, 2, AddTo);
153 void CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO);
157 /// SimplifyDemandedBits - Check the specified integer node value to see if
158 /// it can be simplified or if things it uses can be simplified by bit
159 /// propagation. If so, return true.
160 bool SimplifyDemandedBits(SDValue Op) {
161 unsigned BitWidth = Op.getValueType().getScalarType().getSizeInBits();
162 APInt Demanded = APInt::getAllOnesValue(BitWidth);
163 return SimplifyDemandedBits(Op, Demanded);
166 bool SimplifyDemandedBits(SDValue Op, const APInt &Demanded);
168 bool CombineToPreIndexedLoadStore(SDNode *N);
169 bool CombineToPostIndexedLoadStore(SDNode *N);
170 bool SliceUpLoad(SDNode *N);
172 /// \brief Replace an ISD::EXTRACT_VECTOR_ELT of a load with a narrowed
175 /// \param EVE ISD::EXTRACT_VECTOR_ELT to be replaced.
176 /// \param InVecVT type of the input vector to EVE with bitcasts resolved.
177 /// \param EltNo index of the vector element to load.
178 /// \param OriginalLoad load that EVE came from to be replaced.
179 /// \returns EVE on success SDValue() on failure.
180 SDValue ReplaceExtractVectorEltOfLoadWithNarrowedLoad(
181 SDNode *EVE, EVT InVecVT, SDValue EltNo, LoadSDNode *OriginalLoad);
182 void ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad);
183 SDValue PromoteOperand(SDValue Op, EVT PVT, bool &Replace);
184 SDValue SExtPromoteOperand(SDValue Op, EVT PVT);
185 SDValue ZExtPromoteOperand(SDValue Op, EVT PVT);
186 SDValue PromoteIntBinOp(SDValue Op);
187 SDValue PromoteIntShiftOp(SDValue Op);
188 SDValue PromoteExtend(SDValue Op);
189 bool PromoteLoad(SDValue Op);
191 void ExtendSetCCUses(const SmallVectorImpl<SDNode *> &SetCCs,
192 SDValue Trunc, SDValue ExtLoad, SDLoc DL,
193 ISD::NodeType ExtType);
195 /// combine - call the node-specific routine that knows how to fold each
196 /// particular type of node. If that doesn't do anything, try the
197 /// target-specific DAG combines.
198 SDValue combine(SDNode *N);
200 // Visitation implementation - Implement dag node combining for different
201 // node types. The semantics are as follows:
203 // SDValue.getNode() == 0 - No change was made
204 // SDValue.getNode() == N - N was replaced, is dead and has been handled.
205 // otherwise - N should be replaced by the returned Operand.
207 SDValue visitTokenFactor(SDNode *N);
208 SDValue visitMERGE_VALUES(SDNode *N);
209 SDValue visitADD(SDNode *N);
210 SDValue visitSUB(SDNode *N);
211 SDValue visitADDC(SDNode *N);
212 SDValue visitSUBC(SDNode *N);
213 SDValue visitADDE(SDNode *N);
214 SDValue visitSUBE(SDNode *N);
215 SDValue visitMUL(SDNode *N);
216 SDValue visitSDIV(SDNode *N);
217 SDValue visitUDIV(SDNode *N);
218 SDValue visitSREM(SDNode *N);
219 SDValue visitUREM(SDNode *N);
220 SDValue visitMULHU(SDNode *N);
221 SDValue visitMULHS(SDNode *N);
222 SDValue visitSMUL_LOHI(SDNode *N);
223 SDValue visitUMUL_LOHI(SDNode *N);
224 SDValue visitSMULO(SDNode *N);
225 SDValue visitUMULO(SDNode *N);
226 SDValue visitSDIVREM(SDNode *N);
227 SDValue visitUDIVREM(SDNode *N);
228 SDValue visitAND(SDNode *N);
229 SDValue visitOR(SDNode *N);
230 SDValue visitXOR(SDNode *N);
231 SDValue SimplifyVBinOp(SDNode *N);
232 SDValue SimplifyVUnaryOp(SDNode *N);
233 SDValue visitSHL(SDNode *N);
234 SDValue visitSRA(SDNode *N);
235 SDValue visitSRL(SDNode *N);
236 SDValue visitRotate(SDNode *N);
237 SDValue visitCTLZ(SDNode *N);
238 SDValue visitCTLZ_ZERO_UNDEF(SDNode *N);
239 SDValue visitCTTZ(SDNode *N);
240 SDValue visitCTTZ_ZERO_UNDEF(SDNode *N);
241 SDValue visitCTPOP(SDNode *N);
242 SDValue visitSELECT(SDNode *N);
243 SDValue visitVSELECT(SDNode *N);
244 SDValue visitSELECT_CC(SDNode *N);
245 SDValue visitSETCC(SDNode *N);
246 SDValue visitSIGN_EXTEND(SDNode *N);
247 SDValue visitZERO_EXTEND(SDNode *N);
248 SDValue visitANY_EXTEND(SDNode *N);
249 SDValue visitSIGN_EXTEND_INREG(SDNode *N);
250 SDValue visitTRUNCATE(SDNode *N);
251 SDValue visitBITCAST(SDNode *N);
252 SDValue visitBUILD_PAIR(SDNode *N);
253 SDValue visitFADD(SDNode *N);
254 SDValue visitFSUB(SDNode *N);
255 SDValue visitFMUL(SDNode *N);
256 SDValue visitFMA(SDNode *N);
257 SDValue visitFDIV(SDNode *N);
258 SDValue visitFREM(SDNode *N);
259 SDValue visitFCOPYSIGN(SDNode *N);
260 SDValue visitSINT_TO_FP(SDNode *N);
261 SDValue visitUINT_TO_FP(SDNode *N);
262 SDValue visitFP_TO_SINT(SDNode *N);
263 SDValue visitFP_TO_UINT(SDNode *N);
264 SDValue visitFP_ROUND(SDNode *N);
265 SDValue visitFP_ROUND_INREG(SDNode *N);
266 SDValue visitFP_EXTEND(SDNode *N);
267 SDValue visitFNEG(SDNode *N);
268 SDValue visitFABS(SDNode *N);
269 SDValue visitFCEIL(SDNode *N);
270 SDValue visitFTRUNC(SDNode *N);
271 SDValue visitFFLOOR(SDNode *N);
272 SDValue visitBRCOND(SDNode *N);
273 SDValue visitBR_CC(SDNode *N);
274 SDValue visitLOAD(SDNode *N);
275 SDValue visitSTORE(SDNode *N);
276 SDValue visitINSERT_VECTOR_ELT(SDNode *N);
277 SDValue visitEXTRACT_VECTOR_ELT(SDNode *N);
278 SDValue visitBUILD_VECTOR(SDNode *N);
279 SDValue visitCONCAT_VECTORS(SDNode *N);
280 SDValue visitEXTRACT_SUBVECTOR(SDNode *N);
281 SDValue visitVECTOR_SHUFFLE(SDNode *N);
282 SDValue visitINSERT_SUBVECTOR(SDNode *N);
284 SDValue XformToShuffleWithZero(SDNode *N);
285 SDValue ReassociateOps(unsigned Opc, SDLoc DL, SDValue LHS, SDValue RHS);
287 SDValue visitShiftByConstant(SDNode *N, ConstantSDNode *Amt);
289 bool SimplifySelectOps(SDNode *SELECT, SDValue LHS, SDValue RHS);
290 SDValue SimplifyBinOpWithSameOpcodeHands(SDNode *N);
291 SDValue SimplifySelect(SDLoc DL, SDValue N0, SDValue N1, SDValue N2);
292 SDValue SimplifySelectCC(SDLoc DL, SDValue N0, SDValue N1, SDValue N2,
293 SDValue N3, ISD::CondCode CC,
294 bool NotExtCompare = false);
295 SDValue SimplifySetCC(EVT VT, SDValue N0, SDValue N1, ISD::CondCode Cond,
296 SDLoc DL, bool foldBooleans = true);
298 bool isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS,
300 bool isOneUseSetCC(SDValue N) const;
302 SDValue SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
304 SDValue CombineConsecutiveLoads(SDNode *N, EVT VT);
305 SDValue ConstantFoldBITCASTofBUILD_VECTOR(SDNode *, EVT);
306 SDValue BuildSDIV(SDNode *N);
307 SDValue BuildUDIV(SDNode *N);
308 SDValue MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
309 bool DemandHighBits = true);
310 SDValue MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1);
311 SDNode *MatchRotatePosNeg(SDValue Shifted, SDValue Pos, SDValue Neg,
312 SDValue InnerPos, SDValue InnerNeg,
313 unsigned PosOpcode, unsigned NegOpcode,
315 SDNode *MatchRotate(SDValue LHS, SDValue RHS, SDLoc DL);
316 SDValue ReduceLoadWidth(SDNode *N);
317 SDValue ReduceLoadOpStoreWidth(SDNode *N);
318 SDValue TransformFPLoadStorePair(SDNode *N);
319 SDValue reduceBuildVecExtToExtBuildVec(SDNode *N);
320 SDValue reduceBuildVecConvertToConvertBuildVec(SDNode *N);
322 SDValue GetDemandedBits(SDValue V, const APInt &Mask);
324 /// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes,
325 /// looking for aliasing nodes and adding them to the Aliases vector.
326 void GatherAllAliases(SDNode *N, SDValue OriginalChain,
327 SmallVectorImpl<SDValue> &Aliases);
329 /// isAlias - Return true if there is any possibility that the two addresses
331 bool isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) const;
333 /// FindBetterChain - Walk up chain skipping non-aliasing memory nodes,
334 /// looking for a better chain (aliasing node.)
335 SDValue FindBetterChain(SDNode *N, SDValue Chain);
337 /// Merge consecutive store operations into a wide store.
338 /// This optimization uses wide integers or vectors when possible.
339 /// \return True if some memory operations were changed.
340 bool MergeConsecutiveStores(StoreSDNode *N);
342 /// \brief Try to transform a truncation where C is a constant:
343 /// (trunc (and X, C)) -> (and (trunc X), (trunc C))
345 /// \p N needs to be a truncation and its first operand an AND. Other
346 /// requirements are checked by the function (e.g. that trunc is
347 /// single-use) and if missed an empty SDValue is returned.
348 SDValue distributeTruncateThroughAnd(SDNode *N);
351 DAGCombiner(SelectionDAG &D, AliasAnalysis &A, CodeGenOpt::Level OL)
352 : DAG(D), TLI(D.getTargetLoweringInfo()), Level(BeforeLegalizeTypes),
353 OptLevel(OL), LegalOperations(false), LegalTypes(false), AA(A) {
354 AttributeSet FnAttrs =
355 DAG.getMachineFunction().getFunction()->getAttributes();
357 FnAttrs.hasAttribute(AttributeSet::FunctionIndex,
358 Attribute::OptimizeForSize) ||
359 FnAttrs.hasAttribute(AttributeSet::FunctionIndex, Attribute::MinSize);
362 /// Run - runs the dag combiner on all nodes in the work list
363 void Run(CombineLevel AtLevel);
365 SelectionDAG &getDAG() const { return DAG; }
367 /// getShiftAmountTy - Returns a type large enough to hold any valid
368 /// shift amount - before type legalization these can be huge.
369 EVT getShiftAmountTy(EVT LHSTy) {
370 assert(LHSTy.isInteger() && "Shift amount is not an integer type!");
371 if (LHSTy.isVector())
373 return LegalTypes ? TLI.getScalarShiftAmountTy(LHSTy)
374 : TLI.getPointerTy();
377 /// isTypeLegal - This method returns true if we are running before type
378 /// legalization or if the specified VT is legal.
379 bool isTypeLegal(const EVT &VT) {
380 if (!LegalTypes) return true;
381 return TLI.isTypeLegal(VT);
384 /// getSetCCResultType - Convenience wrapper around
385 /// TargetLowering::getSetCCResultType
386 EVT getSetCCResultType(EVT VT) const {
387 return TLI.getSetCCResultType(*DAG.getContext(), VT);
394 /// WorkListRemover - This class is a DAGUpdateListener that removes any deleted
395 /// nodes from the worklist.
396 class WorkListRemover : public SelectionDAG::DAGUpdateListener {
399 explicit WorkListRemover(DAGCombiner &dc)
400 : SelectionDAG::DAGUpdateListener(dc.getDAG()), DC(dc) {}
402 void NodeDeleted(SDNode *N, SDNode *E) override {
403 DC.removeFromWorkList(N);
408 //===----------------------------------------------------------------------===//
409 // TargetLowering::DAGCombinerInfo implementation
410 //===----------------------------------------------------------------------===//
412 void TargetLowering::DAGCombinerInfo::AddToWorklist(SDNode *N) {
413 ((DAGCombiner*)DC)->AddToWorkList(N);
416 void TargetLowering::DAGCombinerInfo::RemoveFromWorklist(SDNode *N) {
417 ((DAGCombiner*)DC)->removeFromWorkList(N);
420 SDValue TargetLowering::DAGCombinerInfo::
421 CombineTo(SDNode *N, const std::vector<SDValue> &To, bool AddTo) {
422 return ((DAGCombiner*)DC)->CombineTo(N, &To[0], To.size(), AddTo);
425 SDValue TargetLowering::DAGCombinerInfo::
426 CombineTo(SDNode *N, SDValue Res, bool AddTo) {
427 return ((DAGCombiner*)DC)->CombineTo(N, Res, AddTo);
431 SDValue TargetLowering::DAGCombinerInfo::
432 CombineTo(SDNode *N, SDValue Res0, SDValue Res1, bool AddTo) {
433 return ((DAGCombiner*)DC)->CombineTo(N, Res0, Res1, AddTo);
436 void TargetLowering::DAGCombinerInfo::
437 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
438 return ((DAGCombiner*)DC)->CommitTargetLoweringOpt(TLO);
441 //===----------------------------------------------------------------------===//
443 //===----------------------------------------------------------------------===//
445 /// isNegatibleForFree - Return 1 if we can compute the negated form of the
446 /// specified expression for the same cost as the expression itself, or 2 if we
447 /// can compute the negated form more cheaply than the expression itself.
448 static char isNegatibleForFree(SDValue Op, bool LegalOperations,
449 const TargetLowering &TLI,
450 const TargetOptions *Options,
451 unsigned Depth = 0) {
452 // fneg is removable even if it has multiple uses.
453 if (Op.getOpcode() == ISD::FNEG) return 2;
455 // Don't allow anything with multiple uses.
456 if (!Op.hasOneUse()) return 0;
458 // Don't recurse exponentially.
459 if (Depth > 6) return 0;
461 switch (Op.getOpcode()) {
462 default: return false;
463 case ISD::ConstantFP:
464 // Don't invert constant FP values after legalize. The negated constant
465 // isn't necessarily legal.
466 return LegalOperations ? 0 : 1;
468 // FIXME: determine better conditions for this xform.
469 if (!Options->UnsafeFPMath) return 0;
471 // After operation legalization, it might not be legal to create new FSUBs.
472 if (LegalOperations &&
473 !TLI.isOperationLegalOrCustom(ISD::FSUB, Op.getValueType()))
476 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B)
477 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI,
480 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
481 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options,
484 // We can't turn -(A-B) into B-A when we honor signed zeros.
485 if (!Options->UnsafeFPMath) return 0;
487 // fold (fneg (fsub A, B)) -> (fsub B, A)
492 if (Options->HonorSignDependentRoundingFPMath()) return 0;
494 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y) or (fmul X, (fneg Y))
495 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI,
499 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options,
505 return isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI, Options,
510 /// GetNegatedExpression - If isNegatibleForFree returns true, this function
511 /// returns the newly negated expression.
512 static SDValue GetNegatedExpression(SDValue Op, SelectionDAG &DAG,
513 bool LegalOperations, unsigned Depth = 0) {
514 // fneg is removable even if it has multiple uses.
515 if (Op.getOpcode() == ISD::FNEG) return Op.getOperand(0);
517 // Don't allow anything with multiple uses.
518 assert(Op.hasOneUse() && "Unknown reuse!");
520 assert(Depth <= 6 && "GetNegatedExpression doesn't match isNegatibleForFree");
521 switch (Op.getOpcode()) {
522 default: llvm_unreachable("Unknown code");
523 case ISD::ConstantFP: {
524 APFloat V = cast<ConstantFPSDNode>(Op)->getValueAPF();
526 return DAG.getConstantFP(V, Op.getValueType());
529 // FIXME: determine better conditions for this xform.
530 assert(DAG.getTarget().Options.UnsafeFPMath);
532 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B)
533 if (isNegatibleForFree(Op.getOperand(0), LegalOperations,
534 DAG.getTargetLoweringInfo(),
535 &DAG.getTarget().Options, Depth+1))
536 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
537 GetNegatedExpression(Op.getOperand(0), DAG,
538 LegalOperations, Depth+1),
540 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
541 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
542 GetNegatedExpression(Op.getOperand(1), DAG,
543 LegalOperations, Depth+1),
546 // We can't turn -(A-B) into B-A when we honor signed zeros.
547 assert(DAG.getTarget().Options.UnsafeFPMath);
549 // fold (fneg (fsub 0, B)) -> B
550 if (ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(Op.getOperand(0)))
551 if (N0CFP->getValueAPF().isZero())
552 return Op.getOperand(1);
554 // fold (fneg (fsub A, B)) -> (fsub B, A)
555 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
556 Op.getOperand(1), Op.getOperand(0));
560 assert(!DAG.getTarget().Options.HonorSignDependentRoundingFPMath());
562 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y)
563 if (isNegatibleForFree(Op.getOperand(0), LegalOperations,
564 DAG.getTargetLoweringInfo(),
565 &DAG.getTarget().Options, Depth+1))
566 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
567 GetNegatedExpression(Op.getOperand(0), DAG,
568 LegalOperations, Depth+1),
571 // fold (fneg (fmul X, Y)) -> (fmul X, (fneg Y))
572 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
574 GetNegatedExpression(Op.getOperand(1), DAG,
575 LegalOperations, Depth+1));
579 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
580 GetNegatedExpression(Op.getOperand(0), DAG,
581 LegalOperations, Depth+1));
583 return DAG.getNode(ISD::FP_ROUND, SDLoc(Op), Op.getValueType(),
584 GetNegatedExpression(Op.getOperand(0), DAG,
585 LegalOperations, Depth+1),
590 // isSetCCEquivalent - Return true if this node is a setcc, or is a select_cc
591 // that selects between the target values used for true and false, making it
592 // equivalent to a setcc. Also, set the incoming LHS, RHS, and CC references to
593 // the appropriate nodes based on the type of node we are checking. This
594 // simplifies life a bit for the callers.
595 bool DAGCombiner::isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS,
597 if (N.getOpcode() == ISD::SETCC) {
598 LHS = N.getOperand(0);
599 RHS = N.getOperand(1);
600 CC = N.getOperand(2);
604 if (N.getOpcode() != ISD::SELECT_CC ||
605 !TLI.isConstTrueVal(N.getOperand(2).getNode()) ||
606 !TLI.isConstFalseVal(N.getOperand(3).getNode()))
609 LHS = N.getOperand(0);
610 RHS = N.getOperand(1);
611 CC = N.getOperand(4);
615 // isOneUseSetCC - Return true if this is a SetCC-equivalent operation with only
616 // one use. If this is true, it allows the users to invert the operation for
617 // free when it is profitable to do so.
618 bool DAGCombiner::isOneUseSetCC(SDValue N) const {
620 if (isSetCCEquivalent(N, N0, N1, N2) && N.getNode()->hasOneUse())
625 /// isConstantSplatVector - Returns true if N is a BUILD_VECTOR node whose
626 /// elements are all the same constant or undefined.
627 static bool isConstantSplatVector(SDNode *N, APInt& SplatValue) {
628 BuildVectorSDNode *C = dyn_cast<BuildVectorSDNode>(N);
633 unsigned SplatBitSize;
635 EVT EltVT = N->getValueType(0).getVectorElementType();
636 return (C->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
638 EltVT.getSizeInBits() >= SplatBitSize);
641 // \brief Returns the SDNode if it is a constant BuildVector or constant.
642 static SDNode *isConstantBuildVectorOrConstantInt(SDValue N) {
643 if (isa<ConstantSDNode>(N))
645 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N);
646 if(BV && BV->isConstant())
651 // \brief Returns the SDNode if it is a constant splat BuildVector or constant
653 static ConstantSDNode *isConstOrConstSplat(SDValue N) {
654 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N))
657 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N)) {
658 BitVector UndefElements;
659 ConstantSDNode *CN = BV->getConstantSplatNode(&UndefElements);
661 // BuildVectors can truncate their operands. Ignore that case here.
662 // FIXME: We blindly ignore splats which include undef which is overly
664 if (CN && UndefElements.none() &&
665 CN->getValueType(0) == N.getValueType().getScalarType())
672 SDValue DAGCombiner::ReassociateOps(unsigned Opc, SDLoc DL,
673 SDValue N0, SDValue N1) {
674 EVT VT = N0.getValueType();
675 if (N0.getOpcode() == Opc) {
676 if (SDNode *L = isConstantBuildVectorOrConstantInt(N0.getOperand(1))) {
677 if (SDNode *R = isConstantBuildVectorOrConstantInt(N1)) {
678 // reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2))
679 SDValue OpNode = DAG.FoldConstantArithmetic(Opc, VT, L, R);
680 if (!OpNode.getNode())
682 return DAG.getNode(Opc, DL, VT, N0.getOperand(0), OpNode);
684 if (N0.hasOneUse()) {
685 // reassoc. (op (op x, c1), y) -> (op (op x, y), c1) iff x+c1 has one
687 SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N0.getOperand(0), N1);
688 if (!OpNode.getNode())
690 AddToWorkList(OpNode.getNode());
691 return DAG.getNode(Opc, DL, VT, OpNode, N0.getOperand(1));
696 if (N1.getOpcode() == Opc) {
697 if (SDNode *R = isConstantBuildVectorOrConstantInt(N1.getOperand(1))) {
698 if (SDNode *L = isConstantBuildVectorOrConstantInt(N0)) {
699 // reassoc. (op c2, (op x, c1)) -> (op x, (op c1, c2))
700 SDValue OpNode = DAG.FoldConstantArithmetic(Opc, VT, R, L);
701 if (!OpNode.getNode())
703 return DAG.getNode(Opc, DL, VT, N1.getOperand(0), OpNode);
705 if (N1.hasOneUse()) {
706 // reassoc. (op y, (op x, c1)) -> (op (op x, y), c1) iff x+c1 has one
708 SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N1.getOperand(0), N0);
709 if (!OpNode.getNode())
711 AddToWorkList(OpNode.getNode());
712 return DAG.getNode(Opc, DL, VT, OpNode, N1.getOperand(1));
720 SDValue DAGCombiner::CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
722 assert(N->getNumValues() == NumTo && "Broken CombineTo call!");
724 DEBUG(dbgs() << "\nReplacing.1 ";
726 dbgs() << "\nWith: ";
727 To[0].getNode()->dump(&DAG);
728 dbgs() << " and " << NumTo-1 << " other values\n";
729 for (unsigned i = 0, e = NumTo; i != e; ++i)
730 assert((!To[i].getNode() ||
731 N->getValueType(i) == To[i].getValueType()) &&
732 "Cannot combine value to value of different type!"));
733 WorkListRemover DeadNodes(*this);
734 DAG.ReplaceAllUsesWith(N, To);
736 // Push the new nodes and any users onto the worklist
737 for (unsigned i = 0, e = NumTo; i != e; ++i) {
738 if (To[i].getNode()) {
739 AddToWorkList(To[i].getNode());
740 AddUsersToWorkList(To[i].getNode());
745 // Finally, if the node is now dead, remove it from the graph. The node
746 // may not be dead if the replacement process recursively simplified to
747 // something else needing this node.
748 if (N->use_empty()) {
749 // Nodes can be reintroduced into the worklist. Make sure we do not
750 // process a node that has been replaced.
751 removeFromWorkList(N);
753 // Finally, since the node is now dead, remove it from the graph.
756 return SDValue(N, 0);
760 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
761 // Replace all uses. If any nodes become isomorphic to other nodes and
762 // are deleted, make sure to remove them from our worklist.
763 WorkListRemover DeadNodes(*this);
764 DAG.ReplaceAllUsesOfValueWith(TLO.Old, TLO.New);
766 // Push the new node and any (possibly new) users onto the worklist.
767 AddToWorkList(TLO.New.getNode());
768 AddUsersToWorkList(TLO.New.getNode());
770 // Finally, if the node is now dead, remove it from the graph. The node
771 // may not be dead if the replacement process recursively simplified to
772 // something else needing this node.
773 if (TLO.Old.getNode()->use_empty()) {
774 removeFromWorkList(TLO.Old.getNode());
776 // If the operands of this node are only used by the node, they will now
777 // be dead. Make sure to visit them first to delete dead nodes early.
778 for (unsigned i = 0, e = TLO.Old.getNode()->getNumOperands(); i != e; ++i)
779 if (TLO.Old.getNode()->getOperand(i).getNode()->hasOneUse())
780 AddToWorkList(TLO.Old.getNode()->getOperand(i).getNode());
782 DAG.DeleteNode(TLO.Old.getNode());
786 /// SimplifyDemandedBits - Check the specified integer node value to see if
787 /// it can be simplified or if things it uses can be simplified by bit
788 /// propagation. If so, return true.
789 bool DAGCombiner::SimplifyDemandedBits(SDValue Op, const APInt &Demanded) {
790 TargetLowering::TargetLoweringOpt TLO(DAG, LegalTypes, LegalOperations);
791 APInt KnownZero, KnownOne;
792 if (!TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
796 AddToWorkList(Op.getNode());
798 // Replace the old value with the new one.
800 DEBUG(dbgs() << "\nReplacing.2 ";
801 TLO.Old.getNode()->dump(&DAG);
802 dbgs() << "\nWith: ";
803 TLO.New.getNode()->dump(&DAG);
806 CommitTargetLoweringOpt(TLO);
810 void DAGCombiner::ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad) {
812 EVT VT = Load->getValueType(0);
813 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, VT, SDValue(ExtLoad, 0));
815 DEBUG(dbgs() << "\nReplacing.9 ";
817 dbgs() << "\nWith: ";
818 Trunc.getNode()->dump(&DAG);
820 WorkListRemover DeadNodes(*this);
821 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 0), Trunc);
822 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), SDValue(ExtLoad, 1));
823 removeFromWorkList(Load);
824 DAG.DeleteNode(Load);
825 AddToWorkList(Trunc.getNode());
828 SDValue DAGCombiner::PromoteOperand(SDValue Op, EVT PVT, bool &Replace) {
831 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Op)) {
832 EVT MemVT = LD->getMemoryVT();
833 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
834 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT) ? ISD::ZEXTLOAD
836 : LD->getExtensionType();
838 return DAG.getExtLoad(ExtType, dl, PVT,
839 LD->getChain(), LD->getBasePtr(),
840 MemVT, LD->getMemOperand());
843 unsigned Opc = Op.getOpcode();
846 case ISD::AssertSext:
847 return DAG.getNode(ISD::AssertSext, dl, PVT,
848 SExtPromoteOperand(Op.getOperand(0), PVT),
850 case ISD::AssertZext:
851 return DAG.getNode(ISD::AssertZext, dl, PVT,
852 ZExtPromoteOperand(Op.getOperand(0), PVT),
854 case ISD::Constant: {
856 Op.getValueType().isByteSized() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
857 return DAG.getNode(ExtOpc, dl, PVT, Op);
861 if (!TLI.isOperationLegal(ISD::ANY_EXTEND, PVT))
863 return DAG.getNode(ISD::ANY_EXTEND, dl, PVT, Op);
866 SDValue DAGCombiner::SExtPromoteOperand(SDValue Op, EVT PVT) {
867 if (!TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, PVT))
869 EVT OldVT = Op.getValueType();
871 bool Replace = false;
872 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
873 if (!NewOp.getNode())
875 AddToWorkList(NewOp.getNode());
878 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
879 return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, NewOp.getValueType(), NewOp,
880 DAG.getValueType(OldVT));
883 SDValue DAGCombiner::ZExtPromoteOperand(SDValue Op, EVT PVT) {
884 EVT OldVT = Op.getValueType();
886 bool Replace = false;
887 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
888 if (!NewOp.getNode())
890 AddToWorkList(NewOp.getNode());
893 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
894 return DAG.getZeroExtendInReg(NewOp, dl, OldVT);
897 /// PromoteIntBinOp - Promote the specified integer binary operation if the
898 /// target indicates it is beneficial. e.g. On x86, it's usually better to
899 /// promote i16 operations to i32 since i16 instructions are longer.
900 SDValue DAGCombiner::PromoteIntBinOp(SDValue Op) {
901 if (!LegalOperations)
904 EVT VT = Op.getValueType();
905 if (VT.isVector() || !VT.isInteger())
908 // If operation type is 'undesirable', e.g. i16 on x86, consider
910 unsigned Opc = Op.getOpcode();
911 if (TLI.isTypeDesirableForOp(Opc, VT))
915 // Consult target whether it is a good idea to promote this operation and
916 // what's the right type to promote it to.
917 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
918 assert(PVT != VT && "Don't know what type to promote to!");
920 bool Replace0 = false;
921 SDValue N0 = Op.getOperand(0);
922 SDValue NN0 = PromoteOperand(N0, PVT, Replace0);
926 bool Replace1 = false;
927 SDValue N1 = Op.getOperand(1);
932 NN1 = PromoteOperand(N1, PVT, Replace1);
937 AddToWorkList(NN0.getNode());
939 AddToWorkList(NN1.getNode());
942 ReplaceLoadWithPromotedLoad(N0.getNode(), NN0.getNode());
944 ReplaceLoadWithPromotedLoad(N1.getNode(), NN1.getNode());
946 DEBUG(dbgs() << "\nPromoting ";
947 Op.getNode()->dump(&DAG));
949 return DAG.getNode(ISD::TRUNCATE, dl, VT,
950 DAG.getNode(Opc, dl, PVT, NN0, NN1));
955 /// PromoteIntShiftOp - Promote the specified integer shift operation if the
956 /// target indicates it is beneficial. e.g. On x86, it's usually better to
957 /// promote i16 operations to i32 since i16 instructions are longer.
958 SDValue DAGCombiner::PromoteIntShiftOp(SDValue Op) {
959 if (!LegalOperations)
962 EVT VT = Op.getValueType();
963 if (VT.isVector() || !VT.isInteger())
966 // If operation type is 'undesirable', e.g. i16 on x86, consider
968 unsigned Opc = Op.getOpcode();
969 if (TLI.isTypeDesirableForOp(Opc, VT))
973 // Consult target whether it is a good idea to promote this operation and
974 // what's the right type to promote it to.
975 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
976 assert(PVT != VT && "Don't know what type to promote to!");
978 bool Replace = false;
979 SDValue N0 = Op.getOperand(0);
981 N0 = SExtPromoteOperand(Op.getOperand(0), PVT);
982 else if (Opc == ISD::SRL)
983 N0 = ZExtPromoteOperand(Op.getOperand(0), PVT);
985 N0 = PromoteOperand(N0, PVT, Replace);
989 AddToWorkList(N0.getNode());
991 ReplaceLoadWithPromotedLoad(Op.getOperand(0).getNode(), N0.getNode());
993 DEBUG(dbgs() << "\nPromoting ";
994 Op.getNode()->dump(&DAG));
996 return DAG.getNode(ISD::TRUNCATE, dl, VT,
997 DAG.getNode(Opc, dl, PVT, N0, Op.getOperand(1)));
1002 SDValue DAGCombiner::PromoteExtend(SDValue Op) {
1003 if (!LegalOperations)
1006 EVT VT = Op.getValueType();
1007 if (VT.isVector() || !VT.isInteger())
1010 // If operation type is 'undesirable', e.g. i16 on x86, consider
1012 unsigned Opc = Op.getOpcode();
1013 if (TLI.isTypeDesirableForOp(Opc, VT))
1017 // Consult target whether it is a good idea to promote this operation and
1018 // what's the right type to promote it to.
1019 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1020 assert(PVT != VT && "Don't know what type to promote to!");
1021 // fold (aext (aext x)) -> (aext x)
1022 // fold (aext (zext x)) -> (zext x)
1023 // fold (aext (sext x)) -> (sext x)
1024 DEBUG(dbgs() << "\nPromoting ";
1025 Op.getNode()->dump(&DAG));
1026 return DAG.getNode(Op.getOpcode(), SDLoc(Op), VT, Op.getOperand(0));
1031 bool DAGCombiner::PromoteLoad(SDValue Op) {
1032 if (!LegalOperations)
1035 EVT VT = Op.getValueType();
1036 if (VT.isVector() || !VT.isInteger())
1039 // If operation type is 'undesirable', e.g. i16 on x86, consider
1041 unsigned Opc = Op.getOpcode();
1042 if (TLI.isTypeDesirableForOp(Opc, VT))
1046 // Consult target whether it is a good idea to promote this operation and
1047 // what's the right type to promote it to.
1048 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1049 assert(PVT != VT && "Don't know what type to promote to!");
1052 SDNode *N = Op.getNode();
1053 LoadSDNode *LD = cast<LoadSDNode>(N);
1054 EVT MemVT = LD->getMemoryVT();
1055 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
1056 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT) ? ISD::ZEXTLOAD
1058 : LD->getExtensionType();
1059 SDValue NewLD = DAG.getExtLoad(ExtType, dl, PVT,
1060 LD->getChain(), LD->getBasePtr(),
1061 MemVT, LD->getMemOperand());
1062 SDValue Result = DAG.getNode(ISD::TRUNCATE, dl, VT, NewLD);
1064 DEBUG(dbgs() << "\nPromoting ";
1067 Result.getNode()->dump(&DAG);
1069 WorkListRemover DeadNodes(*this);
1070 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result);
1071 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), NewLD.getValue(1));
1072 removeFromWorkList(N);
1074 AddToWorkList(Result.getNode());
1081 //===----------------------------------------------------------------------===//
1082 // Main DAG Combiner implementation
1083 //===----------------------------------------------------------------------===//
1085 void DAGCombiner::Run(CombineLevel AtLevel) {
1086 // set the instance variables, so that the various visit routines may use it.
1088 LegalOperations = Level >= AfterLegalizeVectorOps;
1089 LegalTypes = Level >= AfterLegalizeTypes;
1091 // Add all the dag nodes to the worklist.
1092 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
1093 E = DAG.allnodes_end(); I != E; ++I)
1096 // Create a dummy node (which is not added to allnodes), that adds a reference
1097 // to the root node, preventing it from being deleted, and tracking any
1098 // changes of the root.
1099 HandleSDNode Dummy(DAG.getRoot());
1101 // The root of the dag may dangle to deleted nodes until the dag combiner is
1102 // done. Set it to null to avoid confusion.
1103 DAG.setRoot(SDValue());
1105 // while the worklist isn't empty, find a node and
1106 // try and combine it.
1107 while (!WorkListContents.empty()) {
1109 // The WorkListOrder holds the SDNodes in order, but it may contain
1111 // In order to avoid a linear scan, we use a set (O(log N)) to hold what the
1112 // worklist *should* contain, and check the node we want to visit is should
1113 // actually be visited.
1115 N = WorkListOrder.pop_back_val();
1116 } while (!WorkListContents.erase(N));
1118 // If N has no uses, it is dead. Make sure to revisit all N's operands once
1119 // N is deleted from the DAG, since they too may now be dead or may have a
1120 // reduced number of uses, allowing other xforms.
1121 if (N->use_empty() && N != &Dummy) {
1122 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1123 AddToWorkList(N->getOperand(i).getNode());
1129 SDValue RV = combine(N);
1136 // If we get back the same node we passed in, rather than a new node or
1137 // zero, we know that the node must have defined multiple values and
1138 // CombineTo was used. Since CombineTo takes care of the worklist
1139 // mechanics for us, we have no work to do in this case.
1140 if (RV.getNode() == N)
1143 assert(N->getOpcode() != ISD::DELETED_NODE &&
1144 RV.getNode()->getOpcode() != ISD::DELETED_NODE &&
1145 "Node was deleted but visit returned new node!");
1147 DEBUG(dbgs() << "\nReplacing.3 ";
1149 dbgs() << "\nWith: ";
1150 RV.getNode()->dump(&DAG);
1153 // Transfer debug value.
1154 DAG.TransferDbgValues(SDValue(N, 0), RV);
1155 WorkListRemover DeadNodes(*this);
1156 if (N->getNumValues() == RV.getNode()->getNumValues())
1157 DAG.ReplaceAllUsesWith(N, RV.getNode());
1159 assert(N->getValueType(0) == RV.getValueType() &&
1160 N->getNumValues() == 1 && "Type mismatch");
1162 DAG.ReplaceAllUsesWith(N, &OpV);
1165 // Push the new node and any users onto the worklist
1166 AddToWorkList(RV.getNode());
1167 AddUsersToWorkList(RV.getNode());
1169 // Add any uses of the old node to the worklist in case this node is the
1170 // last one that uses them. They may become dead after this node is
1172 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1173 AddToWorkList(N->getOperand(i).getNode());
1175 // Finally, if the node is now dead, remove it from the graph. The node
1176 // may not be dead if the replacement process recursively simplified to
1177 // something else needing this node.
1178 if (N->use_empty()) {
1179 // Nodes can be reintroduced into the worklist. Make sure we do not
1180 // process a node that has been replaced.
1181 removeFromWorkList(N);
1183 // Finally, since the node is now dead, remove it from the graph.
1188 // If the root changed (e.g. it was a dead load, update the root).
1189 DAG.setRoot(Dummy.getValue());
1190 DAG.RemoveDeadNodes();
1193 SDValue DAGCombiner::visit(SDNode *N) {
1194 switch (N->getOpcode()) {
1196 case ISD::TokenFactor: return visitTokenFactor(N);
1197 case ISD::MERGE_VALUES: return visitMERGE_VALUES(N);
1198 case ISD::ADD: return visitADD(N);
1199 case ISD::SUB: return visitSUB(N);
1200 case ISD::ADDC: return visitADDC(N);
1201 case ISD::SUBC: return visitSUBC(N);
1202 case ISD::ADDE: return visitADDE(N);
1203 case ISD::SUBE: return visitSUBE(N);
1204 case ISD::MUL: return visitMUL(N);
1205 case ISD::SDIV: return visitSDIV(N);
1206 case ISD::UDIV: return visitUDIV(N);
1207 case ISD::SREM: return visitSREM(N);
1208 case ISD::UREM: return visitUREM(N);
1209 case ISD::MULHU: return visitMULHU(N);
1210 case ISD::MULHS: return visitMULHS(N);
1211 case ISD::SMUL_LOHI: return visitSMUL_LOHI(N);
1212 case ISD::UMUL_LOHI: return visitUMUL_LOHI(N);
1213 case ISD::SMULO: return visitSMULO(N);
1214 case ISD::UMULO: return visitUMULO(N);
1215 case ISD::SDIVREM: return visitSDIVREM(N);
1216 case ISD::UDIVREM: return visitUDIVREM(N);
1217 case ISD::AND: return visitAND(N);
1218 case ISD::OR: return visitOR(N);
1219 case ISD::XOR: return visitXOR(N);
1220 case ISD::SHL: return visitSHL(N);
1221 case ISD::SRA: return visitSRA(N);
1222 case ISD::SRL: return visitSRL(N);
1224 case ISD::ROTL: return visitRotate(N);
1225 case ISD::CTLZ: return visitCTLZ(N);
1226 case ISD::CTLZ_ZERO_UNDEF: return visitCTLZ_ZERO_UNDEF(N);
1227 case ISD::CTTZ: return visitCTTZ(N);
1228 case ISD::CTTZ_ZERO_UNDEF: return visitCTTZ_ZERO_UNDEF(N);
1229 case ISD::CTPOP: return visitCTPOP(N);
1230 case ISD::SELECT: return visitSELECT(N);
1231 case ISD::VSELECT: return visitVSELECT(N);
1232 case ISD::SELECT_CC: return visitSELECT_CC(N);
1233 case ISD::SETCC: return visitSETCC(N);
1234 case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N);
1235 case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N);
1236 case ISD::ANY_EXTEND: return visitANY_EXTEND(N);
1237 case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N);
1238 case ISD::TRUNCATE: return visitTRUNCATE(N);
1239 case ISD::BITCAST: return visitBITCAST(N);
1240 case ISD::BUILD_PAIR: return visitBUILD_PAIR(N);
1241 case ISD::FADD: return visitFADD(N);
1242 case ISD::FSUB: return visitFSUB(N);
1243 case ISD::FMUL: return visitFMUL(N);
1244 case ISD::FMA: return visitFMA(N);
1245 case ISD::FDIV: return visitFDIV(N);
1246 case ISD::FREM: return visitFREM(N);
1247 case ISD::FCOPYSIGN: return visitFCOPYSIGN(N);
1248 case ISD::SINT_TO_FP: return visitSINT_TO_FP(N);
1249 case ISD::UINT_TO_FP: return visitUINT_TO_FP(N);
1250 case ISD::FP_TO_SINT: return visitFP_TO_SINT(N);
1251 case ISD::FP_TO_UINT: return visitFP_TO_UINT(N);
1252 case ISD::FP_ROUND: return visitFP_ROUND(N);
1253 case ISD::FP_ROUND_INREG: return visitFP_ROUND_INREG(N);
1254 case ISD::FP_EXTEND: return visitFP_EXTEND(N);
1255 case ISD::FNEG: return visitFNEG(N);
1256 case ISD::FABS: return visitFABS(N);
1257 case ISD::FFLOOR: return visitFFLOOR(N);
1258 case ISD::FCEIL: return visitFCEIL(N);
1259 case ISD::FTRUNC: return visitFTRUNC(N);
1260 case ISD::BRCOND: return visitBRCOND(N);
1261 case ISD::BR_CC: return visitBR_CC(N);
1262 case ISD::LOAD: return visitLOAD(N);
1263 case ISD::STORE: return visitSTORE(N);
1264 case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N);
1265 case ISD::EXTRACT_VECTOR_ELT: return visitEXTRACT_VECTOR_ELT(N);
1266 case ISD::BUILD_VECTOR: return visitBUILD_VECTOR(N);
1267 case ISD::CONCAT_VECTORS: return visitCONCAT_VECTORS(N);
1268 case ISD::EXTRACT_SUBVECTOR: return visitEXTRACT_SUBVECTOR(N);
1269 case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N);
1270 case ISD::INSERT_SUBVECTOR: return visitINSERT_SUBVECTOR(N);
1275 SDValue DAGCombiner::combine(SDNode *N) {
1276 SDValue RV = visit(N);
1278 // If nothing happened, try a target-specific DAG combine.
1279 if (!RV.getNode()) {
1280 assert(N->getOpcode() != ISD::DELETED_NODE &&
1281 "Node was deleted but visit returned NULL!");
1283 if (N->getOpcode() >= ISD::BUILTIN_OP_END ||
1284 TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode())) {
1286 // Expose the DAG combiner to the target combiner impls.
1287 TargetLowering::DAGCombinerInfo
1288 DagCombineInfo(DAG, Level, false, this);
1290 RV = TLI.PerformDAGCombine(N, DagCombineInfo);
1294 // If nothing happened still, try promoting the operation.
1295 if (!RV.getNode()) {
1296 switch (N->getOpcode()) {
1304 RV = PromoteIntBinOp(SDValue(N, 0));
1309 RV = PromoteIntShiftOp(SDValue(N, 0));
1311 case ISD::SIGN_EXTEND:
1312 case ISD::ZERO_EXTEND:
1313 case ISD::ANY_EXTEND:
1314 RV = PromoteExtend(SDValue(N, 0));
1317 if (PromoteLoad(SDValue(N, 0)))
1323 // If N is a commutative binary node, try commuting it to enable more
1325 if (!RV.getNode() && SelectionDAG::isCommutativeBinOp(N->getOpcode()) &&
1326 N->getNumValues() == 1) {
1327 SDValue N0 = N->getOperand(0);
1328 SDValue N1 = N->getOperand(1);
1330 // Constant operands are canonicalized to RHS.
1331 if (isa<ConstantSDNode>(N0) || !isa<ConstantSDNode>(N1)) {
1332 SDValue Ops[] = {N1, N0};
1334 if (const BinaryWithFlagsSDNode *BinNode =
1335 dyn_cast<BinaryWithFlagsSDNode>(N)) {
1336 CSENode = DAG.getNodeIfExists(
1337 N->getOpcode(), N->getVTList(), Ops, BinNode->hasNoUnsignedWrap(),
1338 BinNode->hasNoSignedWrap(), BinNode->isExact());
1340 CSENode = DAG.getNodeIfExists(N->getOpcode(), N->getVTList(), Ops);
1343 return SDValue(CSENode, 0);
1350 /// getInputChainForNode - Given a node, return its input chain if it has one,
1351 /// otherwise return a null sd operand.
1352 static SDValue getInputChainForNode(SDNode *N) {
1353 if (unsigned NumOps = N->getNumOperands()) {
1354 if (N->getOperand(0).getValueType() == MVT::Other)
1355 return N->getOperand(0);
1356 if (N->getOperand(NumOps-1).getValueType() == MVT::Other)
1357 return N->getOperand(NumOps-1);
1358 for (unsigned i = 1; i < NumOps-1; ++i)
1359 if (N->getOperand(i).getValueType() == MVT::Other)
1360 return N->getOperand(i);
1365 SDValue DAGCombiner::visitTokenFactor(SDNode *N) {
1366 // If N has two operands, where one has an input chain equal to the other,
1367 // the 'other' chain is redundant.
1368 if (N->getNumOperands() == 2) {
1369 if (getInputChainForNode(N->getOperand(0).getNode()) == N->getOperand(1))
1370 return N->getOperand(0);
1371 if (getInputChainForNode(N->getOperand(1).getNode()) == N->getOperand(0))
1372 return N->getOperand(1);
1375 SmallVector<SDNode *, 8> TFs; // List of token factors to visit.
1376 SmallVector<SDValue, 8> Ops; // Ops for replacing token factor.
1377 SmallPtrSet<SDNode*, 16> SeenOps;
1378 bool Changed = false; // If we should replace this token factor.
1380 // Start out with this token factor.
1383 // Iterate through token factors. The TFs grows when new token factors are
1385 for (unsigned i = 0; i < TFs.size(); ++i) {
1386 SDNode *TF = TFs[i];
1388 // Check each of the operands.
1389 for (unsigned i = 0, ie = TF->getNumOperands(); i != ie; ++i) {
1390 SDValue Op = TF->getOperand(i);
1392 switch (Op.getOpcode()) {
1393 case ISD::EntryToken:
1394 // Entry tokens don't need to be added to the list. They are
1399 case ISD::TokenFactor:
1400 if (Op.hasOneUse() &&
1401 std::find(TFs.begin(), TFs.end(), Op.getNode()) == TFs.end()) {
1402 // Queue up for processing.
1403 TFs.push_back(Op.getNode());
1404 // Clean up in case the token factor is removed.
1405 AddToWorkList(Op.getNode());
1412 // Only add if it isn't already in the list.
1413 if (SeenOps.insert(Op.getNode()))
1424 // If we've change things around then replace token factor.
1427 // The entry token is the only possible outcome.
1428 Result = DAG.getEntryNode();
1430 // New and improved token factor.
1431 Result = DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, Ops);
1434 // Don't add users to work list.
1435 return CombineTo(N, Result, false);
1441 /// MERGE_VALUES can always be eliminated.
1442 SDValue DAGCombiner::visitMERGE_VALUES(SDNode *N) {
1443 WorkListRemover DeadNodes(*this);
1444 // Replacing results may cause a different MERGE_VALUES to suddenly
1445 // be CSE'd with N, and carry its uses with it. Iterate until no
1446 // uses remain, to ensure that the node can be safely deleted.
1447 // First add the users of this node to the work list so that they
1448 // can be tried again once they have new operands.
1449 AddUsersToWorkList(N);
1451 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1452 DAG.ReplaceAllUsesOfValueWith(SDValue(N, i), N->getOperand(i));
1453 } while (!N->use_empty());
1454 removeFromWorkList(N);
1456 return SDValue(N, 0); // Return N so it doesn't get rechecked!
1460 SDValue combineShlAddConstant(SDLoc DL, SDValue N0, SDValue N1,
1461 SelectionDAG &DAG) {
1462 EVT VT = N0.getValueType();
1463 SDValue N00 = N0.getOperand(0);
1464 SDValue N01 = N0.getOperand(1);
1465 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N01);
1467 if (N01C && N00.getOpcode() == ISD::ADD && N00.getNode()->hasOneUse() &&
1468 isa<ConstantSDNode>(N00.getOperand(1))) {
1469 // fold (add (shl (add x, c1), c2), ) -> (add (add (shl x, c2), c1<<c2), )
1470 N0 = DAG.getNode(ISD::ADD, SDLoc(N0), VT,
1471 DAG.getNode(ISD::SHL, SDLoc(N00), VT,
1472 N00.getOperand(0), N01),
1473 DAG.getNode(ISD::SHL, SDLoc(N01), VT,
1474 N00.getOperand(1), N01));
1475 return DAG.getNode(ISD::ADD, DL, VT, N0, N1);
1481 SDValue DAGCombiner::visitADD(SDNode *N) {
1482 SDValue N0 = N->getOperand(0);
1483 SDValue N1 = N->getOperand(1);
1484 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1485 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1486 EVT VT = N0.getValueType();
1489 if (VT.isVector()) {
1490 SDValue FoldedVOp = SimplifyVBinOp(N);
1491 if (FoldedVOp.getNode()) return FoldedVOp;
1493 // fold (add x, 0) -> x, vector edition
1494 if (ISD::isBuildVectorAllZeros(N1.getNode()))
1496 if (ISD::isBuildVectorAllZeros(N0.getNode()))
1500 // fold (add x, undef) -> undef
1501 if (N0.getOpcode() == ISD::UNDEF)
1503 if (N1.getOpcode() == ISD::UNDEF)
1505 // fold (add c1, c2) -> c1+c2
1507 return DAG.FoldConstantArithmetic(ISD::ADD, VT, N0C, N1C);
1508 // canonicalize constant to RHS
1510 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N1, N0);
1511 // fold (add x, 0) -> x
1512 if (N1C && N1C->isNullValue())
1514 // fold (add Sym, c) -> Sym+c
1515 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1516 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA) && N1C &&
1517 GA->getOpcode() == ISD::GlobalAddress)
1518 return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT,
1520 (uint64_t)N1C->getSExtValue());
1521 // fold ((c1-A)+c2) -> (c1+c2)-A
1522 if (N1C && N0.getOpcode() == ISD::SUB)
1523 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getOperand(0)))
1524 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1525 DAG.getConstant(N1C->getAPIntValue()+
1526 N0C->getAPIntValue(), VT),
1529 SDValue RADD = ReassociateOps(ISD::ADD, SDLoc(N), N0, N1);
1532 // fold ((0-A) + B) -> B-A
1533 if (N0.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N0.getOperand(0)) &&
1534 cast<ConstantSDNode>(N0.getOperand(0))->isNullValue())
1535 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1, N0.getOperand(1));
1536 // fold (A + (0-B)) -> A-B
1537 if (N1.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N1.getOperand(0)) &&
1538 cast<ConstantSDNode>(N1.getOperand(0))->isNullValue())
1539 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, N1.getOperand(1));
1540 // fold (A+(B-A)) -> B
1541 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1))
1542 return N1.getOperand(0);
1543 // fold ((B-A)+A) -> B
1544 if (N0.getOpcode() == ISD::SUB && N1 == N0.getOperand(1))
1545 return N0.getOperand(0);
1546 // fold (A+(B-(A+C))) to (B-C)
1547 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1548 N0 == N1.getOperand(1).getOperand(0))
1549 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1.getOperand(0),
1550 N1.getOperand(1).getOperand(1));
1551 // fold (A+(B-(C+A))) to (B-C)
1552 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1553 N0 == N1.getOperand(1).getOperand(1))
1554 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1.getOperand(0),
1555 N1.getOperand(1).getOperand(0));
1556 // fold (A+((B-A)+or-C)) to (B+or-C)
1557 if ((N1.getOpcode() == ISD::SUB || N1.getOpcode() == ISD::ADD) &&
1558 N1.getOperand(0).getOpcode() == ISD::SUB &&
1559 N0 == N1.getOperand(0).getOperand(1))
1560 return DAG.getNode(N1.getOpcode(), SDLoc(N), VT,
1561 N1.getOperand(0).getOperand(0), N1.getOperand(1));
1563 // fold (A-B)+(C-D) to (A+C)-(B+D) when A or C is constant
1564 if (N0.getOpcode() == ISD::SUB && N1.getOpcode() == ISD::SUB) {
1565 SDValue N00 = N0.getOperand(0);
1566 SDValue N01 = N0.getOperand(1);
1567 SDValue N10 = N1.getOperand(0);
1568 SDValue N11 = N1.getOperand(1);
1570 if (isa<ConstantSDNode>(N00) || isa<ConstantSDNode>(N10))
1571 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1572 DAG.getNode(ISD::ADD, SDLoc(N0), VT, N00, N10),
1573 DAG.getNode(ISD::ADD, SDLoc(N1), VT, N01, N11));
1576 if (!VT.isVector() && SimplifyDemandedBits(SDValue(N, 0)))
1577 return SDValue(N, 0);
1579 // fold (a+b) -> (a|b) iff a and b share no bits.
1580 if (VT.isInteger() && !VT.isVector()) {
1581 APInt LHSZero, LHSOne;
1582 APInt RHSZero, RHSOne;
1583 DAG.computeKnownBits(N0, LHSZero, LHSOne);
1585 if (LHSZero.getBoolValue()) {
1586 DAG.computeKnownBits(N1, RHSZero, RHSOne);
1588 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1589 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1590 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero){
1591 if (!LegalOperations || TLI.isOperationLegal(ISD::OR, VT))
1592 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1);
1597 // fold (add (shl (add x, c1), c2), ) -> (add (add (shl x, c2), c1<<c2), )
1598 if (N0.getOpcode() == ISD::SHL && N0.getNode()->hasOneUse()) {
1599 SDValue Result = combineShlAddConstant(SDLoc(N), N0, N1, DAG);
1600 if (Result.getNode()) return Result;
1602 if (N1.getOpcode() == ISD::SHL && N1.getNode()->hasOneUse()) {
1603 SDValue Result = combineShlAddConstant(SDLoc(N), N1, N0, DAG);
1604 if (Result.getNode()) return Result;
1607 // fold (add x, shl(0 - y, n)) -> sub(x, shl(y, n))
1608 if (N1.getOpcode() == ISD::SHL &&
1609 N1.getOperand(0).getOpcode() == ISD::SUB)
1610 if (ConstantSDNode *C =
1611 dyn_cast<ConstantSDNode>(N1.getOperand(0).getOperand(0)))
1612 if (C->getAPIntValue() == 0)
1613 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N0,
1614 DAG.getNode(ISD::SHL, SDLoc(N), VT,
1615 N1.getOperand(0).getOperand(1),
1617 if (N0.getOpcode() == ISD::SHL &&
1618 N0.getOperand(0).getOpcode() == ISD::SUB)
1619 if (ConstantSDNode *C =
1620 dyn_cast<ConstantSDNode>(N0.getOperand(0).getOperand(0)))
1621 if (C->getAPIntValue() == 0)
1622 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1,
1623 DAG.getNode(ISD::SHL, SDLoc(N), VT,
1624 N0.getOperand(0).getOperand(1),
1627 if (N1.getOpcode() == ISD::AND) {
1628 SDValue AndOp0 = N1.getOperand(0);
1629 ConstantSDNode *AndOp1 = dyn_cast<ConstantSDNode>(N1->getOperand(1));
1630 unsigned NumSignBits = DAG.ComputeNumSignBits(AndOp0);
1631 unsigned DestBits = VT.getScalarType().getSizeInBits();
1633 // (add z, (and (sbbl x, x), 1)) -> (sub z, (sbbl x, x))
1634 // and similar xforms where the inner op is either ~0 or 0.
1635 if (NumSignBits == DestBits && AndOp1 && AndOp1->isOne()) {
1637 return DAG.getNode(ISD::SUB, DL, VT, N->getOperand(0), AndOp0);
1641 // add (sext i1), X -> sub X, (zext i1)
1642 if (N0.getOpcode() == ISD::SIGN_EXTEND &&
1643 N0.getOperand(0).getValueType() == MVT::i1 &&
1644 !TLI.isOperationLegal(ISD::SIGN_EXTEND, MVT::i1)) {
1646 SDValue ZExt = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0));
1647 return DAG.getNode(ISD::SUB, DL, VT, N1, ZExt);
1653 SDValue DAGCombiner::visitADDC(SDNode *N) {
1654 SDValue N0 = N->getOperand(0);
1655 SDValue N1 = N->getOperand(1);
1656 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1657 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1658 EVT VT = N0.getValueType();
1660 // If the flag result is dead, turn this into an ADD.
1661 if (!N->hasAnyUseOfValue(1))
1662 return CombineTo(N, DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N1),
1663 DAG.getNode(ISD::CARRY_FALSE,
1664 SDLoc(N), MVT::Glue));
1666 // canonicalize constant to RHS.
1668 return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N1, N0);
1670 // fold (addc x, 0) -> x + no carry out
1671 if (N1C && N1C->isNullValue())
1672 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE,
1673 SDLoc(N), MVT::Glue));
1675 // fold (addc a, b) -> (or a, b), CARRY_FALSE iff a and b share no bits.
1676 APInt LHSZero, LHSOne;
1677 APInt RHSZero, RHSOne;
1678 DAG.computeKnownBits(N0, LHSZero, LHSOne);
1680 if (LHSZero.getBoolValue()) {
1681 DAG.computeKnownBits(N1, RHSZero, RHSOne);
1683 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1684 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1685 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero)
1686 return CombineTo(N, DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1),
1687 DAG.getNode(ISD::CARRY_FALSE,
1688 SDLoc(N), MVT::Glue));
1694 SDValue DAGCombiner::visitADDE(SDNode *N) {
1695 SDValue N0 = N->getOperand(0);
1696 SDValue N1 = N->getOperand(1);
1697 SDValue CarryIn = N->getOperand(2);
1698 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1699 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1701 // canonicalize constant to RHS
1703 return DAG.getNode(ISD::ADDE, SDLoc(N), N->getVTList(),
1706 // fold (adde x, y, false) -> (addc x, y)
1707 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1708 return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N0, N1);
1713 // Since it may not be valid to emit a fold to zero for vector initializers
1714 // check if we can before folding.
1715 static SDValue tryFoldToZero(SDLoc DL, const TargetLowering &TLI, EVT VT,
1717 bool LegalOperations, bool LegalTypes) {
1719 return DAG.getConstant(0, VT);
1720 if (!LegalOperations || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
1721 return DAG.getConstant(0, VT);
1725 SDValue DAGCombiner::visitSUB(SDNode *N) {
1726 SDValue N0 = N->getOperand(0);
1727 SDValue N1 = N->getOperand(1);
1728 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1729 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1730 ConstantSDNode *N1C1 = N1.getOpcode() != ISD::ADD ? nullptr :
1731 dyn_cast<ConstantSDNode>(N1.getOperand(1).getNode());
1732 EVT VT = N0.getValueType();
1735 if (VT.isVector()) {
1736 SDValue FoldedVOp = SimplifyVBinOp(N);
1737 if (FoldedVOp.getNode()) return FoldedVOp;
1739 // fold (sub x, 0) -> x, vector edition
1740 if (ISD::isBuildVectorAllZeros(N1.getNode()))
1744 // fold (sub x, x) -> 0
1745 // FIXME: Refactor this and xor and other similar operations together.
1747 return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations, LegalTypes);
1748 // fold (sub c1, c2) -> c1-c2
1750 return DAG.FoldConstantArithmetic(ISD::SUB, VT, N0C, N1C);
1751 // fold (sub x, c) -> (add x, -c)
1753 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0,
1754 DAG.getConstant(-N1C->getAPIntValue(), VT));
1755 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1)
1756 if (N0C && N0C->isAllOnesValue())
1757 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0);
1758 // fold A-(A-B) -> B
1759 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(0))
1760 return N1.getOperand(1);
1761 // fold (A+B)-A -> B
1762 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1)
1763 return N0.getOperand(1);
1764 // fold (A+B)-B -> A
1765 if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1)
1766 return N0.getOperand(0);
1767 // fold C2-(A+C1) -> (C2-C1)-A
1768 if (N1.getOpcode() == ISD::ADD && N0C && N1C1) {
1769 SDValue NewC = DAG.getConstant(N0C->getAPIntValue() - N1C1->getAPIntValue(),
1771 return DAG.getNode(ISD::SUB, SDLoc(N), VT, NewC,
1774 // fold ((A+(B+or-C))-B) -> A+or-C
1775 if (N0.getOpcode() == ISD::ADD &&
1776 (N0.getOperand(1).getOpcode() == ISD::SUB ||
1777 N0.getOperand(1).getOpcode() == ISD::ADD) &&
1778 N0.getOperand(1).getOperand(0) == N1)
1779 return DAG.getNode(N0.getOperand(1).getOpcode(), SDLoc(N), VT,
1780 N0.getOperand(0), N0.getOperand(1).getOperand(1));
1781 // fold ((A+(C+B))-B) -> A+C
1782 if (N0.getOpcode() == ISD::ADD &&
1783 N0.getOperand(1).getOpcode() == ISD::ADD &&
1784 N0.getOperand(1).getOperand(1) == N1)
1785 return DAG.getNode(ISD::ADD, SDLoc(N), VT,
1786 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1787 // fold ((A-(B-C))-C) -> A-B
1788 if (N0.getOpcode() == ISD::SUB &&
1789 N0.getOperand(1).getOpcode() == ISD::SUB &&
1790 N0.getOperand(1).getOperand(1) == N1)
1791 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1792 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1794 // If either operand of a sub is undef, the result is undef
1795 if (N0.getOpcode() == ISD::UNDEF)
1797 if (N1.getOpcode() == ISD::UNDEF)
1800 // If the relocation model supports it, consider symbol offsets.
1801 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1802 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA)) {
1803 // fold (sub Sym, c) -> Sym-c
1804 if (N1C && GA->getOpcode() == ISD::GlobalAddress)
1805 return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT,
1807 (uint64_t)N1C->getSExtValue());
1808 // fold (sub Sym+c1, Sym+c2) -> c1-c2
1809 if (GlobalAddressSDNode *GB = dyn_cast<GlobalAddressSDNode>(N1))
1810 if (GA->getGlobal() == GB->getGlobal())
1811 return DAG.getConstant((uint64_t)GA->getOffset() - GB->getOffset(),
1818 SDValue DAGCombiner::visitSUBC(SDNode *N) {
1819 SDValue N0 = N->getOperand(0);
1820 SDValue N1 = N->getOperand(1);
1821 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1822 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1823 EVT VT = N0.getValueType();
1825 // If the flag result is dead, turn this into an SUB.
1826 if (!N->hasAnyUseOfValue(1))
1827 return CombineTo(N, DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, N1),
1828 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1831 // fold (subc x, x) -> 0 + no borrow
1833 return CombineTo(N, DAG.getConstant(0, VT),
1834 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1837 // fold (subc x, 0) -> x + no borrow
1838 if (N1C && N1C->isNullValue())
1839 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1842 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1) + no borrow
1843 if (N0C && N0C->isAllOnesValue())
1844 return CombineTo(N, DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0),
1845 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1851 SDValue DAGCombiner::visitSUBE(SDNode *N) {
1852 SDValue N0 = N->getOperand(0);
1853 SDValue N1 = N->getOperand(1);
1854 SDValue CarryIn = N->getOperand(2);
1856 // fold (sube x, y, false) -> (subc x, y)
1857 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1858 return DAG.getNode(ISD::SUBC, SDLoc(N), N->getVTList(), N0, N1);
1863 SDValue DAGCombiner::visitMUL(SDNode *N) {
1864 SDValue N0 = N->getOperand(0);
1865 SDValue N1 = N->getOperand(1);
1866 EVT VT = N0.getValueType();
1868 // fold (mul x, undef) -> 0
1869 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
1870 return DAG.getConstant(0, VT);
1872 bool N0IsConst = false;
1873 bool N1IsConst = false;
1874 APInt ConstValue0, ConstValue1;
1876 if (VT.isVector()) {
1877 SDValue FoldedVOp = SimplifyVBinOp(N);
1878 if (FoldedVOp.getNode()) return FoldedVOp;
1880 N0IsConst = isConstantSplatVector(N0.getNode(), ConstValue0);
1881 N1IsConst = isConstantSplatVector(N1.getNode(), ConstValue1);
1883 N0IsConst = dyn_cast<ConstantSDNode>(N0) != nullptr;
1884 ConstValue0 = N0IsConst ? (dyn_cast<ConstantSDNode>(N0))->getAPIntValue()
1886 N1IsConst = dyn_cast<ConstantSDNode>(N1) != nullptr;
1887 ConstValue1 = N1IsConst ? (dyn_cast<ConstantSDNode>(N1))->getAPIntValue()
1891 // fold (mul c1, c2) -> c1*c2
1892 if (N0IsConst && N1IsConst)
1893 return DAG.FoldConstantArithmetic(ISD::MUL, VT, N0.getNode(), N1.getNode());
1895 // canonicalize constant to RHS
1896 if (N0IsConst && !N1IsConst)
1897 return DAG.getNode(ISD::MUL, SDLoc(N), VT, N1, N0);
1898 // fold (mul x, 0) -> 0
1899 if (N1IsConst && ConstValue1 == 0)
1901 // We require a splat of the entire scalar bit width for non-contiguous
1904 ConstValue1.getBitWidth() == VT.getScalarType().getSizeInBits();
1905 // fold (mul x, 1) -> x
1906 if (N1IsConst && ConstValue1 == 1 && IsFullSplat)
1908 // fold (mul x, -1) -> 0-x
1909 if (N1IsConst && ConstValue1.isAllOnesValue())
1910 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1911 DAG.getConstant(0, VT), N0);
1912 // fold (mul x, (1 << c)) -> x << c
1913 if (N1IsConst && ConstValue1.isPowerOf2() && IsFullSplat)
1914 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0,
1915 DAG.getConstant(ConstValue1.logBase2(),
1916 getShiftAmountTy(N0.getValueType())));
1917 // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c
1918 if (N1IsConst && (-ConstValue1).isPowerOf2() && IsFullSplat) {
1919 unsigned Log2Val = (-ConstValue1).logBase2();
1920 // FIXME: If the input is something that is easily negated (e.g. a
1921 // single-use add), we should put the negate there.
1922 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1923 DAG.getConstant(0, VT),
1924 DAG.getNode(ISD::SHL, SDLoc(N), VT, N0,
1925 DAG.getConstant(Log2Val,
1926 getShiftAmountTy(N0.getValueType()))));
1930 // (mul (shl X, c1), c2) -> (mul X, c2 << c1)
1931 if (N1IsConst && N0.getOpcode() == ISD::SHL &&
1932 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
1933 isa<ConstantSDNode>(N0.getOperand(1)))) {
1934 SDValue C3 = DAG.getNode(ISD::SHL, SDLoc(N), VT,
1935 N1, N0.getOperand(1));
1936 AddToWorkList(C3.getNode());
1937 return DAG.getNode(ISD::MUL, SDLoc(N), VT,
1938 N0.getOperand(0), C3);
1941 // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one
1944 SDValue Sh(nullptr,0), Y(nullptr,0);
1945 // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)).
1946 if (N0.getOpcode() == ISD::SHL &&
1947 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
1948 isa<ConstantSDNode>(N0.getOperand(1))) &&
1949 N0.getNode()->hasOneUse()) {
1951 } else if (N1.getOpcode() == ISD::SHL &&
1952 isa<ConstantSDNode>(N1.getOperand(1)) &&
1953 N1.getNode()->hasOneUse()) {
1958 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
1959 Sh.getOperand(0), Y);
1960 return DAG.getNode(ISD::SHL, SDLoc(N), VT,
1961 Mul, Sh.getOperand(1));
1965 // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2)
1966 if (N1IsConst && N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() &&
1967 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
1968 isa<ConstantSDNode>(N0.getOperand(1))))
1969 return DAG.getNode(ISD::ADD, SDLoc(N), VT,
1970 DAG.getNode(ISD::MUL, SDLoc(N0), VT,
1971 N0.getOperand(0), N1),
1972 DAG.getNode(ISD::MUL, SDLoc(N1), VT,
1973 N0.getOperand(1), N1));
1976 SDValue RMUL = ReassociateOps(ISD::MUL, SDLoc(N), N0, N1);
1983 SDValue DAGCombiner::visitSDIV(SDNode *N) {
1984 SDValue N0 = N->getOperand(0);
1985 SDValue N1 = N->getOperand(1);
1986 ConstantSDNode *N0C = isConstOrConstSplat(N0);
1987 ConstantSDNode *N1C = isConstOrConstSplat(N1);
1988 EVT VT = N->getValueType(0);
1991 if (VT.isVector()) {
1992 SDValue FoldedVOp = SimplifyVBinOp(N);
1993 if (FoldedVOp.getNode()) return FoldedVOp;
1996 // fold (sdiv c1, c2) -> c1/c2
1997 if (N0C && N1C && !N1C->isNullValue())
1998 return DAG.FoldConstantArithmetic(ISD::SDIV, VT, N0C, N1C);
1999 // fold (sdiv X, 1) -> X
2000 if (N1C && N1C->getAPIntValue() == 1LL)
2002 // fold (sdiv X, -1) -> 0-X
2003 if (N1C && N1C->isAllOnesValue())
2004 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
2005 DAG.getConstant(0, VT), N0);
2006 // If we know the sign bits of both operands are zero, strength reduce to a
2007 // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2
2008 if (!VT.isVector()) {
2009 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
2010 return DAG.getNode(ISD::UDIV, SDLoc(N), N1.getValueType(),
2014 // fold (sdiv X, pow2) -> simple ops after legalize
2015 if (N1C && !N1C->isNullValue() && (N1C->getAPIntValue().isPowerOf2() ||
2016 (-N1C->getAPIntValue()).isPowerOf2())) {
2017 // If dividing by powers of two is cheap, then don't perform the following
2019 if (TLI.isPow2DivCheap())
2022 unsigned lg2 = N1C->getAPIntValue().countTrailingZeros();
2024 // Splat the sign bit into the register
2026 DAG.getNode(ISD::SRA, SDLoc(N), VT, N0,
2027 DAG.getConstant(VT.getScalarSizeInBits() - 1,
2028 getShiftAmountTy(N0.getValueType())));
2029 AddToWorkList(SGN.getNode());
2031 // Add (N0 < 0) ? abs2 - 1 : 0;
2033 DAG.getNode(ISD::SRL, SDLoc(N), VT, SGN,
2034 DAG.getConstant(VT.getScalarSizeInBits() - lg2,
2035 getShiftAmountTy(SGN.getValueType())));
2036 SDValue ADD = DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, SRL);
2037 AddToWorkList(SRL.getNode());
2038 AddToWorkList(ADD.getNode()); // Divide by pow2
2039 SDValue SRA = DAG.getNode(ISD::SRA, SDLoc(N), VT, ADD,
2040 DAG.getConstant(lg2, getShiftAmountTy(ADD.getValueType())));
2042 // If we're dividing by a positive value, we're done. Otherwise, we must
2043 // negate the result.
2044 if (N1C->getAPIntValue().isNonNegative())
2047 AddToWorkList(SRA.getNode());
2048 return DAG.getNode(ISD::SUB, SDLoc(N), VT, DAG.getConstant(0, VT), SRA);
2051 // if integer divide is expensive and we satisfy the requirements, emit an
2052 // alternate sequence.
2053 if (N1C && !TLI.isIntDivCheap()) {
2054 SDValue Op = BuildSDIV(N);
2055 if (Op.getNode()) return Op;
2059 if (N0.getOpcode() == ISD::UNDEF)
2060 return DAG.getConstant(0, VT);
2061 // X / undef -> undef
2062 if (N1.getOpcode() == ISD::UNDEF)
2068 SDValue DAGCombiner::visitUDIV(SDNode *N) {
2069 SDValue N0 = N->getOperand(0);
2070 SDValue N1 = N->getOperand(1);
2071 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2072 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2073 EVT VT = N->getValueType(0);
2076 if (VT.isVector()) {
2077 SDValue FoldedVOp = SimplifyVBinOp(N);
2078 if (FoldedVOp.getNode()) return FoldedVOp;
2081 // fold (udiv c1, c2) -> c1/c2
2082 if (N0C && N1C && !N1C->isNullValue())
2083 return DAG.FoldConstantArithmetic(ISD::UDIV, VT, N0C, N1C);
2084 // fold (udiv x, (1 << c)) -> x >>u c
2085 if (N1C && N1C->getAPIntValue().isPowerOf2())
2086 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0,
2087 DAG.getConstant(N1C->getAPIntValue().logBase2(),
2088 getShiftAmountTy(N0.getValueType())));
2089 // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2
2090 if (N1.getOpcode() == ISD::SHL) {
2091 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
2092 if (SHC->getAPIntValue().isPowerOf2()) {
2093 EVT ADDVT = N1.getOperand(1).getValueType();
2094 SDValue Add = DAG.getNode(ISD::ADD, SDLoc(N), ADDVT,
2096 DAG.getConstant(SHC->getAPIntValue()
2099 AddToWorkList(Add.getNode());
2100 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, Add);
2104 // fold (udiv x, c) -> alternate
2105 if (N1C && !TLI.isIntDivCheap()) {
2106 SDValue Op = BuildUDIV(N);
2107 if (Op.getNode()) return Op;
2111 if (N0.getOpcode() == ISD::UNDEF)
2112 return DAG.getConstant(0, VT);
2113 // X / undef -> undef
2114 if (N1.getOpcode() == ISD::UNDEF)
2120 SDValue DAGCombiner::visitSREM(SDNode *N) {
2121 SDValue N0 = N->getOperand(0);
2122 SDValue N1 = N->getOperand(1);
2123 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2124 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2125 EVT VT = N->getValueType(0);
2127 // fold (srem c1, c2) -> c1%c2
2128 if (N0C && N1C && !N1C->isNullValue())
2129 return DAG.FoldConstantArithmetic(ISD::SREM, VT, N0C, N1C);
2130 // If we know the sign bits of both operands are zero, strength reduce to a
2131 // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15
2132 if (!VT.isVector()) {
2133 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
2134 return DAG.getNode(ISD::UREM, SDLoc(N), VT, N0, N1);
2137 // If X/C can be simplified by the division-by-constant logic, lower
2138 // X%C to the equivalent of X-X/C*C.
2139 if (N1C && !N1C->isNullValue()) {
2140 SDValue Div = DAG.getNode(ISD::SDIV, SDLoc(N), VT, N0, N1);
2141 AddToWorkList(Div.getNode());
2142 SDValue OptimizedDiv = combine(Div.getNode());
2143 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2144 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2146 SDValue Sub = DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, Mul);
2147 AddToWorkList(Mul.getNode());
2153 if (N0.getOpcode() == ISD::UNDEF)
2154 return DAG.getConstant(0, VT);
2155 // X % undef -> undef
2156 if (N1.getOpcode() == ISD::UNDEF)
2162 SDValue DAGCombiner::visitUREM(SDNode *N) {
2163 SDValue N0 = N->getOperand(0);
2164 SDValue N1 = N->getOperand(1);
2165 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2166 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2167 EVT VT = N->getValueType(0);
2169 // fold (urem c1, c2) -> c1%c2
2170 if (N0C && N1C && !N1C->isNullValue())
2171 return DAG.FoldConstantArithmetic(ISD::UREM, VT, N0C, N1C);
2172 // fold (urem x, pow2) -> (and x, pow2-1)
2173 if (N1C && !N1C->isNullValue() && N1C->getAPIntValue().isPowerOf2())
2174 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0,
2175 DAG.getConstant(N1C->getAPIntValue()-1,VT));
2176 // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1))
2177 if (N1.getOpcode() == ISD::SHL) {
2178 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
2179 if (SHC->getAPIntValue().isPowerOf2()) {
2181 DAG.getNode(ISD::ADD, SDLoc(N), VT, N1,
2182 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()),
2184 AddToWorkList(Add.getNode());
2185 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, Add);
2190 // If X/C can be simplified by the division-by-constant logic, lower
2191 // X%C to the equivalent of X-X/C*C.
2192 if (N1C && !N1C->isNullValue()) {
2193 SDValue Div = DAG.getNode(ISD::UDIV, SDLoc(N), VT, N0, N1);
2194 AddToWorkList(Div.getNode());
2195 SDValue OptimizedDiv = combine(Div.getNode());
2196 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2197 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2199 SDValue Sub = DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, Mul);
2200 AddToWorkList(Mul.getNode());
2206 if (N0.getOpcode() == ISD::UNDEF)
2207 return DAG.getConstant(0, VT);
2208 // X % undef -> undef
2209 if (N1.getOpcode() == ISD::UNDEF)
2215 SDValue DAGCombiner::visitMULHS(SDNode *N) {
2216 SDValue N0 = N->getOperand(0);
2217 SDValue N1 = N->getOperand(1);
2218 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2219 EVT VT = N->getValueType(0);
2222 // fold (mulhs x, 0) -> 0
2223 if (N1C && N1C->isNullValue())
2225 // fold (mulhs x, 1) -> (sra x, size(x)-1)
2226 if (N1C && N1C->getAPIntValue() == 1)
2227 return DAG.getNode(ISD::SRA, SDLoc(N), N0.getValueType(), N0,
2228 DAG.getConstant(N0.getValueType().getSizeInBits() - 1,
2229 getShiftAmountTy(N0.getValueType())));
2230 // fold (mulhs x, undef) -> 0
2231 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2232 return DAG.getConstant(0, VT);
2234 // If the type twice as wide is legal, transform the mulhs to a wider multiply
2236 if (VT.isSimple() && !VT.isVector()) {
2237 MVT Simple = VT.getSimpleVT();
2238 unsigned SimpleSize = Simple.getSizeInBits();
2239 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2240 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2241 N0 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N0);
2242 N1 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N1);
2243 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2244 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2245 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2246 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2253 SDValue DAGCombiner::visitMULHU(SDNode *N) {
2254 SDValue N0 = N->getOperand(0);
2255 SDValue N1 = N->getOperand(1);
2256 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2257 EVT VT = N->getValueType(0);
2260 // fold (mulhu x, 0) -> 0
2261 if (N1C && N1C->isNullValue())
2263 // fold (mulhu x, 1) -> 0
2264 if (N1C && N1C->getAPIntValue() == 1)
2265 return DAG.getConstant(0, N0.getValueType());
2266 // fold (mulhu x, undef) -> 0
2267 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2268 return DAG.getConstant(0, VT);
2270 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2272 if (VT.isSimple() && !VT.isVector()) {
2273 MVT Simple = VT.getSimpleVT();
2274 unsigned SimpleSize = Simple.getSizeInBits();
2275 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2276 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2277 N0 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N0);
2278 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N1);
2279 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2280 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2281 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2282 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2289 /// SimplifyNodeWithTwoResults - Perform optimizations common to nodes that
2290 /// compute two values. LoOp and HiOp give the opcodes for the two computations
2291 /// that are being performed. Return true if a simplification was made.
2293 SDValue DAGCombiner::SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
2295 // If the high half is not needed, just compute the low half.
2296 bool HiExists = N->hasAnyUseOfValue(1);
2298 (!LegalOperations ||
2299 TLI.isOperationLegalOrCustom(LoOp, N->getValueType(0)))) {
2300 SDValue Res = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0),
2301 ArrayRef<SDUse>(N->op_begin(), N->op_end()));
2302 return CombineTo(N, Res, Res);
2305 // If the low half is not needed, just compute the high half.
2306 bool LoExists = N->hasAnyUseOfValue(0);
2308 (!LegalOperations ||
2309 TLI.isOperationLegal(HiOp, N->getValueType(1)))) {
2310 SDValue Res = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1),
2311 ArrayRef<SDUse>(N->op_begin(), N->op_end()));
2312 return CombineTo(N, Res, Res);
2315 // If both halves are used, return as it is.
2316 if (LoExists && HiExists)
2319 // If the two computed results can be simplified separately, separate them.
2321 SDValue Lo = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0),
2322 ArrayRef<SDUse>(N->op_begin(), N->op_end()));
2323 AddToWorkList(Lo.getNode());
2324 SDValue LoOpt = combine(Lo.getNode());
2325 if (LoOpt.getNode() && LoOpt.getNode() != Lo.getNode() &&
2326 (!LegalOperations ||
2327 TLI.isOperationLegal(LoOpt.getOpcode(), LoOpt.getValueType())))
2328 return CombineTo(N, LoOpt, LoOpt);
2332 SDValue Hi = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1),
2333 ArrayRef<SDUse>(N->op_begin(), N->op_end()));
2334 AddToWorkList(Hi.getNode());
2335 SDValue HiOpt = combine(Hi.getNode());
2336 if (HiOpt.getNode() && HiOpt != Hi &&
2337 (!LegalOperations ||
2338 TLI.isOperationLegal(HiOpt.getOpcode(), HiOpt.getValueType())))
2339 return CombineTo(N, HiOpt, HiOpt);
2345 SDValue DAGCombiner::visitSMUL_LOHI(SDNode *N) {
2346 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHS);
2347 if (Res.getNode()) return Res;
2349 EVT VT = N->getValueType(0);
2352 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2354 if (VT.isSimple() && !VT.isVector()) {
2355 MVT Simple = VT.getSimpleVT();
2356 unsigned SimpleSize = Simple.getSizeInBits();
2357 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2358 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2359 SDValue Lo = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(0));
2360 SDValue Hi = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(1));
2361 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2362 // Compute the high part as N1.
2363 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2364 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2365 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2366 // Compute the low part as N0.
2367 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2368 return CombineTo(N, Lo, Hi);
2375 SDValue DAGCombiner::visitUMUL_LOHI(SDNode *N) {
2376 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHU);
2377 if (Res.getNode()) return Res;
2379 EVT VT = N->getValueType(0);
2382 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2384 if (VT.isSimple() && !VT.isVector()) {
2385 MVT Simple = VT.getSimpleVT();
2386 unsigned SimpleSize = Simple.getSizeInBits();
2387 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2388 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2389 SDValue Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(0));
2390 SDValue Hi = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(1));
2391 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2392 // Compute the high part as N1.
2393 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2394 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2395 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2396 // Compute the low part as N0.
2397 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2398 return CombineTo(N, Lo, Hi);
2405 SDValue DAGCombiner::visitSMULO(SDNode *N) {
2406 // (smulo x, 2) -> (saddo x, x)
2407 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2408 if (C2->getAPIntValue() == 2)
2409 return DAG.getNode(ISD::SADDO, SDLoc(N), N->getVTList(),
2410 N->getOperand(0), N->getOperand(0));
2415 SDValue DAGCombiner::visitUMULO(SDNode *N) {
2416 // (umulo x, 2) -> (uaddo x, x)
2417 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2418 if (C2->getAPIntValue() == 2)
2419 return DAG.getNode(ISD::UADDO, SDLoc(N), N->getVTList(),
2420 N->getOperand(0), N->getOperand(0));
2425 SDValue DAGCombiner::visitSDIVREM(SDNode *N) {
2426 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::SDIV, ISD::SREM);
2427 if (Res.getNode()) return Res;
2432 SDValue DAGCombiner::visitUDIVREM(SDNode *N) {
2433 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::UDIV, ISD::UREM);
2434 if (Res.getNode()) return Res;
2439 /// SimplifyBinOpWithSameOpcodeHands - If this is a binary operator with
2440 /// two operands of the same opcode, try to simplify it.
2441 SDValue DAGCombiner::SimplifyBinOpWithSameOpcodeHands(SDNode *N) {
2442 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
2443 EVT VT = N0.getValueType();
2444 assert(N0.getOpcode() == N1.getOpcode() && "Bad input!");
2446 // Bail early if none of these transforms apply.
2447 if (N0.getNode()->getNumOperands() == 0) return SDValue();
2449 // For each of OP in AND/OR/XOR:
2450 // fold (OP (zext x), (zext y)) -> (zext (OP x, y))
2451 // fold (OP (sext x), (sext y)) -> (sext (OP x, y))
2452 // fold (OP (aext x), (aext y)) -> (aext (OP x, y))
2453 // fold (OP (trunc x), (trunc y)) -> (trunc (OP x, y)) (if trunc isn't free)
2455 // do not sink logical op inside of a vector extend, since it may combine
2457 EVT Op0VT = N0.getOperand(0).getValueType();
2458 if ((N0.getOpcode() == ISD::ZERO_EXTEND ||
2459 N0.getOpcode() == ISD::SIGN_EXTEND ||
2460 // Avoid infinite looping with PromoteIntBinOp.
2461 (N0.getOpcode() == ISD::ANY_EXTEND &&
2462 (!LegalTypes || TLI.isTypeDesirableForOp(N->getOpcode(), Op0VT))) ||
2463 (N0.getOpcode() == ISD::TRUNCATE &&
2464 (!TLI.isZExtFree(VT, Op0VT) ||
2465 !TLI.isTruncateFree(Op0VT, VT)) &&
2466 TLI.isTypeLegal(Op0VT))) &&
2468 Op0VT == N1.getOperand(0).getValueType() &&
2469 (!LegalOperations || TLI.isOperationLegal(N->getOpcode(), Op0VT))) {
2470 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0),
2471 N0.getOperand(0).getValueType(),
2472 N0.getOperand(0), N1.getOperand(0));
2473 AddToWorkList(ORNode.getNode());
2474 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, ORNode);
2477 // For each of OP in SHL/SRL/SRA/AND...
2478 // fold (and (OP x, z), (OP y, z)) -> (OP (and x, y), z)
2479 // fold (or (OP x, z), (OP y, z)) -> (OP (or x, y), z)
2480 // fold (xor (OP x, z), (OP y, z)) -> (OP (xor x, y), z)
2481 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL ||
2482 N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::AND) &&
2483 N0.getOperand(1) == N1.getOperand(1)) {
2484 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0),
2485 N0.getOperand(0).getValueType(),
2486 N0.getOperand(0), N1.getOperand(0));
2487 AddToWorkList(ORNode.getNode());
2488 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
2489 ORNode, N0.getOperand(1));
2492 // Simplify xor/and/or (bitcast(A), bitcast(B)) -> bitcast(op (A,B))
2493 // Only perform this optimization after type legalization and before
2494 // LegalizeVectorOprs. LegalizeVectorOprs promotes vector operations by
2495 // adding bitcasts. For example (xor v4i32) is promoted to (v2i64), and
2496 // we don't want to undo this promotion.
2497 // We also handle SCALAR_TO_VECTOR because xor/or/and operations are cheaper
2499 if ((N0.getOpcode() == ISD::BITCAST ||
2500 N0.getOpcode() == ISD::SCALAR_TO_VECTOR) &&
2501 Level == AfterLegalizeTypes) {
2502 SDValue In0 = N0.getOperand(0);
2503 SDValue In1 = N1.getOperand(0);
2504 EVT In0Ty = In0.getValueType();
2505 EVT In1Ty = In1.getValueType();
2507 // If both incoming values are integers, and the original types are the
2509 if (In0Ty.isInteger() && In1Ty.isInteger() && In0Ty == In1Ty) {
2510 SDValue Op = DAG.getNode(N->getOpcode(), DL, In0Ty, In0, In1);
2511 SDValue BC = DAG.getNode(N0.getOpcode(), DL, VT, Op);
2512 AddToWorkList(Op.getNode());
2517 // Xor/and/or are indifferent to the swizzle operation (shuffle of one value).
2518 // Simplify xor/and/or (shuff(A), shuff(B)) -> shuff(op (A,B))
2519 // If both shuffles use the same mask, and both shuffle within a single
2520 // vector, then it is worthwhile to move the swizzle after the operation.
2521 // The type-legalizer generates this pattern when loading illegal
2522 // vector types from memory. In many cases this allows additional shuffle
2524 // There are other cases where moving the shuffle after the xor/and/or
2525 // is profitable even if shuffles don't perform a swizzle.
2526 // If both shuffles use the same mask, and both shuffles have the same first
2527 // or second operand, then it might still be profitable to move the shuffle
2528 // after the xor/and/or operation.
2529 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG) {
2530 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(N0);
2531 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(N1);
2533 assert(N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType() &&
2534 "Inputs to shuffles are not the same type");
2536 // Check that both shuffles use the same mask. The masks are known to be of
2537 // the same length because the result vector type is the same.
2538 // Check also that shuffles have only one use to avoid introducing extra
2540 if (SVN0->hasOneUse() && SVN1->hasOneUse() &&
2541 SVN0->getMask().equals(SVN1->getMask())) {
2542 SDValue ShOp = N0->getOperand(1);
2544 // Don't try to fold this node if it requires introducing a
2545 // build vector of all zeros that might be illegal at this stage.
2546 if (N->getOpcode() == ISD::XOR && ShOp.getOpcode() != ISD::UNDEF) {
2548 ShOp = DAG.getConstant(0, VT);
2553 // (AND (shuf (A, C), shuf (B, C)) -> shuf (AND (A, B), C)
2554 // (OR (shuf (A, C), shuf (B, C)) -> shuf (OR (A, B), C)
2555 // (XOR (shuf (A, C), shuf (B, C)) -> shuf (XOR (A, B), V_0)
2556 if (N0.getOperand(1) == N1.getOperand(1) && ShOp.getNode()) {
2557 SDValue NewNode = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
2558 N0->getOperand(0), N1->getOperand(0));
2559 AddToWorkList(NewNode.getNode());
2560 return DAG.getVectorShuffle(VT, SDLoc(N), NewNode, ShOp,
2561 &SVN0->getMask()[0]);
2564 // Don't try to fold this node if it requires introducing a
2565 // build vector of all zeros that might be illegal at this stage.
2566 ShOp = N0->getOperand(0);
2567 if (N->getOpcode() == ISD::XOR && ShOp.getOpcode() != ISD::UNDEF) {
2569 ShOp = DAG.getConstant(0, VT);
2574 // (AND (shuf (C, A), shuf (C, B)) -> shuf (C, AND (A, B))
2575 // (OR (shuf (C, A), shuf (C, B)) -> shuf (C, OR (A, B))
2576 // (XOR (shuf (C, A), shuf (C, B)) -> shuf (V_0, XOR (A, B))
2577 if (N0->getOperand(0) == N1->getOperand(0) && ShOp.getNode()) {
2578 SDValue NewNode = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
2579 N0->getOperand(1), N1->getOperand(1));
2580 AddToWorkList(NewNode.getNode());
2581 return DAG.getVectorShuffle(VT, SDLoc(N), ShOp, NewNode,
2582 &SVN0->getMask()[0]);
2590 SDValue DAGCombiner::visitAND(SDNode *N) {
2591 SDValue N0 = N->getOperand(0);
2592 SDValue N1 = N->getOperand(1);
2593 SDValue LL, LR, RL, RR, CC0, CC1;
2594 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2595 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2596 EVT VT = N1.getValueType();
2597 unsigned BitWidth = VT.getScalarType().getSizeInBits();
2600 if (VT.isVector()) {
2601 SDValue FoldedVOp = SimplifyVBinOp(N);
2602 if (FoldedVOp.getNode()) return FoldedVOp;
2604 // fold (and x, 0) -> 0, vector edition
2605 if (ISD::isBuildVectorAllZeros(N0.getNode()))
2607 if (ISD::isBuildVectorAllZeros(N1.getNode()))
2610 // fold (and x, -1) -> x, vector edition
2611 if (ISD::isBuildVectorAllOnes(N0.getNode()))
2613 if (ISD::isBuildVectorAllOnes(N1.getNode()))
2617 // fold (and x, undef) -> 0
2618 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2619 return DAG.getConstant(0, VT);
2620 // fold (and c1, c2) -> c1&c2
2622 return DAG.FoldConstantArithmetic(ISD::AND, VT, N0C, N1C);
2623 // canonicalize constant to RHS
2625 return DAG.getNode(ISD::AND, SDLoc(N), VT, N1, N0);
2626 // fold (and x, -1) -> x
2627 if (N1C && N1C->isAllOnesValue())
2629 // if (and x, c) is known to be zero, return 0
2630 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
2631 APInt::getAllOnesValue(BitWidth)))
2632 return DAG.getConstant(0, VT);
2634 SDValue RAND = ReassociateOps(ISD::AND, SDLoc(N), N0, N1);
2637 // fold (and (or x, C), D) -> D if (C & D) == D
2638 if (N1C && N0.getOpcode() == ISD::OR)
2639 if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
2640 if ((ORI->getAPIntValue() & N1C->getAPIntValue()) == N1C->getAPIntValue())
2642 // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits.
2643 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
2644 SDValue N0Op0 = N0.getOperand(0);
2645 APInt Mask = ~N1C->getAPIntValue();
2646 Mask = Mask.trunc(N0Op0.getValueSizeInBits());
2647 if (DAG.MaskedValueIsZero(N0Op0, Mask)) {
2648 SDValue Zext = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N),
2649 N0.getValueType(), N0Op0);
2651 // Replace uses of the AND with uses of the Zero extend node.
2654 // We actually want to replace all uses of the any_extend with the
2655 // zero_extend, to avoid duplicating things. This will later cause this
2656 // AND to be folded.
2657 CombineTo(N0.getNode(), Zext);
2658 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2661 // similarly fold (and (X (load ([non_ext|any_ext|zero_ext] V))), c) ->
2662 // (X (load ([non_ext|zero_ext] V))) if 'and' only clears top bits which must
2663 // already be zero by virtue of the width of the base type of the load.
2665 // the 'X' node here can either be nothing or an extract_vector_elt to catch
2667 if ((N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
2668 N0.getOperand(0).getOpcode() == ISD::LOAD) ||
2669 N0.getOpcode() == ISD::LOAD) {
2670 LoadSDNode *Load = cast<LoadSDNode>( (N0.getOpcode() == ISD::LOAD) ?
2671 N0 : N0.getOperand(0) );
2673 // Get the constant (if applicable) the zero'th operand is being ANDed with.
2674 // This can be a pure constant or a vector splat, in which case we treat the
2675 // vector as a scalar and use the splat value.
2676 APInt Constant = APInt::getNullValue(1);
2677 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
2678 Constant = C->getAPIntValue();
2679 } else if (BuildVectorSDNode *Vector = dyn_cast<BuildVectorSDNode>(N1)) {
2680 APInt SplatValue, SplatUndef;
2681 unsigned SplatBitSize;
2683 bool IsSplat = Vector->isConstantSplat(SplatValue, SplatUndef,
2684 SplatBitSize, HasAnyUndefs);
2686 // Undef bits can contribute to a possible optimisation if set, so
2688 SplatValue |= SplatUndef;
2690 // The splat value may be something like "0x00FFFFFF", which means 0 for
2691 // the first vector value and FF for the rest, repeating. We need a mask
2692 // that will apply equally to all members of the vector, so AND all the
2693 // lanes of the constant together.
2694 EVT VT = Vector->getValueType(0);
2695 unsigned BitWidth = VT.getVectorElementType().getSizeInBits();
2697 // If the splat value has been compressed to a bitlength lower
2698 // than the size of the vector lane, we need to re-expand it to
2700 if (BitWidth > SplatBitSize)
2701 for (SplatValue = SplatValue.zextOrTrunc(BitWidth);
2702 SplatBitSize < BitWidth;
2703 SplatBitSize = SplatBitSize * 2)
2704 SplatValue |= SplatValue.shl(SplatBitSize);
2706 Constant = APInt::getAllOnesValue(BitWidth);
2707 for (unsigned i = 0, n = SplatBitSize/BitWidth; i < n; ++i)
2708 Constant &= SplatValue.lshr(i*BitWidth).zextOrTrunc(BitWidth);
2712 // If we want to change an EXTLOAD to a ZEXTLOAD, ensure a ZEXTLOAD is
2713 // actually legal and isn't going to get expanded, else this is a false
2715 bool CanZextLoadProfitably = TLI.isLoadExtLegal(ISD::ZEXTLOAD,
2716 Load->getMemoryVT());
2718 // Resize the constant to the same size as the original memory access before
2719 // extension. If it is still the AllOnesValue then this AND is completely
2722 Constant.zextOrTrunc(Load->getMemoryVT().getScalarType().getSizeInBits());
2725 switch (Load->getExtensionType()) {
2726 default: B = false; break;
2727 case ISD::EXTLOAD: B = CanZextLoadProfitably; break;
2729 case ISD::NON_EXTLOAD: B = true; break;
2732 if (B && Constant.isAllOnesValue()) {
2733 // If the load type was an EXTLOAD, convert to ZEXTLOAD in order to
2734 // preserve semantics once we get rid of the AND.
2735 SDValue NewLoad(Load, 0);
2736 if (Load->getExtensionType() == ISD::EXTLOAD) {
2737 NewLoad = DAG.getLoad(Load->getAddressingMode(), ISD::ZEXTLOAD,
2738 Load->getValueType(0), SDLoc(Load),
2739 Load->getChain(), Load->getBasePtr(),
2740 Load->getOffset(), Load->getMemoryVT(),
2741 Load->getMemOperand());
2742 // Replace uses of the EXTLOAD with the new ZEXTLOAD.
2743 if (Load->getNumValues() == 3) {
2744 // PRE/POST_INC loads have 3 values.
2745 SDValue To[] = { NewLoad.getValue(0), NewLoad.getValue(1),
2746 NewLoad.getValue(2) };
2747 CombineTo(Load, To, 3, true);
2749 CombineTo(Load, NewLoad.getValue(0), NewLoad.getValue(1));
2753 // Fold the AND away, taking care not to fold to the old load node if we
2755 CombineTo(N, (N0.getNode() == Load) ? NewLoad : N0);
2757 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2760 // fold (and (setcc x), (setcc y)) -> (setcc (and x, y))
2761 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
2762 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
2763 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
2765 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
2766 LL.getValueType().isInteger()) {
2767 // fold (and (seteq X, 0), (seteq Y, 0)) -> (seteq (or X, Y), 0)
2768 if (cast<ConstantSDNode>(LR)->isNullValue() && Op1 == ISD::SETEQ) {
2769 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(N0),
2770 LR.getValueType(), LL, RL);
2771 AddToWorkList(ORNode.getNode());
2772 return DAG.getSetCC(SDLoc(N), VT, ORNode, LR, Op1);
2774 // fold (and (seteq X, -1), (seteq Y, -1)) -> (seteq (and X, Y), -1)
2775 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETEQ) {
2776 SDValue ANDNode = DAG.getNode(ISD::AND, SDLoc(N0),
2777 LR.getValueType(), LL, RL);
2778 AddToWorkList(ANDNode.getNode());
2779 return DAG.getSetCC(SDLoc(N), VT, ANDNode, LR, Op1);
2781 // fold (and (setgt X, -1), (setgt Y, -1)) -> (setgt (or X, Y), -1)
2782 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETGT) {
2783 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(N0),
2784 LR.getValueType(), LL, RL);
2785 AddToWorkList(ORNode.getNode());
2786 return DAG.getSetCC(SDLoc(N), VT, ORNode, LR, Op1);
2789 // Simplify (and (setne X, 0), (setne X, -1)) -> (setuge (add X, 1), 2)
2790 if (LL == RL && isa<ConstantSDNode>(LR) && isa<ConstantSDNode>(RR) &&
2791 Op0 == Op1 && LL.getValueType().isInteger() &&
2792 Op0 == ISD::SETNE && ((cast<ConstantSDNode>(LR)->isNullValue() &&
2793 cast<ConstantSDNode>(RR)->isAllOnesValue()) ||
2794 (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
2795 cast<ConstantSDNode>(RR)->isNullValue()))) {
2796 SDValue ADDNode = DAG.getNode(ISD::ADD, SDLoc(N0), LL.getValueType(),
2797 LL, DAG.getConstant(1, LL.getValueType()));
2798 AddToWorkList(ADDNode.getNode());
2799 return DAG.getSetCC(SDLoc(N), VT, ADDNode,
2800 DAG.getConstant(2, LL.getValueType()), ISD::SETUGE);
2802 // canonicalize equivalent to ll == rl
2803 if (LL == RR && LR == RL) {
2804 Op1 = ISD::getSetCCSwappedOperands(Op1);
2807 if (LL == RL && LR == RR) {
2808 bool isInteger = LL.getValueType().isInteger();
2809 ISD::CondCode Result = ISD::getSetCCAndOperation(Op0, Op1, isInteger);
2810 if (Result != ISD::SETCC_INVALID &&
2811 (!LegalOperations ||
2812 (TLI.isCondCodeLegal(Result, LL.getSimpleValueType()) &&
2813 TLI.isOperationLegal(ISD::SETCC,
2814 getSetCCResultType(N0.getSimpleValueType())))))
2815 return DAG.getSetCC(SDLoc(N), N0.getValueType(),
2820 // Simplify: (and (op x...), (op y...)) -> (op (and x, y))
2821 if (N0.getOpcode() == N1.getOpcode()) {
2822 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
2823 if (Tmp.getNode()) return Tmp;
2826 // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1)
2827 // fold (and (sra)) -> (and (srl)) when possible.
2828 if (!VT.isVector() &&
2829 SimplifyDemandedBits(SDValue(N, 0)))
2830 return SDValue(N, 0);
2832 // fold (zext_inreg (extload x)) -> (zextload x)
2833 if (ISD::isEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode())) {
2834 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2835 EVT MemVT = LN0->getMemoryVT();
2836 // If we zero all the possible extended bits, then we can turn this into
2837 // a zextload if we are running before legalize or the operation is legal.
2838 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
2839 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
2840 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
2841 ((!LegalOperations && !LN0->isVolatile()) ||
2842 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
2843 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT,
2844 LN0->getChain(), LN0->getBasePtr(),
2845 MemVT, LN0->getMemOperand());
2847 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
2848 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2851 // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use
2852 if (ISD::isSEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
2854 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2855 EVT MemVT = LN0->getMemoryVT();
2856 // If we zero all the possible extended bits, then we can turn this into
2857 // a zextload if we are running before legalize or the operation is legal.
2858 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
2859 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
2860 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
2861 ((!LegalOperations && !LN0->isVolatile()) ||
2862 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
2863 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT,
2864 LN0->getChain(), LN0->getBasePtr(),
2865 MemVT, LN0->getMemOperand());
2867 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
2868 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2872 // fold (and (load x), 255) -> (zextload x, i8)
2873 // fold (and (extload x, i16), 255) -> (zextload x, i8)
2874 // fold (and (any_ext (extload x, i16)), 255) -> (zextload x, i8)
2875 if (N1C && (N0.getOpcode() == ISD::LOAD ||
2876 (N0.getOpcode() == ISD::ANY_EXTEND &&
2877 N0.getOperand(0).getOpcode() == ISD::LOAD))) {
2878 bool HasAnyExt = N0.getOpcode() == ISD::ANY_EXTEND;
2879 LoadSDNode *LN0 = HasAnyExt
2880 ? cast<LoadSDNode>(N0.getOperand(0))
2881 : cast<LoadSDNode>(N0);
2882 if (LN0->getExtensionType() != ISD::SEXTLOAD &&
2883 LN0->isUnindexed() && N0.hasOneUse() && SDValue(LN0, 0).hasOneUse()) {
2884 uint32_t ActiveBits = N1C->getAPIntValue().getActiveBits();
2885 if (ActiveBits > 0 && APIntOps::isMask(ActiveBits, N1C->getAPIntValue())){
2886 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits);
2887 EVT LoadedVT = LN0->getMemoryVT();
2889 if (ExtVT == LoadedVT &&
2890 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) {
2891 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
2894 DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy,
2895 LN0->getChain(), LN0->getBasePtr(), ExtVT,
2896 LN0->getMemOperand());
2898 CombineTo(LN0, NewLoad, NewLoad.getValue(1));
2899 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2902 // Do not change the width of a volatile load.
2903 // Do not generate loads of non-round integer types since these can
2904 // be expensive (and would be wrong if the type is not byte sized).
2905 if (!LN0->isVolatile() && LoadedVT.bitsGT(ExtVT) && ExtVT.isRound() &&
2906 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) {
2907 EVT PtrType = LN0->getOperand(1).getValueType();
2909 unsigned Alignment = LN0->getAlignment();
2910 SDValue NewPtr = LN0->getBasePtr();
2912 // For big endian targets, we need to add an offset to the pointer
2913 // to load the correct bytes. For little endian systems, we merely
2914 // need to read fewer bytes from the same pointer.
2915 if (TLI.isBigEndian()) {
2916 unsigned LVTStoreBytes = LoadedVT.getStoreSize();
2917 unsigned EVTStoreBytes = ExtVT.getStoreSize();
2918 unsigned PtrOff = LVTStoreBytes - EVTStoreBytes;
2919 NewPtr = DAG.getNode(ISD::ADD, SDLoc(LN0), PtrType,
2920 NewPtr, DAG.getConstant(PtrOff, PtrType));
2921 Alignment = MinAlign(Alignment, PtrOff);
2924 AddToWorkList(NewPtr.getNode());
2926 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
2928 DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy,
2929 LN0->getChain(), NewPtr,
2930 LN0->getPointerInfo(),
2931 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
2932 Alignment, LN0->getTBAAInfo());
2934 CombineTo(LN0, Load, Load.getValue(1));
2935 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2941 if (N0.getOpcode() == ISD::ADD && N1.getOpcode() == ISD::SRL &&
2942 VT.getSizeInBits() <= 64) {
2943 if (ConstantSDNode *ADDI = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
2944 APInt ADDC = ADDI->getAPIntValue();
2945 if (!TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
2946 // Look for (and (add x, c1), (lshr y, c2)). If C1 wasn't a legal
2947 // immediate for an add, but it is legal if its top c2 bits are set,
2948 // transform the ADD so the immediate doesn't need to be materialized
2950 if (ConstantSDNode *SRLI = dyn_cast<ConstantSDNode>(N1.getOperand(1))) {
2951 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(),
2952 SRLI->getZExtValue());
2953 if (DAG.MaskedValueIsZero(N0.getOperand(1), Mask)) {
2955 if (TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
2957 DAG.getNode(ISD::ADD, SDLoc(N0), VT,
2958 N0.getOperand(0), DAG.getConstant(ADDC, VT));
2959 CombineTo(N0.getNode(), NewAdd);
2960 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2968 // fold (and (or (srl N, 8), (shl N, 8)), 0xffff) -> (srl (bswap N), const)
2969 if (N1C && N1C->getAPIntValue() == 0xffff && N0.getOpcode() == ISD::OR) {
2970 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
2971 N0.getOperand(1), false);
2972 if (BSwap.getNode())
2979 /// MatchBSwapHWord - Match (a >> 8) | (a << 8) as (bswap a) >> 16
2981 SDValue DAGCombiner::MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
2982 bool DemandHighBits) {
2983 if (!LegalOperations)
2986 EVT VT = N->getValueType(0);
2987 if (VT != MVT::i64 && VT != MVT::i32 && VT != MVT::i16)
2989 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
2992 // Recognize (and (shl a, 8), 0xff), (and (srl a, 8), 0xff00)
2993 bool LookPassAnd0 = false;
2994 bool LookPassAnd1 = false;
2995 if (N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode() == ISD::SRL)
2997 if (N1.getOpcode() == ISD::AND && N1.getOperand(0).getOpcode() == ISD::SHL)
2999 if (N0.getOpcode() == ISD::AND) {
3000 if (!N0.getNode()->hasOneUse())
3002 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3003 if (!N01C || N01C->getZExtValue() != 0xFF00)
3005 N0 = N0.getOperand(0);
3006 LookPassAnd0 = true;
3009 if (N1.getOpcode() == ISD::AND) {
3010 if (!N1.getNode()->hasOneUse())
3012 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
3013 if (!N11C || N11C->getZExtValue() != 0xFF)
3015 N1 = N1.getOperand(0);
3016 LookPassAnd1 = true;
3019 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
3021 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
3023 if (!N0.getNode()->hasOneUse() ||
3024 !N1.getNode()->hasOneUse())
3027 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3028 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
3031 if (N01C->getZExtValue() != 8 || N11C->getZExtValue() != 8)
3034 // Look for (shl (and a, 0xff), 8), (srl (and a, 0xff00), 8)
3035 SDValue N00 = N0->getOperand(0);
3036 if (!LookPassAnd0 && N00.getOpcode() == ISD::AND) {
3037 if (!N00.getNode()->hasOneUse())
3039 ConstantSDNode *N001C = dyn_cast<ConstantSDNode>(N00.getOperand(1));
3040 if (!N001C || N001C->getZExtValue() != 0xFF)
3042 N00 = N00.getOperand(0);
3043 LookPassAnd0 = true;
3046 SDValue N10 = N1->getOperand(0);
3047 if (!LookPassAnd1 && N10.getOpcode() == ISD::AND) {
3048 if (!N10.getNode()->hasOneUse())
3050 ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N10.getOperand(1));
3051 if (!N101C || N101C->getZExtValue() != 0xFF00)
3053 N10 = N10.getOperand(0);
3054 LookPassAnd1 = true;
3060 // Make sure everything beyond the low halfword gets set to zero since the SRL
3061 // 16 will clear the top bits.
3062 unsigned OpSizeInBits = VT.getSizeInBits();
3063 if (DemandHighBits && OpSizeInBits > 16) {
3064 // If the left-shift isn't masked out then the only way this is a bswap is
3065 // if all bits beyond the low 8 are 0. In that case the entire pattern
3066 // reduces to a left shift anyway: leave it for other parts of the combiner.
3070 // However, if the right shift isn't masked out then it might be because
3071 // it's not needed. See if we can spot that too.
3072 if (!LookPassAnd1 &&
3073 !DAG.MaskedValueIsZero(
3074 N10, APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - 16)))
3078 SDValue Res = DAG.getNode(ISD::BSWAP, SDLoc(N), VT, N00);
3079 if (OpSizeInBits > 16)
3080 Res = DAG.getNode(ISD::SRL, SDLoc(N), VT, Res,
3081 DAG.getConstant(OpSizeInBits-16, getShiftAmountTy(VT)));
3085 /// isBSwapHWordElement - Return true if the specified node is an element
3086 /// that makes up a 32-bit packed halfword byteswap. i.e.
3087 /// ((x&0xff)<<8)|((x&0xff00)>>8)|((x&0x00ff0000)<<8)|((x&0xff000000)>>8)
3088 static bool isBSwapHWordElement(SDValue N, SmallVectorImpl<SDNode *> &Parts) {
3089 if (!N.getNode()->hasOneUse())
3092 unsigned Opc = N.getOpcode();
3093 if (Opc != ISD::AND && Opc != ISD::SHL && Opc != ISD::SRL)
3096 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3101 switch (N1C->getZExtValue()) {
3104 case 0xFF: Num = 0; break;
3105 case 0xFF00: Num = 1; break;
3106 case 0xFF0000: Num = 2; break;
3107 case 0xFF000000: Num = 3; break;
3110 // Look for (x & 0xff) << 8 as well as ((x << 8) & 0xff00).
3111 SDValue N0 = N.getOperand(0);
3112 if (Opc == ISD::AND) {
3113 if (Num == 0 || Num == 2) {
3115 // (x >> 8) & 0xff0000
3116 if (N0.getOpcode() != ISD::SRL)
3118 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3119 if (!C || C->getZExtValue() != 8)
3122 // (x << 8) & 0xff00
3123 // (x << 8) & 0xff000000
3124 if (N0.getOpcode() != ISD::SHL)
3126 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3127 if (!C || C->getZExtValue() != 8)
3130 } else if (Opc == ISD::SHL) {
3132 // (x & 0xff0000) << 8
3133 if (Num != 0 && Num != 2)
3135 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3136 if (!C || C->getZExtValue() != 8)
3138 } else { // Opc == ISD::SRL
3139 // (x & 0xff00) >> 8
3140 // (x & 0xff000000) >> 8
3141 if (Num != 1 && Num != 3)
3143 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3144 if (!C || C->getZExtValue() != 8)
3151 Parts[Num] = N0.getOperand(0).getNode();
3155 /// MatchBSwapHWord - Match a 32-bit packed halfword bswap. That is
3156 /// ((x&0xff)<<8)|((x&0xff00)>>8)|((x&0x00ff0000)<<8)|((x&0xff000000)>>8)
3157 /// => (rotl (bswap x), 16)
3158 SDValue DAGCombiner::MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1) {
3159 if (!LegalOperations)
3162 EVT VT = N->getValueType(0);
3165 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
3168 SmallVector<SDNode*,4> Parts(4, (SDNode*)nullptr);
3170 // (or (or (and), (and)), (or (and), (and)))
3171 // (or (or (or (and), (and)), (and)), (and))
3172 if (N0.getOpcode() != ISD::OR)
3174 SDValue N00 = N0.getOperand(0);
3175 SDValue N01 = N0.getOperand(1);
3177 if (N1.getOpcode() == ISD::OR &&
3178 N00.getNumOperands() == 2 && N01.getNumOperands() == 2) {
3179 // (or (or (and), (and)), (or (and), (and)))
3180 SDValue N000 = N00.getOperand(0);
3181 if (!isBSwapHWordElement(N000, Parts))
3184 SDValue N001 = N00.getOperand(1);
3185 if (!isBSwapHWordElement(N001, Parts))
3187 SDValue N010 = N01.getOperand(0);
3188 if (!isBSwapHWordElement(N010, Parts))
3190 SDValue N011 = N01.getOperand(1);
3191 if (!isBSwapHWordElement(N011, Parts))
3194 // (or (or (or (and), (and)), (and)), (and))
3195 if (!isBSwapHWordElement(N1, Parts))
3197 if (!isBSwapHWordElement(N01, Parts))
3199 if (N00.getOpcode() != ISD::OR)
3201 SDValue N000 = N00.getOperand(0);
3202 if (!isBSwapHWordElement(N000, Parts))
3204 SDValue N001 = N00.getOperand(1);
3205 if (!isBSwapHWordElement(N001, Parts))
3209 // Make sure the parts are all coming from the same node.
3210 if (Parts[0] != Parts[1] || Parts[0] != Parts[2] || Parts[0] != Parts[3])
3213 SDValue BSwap = DAG.getNode(ISD::BSWAP, SDLoc(N), VT,
3214 SDValue(Parts[0],0));
3216 // Result of the bswap should be rotated by 16. If it's not legal, then
3217 // do (x << 16) | (x >> 16).
3218 SDValue ShAmt = DAG.getConstant(16, getShiftAmountTy(VT));
3219 if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT))
3220 return DAG.getNode(ISD::ROTL, SDLoc(N), VT, BSwap, ShAmt);
3221 if (TLI.isOperationLegalOrCustom(ISD::ROTR, VT))
3222 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, BSwap, ShAmt);
3223 return DAG.getNode(ISD::OR, SDLoc(N), VT,
3224 DAG.getNode(ISD::SHL, SDLoc(N), VT, BSwap, ShAmt),
3225 DAG.getNode(ISD::SRL, SDLoc(N), VT, BSwap, ShAmt));
3228 SDValue DAGCombiner::visitOR(SDNode *N) {
3229 SDValue N0 = N->getOperand(0);
3230 SDValue N1 = N->getOperand(1);
3231 SDValue LL, LR, RL, RR, CC0, CC1;
3232 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3233 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3234 EVT VT = N1.getValueType();
3237 if (VT.isVector()) {
3238 SDValue FoldedVOp = SimplifyVBinOp(N);
3239 if (FoldedVOp.getNode()) return FoldedVOp;
3241 // fold (or x, 0) -> x, vector edition
3242 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3244 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3247 // fold (or x, -1) -> -1, vector edition
3248 if (ISD::isBuildVectorAllOnes(N0.getNode()))
3250 if (ISD::isBuildVectorAllOnes(N1.getNode()))
3253 // fold (or (shuf A, V_0, MA), (shuf B, V_0, MB)) -> (shuf A, B, Mask1)
3254 // fold (or (shuf A, V_0, MA), (shuf B, V_0, MB)) -> (shuf B, A, Mask2)
3255 // Do this only if the resulting shuffle is legal.
3256 if (isa<ShuffleVectorSDNode>(N0) &&
3257 isa<ShuffleVectorSDNode>(N1) &&
3258 // Avoid folding a node with illegal type.
3259 TLI.isTypeLegal(VT) &&
3260 N0->getOperand(1) == N1->getOperand(1) &&
3261 ISD::isBuildVectorAllZeros(N0.getOperand(1).getNode())) {
3262 bool CanFold = true;
3263 unsigned NumElts = VT.getVectorNumElements();
3264 const ShuffleVectorSDNode *SV0 = cast<ShuffleVectorSDNode>(N0);
3265 const ShuffleVectorSDNode *SV1 = cast<ShuffleVectorSDNode>(N1);
3266 // We construct two shuffle masks:
3267 // - Mask1 is a shuffle mask for a shuffle with N0 as the first operand
3268 // and N1 as the second operand.
3269 // - Mask2 is a shuffle mask for a shuffle with N1 as the first operand
3270 // and N0 as the second operand.
3271 // We do this because OR is commutable and therefore there might be
3272 // two ways to fold this node into a shuffle.
3273 SmallVector<int,4> Mask1;
3274 SmallVector<int,4> Mask2;
3276 for (unsigned i = 0; i != NumElts && CanFold; ++i) {
3277 int M0 = SV0->getMaskElt(i);
3278 int M1 = SV1->getMaskElt(i);
3280 // Both shuffle indexes are undef. Propagate Undef.
3281 if (M0 < 0 && M1 < 0) {
3282 Mask1.push_back(M0);
3283 Mask2.push_back(M0);
3287 if (M0 < 0 || M1 < 0 ||
3288 (M0 < (int)NumElts && M1 < (int)NumElts) ||
3289 (M0 >= (int)NumElts && M1 >= (int)NumElts)) {
3294 Mask1.push_back(M0 < (int)NumElts ? M0 : M1 + NumElts);
3295 Mask2.push_back(M1 < (int)NumElts ? M1 : M0 + NumElts);
3299 // Fold this sequence only if the resulting shuffle is 'legal'.
3300 if (TLI.isShuffleMaskLegal(Mask1, VT))
3301 return DAG.getVectorShuffle(VT, SDLoc(N), N0->getOperand(0),
3302 N1->getOperand(0), &Mask1[0]);
3303 if (TLI.isShuffleMaskLegal(Mask2, VT))
3304 return DAG.getVectorShuffle(VT, SDLoc(N), N1->getOperand(0),
3305 N0->getOperand(0), &Mask2[0]);
3310 // fold (or x, undef) -> -1
3311 if (!LegalOperations &&
3312 (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)) {
3313 EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT;
3314 return DAG.getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()), VT);
3316 // fold (or c1, c2) -> c1|c2
3318 return DAG.FoldConstantArithmetic(ISD::OR, VT, N0C, N1C);
3319 // canonicalize constant to RHS
3321 return DAG.getNode(ISD::OR, SDLoc(N), VT, N1, N0);
3322 // fold (or x, 0) -> x
3323 if (N1C && N1C->isNullValue())
3325 // fold (or x, -1) -> -1
3326 if (N1C && N1C->isAllOnesValue())
3328 // fold (or x, c) -> c iff (x & ~c) == 0
3329 if (N1C && DAG.MaskedValueIsZero(N0, ~N1C->getAPIntValue()))
3332 // Recognize halfword bswaps as (bswap + rotl 16) or (bswap + shl 16)
3333 SDValue BSwap = MatchBSwapHWord(N, N0, N1);
3334 if (BSwap.getNode())
3336 BSwap = MatchBSwapHWordLow(N, N0, N1);
3337 if (BSwap.getNode())
3341 SDValue ROR = ReassociateOps(ISD::OR, SDLoc(N), N0, N1);
3344 // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2)
3345 // iff (c1 & c2) == 0.
3346 if (N1C && N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
3347 isa<ConstantSDNode>(N0.getOperand(1))) {
3348 ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1));
3349 if ((C1->getAPIntValue() & N1C->getAPIntValue()) != 0) {
3350 SDValue COR = DAG.FoldConstantArithmetic(ISD::OR, VT, N1C, C1);
3353 return DAG.getNode(ISD::AND, SDLoc(N), VT,
3354 DAG.getNode(ISD::OR, SDLoc(N0), VT,
3355 N0.getOperand(0), N1), COR);
3358 // fold (or (setcc x), (setcc y)) -> (setcc (or x, y))
3359 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
3360 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
3361 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
3363 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
3364 LL.getValueType().isInteger()) {
3365 // fold (or (setne X, 0), (setne Y, 0)) -> (setne (or X, Y), 0)
3366 // fold (or (setlt X, 0), (setlt Y, 0)) -> (setne (or X, Y), 0)
3367 if (cast<ConstantSDNode>(LR)->isNullValue() &&
3368 (Op1 == ISD::SETNE || Op1 == ISD::SETLT)) {
3369 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(LR),
3370 LR.getValueType(), LL, RL);
3371 AddToWorkList(ORNode.getNode());
3372 return DAG.getSetCC(SDLoc(N), VT, ORNode, LR, Op1);
3374 // fold (or (setne X, -1), (setne Y, -1)) -> (setne (and X, Y), -1)
3375 // fold (or (setgt X, -1), (setgt Y -1)) -> (setgt (and X, Y), -1)
3376 if (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
3377 (Op1 == ISD::SETNE || Op1 == ISD::SETGT)) {
3378 SDValue ANDNode = DAG.getNode(ISD::AND, SDLoc(LR),
3379 LR.getValueType(), LL, RL);
3380 AddToWorkList(ANDNode.getNode());
3381 return DAG.getSetCC(SDLoc(N), VT, ANDNode, LR, Op1);
3384 // canonicalize equivalent to ll == rl
3385 if (LL == RR && LR == RL) {
3386 Op1 = ISD::getSetCCSwappedOperands(Op1);
3389 if (LL == RL && LR == RR) {
3390 bool isInteger = LL.getValueType().isInteger();
3391 ISD::CondCode Result = ISD::getSetCCOrOperation(Op0, Op1, isInteger);
3392 if (Result != ISD::SETCC_INVALID &&
3393 (!LegalOperations ||
3394 (TLI.isCondCodeLegal(Result, LL.getSimpleValueType()) &&
3395 TLI.isOperationLegal(ISD::SETCC,
3396 getSetCCResultType(N0.getValueType())))))
3397 return DAG.getSetCC(SDLoc(N), N0.getValueType(),
3402 // Simplify: (or (op x...), (op y...)) -> (op (or x, y))
3403 if (N0.getOpcode() == N1.getOpcode()) {
3404 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3405 if (Tmp.getNode()) return Tmp;
3408 // (or (and X, C1), (and Y, C2)) -> (and (or X, Y), C3) if possible.
3409 if (N0.getOpcode() == ISD::AND &&
3410 N1.getOpcode() == ISD::AND &&
3411 N0.getOperand(1).getOpcode() == ISD::Constant &&
3412 N1.getOperand(1).getOpcode() == ISD::Constant &&
3413 // Don't increase # computations.
3414 (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) {
3415 // We can only do this xform if we know that bits from X that are set in C2
3416 // but not in C1 are already zero. Likewise for Y.
3417 const APInt &LHSMask =
3418 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
3419 const APInt &RHSMask =
3420 cast<ConstantSDNode>(N1.getOperand(1))->getAPIntValue();
3422 if (DAG.MaskedValueIsZero(N0.getOperand(0), RHSMask&~LHSMask) &&
3423 DAG.MaskedValueIsZero(N1.getOperand(0), LHSMask&~RHSMask)) {
3424 SDValue X = DAG.getNode(ISD::OR, SDLoc(N0), VT,
3425 N0.getOperand(0), N1.getOperand(0));
3426 return DAG.getNode(ISD::AND, SDLoc(N), VT, X,
3427 DAG.getConstant(LHSMask | RHSMask, VT));
3431 // See if this is some rotate idiom.
3432 if (SDNode *Rot = MatchRotate(N0, N1, SDLoc(N)))
3433 return SDValue(Rot, 0);
3435 // Simplify the operands using demanded-bits information.
3436 if (!VT.isVector() &&
3437 SimplifyDemandedBits(SDValue(N, 0)))
3438 return SDValue(N, 0);
3443 /// MatchRotateHalf - Match "(X shl/srl V1) & V2" where V2 may not be present.
3444 static bool MatchRotateHalf(SDValue Op, SDValue &Shift, SDValue &Mask) {
3445 if (Op.getOpcode() == ISD::AND) {
3446 if (isa<ConstantSDNode>(Op.getOperand(1))) {
3447 Mask = Op.getOperand(1);
3448 Op = Op.getOperand(0);
3454 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SHL) {
3462 // Return true if we can prove that, whenever Neg and Pos are both in the
3463 // range [0, OpSize), Neg == (Pos == 0 ? 0 : OpSize - Pos). This means that
3464 // for two opposing shifts shift1 and shift2 and a value X with OpBits bits:
3466 // (or (shift1 X, Neg), (shift2 X, Pos))
3468 // reduces to a rotate in direction shift2 by Pos or (equivalently) a rotate
3469 // in direction shift1 by Neg. The range [0, OpSize) means that we only need
3470 // to consider shift amounts with defined behavior.
3471 static bool matchRotateSub(SDValue Pos, SDValue Neg, unsigned OpSize) {
3472 // If OpSize is a power of 2 then:
3474 // (a) (Pos == 0 ? 0 : OpSize - Pos) == (OpSize - Pos) & (OpSize - 1)
3475 // (b) Neg == Neg & (OpSize - 1) whenever Neg is in [0, OpSize).
3477 // So if OpSize is a power of 2 and Neg is (and Neg', OpSize-1), we check
3478 // for the stronger condition:
3480 // Neg & (OpSize - 1) == (OpSize - Pos) & (OpSize - 1) [A]
3482 // for all Neg and Pos. Since Neg & (OpSize - 1) == Neg' & (OpSize - 1)
3483 // we can just replace Neg with Neg' for the rest of the function.
3485 // In other cases we check for the even stronger condition:
3487 // Neg == OpSize - Pos [B]
3489 // for all Neg and Pos. Note that the (or ...) then invokes undefined
3490 // behavior if Pos == 0 (and consequently Neg == OpSize).
3492 // We could actually use [A] whenever OpSize is a power of 2, but the
3493 // only extra cases that it would match are those uninteresting ones
3494 // where Neg and Pos are never in range at the same time. E.g. for
3495 // OpSize == 32, using [A] would allow a Neg of the form (sub 64, Pos)
3496 // as well as (sub 32, Pos), but:
3498 // (or (shift1 X, (sub 64, Pos)), (shift2 X, Pos))
3500 // always invokes undefined behavior for 32-bit X.
3502 // Below, Mask == OpSize - 1 when using [A] and is all-ones otherwise.
3503 unsigned MaskLoBits = 0;
3504 if (Neg.getOpcode() == ISD::AND &&
3505 isPowerOf2_64(OpSize) &&
3506 Neg.getOperand(1).getOpcode() == ISD::Constant &&
3507 cast<ConstantSDNode>(Neg.getOperand(1))->getAPIntValue() == OpSize - 1) {
3508 Neg = Neg.getOperand(0);
3509 MaskLoBits = Log2_64(OpSize);
3512 // Check whether Neg has the form (sub NegC, NegOp1) for some NegC and NegOp1.
3513 if (Neg.getOpcode() != ISD::SUB)
3515 ConstantSDNode *NegC = dyn_cast<ConstantSDNode>(Neg.getOperand(0));
3518 SDValue NegOp1 = Neg.getOperand(1);
3520 // On the RHS of [A], if Pos is Pos' & (OpSize - 1), just replace Pos with
3521 // Pos'. The truncation is redundant for the purpose of the equality.
3523 Pos.getOpcode() == ISD::AND &&
3524 Pos.getOperand(1).getOpcode() == ISD::Constant &&
3525 cast<ConstantSDNode>(Pos.getOperand(1))->getAPIntValue() == OpSize - 1)
3526 Pos = Pos.getOperand(0);
3528 // The condition we need is now:
3530 // (NegC - NegOp1) & Mask == (OpSize - Pos) & Mask
3532 // If NegOp1 == Pos then we need:
3534 // OpSize & Mask == NegC & Mask
3536 // (because "x & Mask" is a truncation and distributes through subtraction).
3539 Width = NegC->getAPIntValue();
3540 // Check for cases where Pos has the form (add NegOp1, PosC) for some PosC.
3541 // Then the condition we want to prove becomes:
3543 // (NegC - NegOp1) & Mask == (OpSize - (NegOp1 + PosC)) & Mask
3545 // which, again because "x & Mask" is a truncation, becomes:
3547 // NegC & Mask == (OpSize - PosC) & Mask
3548 // OpSize & Mask == (NegC + PosC) & Mask
3549 else if (Pos.getOpcode() == ISD::ADD &&
3550 Pos.getOperand(0) == NegOp1 &&
3551 Pos.getOperand(1).getOpcode() == ISD::Constant)
3552 Width = (cast<ConstantSDNode>(Pos.getOperand(1))->getAPIntValue() +
3553 NegC->getAPIntValue());
3557 // Now we just need to check that OpSize & Mask == Width & Mask.
3559 // Opsize & Mask is 0 since Mask is Opsize - 1.
3560 return Width.getLoBits(MaskLoBits) == 0;
3561 return Width == OpSize;
3564 // A subroutine of MatchRotate used once we have found an OR of two opposite
3565 // shifts of Shifted. If Neg == <operand size> - Pos then the OR reduces
3566 // to both (PosOpcode Shifted, Pos) and (NegOpcode Shifted, Neg), with the
3567 // former being preferred if supported. InnerPos and InnerNeg are Pos and
3568 // Neg with outer conversions stripped away.
3569 SDNode *DAGCombiner::MatchRotatePosNeg(SDValue Shifted, SDValue Pos,
3570 SDValue Neg, SDValue InnerPos,
3571 SDValue InnerNeg, unsigned PosOpcode,
3572 unsigned NegOpcode, SDLoc DL) {
3573 // fold (or (shl x, (*ext y)),
3574 // (srl x, (*ext (sub 32, y)))) ->
3575 // (rotl x, y) or (rotr x, (sub 32, y))
3577 // fold (or (shl x, (*ext (sub 32, y))),
3578 // (srl x, (*ext y))) ->
3579 // (rotr x, y) or (rotl x, (sub 32, y))
3580 EVT VT = Shifted.getValueType();
3581 if (matchRotateSub(InnerPos, InnerNeg, VT.getSizeInBits())) {
3582 bool HasPos = TLI.isOperationLegalOrCustom(PosOpcode, VT);
3583 return DAG.getNode(HasPos ? PosOpcode : NegOpcode, DL, VT, Shifted,
3584 HasPos ? Pos : Neg).getNode();
3590 // MatchRotate - Handle an 'or' of two operands. If this is one of the many
3591 // idioms for rotate, and if the target supports rotation instructions, generate
3593 SDNode *DAGCombiner::MatchRotate(SDValue LHS, SDValue RHS, SDLoc DL) {
3594 // Must be a legal type. Expanded 'n promoted things won't work with rotates.
3595 EVT VT = LHS.getValueType();
3596 if (!TLI.isTypeLegal(VT)) return nullptr;
3598 // The target must have at least one rotate flavor.
3599 bool HasROTL = TLI.isOperationLegalOrCustom(ISD::ROTL, VT);
3600 bool HasROTR = TLI.isOperationLegalOrCustom(ISD::ROTR, VT);
3601 if (!HasROTL && !HasROTR) return nullptr;
3603 // Match "(X shl/srl V1) & V2" where V2 may not be present.
3604 SDValue LHSShift; // The shift.
3605 SDValue LHSMask; // AND value if any.
3606 if (!MatchRotateHalf(LHS, LHSShift, LHSMask))
3607 return nullptr; // Not part of a rotate.
3609 SDValue RHSShift; // The shift.
3610 SDValue RHSMask; // AND value if any.
3611 if (!MatchRotateHalf(RHS, RHSShift, RHSMask))
3612 return nullptr; // Not part of a rotate.
3614 if (LHSShift.getOperand(0) != RHSShift.getOperand(0))
3615 return nullptr; // Not shifting the same value.
3617 if (LHSShift.getOpcode() == RHSShift.getOpcode())
3618 return nullptr; // Shifts must disagree.
3620 // Canonicalize shl to left side in a shl/srl pair.
3621 if (RHSShift.getOpcode() == ISD::SHL) {
3622 std::swap(LHS, RHS);
3623 std::swap(LHSShift, RHSShift);
3624 std::swap(LHSMask , RHSMask );
3627 unsigned OpSizeInBits = VT.getSizeInBits();
3628 SDValue LHSShiftArg = LHSShift.getOperand(0);
3629 SDValue LHSShiftAmt = LHSShift.getOperand(1);
3630 SDValue RHSShiftArg = RHSShift.getOperand(0);
3631 SDValue RHSShiftAmt = RHSShift.getOperand(1);
3633 // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1)
3634 // fold (or (shl x, C1), (srl x, C2)) -> (rotr x, C2)
3635 if (LHSShiftAmt.getOpcode() == ISD::Constant &&
3636 RHSShiftAmt.getOpcode() == ISD::Constant) {
3637 uint64_t LShVal = cast<ConstantSDNode>(LHSShiftAmt)->getZExtValue();
3638 uint64_t RShVal = cast<ConstantSDNode>(RHSShiftAmt)->getZExtValue();
3639 if ((LShVal + RShVal) != OpSizeInBits)
3642 SDValue Rot = DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT,
3643 LHSShiftArg, HasROTL ? LHSShiftAmt : RHSShiftAmt);
3645 // If there is an AND of either shifted operand, apply it to the result.
3646 if (LHSMask.getNode() || RHSMask.getNode()) {
3647 APInt Mask = APInt::getAllOnesValue(OpSizeInBits);
3649 if (LHSMask.getNode()) {
3650 APInt RHSBits = APInt::getLowBitsSet(OpSizeInBits, LShVal);
3651 Mask &= cast<ConstantSDNode>(LHSMask)->getAPIntValue() | RHSBits;
3653 if (RHSMask.getNode()) {
3654 APInt LHSBits = APInt::getHighBitsSet(OpSizeInBits, RShVal);
3655 Mask &= cast<ConstantSDNode>(RHSMask)->getAPIntValue() | LHSBits;
3658 Rot = DAG.getNode(ISD::AND, DL, VT, Rot, DAG.getConstant(Mask, VT));
3661 return Rot.getNode();
3664 // If there is a mask here, and we have a variable shift, we can't be sure
3665 // that we're masking out the right stuff.
3666 if (LHSMask.getNode() || RHSMask.getNode())
3669 // If the shift amount is sign/zext/any-extended just peel it off.
3670 SDValue LExtOp0 = LHSShiftAmt;
3671 SDValue RExtOp0 = RHSShiftAmt;
3672 if ((LHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3673 LHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3674 LHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3675 LHSShiftAmt.getOpcode() == ISD::TRUNCATE) &&
3676 (RHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3677 RHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3678 RHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3679 RHSShiftAmt.getOpcode() == ISD::TRUNCATE)) {
3680 LExtOp0 = LHSShiftAmt.getOperand(0);
3681 RExtOp0 = RHSShiftAmt.getOperand(0);
3684 SDNode *TryL = MatchRotatePosNeg(LHSShiftArg, LHSShiftAmt, RHSShiftAmt,
3685 LExtOp0, RExtOp0, ISD::ROTL, ISD::ROTR, DL);
3689 SDNode *TryR = MatchRotatePosNeg(RHSShiftArg, RHSShiftAmt, LHSShiftAmt,
3690 RExtOp0, LExtOp0, ISD::ROTR, ISD::ROTL, DL);
3697 SDValue DAGCombiner::visitXOR(SDNode *N) {
3698 SDValue N0 = N->getOperand(0);
3699 SDValue N1 = N->getOperand(1);
3700 SDValue LHS, RHS, CC;
3701 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3702 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3703 EVT VT = N0.getValueType();
3706 if (VT.isVector()) {
3707 SDValue FoldedVOp = SimplifyVBinOp(N);
3708 if (FoldedVOp.getNode()) return FoldedVOp;
3710 // fold (xor x, 0) -> x, vector edition
3711 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3713 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3717 // fold (xor undef, undef) -> 0. This is a common idiom (misuse).
3718 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
3719 return DAG.getConstant(0, VT);
3720 // fold (xor x, undef) -> undef
3721 if (N0.getOpcode() == ISD::UNDEF)
3723 if (N1.getOpcode() == ISD::UNDEF)
3725 // fold (xor c1, c2) -> c1^c2
3727 return DAG.FoldConstantArithmetic(ISD::XOR, VT, N0C, N1C);
3728 // canonicalize constant to RHS
3730 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0);
3731 // fold (xor x, 0) -> x
3732 if (N1C && N1C->isNullValue())
3735 SDValue RXOR = ReassociateOps(ISD::XOR, SDLoc(N), N0, N1);
3739 // fold !(x cc y) -> (x !cc y)
3740 if (N1C && N1C->getAPIntValue() == 1 && isSetCCEquivalent(N0, LHS, RHS, CC)) {
3741 bool isInt = LHS.getValueType().isInteger();
3742 ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
3745 if (!LegalOperations ||
3746 TLI.isCondCodeLegal(NotCC, LHS.getSimpleValueType())) {
3747 switch (N0.getOpcode()) {
3749 llvm_unreachable("Unhandled SetCC Equivalent!");
3751 return DAG.getSetCC(SDLoc(N), VT, LHS, RHS, NotCC);
3752 case ISD::SELECT_CC:
3753 return DAG.getSelectCC(SDLoc(N), LHS, RHS, N0.getOperand(2),
3754 N0.getOperand(3), NotCC);
3759 // fold (not (zext (setcc x, y))) -> (zext (not (setcc x, y)))
3760 if (N1C && N1C->getAPIntValue() == 1 && N0.getOpcode() == ISD::ZERO_EXTEND &&
3761 N0.getNode()->hasOneUse() &&
3762 isSetCCEquivalent(N0.getOperand(0), LHS, RHS, CC)){
3763 SDValue V = N0.getOperand(0);
3764 V = DAG.getNode(ISD::XOR, SDLoc(N0), V.getValueType(), V,
3765 DAG.getConstant(1, V.getValueType()));
3766 AddToWorkList(V.getNode());
3767 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, V);
3770 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are setcc
3771 if (N1C && N1C->getAPIntValue() == 1 && VT == MVT::i1 &&
3772 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3773 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3774 if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) {
3775 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3776 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS
3777 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS
3778 AddToWorkList(LHS.getNode()); AddToWorkList(RHS.getNode());
3779 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS);
3782 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are constants
3783 if (N1C && N1C->isAllOnesValue() &&
3784 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3785 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3786 if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) {
3787 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3788 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS
3789 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS
3790 AddToWorkList(LHS.getNode()); AddToWorkList(RHS.getNode());
3791 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS);
3794 // fold (xor (and x, y), y) -> (and (not x), y)
3795 if (N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
3796 N0->getOperand(1) == N1) {
3797 SDValue X = N0->getOperand(0);
3798 SDValue NotX = DAG.getNOT(SDLoc(X), X, VT);
3799 AddToWorkList(NotX.getNode());
3800 return DAG.getNode(ISD::AND, SDLoc(N), VT, NotX, N1);
3802 // fold (xor (xor x, c1), c2) -> (xor x, (xor c1, c2))
3803 if (N1C && N0.getOpcode() == ISD::XOR) {
3804 ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0));
3805 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3807 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N0.getOperand(1),
3808 DAG.getConstant(N1C->getAPIntValue() ^
3809 N00C->getAPIntValue(), VT));
3811 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N0.getOperand(0),
3812 DAG.getConstant(N1C->getAPIntValue() ^
3813 N01C->getAPIntValue(), VT));
3815 // fold (xor x, x) -> 0
3817 return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations, LegalTypes);
3819 // Simplify: xor (op x...), (op y...) -> (op (xor x, y))
3820 if (N0.getOpcode() == N1.getOpcode()) {
3821 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3822 if (Tmp.getNode()) return Tmp;
3825 // Simplify the expression using non-local knowledge.
3826 if (!VT.isVector() &&
3827 SimplifyDemandedBits(SDValue(N, 0)))
3828 return SDValue(N, 0);
3833 /// visitShiftByConstant - Handle transforms common to the three shifts, when
3834 /// the shift amount is a constant.
3835 SDValue DAGCombiner::visitShiftByConstant(SDNode *N, ConstantSDNode *Amt) {
3836 // We can't and shouldn't fold opaque constants.
3837 if (Amt->isOpaque())
3840 SDNode *LHS = N->getOperand(0).getNode();
3841 if (!LHS->hasOneUse()) return SDValue();
3843 // We want to pull some binops through shifts, so that we have (and (shift))
3844 // instead of (shift (and)), likewise for add, or, xor, etc. This sort of
3845 // thing happens with address calculations, so it's important to canonicalize
3847 bool HighBitSet = false; // Can we transform this if the high bit is set?
3849 switch (LHS->getOpcode()) {
3850 default: return SDValue();
3853 HighBitSet = false; // We can only transform sra if the high bit is clear.
3856 HighBitSet = true; // We can only transform sra if the high bit is set.
3859 if (N->getOpcode() != ISD::SHL)
3860 return SDValue(); // only shl(add) not sr[al](add).
3861 HighBitSet = false; // We can only transform sra if the high bit is clear.
3865 // We require the RHS of the binop to be a constant and not opaque as well.
3866 ConstantSDNode *BinOpCst = dyn_cast<ConstantSDNode>(LHS->getOperand(1));
3867 if (!BinOpCst || BinOpCst->isOpaque()) return SDValue();
3869 // FIXME: disable this unless the input to the binop is a shift by a constant.
3870 // If it is not a shift, it pessimizes some common cases like:
3872 // void foo(int *X, int i) { X[i & 1235] = 1; }
3873 // int bar(int *X, int i) { return X[i & 255]; }
3874 SDNode *BinOpLHSVal = LHS->getOperand(0).getNode();
3875 if ((BinOpLHSVal->getOpcode() != ISD::SHL &&
3876 BinOpLHSVal->getOpcode() != ISD::SRA &&
3877 BinOpLHSVal->getOpcode() != ISD::SRL) ||
3878 !isa<ConstantSDNode>(BinOpLHSVal->getOperand(1)))
3881 EVT VT = N->getValueType(0);
3883 // If this is a signed shift right, and the high bit is modified by the
3884 // logical operation, do not perform the transformation. The highBitSet
3885 // boolean indicates the value of the high bit of the constant which would
3886 // cause it to be modified for this operation.
3887 if (N->getOpcode() == ISD::SRA) {
3888 bool BinOpRHSSignSet = BinOpCst->getAPIntValue().isNegative();
3889 if (BinOpRHSSignSet != HighBitSet)
3893 if (!TLI.isDesirableToCommuteWithShift(LHS))
3896 // Fold the constants, shifting the binop RHS by the shift amount.
3897 SDValue NewRHS = DAG.getNode(N->getOpcode(), SDLoc(LHS->getOperand(1)),
3899 LHS->getOperand(1), N->getOperand(1));
3900 assert(isa<ConstantSDNode>(NewRHS) && "Folding was not successful!");
3902 // Create the new shift.
3903 SDValue NewShift = DAG.getNode(N->getOpcode(),
3904 SDLoc(LHS->getOperand(0)),
3905 VT, LHS->getOperand(0), N->getOperand(1));
3907 // Create the new binop.
3908 return DAG.getNode(LHS->getOpcode(), SDLoc(N), VT, NewShift, NewRHS);
3911 SDValue DAGCombiner::distributeTruncateThroughAnd(SDNode *N) {
3912 assert(N->getOpcode() == ISD::TRUNCATE);
3913 assert(N->getOperand(0).getOpcode() == ISD::AND);
3915 // (truncate:TruncVT (and N00, N01C)) -> (and (truncate:TruncVT N00), TruncC)
3916 if (N->hasOneUse() && N->getOperand(0).hasOneUse()) {
3917 SDValue N01 = N->getOperand(0).getOperand(1);
3919 if (ConstantSDNode *N01C = isConstOrConstSplat(N01)) {
3920 EVT TruncVT = N->getValueType(0);
3921 SDValue N00 = N->getOperand(0).getOperand(0);
3922 APInt TruncC = N01C->getAPIntValue();
3923 TruncC = TruncC.trunc(TruncVT.getScalarSizeInBits());
3925 return DAG.getNode(ISD::AND, SDLoc(N), TruncVT,
3926 DAG.getNode(ISD::TRUNCATE, SDLoc(N), TruncVT, N00),
3927 DAG.getConstant(TruncC, TruncVT));
3934 SDValue DAGCombiner::visitRotate(SDNode *N) {
3935 // fold (rot* x, (trunc (and y, c))) -> (rot* x, (and (trunc y), (trunc c))).
3936 if (N->getOperand(1).getOpcode() == ISD::TRUNCATE &&
3937 N->getOperand(1).getOperand(0).getOpcode() == ISD::AND) {
3938 SDValue NewOp1 = distributeTruncateThroughAnd(N->getOperand(1).getNode());
3939 if (NewOp1.getNode())
3940 return DAG.getNode(N->getOpcode(), SDLoc(N), N->getValueType(0),
3941 N->getOperand(0), NewOp1);
3946 SDValue DAGCombiner::visitSHL(SDNode *N) {
3947 SDValue N0 = N->getOperand(0);
3948 SDValue N1 = N->getOperand(1);
3949 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3950 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3951 EVT VT = N0.getValueType();
3952 unsigned OpSizeInBits = VT.getScalarSizeInBits();
3955 if (VT.isVector()) {
3956 SDValue FoldedVOp = SimplifyVBinOp(N);
3957 if (FoldedVOp.getNode()) return FoldedVOp;
3959 BuildVectorSDNode *N1CV = dyn_cast<BuildVectorSDNode>(N1);
3960 // If setcc produces all-one true value then:
3961 // (shl (and (setcc) N01CV) N1CV) -> (and (setcc) N01CV<<N1CV)
3962 if (N1CV && N1CV->isConstant()) {
3963 if (N0.getOpcode() == ISD::AND) {
3964 SDValue N00 = N0->getOperand(0);
3965 SDValue N01 = N0->getOperand(1);
3966 BuildVectorSDNode *N01CV = dyn_cast<BuildVectorSDNode>(N01);
3968 if (N01CV && N01CV->isConstant() && N00.getOpcode() == ISD::SETCC &&
3969 TLI.getBooleanContents(N00.getOperand(0).getValueType()) ==
3970 TargetLowering::ZeroOrNegativeOneBooleanContent) {
3971 SDValue C = DAG.FoldConstantArithmetic(ISD::SHL, VT, N01CV, N1CV);
3973 return DAG.getNode(ISD::AND, SDLoc(N), VT, N00, C);
3976 N1C = isConstOrConstSplat(N1);
3981 // fold (shl c1, c2) -> c1<<c2
3983 return DAG.FoldConstantArithmetic(ISD::SHL, VT, N0C, N1C);
3984 // fold (shl 0, x) -> 0
3985 if (N0C && N0C->isNullValue())
3987 // fold (shl x, c >= size(x)) -> undef
3988 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
3989 return DAG.getUNDEF(VT);
3990 // fold (shl x, 0) -> x
3991 if (N1C && N1C->isNullValue())
3993 // fold (shl undef, x) -> 0
3994 if (N0.getOpcode() == ISD::UNDEF)
3995 return DAG.getConstant(0, VT);
3996 // if (shl x, c) is known to be zero, return 0
3997 if (DAG.MaskedValueIsZero(SDValue(N, 0),
3998 APInt::getAllOnesValue(OpSizeInBits)))
3999 return DAG.getConstant(0, VT);
4000 // fold (shl x, (trunc (and y, c))) -> (shl x, (and (trunc y), (trunc c))).
4001 if (N1.getOpcode() == ISD::TRUNCATE &&
4002 N1.getOperand(0).getOpcode() == ISD::AND) {
4003 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4004 if (NewOp1.getNode())
4005 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0, NewOp1);
4008 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4009 return SDValue(N, 0);
4011 // fold (shl (shl x, c1), c2) -> 0 or (shl x, (add c1, c2))
4012 if (N1C && N0.getOpcode() == ISD::SHL) {
4013 if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
4014 uint64_t c1 = N0C1->getZExtValue();
4015 uint64_t c2 = N1C->getZExtValue();
4016 if (c1 + c2 >= OpSizeInBits)
4017 return DAG.getConstant(0, VT);
4018 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0),
4019 DAG.getConstant(c1 + c2, N1.getValueType()));
4023 // fold (shl (ext (shl x, c1)), c2) -> (ext (shl x, (add c1, c2)))
4024 // For this to be valid, the second form must not preserve any of the bits
4025 // that are shifted out by the inner shift in the first form. This means
4026 // the outer shift size must be >= the number of bits added by the ext.
4027 // As a corollary, we don't care what kind of ext it is.
4028 if (N1C && (N0.getOpcode() == ISD::ZERO_EXTEND ||
4029 N0.getOpcode() == ISD::ANY_EXTEND ||
4030 N0.getOpcode() == ISD::SIGN_EXTEND) &&
4031 N0.getOperand(0).getOpcode() == ISD::SHL) {
4032 SDValue N0Op0 = N0.getOperand(0);
4033 if (ConstantSDNode *N0Op0C1 = isConstOrConstSplat(N0Op0.getOperand(1))) {
4034 uint64_t c1 = N0Op0C1->getZExtValue();
4035 uint64_t c2 = N1C->getZExtValue();
4036 EVT InnerShiftVT = N0Op0.getValueType();
4037 uint64_t InnerShiftSize = InnerShiftVT.getScalarSizeInBits();
4038 if (c2 >= OpSizeInBits - InnerShiftSize) {
4039 if (c1 + c2 >= OpSizeInBits)
4040 return DAG.getConstant(0, VT);
4041 return DAG.getNode(ISD::SHL, SDLoc(N0), VT,
4042 DAG.getNode(N0.getOpcode(), SDLoc(N0), VT,
4043 N0Op0->getOperand(0)),
4044 DAG.getConstant(c1 + c2, N1.getValueType()));
4049 // fold (shl (zext (srl x, C)), C) -> (zext (shl (srl x, C), C))
4050 // Only fold this if the inner zext has no other uses to avoid increasing
4051 // the total number of instructions.
4052 if (N1C && N0.getOpcode() == ISD::ZERO_EXTEND && N0.hasOneUse() &&
4053 N0.getOperand(0).getOpcode() == ISD::SRL) {
4054 SDValue N0Op0 = N0.getOperand(0);
4055 if (ConstantSDNode *N0Op0C1 = isConstOrConstSplat(N0Op0.getOperand(1))) {
4056 uint64_t c1 = N0Op0C1->getZExtValue();
4057 if (c1 < VT.getScalarSizeInBits()) {
4058 uint64_t c2 = N1C->getZExtValue();
4060 SDValue NewOp0 = N0.getOperand(0);
4061 EVT CountVT = NewOp0.getOperand(1).getValueType();
4062 SDValue NewSHL = DAG.getNode(ISD::SHL, SDLoc(N), NewOp0.getValueType(),
4063 NewOp0, DAG.getConstant(c2, CountVT));
4064 AddToWorkList(NewSHL.getNode());
4065 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N0), VT, NewSHL);
4071 // fold (shl (srl x, c1), c2) -> (and (shl x, (sub c2, c1), MASK) or
4072 // (and (srl x, (sub c1, c2), MASK)
4073 // Only fold this if the inner shift has no other uses -- if it does, folding
4074 // this will increase the total number of instructions.
4075 if (N1C && N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
4076 if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
4077 uint64_t c1 = N0C1->getZExtValue();
4078 if (c1 < OpSizeInBits) {
4079 uint64_t c2 = N1C->getZExtValue();
4080 APInt Mask = APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - c1);
4083 Mask = Mask.shl(c2 - c1);
4084 Shift = DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0),
4085 DAG.getConstant(c2 - c1, N1.getValueType()));
4087 Mask = Mask.lshr(c1 - c2);
4088 Shift = DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0),
4089 DAG.getConstant(c1 - c2, N1.getValueType()));
4091 return DAG.getNode(ISD::AND, SDLoc(N0), VT, Shift,
4092 DAG.getConstant(Mask, VT));
4096 // fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1))
4097 if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1)) {
4098 unsigned BitSize = VT.getScalarSizeInBits();
4099 SDValue HiBitsMask =
4100 DAG.getConstant(APInt::getHighBitsSet(BitSize,
4101 BitSize - N1C->getZExtValue()), VT);
4102 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0.getOperand(0),
4107 SDValue NewSHL = visitShiftByConstant(N, N1C);
4108 if (NewSHL.getNode())
4115 SDValue DAGCombiner::visitSRA(SDNode *N) {
4116 SDValue N0 = N->getOperand(0);
4117 SDValue N1 = N->getOperand(1);
4118 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4119 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4120 EVT VT = N0.getValueType();
4121 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
4124 if (VT.isVector()) {
4125 SDValue FoldedVOp = SimplifyVBinOp(N);
4126 if (FoldedVOp.getNode()) return FoldedVOp;
4128 N1C = isConstOrConstSplat(N1);
4131 // fold (sra c1, c2) -> (sra c1, c2)
4133 return DAG.FoldConstantArithmetic(ISD::SRA, VT, N0C, N1C);
4134 // fold (sra 0, x) -> 0
4135 if (N0C && N0C->isNullValue())
4137 // fold (sra -1, x) -> -1
4138 if (N0C && N0C->isAllOnesValue())
4140 // fold (sra x, (setge c, size(x))) -> undef
4141 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4142 return DAG.getUNDEF(VT);
4143 // fold (sra x, 0) -> x
4144 if (N1C && N1C->isNullValue())
4146 // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports
4148 if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) {
4149 unsigned LowBits = OpSizeInBits - (unsigned)N1C->getZExtValue();
4150 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), LowBits);
4152 ExtVT = EVT::getVectorVT(*DAG.getContext(),
4153 ExtVT, VT.getVectorNumElements());
4154 if ((!LegalOperations ||
4155 TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, ExtVT)))
4156 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
4157 N0.getOperand(0), DAG.getValueType(ExtVT));
4160 // fold (sra (sra x, c1), c2) -> (sra x, (add c1, c2))
4161 if (N1C && N0.getOpcode() == ISD::SRA) {
4162 if (ConstantSDNode *C1 = isConstOrConstSplat(N0.getOperand(1))) {
4163 unsigned Sum = N1C->getZExtValue() + C1->getZExtValue();
4164 if (Sum >= OpSizeInBits)
4165 Sum = OpSizeInBits - 1;
4166 return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0.getOperand(0),
4167 DAG.getConstant(Sum, N1.getValueType()));
4171 // fold (sra (shl X, m), (sub result_size, n))
4172 // -> (sign_extend (trunc (shl X, (sub (sub result_size, n), m)))) for
4173 // result_size - n != m.
4174 // If truncate is free for the target sext(shl) is likely to result in better
4176 if (N0.getOpcode() == ISD::SHL && N1C) {
4177 // Get the two constanst of the shifts, CN0 = m, CN = n.
4178 const ConstantSDNode *N01C = isConstOrConstSplat(N0.getOperand(1));
4180 LLVMContext &Ctx = *DAG.getContext();
4181 // Determine what the truncate's result bitsize and type would be.
4182 EVT TruncVT = EVT::getIntegerVT(Ctx, OpSizeInBits - N1C->getZExtValue());
4185 TruncVT = EVT::getVectorVT(Ctx, TruncVT, VT.getVectorNumElements());
4187 // Determine the residual right-shift amount.
4188 signed ShiftAmt = N1C->getZExtValue() - N01C->getZExtValue();
4190 // If the shift is not a no-op (in which case this should be just a sign
4191 // extend already), the truncated to type is legal, sign_extend is legal
4192 // on that type, and the truncate to that type is both legal and free,
4193 // perform the transform.
4194 if ((ShiftAmt > 0) &&
4195 TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND, TruncVT) &&
4196 TLI.isOperationLegalOrCustom(ISD::TRUNCATE, VT) &&
4197 TLI.isTruncateFree(VT, TruncVT)) {
4199 SDValue Amt = DAG.getConstant(ShiftAmt,
4200 getShiftAmountTy(N0.getOperand(0).getValueType()));
4201 SDValue Shift = DAG.getNode(ISD::SRL, SDLoc(N0), VT,
4202 N0.getOperand(0), Amt);
4203 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), TruncVT,
4205 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N),
4206 N->getValueType(0), Trunc);
4211 // fold (sra x, (trunc (and y, c))) -> (sra x, (and (trunc y), (trunc c))).
4212 if (N1.getOpcode() == ISD::TRUNCATE &&
4213 N1.getOperand(0).getOpcode() == ISD::AND) {
4214 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4215 if (NewOp1.getNode())
4216 return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0, NewOp1);
4219 // fold (sra (trunc (srl x, c1)), c2) -> (trunc (sra x, c1 + c2))
4220 // if c1 is equal to the number of bits the trunc removes
4221 if (N0.getOpcode() == ISD::TRUNCATE &&
4222 (N0.getOperand(0).getOpcode() == ISD::SRL ||
4223 N0.getOperand(0).getOpcode() == ISD::SRA) &&
4224 N0.getOperand(0).hasOneUse() &&
4225 N0.getOperand(0).getOperand(1).hasOneUse() &&
4227 SDValue N0Op0 = N0.getOperand(0);
4228 if (ConstantSDNode *LargeShift = isConstOrConstSplat(N0Op0.getOperand(1))) {
4229 unsigned LargeShiftVal = LargeShift->getZExtValue();
4230 EVT LargeVT = N0Op0.getValueType();
4232 if (LargeVT.getScalarSizeInBits() - OpSizeInBits == LargeShiftVal) {
4234 DAG.getConstant(LargeShiftVal + N1C->getZExtValue(),
4235 getShiftAmountTy(N0Op0.getOperand(0).getValueType()));
4236 SDValue SRA = DAG.getNode(ISD::SRA, SDLoc(N), LargeVT,
4237 N0Op0.getOperand(0), Amt);
4238 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, SRA);
4243 // Simplify, based on bits shifted out of the LHS.
4244 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4245 return SDValue(N, 0);
4248 // If the sign bit is known to be zero, switch this to a SRL.
4249 if (DAG.SignBitIsZero(N0))
4250 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, N1);
4253 SDValue NewSRA = visitShiftByConstant(N, N1C);
4254 if (NewSRA.getNode())
4261 SDValue DAGCombiner::visitSRL(SDNode *N) {
4262 SDValue N0 = N->getOperand(0);
4263 SDValue N1 = N->getOperand(1);
4264 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4265 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4266 EVT VT = N0.getValueType();
4267 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
4270 if (VT.isVector()) {
4271 SDValue FoldedVOp = SimplifyVBinOp(N);
4272 if (FoldedVOp.getNode()) return FoldedVOp;
4274 N1C = isConstOrConstSplat(N1);
4277 // fold (srl c1, c2) -> c1 >>u c2
4279 return DAG.FoldConstantArithmetic(ISD::SRL, VT, N0C, N1C);
4280 // fold (srl 0, x) -> 0
4281 if (N0C && N0C->isNullValue())
4283 // fold (srl x, c >= size(x)) -> undef
4284 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4285 return DAG.getUNDEF(VT);
4286 // fold (srl x, 0) -> x
4287 if (N1C && N1C->isNullValue())
4289 // if (srl x, c) is known to be zero, return 0
4290 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
4291 APInt::getAllOnesValue(OpSizeInBits)))
4292 return DAG.getConstant(0, VT);
4294 // fold (srl (srl x, c1), c2) -> 0 or (srl x, (add c1, c2))
4295 if (N1C && N0.getOpcode() == ISD::SRL) {
4296 if (ConstantSDNode *N01C = isConstOrConstSplat(N0.getOperand(1))) {
4297 uint64_t c1 = N01C->getZExtValue();
4298 uint64_t c2 = N1C->getZExtValue();
4299 if (c1 + c2 >= OpSizeInBits)
4300 return DAG.getConstant(0, VT);
4301 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0),
4302 DAG.getConstant(c1 + c2, N1.getValueType()));
4306 // fold (srl (trunc (srl x, c1)), c2) -> 0 or (trunc (srl x, (add c1, c2)))
4307 if (N1C && N0.getOpcode() == ISD::TRUNCATE &&
4308 N0.getOperand(0).getOpcode() == ISD::SRL &&
4309 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
4311 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
4312 uint64_t c2 = N1C->getZExtValue();
4313 EVT InnerShiftVT = N0.getOperand(0).getValueType();
4314 EVT ShiftCountVT = N0.getOperand(0)->getOperand(1).getValueType();
4315 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
4316 // This is only valid if the OpSizeInBits + c1 = size of inner shift.
4317 if (c1 + OpSizeInBits == InnerShiftSize) {
4318 if (c1 + c2 >= InnerShiftSize)
4319 return DAG.getConstant(0, VT);
4320 return DAG.getNode(ISD::TRUNCATE, SDLoc(N0), VT,
4321 DAG.getNode(ISD::SRL, SDLoc(N0), InnerShiftVT,
4322 N0.getOperand(0)->getOperand(0),
4323 DAG.getConstant(c1 + c2, ShiftCountVT)));
4327 // fold (srl (shl x, c), c) -> (and x, cst2)
4328 if (N1C && N0.getOpcode() == ISD::SHL && N0.getOperand(1) == N1) {
4329 unsigned BitSize = N0.getScalarValueSizeInBits();
4330 if (BitSize <= 64) {
4331 uint64_t ShAmt = N1C->getZExtValue() + 64 - BitSize;
4332 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0.getOperand(0),
4333 DAG.getConstant(~0ULL >> ShAmt, VT));
4337 // fold (srl (anyextend x), c) -> (and (anyextend (srl x, c)), mask)
4338 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
4339 // Shifting in all undef bits?
4340 EVT SmallVT = N0.getOperand(0).getValueType();
4341 unsigned BitSize = SmallVT.getScalarSizeInBits();
4342 if (N1C->getZExtValue() >= BitSize)
4343 return DAG.getUNDEF(VT);
4345 if (!LegalTypes || TLI.isTypeDesirableForOp(ISD::SRL, SmallVT)) {
4346 uint64_t ShiftAmt = N1C->getZExtValue();
4347 SDValue SmallShift = DAG.getNode(ISD::SRL, SDLoc(N0), SmallVT,
4349 DAG.getConstant(ShiftAmt, getShiftAmountTy(SmallVT)));
4350 AddToWorkList(SmallShift.getNode());
4351 APInt Mask = APInt::getAllOnesValue(OpSizeInBits).lshr(ShiftAmt);
4352 return DAG.getNode(ISD::AND, SDLoc(N), VT,
4353 DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, SmallShift),
4354 DAG.getConstant(Mask, VT));
4358 // fold (srl (sra X, Y), 31) -> (srl X, 31). This srl only looks at the sign
4359 // bit, which is unmodified by sra.
4360 if (N1C && N1C->getZExtValue() + 1 == OpSizeInBits) {
4361 if (N0.getOpcode() == ISD::SRA)
4362 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0), N1);
4365 // fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit).
4366 if (N1C && N0.getOpcode() == ISD::CTLZ &&
4367 N1C->getAPIntValue() == Log2_32(OpSizeInBits)) {
4368 APInt KnownZero, KnownOne;
4369 DAG.computeKnownBits(N0.getOperand(0), KnownZero, KnownOne);
4371 // If any of the input bits are KnownOne, then the input couldn't be all
4372 // zeros, thus the result of the srl will always be zero.
4373 if (KnownOne.getBoolValue()) return DAG.getConstant(0, VT);
4375 // If all of the bits input the to ctlz node are known to be zero, then
4376 // the result of the ctlz is "32" and the result of the shift is one.
4377 APInt UnknownBits = ~KnownZero;
4378 if (UnknownBits == 0) return DAG.getConstant(1, VT);
4380 // Otherwise, check to see if there is exactly one bit input to the ctlz.
4381 if ((UnknownBits & (UnknownBits - 1)) == 0) {
4382 // Okay, we know that only that the single bit specified by UnknownBits
4383 // could be set on input to the CTLZ node. If this bit is set, the SRL
4384 // will return 0, if it is clear, it returns 1. Change the CTLZ/SRL pair
4385 // to an SRL/XOR pair, which is likely to simplify more.
4386 unsigned ShAmt = UnknownBits.countTrailingZeros();
4387 SDValue Op = N0.getOperand(0);
4390 Op = DAG.getNode(ISD::SRL, SDLoc(N0), VT, Op,
4391 DAG.getConstant(ShAmt, getShiftAmountTy(Op.getValueType())));
4392 AddToWorkList(Op.getNode());
4395 return DAG.getNode(ISD::XOR, SDLoc(N), VT,
4396 Op, DAG.getConstant(1, VT));
4400 // fold (srl x, (trunc (and y, c))) -> (srl x, (and (trunc y), (trunc c))).
4401 if (N1.getOpcode() == ISD::TRUNCATE &&
4402 N1.getOperand(0).getOpcode() == ISD::AND) {
4403 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4404 if (NewOp1.getNode())
4405 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, NewOp1);
4408 // fold operands of srl based on knowledge that the low bits are not
4410 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4411 return SDValue(N, 0);
4414 SDValue NewSRL = visitShiftByConstant(N, N1C);
4415 if (NewSRL.getNode())
4419 // Attempt to convert a srl of a load into a narrower zero-extending load.
4420 SDValue NarrowLoad = ReduceLoadWidth(N);
4421 if (NarrowLoad.getNode())
4424 // Here is a common situation. We want to optimize:
4427 // %b = and i32 %a, 2
4428 // %c = srl i32 %b, 1
4429 // brcond i32 %c ...
4435 // %c = setcc eq %b, 0
4438 // However when after the source operand of SRL is optimized into AND, the SRL
4439 // itself may not be optimized further. Look for it and add the BRCOND into
4441 if (N->hasOneUse()) {
4442 SDNode *Use = *N->use_begin();
4443 if (Use->getOpcode() == ISD::BRCOND)
4445 else if (Use->getOpcode() == ISD::TRUNCATE && Use->hasOneUse()) {
4446 // Also look pass the truncate.
4447 Use = *Use->use_begin();
4448 if (Use->getOpcode() == ISD::BRCOND)
4456 SDValue DAGCombiner::visitCTLZ(SDNode *N) {
4457 SDValue N0 = N->getOperand(0);
4458 EVT VT = N->getValueType(0);
4460 // fold (ctlz c1) -> c2
4461 if (isa<ConstantSDNode>(N0))
4462 return DAG.getNode(ISD::CTLZ, SDLoc(N), VT, N0);
4466 SDValue DAGCombiner::visitCTLZ_ZERO_UNDEF(SDNode *N) {
4467 SDValue N0 = N->getOperand(0);
4468 EVT VT = N->getValueType(0);
4470 // fold (ctlz_zero_undef c1) -> c2
4471 if (isa<ConstantSDNode>(N0))
4472 return DAG.getNode(ISD::CTLZ_ZERO_UNDEF, SDLoc(N), VT, N0);
4476 SDValue DAGCombiner::visitCTTZ(SDNode *N) {
4477 SDValue N0 = N->getOperand(0);
4478 EVT VT = N->getValueType(0);
4480 // fold (cttz c1) -> c2
4481 if (isa<ConstantSDNode>(N0))
4482 return DAG.getNode(ISD::CTTZ, SDLoc(N), VT, N0);
4486 SDValue DAGCombiner::visitCTTZ_ZERO_UNDEF(SDNode *N) {
4487 SDValue N0 = N->getOperand(0);
4488 EVT VT = N->getValueType(0);
4490 // fold (cttz_zero_undef c1) -> c2
4491 if (isa<ConstantSDNode>(N0))
4492 return DAG.getNode(ISD::CTTZ_ZERO_UNDEF, SDLoc(N), VT, N0);
4496 SDValue DAGCombiner::visitCTPOP(SDNode *N) {
4497 SDValue N0 = N->getOperand(0);
4498 EVT VT = N->getValueType(0);
4500 // fold (ctpop c1) -> c2
4501 if (isa<ConstantSDNode>(N0))
4502 return DAG.getNode(ISD::CTPOP, SDLoc(N), VT, N0);
4506 SDValue DAGCombiner::visitSELECT(SDNode *N) {
4507 SDValue N0 = N->getOperand(0);
4508 SDValue N1 = N->getOperand(1);
4509 SDValue N2 = N->getOperand(2);
4510 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4511 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4512 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
4513 EVT VT = N->getValueType(0);
4514 EVT VT0 = N0.getValueType();
4516 // fold (select C, X, X) -> X
4519 // fold (select true, X, Y) -> X
4520 if (N0C && !N0C->isNullValue())
4522 // fold (select false, X, Y) -> Y
4523 if (N0C && N0C->isNullValue())
4525 // fold (select C, 1, X) -> (or C, X)
4526 if (VT == MVT::i1 && N1C && N1C->getAPIntValue() == 1)
4527 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N2);
4528 // fold (select C, 0, 1) -> (xor C, 1)
4529 // We can't do this reliably if integer based booleans have different contents
4530 // to floating point based booleans. This is because we can't tell whether we
4531 // have an integer-based boolean or a floating-point-based boolean unless we
4532 // can find the SETCC that produced it and inspect its operands. This is
4533 // fairly easy if C is the SETCC node, but it can potentially be
4534 // undiscoverable (or not reasonably discoverable). For example, it could be
4535 // in another basic block or it could require searching a complicated
4537 if (VT.isInteger() &&
4538 (VT0 == MVT::i1 || (VT0.isInteger() &&
4539 TLI.getBooleanContents(false, false) ==
4540 TLI.getBooleanContents(false, true) &&
4541 TLI.getBooleanContents(false, false) ==
4542 TargetLowering::ZeroOrOneBooleanContent)) &&
4543 N1C && N2C && N1C->isNullValue() && N2C->getAPIntValue() == 1) {
4546 return DAG.getNode(ISD::XOR, SDLoc(N), VT0,
4547 N0, DAG.getConstant(1, VT0));
4548 XORNode = DAG.getNode(ISD::XOR, SDLoc(N0), VT0,
4549 N0, DAG.getConstant(1, VT0));
4550 AddToWorkList(XORNode.getNode());
4552 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, XORNode);
4553 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, XORNode);
4555 // fold (select C, 0, X) -> (and (not C), X)
4556 if (VT == VT0 && VT == MVT::i1 && N1C && N1C->isNullValue()) {
4557 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT);
4558 AddToWorkList(NOTNode.getNode());
4559 return DAG.getNode(ISD::AND, SDLoc(N), VT, NOTNode, N2);
4561 // fold (select C, X, 1) -> (or (not C), X)
4562 if (VT == VT0 && VT == MVT::i1 && N2C && N2C->getAPIntValue() == 1) {
4563 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT);
4564 AddToWorkList(NOTNode.getNode());
4565 return DAG.getNode(ISD::OR, SDLoc(N), VT, NOTNode, N1);
4567 // fold (select C, X, 0) -> (and C, X)
4568 if (VT == MVT::i1 && N2C && N2C->isNullValue())
4569 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, N1);
4570 // fold (select X, X, Y) -> (or X, Y)
4571 // fold (select X, 1, Y) -> (or X, Y)
4572 if (VT == MVT::i1 && (N0 == N1 || (N1C && N1C->getAPIntValue() == 1)))
4573 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N2);
4574 // fold (select X, Y, X) -> (and X, Y)
4575 // fold (select X, Y, 0) -> (and X, Y)
4576 if (VT == MVT::i1 && (N0 == N2 || (N2C && N2C->getAPIntValue() == 0)))
4577 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, N1);
4579 // If we can fold this based on the true/false value, do so.
4580 if (SimplifySelectOps(N, N1, N2))
4581 return SDValue(N, 0); // Don't revisit N.
4583 // fold selects based on a setcc into other things, such as min/max/abs
4584 if (N0.getOpcode() == ISD::SETCC) {
4585 if ((!LegalOperations &&
4586 TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT)) ||
4587 TLI.isOperationLegal(ISD::SELECT_CC, VT))
4588 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT,
4589 N0.getOperand(0), N0.getOperand(1),
4590 N1, N2, N0.getOperand(2));
4591 return SimplifySelect(SDLoc(N), N0, N1, N2);
4598 std::pair<SDValue, SDValue> SplitVSETCC(const SDNode *N, SelectionDAG &DAG) {
4601 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0));
4603 // Split the inputs.
4604 SDValue Lo, Hi, LL, LH, RL, RH;
4605 std::tie(LL, LH) = DAG.SplitVectorOperand(N, 0);
4606 std::tie(RL, RH) = DAG.SplitVectorOperand(N, 1);
4608 Lo = DAG.getNode(N->getOpcode(), DL, LoVT, LL, RL, N->getOperand(2));
4609 Hi = DAG.getNode(N->getOpcode(), DL, HiVT, LH, RH, N->getOperand(2));
4611 return std::make_pair(Lo, Hi);
4614 // This function assumes all the vselect's arguments are CONCAT_VECTOR
4615 // nodes and that the condition is a BV of ConstantSDNodes (or undefs).
4616 static SDValue ConvertSelectToConcatVector(SDNode *N, SelectionDAG &DAG) {
4618 SDValue Cond = N->getOperand(0);
4619 SDValue LHS = N->getOperand(1);
4620 SDValue RHS = N->getOperand(2);
4621 MVT VT = N->getSimpleValueType(0);
4622 int NumElems = VT.getVectorNumElements();
4623 assert(LHS.getOpcode() == ISD::CONCAT_VECTORS &&
4624 RHS.getOpcode() == ISD::CONCAT_VECTORS &&
4625 Cond.getOpcode() == ISD::BUILD_VECTOR);
4627 // We're sure we have an even number of elements due to the
4628 // concat_vectors we have as arguments to vselect.
4629 // Skip BV elements until we find one that's not an UNDEF
4630 // After we find an UNDEF element, keep looping until we get to half the
4631 // length of the BV and see if all the non-undef nodes are the same.
4632 ConstantSDNode *BottomHalf = nullptr;
4633 for (int i = 0; i < NumElems / 2; ++i) {
4634 if (Cond->getOperand(i)->getOpcode() == ISD::UNDEF)
4637 if (BottomHalf == nullptr)
4638 BottomHalf = cast<ConstantSDNode>(Cond.getOperand(i));
4639 else if (Cond->getOperand(i).getNode() != BottomHalf)
4643 // Do the same for the second half of the BuildVector
4644 ConstantSDNode *TopHalf = nullptr;
4645 for (int i = NumElems / 2; i < NumElems; ++i) {
4646 if (Cond->getOperand(i)->getOpcode() == ISD::UNDEF)
4649 if (TopHalf == nullptr)
4650 TopHalf = cast<ConstantSDNode>(Cond.getOperand(i));
4651 else if (Cond->getOperand(i).getNode() != TopHalf)
4655 assert(TopHalf && BottomHalf &&
4656 "One half of the selector was all UNDEFs and the other was all the "
4657 "same value. This should have been addressed before this function.");
4659 ISD::CONCAT_VECTORS, dl, VT,
4660 BottomHalf->isNullValue() ? RHS->getOperand(0) : LHS->getOperand(0),
4661 TopHalf->isNullValue() ? RHS->getOperand(1) : LHS->getOperand(1));
4664 SDValue DAGCombiner::visitVSELECT(SDNode *N) {
4665 SDValue N0 = N->getOperand(0);
4666 SDValue N1 = N->getOperand(1);
4667 SDValue N2 = N->getOperand(2);
4670 // Canonicalize integer abs.
4671 // vselect (setg[te] X, 0), X, -X ->
4672 // vselect (setgt X, -1), X, -X ->
4673 // vselect (setl[te] X, 0), -X, X ->
4674 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
4675 if (N0.getOpcode() == ISD::SETCC) {
4676 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
4677 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
4679 bool RHSIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
4681 if (((RHSIsAllZeros && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
4682 (ISD::isBuildVectorAllOnes(RHS.getNode()) && CC == ISD::SETGT)) &&
4683 N1 == LHS && N2.getOpcode() == ISD::SUB && N1 == N2.getOperand(1))
4684 isAbs = ISD::isBuildVectorAllZeros(N2.getOperand(0).getNode());
4685 else if ((RHSIsAllZeros && (CC == ISD::SETLT || CC == ISD::SETLE)) &&
4686 N2 == LHS && N1.getOpcode() == ISD::SUB && N2 == N1.getOperand(1))
4687 isAbs = ISD::isBuildVectorAllZeros(N1.getOperand(0).getNode());
4690 EVT VT = LHS.getValueType();
4691 SDValue Shift = DAG.getNode(
4692 ISD::SRA, DL, VT, LHS,
4693 DAG.getConstant(VT.getScalarType().getSizeInBits() - 1, VT));
4694 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, LHS, Shift);
4695 AddToWorkList(Shift.getNode());
4696 AddToWorkList(Add.getNode());
4697 return DAG.getNode(ISD::XOR, DL, VT, Add, Shift);
4701 // If the VSELECT result requires splitting and the mask is provided by a
4702 // SETCC, then split both nodes and its operands before legalization. This
4703 // prevents the type legalizer from unrolling SETCC into scalar comparisons
4704 // and enables future optimizations (e.g. min/max pattern matching on X86).
4705 if (N0.getOpcode() == ISD::SETCC) {
4706 EVT VT = N->getValueType(0);
4708 // Check if any splitting is required.
4709 if (TLI.getTypeAction(*DAG.getContext(), VT) !=
4710 TargetLowering::TypeSplitVector)
4713 SDValue Lo, Hi, CCLo, CCHi, LL, LH, RL, RH;
4714 std::tie(CCLo, CCHi) = SplitVSETCC(N0.getNode(), DAG);
4715 std::tie(LL, LH) = DAG.SplitVectorOperand(N, 1);
4716 std::tie(RL, RH) = DAG.SplitVectorOperand(N, 2);
4718 Lo = DAG.getNode(N->getOpcode(), DL, LL.getValueType(), CCLo, LL, RL);
4719 Hi = DAG.getNode(N->getOpcode(), DL, LH.getValueType(), CCHi, LH, RH);
4721 // Add the new VSELECT nodes to the work list in case they need to be split
4723 AddToWorkList(Lo.getNode());
4724 AddToWorkList(Hi.getNode());
4726 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
4729 // Fold (vselect (build_vector all_ones), N1, N2) -> N1
4730 if (ISD::isBuildVectorAllOnes(N0.getNode()))
4732 // Fold (vselect (build_vector all_zeros), N1, N2) -> N2
4733 if (ISD::isBuildVectorAllZeros(N0.getNode()))
4736 // The ConvertSelectToConcatVector function is assuming both the above
4737 // checks for (vselect (build_vector all{ones,zeros) ...) have been made
4739 if (N1.getOpcode() == ISD::CONCAT_VECTORS &&
4740 N2.getOpcode() == ISD::CONCAT_VECTORS &&
4741 ISD::isBuildVectorOfConstantSDNodes(N0.getNode())) {
4742 SDValue CV = ConvertSelectToConcatVector(N, DAG);
4750 SDValue DAGCombiner::visitSELECT_CC(SDNode *N) {
4751 SDValue N0 = N->getOperand(0);
4752 SDValue N1 = N->getOperand(1);
4753 SDValue N2 = N->getOperand(2);
4754 SDValue N3 = N->getOperand(3);
4755 SDValue N4 = N->getOperand(4);
4756 ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get();
4758 // fold select_cc lhs, rhs, x, x, cc -> x
4762 // Determine if the condition we're dealing with is constant
4763 SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()),
4764 N0, N1, CC, SDLoc(N), false);
4765 if (SCC.getNode()) {
4766 AddToWorkList(SCC.getNode());
4768 if (ConstantSDNode *SCCC = dyn_cast<ConstantSDNode>(SCC.getNode())) {
4769 if (!SCCC->isNullValue())
4770 return N2; // cond always true -> true val
4772 return N3; // cond always false -> false val
4775 // Fold to a simpler select_cc
4776 if (SCC.getOpcode() == ISD::SETCC)
4777 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), N2.getValueType(),
4778 SCC.getOperand(0), SCC.getOperand(1), N2, N3,
4782 // If we can fold this based on the true/false value, do so.
4783 if (SimplifySelectOps(N, N2, N3))
4784 return SDValue(N, 0); // Don't revisit N.
4786 // fold select_cc into other things, such as min/max/abs
4787 return SimplifySelectCC(SDLoc(N), N0, N1, N2, N3, CC);
4790 SDValue DAGCombiner::visitSETCC(SDNode *N) {
4791 return SimplifySetCC(N->getValueType(0), N->getOperand(0), N->getOperand(1),
4792 cast<CondCodeSDNode>(N->getOperand(2))->get(),
4796 // tryToFoldExtendOfConstant - Try to fold a sext/zext/aext
4797 // dag node into a ConstantSDNode or a build_vector of constants.
4798 // This function is called by the DAGCombiner when visiting sext/zext/aext
4799 // dag nodes (see for example method DAGCombiner::visitSIGN_EXTEND).
4800 // Vector extends are not folded if operations are legal; this is to
4801 // avoid introducing illegal build_vector dag nodes.
4802 static SDNode *tryToFoldExtendOfConstant(SDNode *N, const TargetLowering &TLI,
4803 SelectionDAG &DAG, bool LegalTypes,
4804 bool LegalOperations) {
4805 unsigned Opcode = N->getOpcode();
4806 SDValue N0 = N->getOperand(0);
4807 EVT VT = N->getValueType(0);
4809 assert((Opcode == ISD::SIGN_EXTEND || Opcode == ISD::ZERO_EXTEND ||
4810 Opcode == ISD::ANY_EXTEND) && "Expected EXTEND dag node in input!");
4812 // fold (sext c1) -> c1
4813 // fold (zext c1) -> c1
4814 // fold (aext c1) -> c1
4815 if (isa<ConstantSDNode>(N0))
4816 return DAG.getNode(Opcode, SDLoc(N), VT, N0).getNode();
4818 // fold (sext (build_vector AllConstants) -> (build_vector AllConstants)
4819 // fold (zext (build_vector AllConstants) -> (build_vector AllConstants)
4820 // fold (aext (build_vector AllConstants) -> (build_vector AllConstants)
4821 EVT SVT = VT.getScalarType();
4822 if (!(VT.isVector() &&
4823 (!LegalTypes || (!LegalOperations && TLI.isTypeLegal(SVT))) &&
4824 ISD::isBuildVectorOfConstantSDNodes(N0.getNode())))
4827 // We can fold this node into a build_vector.
4828 unsigned VTBits = SVT.getSizeInBits();
4829 unsigned EVTBits = N0->getValueType(0).getScalarType().getSizeInBits();
4830 unsigned ShAmt = VTBits - EVTBits;
4831 SmallVector<SDValue, 8> Elts;
4832 unsigned NumElts = N0->getNumOperands();
4835 for (unsigned i=0; i != NumElts; ++i) {
4836 SDValue Op = N0->getOperand(i);
4837 if (Op->getOpcode() == ISD::UNDEF) {
4838 Elts.push_back(DAG.getUNDEF(SVT));
4842 ConstantSDNode *CurrentND = cast<ConstantSDNode>(Op);
4843 const APInt &C = APInt(VTBits, CurrentND->getAPIntValue().getZExtValue());
4844 if (Opcode == ISD::SIGN_EXTEND)
4845 Elts.push_back(DAG.getConstant(C.shl(ShAmt).ashr(ShAmt).getZExtValue(),
4848 Elts.push_back(DAG.getConstant(C.shl(ShAmt).lshr(ShAmt).getZExtValue(),
4852 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Elts).getNode();
4855 // ExtendUsesToFormExtLoad - Trying to extend uses of a load to enable this:
4856 // "fold ({s|z|a}ext (load x)) -> ({s|z|a}ext (truncate ({s|z|a}extload x)))"
4857 // transformation. Returns true if extension are possible and the above
4858 // mentioned transformation is profitable.
4859 static bool ExtendUsesToFormExtLoad(SDNode *N, SDValue N0,
4861 SmallVectorImpl<SDNode *> &ExtendNodes,
4862 const TargetLowering &TLI) {
4863 bool HasCopyToRegUses = false;
4864 bool isTruncFree = TLI.isTruncateFree(N->getValueType(0), N0.getValueType());
4865 for (SDNode::use_iterator UI = N0.getNode()->use_begin(),
4866 UE = N0.getNode()->use_end();
4871 if (UI.getUse().getResNo() != N0.getResNo())
4873 // FIXME: Only extend SETCC N, N and SETCC N, c for now.
4874 if (ExtOpc != ISD::ANY_EXTEND && User->getOpcode() == ISD::SETCC) {
4875 ISD::CondCode CC = cast<CondCodeSDNode>(User->getOperand(2))->get();
4876 if (ExtOpc == ISD::ZERO_EXTEND && ISD::isSignedIntSetCC(CC))
4877 // Sign bits will be lost after a zext.
4880 for (unsigned i = 0; i != 2; ++i) {
4881 SDValue UseOp = User->getOperand(i);
4884 if (!isa<ConstantSDNode>(UseOp))
4889 ExtendNodes.push_back(User);
4892 // If truncates aren't free and there are users we can't
4893 // extend, it isn't worthwhile.
4896 // Remember if this value is live-out.
4897 if (User->getOpcode() == ISD::CopyToReg)
4898 HasCopyToRegUses = true;
4901 if (HasCopyToRegUses) {
4902 bool BothLiveOut = false;
4903 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
4905 SDUse &Use = UI.getUse();
4906 if (Use.getResNo() == 0 && Use.getUser()->getOpcode() == ISD::CopyToReg) {
4912 // Both unextended and extended values are live out. There had better be
4913 // a good reason for the transformation.
4914 return ExtendNodes.size();
4919 void DAGCombiner::ExtendSetCCUses(const SmallVectorImpl<SDNode *> &SetCCs,
4920 SDValue Trunc, SDValue ExtLoad, SDLoc DL,
4921 ISD::NodeType ExtType) {
4922 // Extend SetCC uses if necessary.
4923 for (unsigned i = 0, e = SetCCs.size(); i != e; ++i) {
4924 SDNode *SetCC = SetCCs[i];
4925 SmallVector<SDValue, 4> Ops;
4927 for (unsigned j = 0; j != 2; ++j) {
4928 SDValue SOp = SetCC->getOperand(j);
4930 Ops.push_back(ExtLoad);
4932 Ops.push_back(DAG.getNode(ExtType, DL, ExtLoad->getValueType(0), SOp));
4935 Ops.push_back(SetCC->getOperand(2));
4936 CombineTo(SetCC, DAG.getNode(ISD::SETCC, DL, SetCC->getValueType(0), Ops));
4940 SDValue DAGCombiner::visitSIGN_EXTEND(SDNode *N) {
4941 SDValue N0 = N->getOperand(0);
4942 EVT VT = N->getValueType(0);
4944 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
4946 return SDValue(Res, 0);
4948 // fold (sext (sext x)) -> (sext x)
4949 // fold (sext (aext x)) -> (sext x)
4950 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
4951 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT,
4954 if (N0.getOpcode() == ISD::TRUNCATE) {
4955 // fold (sext (truncate (load x))) -> (sext (smaller load x))
4956 // fold (sext (truncate (srl (load x), c))) -> (sext (smaller load (x+c/n)))
4957 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
4958 if (NarrowLoad.getNode()) {
4959 SDNode* oye = N0.getNode()->getOperand(0).getNode();
4960 if (NarrowLoad.getNode() != N0.getNode()) {
4961 CombineTo(N0.getNode(), NarrowLoad);
4962 // CombineTo deleted the truncate, if needed, but not what's under it.
4965 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4968 // See if the value being truncated is already sign extended. If so, just
4969 // eliminate the trunc/sext pair.
4970 SDValue Op = N0.getOperand(0);
4971 unsigned OpBits = Op.getValueType().getScalarType().getSizeInBits();
4972 unsigned MidBits = N0.getValueType().getScalarType().getSizeInBits();
4973 unsigned DestBits = VT.getScalarType().getSizeInBits();
4974 unsigned NumSignBits = DAG.ComputeNumSignBits(Op);
4976 if (OpBits == DestBits) {
4977 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
4978 // bits, it is already ready.
4979 if (NumSignBits > DestBits-MidBits)
4981 } else if (OpBits < DestBits) {
4982 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
4983 // bits, just sext from i32.
4984 if (NumSignBits > OpBits-MidBits)
4985 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, Op);
4987 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
4988 // bits, just truncate to i32.
4989 if (NumSignBits > OpBits-MidBits)
4990 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
4993 // fold (sext (truncate x)) -> (sextinreg x).
4994 if (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG,
4995 N0.getValueType())) {
4996 if (OpBits < DestBits)
4997 Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N0), VT, Op);
4998 else if (OpBits > DestBits)
4999 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), VT, Op);
5000 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, Op,
5001 DAG.getValueType(N0.getValueType()));
5005 // fold (sext (load x)) -> (sext (truncate (sextload x)))
5006 // None of the supported targets knows how to perform load and sign extend
5007 // on vectors in one instruction. We only perform this transformation on
5009 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
5010 ISD::isUNINDEXEDLoad(N0.getNode()) &&
5011 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5012 TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()))) {
5013 bool DoXform = true;
5014 SmallVector<SDNode*, 4> SetCCs;
5015 if (!N0.hasOneUse())
5016 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::SIGN_EXTEND, SetCCs, TLI);
5018 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5019 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5021 LN0->getBasePtr(), N0.getValueType(),
5022 LN0->getMemOperand());
5023 CombineTo(N, ExtLoad);
5024 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5025 N0.getValueType(), ExtLoad);
5026 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
5027 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5029 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5033 // fold (sext (sextload x)) -> (sext (truncate (sextload x)))
5034 // fold (sext ( extload x)) -> (sext (truncate (sextload x)))
5035 if ((ISD::isSEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
5036 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
5037 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5038 EVT MemVT = LN0->getMemoryVT();
5039 if ((!LegalOperations && !LN0->isVolatile()) ||
5040 TLI.isLoadExtLegal(ISD::SEXTLOAD, MemVT)) {
5041 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5043 LN0->getBasePtr(), MemVT,
5044 LN0->getMemOperand());
5045 CombineTo(N, ExtLoad);
5046 CombineTo(N0.getNode(),
5047 DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5048 N0.getValueType(), ExtLoad),
5049 ExtLoad.getValue(1));
5050 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5054 // fold (sext (and/or/xor (load x), cst)) ->
5055 // (and/or/xor (sextload x), (sext cst))
5056 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
5057 N0.getOpcode() == ISD::XOR) &&
5058 isa<LoadSDNode>(N0.getOperand(0)) &&
5059 N0.getOperand(1).getOpcode() == ISD::Constant &&
5060 TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()) &&
5061 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
5062 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
5063 if (LN0->getExtensionType() != ISD::ZEXTLOAD && LN0->isUnindexed()) {
5064 bool DoXform = true;
5065 SmallVector<SDNode*, 4> SetCCs;
5066 if (!N0.hasOneUse())
5067 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::SIGN_EXTEND,
5070 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(LN0), VT,
5071 LN0->getChain(), LN0->getBasePtr(),
5073 LN0->getMemOperand());
5074 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5075 Mask = Mask.sext(VT.getSizeInBits());
5076 SDValue And = DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
5077 ExtLoad, DAG.getConstant(Mask, VT));
5078 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
5079 SDLoc(N0.getOperand(0)),
5080 N0.getOperand(0).getValueType(), ExtLoad);
5082 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
5083 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5085 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5090 if (N0.getOpcode() == ISD::SETCC) {
5091 EVT N0VT = N0.getOperand(0).getValueType();
5092 // sext(setcc) -> sext_in_reg(vsetcc) for vectors.
5093 // Only do this before legalize for now.
5094 if (VT.isVector() && !LegalOperations &&
5095 TLI.getBooleanContents(N0VT) ==
5096 TargetLowering::ZeroOrNegativeOneBooleanContent) {
5097 // On some architectures (such as SSE/NEON/etc) the SETCC result type is
5098 // of the same size as the compared operands. Only optimize sext(setcc())
5099 // if this is the case.
5100 EVT SVT = getSetCCResultType(N0VT);
5102 // We know that the # elements of the results is the same as the
5103 // # elements of the compare (and the # elements of the compare result
5104 // for that matter). Check to see that they are the same size. If so,
5105 // we know that the element size of the sext'd result matches the
5106 // element size of the compare operands.
5107 if (VT.getSizeInBits() == SVT.getSizeInBits())
5108 return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
5110 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5112 // If the desired elements are smaller or larger than the source
5113 // elements we can use a matching integer vector type and then
5114 // truncate/sign extend
5115 EVT MatchingVectorType = N0VT.changeVectorElementTypeToInteger();
5116 if (SVT == MatchingVectorType) {
5117 SDValue VsetCC = DAG.getSetCC(SDLoc(N), MatchingVectorType,
5118 N0.getOperand(0), N0.getOperand(1),
5119 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5120 return DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT);
5124 // sext(setcc x, y, cc) -> (select (setcc x, y, cc), -1, 0)
5125 unsigned ElementWidth = VT.getScalarType().getSizeInBits();
5127 DAG.getConstant(APInt::getAllOnesValue(ElementWidth), VT);
5129 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
5130 NegOne, DAG.getConstant(0, VT),
5131 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
5132 if (SCC.getNode()) return SCC;
5134 if (!VT.isVector()) {
5135 EVT SetCCVT = getSetCCResultType(N0.getOperand(0).getValueType());
5136 if (!LegalOperations || TLI.isOperationLegal(ISD::SETCC, SetCCVT)) {
5138 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
5139 SDValue SetCC = DAG.getSetCC(DL,
5141 N0.getOperand(0), N0.getOperand(1), CC);
5142 EVT SelectVT = getSetCCResultType(VT);
5143 return DAG.getSelect(DL, VT,
5144 DAG.getSExtOrTrunc(SetCC, DL, SelectVT),
5145 NegOne, DAG.getConstant(0, VT));
5151 // fold (sext x) -> (zext x) if the sign bit is known zero.
5152 if ((!LegalOperations || TLI.isOperationLegal(ISD::ZERO_EXTEND, VT)) &&
5153 DAG.SignBitIsZero(N0))
5154 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, N0);
5159 // isTruncateOf - If N is a truncate of some other value, return true, record
5160 // the value being truncated in Op and which of Op's bits are zero in KnownZero.
5161 // This function computes KnownZero to avoid a duplicated call to
5162 // computeKnownBits in the caller.
5163 static bool isTruncateOf(SelectionDAG &DAG, SDValue N, SDValue &Op,
5166 if (N->getOpcode() == ISD::TRUNCATE) {
5167 Op = N->getOperand(0);
5168 DAG.computeKnownBits(Op, KnownZero, KnownOne);
5172 if (N->getOpcode() != ISD::SETCC || N->getValueType(0) != MVT::i1 ||
5173 cast<CondCodeSDNode>(N->getOperand(2))->get() != ISD::SETNE)
5176 SDValue Op0 = N->getOperand(0);
5177 SDValue Op1 = N->getOperand(1);
5178 assert(Op0.getValueType() == Op1.getValueType());
5180 ConstantSDNode *COp0 = dyn_cast<ConstantSDNode>(Op0);
5181 ConstantSDNode *COp1 = dyn_cast<ConstantSDNode>(Op1);
5182 if (COp0 && COp0->isNullValue())
5184 else if (COp1 && COp1->isNullValue())
5189 DAG.computeKnownBits(Op, KnownZero, KnownOne);
5191 if (!(KnownZero | APInt(Op.getValueSizeInBits(), 1)).isAllOnesValue())
5197 SDValue DAGCombiner::visitZERO_EXTEND(SDNode *N) {
5198 SDValue N0 = N->getOperand(0);
5199 EVT VT = N->getValueType(0);
5201 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
5203 return SDValue(Res, 0);
5205 // fold (zext (zext x)) -> (zext x)
5206 // fold (zext (aext x)) -> (zext x)
5207 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
5208 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT,
5211 // fold (zext (truncate x)) -> (zext x) or
5212 // (zext (truncate x)) -> (truncate x)
5213 // This is valid when the truncated bits of x are already zero.
5214 // FIXME: We should extend this to work for vectors too.
5217 if (!VT.isVector() && isTruncateOf(DAG, N0, Op, KnownZero)) {
5218 APInt TruncatedBits =
5219 (Op.getValueSizeInBits() == N0.getValueSizeInBits()) ?
5220 APInt(Op.getValueSizeInBits(), 0) :
5221 APInt::getBitsSet(Op.getValueSizeInBits(),
5222 N0.getValueSizeInBits(),
5223 std::min(Op.getValueSizeInBits(),
5224 VT.getSizeInBits()));
5225 if (TruncatedBits == (KnownZero & TruncatedBits)) {
5226 if (VT.bitsGT(Op.getValueType()))
5227 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, Op);
5228 if (VT.bitsLT(Op.getValueType()))
5229 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
5235 // fold (zext (truncate (load x))) -> (zext (smaller load x))
5236 // fold (zext (truncate (srl (load x), c))) -> (zext (small load (x+c/n)))
5237 if (N0.getOpcode() == ISD::TRUNCATE) {
5238 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5239 if (NarrowLoad.getNode()) {
5240 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5241 if (NarrowLoad.getNode() != N0.getNode()) {
5242 CombineTo(N0.getNode(), NarrowLoad);
5243 // CombineTo deleted the truncate, if needed, but not what's under it.
5246 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5250 // fold (zext (truncate x)) -> (and x, mask)
5251 if (N0.getOpcode() == ISD::TRUNCATE &&
5252 (!LegalOperations || TLI.isOperationLegal(ISD::AND, VT))) {
5254 // fold (zext (truncate (load x))) -> (zext (smaller load x))
5255 // fold (zext (truncate (srl (load x), c))) -> (zext (smaller load (x+c/n)))
5256 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5257 if (NarrowLoad.getNode()) {
5258 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5259 if (NarrowLoad.getNode() != N0.getNode()) {
5260 CombineTo(N0.getNode(), NarrowLoad);
5261 // CombineTo deleted the truncate, if needed, but not what's under it.
5264 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5267 SDValue Op = N0.getOperand(0);
5268 if (Op.getValueType().bitsLT(VT)) {
5269 Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, Op);
5270 AddToWorkList(Op.getNode());
5271 } else if (Op.getValueType().bitsGT(VT)) {
5272 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
5273 AddToWorkList(Op.getNode());
5275 return DAG.getZeroExtendInReg(Op, SDLoc(N),
5276 N0.getValueType().getScalarType());
5279 // Fold (zext (and (trunc x), cst)) -> (and x, cst),
5280 // if either of the casts is not free.
5281 if (N0.getOpcode() == ISD::AND &&
5282 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
5283 N0.getOperand(1).getOpcode() == ISD::Constant &&
5284 (!TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
5285 N0.getValueType()) ||
5286 !TLI.isZExtFree(N0.getValueType(), VT))) {
5287 SDValue X = N0.getOperand(0).getOperand(0);
5288 if (X.getValueType().bitsLT(VT)) {
5289 X = DAG.getNode(ISD::ANY_EXTEND, SDLoc(X), VT, X);
5290 } else if (X.getValueType().bitsGT(VT)) {
5291 X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X);
5293 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5294 Mask = Mask.zext(VT.getSizeInBits());
5295 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5296 X, DAG.getConstant(Mask, VT));
5299 // fold (zext (load x)) -> (zext (truncate (zextload x)))
5300 // None of the supported targets knows how to perform load and vector_zext
5301 // on vectors in one instruction. We only perform this transformation on
5303 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
5304 ISD::isUNINDEXEDLoad(N0.getNode()) &&
5305 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5306 TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()))) {
5307 bool DoXform = true;
5308 SmallVector<SDNode*, 4> SetCCs;
5309 if (!N0.hasOneUse())
5310 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ZERO_EXTEND, SetCCs, TLI);
5312 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5313 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT,
5315 LN0->getBasePtr(), N0.getValueType(),
5316 LN0->getMemOperand());
5317 CombineTo(N, ExtLoad);
5318 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5319 N0.getValueType(), ExtLoad);
5320 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
5322 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5324 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5328 // fold (zext (and/or/xor (load x), cst)) ->
5329 // (and/or/xor (zextload x), (zext cst))
5330 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
5331 N0.getOpcode() == ISD::XOR) &&
5332 isa<LoadSDNode>(N0.getOperand(0)) &&
5333 N0.getOperand(1).getOpcode() == ISD::Constant &&
5334 TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()) &&
5335 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
5336 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
5337 if (LN0->getExtensionType() != ISD::SEXTLOAD && LN0->isUnindexed()) {
5338 bool DoXform = true;
5339 SmallVector<SDNode*, 4> SetCCs;
5340 if (!N0.hasOneUse())
5341 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::ZERO_EXTEND,
5344 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), VT,
5345 LN0->getChain(), LN0->getBasePtr(),
5347 LN0->getMemOperand());
5348 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5349 Mask = Mask.zext(VT.getSizeInBits());
5350 SDValue And = DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
5351 ExtLoad, DAG.getConstant(Mask, VT));
5352 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
5353 SDLoc(N0.getOperand(0)),
5354 N0.getOperand(0).getValueType(), ExtLoad);
5356 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
5357 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5359 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5364 // fold (zext (zextload x)) -> (zext (truncate (zextload x)))
5365 // fold (zext ( extload x)) -> (zext (truncate (zextload x)))
5366 if ((ISD::isZEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
5367 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
5368 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5369 EVT MemVT = LN0->getMemoryVT();
5370 if ((!LegalOperations && !LN0->isVolatile()) ||
5371 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT)) {
5372 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT,
5374 LN0->getBasePtr(), MemVT,
5375 LN0->getMemOperand());
5376 CombineTo(N, ExtLoad);
5377 CombineTo(N0.getNode(),
5378 DAG.getNode(ISD::TRUNCATE, SDLoc(N0), N0.getValueType(),
5380 ExtLoad.getValue(1));
5381 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5385 if (N0.getOpcode() == ISD::SETCC) {
5386 if (!LegalOperations && VT.isVector() &&
5387 N0.getValueType().getVectorElementType() == MVT::i1) {
5388 EVT N0VT = N0.getOperand(0).getValueType();
5389 if (getSetCCResultType(N0VT) == N0.getValueType())
5392 // zext(setcc) -> (and (vsetcc), (1, 1, ...) for vectors.
5393 // Only do this before legalize for now.
5394 EVT EltVT = VT.getVectorElementType();
5395 SmallVector<SDValue,8> OneOps(VT.getVectorNumElements(),
5396 DAG.getConstant(1, EltVT));
5397 if (VT.getSizeInBits() == N0VT.getSizeInBits())
5398 // We know that the # elements of the results is the same as the
5399 // # elements of the compare (and the # elements of the compare result
5400 // for that matter). Check to see that they are the same size. If so,
5401 // we know that the element size of the sext'd result matches the
5402 // element size of the compare operands.
5403 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5404 DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
5406 cast<CondCodeSDNode>(N0.getOperand(2))->get()),
5407 DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT,
5410 // If the desired elements are smaller or larger than the source
5411 // elements we can use a matching integer vector type and then
5412 // truncate/sign extend
5413 EVT MatchingElementType =
5414 EVT::getIntegerVT(*DAG.getContext(),
5415 N0VT.getScalarType().getSizeInBits());
5416 EVT MatchingVectorType =
5417 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
5418 N0VT.getVectorNumElements());
5420 DAG.getSetCC(SDLoc(N), MatchingVectorType, N0.getOperand(0),
5422 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5423 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5424 DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT),
5425 DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, OneOps));
5428 // zext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
5430 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
5431 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
5432 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
5433 if (SCC.getNode()) return SCC;
5436 // (zext (shl (zext x), cst)) -> (shl (zext x), cst)
5437 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL) &&
5438 isa<ConstantSDNode>(N0.getOperand(1)) &&
5439 N0.getOperand(0).getOpcode() == ISD::ZERO_EXTEND &&
5441 SDValue ShAmt = N0.getOperand(1);
5442 unsigned ShAmtVal = cast<ConstantSDNode>(ShAmt)->getZExtValue();
5443 if (N0.getOpcode() == ISD::SHL) {
5444 SDValue InnerZExt = N0.getOperand(0);
5445 // If the original shl may be shifting out bits, do not perform this
5447 unsigned KnownZeroBits = InnerZExt.getValueType().getSizeInBits() -
5448 InnerZExt.getOperand(0).getValueType().getSizeInBits();
5449 if (ShAmtVal > KnownZeroBits)
5455 // Ensure that the shift amount is wide enough for the shifted value.
5456 if (VT.getSizeInBits() >= 256)
5457 ShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, ShAmt);
5459 return DAG.getNode(N0.getOpcode(), DL, VT,
5460 DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0)),
5467 SDValue DAGCombiner::visitANY_EXTEND(SDNode *N) {
5468 SDValue N0 = N->getOperand(0);
5469 EVT VT = N->getValueType(0);
5471 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
5473 return SDValue(Res, 0);
5475 // fold (aext (aext x)) -> (aext x)
5476 // fold (aext (zext x)) -> (zext x)
5477 // fold (aext (sext x)) -> (sext x)
5478 if (N0.getOpcode() == ISD::ANY_EXTEND ||
5479 N0.getOpcode() == ISD::ZERO_EXTEND ||
5480 N0.getOpcode() == ISD::SIGN_EXTEND)
5481 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, N0.getOperand(0));
5483 // fold (aext (truncate (load x))) -> (aext (smaller load x))
5484 // fold (aext (truncate (srl (load x), c))) -> (aext (small load (x+c/n)))
5485 if (N0.getOpcode() == ISD::TRUNCATE) {
5486 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5487 if (NarrowLoad.getNode()) {
5488 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5489 if (NarrowLoad.getNode() != N0.getNode()) {
5490 CombineTo(N0.getNode(), NarrowLoad);
5491 // CombineTo deleted the truncate, if needed, but not what's under it.
5494 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5498 // fold (aext (truncate x))
5499 if (N0.getOpcode() == ISD::TRUNCATE) {
5500 SDValue TruncOp = N0.getOperand(0);
5501 if (TruncOp.getValueType() == VT)
5502 return TruncOp; // x iff x size == zext size.
5503 if (TruncOp.getValueType().bitsGT(VT))
5504 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, TruncOp);
5505 return DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, TruncOp);
5508 // Fold (aext (and (trunc x), cst)) -> (and x, cst)
5509 // if the trunc is not free.
5510 if (N0.getOpcode() == ISD::AND &&
5511 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
5512 N0.getOperand(1).getOpcode() == ISD::Constant &&
5513 !TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
5514 N0.getValueType())) {
5515 SDValue X = N0.getOperand(0).getOperand(0);
5516 if (X.getValueType().bitsLT(VT)) {
5517 X = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, X);
5518 } else if (X.getValueType().bitsGT(VT)) {
5519 X = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, X);
5521 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5522 Mask = Mask.zext(VT.getSizeInBits());
5523 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5524 X, DAG.getConstant(Mask, VT));
5527 // fold (aext (load x)) -> (aext (truncate (extload x)))
5528 // None of the supported targets knows how to perform load and any_ext
5529 // on vectors in one instruction. We only perform this transformation on
5531 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
5532 ISD::isUNINDEXEDLoad(N0.getNode()) &&
5533 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5534 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType()))) {
5535 bool DoXform = true;
5536 SmallVector<SDNode*, 4> SetCCs;
5537 if (!N0.hasOneUse())
5538 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ANY_EXTEND, SetCCs, TLI);
5540 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5541 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT,
5543 LN0->getBasePtr(), N0.getValueType(),
5544 LN0->getMemOperand());
5545 CombineTo(N, ExtLoad);
5546 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5547 N0.getValueType(), ExtLoad);
5548 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
5549 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5551 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5555 // fold (aext (zextload x)) -> (aext (truncate (zextload x)))
5556 // fold (aext (sextload x)) -> (aext (truncate (sextload x)))
5557 // fold (aext ( extload x)) -> (aext (truncate (extload x)))
5558 if (N0.getOpcode() == ISD::LOAD &&
5559 !ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
5561 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5562 ISD::LoadExtType ExtType = LN0->getExtensionType();
5563 EVT MemVT = LN0->getMemoryVT();
5564 if (!LegalOperations || TLI.isLoadExtLegal(ExtType, MemVT)) {
5565 SDValue ExtLoad = DAG.getExtLoad(ExtType, SDLoc(N),
5566 VT, LN0->getChain(), LN0->getBasePtr(),
5567 MemVT, LN0->getMemOperand());
5568 CombineTo(N, ExtLoad);
5569 CombineTo(N0.getNode(),
5570 DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5571 N0.getValueType(), ExtLoad),
5572 ExtLoad.getValue(1));
5573 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5577 if (N0.getOpcode() == ISD::SETCC) {
5579 // aext(setcc) -> vsetcc
5580 // aext(setcc) -> truncate(vsetcc)
5581 // aext(setcc) -> aext(vsetcc)
5582 // Only do this before legalize for now.
5583 if (VT.isVector() && !LegalOperations) {
5584 EVT N0VT = N0.getOperand(0).getValueType();
5585 // We know that the # elements of the results is the same as the
5586 // # elements of the compare (and the # elements of the compare result
5587 // for that matter). Check to see that they are the same size. If so,
5588 // we know that the element size of the sext'd result matches the
5589 // element size of the compare operands.
5590 if (VT.getSizeInBits() == N0VT.getSizeInBits())
5591 return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
5593 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5594 // If the desired elements are smaller or larger than the source
5595 // elements we can use a matching integer vector type and then
5596 // truncate/any extend
5598 EVT MatchingVectorType = N0VT.changeVectorElementTypeToInteger();
5600 DAG.getSetCC(SDLoc(N), MatchingVectorType, N0.getOperand(0),
5602 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5603 return DAG.getAnyExtOrTrunc(VsetCC, SDLoc(N), VT);
5607 // aext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
5609 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
5610 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
5611 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
5619 /// GetDemandedBits - See if the specified operand can be simplified with the
5620 /// knowledge that only the bits specified by Mask are used. If so, return the
5621 /// simpler operand, otherwise return a null SDValue.
5622 SDValue DAGCombiner::GetDemandedBits(SDValue V, const APInt &Mask) {
5623 switch (V.getOpcode()) {
5625 case ISD::Constant: {
5626 const ConstantSDNode *CV = cast<ConstantSDNode>(V.getNode());
5627 assert(CV && "Const value should be ConstSDNode.");
5628 const APInt &CVal = CV->getAPIntValue();
5629 APInt NewVal = CVal & Mask;
5631 return DAG.getConstant(NewVal, V.getValueType());
5636 // If the LHS or RHS don't contribute bits to the or, drop them.
5637 if (DAG.MaskedValueIsZero(V.getOperand(0), Mask))
5638 return V.getOperand(1);
5639 if (DAG.MaskedValueIsZero(V.getOperand(1), Mask))
5640 return V.getOperand(0);
5643 // Only look at single-use SRLs.
5644 if (!V.getNode()->hasOneUse())
5646 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(V.getOperand(1))) {
5647 // See if we can recursively simplify the LHS.
5648 unsigned Amt = RHSC->getZExtValue();
5650 // Watch out for shift count overflow though.
5651 if (Amt >= Mask.getBitWidth()) break;
5652 APInt NewMask = Mask << Amt;
5653 SDValue SimplifyLHS = GetDemandedBits(V.getOperand(0), NewMask);
5654 if (SimplifyLHS.getNode())
5655 return DAG.getNode(ISD::SRL, SDLoc(V), V.getValueType(),
5656 SimplifyLHS, V.getOperand(1));
5662 /// ReduceLoadWidth - If the result of a wider load is shifted to right of N
5663 /// bits and then truncated to a narrower type and where N is a multiple
5664 /// of number of bits of the narrower type, transform it to a narrower load
5665 /// from address + N / num of bits of new type. If the result is to be
5666 /// extended, also fold the extension to form a extending load.
5667 SDValue DAGCombiner::ReduceLoadWidth(SDNode *N) {
5668 unsigned Opc = N->getOpcode();
5670 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
5671 SDValue N0 = N->getOperand(0);
5672 EVT VT = N->getValueType(0);
5675 // This transformation isn't valid for vector loads.
5679 // Special case: SIGN_EXTEND_INREG is basically truncating to ExtVT then
5681 if (Opc == ISD::SIGN_EXTEND_INREG) {
5682 ExtType = ISD::SEXTLOAD;
5683 ExtVT = cast<VTSDNode>(N->getOperand(1))->getVT();
5684 } else if (Opc == ISD::SRL) {
5685 // Another special-case: SRL is basically zero-extending a narrower value.
5686 ExtType = ISD::ZEXTLOAD;
5688 ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1));
5689 if (!N01) return SDValue();
5690 ExtVT = EVT::getIntegerVT(*DAG.getContext(),
5691 VT.getSizeInBits() - N01->getZExtValue());
5693 if (LegalOperations && !TLI.isLoadExtLegal(ExtType, ExtVT))
5696 unsigned EVTBits = ExtVT.getSizeInBits();
5698 // Do not generate loads of non-round integer types since these can
5699 // be expensive (and would be wrong if the type is not byte sized).
5700 if (!ExtVT.isRound())
5704 if (N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
5705 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
5706 ShAmt = N01->getZExtValue();
5707 // Is the shift amount a multiple of size of VT?
5708 if ((ShAmt & (EVTBits-1)) == 0) {
5709 N0 = N0.getOperand(0);
5710 // Is the load width a multiple of size of VT?
5711 if ((N0.getValueType().getSizeInBits() & (EVTBits-1)) != 0)
5715 // At this point, we must have a load or else we can't do the transform.
5716 if (!isa<LoadSDNode>(N0)) return SDValue();
5718 // Because a SRL must be assumed to *need* to zero-extend the high bits
5719 // (as opposed to anyext the high bits), we can't combine the zextload
5720 // lowering of SRL and an sextload.
5721 if (cast<LoadSDNode>(N0)->getExtensionType() == ISD::SEXTLOAD)
5724 // If the shift amount is larger than the input type then we're not
5725 // accessing any of the loaded bytes. If the load was a zextload/extload
5726 // then the result of the shift+trunc is zero/undef (handled elsewhere).
5727 if (ShAmt >= cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits())
5732 // If the load is shifted left (and the result isn't shifted back right),
5733 // we can fold the truncate through the shift.
5734 unsigned ShLeftAmt = 0;
5735 if (ShAmt == 0 && N0.getOpcode() == ISD::SHL && N0.hasOneUse() &&
5736 ExtVT == VT && TLI.isNarrowingProfitable(N0.getValueType(), VT)) {
5737 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
5738 ShLeftAmt = N01->getZExtValue();
5739 N0 = N0.getOperand(0);
5743 // If we haven't found a load, we can't narrow it. Don't transform one with
5744 // multiple uses, this would require adding a new load.
5745 if (!isa<LoadSDNode>(N0) || !N0.hasOneUse())
5748 // Don't change the width of a volatile load.
5749 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5750 if (LN0->isVolatile())
5753 // Verify that we are actually reducing a load width here.
5754 if (LN0->getMemoryVT().getSizeInBits() < EVTBits)
5757 // For the transform to be legal, the load must produce only two values
5758 // (the value loaded and the chain). Don't transform a pre-increment
5759 // load, for example, which produces an extra value. Otherwise the
5760 // transformation is not equivalent, and the downstream logic to replace
5761 // uses gets things wrong.
5762 if (LN0->getNumValues() > 2)
5765 // If the load that we're shrinking is an extload and we're not just
5766 // discarding the extension we can't simply shrink the load. Bail.
5767 // TODO: It would be possible to merge the extensions in some cases.
5768 if (LN0->getExtensionType() != ISD::NON_EXTLOAD &&
5769 LN0->getMemoryVT().getSizeInBits() < ExtVT.getSizeInBits() + ShAmt)
5772 EVT PtrType = N0.getOperand(1).getValueType();
5774 if (PtrType == MVT::Untyped || PtrType.isExtended())
5775 // It's not possible to generate a constant of extended or untyped type.
5778 // For big endian targets, we need to adjust the offset to the pointer to
5779 // load the correct bytes.
5780 if (TLI.isBigEndian()) {
5781 unsigned LVTStoreBits = LN0->getMemoryVT().getStoreSizeInBits();
5782 unsigned EVTStoreBits = ExtVT.getStoreSizeInBits();
5783 ShAmt = LVTStoreBits - EVTStoreBits - ShAmt;
5786 uint64_t PtrOff = ShAmt / 8;
5787 unsigned NewAlign = MinAlign(LN0->getAlignment(), PtrOff);
5788 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(LN0),
5789 PtrType, LN0->getBasePtr(),
5790 DAG.getConstant(PtrOff, PtrType));
5791 AddToWorkList(NewPtr.getNode());
5794 if (ExtType == ISD::NON_EXTLOAD)
5795 Load = DAG.getLoad(VT, SDLoc(N0), LN0->getChain(), NewPtr,
5796 LN0->getPointerInfo().getWithOffset(PtrOff),
5797 LN0->isVolatile(), LN0->isNonTemporal(),
5798 LN0->isInvariant(), NewAlign, LN0->getTBAAInfo());
5800 Load = DAG.getExtLoad(ExtType, SDLoc(N0), VT, LN0->getChain(),NewPtr,
5801 LN0->getPointerInfo().getWithOffset(PtrOff),
5802 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
5803 NewAlign, LN0->getTBAAInfo());
5805 // Replace the old load's chain with the new load's chain.
5806 WorkListRemover DeadNodes(*this);
5807 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1));
5809 // Shift the result left, if we've swallowed a left shift.
5810 SDValue Result = Load;
5811 if (ShLeftAmt != 0) {
5812 EVT ShImmTy = getShiftAmountTy(Result.getValueType());
5813 if (!isUIntN(ShImmTy.getSizeInBits(), ShLeftAmt))
5815 // If the shift amount is as large as the result size (but, presumably,
5816 // no larger than the source) then the useful bits of the result are
5817 // zero; we can't simply return the shortened shift, because the result
5818 // of that operation is undefined.
5819 if (ShLeftAmt >= VT.getSizeInBits())
5820 Result = DAG.getConstant(0, VT);
5822 Result = DAG.getNode(ISD::SHL, SDLoc(N0), VT,
5823 Result, DAG.getConstant(ShLeftAmt, ShImmTy));
5826 // Return the new loaded value.
5830 SDValue DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
5831 SDValue N0 = N->getOperand(0);
5832 SDValue N1 = N->getOperand(1);
5833 EVT VT = N->getValueType(0);
5834 EVT EVT = cast<VTSDNode>(N1)->getVT();
5835 unsigned VTBits = VT.getScalarType().getSizeInBits();
5836 unsigned EVTBits = EVT.getScalarType().getSizeInBits();
5838 // fold (sext_in_reg c1) -> c1
5839 if (isa<ConstantSDNode>(N0) || N0.getOpcode() == ISD::UNDEF)
5840 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, N0, N1);
5842 // If the input is already sign extended, just drop the extension.
5843 if (DAG.ComputeNumSignBits(N0) >= VTBits-EVTBits+1)
5846 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2
5847 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
5848 EVT.bitsLT(cast<VTSDNode>(N0.getOperand(1))->getVT()))
5849 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
5850 N0.getOperand(0), N1);
5852 // fold (sext_in_reg (sext x)) -> (sext x)
5853 // fold (sext_in_reg (aext x)) -> (sext x)
5854 // if x is small enough.
5855 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) {
5856 SDValue N00 = N0.getOperand(0);
5857 if (N00.getValueType().getScalarType().getSizeInBits() <= EVTBits &&
5858 (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND, VT)))
5859 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, N00, N1);
5862 // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is known zero.
5863 if (DAG.MaskedValueIsZero(N0, APInt::getBitsSet(VTBits, EVTBits-1, EVTBits)))
5864 return DAG.getZeroExtendInReg(N0, SDLoc(N), EVT);
5866 // fold operands of sext_in_reg based on knowledge that the top bits are not
5868 if (SimplifyDemandedBits(SDValue(N, 0)))
5869 return SDValue(N, 0);
5871 // fold (sext_in_reg (load x)) -> (smaller sextload x)
5872 // fold (sext_in_reg (srl (load x), c)) -> (smaller sextload (x+c/evtbits))
5873 SDValue NarrowLoad = ReduceLoadWidth(N);
5874 if (NarrowLoad.getNode())
5877 // fold (sext_in_reg (srl X, 24), i8) -> (sra X, 24)
5878 // fold (sext_in_reg (srl X, 23), i8) -> (sra X, 23) iff possible.
5879 // We already fold "(sext_in_reg (srl X, 25), i8) -> srl X, 25" above.
5880 if (N0.getOpcode() == ISD::SRL) {
5881 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
5882 if (ShAmt->getZExtValue()+EVTBits <= VTBits) {
5883 // We can turn this into an SRA iff the input to the SRL is already sign
5885 unsigned InSignBits = DAG.ComputeNumSignBits(N0.getOperand(0));
5886 if (VTBits-(ShAmt->getZExtValue()+EVTBits) < InSignBits)
5887 return DAG.getNode(ISD::SRA, SDLoc(N), VT,
5888 N0.getOperand(0), N0.getOperand(1));
5892 // fold (sext_inreg (extload x)) -> (sextload x)
5893 if (ISD::isEXTLoad(N0.getNode()) &&
5894 ISD::isUNINDEXEDLoad(N0.getNode()) &&
5895 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
5896 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5897 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) {
5898 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5899 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5901 LN0->getBasePtr(), EVT,
5902 LN0->getMemOperand());
5903 CombineTo(N, ExtLoad);
5904 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
5905 AddToWorkList(ExtLoad.getNode());
5906 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5908 // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use
5909 if (ISD::isZEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
5911 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
5912 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5913 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) {
5914 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5915 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5917 LN0->getBasePtr(), EVT,
5918 LN0->getMemOperand());
5919 CombineTo(N, ExtLoad);
5920 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
5921 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5924 // Form (sext_inreg (bswap >> 16)) or (sext_inreg (rotl (bswap) 16))
5925 if (EVTBits <= 16 && N0.getOpcode() == ISD::OR) {
5926 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
5927 N0.getOperand(1), false);
5928 if (BSwap.getNode())
5929 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
5933 // Fold a sext_inreg of a build_vector of ConstantSDNodes or undefs
5934 // into a build_vector.
5935 if (ISD::isBuildVectorOfConstantSDNodes(N0.getNode())) {
5936 SmallVector<SDValue, 8> Elts;
5937 unsigned NumElts = N0->getNumOperands();
5938 unsigned ShAmt = VTBits - EVTBits;
5940 for (unsigned i = 0; i != NumElts; ++i) {
5941 SDValue Op = N0->getOperand(i);
5942 if (Op->getOpcode() == ISD::UNDEF) {
5947 ConstantSDNode *CurrentND = cast<ConstantSDNode>(Op);
5948 const APInt &C = APInt(VTBits, CurrentND->getAPIntValue().getZExtValue());
5949 Elts.push_back(DAG.getConstant(C.shl(ShAmt).ashr(ShAmt).getZExtValue(),
5950 Op.getValueType()));
5953 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Elts);
5959 SDValue DAGCombiner::visitTRUNCATE(SDNode *N) {
5960 SDValue N0 = N->getOperand(0);
5961 EVT VT = N->getValueType(0);
5962 bool isLE = TLI.isLittleEndian();
5965 if (N0.getValueType() == N->getValueType(0))
5967 // fold (truncate c1) -> c1
5968 if (isa<ConstantSDNode>(N0))
5969 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0);
5970 // fold (truncate (truncate x)) -> (truncate x)
5971 if (N0.getOpcode() == ISD::TRUNCATE)
5972 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0));
5973 // fold (truncate (ext x)) -> (ext x) or (truncate x) or x
5974 if (N0.getOpcode() == ISD::ZERO_EXTEND ||
5975 N0.getOpcode() == ISD::SIGN_EXTEND ||
5976 N0.getOpcode() == ISD::ANY_EXTEND) {
5977 if (N0.getOperand(0).getValueType().bitsLT(VT))
5978 // if the source is smaller than the dest, we still need an extend
5979 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
5981 if (N0.getOperand(0).getValueType().bitsGT(VT))
5982 // if the source is larger than the dest, than we just need the truncate
5983 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0));
5984 // if the source and dest are the same type, we can drop both the extend
5985 // and the truncate.
5986 return N0.getOperand(0);
5989 // Fold extract-and-trunc into a narrow extract. For example:
5990 // i64 x = EXTRACT_VECTOR_ELT(v2i64 val, i32 1)
5991 // i32 y = TRUNCATE(i64 x)
5993 // v16i8 b = BITCAST (v2i64 val)
5994 // i8 x = EXTRACT_VECTOR_ELT(v16i8 b, i32 8)
5996 // Note: We only run this optimization after type legalization (which often
5997 // creates this pattern) and before operation legalization after which
5998 // we need to be more careful about the vector instructions that we generate.
5999 if (N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
6000 LegalTypes && !LegalOperations && N0->hasOneUse() && VT != MVT::i1) {
6002 EVT VecTy = N0.getOperand(0).getValueType();
6003 EVT ExTy = N0.getValueType();
6004 EVT TrTy = N->getValueType(0);
6006 unsigned NumElem = VecTy.getVectorNumElements();
6007 unsigned SizeRatio = ExTy.getSizeInBits()/TrTy.getSizeInBits();
6009 EVT NVT = EVT::getVectorVT(*DAG.getContext(), TrTy, SizeRatio * NumElem);
6010 assert(NVT.getSizeInBits() == VecTy.getSizeInBits() && "Invalid Size");
6012 SDValue EltNo = N0->getOperand(1);
6013 if (isa<ConstantSDNode>(EltNo) && isTypeLegal(NVT)) {
6014 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
6015 EVT IndexTy = TLI.getVectorIdxTy();
6016 int Index = isLE ? (Elt*SizeRatio) : (Elt*SizeRatio + (SizeRatio-1));
6018 SDValue V = DAG.getNode(ISD::BITCAST, SDLoc(N),
6019 NVT, N0.getOperand(0));
6021 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
6023 DAG.getConstant(Index, IndexTy));
6027 // trunc (select c, a, b) -> select c, (trunc a), (trunc b)
6028 if (N0.getOpcode() == ISD::SELECT) {
6029 EVT SrcVT = N0.getValueType();
6030 if ((!LegalOperations || TLI.isOperationLegal(ISD::SELECT, SrcVT)) &&
6031 TLI.isTruncateFree(SrcVT, VT)) {
6033 SDValue Cond = N0.getOperand(0);
6034 SDValue TruncOp0 = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(1));
6035 SDValue TruncOp1 = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(2));
6036 return DAG.getNode(ISD::SELECT, SDLoc(N), VT, Cond, TruncOp0, TruncOp1);
6040 // Fold a series of buildvector, bitcast, and truncate if possible.
6042 // (2xi32 trunc (bitcast ((4xi32)buildvector x, x, y, y) 2xi64)) to
6043 // (2xi32 (buildvector x, y)).
6044 if (Level == AfterLegalizeVectorOps && VT.isVector() &&
6045 N0.getOpcode() == ISD::BITCAST && N0.hasOneUse() &&
6046 N0.getOperand(0).getOpcode() == ISD::BUILD_VECTOR &&
6047 N0.getOperand(0).hasOneUse()) {
6049 SDValue BuildVect = N0.getOperand(0);
6050 EVT BuildVectEltTy = BuildVect.getValueType().getVectorElementType();
6051 EVT TruncVecEltTy = VT.getVectorElementType();
6053 // Check that the element types match.
6054 if (BuildVectEltTy == TruncVecEltTy) {
6055 // Now we only need to compute the offset of the truncated elements.
6056 unsigned BuildVecNumElts = BuildVect.getNumOperands();
6057 unsigned TruncVecNumElts = VT.getVectorNumElements();
6058 unsigned TruncEltOffset = BuildVecNumElts / TruncVecNumElts;
6060 assert((BuildVecNumElts % TruncVecNumElts) == 0 &&
6061 "Invalid number of elements");
6063 SmallVector<SDValue, 8> Opnds;
6064 for (unsigned i = 0, e = BuildVecNumElts; i != e; i += TruncEltOffset)
6065 Opnds.push_back(BuildVect.getOperand(i));
6067 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Opnds);
6071 // See if we can simplify the input to this truncate through knowledge that
6072 // only the low bits are being used.
6073 // For example "trunc (or (shl x, 8), y)" // -> trunc y
6074 // Currently we only perform this optimization on scalars because vectors
6075 // may have different active low bits.
6076 if (!VT.isVector()) {
6078 GetDemandedBits(N0, APInt::getLowBitsSet(N0.getValueSizeInBits(),
6079 VT.getSizeInBits()));
6080 if (Shorter.getNode())
6081 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Shorter);
6083 // fold (truncate (load x)) -> (smaller load x)
6084 // fold (truncate (srl (load x), c)) -> (smaller load (x+c/evtbits))
6085 if (!LegalTypes || TLI.isTypeDesirableForOp(N0.getOpcode(), VT)) {
6086 SDValue Reduced = ReduceLoadWidth(N);
6087 if (Reduced.getNode())
6089 // Handle the case where the load remains an extending load even
6090 // after truncation.
6091 if (N0.hasOneUse() && ISD::isUNINDEXEDLoad(N0.getNode())) {
6092 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6093 if (!LN0->isVolatile() &&
6094 LN0->getMemoryVT().getStoreSizeInBits() < VT.getSizeInBits()) {
6095 SDValue NewLoad = DAG.getExtLoad(LN0->getExtensionType(), SDLoc(LN0),
6096 VT, LN0->getChain(), LN0->getBasePtr(),
6098 LN0->getMemOperand());
6099 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLoad.getValue(1));
6104 // fold (trunc (concat ... x ...)) -> (concat ..., (trunc x), ...)),
6105 // where ... are all 'undef'.
6106 if (N0.getOpcode() == ISD::CONCAT_VECTORS && !LegalTypes) {
6107 SmallVector<EVT, 8> VTs;
6110 unsigned NumDefs = 0;
6112 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) {
6113 SDValue X = N0.getOperand(i);
6114 if (X.getOpcode() != ISD::UNDEF) {
6119 // Stop if more than one members are non-undef.
6122 VTs.push_back(EVT::getVectorVT(*DAG.getContext(),
6123 VT.getVectorElementType(),
6124 X.getValueType().getVectorNumElements()));
6128 return DAG.getUNDEF(VT);
6131 assert(V.getNode() && "The single defined operand is empty!");
6132 SmallVector<SDValue, 8> Opnds;
6133 for (unsigned i = 0, e = VTs.size(); i != e; ++i) {
6135 Opnds.push_back(DAG.getUNDEF(VTs[i]));
6138 SDValue NV = DAG.getNode(ISD::TRUNCATE, SDLoc(V), VTs[i], V);
6139 AddToWorkList(NV.getNode());
6140 Opnds.push_back(NV);
6142 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Opnds);
6146 // Simplify the operands using demanded-bits information.
6147 if (!VT.isVector() &&
6148 SimplifyDemandedBits(SDValue(N, 0)))
6149 return SDValue(N, 0);
6154 static SDNode *getBuildPairElt(SDNode *N, unsigned i) {
6155 SDValue Elt = N->getOperand(i);
6156 if (Elt.getOpcode() != ISD::MERGE_VALUES)
6157 return Elt.getNode();
6158 return Elt.getOperand(Elt.getResNo()).getNode();
6161 /// CombineConsecutiveLoads - build_pair (load, load) -> load
6162 /// if load locations are consecutive.
6163 SDValue DAGCombiner::CombineConsecutiveLoads(SDNode *N, EVT VT) {
6164 assert(N->getOpcode() == ISD::BUILD_PAIR);
6166 LoadSDNode *LD1 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 0));
6167 LoadSDNode *LD2 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 1));
6168 if (!LD1 || !LD2 || !ISD::isNON_EXTLoad(LD1) || !LD1->hasOneUse() ||
6169 LD1->getAddressSpace() != LD2->getAddressSpace())
6171 EVT LD1VT = LD1->getValueType(0);
6173 if (ISD::isNON_EXTLoad(LD2) &&
6175 // If both are volatile this would reduce the number of volatile loads.
6176 // If one is volatile it might be ok, but play conservative and bail out.
6177 !LD1->isVolatile() &&
6178 !LD2->isVolatile() &&
6179 DAG.isConsecutiveLoad(LD2, LD1, LD1VT.getSizeInBits()/8, 1)) {
6180 unsigned Align = LD1->getAlignment();
6181 unsigned NewAlign = TLI.getDataLayout()->
6182 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
6184 if (NewAlign <= Align &&
6185 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)))
6186 return DAG.getLoad(VT, SDLoc(N), LD1->getChain(),
6187 LD1->getBasePtr(), LD1->getPointerInfo(),
6188 false, false, false, Align);
6194 SDValue DAGCombiner::visitBITCAST(SDNode *N) {
6195 SDValue N0 = N->getOperand(0);
6196 EVT VT = N->getValueType(0);
6198 // If the input is a BUILD_VECTOR with all constant elements, fold this now.
6199 // Only do this before legalize, since afterward the target may be depending
6200 // on the bitconvert.
6201 // First check to see if this is all constant.
6203 N0.getOpcode() == ISD::BUILD_VECTOR && N0.getNode()->hasOneUse() &&
6205 bool isSimple = cast<BuildVectorSDNode>(N0)->isConstant();
6207 EVT DestEltVT = N->getValueType(0).getVectorElementType();
6208 assert(!DestEltVT.isVector() &&
6209 "Element type of vector ValueType must not be vector!");
6211 return ConstantFoldBITCASTofBUILD_VECTOR(N0.getNode(), DestEltVT);
6214 // If the input is a constant, let getNode fold it.
6215 if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) {
6216 SDValue Res = DAG.getNode(ISD::BITCAST, SDLoc(N), VT, N0);
6217 if (Res.getNode() != N) {
6218 if (!LegalOperations ||
6219 TLI.isOperationLegal(Res.getNode()->getOpcode(), VT))
6222 // Folding it resulted in an illegal node, and it's too late to
6223 // do that. Clean up the old node and forego the transformation.
6224 // Ideally this won't happen very often, because instcombine
6225 // and the earlier dagcombine runs (where illegal nodes are
6226 // permitted) should have folded most of them already.
6227 DAG.DeleteNode(Res.getNode());
6231 // (conv (conv x, t1), t2) -> (conv x, t2)
6232 if (N0.getOpcode() == ISD::BITCAST)
6233 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT,
6236 // fold (conv (load x)) -> (load (conv*)x)
6237 // If the resultant load doesn't need a higher alignment than the original!
6238 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
6239 // Do not change the width of a volatile load.
6240 !cast<LoadSDNode>(N0)->isVolatile() &&
6241 // Do not remove the cast if the types differ in endian layout.
6242 TLI.hasBigEndianPartOrdering(N0.getValueType()) ==
6243 TLI.hasBigEndianPartOrdering(VT) &&
6244 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)) &&
6245 TLI.isLoadBitCastBeneficial(N0.getValueType(), VT)) {
6246 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6247 unsigned Align = TLI.getDataLayout()->
6248 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
6249 unsigned OrigAlign = LN0->getAlignment();
6251 if (Align <= OrigAlign) {
6252 SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(),
6253 LN0->getBasePtr(), LN0->getPointerInfo(),
6254 LN0->isVolatile(), LN0->isNonTemporal(),
6255 LN0->isInvariant(), OrigAlign,
6256 LN0->getTBAAInfo());
6258 CombineTo(N0.getNode(),
6259 DAG.getNode(ISD::BITCAST, SDLoc(N0),
6260 N0.getValueType(), Load),
6266 // fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit)
6267 // fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit))
6268 // This often reduces constant pool loads.
6269 if (((N0.getOpcode() == ISD::FNEG && !TLI.isFNegFree(N0.getValueType())) ||
6270 (N0.getOpcode() == ISD::FABS && !TLI.isFAbsFree(N0.getValueType()))) &&
6271 N0.getNode()->hasOneUse() && VT.isInteger() &&
6272 !VT.isVector() && !N0.getValueType().isVector()) {
6273 SDValue NewConv = DAG.getNode(ISD::BITCAST, SDLoc(N0), VT,
6275 AddToWorkList(NewConv.getNode());
6277 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
6278 if (N0.getOpcode() == ISD::FNEG)
6279 return DAG.getNode(ISD::XOR, SDLoc(N), VT,
6280 NewConv, DAG.getConstant(SignBit, VT));
6281 assert(N0.getOpcode() == ISD::FABS);
6282 return DAG.getNode(ISD::AND, SDLoc(N), VT,
6283 NewConv, DAG.getConstant(~SignBit, VT));
6286 // fold (bitconvert (fcopysign cst, x)) ->
6287 // (or (and (bitconvert x), sign), (and cst, (not sign)))
6288 // Note that we don't handle (copysign x, cst) because this can always be
6289 // folded to an fneg or fabs.
6290 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse() &&
6291 isa<ConstantFPSDNode>(N0.getOperand(0)) &&
6292 VT.isInteger() && !VT.isVector()) {
6293 unsigned OrigXWidth = N0.getOperand(1).getValueType().getSizeInBits();
6294 EVT IntXVT = EVT::getIntegerVT(*DAG.getContext(), OrigXWidth);
6295 if (isTypeLegal(IntXVT)) {
6296 SDValue X = DAG.getNode(ISD::BITCAST, SDLoc(N0),
6297 IntXVT, N0.getOperand(1));
6298 AddToWorkList(X.getNode());
6300 // If X has a different width than the result/lhs, sext it or truncate it.
6301 unsigned VTWidth = VT.getSizeInBits();
6302 if (OrigXWidth < VTWidth) {
6303 X = DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, X);
6304 AddToWorkList(X.getNode());
6305 } else if (OrigXWidth > VTWidth) {
6306 // To get the sign bit in the right place, we have to shift it right
6307 // before truncating.
6308 X = DAG.getNode(ISD::SRL, SDLoc(X),
6309 X.getValueType(), X,
6310 DAG.getConstant(OrigXWidth-VTWidth, X.getValueType()));
6311 AddToWorkList(X.getNode());
6312 X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X);
6313 AddToWorkList(X.getNode());
6316 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
6317 X = DAG.getNode(ISD::AND, SDLoc(X), VT,
6318 X, DAG.getConstant(SignBit, VT));
6319 AddToWorkList(X.getNode());
6321 SDValue Cst = DAG.getNode(ISD::BITCAST, SDLoc(N0),
6322 VT, N0.getOperand(0));
6323 Cst = DAG.getNode(ISD::AND, SDLoc(Cst), VT,
6324 Cst, DAG.getConstant(~SignBit, VT));
6325 AddToWorkList(Cst.getNode());
6327 return DAG.getNode(ISD::OR, SDLoc(N), VT, X, Cst);
6331 // bitconvert(build_pair(ld, ld)) -> ld iff load locations are consecutive.
6332 if (N0.getOpcode() == ISD::BUILD_PAIR) {
6333 SDValue CombineLD = CombineConsecutiveLoads(N0.getNode(), VT);
6334 if (CombineLD.getNode())
6341 SDValue DAGCombiner::visitBUILD_PAIR(SDNode *N) {
6342 EVT VT = N->getValueType(0);
6343 return CombineConsecutiveLoads(N, VT);
6346 /// ConstantFoldBITCASTofBUILD_VECTOR - We know that BV is a build_vector
6347 /// node with Constant, ConstantFP or Undef operands. DstEltVT indicates the
6348 /// destination element value type.
6349 SDValue DAGCombiner::
6350 ConstantFoldBITCASTofBUILD_VECTOR(SDNode *BV, EVT DstEltVT) {
6351 EVT SrcEltVT = BV->getValueType(0).getVectorElementType();
6353 // If this is already the right type, we're done.
6354 if (SrcEltVT == DstEltVT) return SDValue(BV, 0);
6356 unsigned SrcBitSize = SrcEltVT.getSizeInBits();
6357 unsigned DstBitSize = DstEltVT.getSizeInBits();
6359 // If this is a conversion of N elements of one type to N elements of another
6360 // type, convert each element. This handles FP<->INT cases.
6361 if (SrcBitSize == DstBitSize) {
6362 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
6363 BV->getValueType(0).getVectorNumElements());
6365 // Due to the FP element handling below calling this routine recursively,
6366 // we can end up with a scalar-to-vector node here.
6367 if (BV->getOpcode() == ISD::SCALAR_TO_VECTOR)
6368 return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(BV), VT,
6369 DAG.getNode(ISD::BITCAST, SDLoc(BV),
6370 DstEltVT, BV->getOperand(0)));
6372 SmallVector<SDValue, 8> Ops;
6373 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
6374 SDValue Op = BV->getOperand(i);
6375 // If the vector element type is not legal, the BUILD_VECTOR operands
6376 // are promoted and implicitly truncated. Make that explicit here.
6377 if (Op.getValueType() != SrcEltVT)
6378 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(BV), SrcEltVT, Op);
6379 Ops.push_back(DAG.getNode(ISD::BITCAST, SDLoc(BV),
6381 AddToWorkList(Ops.back().getNode());
6383 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
6386 // Otherwise, we're growing or shrinking the elements. To avoid having to
6387 // handle annoying details of growing/shrinking FP values, we convert them to
6389 if (SrcEltVT.isFloatingPoint()) {
6390 // Convert the input float vector to a int vector where the elements are the
6392 assert((SrcEltVT == MVT::f32 || SrcEltVT == MVT::f64) && "Unknown FP VT!");
6393 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), SrcEltVT.getSizeInBits());
6394 BV = ConstantFoldBITCASTofBUILD_VECTOR(BV, IntVT).getNode();
6398 // Now we know the input is an integer vector. If the output is a FP type,
6399 // convert to integer first, then to FP of the right size.
6400 if (DstEltVT.isFloatingPoint()) {
6401 assert((DstEltVT == MVT::f32 || DstEltVT == MVT::f64) && "Unknown FP VT!");
6402 EVT TmpVT = EVT::getIntegerVT(*DAG.getContext(), DstEltVT.getSizeInBits());
6403 SDNode *Tmp = ConstantFoldBITCASTofBUILD_VECTOR(BV, TmpVT).getNode();
6405 // Next, convert to FP elements of the same size.
6406 return ConstantFoldBITCASTofBUILD_VECTOR(Tmp, DstEltVT);
6409 // Okay, we know the src/dst types are both integers of differing types.
6410 // Handling growing first.
6411 assert(SrcEltVT.isInteger() && DstEltVT.isInteger());
6412 if (SrcBitSize < DstBitSize) {
6413 unsigned NumInputsPerOutput = DstBitSize/SrcBitSize;
6415 SmallVector<SDValue, 8> Ops;
6416 for (unsigned i = 0, e = BV->getNumOperands(); i != e;
6417 i += NumInputsPerOutput) {
6418 bool isLE = TLI.isLittleEndian();
6419 APInt NewBits = APInt(DstBitSize, 0);
6420 bool EltIsUndef = true;
6421 for (unsigned j = 0; j != NumInputsPerOutput; ++j) {
6422 // Shift the previously computed bits over.
6423 NewBits <<= SrcBitSize;
6424 SDValue Op = BV->getOperand(i+ (isLE ? (NumInputsPerOutput-j-1) : j));
6425 if (Op.getOpcode() == ISD::UNDEF) continue;
6428 NewBits |= cast<ConstantSDNode>(Op)->getAPIntValue().
6429 zextOrTrunc(SrcBitSize).zext(DstBitSize);
6433 Ops.push_back(DAG.getUNDEF(DstEltVT));
6435 Ops.push_back(DAG.getConstant(NewBits, DstEltVT));
6438 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, Ops.size());
6439 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
6442 // Finally, this must be the case where we are shrinking elements: each input
6443 // turns into multiple outputs.
6444 bool isS2V = ISD::isScalarToVector(BV);
6445 unsigned NumOutputsPerInput = SrcBitSize/DstBitSize;
6446 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
6447 NumOutputsPerInput*BV->getNumOperands());
6448 SmallVector<SDValue, 8> Ops;
6450 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
6451 if (BV->getOperand(i).getOpcode() == ISD::UNDEF) {
6452 for (unsigned j = 0; j != NumOutputsPerInput; ++j)
6453 Ops.push_back(DAG.getUNDEF(DstEltVT));
6457 APInt OpVal = cast<ConstantSDNode>(BV->getOperand(i))->
6458 getAPIntValue().zextOrTrunc(SrcBitSize);
6460 for (unsigned j = 0; j != NumOutputsPerInput; ++j) {
6461 APInt ThisVal = OpVal.trunc(DstBitSize);
6462 Ops.push_back(DAG.getConstant(ThisVal, DstEltVT));
6463 if (isS2V && i == 0 && j == 0 && ThisVal.zext(SrcBitSize) == OpVal)
6464 // Simply turn this into a SCALAR_TO_VECTOR of the new type.
6465 return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(BV), VT,
6467 OpVal = OpVal.lshr(DstBitSize);
6470 // For big endian targets, swap the order of the pieces of each element.
6471 if (TLI.isBigEndian())
6472 std::reverse(Ops.end()-NumOutputsPerInput, Ops.end());
6475 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
6478 SDValue DAGCombiner::visitFADD(SDNode *N) {
6479 SDValue N0 = N->getOperand(0);
6480 SDValue N1 = N->getOperand(1);
6481 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6482 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6483 EVT VT = N->getValueType(0);
6486 if (VT.isVector()) {
6487 SDValue FoldedVOp = SimplifyVBinOp(N);
6488 if (FoldedVOp.getNode()) return FoldedVOp;
6491 // fold (fadd c1, c2) -> c1 + c2
6493 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N1);
6494 // canonicalize constant to RHS
6495 if (N0CFP && !N1CFP)
6496 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N1, N0);
6497 // fold (fadd A, 0) -> A
6498 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
6499 N1CFP->getValueAPF().isZero())
6501 // fold (fadd A, (fneg B)) -> (fsub A, B)
6502 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
6503 isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options) == 2)
6504 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N0,
6505 GetNegatedExpression(N1, DAG, LegalOperations));
6506 // fold (fadd (fneg A), B) -> (fsub B, A)
6507 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
6508 isNegatibleForFree(N0, LegalOperations, TLI, &DAG.getTarget().Options) == 2)
6509 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N1,
6510 GetNegatedExpression(N0, DAG, LegalOperations));
6512 // If allowed, fold (fadd (fadd x, c1), c2) -> (fadd x, (fadd c1, c2))
6513 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
6514 N0.getOpcode() == ISD::FADD && N0.getNode()->hasOneUse() &&
6515 isa<ConstantFPSDNode>(N0.getOperand(1)))
6516 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0.getOperand(0),
6517 DAG.getNode(ISD::FADD, SDLoc(N), VT,
6518 N0.getOperand(1), N1));
6520 // No FP constant should be created after legalization as Instruction
6521 // Selection pass has hard time in dealing with FP constant.
6523 // We don't need test this condition for transformation like following, as
6524 // the DAG being transformed implies it is legal to take FP constant as
6527 // (fadd (fmul c, x), x) -> (fmul c+1, x)
6529 bool AllowNewFpConst = (Level < AfterLegalizeDAG);
6531 // If allow, fold (fadd (fneg x), x) -> 0.0
6532 if (AllowNewFpConst && DAG.getTarget().Options.UnsafeFPMath &&
6533 N0.getOpcode() == ISD::FNEG && N0.getOperand(0) == N1)
6534 return DAG.getConstantFP(0.0, VT);
6536 // If allow, fold (fadd x, (fneg x)) -> 0.0
6537 if (AllowNewFpConst && DAG.getTarget().Options.UnsafeFPMath &&
6538 N1.getOpcode() == ISD::FNEG && N1.getOperand(0) == N0)
6539 return DAG.getConstantFP(0.0, VT);
6541 // In unsafe math mode, we can fold chains of FADD's of the same value
6542 // into multiplications. This transform is not safe in general because
6543 // we are reducing the number of rounding steps.
6544 if (DAG.getTarget().Options.UnsafeFPMath &&
6545 TLI.isOperationLegalOrCustom(ISD::FMUL, VT) &&
6547 if (N0.getOpcode() == ISD::FMUL) {
6548 ConstantFPSDNode *CFP00 = dyn_cast<ConstantFPSDNode>(N0.getOperand(0));
6549 ConstantFPSDNode *CFP01 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
6551 // (fadd (fmul c, x), x) -> (fmul x, c+1)
6552 if (CFP00 && !CFP01 && N0.getOperand(1) == N1) {
6553 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6555 DAG.getConstantFP(1.0, VT));
6556 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6560 // (fadd (fmul x, c), x) -> (fmul x, c+1)
6561 if (CFP01 && !CFP00 && N0.getOperand(0) == N1) {
6562 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6564 DAG.getConstantFP(1.0, VT));
6565 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6569 // (fadd (fmul c, x), (fadd x, x)) -> (fmul x, c+2)
6570 if (CFP00 && !CFP01 && N1.getOpcode() == ISD::FADD &&
6571 N1.getOperand(0) == N1.getOperand(1) &&
6572 N0.getOperand(1) == N1.getOperand(0)) {
6573 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6575 DAG.getConstantFP(2.0, VT));
6576 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6577 N0.getOperand(1), NewCFP);
6580 // (fadd (fmul x, c), (fadd x, x)) -> (fmul x, c+2)
6581 if (CFP01 && !CFP00 && N1.getOpcode() == ISD::FADD &&
6582 N1.getOperand(0) == N1.getOperand(1) &&
6583 N0.getOperand(0) == N1.getOperand(0)) {
6584 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6586 DAG.getConstantFP(2.0, VT));
6587 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6588 N0.getOperand(0), NewCFP);
6592 if (N1.getOpcode() == ISD::FMUL) {
6593 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0));
6594 ConstantFPSDNode *CFP11 = dyn_cast<ConstantFPSDNode>(N1.getOperand(1));
6596 // (fadd x, (fmul c, x)) -> (fmul x, c+1)
6597 if (CFP10 && !CFP11 && N1.getOperand(1) == N0) {
6598 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6600 DAG.getConstantFP(1.0, VT));
6601 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6605 // (fadd x, (fmul x, c)) -> (fmul x, c+1)
6606 if (CFP11 && !CFP10 && N1.getOperand(0) == N0) {
6607 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6609 DAG.getConstantFP(1.0, VT));
6610 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6615 // (fadd (fadd x, x), (fmul c, x)) -> (fmul x, c+2)
6616 if (CFP10 && !CFP11 && N0.getOpcode() == ISD::FADD &&
6617 N0.getOperand(0) == N0.getOperand(1) &&
6618 N1.getOperand(1) == N0.getOperand(0)) {
6619 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6621 DAG.getConstantFP(2.0, VT));
6622 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6623 N1.getOperand(1), NewCFP);
6626 // (fadd (fadd x, x), (fmul x, c)) -> (fmul x, c+2)
6627 if (CFP11 && !CFP10 && N0.getOpcode() == ISD::FADD &&
6628 N0.getOperand(0) == N0.getOperand(1) &&
6629 N1.getOperand(0) == N0.getOperand(0)) {
6630 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6632 DAG.getConstantFP(2.0, VT));
6633 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6634 N1.getOperand(0), NewCFP);
6638 if (N0.getOpcode() == ISD::FADD && AllowNewFpConst) {
6639 ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N0.getOperand(0));
6640 // (fadd (fadd x, x), x) -> (fmul x, 3.0)
6641 if (!CFP && N0.getOperand(0) == N0.getOperand(1) &&
6642 (N0.getOperand(0) == N1))
6643 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6644 N1, DAG.getConstantFP(3.0, VT));
6647 if (N1.getOpcode() == ISD::FADD && AllowNewFpConst) {
6648 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0));
6649 // (fadd x, (fadd x, x)) -> (fmul x, 3.0)
6650 if (!CFP10 && N1.getOperand(0) == N1.getOperand(1) &&
6651 N1.getOperand(0) == N0)
6652 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6653 N0, DAG.getConstantFP(3.0, VT));
6656 // (fadd (fadd x, x), (fadd x, x)) -> (fmul x, 4.0)
6657 if (AllowNewFpConst &&
6658 N0.getOpcode() == ISD::FADD && N1.getOpcode() == ISD::FADD &&
6659 N0.getOperand(0) == N0.getOperand(1) &&
6660 N1.getOperand(0) == N1.getOperand(1) &&
6661 N0.getOperand(0) == N1.getOperand(0))
6662 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6664 DAG.getConstantFP(4.0, VT));
6667 // FADD -> FMA combines:
6668 if ((DAG.getTarget().Options.AllowFPOpFusion == FPOpFusion::Fast ||
6669 DAG.getTarget().Options.UnsafeFPMath) &&
6670 DAG.getTarget().getTargetLowering()->isFMAFasterThanFMulAndFAdd(VT) &&
6671 (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FMA, VT))) {
6673 // fold (fadd (fmul x, y), z) -> (fma x, y, z)
6674 if (N0.getOpcode() == ISD::FMUL && N0->hasOneUse())
6675 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
6676 N0.getOperand(0), N0.getOperand(1), N1);
6678 // fold (fadd x, (fmul y, z)) -> (fma y, z, x)
6679 // Note: Commutes FADD operands.
6680 if (N1.getOpcode() == ISD::FMUL && N1->hasOneUse())
6681 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
6682 N1.getOperand(0), N1.getOperand(1), N0);
6688 SDValue DAGCombiner::visitFSUB(SDNode *N) {
6689 SDValue N0 = N->getOperand(0);
6690 SDValue N1 = N->getOperand(1);
6691 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6692 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6693 EVT VT = N->getValueType(0);
6697 if (VT.isVector()) {
6698 SDValue FoldedVOp = SimplifyVBinOp(N);
6699 if (FoldedVOp.getNode()) return FoldedVOp;
6702 // fold (fsub c1, c2) -> c1-c2
6704 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N0, N1);
6705 // fold (fsub A, 0) -> A
6706 if (DAG.getTarget().Options.UnsafeFPMath &&
6707 N1CFP && N1CFP->getValueAPF().isZero())
6709 // fold (fsub 0, B) -> -B
6710 if (DAG.getTarget().Options.UnsafeFPMath &&
6711 N0CFP && N0CFP->getValueAPF().isZero()) {
6712 if (isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options))
6713 return GetNegatedExpression(N1, DAG, LegalOperations);
6714 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
6715 return DAG.getNode(ISD::FNEG, dl, VT, N1);
6717 // fold (fsub A, (fneg B)) -> (fadd A, B)
6718 if (isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options))
6719 return DAG.getNode(ISD::FADD, dl, VT, N0,
6720 GetNegatedExpression(N1, DAG, LegalOperations));
6722 // If 'unsafe math' is enabled, fold
6723 // (fsub x, x) -> 0.0 &
6724 // (fsub x, (fadd x, y)) -> (fneg y) &
6725 // (fsub x, (fadd y, x)) -> (fneg y)
6726 if (DAG.getTarget().Options.UnsafeFPMath) {
6728 return DAG.getConstantFP(0.0f, VT);
6730 if (N1.getOpcode() == ISD::FADD) {
6731 SDValue N10 = N1->getOperand(0);
6732 SDValue N11 = N1->getOperand(1);
6734 if (N10 == N0 && isNegatibleForFree(N11, LegalOperations, TLI,
6735 &DAG.getTarget().Options))
6736 return GetNegatedExpression(N11, DAG, LegalOperations);
6738 if (N11 == N0 && isNegatibleForFree(N10, LegalOperations, TLI,
6739 &DAG.getTarget().Options))
6740 return GetNegatedExpression(N10, DAG, LegalOperations);
6744 // FSUB -> FMA combines:
6745 if ((DAG.getTarget().Options.AllowFPOpFusion == FPOpFusion::Fast ||
6746 DAG.getTarget().Options.UnsafeFPMath) &&
6747 DAG.getTarget().getTargetLowering()->isFMAFasterThanFMulAndFAdd(VT) &&
6748 (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FMA, VT))) {
6750 // fold (fsub (fmul x, y), z) -> (fma x, y, (fneg z))
6751 if (N0.getOpcode() == ISD::FMUL && N0->hasOneUse())
6752 return DAG.getNode(ISD::FMA, dl, VT,
6753 N0.getOperand(0), N0.getOperand(1),
6754 DAG.getNode(ISD::FNEG, dl, VT, N1));
6756 // fold (fsub x, (fmul y, z)) -> (fma (fneg y), z, x)
6757 // Note: Commutes FSUB operands.
6758 if (N1.getOpcode() == ISD::FMUL && N1->hasOneUse())
6759 return DAG.getNode(ISD::FMA, dl, VT,
6760 DAG.getNode(ISD::FNEG, dl, VT,
6762 N1.getOperand(1), N0);
6764 // fold (fsub (fneg (fmul, x, y)), z) -> (fma (fneg x), y, (fneg z))
6765 if (N0.getOpcode() == ISD::FNEG &&
6766 N0.getOperand(0).getOpcode() == ISD::FMUL &&
6767 N0->hasOneUse() && N0.getOperand(0).hasOneUse()) {
6768 SDValue N00 = N0.getOperand(0).getOperand(0);
6769 SDValue N01 = N0.getOperand(0).getOperand(1);
6770 return DAG.getNode(ISD::FMA, dl, VT,
6771 DAG.getNode(ISD::FNEG, dl, VT, N00), N01,
6772 DAG.getNode(ISD::FNEG, dl, VT, N1));
6779 SDValue DAGCombiner::visitFMUL(SDNode *N) {
6780 SDValue N0 = N->getOperand(0);
6781 SDValue N1 = N->getOperand(1);
6782 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6783 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6784 EVT VT = N->getValueType(0);
6785 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6788 if (VT.isVector()) {
6789 SDValue FoldedVOp = SimplifyVBinOp(N);
6790 if (FoldedVOp.getNode()) return FoldedVOp;
6793 // fold (fmul c1, c2) -> c1*c2
6795 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0, N1);
6796 // canonicalize constant to RHS
6797 if (N0CFP && !N1CFP)
6798 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N1, N0);
6799 // fold (fmul A, 0) -> 0
6800 if (DAG.getTarget().Options.UnsafeFPMath &&
6801 N1CFP && N1CFP->getValueAPF().isZero())
6803 // fold (fmul A, 0) -> 0, vector edition.
6804 if (DAG.getTarget().Options.UnsafeFPMath &&
6805 ISD::isBuildVectorAllZeros(N1.getNode()))
6807 // fold (fmul A, 1.0) -> A
6808 if (N1CFP && N1CFP->isExactlyValue(1.0))
6810 // fold (fmul X, 2.0) -> (fadd X, X)
6811 if (N1CFP && N1CFP->isExactlyValue(+2.0))
6812 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N0);
6813 // fold (fmul X, -1.0) -> (fneg X)
6814 if (N1CFP && N1CFP->isExactlyValue(-1.0))
6815 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
6816 return DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0);
6818 // fold (fmul (fneg X), (fneg Y)) -> (fmul X, Y)
6819 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI,
6820 &DAG.getTarget().Options)) {
6821 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI,
6822 &DAG.getTarget().Options)) {
6823 // Both can be negated for free, check to see if at least one is cheaper
6825 if (LHSNeg == 2 || RHSNeg == 2)
6826 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6827 GetNegatedExpression(N0, DAG, LegalOperations),
6828 GetNegatedExpression(N1, DAG, LegalOperations));
6832 // If allowed, fold (fmul (fmul x, c1), c2) -> (fmul x, (fmul c1, c2))
6833 if (DAG.getTarget().Options.UnsafeFPMath &&
6834 N1CFP && N0.getOpcode() == ISD::FMUL &&
6835 N0.getNode()->hasOneUse() && isa<ConstantFPSDNode>(N0.getOperand(1)))
6836 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0.getOperand(0),
6837 DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6838 N0.getOperand(1), N1));
6843 SDValue DAGCombiner::visitFMA(SDNode *N) {
6844 SDValue N0 = N->getOperand(0);
6845 SDValue N1 = N->getOperand(1);
6846 SDValue N2 = N->getOperand(2);
6847 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6848 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6849 EVT VT = N->getValueType(0);
6852 if (DAG.getTarget().Options.UnsafeFPMath) {
6853 if (N0CFP && N0CFP->isZero())
6855 if (N1CFP && N1CFP->isZero())
6858 if (N0CFP && N0CFP->isExactlyValue(1.0))
6859 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N1, N2);
6860 if (N1CFP && N1CFP->isExactlyValue(1.0))
6861 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N2);
6863 // Canonicalize (fma c, x, y) -> (fma x, c, y)
6864 if (N0CFP && !N1CFP)
6865 return DAG.getNode(ISD::FMA, SDLoc(N), VT, N1, N0, N2);
6867 // (fma x, c1, (fmul x, c2)) -> (fmul x, c1+c2)
6868 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
6869 N2.getOpcode() == ISD::FMUL &&
6870 N0 == N2.getOperand(0) &&
6871 N2.getOperand(1).getOpcode() == ISD::ConstantFP) {
6872 return DAG.getNode(ISD::FMUL, dl, VT, N0,
6873 DAG.getNode(ISD::FADD, dl, VT, N1, N2.getOperand(1)));
6877 // (fma (fmul x, c1), c2, y) -> (fma x, c1*c2, y)
6878 if (DAG.getTarget().Options.UnsafeFPMath &&
6879 N0.getOpcode() == ISD::FMUL && N1CFP &&
6880 N0.getOperand(1).getOpcode() == ISD::ConstantFP) {
6881 return DAG.getNode(ISD::FMA, dl, VT,
6883 DAG.getNode(ISD::FMUL, dl, VT, N1, N0.getOperand(1)),
6887 // (fma x, 1, y) -> (fadd x, y)
6888 // (fma x, -1, y) -> (fadd (fneg x), y)
6890 if (N1CFP->isExactlyValue(1.0))
6891 return DAG.getNode(ISD::FADD, dl, VT, N0, N2);
6893 if (N1CFP->isExactlyValue(-1.0) &&
6894 (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))) {
6895 SDValue RHSNeg = DAG.getNode(ISD::FNEG, dl, VT, N0);
6896 AddToWorkList(RHSNeg.getNode());
6897 return DAG.getNode(ISD::FADD, dl, VT, N2, RHSNeg);
6901 // (fma x, c, x) -> (fmul x, (c+1))
6902 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP && N0 == N2)
6903 return DAG.getNode(ISD::FMUL, dl, VT, N0,
6904 DAG.getNode(ISD::FADD, dl, VT,
6905 N1, DAG.getConstantFP(1.0, VT)));
6907 // (fma x, c, (fneg x)) -> (fmul x, (c-1))
6908 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
6909 N2.getOpcode() == ISD::FNEG && N2.getOperand(0) == N0)
6910 return DAG.getNode(ISD::FMUL, dl, VT, N0,
6911 DAG.getNode(ISD::FADD, dl, VT,
6912 N1, DAG.getConstantFP(-1.0, VT)));
6918 SDValue DAGCombiner::visitFDIV(SDNode *N) {
6919 SDValue N0 = N->getOperand(0);
6920 SDValue N1 = N->getOperand(1);
6921 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6922 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6923 EVT VT = N->getValueType(0);
6924 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6927 if (VT.isVector()) {
6928 SDValue FoldedVOp = SimplifyVBinOp(N);
6929 if (FoldedVOp.getNode()) return FoldedVOp;
6932 // fold (fdiv c1, c2) -> c1/c2
6934 return DAG.getNode(ISD::FDIV, SDLoc(N), VT, N0, N1);
6936 // fold (fdiv X, c2) -> fmul X, 1/c2 if losing precision is acceptable.
6937 if (N1CFP && DAG.getTarget().Options.UnsafeFPMath) {
6938 // Compute the reciprocal 1.0 / c2.
6939 APFloat N1APF = N1CFP->getValueAPF();
6940 APFloat Recip(N1APF.getSemantics(), 1); // 1.0
6941 APFloat::opStatus st = Recip.divide(N1APF, APFloat::rmNearestTiesToEven);
6942 // Only do the transform if the reciprocal is a legal fp immediate that
6943 // isn't too nasty (eg NaN, denormal, ...).
6944 if ((st == APFloat::opOK || st == APFloat::opInexact) && // Not too nasty
6945 (!LegalOperations ||
6946 // FIXME: custom lowering of ConstantFP might fail (see e.g. ARM
6947 // backend)... we should handle this gracefully after Legalize.
6948 // TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT) ||
6949 TLI.isOperationLegal(llvm::ISD::ConstantFP, VT) ||
6950 TLI.isFPImmLegal(Recip, VT)))
6951 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0,
6952 DAG.getConstantFP(Recip, VT));
6955 // (fdiv (fneg X), (fneg Y)) -> (fdiv X, Y)
6956 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI,
6957 &DAG.getTarget().Options)) {
6958 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI,
6959 &DAG.getTarget().Options)) {
6960 // Both can be negated for free, check to see if at least one is cheaper
6962 if (LHSNeg == 2 || RHSNeg == 2)
6963 return DAG.getNode(ISD::FDIV, SDLoc(N), VT,
6964 GetNegatedExpression(N0, DAG, LegalOperations),
6965 GetNegatedExpression(N1, DAG, LegalOperations));
6972 SDValue DAGCombiner::visitFREM(SDNode *N) {
6973 SDValue N0 = N->getOperand(0);
6974 SDValue N1 = N->getOperand(1);
6975 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6976 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6977 EVT VT = N->getValueType(0);
6979 // fold (frem c1, c2) -> fmod(c1,c2)
6981 return DAG.getNode(ISD::FREM, SDLoc(N), VT, N0, N1);
6986 SDValue DAGCombiner::visitFCOPYSIGN(SDNode *N) {
6987 SDValue N0 = N->getOperand(0);
6988 SDValue N1 = N->getOperand(1);
6989 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6990 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6991 EVT VT = N->getValueType(0);
6993 if (N0CFP && N1CFP) // Constant fold
6994 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT, N0, N1);
6997 const APFloat& V = N1CFP->getValueAPF();
6998 // copysign(x, c1) -> fabs(x) iff ispos(c1)
6999 // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1)
7000 if (!V.isNegative()) {
7001 if (!LegalOperations || TLI.isOperationLegal(ISD::FABS, VT))
7002 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
7004 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
7005 return DAG.getNode(ISD::FNEG, SDLoc(N), VT,
7006 DAG.getNode(ISD::FABS, SDLoc(N0), VT, N0));
7010 // copysign(fabs(x), y) -> copysign(x, y)
7011 // copysign(fneg(x), y) -> copysign(x, y)
7012 // copysign(copysign(x,z), y) -> copysign(x, y)
7013 if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG ||
7014 N0.getOpcode() == ISD::FCOPYSIGN)
7015 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
7016 N0.getOperand(0), N1);
7018 // copysign(x, abs(y)) -> abs(x)
7019 if (N1.getOpcode() == ISD::FABS)
7020 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
7022 // copysign(x, copysign(y,z)) -> copysign(x, z)
7023 if (N1.getOpcode() == ISD::FCOPYSIGN)
7024 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
7025 N0, N1.getOperand(1));
7027 // copysign(x, fp_extend(y)) -> copysign(x, y)
7028 // copysign(x, fp_round(y)) -> copysign(x, y)
7029 if (N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND)
7030 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
7031 N0, N1.getOperand(0));
7036 SDValue DAGCombiner::visitSINT_TO_FP(SDNode *N) {
7037 SDValue N0 = N->getOperand(0);
7038 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
7039 EVT VT = N->getValueType(0);
7040 EVT OpVT = N0.getValueType();
7042 // fold (sint_to_fp c1) -> c1fp
7044 // ...but only if the target supports immediate floating-point values
7045 (!LegalOperations ||
7046 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
7047 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0);
7049 // If the input is a legal type, and SINT_TO_FP is not legal on this target,
7050 // but UINT_TO_FP is legal on this target, try to convert.
7051 if (!TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT) &&
7052 TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT)) {
7053 // If the sign bit is known to be zero, we can change this to UINT_TO_FP.
7054 if (DAG.SignBitIsZero(N0))
7055 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0);
7058 // The next optimizations are desirable only if SELECT_CC can be lowered.
7059 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT) || !LegalOperations) {
7060 // fold (sint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
7061 if (N0.getOpcode() == ISD::SETCC && N0.getValueType() == MVT::i1 &&
7063 (!LegalOperations ||
7064 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
7066 { N0.getOperand(0), N0.getOperand(1),
7067 DAG.getConstantFP(-1.0, VT) , DAG.getConstantFP(0.0, VT),
7069 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops);
7072 // fold (sint_to_fp (zext (setcc x, y, cc))) ->
7073 // (select_cc x, y, 1.0, 0.0,, cc)
7074 if (N0.getOpcode() == ISD::ZERO_EXTEND &&
7075 N0.getOperand(0).getOpcode() == ISD::SETCC &&!VT.isVector() &&
7076 (!LegalOperations ||
7077 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
7079 { N0.getOperand(0).getOperand(0), N0.getOperand(0).getOperand(1),
7080 DAG.getConstantFP(1.0, VT) , DAG.getConstantFP(0.0, VT),
7081 N0.getOperand(0).getOperand(2) };
7082 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops);
7089 SDValue DAGCombiner::visitUINT_TO_FP(SDNode *N) {
7090 SDValue N0 = N->getOperand(0);
7091 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
7092 EVT VT = N->getValueType(0);
7093 EVT OpVT = N0.getValueType();
7095 // fold (uint_to_fp c1) -> c1fp
7097 // ...but only if the target supports immediate floating-point values
7098 (!LegalOperations ||
7099 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
7100 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0);
7102 // If the input is a legal type, and UINT_TO_FP is not legal on this target,
7103 // but SINT_TO_FP is legal on this target, try to convert.
7104 if (!TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT) &&
7105 TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT)) {
7106 // If the sign bit is known to be zero, we can change this to SINT_TO_FP.
7107 if (DAG.SignBitIsZero(N0))
7108 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0);
7111 // The next optimizations are desirable only if SELECT_CC can be lowered.
7112 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT) || !LegalOperations) {
7113 // fold (uint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
7115 if (N0.getOpcode() == ISD::SETCC && !VT.isVector() &&
7116 (!LegalOperations ||
7117 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
7119 { N0.getOperand(0), N0.getOperand(1),
7120 DAG.getConstantFP(1.0, VT), DAG.getConstantFP(0.0, VT),
7122 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops);
7129 SDValue DAGCombiner::visitFP_TO_SINT(SDNode *N) {
7130 SDValue N0 = N->getOperand(0);
7131 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7132 EVT VT = N->getValueType(0);
7134 // fold (fp_to_sint c1fp) -> c1
7136 return DAG.getNode(ISD::FP_TO_SINT, SDLoc(N), VT, N0);
7141 SDValue DAGCombiner::visitFP_TO_UINT(SDNode *N) {
7142 SDValue N0 = N->getOperand(0);
7143 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7144 EVT VT = N->getValueType(0);
7146 // fold (fp_to_uint c1fp) -> c1
7148 return DAG.getNode(ISD::FP_TO_UINT, SDLoc(N), VT, N0);
7153 SDValue DAGCombiner::visitFP_ROUND(SDNode *N) {
7154 SDValue N0 = N->getOperand(0);
7155 SDValue N1 = N->getOperand(1);
7156 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7157 EVT VT = N->getValueType(0);
7159 // fold (fp_round c1fp) -> c1fp
7161 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, N0, N1);
7163 // fold (fp_round (fp_extend x)) -> x
7164 if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType())
7165 return N0.getOperand(0);
7167 // fold (fp_round (fp_round x)) -> (fp_round x)
7168 if (N0.getOpcode() == ISD::FP_ROUND) {
7169 // This is a value preserving truncation if both round's are.
7170 bool IsTrunc = N->getConstantOperandVal(1) == 1 &&
7171 N0.getNode()->getConstantOperandVal(1) == 1;
7172 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, N0.getOperand(0),
7173 DAG.getIntPtrConstant(IsTrunc));
7176 // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y)
7177 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse()) {
7178 SDValue Tmp = DAG.getNode(ISD::FP_ROUND, SDLoc(N0), VT,
7179 N0.getOperand(0), N1);
7180 AddToWorkList(Tmp.getNode());
7181 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
7182 Tmp, N0.getOperand(1));
7188 SDValue DAGCombiner::visitFP_ROUND_INREG(SDNode *N) {
7189 SDValue N0 = N->getOperand(0);
7190 EVT VT = N->getValueType(0);
7191 EVT EVT = cast<VTSDNode>(N->getOperand(1))->getVT();
7192 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7194 // fold (fp_round_inreg c1fp) -> c1fp
7195 if (N0CFP && isTypeLegal(EVT)) {
7196 SDValue Round = DAG.getConstantFP(*N0CFP->getConstantFPValue(), EVT);
7197 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, Round);
7203 SDValue DAGCombiner::visitFP_EXTEND(SDNode *N) {
7204 SDValue N0 = N->getOperand(0);
7205 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7206 EVT VT = N->getValueType(0);
7208 // If this is fp_round(fpextend), don't fold it, allow ourselves to be folded.
7209 if (N->hasOneUse() &&
7210 N->use_begin()->getOpcode() == ISD::FP_ROUND)
7213 // fold (fp_extend c1fp) -> c1fp
7215 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, N0);
7217 // Turn fp_extend(fp_round(X, 1)) -> x since the fp_round doesn't affect the
7219 if (N0.getOpcode() == ISD::FP_ROUND
7220 && N0.getNode()->getConstantOperandVal(1) == 1) {
7221 SDValue In = N0.getOperand(0);
7222 if (In.getValueType() == VT) return In;
7223 if (VT.bitsLT(In.getValueType()))
7224 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT,
7225 In, N0.getOperand(1));
7226 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, In);
7229 // fold (fpext (load x)) -> (fpext (fptrunc (extload x)))
7230 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
7231 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
7232 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType()))) {
7233 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
7234 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT,
7236 LN0->getBasePtr(), N0.getValueType(),
7237 LN0->getMemOperand());
7238 CombineTo(N, ExtLoad);
7239 CombineTo(N0.getNode(),
7240 DAG.getNode(ISD::FP_ROUND, SDLoc(N0),
7241 N0.getValueType(), ExtLoad, DAG.getIntPtrConstant(1)),
7242 ExtLoad.getValue(1));
7243 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7249 SDValue DAGCombiner::visitFNEG(SDNode *N) {
7250 SDValue N0 = N->getOperand(0);
7251 EVT VT = N->getValueType(0);
7253 if (VT.isVector()) {
7254 SDValue FoldedVOp = SimplifyVUnaryOp(N);
7255 if (FoldedVOp.getNode()) return FoldedVOp;
7258 if (isNegatibleForFree(N0, LegalOperations, DAG.getTargetLoweringInfo(),
7259 &DAG.getTarget().Options))
7260 return GetNegatedExpression(N0, DAG, LegalOperations);
7262 // Transform fneg(bitconvert(x)) -> bitconvert(x^sign) to avoid loading
7263 // constant pool values.
7264 if (!TLI.isFNegFree(VT) && N0.getOpcode() == ISD::BITCAST &&
7266 N0.getNode()->hasOneUse() &&
7267 N0.getOperand(0).getValueType().isInteger()) {
7268 SDValue Int = N0.getOperand(0);
7269 EVT IntVT = Int.getValueType();
7270 if (IntVT.isInteger() && !IntVT.isVector()) {
7271 Int = DAG.getNode(ISD::XOR, SDLoc(N0), IntVT, Int,
7272 DAG.getConstant(APInt::getSignBit(IntVT.getSizeInBits()), IntVT));
7273 AddToWorkList(Int.getNode());
7274 return DAG.getNode(ISD::BITCAST, SDLoc(N),
7279 // (fneg (fmul c, x)) -> (fmul -c, x)
7280 if (N0.getOpcode() == ISD::FMUL) {
7281 ConstantFPSDNode *CFP1 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
7283 APFloat CVal = CFP1->getValueAPF();
7285 if (Level >= AfterLegalizeDAG &&
7286 (TLI.isFPImmLegal(CVal, N->getValueType(0)) ||
7287 TLI.isOperationLegal(ISD::ConstantFP, N->getValueType(0))))
7289 ISD::FMUL, SDLoc(N), VT, N0.getOperand(0),
7290 DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0.getOperand(1)));
7297 SDValue DAGCombiner::visitFCEIL(SDNode *N) {
7298 SDValue N0 = N->getOperand(0);
7299 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7300 EVT VT = N->getValueType(0);
7302 // fold (fceil c1) -> fceil(c1)
7304 return DAG.getNode(ISD::FCEIL, SDLoc(N), VT, N0);
7309 SDValue DAGCombiner::visitFTRUNC(SDNode *N) {
7310 SDValue N0 = N->getOperand(0);
7311 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7312 EVT VT = N->getValueType(0);
7314 // fold (ftrunc c1) -> ftrunc(c1)
7316 return DAG.getNode(ISD::FTRUNC, SDLoc(N), VT, N0);
7321 SDValue DAGCombiner::visitFFLOOR(SDNode *N) {
7322 SDValue N0 = N->getOperand(0);
7323 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7324 EVT VT = N->getValueType(0);
7326 // fold (ffloor c1) -> ffloor(c1)
7328 return DAG.getNode(ISD::FFLOOR, SDLoc(N), VT, N0);
7333 SDValue DAGCombiner::visitFABS(SDNode *N) {
7334 SDValue N0 = N->getOperand(0);
7335 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7336 EVT VT = N->getValueType(0);
7338 if (VT.isVector()) {
7339 SDValue FoldedVOp = SimplifyVUnaryOp(N);
7340 if (FoldedVOp.getNode()) return FoldedVOp;
7343 // fold (fabs c1) -> fabs(c1)
7345 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
7346 // fold (fabs (fabs x)) -> (fabs x)
7347 if (N0.getOpcode() == ISD::FABS)
7348 return N->getOperand(0);
7349 // fold (fabs (fneg x)) -> (fabs x)
7350 // fold (fabs (fcopysign x, y)) -> (fabs x)
7351 if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN)
7352 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0.getOperand(0));
7354 // Transform fabs(bitconvert(x)) -> bitconvert(x&~sign) to avoid loading
7355 // constant pool values.
7356 if (!TLI.isFAbsFree(VT) &&
7357 N0.getOpcode() == ISD::BITCAST && N0.getNode()->hasOneUse() &&
7358 N0.getOperand(0).getValueType().isInteger() &&
7359 !N0.getOperand(0).getValueType().isVector()) {
7360 SDValue Int = N0.getOperand(0);
7361 EVT IntVT = Int.getValueType();
7362 if (IntVT.isInteger() && !IntVT.isVector()) {
7363 Int = DAG.getNode(ISD::AND, SDLoc(N0), IntVT, Int,
7364 DAG.getConstant(~APInt::getSignBit(IntVT.getSizeInBits()), IntVT));
7365 AddToWorkList(Int.getNode());
7366 return DAG.getNode(ISD::BITCAST, SDLoc(N),
7367 N->getValueType(0), Int);
7374 SDValue DAGCombiner::visitBRCOND(SDNode *N) {
7375 SDValue Chain = N->getOperand(0);
7376 SDValue N1 = N->getOperand(1);
7377 SDValue N2 = N->getOperand(2);
7379 // If N is a constant we could fold this into a fallthrough or unconditional
7380 // branch. However that doesn't happen very often in normal code, because
7381 // Instcombine/SimplifyCFG should have handled the available opportunities.
7382 // If we did this folding here, it would be necessary to update the
7383 // MachineBasicBlock CFG, which is awkward.
7385 // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal
7387 if (N1.getOpcode() == ISD::SETCC &&
7388 TLI.isOperationLegalOrCustom(ISD::BR_CC,
7389 N1.getOperand(0).getValueType())) {
7390 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other,
7391 Chain, N1.getOperand(2),
7392 N1.getOperand(0), N1.getOperand(1), N2);
7395 if ((N1.hasOneUse() && N1.getOpcode() == ISD::SRL) ||
7396 ((N1.getOpcode() == ISD::TRUNCATE && N1.hasOneUse()) &&
7397 (N1.getOperand(0).hasOneUse() &&
7398 N1.getOperand(0).getOpcode() == ISD::SRL))) {
7399 SDNode *Trunc = nullptr;
7400 if (N1.getOpcode() == ISD::TRUNCATE) {
7401 // Look pass the truncate.
7402 Trunc = N1.getNode();
7403 N1 = N1.getOperand(0);
7406 // Match this pattern so that we can generate simpler code:
7409 // %b = and i32 %a, 2
7410 // %c = srl i32 %b, 1
7411 // brcond i32 %c ...
7416 // %b = and i32 %a, 2
7417 // %c = setcc eq %b, 0
7420 // This applies only when the AND constant value has one bit set and the
7421 // SRL constant is equal to the log2 of the AND constant. The back-end is
7422 // smart enough to convert the result into a TEST/JMP sequence.
7423 SDValue Op0 = N1.getOperand(0);
7424 SDValue Op1 = N1.getOperand(1);
7426 if (Op0.getOpcode() == ISD::AND &&
7427 Op1.getOpcode() == ISD::Constant) {
7428 SDValue AndOp1 = Op0.getOperand(1);
7430 if (AndOp1.getOpcode() == ISD::Constant) {
7431 const APInt &AndConst = cast<ConstantSDNode>(AndOp1)->getAPIntValue();
7433 if (AndConst.isPowerOf2() &&
7434 cast<ConstantSDNode>(Op1)->getAPIntValue()==AndConst.logBase2()) {
7436 DAG.getSetCC(SDLoc(N),
7437 getSetCCResultType(Op0.getValueType()),
7438 Op0, DAG.getConstant(0, Op0.getValueType()),
7441 SDValue NewBRCond = DAG.getNode(ISD::BRCOND, SDLoc(N),
7442 MVT::Other, Chain, SetCC, N2);
7443 // Don't add the new BRCond into the worklist or else SimplifySelectCC
7444 // will convert it back to (X & C1) >> C2.
7445 CombineTo(N, NewBRCond, false);
7446 // Truncate is dead.
7448 removeFromWorkList(Trunc);
7449 DAG.DeleteNode(Trunc);
7451 // Replace the uses of SRL with SETCC
7452 WorkListRemover DeadNodes(*this);
7453 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
7454 removeFromWorkList(N1.getNode());
7455 DAG.DeleteNode(N1.getNode());
7456 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7462 // Restore N1 if the above transformation doesn't match.
7463 N1 = N->getOperand(1);
7466 // Transform br(xor(x, y)) -> br(x != y)
7467 // Transform br(xor(xor(x,y), 1)) -> br (x == y)
7468 if (N1.hasOneUse() && N1.getOpcode() == ISD::XOR) {
7469 SDNode *TheXor = N1.getNode();
7470 SDValue Op0 = TheXor->getOperand(0);
7471 SDValue Op1 = TheXor->getOperand(1);
7472 if (Op0.getOpcode() == Op1.getOpcode()) {
7473 // Avoid missing important xor optimizations.
7474 SDValue Tmp = visitXOR(TheXor);
7475 if (Tmp.getNode()) {
7476 if (Tmp.getNode() != TheXor) {
7477 DEBUG(dbgs() << "\nReplacing.8 ";
7479 dbgs() << "\nWith: ";
7480 Tmp.getNode()->dump(&DAG);
7482 WorkListRemover DeadNodes(*this);
7483 DAG.ReplaceAllUsesOfValueWith(N1, Tmp);
7484 removeFromWorkList(TheXor);
7485 DAG.DeleteNode(TheXor);
7486 return DAG.getNode(ISD::BRCOND, SDLoc(N),
7487 MVT::Other, Chain, Tmp, N2);
7490 // visitXOR has changed XOR's operands or replaced the XOR completely,
7492 return SDValue(N, 0);
7496 if (Op0.getOpcode() != ISD::SETCC && Op1.getOpcode() != ISD::SETCC) {
7498 if (ConstantSDNode *RHSCI = dyn_cast<ConstantSDNode>(Op0))
7499 if (RHSCI->getAPIntValue() == 1 && Op0.hasOneUse() &&
7500 Op0.getOpcode() == ISD::XOR) {
7501 TheXor = Op0.getNode();
7505 EVT SetCCVT = N1.getValueType();
7507 SetCCVT = getSetCCResultType(SetCCVT);
7508 SDValue SetCC = DAG.getSetCC(SDLoc(TheXor),
7511 Equal ? ISD::SETEQ : ISD::SETNE);
7512 // Replace the uses of XOR with SETCC
7513 WorkListRemover DeadNodes(*this);
7514 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
7515 removeFromWorkList(N1.getNode());
7516 DAG.DeleteNode(N1.getNode());
7517 return DAG.getNode(ISD::BRCOND, SDLoc(N),
7518 MVT::Other, Chain, SetCC, N2);
7525 // Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB.
7527 SDValue DAGCombiner::visitBR_CC(SDNode *N) {
7528 CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1));
7529 SDValue CondLHS = N->getOperand(2), CondRHS = N->getOperand(3);
7531 // If N is a constant we could fold this into a fallthrough or unconditional
7532 // branch. However that doesn't happen very often in normal code, because
7533 // Instcombine/SimplifyCFG should have handled the available opportunities.
7534 // If we did this folding here, it would be necessary to update the
7535 // MachineBasicBlock CFG, which is awkward.
7537 // Use SimplifySetCC to simplify SETCC's.
7538 SDValue Simp = SimplifySetCC(getSetCCResultType(CondLHS.getValueType()),
7539 CondLHS, CondRHS, CC->get(), SDLoc(N),
7541 if (Simp.getNode()) AddToWorkList(Simp.getNode());
7543 // fold to a simpler setcc
7544 if (Simp.getNode() && Simp.getOpcode() == ISD::SETCC)
7545 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other,
7546 N->getOperand(0), Simp.getOperand(2),
7547 Simp.getOperand(0), Simp.getOperand(1),
7553 /// canFoldInAddressingMode - Return true if 'Use' is a load or a store that
7554 /// uses N as its base pointer and that N may be folded in the load / store
7555 /// addressing mode.
7556 static bool canFoldInAddressingMode(SDNode *N, SDNode *Use,
7558 const TargetLowering &TLI) {
7560 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Use)) {
7561 if (LD->isIndexed() || LD->getBasePtr().getNode() != N)
7563 VT = Use->getValueType(0);
7564 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(Use)) {
7565 if (ST->isIndexed() || ST->getBasePtr().getNode() != N)
7567 VT = ST->getValue().getValueType();
7571 TargetLowering::AddrMode AM;
7572 if (N->getOpcode() == ISD::ADD) {
7573 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
7576 AM.BaseOffs = Offset->getSExtValue();
7580 } else if (N->getOpcode() == ISD::SUB) {
7581 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
7584 AM.BaseOffs = -Offset->getSExtValue();
7591 return TLI.isLegalAddressingMode(AM, VT.getTypeForEVT(*DAG.getContext()));
7594 /// CombineToPreIndexedLoadStore - Try turning a load / store into a
7595 /// pre-indexed load / store when the base pointer is an add or subtract
7596 /// and it has other uses besides the load / store. After the
7597 /// transformation, the new indexed load / store has effectively folded
7598 /// the add / subtract in and all of its other uses are redirected to the
7599 /// new load / store.
7600 bool DAGCombiner::CombineToPreIndexedLoadStore(SDNode *N) {
7601 if (Level < AfterLegalizeDAG)
7607 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7608 if (LD->isIndexed())
7610 VT = LD->getMemoryVT();
7611 if (!TLI.isIndexedLoadLegal(ISD::PRE_INC, VT) &&
7612 !TLI.isIndexedLoadLegal(ISD::PRE_DEC, VT))
7614 Ptr = LD->getBasePtr();
7615 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7616 if (ST->isIndexed())
7618 VT = ST->getMemoryVT();
7619 if (!TLI.isIndexedStoreLegal(ISD::PRE_INC, VT) &&
7620 !TLI.isIndexedStoreLegal(ISD::PRE_DEC, VT))
7622 Ptr = ST->getBasePtr();
7628 // If the pointer is not an add/sub, or if it doesn't have multiple uses, bail
7629 // out. There is no reason to make this a preinc/predec.
7630 if ((Ptr.getOpcode() != ISD::ADD && Ptr.getOpcode() != ISD::SUB) ||
7631 Ptr.getNode()->hasOneUse())
7634 // Ask the target to do addressing mode selection.
7637 ISD::MemIndexedMode AM = ISD::UNINDEXED;
7638 if (!TLI.getPreIndexedAddressParts(N, BasePtr, Offset, AM, DAG))
7641 // Backends without true r+i pre-indexed forms may need to pass a
7642 // constant base with a variable offset so that constant coercion
7643 // will work with the patterns in canonical form.
7644 bool Swapped = false;
7645 if (isa<ConstantSDNode>(BasePtr)) {
7646 std::swap(BasePtr, Offset);
7650 // Don't create a indexed load / store with zero offset.
7651 if (isa<ConstantSDNode>(Offset) &&
7652 cast<ConstantSDNode>(Offset)->isNullValue())
7655 // Try turning it into a pre-indexed load / store except when:
7656 // 1) The new base ptr is a frame index.
7657 // 2) If N is a store and the new base ptr is either the same as or is a
7658 // predecessor of the value being stored.
7659 // 3) Another use of old base ptr is a predecessor of N. If ptr is folded
7660 // that would create a cycle.
7661 // 4) All uses are load / store ops that use it as old base ptr.
7663 // Check #1. Preinc'ing a frame index would require copying the stack pointer
7664 // (plus the implicit offset) to a register to preinc anyway.
7665 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
7670 SDValue Val = cast<StoreSDNode>(N)->getValue();
7671 if (Val == BasePtr || BasePtr.getNode()->isPredecessorOf(Val.getNode()))
7675 // If the offset is a constant, there may be other adds of constants that
7676 // can be folded with this one. We should do this to avoid having to keep
7677 // a copy of the original base pointer.
7678 SmallVector<SDNode *, 16> OtherUses;
7679 if (isa<ConstantSDNode>(Offset))
7680 for (SDNode *Use : BasePtr.getNode()->uses()) {
7681 if (Use == Ptr.getNode())
7684 if (Use->isPredecessorOf(N))
7687 if (Use->getOpcode() != ISD::ADD && Use->getOpcode() != ISD::SUB) {
7692 SDValue Op0 = Use->getOperand(0), Op1 = Use->getOperand(1);
7693 if (Op1.getNode() == BasePtr.getNode())
7694 std::swap(Op0, Op1);
7695 assert(Op0.getNode() == BasePtr.getNode() &&
7696 "Use of ADD/SUB but not an operand");
7698 if (!isa<ConstantSDNode>(Op1)) {
7703 // FIXME: In some cases, we can be smarter about this.
7704 if (Op1.getValueType() != Offset.getValueType()) {
7709 OtherUses.push_back(Use);
7713 std::swap(BasePtr, Offset);
7715 // Now check for #3 and #4.
7716 bool RealUse = false;
7718 // Caches for hasPredecessorHelper
7719 SmallPtrSet<const SDNode *, 32> Visited;
7720 SmallVector<const SDNode *, 16> Worklist;
7722 for (SDNode *Use : Ptr.getNode()->uses()) {
7725 if (N->hasPredecessorHelper(Use, Visited, Worklist))
7728 // If Ptr may be folded in addressing mode of other use, then it's
7729 // not profitable to do this transformation.
7730 if (!canFoldInAddressingMode(Ptr.getNode(), Use, DAG, TLI))
7739 Result = DAG.getIndexedLoad(SDValue(N,0), SDLoc(N),
7740 BasePtr, Offset, AM);
7742 Result = DAG.getIndexedStore(SDValue(N,0), SDLoc(N),
7743 BasePtr, Offset, AM);
7746 DEBUG(dbgs() << "\nReplacing.4 ";
7748 dbgs() << "\nWith: ";
7749 Result.getNode()->dump(&DAG);
7751 WorkListRemover DeadNodes(*this);
7753 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
7754 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
7756 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
7759 // Finally, since the node is now dead, remove it from the graph.
7763 std::swap(BasePtr, Offset);
7765 // Replace other uses of BasePtr that can be updated to use Ptr
7766 for (unsigned i = 0, e = OtherUses.size(); i != e; ++i) {
7767 unsigned OffsetIdx = 1;
7768 if (OtherUses[i]->getOperand(OffsetIdx).getNode() == BasePtr.getNode())
7770 assert(OtherUses[i]->getOperand(!OffsetIdx).getNode() ==
7771 BasePtr.getNode() && "Expected BasePtr operand");
7773 // We need to replace ptr0 in the following expression:
7774 // x0 * offset0 + y0 * ptr0 = t0
7776 // x1 * offset1 + y1 * ptr0 = t1 (the indexed load/store)
7778 // where x0, x1, y0 and y1 in {-1, 1} are given by the types of the
7779 // indexed load/store and the expresion that needs to be re-written.
7781 // Therefore, we have:
7782 // t0 = (x0 * offset0 - x1 * y0 * y1 *offset1) + (y0 * y1) * t1
7784 ConstantSDNode *CN =
7785 cast<ConstantSDNode>(OtherUses[i]->getOperand(OffsetIdx));
7787 APInt Offset0 = CN->getAPIntValue();
7788 APInt Offset1 = cast<ConstantSDNode>(Offset)->getAPIntValue();
7790 X0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 1) ? -1 : 1;
7791 Y0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 0) ? -1 : 1;
7792 X1 = (AM == ISD::PRE_DEC && !Swapped) ? -1 : 1;
7793 Y1 = (AM == ISD::PRE_DEC && Swapped) ? -1 : 1;
7795 unsigned Opcode = (Y0 * Y1 < 0) ? ISD::SUB : ISD::ADD;
7797 APInt CNV = Offset0;
7798 if (X0 < 0) CNV = -CNV;
7799 if (X1 * Y0 * Y1 < 0) CNV = CNV + Offset1;
7800 else CNV = CNV - Offset1;
7802 // We can now generate the new expression.
7803 SDValue NewOp1 = DAG.getConstant(CNV, CN->getValueType(0));
7804 SDValue NewOp2 = Result.getValue(isLoad ? 1 : 0);
7806 SDValue NewUse = DAG.getNode(Opcode,
7807 SDLoc(OtherUses[i]),
7808 OtherUses[i]->getValueType(0), NewOp1, NewOp2);
7809 DAG.ReplaceAllUsesOfValueWith(SDValue(OtherUses[i], 0), NewUse);
7810 removeFromWorkList(OtherUses[i]);
7811 DAG.DeleteNode(OtherUses[i]);
7814 // Replace the uses of Ptr with uses of the updated base value.
7815 DAG.ReplaceAllUsesOfValueWith(Ptr, Result.getValue(isLoad ? 1 : 0));
7816 removeFromWorkList(Ptr.getNode());
7817 DAG.DeleteNode(Ptr.getNode());
7822 /// CombineToPostIndexedLoadStore - Try to combine a load / store with a
7823 /// add / sub of the base pointer node into a post-indexed load / store.
7824 /// The transformation folded the add / subtract into the new indexed
7825 /// load / store effectively and all of its uses are redirected to the
7826 /// new load / store.
7827 bool DAGCombiner::CombineToPostIndexedLoadStore(SDNode *N) {
7828 if (Level < AfterLegalizeDAG)
7834 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7835 if (LD->isIndexed())
7837 VT = LD->getMemoryVT();
7838 if (!TLI.isIndexedLoadLegal(ISD::POST_INC, VT) &&
7839 !TLI.isIndexedLoadLegal(ISD::POST_DEC, VT))
7841 Ptr = LD->getBasePtr();
7842 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7843 if (ST->isIndexed())
7845 VT = ST->getMemoryVT();
7846 if (!TLI.isIndexedStoreLegal(ISD::POST_INC, VT) &&
7847 !TLI.isIndexedStoreLegal(ISD::POST_DEC, VT))
7849 Ptr = ST->getBasePtr();
7855 if (Ptr.getNode()->hasOneUse())
7858 for (SDNode *Op : Ptr.getNode()->uses()) {
7860 (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB))
7865 ISD::MemIndexedMode AM = ISD::UNINDEXED;
7866 if (TLI.getPostIndexedAddressParts(N, Op, BasePtr, Offset, AM, DAG)) {
7867 // Don't create a indexed load / store with zero offset.
7868 if (isa<ConstantSDNode>(Offset) &&
7869 cast<ConstantSDNode>(Offset)->isNullValue())
7872 // Try turning it into a post-indexed load / store except when
7873 // 1) All uses are load / store ops that use it as base ptr (and
7874 // it may be folded as addressing mmode).
7875 // 2) Op must be independent of N, i.e. Op is neither a predecessor
7876 // nor a successor of N. Otherwise, if Op is folded that would
7879 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
7883 bool TryNext = false;
7884 for (SDNode *Use : BasePtr.getNode()->uses()) {
7885 if (Use == Ptr.getNode())
7888 // If all the uses are load / store addresses, then don't do the
7890 if (Use->getOpcode() == ISD::ADD || Use->getOpcode() == ISD::SUB){
7891 bool RealUse = false;
7892 for (SDNode *UseUse : Use->uses()) {
7893 if (!canFoldInAddressingMode(Use, UseUse, DAG, TLI))
7908 if (!Op->isPredecessorOf(N) && !N->isPredecessorOf(Op)) {
7909 SDValue Result = isLoad
7910 ? DAG.getIndexedLoad(SDValue(N,0), SDLoc(N),
7911 BasePtr, Offset, AM)
7912 : DAG.getIndexedStore(SDValue(N,0), SDLoc(N),
7913 BasePtr, Offset, AM);
7916 DEBUG(dbgs() << "\nReplacing.5 ";
7918 dbgs() << "\nWith: ";
7919 Result.getNode()->dump(&DAG);
7921 WorkListRemover DeadNodes(*this);
7923 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
7924 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
7926 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
7929 // Finally, since the node is now dead, remove it from the graph.
7932 // Replace the uses of Use with uses of the updated base value.
7933 DAG.ReplaceAllUsesOfValueWith(SDValue(Op, 0),
7934 Result.getValue(isLoad ? 1 : 0));
7935 removeFromWorkList(Op);
7945 SDValue DAGCombiner::visitLOAD(SDNode *N) {
7946 LoadSDNode *LD = cast<LoadSDNode>(N);
7947 SDValue Chain = LD->getChain();
7948 SDValue Ptr = LD->getBasePtr();
7950 // If load is not volatile and there are no uses of the loaded value (and
7951 // the updated indexed value in case of indexed loads), change uses of the
7952 // chain value into uses of the chain input (i.e. delete the dead load).
7953 if (!LD->isVolatile()) {
7954 if (N->getValueType(1) == MVT::Other) {
7956 if (!N->hasAnyUseOfValue(0)) {
7957 // It's not safe to use the two value CombineTo variant here. e.g.
7958 // v1, chain2 = load chain1, loc
7959 // v2, chain3 = load chain2, loc
7961 // Now we replace use of chain2 with chain1. This makes the second load
7962 // isomorphic to the one we are deleting, and thus makes this load live.
7963 DEBUG(dbgs() << "\nReplacing.6 ";
7965 dbgs() << "\nWith chain: ";
7966 Chain.getNode()->dump(&DAG);
7968 WorkListRemover DeadNodes(*this);
7969 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
7971 if (N->use_empty()) {
7972 removeFromWorkList(N);
7976 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7980 assert(N->getValueType(2) == MVT::Other && "Malformed indexed loads?");
7981 if (!N->hasAnyUseOfValue(0) && !N->hasAnyUseOfValue(1)) {
7982 SDValue Undef = DAG.getUNDEF(N->getValueType(0));
7983 DEBUG(dbgs() << "\nReplacing.7 ";
7985 dbgs() << "\nWith: ";
7986 Undef.getNode()->dump(&DAG);
7987 dbgs() << " and 2 other values\n");
7988 WorkListRemover DeadNodes(*this);
7989 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Undef);
7990 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1),
7991 DAG.getUNDEF(N->getValueType(1)));
7992 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 2), Chain);
7993 removeFromWorkList(N);
7995 return SDValue(N, 0); // Return N so it doesn't get rechecked!
8000 // If this load is directly stored, replace the load value with the stored
8002 // TODO: Handle store large -> read small portion.
8003 // TODO: Handle TRUNCSTORE/LOADEXT
8004 if (ISD::isNormalLoad(N) && !LD->isVolatile()) {
8005 if (ISD::isNON_TRUNCStore(Chain.getNode())) {
8006 StoreSDNode *PrevST = cast<StoreSDNode>(Chain);
8007 if (PrevST->getBasePtr() == Ptr &&
8008 PrevST->getValue().getValueType() == N->getValueType(0))
8009 return CombineTo(N, Chain.getOperand(1), Chain);
8013 // Try to infer better alignment information than the load already has.
8014 if (OptLevel != CodeGenOpt::None && LD->isUnindexed()) {
8015 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
8016 if (Align > LD->getMemOperand()->getBaseAlignment()) {
8018 DAG.getExtLoad(LD->getExtensionType(), SDLoc(N),
8019 LD->getValueType(0),
8020 Chain, Ptr, LD->getPointerInfo(),
8022 LD->isVolatile(), LD->isNonTemporal(), Align,
8024 return CombineTo(N, NewLoad, SDValue(NewLoad.getNode(), 1), true);
8029 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA :
8030 TLI.getTargetMachine().getSubtarget<TargetSubtargetInfo>().useAA();
8032 if (CombinerAAOnlyFunc.getNumOccurrences() &&
8033 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
8036 if (UseAA && LD->isUnindexed()) {
8037 // Walk up chain skipping non-aliasing memory nodes.
8038 SDValue BetterChain = FindBetterChain(N, Chain);
8040 // If there is a better chain.
8041 if (Chain != BetterChain) {
8044 // Replace the chain to void dependency.
8045 if (LD->getExtensionType() == ISD::NON_EXTLOAD) {
8046 ReplLoad = DAG.getLoad(N->getValueType(0), SDLoc(LD),
8047 BetterChain, Ptr, LD->getMemOperand());
8049 ReplLoad = DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD),
8050 LD->getValueType(0),
8051 BetterChain, Ptr, LD->getMemoryVT(),
8052 LD->getMemOperand());
8055 // Create token factor to keep old chain connected.
8056 SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N),
8057 MVT::Other, Chain, ReplLoad.getValue(1));
8059 // Make sure the new and old chains are cleaned up.
8060 AddToWorkList(Token.getNode());
8062 // Replace uses with load result and token factor. Don't add users
8064 return CombineTo(N, ReplLoad.getValue(0), Token, false);
8068 // Try transforming N to an indexed load.
8069 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
8070 return SDValue(N, 0);
8072 // Try to slice up N to more direct loads if the slices are mapped to
8073 // different register banks or pairing can take place.
8075 return SDValue(N, 0);
8081 /// \brief Helper structure used to slice a load in smaller loads.
8082 /// Basically a slice is obtained from the following sequence:
8083 /// Origin = load Ty1, Base
8084 /// Shift = srl Ty1 Origin, CstTy Amount
8085 /// Inst = trunc Shift to Ty2
8087 /// Then, it will be rewriten into:
8088 /// Slice = load SliceTy, Base + SliceOffset
8089 /// [Inst = zext Slice to Ty2], only if SliceTy <> Ty2
8091 /// SliceTy is deduced from the number of bits that are actually used to
8093 struct LoadedSlice {
8094 /// \brief Helper structure used to compute the cost of a slice.
8096 /// Are we optimizing for code size.
8101 unsigned CrossRegisterBanksCopies;
8105 Cost(bool ForCodeSize = false)
8106 : ForCodeSize(ForCodeSize), Loads(0), Truncates(0),
8107 CrossRegisterBanksCopies(0), ZExts(0), Shift(0) {}
8109 /// \brief Get the cost of one isolated slice.
8110 Cost(const LoadedSlice &LS, bool ForCodeSize = false)
8111 : ForCodeSize(ForCodeSize), Loads(1), Truncates(0),
8112 CrossRegisterBanksCopies(0), ZExts(0), Shift(0) {
8113 EVT TruncType = LS.Inst->getValueType(0);
8114 EVT LoadedType = LS.getLoadedType();
8115 if (TruncType != LoadedType &&
8116 !LS.DAG->getTargetLoweringInfo().isZExtFree(LoadedType, TruncType))
8120 /// \brief Account for slicing gain in the current cost.
8121 /// Slicing provide a few gains like removing a shift or a
8122 /// truncate. This method allows to grow the cost of the original
8123 /// load with the gain from this slice.
8124 void addSliceGain(const LoadedSlice &LS) {
8125 // Each slice saves a truncate.
8126 const TargetLowering &TLI = LS.DAG->getTargetLoweringInfo();
8127 if (!TLI.isTruncateFree(LS.Inst->getValueType(0),
8128 LS.Inst->getOperand(0).getValueType()))
8130 // If there is a shift amount, this slice gets rid of it.
8133 // If this slice can merge a cross register bank copy, account for it.
8134 if (LS.canMergeExpensiveCrossRegisterBankCopy())
8135 ++CrossRegisterBanksCopies;
8138 Cost &operator+=(const Cost &RHS) {
8140 Truncates += RHS.Truncates;
8141 CrossRegisterBanksCopies += RHS.CrossRegisterBanksCopies;
8147 bool operator==(const Cost &RHS) const {
8148 return Loads == RHS.Loads && Truncates == RHS.Truncates &&
8149 CrossRegisterBanksCopies == RHS.CrossRegisterBanksCopies &&
8150 ZExts == RHS.ZExts && Shift == RHS.Shift;
8153 bool operator!=(const Cost &RHS) const { return !(*this == RHS); }
8155 bool operator<(const Cost &RHS) const {
8156 // Assume cross register banks copies are as expensive as loads.
8157 // FIXME: Do we want some more target hooks?
8158 unsigned ExpensiveOpsLHS = Loads + CrossRegisterBanksCopies;
8159 unsigned ExpensiveOpsRHS = RHS.Loads + RHS.CrossRegisterBanksCopies;
8160 // Unless we are optimizing for code size, consider the
8161 // expensive operation first.
8162 if (!ForCodeSize && ExpensiveOpsLHS != ExpensiveOpsRHS)
8163 return ExpensiveOpsLHS < ExpensiveOpsRHS;
8164 return (Truncates + ZExts + Shift + ExpensiveOpsLHS) <
8165 (RHS.Truncates + RHS.ZExts + RHS.Shift + ExpensiveOpsRHS);
8168 bool operator>(const Cost &RHS) const { return RHS < *this; }
8170 bool operator<=(const Cost &RHS) const { return !(RHS < *this); }
8172 bool operator>=(const Cost &RHS) const { return !(*this < RHS); }
8174 // The last instruction that represent the slice. This should be a
8175 // truncate instruction.
8177 // The original load instruction.
8179 // The right shift amount in bits from the original load.
8181 // The DAG from which Origin came from.
8182 // This is used to get some contextual information about legal types, etc.
8185 LoadedSlice(SDNode *Inst = nullptr, LoadSDNode *Origin = nullptr,
8186 unsigned Shift = 0, SelectionDAG *DAG = nullptr)
8187 : Inst(Inst), Origin(Origin), Shift(Shift), DAG(DAG) {}
8189 LoadedSlice(const LoadedSlice &LS)
8190 : Inst(LS.Inst), Origin(LS.Origin), Shift(LS.Shift), DAG(LS.DAG) {}
8192 /// \brief Get the bits used in a chunk of bits \p BitWidth large.
8193 /// \return Result is \p BitWidth and has used bits set to 1 and
8194 /// not used bits set to 0.
8195 APInt getUsedBits() const {
8196 // Reproduce the trunc(lshr) sequence:
8197 // - Start from the truncated value.
8198 // - Zero extend to the desired bit width.
8200 assert(Origin && "No original load to compare against.");
8201 unsigned BitWidth = Origin->getValueSizeInBits(0);
8202 assert(Inst && "This slice is not bound to an instruction");
8203 assert(Inst->getValueSizeInBits(0) <= BitWidth &&
8204 "Extracted slice is bigger than the whole type!");
8205 APInt UsedBits(Inst->getValueSizeInBits(0), 0);
8206 UsedBits.setAllBits();
8207 UsedBits = UsedBits.zext(BitWidth);
8212 /// \brief Get the size of the slice to be loaded in bytes.
8213 unsigned getLoadedSize() const {
8214 unsigned SliceSize = getUsedBits().countPopulation();
8215 assert(!(SliceSize & 0x7) && "Size is not a multiple of a byte.");
8216 return SliceSize / 8;
8219 /// \brief Get the type that will be loaded for this slice.
8220 /// Note: This may not be the final type for the slice.
8221 EVT getLoadedType() const {
8222 assert(DAG && "Missing context");
8223 LLVMContext &Ctxt = *DAG->getContext();
8224 return EVT::getIntegerVT(Ctxt, getLoadedSize() * 8);
8227 /// \brief Get the alignment of the load used for this slice.
8228 unsigned getAlignment() const {
8229 unsigned Alignment = Origin->getAlignment();
8230 unsigned Offset = getOffsetFromBase();
8232 Alignment = MinAlign(Alignment, Alignment + Offset);
8236 /// \brief Check if this slice can be rewritten with legal operations.
8237 bool isLegal() const {
8238 // An invalid slice is not legal.
8239 if (!Origin || !Inst || !DAG)
8242 // Offsets are for indexed load only, we do not handle that.
8243 if (Origin->getOffset().getOpcode() != ISD::UNDEF)
8246 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
8248 // Check that the type is legal.
8249 EVT SliceType = getLoadedType();
8250 if (!TLI.isTypeLegal(SliceType))
8253 // Check that the load is legal for this type.
8254 if (!TLI.isOperationLegal(ISD::LOAD, SliceType))
8257 // Check that the offset can be computed.
8258 // 1. Check its type.
8259 EVT PtrType = Origin->getBasePtr().getValueType();
8260 if (PtrType == MVT::Untyped || PtrType.isExtended())
8263 // 2. Check that it fits in the immediate.
8264 if (!TLI.isLegalAddImmediate(getOffsetFromBase()))
8267 // 3. Check that the computation is legal.
8268 if (!TLI.isOperationLegal(ISD::ADD, PtrType))
8271 // Check that the zext is legal if it needs one.
8272 EVT TruncateType = Inst->getValueType(0);
8273 if (TruncateType != SliceType &&
8274 !TLI.isOperationLegal(ISD::ZERO_EXTEND, TruncateType))
8280 /// \brief Get the offset in bytes of this slice in the original chunk of
8282 /// \pre DAG != nullptr.
8283 uint64_t getOffsetFromBase() const {
8284 assert(DAG && "Missing context.");
8286 DAG->getTargetLoweringInfo().getDataLayout()->isBigEndian();
8287 assert(!(Shift & 0x7) && "Shifts not aligned on Bytes are not supported.");
8288 uint64_t Offset = Shift / 8;
8289 unsigned TySizeInBytes = Origin->getValueSizeInBits(0) / 8;
8290 assert(!(Origin->getValueSizeInBits(0) & 0x7) &&
8291 "The size of the original loaded type is not a multiple of a"
8293 // If Offset is bigger than TySizeInBytes, it means we are loading all
8294 // zeros. This should have been optimized before in the process.
8295 assert(TySizeInBytes > Offset &&
8296 "Invalid shift amount for given loaded size");
8298 Offset = TySizeInBytes - Offset - getLoadedSize();
8302 /// \brief Generate the sequence of instructions to load the slice
8303 /// represented by this object and redirect the uses of this slice to
8304 /// this new sequence of instructions.
8305 /// \pre this->Inst && this->Origin are valid Instructions and this
8306 /// object passed the legal check: LoadedSlice::isLegal returned true.
8307 /// \return The last instruction of the sequence used to load the slice.
8308 SDValue loadSlice() const {
8309 assert(Inst && Origin && "Unable to replace a non-existing slice.");
8310 const SDValue &OldBaseAddr = Origin->getBasePtr();
8311 SDValue BaseAddr = OldBaseAddr;
8312 // Get the offset in that chunk of bytes w.r.t. the endianess.
8313 int64_t Offset = static_cast<int64_t>(getOffsetFromBase());
8314 assert(Offset >= 0 && "Offset too big to fit in int64_t!");
8316 // BaseAddr = BaseAddr + Offset.
8317 EVT ArithType = BaseAddr.getValueType();
8318 BaseAddr = DAG->getNode(ISD::ADD, SDLoc(Origin), ArithType, BaseAddr,
8319 DAG->getConstant(Offset, ArithType));
8322 // Create the type of the loaded slice according to its size.
8323 EVT SliceType = getLoadedType();
8325 // Create the load for the slice.
8326 SDValue LastInst = DAG->getLoad(
8327 SliceType, SDLoc(Origin), Origin->getChain(), BaseAddr,
8328 Origin->getPointerInfo().getWithOffset(Offset), Origin->isVolatile(),
8329 Origin->isNonTemporal(), Origin->isInvariant(), getAlignment());
8330 // If the final type is not the same as the loaded type, this means that
8331 // we have to pad with zero. Create a zero extend for that.
8332 EVT FinalType = Inst->getValueType(0);
8333 if (SliceType != FinalType)
8335 DAG->getNode(ISD::ZERO_EXTEND, SDLoc(LastInst), FinalType, LastInst);
8339 /// \brief Check if this slice can be merged with an expensive cross register
8340 /// bank copy. E.g.,
8342 /// f = bitcast i32 i to float
8343 bool canMergeExpensiveCrossRegisterBankCopy() const {
8344 if (!Inst || !Inst->hasOneUse())
8346 SDNode *Use = *Inst->use_begin();
8347 if (Use->getOpcode() != ISD::BITCAST)
8349 assert(DAG && "Missing context");
8350 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
8351 EVT ResVT = Use->getValueType(0);
8352 const TargetRegisterClass *ResRC = TLI.getRegClassFor(ResVT.getSimpleVT());
8353 const TargetRegisterClass *ArgRC =
8354 TLI.getRegClassFor(Use->getOperand(0).getValueType().getSimpleVT());
8355 if (ArgRC == ResRC || !TLI.isOperationLegal(ISD::LOAD, ResVT))
8358 // At this point, we know that we perform a cross-register-bank copy.
8359 // Check if it is expensive.
8360 const TargetRegisterInfo *TRI = TLI.getTargetMachine().getRegisterInfo();
8361 // Assume bitcasts are cheap, unless both register classes do not
8362 // explicitly share a common sub class.
8363 if (!TRI || TRI->getCommonSubClass(ArgRC, ResRC))
8366 // Check if it will be merged with the load.
8367 // 1. Check the alignment constraint.
8368 unsigned RequiredAlignment = TLI.getDataLayout()->getABITypeAlignment(
8369 ResVT.getTypeForEVT(*DAG->getContext()));
8371 if (RequiredAlignment > getAlignment())
8374 // 2. Check that the load is a legal operation for that type.
8375 if (!TLI.isOperationLegal(ISD::LOAD, ResVT))
8378 // 3. Check that we do not have a zext in the way.
8379 if (Inst->getValueType(0) != getLoadedType())
8387 /// \brief Check that all bits set in \p UsedBits form a dense region, i.e.,
8388 /// \p UsedBits looks like 0..0 1..1 0..0.
8389 static bool areUsedBitsDense(const APInt &UsedBits) {
8390 // If all the bits are one, this is dense!
8391 if (UsedBits.isAllOnesValue())
8394 // Get rid of the unused bits on the right.
8395 APInt NarrowedUsedBits = UsedBits.lshr(UsedBits.countTrailingZeros());
8396 // Get rid of the unused bits on the left.
8397 if (NarrowedUsedBits.countLeadingZeros())
8398 NarrowedUsedBits = NarrowedUsedBits.trunc(NarrowedUsedBits.getActiveBits());
8399 // Check that the chunk of bits is completely used.
8400 return NarrowedUsedBits.isAllOnesValue();
8403 /// \brief Check whether or not \p First and \p Second are next to each other
8404 /// in memory. This means that there is no hole between the bits loaded
8405 /// by \p First and the bits loaded by \p Second.
8406 static bool areSlicesNextToEachOther(const LoadedSlice &First,
8407 const LoadedSlice &Second) {
8408 assert(First.Origin == Second.Origin && First.Origin &&
8409 "Unable to match different memory origins.");
8410 APInt UsedBits = First.getUsedBits();
8411 assert((UsedBits & Second.getUsedBits()) == 0 &&
8412 "Slices are not supposed to overlap.");
8413 UsedBits |= Second.getUsedBits();
8414 return areUsedBitsDense(UsedBits);
8417 /// \brief Adjust the \p GlobalLSCost according to the target
8418 /// paring capabilities and the layout of the slices.
8419 /// \pre \p GlobalLSCost should account for at least as many loads as
8420 /// there is in the slices in \p LoadedSlices.
8421 static void adjustCostForPairing(SmallVectorImpl<LoadedSlice> &LoadedSlices,
8422 LoadedSlice::Cost &GlobalLSCost) {
8423 unsigned NumberOfSlices = LoadedSlices.size();
8424 // If there is less than 2 elements, no pairing is possible.
8425 if (NumberOfSlices < 2)
8428 // Sort the slices so that elements that are likely to be next to each
8429 // other in memory are next to each other in the list.
8430 std::sort(LoadedSlices.begin(), LoadedSlices.end(),
8431 [](const LoadedSlice &LHS, const LoadedSlice &RHS) {
8432 assert(LHS.Origin == RHS.Origin && "Different bases not implemented.");
8433 return LHS.getOffsetFromBase() < RHS.getOffsetFromBase();
8435 const TargetLowering &TLI = LoadedSlices[0].DAG->getTargetLoweringInfo();
8436 // First (resp. Second) is the first (resp. Second) potentially candidate
8437 // to be placed in a paired load.
8438 const LoadedSlice *First = nullptr;
8439 const LoadedSlice *Second = nullptr;
8440 for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice,
8441 // Set the beginning of the pair.
8444 Second = &LoadedSlices[CurrSlice];
8446 // If First is NULL, it means we start a new pair.
8447 // Get to the next slice.
8451 EVT LoadedType = First->getLoadedType();
8453 // If the types of the slices are different, we cannot pair them.
8454 if (LoadedType != Second->getLoadedType())
8457 // Check if the target supplies paired loads for this type.
8458 unsigned RequiredAlignment = 0;
8459 if (!TLI.hasPairedLoad(LoadedType, RequiredAlignment)) {
8460 // move to the next pair, this type is hopeless.
8464 // Check if we meet the alignment requirement.
8465 if (RequiredAlignment > First->getAlignment())
8468 // Check that both loads are next to each other in memory.
8469 if (!areSlicesNextToEachOther(*First, *Second))
8472 assert(GlobalLSCost.Loads > 0 && "We save more loads than we created!");
8473 --GlobalLSCost.Loads;
8474 // Move to the next pair.
8479 /// \brief Check the profitability of all involved LoadedSlice.
8480 /// Currently, it is considered profitable if there is exactly two
8481 /// involved slices (1) which are (2) next to each other in memory, and
8482 /// whose cost (\see LoadedSlice::Cost) is smaller than the original load (3).
8484 /// Note: The order of the elements in \p LoadedSlices may be modified, but not
8485 /// the elements themselves.
8487 /// FIXME: When the cost model will be mature enough, we can relax
8488 /// constraints (1) and (2).
8489 static bool isSlicingProfitable(SmallVectorImpl<LoadedSlice> &LoadedSlices,
8490 const APInt &UsedBits, bool ForCodeSize) {
8491 unsigned NumberOfSlices = LoadedSlices.size();
8492 if (StressLoadSlicing)
8493 return NumberOfSlices > 1;
8496 if (NumberOfSlices != 2)
8500 if (!areUsedBitsDense(UsedBits))
8504 LoadedSlice::Cost OrigCost(ForCodeSize), GlobalSlicingCost(ForCodeSize);
8505 // The original code has one big load.
8507 for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice) {
8508 const LoadedSlice &LS = LoadedSlices[CurrSlice];
8509 // Accumulate the cost of all the slices.
8510 LoadedSlice::Cost SliceCost(LS, ForCodeSize);
8511 GlobalSlicingCost += SliceCost;
8513 // Account as cost in the original configuration the gain obtained
8514 // with the current slices.
8515 OrigCost.addSliceGain(LS);
8518 // If the target supports paired load, adjust the cost accordingly.
8519 adjustCostForPairing(LoadedSlices, GlobalSlicingCost);
8520 return OrigCost > GlobalSlicingCost;
8523 /// \brief If the given load, \p LI, is used only by trunc or trunc(lshr)
8524 /// operations, split it in the various pieces being extracted.
8526 /// This sort of thing is introduced by SROA.
8527 /// This slicing takes care not to insert overlapping loads.
8528 /// \pre LI is a simple load (i.e., not an atomic or volatile load).
8529 bool DAGCombiner::SliceUpLoad(SDNode *N) {
8530 if (Level < AfterLegalizeDAG)
8533 LoadSDNode *LD = cast<LoadSDNode>(N);
8534 if (LD->isVolatile() || !ISD::isNormalLoad(LD) ||
8535 !LD->getValueType(0).isInteger())
8538 // Keep track of already used bits to detect overlapping values.
8539 // In that case, we will just abort the transformation.
8540 APInt UsedBits(LD->getValueSizeInBits(0), 0);
8542 SmallVector<LoadedSlice, 4> LoadedSlices;
8544 // Check if this load is used as several smaller chunks of bits.
8545 // Basically, look for uses in trunc or trunc(lshr) and record a new chain
8546 // of computation for each trunc.
8547 for (SDNode::use_iterator UI = LD->use_begin(), UIEnd = LD->use_end();
8548 UI != UIEnd; ++UI) {
8549 // Skip the uses of the chain.
8550 if (UI.getUse().getResNo() != 0)
8556 // Check if this is a trunc(lshr).
8557 if (User->getOpcode() == ISD::SRL && User->hasOneUse() &&
8558 isa<ConstantSDNode>(User->getOperand(1))) {
8559 Shift = cast<ConstantSDNode>(User->getOperand(1))->getZExtValue();
8560 User = *User->use_begin();
8563 // At this point, User is a Truncate, iff we encountered, trunc or
8565 if (User->getOpcode() != ISD::TRUNCATE)
8568 // The width of the type must be a power of 2 and greater than 8-bits.
8569 // Otherwise the load cannot be represented in LLVM IR.
8570 // Moreover, if we shifted with a non-8-bits multiple, the slice
8571 // will be across several bytes. We do not support that.
8572 unsigned Width = User->getValueSizeInBits(0);
8573 if (Width < 8 || !isPowerOf2_32(Width) || (Shift & 0x7))
8576 // Build the slice for this chain of computations.
8577 LoadedSlice LS(User, LD, Shift, &DAG);
8578 APInt CurrentUsedBits = LS.getUsedBits();
8580 // Check if this slice overlaps with another.
8581 if ((CurrentUsedBits & UsedBits) != 0)
8583 // Update the bits used globally.
8584 UsedBits |= CurrentUsedBits;
8586 // Check if the new slice would be legal.
8590 // Record the slice.
8591 LoadedSlices.push_back(LS);
8594 // Abort slicing if it does not seem to be profitable.
8595 if (!isSlicingProfitable(LoadedSlices, UsedBits, ForCodeSize))
8600 // Rewrite each chain to use an independent load.
8601 // By construction, each chain can be represented by a unique load.
8603 // Prepare the argument for the new token factor for all the slices.
8604 SmallVector<SDValue, 8> ArgChains;
8605 for (SmallVectorImpl<LoadedSlice>::const_iterator
8606 LSIt = LoadedSlices.begin(),
8607 LSItEnd = LoadedSlices.end();
8608 LSIt != LSItEnd; ++LSIt) {
8609 SDValue SliceInst = LSIt->loadSlice();
8610 CombineTo(LSIt->Inst, SliceInst, true);
8611 if (SliceInst.getNode()->getOpcode() != ISD::LOAD)
8612 SliceInst = SliceInst.getOperand(0);
8613 assert(SliceInst->getOpcode() == ISD::LOAD &&
8614 "It takes more than a zext to get to the loaded slice!!");
8615 ArgChains.push_back(SliceInst.getValue(1));
8618 SDValue Chain = DAG.getNode(ISD::TokenFactor, SDLoc(LD), MVT::Other,
8620 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
8624 /// CheckForMaskedLoad - Check to see if V is (and load (ptr), imm), where the
8625 /// load is having specific bytes cleared out. If so, return the byte size
8626 /// being masked out and the shift amount.
8627 static std::pair<unsigned, unsigned>
8628 CheckForMaskedLoad(SDValue V, SDValue Ptr, SDValue Chain) {
8629 std::pair<unsigned, unsigned> Result(0, 0);
8631 // Check for the structure we're looking for.
8632 if (V->getOpcode() != ISD::AND ||
8633 !isa<ConstantSDNode>(V->getOperand(1)) ||
8634 !ISD::isNormalLoad(V->getOperand(0).getNode()))
8637 // Check the chain and pointer.
8638 LoadSDNode *LD = cast<LoadSDNode>(V->getOperand(0));
8639 if (LD->getBasePtr() != Ptr) return Result; // Not from same pointer.
8641 // The store should be chained directly to the load or be an operand of a
8643 if (LD == Chain.getNode())
8645 else if (Chain->getOpcode() != ISD::TokenFactor)
8646 return Result; // Fail.
8649 for (unsigned i = 0, e = Chain->getNumOperands(); i != e; ++i)
8650 if (Chain->getOperand(i).getNode() == LD) {
8654 if (!isOk) return Result;
8657 // This only handles simple types.
8658 if (V.getValueType() != MVT::i16 &&
8659 V.getValueType() != MVT::i32 &&
8660 V.getValueType() != MVT::i64)
8663 // Check the constant mask. Invert it so that the bits being masked out are
8664 // 0 and the bits being kept are 1. Use getSExtValue so that leading bits
8665 // follow the sign bit for uniformity.
8666 uint64_t NotMask = ~cast<ConstantSDNode>(V->getOperand(1))->getSExtValue();
8667 unsigned NotMaskLZ = countLeadingZeros(NotMask);
8668 if (NotMaskLZ & 7) return Result; // Must be multiple of a byte.
8669 unsigned NotMaskTZ = countTrailingZeros(NotMask);
8670 if (NotMaskTZ & 7) return Result; // Must be multiple of a byte.
8671 if (NotMaskLZ == 64) return Result; // All zero mask.
8673 // See if we have a continuous run of bits. If so, we have 0*1+0*
8674 if (CountTrailingOnes_64(NotMask >> NotMaskTZ)+NotMaskTZ+NotMaskLZ != 64)
8677 // Adjust NotMaskLZ down to be from the actual size of the int instead of i64.
8678 if (V.getValueType() != MVT::i64 && NotMaskLZ)
8679 NotMaskLZ -= 64-V.getValueSizeInBits();
8681 unsigned MaskedBytes = (V.getValueSizeInBits()-NotMaskLZ-NotMaskTZ)/8;
8682 switch (MaskedBytes) {
8686 default: return Result; // All one mask, or 5-byte mask.
8689 // Verify that the first bit starts at a multiple of mask so that the access
8690 // is aligned the same as the access width.
8691 if (NotMaskTZ && NotMaskTZ/8 % MaskedBytes) return Result;
8693 Result.first = MaskedBytes;
8694 Result.second = NotMaskTZ/8;
8699 /// ShrinkLoadReplaceStoreWithStore - Check to see if IVal is something that
8700 /// provides a value as specified by MaskInfo. If so, replace the specified
8701 /// store with a narrower store of truncated IVal.
8703 ShrinkLoadReplaceStoreWithStore(const std::pair<unsigned, unsigned> &MaskInfo,
8704 SDValue IVal, StoreSDNode *St,
8706 unsigned NumBytes = MaskInfo.first;
8707 unsigned ByteShift = MaskInfo.second;
8708 SelectionDAG &DAG = DC->getDAG();
8710 // Check to see if IVal is all zeros in the part being masked in by the 'or'
8711 // that uses this. If not, this is not a replacement.
8712 APInt Mask = ~APInt::getBitsSet(IVal.getValueSizeInBits(),
8713 ByteShift*8, (ByteShift+NumBytes)*8);
8714 if (!DAG.MaskedValueIsZero(IVal, Mask)) return nullptr;
8716 // Check that it is legal on the target to do this. It is legal if the new
8717 // VT we're shrinking to (i8/i16/i32) is legal or we're still before type
8719 MVT VT = MVT::getIntegerVT(NumBytes*8);
8720 if (!DC->isTypeLegal(VT))
8723 // Okay, we can do this! Replace the 'St' store with a store of IVal that is
8724 // shifted by ByteShift and truncated down to NumBytes.
8726 IVal = DAG.getNode(ISD::SRL, SDLoc(IVal), IVal.getValueType(), IVal,
8727 DAG.getConstant(ByteShift*8,
8728 DC->getShiftAmountTy(IVal.getValueType())));
8730 // Figure out the offset for the store and the alignment of the access.
8732 unsigned NewAlign = St->getAlignment();
8734 if (DAG.getTargetLoweringInfo().isLittleEndian())
8735 StOffset = ByteShift;
8737 StOffset = IVal.getValueType().getStoreSize() - ByteShift - NumBytes;
8739 SDValue Ptr = St->getBasePtr();
8741 Ptr = DAG.getNode(ISD::ADD, SDLoc(IVal), Ptr.getValueType(),
8742 Ptr, DAG.getConstant(StOffset, Ptr.getValueType()));
8743 NewAlign = MinAlign(NewAlign, StOffset);
8746 // Truncate down to the new size.
8747 IVal = DAG.getNode(ISD::TRUNCATE, SDLoc(IVal), VT, IVal);
8750 return DAG.getStore(St->getChain(), SDLoc(St), IVal, Ptr,
8751 St->getPointerInfo().getWithOffset(StOffset),
8752 false, false, NewAlign).getNode();
8756 /// ReduceLoadOpStoreWidth - Look for sequence of load / op / store where op is
8757 /// one of 'or', 'xor', and 'and' of immediates. If 'op' is only touching some
8758 /// of the loaded bits, try narrowing the load and store if it would end up
8759 /// being a win for performance or code size.
8760 SDValue DAGCombiner::ReduceLoadOpStoreWidth(SDNode *N) {
8761 StoreSDNode *ST = cast<StoreSDNode>(N);
8762 if (ST->isVolatile())
8765 SDValue Chain = ST->getChain();
8766 SDValue Value = ST->getValue();
8767 SDValue Ptr = ST->getBasePtr();
8768 EVT VT = Value.getValueType();
8770 if (ST->isTruncatingStore() || VT.isVector() || !Value.hasOneUse())
8773 unsigned Opc = Value.getOpcode();
8775 // If this is "store (or X, Y), P" and X is "(and (load P), cst)", where cst
8776 // is a byte mask indicating a consecutive number of bytes, check to see if
8777 // Y is known to provide just those bytes. If so, we try to replace the
8778 // load + replace + store sequence with a single (narrower) store, which makes
8780 if (Opc == ISD::OR) {
8781 std::pair<unsigned, unsigned> MaskedLoad;
8782 MaskedLoad = CheckForMaskedLoad(Value.getOperand(0), Ptr, Chain);
8783 if (MaskedLoad.first)
8784 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
8785 Value.getOperand(1), ST,this))
8786 return SDValue(NewST, 0);
8788 // Or is commutative, so try swapping X and Y.
8789 MaskedLoad = CheckForMaskedLoad(Value.getOperand(1), Ptr, Chain);
8790 if (MaskedLoad.first)
8791 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
8792 Value.getOperand(0), ST,this))
8793 return SDValue(NewST, 0);
8796 if ((Opc != ISD::OR && Opc != ISD::XOR && Opc != ISD::AND) ||
8797 Value.getOperand(1).getOpcode() != ISD::Constant)
8800 SDValue N0 = Value.getOperand(0);
8801 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
8802 Chain == SDValue(N0.getNode(), 1)) {
8803 LoadSDNode *LD = cast<LoadSDNode>(N0);
8804 if (LD->getBasePtr() != Ptr ||
8805 LD->getPointerInfo().getAddrSpace() !=
8806 ST->getPointerInfo().getAddrSpace())
8809 // Find the type to narrow it the load / op / store to.
8810 SDValue N1 = Value.getOperand(1);
8811 unsigned BitWidth = N1.getValueSizeInBits();
8812 APInt Imm = cast<ConstantSDNode>(N1)->getAPIntValue();
8813 if (Opc == ISD::AND)
8814 Imm ^= APInt::getAllOnesValue(BitWidth);
8815 if (Imm == 0 || Imm.isAllOnesValue())
8817 unsigned ShAmt = Imm.countTrailingZeros();
8818 unsigned MSB = BitWidth - Imm.countLeadingZeros() - 1;
8819 unsigned NewBW = NextPowerOf2(MSB - ShAmt);
8820 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
8821 while (NewBW < BitWidth &&
8822 !(TLI.isOperationLegalOrCustom(Opc, NewVT) &&
8823 TLI.isNarrowingProfitable(VT, NewVT))) {
8824 NewBW = NextPowerOf2(NewBW);
8825 NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
8827 if (NewBW >= BitWidth)
8830 // If the lsb changed does not start at the type bitwidth boundary,
8831 // start at the previous one.
8833 ShAmt = (((ShAmt + NewBW - 1) / NewBW) * NewBW) - NewBW;
8834 APInt Mask = APInt::getBitsSet(BitWidth, ShAmt,
8835 std::min(BitWidth, ShAmt + NewBW));
8836 if ((Imm & Mask) == Imm) {
8837 APInt NewImm = (Imm & Mask).lshr(ShAmt).trunc(NewBW);
8838 if (Opc == ISD::AND)
8839 NewImm ^= APInt::getAllOnesValue(NewBW);
8840 uint64_t PtrOff = ShAmt / 8;
8841 // For big endian targets, we need to adjust the offset to the pointer to
8842 // load the correct bytes.
8843 if (TLI.isBigEndian())
8844 PtrOff = (BitWidth + 7 - NewBW) / 8 - PtrOff;
8846 unsigned NewAlign = MinAlign(LD->getAlignment(), PtrOff);
8847 Type *NewVTTy = NewVT.getTypeForEVT(*DAG.getContext());
8848 if (NewAlign < TLI.getDataLayout()->getABITypeAlignment(NewVTTy))
8851 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(LD),
8852 Ptr.getValueType(), Ptr,
8853 DAG.getConstant(PtrOff, Ptr.getValueType()));
8854 SDValue NewLD = DAG.getLoad(NewVT, SDLoc(N0),
8855 LD->getChain(), NewPtr,
8856 LD->getPointerInfo().getWithOffset(PtrOff),
8857 LD->isVolatile(), LD->isNonTemporal(),
8858 LD->isInvariant(), NewAlign,
8860 SDValue NewVal = DAG.getNode(Opc, SDLoc(Value), NewVT, NewLD,
8861 DAG.getConstant(NewImm, NewVT));
8862 SDValue NewST = DAG.getStore(Chain, SDLoc(N),
8864 ST->getPointerInfo().getWithOffset(PtrOff),
8865 false, false, NewAlign);
8867 AddToWorkList(NewPtr.getNode());
8868 AddToWorkList(NewLD.getNode());
8869 AddToWorkList(NewVal.getNode());
8870 WorkListRemover DeadNodes(*this);
8871 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLD.getValue(1));
8880 /// TransformFPLoadStorePair - For a given floating point load / store pair,
8881 /// if the load value isn't used by any other operations, then consider
8882 /// transforming the pair to integer load / store operations if the target
8883 /// deems the transformation profitable.
8884 SDValue DAGCombiner::TransformFPLoadStorePair(SDNode *N) {
8885 StoreSDNode *ST = cast<StoreSDNode>(N);
8886 SDValue Chain = ST->getChain();
8887 SDValue Value = ST->getValue();
8888 if (ISD::isNormalStore(ST) && ISD::isNormalLoad(Value.getNode()) &&
8889 Value.hasOneUse() &&
8890 Chain == SDValue(Value.getNode(), 1)) {
8891 LoadSDNode *LD = cast<LoadSDNode>(Value);
8892 EVT VT = LD->getMemoryVT();
8893 if (!VT.isFloatingPoint() ||
8894 VT != ST->getMemoryVT() ||
8895 LD->isNonTemporal() ||
8896 ST->isNonTemporal() ||
8897 LD->getPointerInfo().getAddrSpace() != 0 ||
8898 ST->getPointerInfo().getAddrSpace() != 0)
8901 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
8902 if (!TLI.isOperationLegal(ISD::LOAD, IntVT) ||
8903 !TLI.isOperationLegal(ISD::STORE, IntVT) ||
8904 !TLI.isDesirableToTransformToIntegerOp(ISD::LOAD, VT) ||
8905 !TLI.isDesirableToTransformToIntegerOp(ISD::STORE, VT))
8908 unsigned LDAlign = LD->getAlignment();
8909 unsigned STAlign = ST->getAlignment();
8910 Type *IntVTTy = IntVT.getTypeForEVT(*DAG.getContext());
8911 unsigned ABIAlign = TLI.getDataLayout()->getABITypeAlignment(IntVTTy);
8912 if (LDAlign < ABIAlign || STAlign < ABIAlign)
8915 SDValue NewLD = DAG.getLoad(IntVT, SDLoc(Value),
8916 LD->getChain(), LD->getBasePtr(),
8917 LD->getPointerInfo(),
8918 false, false, false, LDAlign);
8920 SDValue NewST = DAG.getStore(NewLD.getValue(1), SDLoc(N),
8921 NewLD, ST->getBasePtr(),
8922 ST->getPointerInfo(),
8923 false, false, STAlign);
8925 AddToWorkList(NewLD.getNode());
8926 AddToWorkList(NewST.getNode());
8927 WorkListRemover DeadNodes(*this);
8928 DAG.ReplaceAllUsesOfValueWith(Value.getValue(1), NewLD.getValue(1));
8936 /// Helper struct to parse and store a memory address as base + index + offset.
8937 /// We ignore sign extensions when it is safe to do so.
8938 /// The following two expressions are not equivalent. To differentiate we need
8939 /// to store whether there was a sign extension involved in the index
8941 /// (load (i64 add (i64 copyfromreg %c)
8942 /// (i64 signextend (add (i8 load %index)
8946 /// (load (i64 add (i64 copyfromreg %c)
8947 /// (i64 signextend (i32 add (i32 signextend (i8 load %index))
8949 struct BaseIndexOffset {
8953 bool IsIndexSignExt;
8955 BaseIndexOffset() : Offset(0), IsIndexSignExt(false) {}
8957 BaseIndexOffset(SDValue Base, SDValue Index, int64_t Offset,
8958 bool IsIndexSignExt) :
8959 Base(Base), Index(Index), Offset(Offset), IsIndexSignExt(IsIndexSignExt) {}
8961 bool equalBaseIndex(const BaseIndexOffset &Other) {
8962 return Other.Base == Base && Other.Index == Index &&
8963 Other.IsIndexSignExt == IsIndexSignExt;
8966 /// Parses tree in Ptr for base, index, offset addresses.
8967 static BaseIndexOffset match(SDValue Ptr) {
8968 bool IsIndexSignExt = false;
8970 // We only can pattern match BASE + INDEX + OFFSET. If Ptr is not an ADD
8971 // instruction, then it could be just the BASE or everything else we don't
8972 // know how to handle. Just use Ptr as BASE and give up.
8973 if (Ptr->getOpcode() != ISD::ADD)
8974 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
8976 // We know that we have at least an ADD instruction. Try to pattern match
8977 // the simple case of BASE + OFFSET.
8978 if (isa<ConstantSDNode>(Ptr->getOperand(1))) {
8979 int64_t Offset = cast<ConstantSDNode>(Ptr->getOperand(1))->getSExtValue();
8980 return BaseIndexOffset(Ptr->getOperand(0), SDValue(), Offset,
8984 // Inside a loop the current BASE pointer is calculated using an ADD and a
8985 // MUL instruction. In this case Ptr is the actual BASE pointer.
8986 // (i64 add (i64 %array_ptr)
8987 // (i64 mul (i64 %induction_var)
8988 // (i64 %element_size)))
8989 if (Ptr->getOperand(1)->getOpcode() == ISD::MUL)
8990 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
8992 // Look at Base + Index + Offset cases.
8993 SDValue Base = Ptr->getOperand(0);
8994 SDValue IndexOffset = Ptr->getOperand(1);
8996 // Skip signextends.
8997 if (IndexOffset->getOpcode() == ISD::SIGN_EXTEND) {
8998 IndexOffset = IndexOffset->getOperand(0);
8999 IsIndexSignExt = true;
9002 // Either the case of Base + Index (no offset) or something else.
9003 if (IndexOffset->getOpcode() != ISD::ADD)
9004 return BaseIndexOffset(Base, IndexOffset, 0, IsIndexSignExt);
9006 // Now we have the case of Base + Index + offset.
9007 SDValue Index = IndexOffset->getOperand(0);
9008 SDValue Offset = IndexOffset->getOperand(1);
9010 if (!isa<ConstantSDNode>(Offset))
9011 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
9013 // Ignore signextends.
9014 if (Index->getOpcode() == ISD::SIGN_EXTEND) {
9015 Index = Index->getOperand(0);
9016 IsIndexSignExt = true;
9017 } else IsIndexSignExt = false;
9019 int64_t Off = cast<ConstantSDNode>(Offset)->getSExtValue();
9020 return BaseIndexOffset(Base, Index, Off, IsIndexSignExt);
9024 /// Holds a pointer to an LSBaseSDNode as well as information on where it
9025 /// is located in a sequence of memory operations connected by a chain.
9027 MemOpLink (LSBaseSDNode *N, int64_t Offset, unsigned Seq):
9028 MemNode(N), OffsetFromBase(Offset), SequenceNum(Seq) { }
9029 // Ptr to the mem node.
9030 LSBaseSDNode *MemNode;
9031 // Offset from the base ptr.
9032 int64_t OffsetFromBase;
9033 // What is the sequence number of this mem node.
9034 // Lowest mem operand in the DAG starts at zero.
9035 unsigned SequenceNum;
9038 bool DAGCombiner::MergeConsecutiveStores(StoreSDNode* St) {
9039 EVT MemVT = St->getMemoryVT();
9040 int64_t ElementSizeBytes = MemVT.getSizeInBits()/8;
9041 bool NoVectors = DAG.getMachineFunction().getFunction()->getAttributes().
9042 hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat);
9044 // Don't merge vectors into wider inputs.
9045 if (MemVT.isVector() || !MemVT.isSimple())
9048 // Perform an early exit check. Do not bother looking at stored values that
9049 // are not constants or loads.
9050 SDValue StoredVal = St->getValue();
9051 bool IsLoadSrc = isa<LoadSDNode>(StoredVal);
9052 if (!isa<ConstantSDNode>(StoredVal) && !isa<ConstantFPSDNode>(StoredVal) &&
9056 // Only look at ends of store sequences.
9057 SDValue Chain = SDValue(St, 1);
9058 if (Chain->hasOneUse() && Chain->use_begin()->getOpcode() == ISD::STORE)
9061 // This holds the base pointer, index, and the offset in bytes from the base
9063 BaseIndexOffset BasePtr = BaseIndexOffset::match(St->getBasePtr());
9065 // We must have a base and an offset.
9066 if (!BasePtr.Base.getNode())
9069 // Do not handle stores to undef base pointers.
9070 if (BasePtr.Base.getOpcode() == ISD::UNDEF)
9073 // Save the LoadSDNodes that we find in the chain.
9074 // We need to make sure that these nodes do not interfere with
9075 // any of the store nodes.
9076 SmallVector<LSBaseSDNode*, 8> AliasLoadNodes;
9078 // Save the StoreSDNodes that we find in the chain.
9079 SmallVector<MemOpLink, 8> StoreNodes;
9081 // Walk up the chain and look for nodes with offsets from the same
9082 // base pointer. Stop when reaching an instruction with a different kind
9083 // or instruction which has a different base pointer.
9085 StoreSDNode *Index = St;
9087 // If the chain has more than one use, then we can't reorder the mem ops.
9088 if (Index != St && !SDValue(Index, 1)->hasOneUse())
9091 // Find the base pointer and offset for this memory node.
9092 BaseIndexOffset Ptr = BaseIndexOffset::match(Index->getBasePtr());
9094 // Check that the base pointer is the same as the original one.
9095 if (!Ptr.equalBaseIndex(BasePtr))
9098 // Check that the alignment is the same.
9099 if (Index->getAlignment() != St->getAlignment())
9102 // The memory operands must not be volatile.
9103 if (Index->isVolatile() || Index->isIndexed())
9107 if (StoreSDNode *St = dyn_cast<StoreSDNode>(Index))
9108 if (St->isTruncatingStore())
9111 // The stored memory type must be the same.
9112 if (Index->getMemoryVT() != MemVT)
9115 // We do not allow unaligned stores because we want to prevent overriding
9117 if (Index->getAlignment()*8 != MemVT.getSizeInBits())
9120 // We found a potential memory operand to merge.
9121 StoreNodes.push_back(MemOpLink(Index, Ptr.Offset, Seq++));
9123 // Find the next memory operand in the chain. If the next operand in the
9124 // chain is a store then move up and continue the scan with the next
9125 // memory operand. If the next operand is a load save it and use alias
9126 // information to check if it interferes with anything.
9127 SDNode *NextInChain = Index->getChain().getNode();
9129 if (StoreSDNode *STn = dyn_cast<StoreSDNode>(NextInChain)) {
9130 // We found a store node. Use it for the next iteration.
9133 } else if (LoadSDNode *Ldn = dyn_cast<LoadSDNode>(NextInChain)) {
9134 if (Ldn->isVolatile()) {
9139 // Save the load node for later. Continue the scan.
9140 AliasLoadNodes.push_back(Ldn);
9141 NextInChain = Ldn->getChain().getNode();
9150 // Check if there is anything to merge.
9151 if (StoreNodes.size() < 2)
9154 // Sort the memory operands according to their distance from the base pointer.
9155 std::sort(StoreNodes.begin(), StoreNodes.end(),
9156 [](MemOpLink LHS, MemOpLink RHS) {
9157 return LHS.OffsetFromBase < RHS.OffsetFromBase ||
9158 (LHS.OffsetFromBase == RHS.OffsetFromBase &&
9159 LHS.SequenceNum > RHS.SequenceNum);
9162 // Scan the memory operations on the chain and find the first non-consecutive
9163 // store memory address.
9164 unsigned LastConsecutiveStore = 0;
9165 int64_t StartAddress = StoreNodes[0].OffsetFromBase;
9166 for (unsigned i = 0, e = StoreNodes.size(); i < e; ++i) {
9168 // Check that the addresses are consecutive starting from the second
9169 // element in the list of stores.
9171 int64_t CurrAddress = StoreNodes[i].OffsetFromBase;
9172 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
9177 // Check if this store interferes with any of the loads that we found.
9178 for (unsigned ld = 0, lde = AliasLoadNodes.size(); ld < lde; ++ld)
9179 if (isAlias(AliasLoadNodes[ld], StoreNodes[i].MemNode)) {
9183 // We found a load that alias with this store. Stop the sequence.
9187 // Mark this node as useful.
9188 LastConsecutiveStore = i;
9191 // The node with the lowest store address.
9192 LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode;
9194 // Store the constants into memory as one consecutive store.
9196 unsigned LastLegalType = 0;
9197 unsigned LastLegalVectorType = 0;
9198 bool NonZero = false;
9199 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
9200 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
9201 SDValue StoredVal = St->getValue();
9203 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(StoredVal)) {
9204 NonZero |= !C->isNullValue();
9205 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(StoredVal)) {
9206 NonZero |= !C->getConstantFPValue()->isNullValue();
9212 // Find a legal type for the constant store.
9213 unsigned StoreBW = (i+1) * ElementSizeBytes * 8;
9214 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
9215 if (TLI.isTypeLegal(StoreTy))
9216 LastLegalType = i+1;
9217 // Or check whether a truncstore is legal.
9218 else if (TLI.getTypeAction(*DAG.getContext(), StoreTy) ==
9219 TargetLowering::TypePromoteInteger) {
9220 EVT LegalizedStoredValueTy =
9221 TLI.getTypeToTransformTo(*DAG.getContext(), StoredVal.getValueType());
9222 if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy))
9223 LastLegalType = i+1;
9226 // Find a legal type for the vector store.
9227 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
9228 if (TLI.isTypeLegal(Ty))
9229 LastLegalVectorType = i + 1;
9232 // We only use vectors if the constant is known to be zero and the
9233 // function is not marked with the noimplicitfloat attribute.
9234 if (NonZero || NoVectors)
9235 LastLegalVectorType = 0;
9237 // Check if we found a legal integer type to store.
9238 if (LastLegalType == 0 && LastLegalVectorType == 0)
9241 bool UseVector = (LastLegalVectorType > LastLegalType) && !NoVectors;
9242 unsigned NumElem = UseVector ? LastLegalVectorType : LastLegalType;
9244 // Make sure we have something to merge.
9248 unsigned EarliestNodeUsed = 0;
9249 for (unsigned i=0; i < NumElem; ++i) {
9250 // Find a chain for the new wide-store operand. Notice that some
9251 // of the store nodes that we found may not be selected for inclusion
9252 // in the wide store. The chain we use needs to be the chain of the
9253 // earliest store node which is *used* and replaced by the wide store.
9254 if (StoreNodes[i].SequenceNum > StoreNodes[EarliestNodeUsed].SequenceNum)
9255 EarliestNodeUsed = i;
9258 // The earliest Node in the DAG.
9259 LSBaseSDNode *EarliestOp = StoreNodes[EarliestNodeUsed].MemNode;
9260 SDLoc DL(StoreNodes[0].MemNode);
9264 // Find a legal type for the vector store.
9265 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
9266 assert(TLI.isTypeLegal(Ty) && "Illegal vector store");
9267 StoredVal = DAG.getConstant(0, Ty);
9269 unsigned StoreBW = NumElem * ElementSizeBytes * 8;
9270 APInt StoreInt(StoreBW, 0);
9272 // Construct a single integer constant which is made of the smaller
9274 bool IsLE = TLI.isLittleEndian();
9275 for (unsigned i = 0; i < NumElem ; ++i) {
9276 unsigned Idx = IsLE ?(NumElem - 1 - i) : i;
9277 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[Idx].MemNode);
9278 SDValue Val = St->getValue();
9279 StoreInt<<=ElementSizeBytes*8;
9280 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val)) {
9281 StoreInt|=C->getAPIntValue().zext(StoreBW);
9282 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Val)) {
9283 StoreInt|= C->getValueAPF().bitcastToAPInt().zext(StoreBW);
9285 assert(false && "Invalid constant element type");
9289 // Create the new Load and Store operations.
9290 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
9291 StoredVal = DAG.getConstant(StoreInt, StoreTy);
9294 SDValue NewStore = DAG.getStore(EarliestOp->getChain(), DL, StoredVal,
9295 FirstInChain->getBasePtr(),
9296 FirstInChain->getPointerInfo(),
9298 FirstInChain->getAlignment());
9300 // Replace the first store with the new store
9301 CombineTo(EarliestOp, NewStore);
9302 // Erase all other stores.
9303 for (unsigned i = 0; i < NumElem ; ++i) {
9304 if (StoreNodes[i].MemNode == EarliestOp)
9306 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
9307 // ReplaceAllUsesWith will replace all uses that existed when it was
9308 // called, but graph optimizations may cause new ones to appear. For
9309 // example, the case in pr14333 looks like
9311 // St's chain -> St -> another store -> X
9313 // And the only difference from St to the other store is the chain.
9314 // When we change it's chain to be St's chain they become identical,
9315 // get CSEed and the net result is that X is now a use of St.
9316 // Since we know that St is redundant, just iterate.
9317 while (!St->use_empty())
9318 DAG.ReplaceAllUsesWith(SDValue(St, 0), St->getChain());
9319 removeFromWorkList(St);
9326 // Below we handle the case of multiple consecutive stores that
9327 // come from multiple consecutive loads. We merge them into a single
9328 // wide load and a single wide store.
9330 // Look for load nodes which are used by the stored values.
9331 SmallVector<MemOpLink, 8> LoadNodes;
9333 // Find acceptable loads. Loads need to have the same chain (token factor),
9334 // must not be zext, volatile, indexed, and they must be consecutive.
9335 BaseIndexOffset LdBasePtr;
9336 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
9337 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
9338 LoadSDNode *Ld = dyn_cast<LoadSDNode>(St->getValue());
9341 // Loads must only have one use.
9342 if (!Ld->hasNUsesOfValue(1, 0))
9345 // Check that the alignment is the same as the stores.
9346 if (Ld->getAlignment() != St->getAlignment())
9349 // The memory operands must not be volatile.
9350 if (Ld->isVolatile() || Ld->isIndexed())
9353 // We do not accept ext loads.
9354 if (Ld->getExtensionType() != ISD::NON_EXTLOAD)
9357 // The stored memory type must be the same.
9358 if (Ld->getMemoryVT() != MemVT)
9361 BaseIndexOffset LdPtr = BaseIndexOffset::match(Ld->getBasePtr());
9362 // If this is not the first ptr that we check.
9363 if (LdBasePtr.Base.getNode()) {
9364 // The base ptr must be the same.
9365 if (!LdPtr.equalBaseIndex(LdBasePtr))
9368 // Check that all other base pointers are the same as this one.
9372 // We found a potential memory operand to merge.
9373 LoadNodes.push_back(MemOpLink(Ld, LdPtr.Offset, 0));
9376 if (LoadNodes.size() < 2)
9379 // Scan the memory operations on the chain and find the first non-consecutive
9380 // load memory address. These variables hold the index in the store node
9382 unsigned LastConsecutiveLoad = 0;
9383 // This variable refers to the size and not index in the array.
9384 unsigned LastLegalVectorType = 0;
9385 unsigned LastLegalIntegerType = 0;
9386 StartAddress = LoadNodes[0].OffsetFromBase;
9387 SDValue FirstChain = LoadNodes[0].MemNode->getChain();
9388 for (unsigned i = 1; i < LoadNodes.size(); ++i) {
9389 // All loads much share the same chain.
9390 if (LoadNodes[i].MemNode->getChain() != FirstChain)
9393 int64_t CurrAddress = LoadNodes[i].OffsetFromBase;
9394 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
9396 LastConsecutiveLoad = i;
9398 // Find a legal type for the vector store.
9399 EVT StoreTy = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
9400 if (TLI.isTypeLegal(StoreTy))
9401 LastLegalVectorType = i + 1;
9403 // Find a legal type for the integer store.
9404 unsigned StoreBW = (i+1) * ElementSizeBytes * 8;
9405 StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
9406 if (TLI.isTypeLegal(StoreTy))
9407 LastLegalIntegerType = i + 1;
9408 // Or check whether a truncstore and extload is legal.
9409 else if (TLI.getTypeAction(*DAG.getContext(), StoreTy) ==
9410 TargetLowering::TypePromoteInteger) {
9411 EVT LegalizedStoredValueTy =
9412 TLI.getTypeToTransformTo(*DAG.getContext(), StoreTy);
9413 if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy) &&
9414 TLI.isLoadExtLegal(ISD::ZEXTLOAD, StoreTy) &&
9415 TLI.isLoadExtLegal(ISD::SEXTLOAD, StoreTy) &&
9416 TLI.isLoadExtLegal(ISD::EXTLOAD, StoreTy))
9417 LastLegalIntegerType = i+1;
9421 // Only use vector types if the vector type is larger than the integer type.
9422 // If they are the same, use integers.
9423 bool UseVectorTy = LastLegalVectorType > LastLegalIntegerType && !NoVectors;
9424 unsigned LastLegalType = std::max(LastLegalVectorType, LastLegalIntegerType);
9426 // We add +1 here because the LastXXX variables refer to location while
9427 // the NumElem refers to array/index size.
9428 unsigned NumElem = std::min(LastConsecutiveStore, LastConsecutiveLoad) + 1;
9429 NumElem = std::min(LastLegalType, NumElem);
9434 // The earliest Node in the DAG.
9435 unsigned EarliestNodeUsed = 0;
9436 LSBaseSDNode *EarliestOp = StoreNodes[EarliestNodeUsed].MemNode;
9437 for (unsigned i=1; i<NumElem; ++i) {
9438 // Find a chain for the new wide-store operand. Notice that some
9439 // of the store nodes that we found may not be selected for inclusion
9440 // in the wide store. The chain we use needs to be the chain of the
9441 // earliest store node which is *used* and replaced by the wide store.
9442 if (StoreNodes[i].SequenceNum > StoreNodes[EarliestNodeUsed].SequenceNum)
9443 EarliestNodeUsed = i;
9446 // Find if it is better to use vectors or integers to load and store
9450 JointMemOpVT = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
9452 unsigned StoreBW = NumElem * ElementSizeBytes * 8;
9453 JointMemOpVT = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
9456 SDLoc LoadDL(LoadNodes[0].MemNode);
9457 SDLoc StoreDL(StoreNodes[0].MemNode);
9459 LoadSDNode *FirstLoad = cast<LoadSDNode>(LoadNodes[0].MemNode);
9460 SDValue NewLoad = DAG.getLoad(JointMemOpVT, LoadDL,
9461 FirstLoad->getChain(),
9462 FirstLoad->getBasePtr(),
9463 FirstLoad->getPointerInfo(),
9464 false, false, false,
9465 FirstLoad->getAlignment());
9467 SDValue NewStore = DAG.getStore(EarliestOp->getChain(), StoreDL, NewLoad,
9468 FirstInChain->getBasePtr(),
9469 FirstInChain->getPointerInfo(), false, false,
9470 FirstInChain->getAlignment());
9472 // Replace one of the loads with the new load.
9473 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[0].MemNode);
9474 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1),
9475 SDValue(NewLoad.getNode(), 1));
9477 // Remove the rest of the load chains.
9478 for (unsigned i = 1; i < NumElem ; ++i) {
9479 // Replace all chain users of the old load nodes with the chain of the new
9481 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[i].MemNode);
9482 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Ld->getChain());
9485 // Replace the first store with the new store.
9486 CombineTo(EarliestOp, NewStore);
9487 // Erase all other stores.
9488 for (unsigned i = 0; i < NumElem ; ++i) {
9489 // Remove all Store nodes.
9490 if (StoreNodes[i].MemNode == EarliestOp)
9492 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
9493 DAG.ReplaceAllUsesOfValueWith(SDValue(St, 0), St->getChain());
9494 removeFromWorkList(St);
9501 SDValue DAGCombiner::visitSTORE(SDNode *N) {
9502 StoreSDNode *ST = cast<StoreSDNode>(N);
9503 SDValue Chain = ST->getChain();
9504 SDValue Value = ST->getValue();
9505 SDValue Ptr = ST->getBasePtr();
9507 // If this is a store of a bit convert, store the input value if the
9508 // resultant store does not need a higher alignment than the original.
9509 if (Value.getOpcode() == ISD::BITCAST && !ST->isTruncatingStore() &&
9510 ST->isUnindexed()) {
9511 unsigned OrigAlign = ST->getAlignment();
9512 EVT SVT = Value.getOperand(0).getValueType();
9513 unsigned Align = TLI.getDataLayout()->
9514 getABITypeAlignment(SVT.getTypeForEVT(*DAG.getContext()));
9515 if (Align <= OrigAlign &&
9516 ((!LegalOperations && !ST->isVolatile()) ||
9517 TLI.isOperationLegalOrCustom(ISD::STORE, SVT)))
9518 return DAG.getStore(Chain, SDLoc(N), Value.getOperand(0),
9519 Ptr, ST->getPointerInfo(), ST->isVolatile(),
9520 ST->isNonTemporal(), OrigAlign,
9524 // Turn 'store undef, Ptr' -> nothing.
9525 if (Value.getOpcode() == ISD::UNDEF && ST->isUnindexed())
9528 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
9529 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Value)) {
9530 // NOTE: If the original store is volatile, this transform must not increase
9531 // the number of stores. For example, on x86-32 an f64 can be stored in one
9532 // processor operation but an i64 (which is not legal) requires two. So the
9533 // transform should not be done in this case.
9534 if (Value.getOpcode() != ISD::TargetConstantFP) {
9536 switch (CFP->getSimpleValueType(0).SimpleTy) {
9537 default: llvm_unreachable("Unknown FP type");
9538 case MVT::f16: // We don't do this for these yet.
9544 if ((isTypeLegal(MVT::i32) && !LegalOperations && !ST->isVolatile()) ||
9545 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
9546 Tmp = DAG.getConstant((uint32_t)CFP->getValueAPF().
9547 bitcastToAPInt().getZExtValue(), MVT::i32);
9548 return DAG.getStore(Chain, SDLoc(N), Tmp,
9549 Ptr, ST->getMemOperand());
9553 if ((TLI.isTypeLegal(MVT::i64) && !LegalOperations &&
9554 !ST->isVolatile()) ||
9555 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i64)) {
9556 Tmp = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
9557 getZExtValue(), MVT::i64);
9558 return DAG.getStore(Chain, SDLoc(N), Tmp,
9559 Ptr, ST->getMemOperand());
9562 if (!ST->isVolatile() &&
9563 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
9564 // Many FP stores are not made apparent until after legalize, e.g. for
9565 // argument passing. Since this is so common, custom legalize the
9566 // 64-bit integer store into two 32-bit stores.
9567 uint64_t Val = CFP->getValueAPF().bitcastToAPInt().getZExtValue();
9568 SDValue Lo = DAG.getConstant(Val & 0xFFFFFFFF, MVT::i32);
9569 SDValue Hi = DAG.getConstant(Val >> 32, MVT::i32);
9570 if (TLI.isBigEndian()) std::swap(Lo, Hi);
9572 unsigned Alignment = ST->getAlignment();
9573 bool isVolatile = ST->isVolatile();
9574 bool isNonTemporal = ST->isNonTemporal();
9575 const MDNode *TBAAInfo = ST->getTBAAInfo();
9577 SDValue St0 = DAG.getStore(Chain, SDLoc(ST), Lo,
9578 Ptr, ST->getPointerInfo(),
9579 isVolatile, isNonTemporal,
9580 ST->getAlignment(), TBAAInfo);
9581 Ptr = DAG.getNode(ISD::ADD, SDLoc(N), Ptr.getValueType(), Ptr,
9582 DAG.getConstant(4, Ptr.getValueType()));
9583 Alignment = MinAlign(Alignment, 4U);
9584 SDValue St1 = DAG.getStore(Chain, SDLoc(ST), Hi,
9585 Ptr, ST->getPointerInfo().getWithOffset(4),
9586 isVolatile, isNonTemporal,
9587 Alignment, TBAAInfo);
9588 return DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other,
9597 // Try to infer better alignment information than the store already has.
9598 if (OptLevel != CodeGenOpt::None && ST->isUnindexed()) {
9599 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
9600 if (Align > ST->getAlignment())
9601 return DAG.getTruncStore(Chain, SDLoc(N), Value,
9602 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
9603 ST->isVolatile(), ST->isNonTemporal(), Align,
9608 // Try transforming a pair floating point load / store ops to integer
9609 // load / store ops.
9610 SDValue NewST = TransformFPLoadStorePair(N);
9611 if (NewST.getNode())
9614 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA :
9615 TLI.getTargetMachine().getSubtarget<TargetSubtargetInfo>().useAA();
9617 if (CombinerAAOnlyFunc.getNumOccurrences() &&
9618 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
9621 if (UseAA && ST->isUnindexed()) {
9622 // Walk up chain skipping non-aliasing memory nodes.
9623 SDValue BetterChain = FindBetterChain(N, Chain);
9625 // If there is a better chain.
9626 if (Chain != BetterChain) {
9629 // Replace the chain to avoid dependency.
9630 if (ST->isTruncatingStore()) {
9631 ReplStore = DAG.getTruncStore(BetterChain, SDLoc(N), Value, Ptr,
9632 ST->getMemoryVT(), ST->getMemOperand());
9634 ReplStore = DAG.getStore(BetterChain, SDLoc(N), Value, Ptr,
9635 ST->getMemOperand());
9638 // Create token to keep both nodes around.
9639 SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N),
9640 MVT::Other, Chain, ReplStore);
9642 // Make sure the new and old chains are cleaned up.
9643 AddToWorkList(Token.getNode());
9645 // Don't add users to work list.
9646 return CombineTo(N, Token, false);
9650 // Try transforming N to an indexed store.
9651 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
9652 return SDValue(N, 0);
9654 // FIXME: is there such a thing as a truncating indexed store?
9655 if (ST->isTruncatingStore() && ST->isUnindexed() &&
9656 Value.getValueType().isInteger()) {
9657 // See if we can simplify the input to this truncstore with knowledge that
9658 // only the low bits are being used. For example:
9659 // "truncstore (or (shl x, 8), y), i8" -> "truncstore y, i8"
9661 GetDemandedBits(Value,
9662 APInt::getLowBitsSet(
9663 Value.getValueType().getScalarType().getSizeInBits(),
9664 ST->getMemoryVT().getScalarType().getSizeInBits()));
9665 AddToWorkList(Value.getNode());
9666 if (Shorter.getNode())
9667 return DAG.getTruncStore(Chain, SDLoc(N), Shorter,
9668 Ptr, ST->getMemoryVT(), ST->getMemOperand());
9670 // Otherwise, see if we can simplify the operation with
9671 // SimplifyDemandedBits, which only works if the value has a single use.
9672 if (SimplifyDemandedBits(Value,
9673 APInt::getLowBitsSet(
9674 Value.getValueType().getScalarType().getSizeInBits(),
9675 ST->getMemoryVT().getScalarType().getSizeInBits())))
9676 return SDValue(N, 0);
9679 // If this is a load followed by a store to the same location, then the store
9681 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Value)) {
9682 if (Ld->getBasePtr() == Ptr && ST->getMemoryVT() == Ld->getMemoryVT() &&
9683 ST->isUnindexed() && !ST->isVolatile() &&
9684 // There can't be any side effects between the load and store, such as
9686 Chain.reachesChainWithoutSideEffects(SDValue(Ld, 1))) {
9687 // The store is dead, remove it.
9692 // If this is an FP_ROUND or TRUNC followed by a store, fold this into a
9693 // truncating store. We can do this even if this is already a truncstore.
9694 if ((Value.getOpcode() == ISD::FP_ROUND || Value.getOpcode() == ISD::TRUNCATE)
9695 && Value.getNode()->hasOneUse() && ST->isUnindexed() &&
9696 TLI.isTruncStoreLegal(Value.getOperand(0).getValueType(),
9697 ST->getMemoryVT())) {
9698 return DAG.getTruncStore(Chain, SDLoc(N), Value.getOperand(0),
9699 Ptr, ST->getMemoryVT(), ST->getMemOperand());
9702 // Only perform this optimization before the types are legal, because we
9703 // don't want to perform this optimization on every DAGCombine invocation.
9705 bool EverChanged = false;
9708 // There can be multiple store sequences on the same chain.
9709 // Keep trying to merge store sequences until we are unable to do so
9710 // or until we merge the last store on the chain.
9711 bool Changed = MergeConsecutiveStores(ST);
9712 EverChanged |= Changed;
9713 if (!Changed) break;
9714 } while (ST->getOpcode() != ISD::DELETED_NODE);
9717 return SDValue(N, 0);
9720 return ReduceLoadOpStoreWidth(N);
9723 SDValue DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) {
9724 SDValue InVec = N->getOperand(0);
9725 SDValue InVal = N->getOperand(1);
9726 SDValue EltNo = N->getOperand(2);
9729 // If the inserted element is an UNDEF, just use the input vector.
9730 if (InVal.getOpcode() == ISD::UNDEF)
9733 EVT VT = InVec.getValueType();
9735 // If we can't generate a legal BUILD_VECTOR, exit
9736 if (LegalOperations && !TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
9739 // Check that we know which element is being inserted
9740 if (!isa<ConstantSDNode>(EltNo))
9742 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
9744 // Canonicalize insert_vector_elt dag nodes.
9746 // (insert_vector_elt (insert_vector_elt A, Idx0), Idx1)
9747 // -> (insert_vector_elt (insert_vector_elt A, Idx1), Idx0)
9749 // Do this only if the child insert_vector node has one use; also
9750 // do this only if indices are both constants and Idx1 < Idx0.
9751 if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT && InVec.hasOneUse()
9752 && isa<ConstantSDNode>(InVec.getOperand(2))) {
9754 cast<ConstantSDNode>(InVec.getOperand(2))->getZExtValue();
9755 if (Elt < OtherElt) {
9757 SDValue NewOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(N), VT,
9758 InVec.getOperand(0), InVal, EltNo);
9759 AddToWorkList(NewOp.getNode());
9760 return DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(InVec.getNode()),
9761 VT, NewOp, InVec.getOperand(1), InVec.getOperand(2));
9765 // Check that the operand is a BUILD_VECTOR (or UNDEF, which can essentially
9766 // be converted to a BUILD_VECTOR). Fill in the Ops vector with the
9768 SmallVector<SDValue, 8> Ops;
9769 // Do not combine these two vectors if the output vector will not replace
9770 // the input vector.
9771 if (InVec.getOpcode() == ISD::BUILD_VECTOR && InVec.hasOneUse()) {
9772 Ops.append(InVec.getNode()->op_begin(),
9773 InVec.getNode()->op_end());
9774 } else if (InVec.getOpcode() == ISD::UNDEF) {
9775 unsigned NElts = VT.getVectorNumElements();
9776 Ops.append(NElts, DAG.getUNDEF(InVal.getValueType()));
9781 // Insert the element
9782 if (Elt < Ops.size()) {
9783 // All the operands of BUILD_VECTOR must have the same type;
9784 // we enforce that here.
9785 EVT OpVT = Ops[0].getValueType();
9786 if (InVal.getValueType() != OpVT)
9787 InVal = OpVT.bitsGT(InVal.getValueType()) ?
9788 DAG.getNode(ISD::ANY_EXTEND, dl, OpVT, InVal) :
9789 DAG.getNode(ISD::TRUNCATE, dl, OpVT, InVal);
9793 // Return the new vector
9794 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
9797 SDValue DAGCombiner::ReplaceExtractVectorEltOfLoadWithNarrowedLoad(
9798 SDNode *EVE, EVT InVecVT, SDValue EltNo, LoadSDNode *OriginalLoad) {
9799 EVT ResultVT = EVE->getValueType(0);
9800 EVT VecEltVT = InVecVT.getVectorElementType();
9801 unsigned Align = OriginalLoad->getAlignment();
9802 unsigned NewAlign = TLI.getDataLayout()->getABITypeAlignment(
9803 VecEltVT.getTypeForEVT(*DAG.getContext()));
9805 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, VecEltVT))
9810 SDValue NewPtr = OriginalLoad->getBasePtr();
9812 EVT PtrType = NewPtr.getValueType();
9813 MachinePointerInfo MPI;
9814 if (auto *ConstEltNo = dyn_cast<ConstantSDNode>(EltNo)) {
9815 int Elt = ConstEltNo->getZExtValue();
9816 unsigned PtrOff = VecEltVT.getSizeInBits() * Elt / 8;
9817 if (TLI.isBigEndian())
9818 PtrOff = InVecVT.getSizeInBits() / 8 - PtrOff;
9819 Offset = DAG.getConstant(PtrOff, PtrType);
9820 MPI = OriginalLoad->getPointerInfo().getWithOffset(PtrOff);
9822 Offset = DAG.getNode(
9823 ISD::MUL, SDLoc(EVE), EltNo.getValueType(), EltNo,
9824 DAG.getConstant(VecEltVT.getStoreSize(), EltNo.getValueType()));
9825 if (TLI.isBigEndian())
9826 Offset = DAG.getNode(
9827 ISD::SUB, SDLoc(EVE), EltNo.getValueType(),
9828 DAG.getConstant(InVecVT.getStoreSize(), EltNo.getValueType()), Offset);
9829 MPI = OriginalLoad->getPointerInfo();
9831 NewPtr = DAG.getNode(ISD::ADD, SDLoc(EVE), PtrType, NewPtr, Offset);
9833 // The replacement we need to do here is a little tricky: we need to
9834 // replace an extractelement of a load with a load.
9835 // Use ReplaceAllUsesOfValuesWith to do the replacement.
9836 // Note that this replacement assumes that the extractvalue is the only
9837 // use of the load; that's okay because we don't want to perform this
9838 // transformation in other cases anyway.
9841 if (ResultVT.bitsGT(VecEltVT)) {
9842 // If the result type of vextract is wider than the load, then issue an
9843 // extending load instead.
9844 ISD::LoadExtType ExtType = TLI.isLoadExtLegal(ISD::ZEXTLOAD, VecEltVT)
9847 Load = DAG.getExtLoad(ExtType, SDLoc(EVE), ResultVT, OriginalLoad->getChain(),
9848 NewPtr, MPI, VecEltVT, OriginalLoad->isVolatile(),
9849 OriginalLoad->isNonTemporal(), Align,
9850 OriginalLoad->getTBAAInfo());
9851 Chain = Load.getValue(1);
9854 VecEltVT, SDLoc(EVE), OriginalLoad->getChain(), NewPtr, MPI,
9855 OriginalLoad->isVolatile(), OriginalLoad->isNonTemporal(),
9856 OriginalLoad->isInvariant(), Align, OriginalLoad->getTBAAInfo());
9857 Chain = Load.getValue(1);
9858 if (ResultVT.bitsLT(VecEltVT))
9859 Load = DAG.getNode(ISD::TRUNCATE, SDLoc(EVE), ResultVT, Load);
9861 Load = DAG.getNode(ISD::BITCAST, SDLoc(EVE), ResultVT, Load);
9863 WorkListRemover DeadNodes(*this);
9864 SDValue From[] = { SDValue(EVE, 0), SDValue(OriginalLoad, 1) };
9865 SDValue To[] = { Load, Chain };
9866 DAG.ReplaceAllUsesOfValuesWith(From, To, 2);
9867 // Since we're explicitly calling ReplaceAllUses, add the new node to the
9868 // worklist explicitly as well.
9869 AddToWorkList(Load.getNode());
9870 AddUsersToWorkList(Load.getNode()); // Add users too
9871 // Make sure to revisit this node to clean it up; it will usually be dead.
9874 return SDValue(EVE, 0);
9877 SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) {
9878 // (vextract (scalar_to_vector val, 0) -> val
9879 SDValue InVec = N->getOperand(0);
9880 EVT VT = InVec.getValueType();
9881 EVT NVT = N->getValueType(0);
9883 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
9884 // Check if the result type doesn't match the inserted element type. A
9885 // SCALAR_TO_VECTOR may truncate the inserted element and the
9886 // EXTRACT_VECTOR_ELT may widen the extracted vector.
9887 SDValue InOp = InVec.getOperand(0);
9888 if (InOp.getValueType() != NVT) {
9889 assert(InOp.getValueType().isInteger() && NVT.isInteger());
9890 return DAG.getSExtOrTrunc(InOp, SDLoc(InVec), NVT);
9895 SDValue EltNo = N->getOperand(1);
9896 bool ConstEltNo = isa<ConstantSDNode>(EltNo);
9898 // Transform: (EXTRACT_VECTOR_ELT( VECTOR_SHUFFLE )) -> EXTRACT_VECTOR_ELT.
9899 // We only perform this optimization before the op legalization phase because
9900 // we may introduce new vector instructions which are not backed by TD
9901 // patterns. For example on AVX, extracting elements from a wide vector
9902 // without using extract_subvector. However, if we can find an underlying
9903 // scalar value, then we can always use that.
9904 if (InVec.getOpcode() == ISD::VECTOR_SHUFFLE
9906 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
9907 int NumElem = VT.getVectorNumElements();
9908 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(InVec);
9909 // Find the new index to extract from.
9910 int OrigElt = SVOp->getMaskElt(Elt);
9912 // Extracting an undef index is undef.
9914 return DAG.getUNDEF(NVT);
9916 // Select the right vector half to extract from.
9918 if (OrigElt < NumElem) {
9919 SVInVec = InVec->getOperand(0);
9921 SVInVec = InVec->getOperand(1);
9925 if (SVInVec.getOpcode() == ISD::BUILD_VECTOR) {
9926 SDValue InOp = SVInVec.getOperand(OrigElt);
9927 if (InOp.getValueType() != NVT) {
9928 assert(InOp.getValueType().isInteger() && NVT.isInteger());
9929 InOp = DAG.getSExtOrTrunc(InOp, SDLoc(SVInVec), NVT);
9935 // FIXME: We should handle recursing on other vector shuffles and
9936 // scalar_to_vector here as well.
9938 if (!LegalOperations) {
9939 EVT IndexTy = TLI.getVectorIdxTy();
9940 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(N), NVT,
9941 SVInVec, DAG.getConstant(OrigElt, IndexTy));
9945 bool BCNumEltsChanged = false;
9946 EVT ExtVT = VT.getVectorElementType();
9949 // If the result of load has to be truncated, then it's not necessarily
9951 if (NVT.bitsLT(LVT) && !TLI.isTruncateFree(LVT, NVT))
9954 if (InVec.getOpcode() == ISD::BITCAST) {
9955 // Don't duplicate a load with other uses.
9956 if (!InVec.hasOneUse())
9959 EVT BCVT = InVec.getOperand(0).getValueType();
9960 if (!BCVT.isVector() || ExtVT.bitsGT(BCVT.getVectorElementType()))
9962 if (VT.getVectorNumElements() != BCVT.getVectorNumElements())
9963 BCNumEltsChanged = true;
9964 InVec = InVec.getOperand(0);
9965 ExtVT = BCVT.getVectorElementType();
9968 // (vextract (vN[if]M load $addr), i) -> ([if]M load $addr + i * size)
9969 if (!LegalOperations && !ConstEltNo && InVec.hasOneUse() &&
9970 ISD::isNormalLoad(InVec.getNode())) {
9971 SDValue Index = N->getOperand(1);
9972 if (LoadSDNode *OrigLoad = dyn_cast<LoadSDNode>(InVec))
9973 return ReplaceExtractVectorEltOfLoadWithNarrowedLoad(N, VT, Index,
9977 // Perform only after legalization to ensure build_vector / vector_shuffle
9978 // optimizations have already been done.
9979 if (!LegalOperations) return SDValue();
9981 // (vextract (v4f32 load $addr), c) -> (f32 load $addr+c*size)
9982 // (vextract (v4f32 s2v (f32 load $addr)), c) -> (f32 load $addr+c*size)
9983 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), 0) -> (f32 load $addr)
9986 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
9988 LoadSDNode *LN0 = nullptr;
9989 const ShuffleVectorSDNode *SVN = nullptr;
9990 if (ISD::isNormalLoad(InVec.getNode())) {
9991 LN0 = cast<LoadSDNode>(InVec);
9992 } else if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR &&
9993 InVec.getOperand(0).getValueType() == ExtVT &&
9994 ISD::isNormalLoad(InVec.getOperand(0).getNode())) {
9995 // Don't duplicate a load with other uses.
9996 if (!InVec.hasOneUse())
9999 LN0 = cast<LoadSDNode>(InVec.getOperand(0));
10000 } else if ((SVN = dyn_cast<ShuffleVectorSDNode>(InVec))) {
10001 // (vextract (vector_shuffle (load $addr), v2, <1, u, u, u>), 1)
10003 // (load $addr+1*size)
10005 // Don't duplicate a load with other uses.
10006 if (!InVec.hasOneUse())
10009 // If the bit convert changed the number of elements, it is unsafe
10010 // to examine the mask.
10011 if (BCNumEltsChanged)
10014 // Select the input vector, guarding against out of range extract vector.
10015 unsigned NumElems = VT.getVectorNumElements();
10016 int Idx = (Elt > (int)NumElems) ? -1 : SVN->getMaskElt(Elt);
10017 InVec = (Idx < (int)NumElems) ? InVec.getOperand(0) : InVec.getOperand(1);
10019 if (InVec.getOpcode() == ISD::BITCAST) {
10020 // Don't duplicate a load with other uses.
10021 if (!InVec.hasOneUse())
10024 InVec = InVec.getOperand(0);
10026 if (ISD::isNormalLoad(InVec.getNode())) {
10027 LN0 = cast<LoadSDNode>(InVec);
10028 Elt = (Idx < (int)NumElems) ? Idx : Idx - (int)NumElems;
10029 EltNo = DAG.getConstant(Elt, EltNo.getValueType());
10033 // Make sure we found a non-volatile load and the extractelement is
10035 if (!LN0 || !LN0->hasNUsesOfValue(1,0) || LN0->isVolatile())
10038 // If Idx was -1 above, Elt is going to be -1, so just return undef.
10040 return DAG.getUNDEF(LVT);
10042 return ReplaceExtractVectorEltOfLoadWithNarrowedLoad(N, VT, EltNo, LN0);
10048 // Simplify (build_vec (ext )) to (bitcast (build_vec ))
10049 SDValue DAGCombiner::reduceBuildVecExtToExtBuildVec(SDNode *N) {
10050 // We perform this optimization post type-legalization because
10051 // the type-legalizer often scalarizes integer-promoted vectors.
10052 // Performing this optimization before may create bit-casts which
10053 // will be type-legalized to complex code sequences.
10054 // We perform this optimization only before the operation legalizer because we
10055 // may introduce illegal operations.
10056 if (Level != AfterLegalizeVectorOps && Level != AfterLegalizeTypes)
10059 unsigned NumInScalars = N->getNumOperands();
10061 EVT VT = N->getValueType(0);
10063 // Check to see if this is a BUILD_VECTOR of a bunch of values
10064 // which come from any_extend or zero_extend nodes. If so, we can create
10065 // a new BUILD_VECTOR using bit-casts which may enable other BUILD_VECTOR
10066 // optimizations. We do not handle sign-extend because we can't fill the sign
10068 EVT SourceType = MVT::Other;
10069 bool AllAnyExt = true;
10071 for (unsigned i = 0; i != NumInScalars; ++i) {
10072 SDValue In = N->getOperand(i);
10073 // Ignore undef inputs.
10074 if (In.getOpcode() == ISD::UNDEF) continue;
10076 bool AnyExt = In.getOpcode() == ISD::ANY_EXTEND;
10077 bool ZeroExt = In.getOpcode() == ISD::ZERO_EXTEND;
10079 // Abort if the element is not an extension.
10080 if (!ZeroExt && !AnyExt) {
10081 SourceType = MVT::Other;
10085 // The input is a ZeroExt or AnyExt. Check the original type.
10086 EVT InTy = In.getOperand(0).getValueType();
10088 // Check that all of the widened source types are the same.
10089 if (SourceType == MVT::Other)
10092 else if (InTy != SourceType) {
10093 // Multiple income types. Abort.
10094 SourceType = MVT::Other;
10098 // Check if all of the extends are ANY_EXTENDs.
10099 AllAnyExt &= AnyExt;
10102 // In order to have valid types, all of the inputs must be extended from the
10103 // same source type and all of the inputs must be any or zero extend.
10104 // Scalar sizes must be a power of two.
10105 EVT OutScalarTy = VT.getScalarType();
10106 bool ValidTypes = SourceType != MVT::Other &&
10107 isPowerOf2_32(OutScalarTy.getSizeInBits()) &&
10108 isPowerOf2_32(SourceType.getSizeInBits());
10110 // Create a new simpler BUILD_VECTOR sequence which other optimizations can
10111 // turn into a single shuffle instruction.
10115 bool isLE = TLI.isLittleEndian();
10116 unsigned ElemRatio = OutScalarTy.getSizeInBits()/SourceType.getSizeInBits();
10117 assert(ElemRatio > 1 && "Invalid element size ratio");
10118 SDValue Filler = AllAnyExt ? DAG.getUNDEF(SourceType):
10119 DAG.getConstant(0, SourceType);
10121 unsigned NewBVElems = ElemRatio * VT.getVectorNumElements();
10122 SmallVector<SDValue, 8> Ops(NewBVElems, Filler);
10124 // Populate the new build_vector
10125 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
10126 SDValue Cast = N->getOperand(i);
10127 assert((Cast.getOpcode() == ISD::ANY_EXTEND ||
10128 Cast.getOpcode() == ISD::ZERO_EXTEND ||
10129 Cast.getOpcode() == ISD::UNDEF) && "Invalid cast opcode");
10131 if (Cast.getOpcode() == ISD::UNDEF)
10132 In = DAG.getUNDEF(SourceType);
10134 In = Cast->getOperand(0);
10135 unsigned Index = isLE ? (i * ElemRatio) :
10136 (i * ElemRatio + (ElemRatio - 1));
10138 assert(Index < Ops.size() && "Invalid index");
10142 // The type of the new BUILD_VECTOR node.
10143 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), SourceType, NewBVElems);
10144 assert(VecVT.getSizeInBits() == VT.getSizeInBits() &&
10145 "Invalid vector size");
10146 // Check if the new vector type is legal.
10147 if (!isTypeLegal(VecVT)) return SDValue();
10149 // Make the new BUILD_VECTOR.
10150 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
10152 // The new BUILD_VECTOR node has the potential to be further optimized.
10153 AddToWorkList(BV.getNode());
10154 // Bitcast to the desired type.
10155 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
10158 SDValue DAGCombiner::reduceBuildVecConvertToConvertBuildVec(SDNode *N) {
10159 EVT VT = N->getValueType(0);
10161 unsigned NumInScalars = N->getNumOperands();
10164 EVT SrcVT = MVT::Other;
10165 unsigned Opcode = ISD::DELETED_NODE;
10166 unsigned NumDefs = 0;
10168 for (unsigned i = 0; i != NumInScalars; ++i) {
10169 SDValue In = N->getOperand(i);
10170 unsigned Opc = In.getOpcode();
10172 if (Opc == ISD::UNDEF)
10175 // If all scalar values are floats and converted from integers.
10176 if (Opcode == ISD::DELETED_NODE &&
10177 (Opc == ISD::UINT_TO_FP || Opc == ISD::SINT_TO_FP)) {
10184 EVT InVT = In.getOperand(0).getValueType();
10186 // If all scalar values are typed differently, bail out. It's chosen to
10187 // simplify BUILD_VECTOR of integer types.
10188 if (SrcVT == MVT::Other)
10195 // If the vector has just one element defined, it's not worth to fold it into
10196 // a vectorized one.
10200 assert((Opcode == ISD::UINT_TO_FP || Opcode == ISD::SINT_TO_FP)
10201 && "Should only handle conversion from integer to float.");
10202 assert(SrcVT != MVT::Other && "Cannot determine source type!");
10204 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SrcVT, NumInScalars);
10206 if (!TLI.isOperationLegalOrCustom(Opcode, NVT))
10209 SmallVector<SDValue, 8> Opnds;
10210 for (unsigned i = 0; i != NumInScalars; ++i) {
10211 SDValue In = N->getOperand(i);
10213 if (In.getOpcode() == ISD::UNDEF)
10214 Opnds.push_back(DAG.getUNDEF(SrcVT));
10216 Opnds.push_back(In.getOperand(0));
10218 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NVT, Opnds);
10219 AddToWorkList(BV.getNode());
10221 return DAG.getNode(Opcode, dl, VT, BV);
10224 SDValue DAGCombiner::visitBUILD_VECTOR(SDNode *N) {
10225 unsigned NumInScalars = N->getNumOperands();
10227 EVT VT = N->getValueType(0);
10229 // A vector built entirely of undefs is undef.
10230 if (ISD::allOperandsUndef(N))
10231 return DAG.getUNDEF(VT);
10233 SDValue V = reduceBuildVecExtToExtBuildVec(N);
10237 V = reduceBuildVecConvertToConvertBuildVec(N);
10241 // Check to see if this is a BUILD_VECTOR of a bunch of EXTRACT_VECTOR_ELT
10242 // operations. If so, and if the EXTRACT_VECTOR_ELT vector inputs come from
10243 // at most two distinct vectors, turn this into a shuffle node.
10245 // May only combine to shuffle after legalize if shuffle is legal.
10246 if (LegalOperations &&
10247 !TLI.isOperationLegalOrCustom(ISD::VECTOR_SHUFFLE, VT))
10250 SDValue VecIn1, VecIn2;
10251 for (unsigned i = 0; i != NumInScalars; ++i) {
10252 // Ignore undef inputs.
10253 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
10255 // If this input is something other than a EXTRACT_VECTOR_ELT with a
10256 // constant index, bail out.
10257 if (N->getOperand(i).getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
10258 !isa<ConstantSDNode>(N->getOperand(i).getOperand(1))) {
10259 VecIn1 = VecIn2 = SDValue(nullptr, 0);
10263 // We allow up to two distinct input vectors.
10264 SDValue ExtractedFromVec = N->getOperand(i).getOperand(0);
10265 if (ExtractedFromVec == VecIn1 || ExtractedFromVec == VecIn2)
10268 if (!VecIn1.getNode()) {
10269 VecIn1 = ExtractedFromVec;
10270 } else if (!VecIn2.getNode()) {
10271 VecIn2 = ExtractedFromVec;
10273 // Too many inputs.
10274 VecIn1 = VecIn2 = SDValue(nullptr, 0);
10279 // If everything is good, we can make a shuffle operation.
10280 if (VecIn1.getNode()) {
10281 SmallVector<int, 8> Mask;
10282 for (unsigned i = 0; i != NumInScalars; ++i) {
10283 if (N->getOperand(i).getOpcode() == ISD::UNDEF) {
10284 Mask.push_back(-1);
10288 // If extracting from the first vector, just use the index directly.
10289 SDValue Extract = N->getOperand(i);
10290 SDValue ExtVal = Extract.getOperand(1);
10291 if (Extract.getOperand(0) == VecIn1) {
10292 unsigned ExtIndex = cast<ConstantSDNode>(ExtVal)->getZExtValue();
10293 if (ExtIndex > VT.getVectorNumElements())
10296 Mask.push_back(ExtIndex);
10300 // Otherwise, use InIdx + VecSize
10301 unsigned Idx = cast<ConstantSDNode>(ExtVal)->getZExtValue();
10302 Mask.push_back(Idx+NumInScalars);
10305 // We can't generate a shuffle node with mismatched input and output types.
10306 // Attempt to transform a single input vector to the correct type.
10307 if ((VT != VecIn1.getValueType())) {
10308 // We don't support shuffeling between TWO values of different types.
10309 if (VecIn2.getNode())
10312 // We only support widening of vectors which are half the size of the
10313 // output registers. For example XMM->YMM widening on X86 with AVX.
10314 if (VecIn1.getValueType().getSizeInBits()*2 != VT.getSizeInBits())
10317 // If the input vector type has a different base type to the output
10318 // vector type, bail out.
10319 if (VecIn1.getValueType().getVectorElementType() !=
10320 VT.getVectorElementType())
10323 // Widen the input vector by adding undef values.
10324 VecIn1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
10325 VecIn1, DAG.getUNDEF(VecIn1.getValueType()));
10328 // If VecIn2 is unused then change it to undef.
10329 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
10331 // Check that we were able to transform all incoming values to the same
10333 if (VecIn2.getValueType() != VecIn1.getValueType() ||
10334 VecIn1.getValueType() != VT)
10337 // Only type-legal BUILD_VECTOR nodes are converted to shuffle nodes.
10338 if (!isTypeLegal(VT))
10341 // Return the new VECTOR_SHUFFLE node.
10345 return DAG.getVectorShuffle(VT, dl, Ops[0], Ops[1], &Mask[0]);
10351 SDValue DAGCombiner::visitCONCAT_VECTORS(SDNode *N) {
10352 // TODO: Check to see if this is a CONCAT_VECTORS of a bunch of
10353 // EXTRACT_SUBVECTOR operations. If so, and if the EXTRACT_SUBVECTOR vector
10354 // inputs come from at most two distinct vectors, turn this into a shuffle
10357 // If we only have one input vector, we don't need to do any concatenation.
10358 if (N->getNumOperands() == 1)
10359 return N->getOperand(0);
10361 // Check if all of the operands are undefs.
10362 EVT VT = N->getValueType(0);
10363 if (ISD::allOperandsUndef(N))
10364 return DAG.getUNDEF(VT);
10366 // Optimize concat_vectors where one of the vectors is undef.
10367 if (N->getNumOperands() == 2 &&
10368 N->getOperand(1)->getOpcode() == ISD::UNDEF) {
10369 SDValue In = N->getOperand(0);
10370 assert(In.getValueType().isVector() && "Must concat vectors");
10372 // Transform: concat_vectors(scalar, undef) -> scalar_to_vector(sclr).
10373 if (In->getOpcode() == ISD::BITCAST &&
10374 !In->getOperand(0)->getValueType(0).isVector()) {
10375 SDValue Scalar = In->getOperand(0);
10376 EVT SclTy = Scalar->getValueType(0);
10378 if (!SclTy.isFloatingPoint() && !SclTy.isInteger())
10381 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SclTy,
10382 VT.getSizeInBits() / SclTy.getSizeInBits());
10383 if (!TLI.isTypeLegal(NVT) || !TLI.isTypeLegal(Scalar.getValueType()))
10386 SDLoc dl = SDLoc(N);
10387 SDValue Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, NVT, Scalar);
10388 return DAG.getNode(ISD::BITCAST, dl, VT, Res);
10392 // fold (concat_vectors (BUILD_VECTOR A, B, ...), (BUILD_VECTOR C, D, ...))
10393 // -> (BUILD_VECTOR A, B, ..., C, D, ...)
10394 if (N->getNumOperands() == 2 &&
10395 N->getOperand(0).getOpcode() == ISD::BUILD_VECTOR &&
10396 N->getOperand(1).getOpcode() == ISD::BUILD_VECTOR) {
10397 EVT VT = N->getValueType(0);
10398 SDValue N0 = N->getOperand(0);
10399 SDValue N1 = N->getOperand(1);
10400 SmallVector<SDValue, 8> Opnds;
10401 unsigned BuildVecNumElts = N0.getNumOperands();
10403 EVT SclTy0 = N0.getOperand(0)->getValueType(0);
10404 EVT SclTy1 = N1.getOperand(0)->getValueType(0);
10405 if (SclTy0.isFloatingPoint()) {
10406 for (unsigned i = 0; i != BuildVecNumElts; ++i)
10407 Opnds.push_back(N0.getOperand(i));
10408 for (unsigned i = 0; i != BuildVecNumElts; ++i)
10409 Opnds.push_back(N1.getOperand(i));
10411 // If BUILD_VECTOR are from built from integer, they may have different
10412 // operand types. Get the smaller type and truncate all operands to it.
10413 EVT MinTy = SclTy0.bitsLE(SclTy1) ? SclTy0 : SclTy1;
10414 for (unsigned i = 0; i != BuildVecNumElts; ++i)
10415 Opnds.push_back(DAG.getNode(ISD::TRUNCATE, SDLoc(N), MinTy,
10416 N0.getOperand(i)));
10417 for (unsigned i = 0; i != BuildVecNumElts; ++i)
10418 Opnds.push_back(DAG.getNode(ISD::TRUNCATE, SDLoc(N), MinTy,
10419 N1.getOperand(i)));
10422 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Opnds);
10425 // Type legalization of vectors and DAG canonicalization of SHUFFLE_VECTOR
10426 // nodes often generate nop CONCAT_VECTOR nodes.
10427 // Scan the CONCAT_VECTOR operands and look for a CONCAT operations that
10428 // place the incoming vectors at the exact same location.
10429 SDValue SingleSource = SDValue();
10430 unsigned PartNumElem = N->getOperand(0).getValueType().getVectorNumElements();
10432 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
10433 SDValue Op = N->getOperand(i);
10435 if (Op.getOpcode() == ISD::UNDEF)
10438 // Check if this is the identity extract:
10439 if (Op.getOpcode() != ISD::EXTRACT_SUBVECTOR)
10442 // Find the single incoming vector for the extract_subvector.
10443 if (SingleSource.getNode()) {
10444 if (Op.getOperand(0) != SingleSource)
10447 SingleSource = Op.getOperand(0);
10449 // Check the source type is the same as the type of the result.
10450 // If not, this concat may extend the vector, so we can not
10451 // optimize it away.
10452 if (SingleSource.getValueType() != N->getValueType(0))
10456 unsigned IdentityIndex = i * PartNumElem;
10457 ConstantSDNode *CS = dyn_cast<ConstantSDNode>(Op.getOperand(1));
10458 // The extract index must be constant.
10462 // Check that we are reading from the identity index.
10463 if (CS->getZExtValue() != IdentityIndex)
10467 if (SingleSource.getNode())
10468 return SingleSource;
10473 SDValue DAGCombiner::visitEXTRACT_SUBVECTOR(SDNode* N) {
10474 EVT NVT = N->getValueType(0);
10475 SDValue V = N->getOperand(0);
10477 if (V->getOpcode() == ISD::CONCAT_VECTORS) {
10479 // (extract_subvec (concat V1, V2, ...), i)
10482 // Only operand 0 is checked as 'concat' assumes all inputs of the same
10484 if (V->getOperand(0).getValueType() != NVT)
10486 unsigned Idx = dyn_cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
10487 unsigned NumElems = NVT.getVectorNumElements();
10488 assert((Idx % NumElems) == 0 &&
10489 "IDX in concat is not a multiple of the result vector length.");
10490 return V->getOperand(Idx / NumElems);
10494 if (V->getOpcode() == ISD::BITCAST)
10495 V = V.getOperand(0);
10497 if (V->getOpcode() == ISD::INSERT_SUBVECTOR) {
10499 // Handle only simple case where vector being inserted and vector
10500 // being extracted are of same type, and are half size of larger vectors.
10501 EVT BigVT = V->getOperand(0).getValueType();
10502 EVT SmallVT = V->getOperand(1).getValueType();
10503 if (!NVT.bitsEq(SmallVT) || NVT.getSizeInBits()*2 != BigVT.getSizeInBits())
10506 // Only handle cases where both indexes are constants with the same type.
10507 ConstantSDNode *ExtIdx = dyn_cast<ConstantSDNode>(N->getOperand(1));
10508 ConstantSDNode *InsIdx = dyn_cast<ConstantSDNode>(V->getOperand(2));
10510 if (InsIdx && ExtIdx &&
10511 InsIdx->getValueType(0).getSizeInBits() <= 64 &&
10512 ExtIdx->getValueType(0).getSizeInBits() <= 64) {
10514 // (extract_subvec (insert_subvec V1, V2, InsIdx), ExtIdx)
10516 // indices are equal or bit offsets are equal => V1
10517 // otherwise => (extract_subvec V1, ExtIdx)
10518 if (InsIdx->getZExtValue() * SmallVT.getScalarType().getSizeInBits() ==
10519 ExtIdx->getZExtValue() * NVT.getScalarType().getSizeInBits())
10520 return DAG.getNode(ISD::BITCAST, dl, NVT, V->getOperand(1));
10521 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, NVT,
10522 DAG.getNode(ISD::BITCAST, dl,
10523 N->getOperand(0).getValueType(),
10524 V->getOperand(0)), N->getOperand(1));
10531 // Tries to turn a shuffle of two CONCAT_VECTORS into a single concat.
10532 static SDValue partitionShuffleOfConcats(SDNode *N, SelectionDAG &DAG) {
10533 EVT VT = N->getValueType(0);
10534 unsigned NumElts = VT.getVectorNumElements();
10536 SDValue N0 = N->getOperand(0);
10537 SDValue N1 = N->getOperand(1);
10538 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
10540 SmallVector<SDValue, 4> Ops;
10541 EVT ConcatVT = N0.getOperand(0).getValueType();
10542 unsigned NumElemsPerConcat = ConcatVT.getVectorNumElements();
10543 unsigned NumConcats = NumElts / NumElemsPerConcat;
10545 // Look at every vector that's inserted. We're looking for exact
10546 // subvector-sized copies from a concatenated vector
10547 for (unsigned I = 0; I != NumConcats; ++I) {
10548 // Make sure we're dealing with a copy.
10549 unsigned Begin = I * NumElemsPerConcat;
10550 bool AllUndef = true, NoUndef = true;
10551 for (unsigned J = Begin; J != Begin + NumElemsPerConcat; ++J) {
10552 if (SVN->getMaskElt(J) >= 0)
10559 if (SVN->getMaskElt(Begin) % NumElemsPerConcat != 0)
10562 for (unsigned J = 1; J != NumElemsPerConcat; ++J)
10563 if (SVN->getMaskElt(Begin + J - 1) + 1 != SVN->getMaskElt(Begin + J))
10566 unsigned FirstElt = SVN->getMaskElt(Begin) / NumElemsPerConcat;
10567 if (FirstElt < N0.getNumOperands())
10568 Ops.push_back(N0.getOperand(FirstElt));
10570 Ops.push_back(N1.getOperand(FirstElt - N0.getNumOperands()));
10572 } else if (AllUndef) {
10573 Ops.push_back(DAG.getUNDEF(N0.getOperand(0).getValueType()));
10574 } else { // Mixed with general masks and undefs, can't do optimization.
10579 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Ops);
10582 SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
10583 EVT VT = N->getValueType(0);
10584 unsigned NumElts = VT.getVectorNumElements();
10586 SDValue N0 = N->getOperand(0);
10587 SDValue N1 = N->getOperand(1);
10589 assert(N0.getValueType() == VT && "Vector shuffle must be normalized in DAG");
10591 // Canonicalize shuffle undef, undef -> undef
10592 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
10593 return DAG.getUNDEF(VT);
10595 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
10597 // Canonicalize shuffle v, v -> v, undef
10599 SmallVector<int, 8> NewMask;
10600 for (unsigned i = 0; i != NumElts; ++i) {
10601 int Idx = SVN->getMaskElt(i);
10602 if (Idx >= (int)NumElts) Idx -= NumElts;
10603 NewMask.push_back(Idx);
10605 return DAG.getVectorShuffle(VT, SDLoc(N), N0, DAG.getUNDEF(VT),
10609 // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask.
10610 if (N0.getOpcode() == ISD::UNDEF) {
10611 SmallVector<int, 8> NewMask;
10612 for (unsigned i = 0; i != NumElts; ++i) {
10613 int Idx = SVN->getMaskElt(i);
10615 if (Idx >= (int)NumElts)
10618 Idx = -1; // remove reference to lhs
10620 NewMask.push_back(Idx);
10622 return DAG.getVectorShuffle(VT, SDLoc(N), N1, DAG.getUNDEF(VT),
10626 // Remove references to rhs if it is undef
10627 if (N1.getOpcode() == ISD::UNDEF) {
10628 bool Changed = false;
10629 SmallVector<int, 8> NewMask;
10630 for (unsigned i = 0; i != NumElts; ++i) {
10631 int Idx = SVN->getMaskElt(i);
10632 if (Idx >= (int)NumElts) {
10636 NewMask.push_back(Idx);
10639 return DAG.getVectorShuffle(VT, SDLoc(N), N0, N1, &NewMask[0]);
10642 // If it is a splat, check if the argument vector is another splat or a
10643 // build_vector with all scalar elements the same.
10644 if (SVN->isSplat() && SVN->getSplatIndex() < (int)NumElts) {
10645 SDNode *V = N0.getNode();
10647 // If this is a bit convert that changes the element type of the vector but
10648 // not the number of vector elements, look through it. Be careful not to
10649 // look though conversions that change things like v4f32 to v2f64.
10650 if (V->getOpcode() == ISD::BITCAST) {
10651 SDValue ConvInput = V->getOperand(0);
10652 if (ConvInput.getValueType().isVector() &&
10653 ConvInput.getValueType().getVectorNumElements() == NumElts)
10654 V = ConvInput.getNode();
10657 if (V->getOpcode() == ISD::BUILD_VECTOR) {
10658 assert(V->getNumOperands() == NumElts &&
10659 "BUILD_VECTOR has wrong number of operands");
10661 bool AllSame = true;
10662 for (unsigned i = 0; i != NumElts; ++i) {
10663 if (V->getOperand(i).getOpcode() != ISD::UNDEF) {
10664 Base = V->getOperand(i);
10668 // Splat of <u, u, u, u>, return <u, u, u, u>
10669 if (!Base.getNode())
10671 for (unsigned i = 0; i != NumElts; ++i) {
10672 if (V->getOperand(i) != Base) {
10677 // Splat of <x, x, x, x>, return <x, x, x, x>
10683 if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
10684 Level < AfterLegalizeVectorOps &&
10685 (N1.getOpcode() == ISD::UNDEF ||
10686 (N1.getOpcode() == ISD::CONCAT_VECTORS &&
10687 N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()))) {
10688 SDValue V = partitionShuffleOfConcats(N, DAG);
10694 // If this shuffle node is simply a swizzle of another shuffle node,
10695 // then try to simplify it.
10696 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
10697 N1.getOpcode() == ISD::UNDEF) {
10699 ShuffleVectorSDNode *OtherSV = cast<ShuffleVectorSDNode>(N0);
10701 // The incoming shuffle must be of the same type as the result of the
10702 // current shuffle.
10703 assert(OtherSV->getOperand(0).getValueType() == VT &&
10704 "Shuffle types don't match");
10706 SmallVector<int, 4> Mask;
10707 // Compute the combined shuffle mask.
10708 for (unsigned i = 0; i != NumElts; ++i) {
10709 int Idx = SVN->getMaskElt(i);
10710 assert(Idx < (int)NumElts && "Index references undef operand");
10711 // Next, this index comes from the first value, which is the incoming
10712 // shuffle. Adopt the incoming index.
10714 Idx = OtherSV->getMaskElt(Idx);
10715 Mask.push_back(Idx);
10718 bool CommuteOperands = false;
10719 if (N0.getOperand(1).getOpcode() != ISD::UNDEF) {
10720 // To be valid, the combine shuffle mask should only reference elements
10721 // from one of the two vectors in input to the inner shufflevector.
10722 bool IsValidMask = true;
10723 for (unsigned i = 0; i != NumElts && IsValidMask; ++i)
10724 // See if the combined mask only reference undefs or elements coming
10725 // from the first shufflevector operand.
10726 IsValidMask = Mask[i] < 0 || (unsigned)Mask[i] < NumElts;
10728 if (!IsValidMask) {
10729 IsValidMask = true;
10730 for (unsigned i = 0; i != NumElts && IsValidMask; ++i)
10731 // Check that all the elements come from the second shuffle operand.
10732 IsValidMask = Mask[i] < 0 || (unsigned)Mask[i] >= NumElts;
10733 CommuteOperands = IsValidMask;
10736 // Early exit if the combined shuffle mask is not valid.
10741 // See if this pair of shuffles can be safely folded according to either
10742 // of the following rules:
10743 // shuffle(shuffle(x, y), undef) -> x
10744 // shuffle(shuffle(x, undef), undef) -> x
10745 // shuffle(shuffle(x, y), undef) -> y
10746 bool IsIdentityMask = true;
10747 unsigned BaseMaskIndex = CommuteOperands ? NumElts : 0;
10748 for (unsigned i = 0; i != NumElts && IsIdentityMask; ++i) {
10753 // The combined shuffle must map each index to itself.
10754 IsIdentityMask = (unsigned)Mask[i] == i + BaseMaskIndex;
10757 if (IsIdentityMask) {
10758 if (CommuteOperands)
10759 // optimize shuffle(shuffle(x, y), undef) -> y.
10760 return OtherSV->getOperand(1);
10762 // optimize shuffle(shuffle(x, undef), undef) -> x
10763 // optimize shuffle(shuffle(x, y), undef) -> x
10764 return OtherSV->getOperand(0);
10767 // It may still be beneficial to combine the two shuffles if the
10768 // resulting shuffle is legal.
10769 if (TLI.isTypeLegal(VT) && TLI.isShuffleMaskLegal(Mask, VT)) {
10770 if (!CommuteOperands)
10771 // shuffle(shuffle(x, undef, M1), undef, M2) -> shuffle(x, undef, M3).
10772 // shuffle(shuffle(x, y, M1), undef, M2) -> shuffle(x, undef, M3)
10773 return DAG.getVectorShuffle(VT, SDLoc(N), N0->getOperand(0), N1,
10776 // shuffle(shuffle(x, y, M1), undef, M2) -> shuffle(undef, y, M3)
10777 return DAG.getVectorShuffle(VT, SDLoc(N), N1, N0->getOperand(1),
10782 // Try to fold according to rules:
10783 // shuffle(shuffle(A, B, M0), B, M1) -> shuffle(A, B, M2)
10784 // shuffle(shuffle(A, B, M0), A, M1) -> shuffle(A, B, M2)
10785 // shuffle(shuffle(A, Undef, M0), B, M1) -> shuffle(A, B, M2)
10786 // shuffle(shuffle(A, Undef, M0), A, M1) -> shuffle(A, Undef, M2)
10787 // Don't try to fold shuffles with illegal type.
10788 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
10789 N1.getOpcode() != ISD::UNDEF && TLI.isTypeLegal(VT)) {
10790 ShuffleVectorSDNode *OtherSV = cast<ShuffleVectorSDNode>(N0);
10792 // The incoming shuffle must be of the same type as the result of the
10793 // current shuffle.
10794 assert(OtherSV->getOperand(0).getValueType() == VT &&
10795 "Shuffle types don't match");
10797 SDValue SV0 = OtherSV->getOperand(0);
10798 SDValue SV1 = OtherSV->getOperand(1);
10799 bool HasSameOp0 = N1 == SV0;
10800 bool IsSV1Undef = SV1.getOpcode() == ISD::UNDEF;
10801 if (!HasSameOp0 && !IsSV1Undef && N1 != SV1)
10805 SmallVector<int, 4> Mask;
10806 // Compute the combined shuffle mask for a shuffle with SV0 as the first
10807 // operand, and SV1 as the second operand.
10808 for (unsigned i = 0; i != NumElts; ++i) {
10809 int Idx = SVN->getMaskElt(i);
10811 // Propagate Undef.
10812 Mask.push_back(Idx);
10816 if (Idx < (int)NumElts) {
10817 Idx = OtherSV->getMaskElt(Idx);
10818 if (IsSV1Undef && Idx >= (int) NumElts)
10819 Idx = -1; // Propagate Undef.
10821 Idx = HasSameOp0 ? Idx - NumElts : Idx;
10823 Mask.push_back(Idx);
10826 // Avoid introducing shuffles with illegal mask.
10827 if (TLI.isShuffleMaskLegal(Mask, VT)) {
10829 // shuffle(shuffle(A, Undef, M0), B, M1) -> shuffle(A, B, M2)
10830 // shuffle(shuffle(A, Undef, M0), A, M1) -> shuffle(A, Undef, M2)
10831 return DAG.getVectorShuffle(VT, SDLoc(N), SV0, N1, &Mask[0]);
10832 return DAG.getVectorShuffle(VT, SDLoc(N), SV0, SV1, &Mask[0]);
10839 SDValue DAGCombiner::visitINSERT_SUBVECTOR(SDNode *N) {
10840 SDValue N0 = N->getOperand(0);
10841 SDValue N2 = N->getOperand(2);
10843 // If the input vector is a concatenation, and the insert replaces
10844 // one of the halves, we can optimize into a single concat_vectors.
10845 if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
10846 N0->getNumOperands() == 2 && N2.getOpcode() == ISD::Constant) {
10847 APInt InsIdx = cast<ConstantSDNode>(N2)->getAPIntValue();
10848 EVT VT = N->getValueType(0);
10850 // Lower half: fold (insert_subvector (concat_vectors X, Y), Z) ->
10851 // (concat_vectors Z, Y)
10853 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
10854 N->getOperand(1), N0.getOperand(1));
10856 // Upper half: fold (insert_subvector (concat_vectors X, Y), Z) ->
10857 // (concat_vectors X, Z)
10858 if (InsIdx == VT.getVectorNumElements()/2)
10859 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
10860 N0.getOperand(0), N->getOperand(1));
10866 /// XformToShuffleWithZero - Returns a vector_shuffle if it able to transform
10867 /// an AND to a vector_shuffle with the destination vector and a zero vector.
10868 /// e.g. AND V, <0xffffffff, 0, 0xffffffff, 0>. ==>
10869 /// vector_shuffle V, Zero, <0, 4, 2, 4>
10870 SDValue DAGCombiner::XformToShuffleWithZero(SDNode *N) {
10871 EVT VT = N->getValueType(0);
10873 SDValue LHS = N->getOperand(0);
10874 SDValue RHS = N->getOperand(1);
10875 if (N->getOpcode() == ISD::AND) {
10876 if (RHS.getOpcode() == ISD::BITCAST)
10877 RHS = RHS.getOperand(0);
10878 if (RHS.getOpcode() == ISD::BUILD_VECTOR) {
10879 SmallVector<int, 8> Indices;
10880 unsigned NumElts = RHS.getNumOperands();
10881 for (unsigned i = 0; i != NumElts; ++i) {
10882 SDValue Elt = RHS.getOperand(i);
10883 if (!isa<ConstantSDNode>(Elt))
10886 if (cast<ConstantSDNode>(Elt)->isAllOnesValue())
10887 Indices.push_back(i);
10888 else if (cast<ConstantSDNode>(Elt)->isNullValue())
10889 Indices.push_back(NumElts);
10894 // Let's see if the target supports this vector_shuffle.
10895 EVT RVT = RHS.getValueType();
10896 if (!TLI.isVectorClearMaskLegal(Indices, RVT))
10899 // Return the new VECTOR_SHUFFLE node.
10900 EVT EltVT = RVT.getVectorElementType();
10901 SmallVector<SDValue,8> ZeroOps(RVT.getVectorNumElements(),
10902 DAG.getConstant(0, EltVT));
10903 SDValue Zero = DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), RVT, ZeroOps);
10904 LHS = DAG.getNode(ISD::BITCAST, dl, RVT, LHS);
10905 SDValue Shuf = DAG.getVectorShuffle(RVT, dl, LHS, Zero, &Indices[0]);
10906 return DAG.getNode(ISD::BITCAST, dl, VT, Shuf);
10913 /// SimplifyVBinOp - Visit a binary vector operation, like ADD.
10914 SDValue DAGCombiner::SimplifyVBinOp(SDNode *N) {
10915 assert(N->getValueType(0).isVector() &&
10916 "SimplifyVBinOp only works on vectors!");
10918 SDValue LHS = N->getOperand(0);
10919 SDValue RHS = N->getOperand(1);
10920 SDValue Shuffle = XformToShuffleWithZero(N);
10921 if (Shuffle.getNode()) return Shuffle;
10923 // If the LHS and RHS are BUILD_VECTOR nodes, see if we can constant fold
10925 if (LHS.getOpcode() == ISD::BUILD_VECTOR &&
10926 RHS.getOpcode() == ISD::BUILD_VECTOR) {
10927 // Check if both vectors are constants. If not bail out.
10928 if (!(cast<BuildVectorSDNode>(LHS)->isConstant() &&
10929 cast<BuildVectorSDNode>(RHS)->isConstant()))
10932 SmallVector<SDValue, 8> Ops;
10933 for (unsigned i = 0, e = LHS.getNumOperands(); i != e; ++i) {
10934 SDValue LHSOp = LHS.getOperand(i);
10935 SDValue RHSOp = RHS.getOperand(i);
10937 // Can't fold divide by zero.
10938 if (N->getOpcode() == ISD::SDIV || N->getOpcode() == ISD::UDIV ||
10939 N->getOpcode() == ISD::FDIV) {
10940 if ((RHSOp.getOpcode() == ISD::Constant &&
10941 cast<ConstantSDNode>(RHSOp.getNode())->isNullValue()) ||
10942 (RHSOp.getOpcode() == ISD::ConstantFP &&
10943 cast<ConstantFPSDNode>(RHSOp.getNode())->getValueAPF().isZero()))
10947 EVT VT = LHSOp.getValueType();
10948 EVT RVT = RHSOp.getValueType();
10950 // Integer BUILD_VECTOR operands may have types larger than the element
10951 // size (e.g., when the element type is not legal). Prior to type
10952 // legalization, the types may not match between the two BUILD_VECTORS.
10953 // Truncate one of the operands to make them match.
10954 if (RVT.getSizeInBits() > VT.getSizeInBits()) {
10955 RHSOp = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, RHSOp);
10957 LHSOp = DAG.getNode(ISD::TRUNCATE, SDLoc(N), RVT, LHSOp);
10961 SDValue FoldOp = DAG.getNode(N->getOpcode(), SDLoc(LHS), VT,
10963 if (FoldOp.getOpcode() != ISD::UNDEF &&
10964 FoldOp.getOpcode() != ISD::Constant &&
10965 FoldOp.getOpcode() != ISD::ConstantFP)
10967 Ops.push_back(FoldOp);
10968 AddToWorkList(FoldOp.getNode());
10971 if (Ops.size() == LHS.getNumOperands())
10972 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), LHS.getValueType(), Ops);
10975 // Type legalization might introduce new shuffles in the DAG.
10976 // Fold (VBinOp (shuffle (A, Undef, Mask)), (shuffle (B, Undef, Mask)))
10977 // -> (shuffle (VBinOp (A, B)), Undef, Mask).
10978 if (LegalTypes && isa<ShuffleVectorSDNode>(LHS) &&
10979 isa<ShuffleVectorSDNode>(RHS) && LHS.hasOneUse() && RHS.hasOneUse() &&
10980 LHS.getOperand(1).getOpcode() == ISD::UNDEF &&
10981 RHS.getOperand(1).getOpcode() == ISD::UNDEF) {
10982 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(LHS);
10983 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(RHS);
10985 if (SVN0->getMask().equals(SVN1->getMask())) {
10986 EVT VT = N->getValueType(0);
10987 SDValue UndefVector = LHS.getOperand(1);
10988 SDValue NewBinOp = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
10989 LHS.getOperand(0), RHS.getOperand(0));
10990 AddUsersToWorkList(N);
10991 return DAG.getVectorShuffle(VT, SDLoc(N), NewBinOp, UndefVector,
10992 &SVN0->getMask()[0]);
10999 /// SimplifyVUnaryOp - Visit a binary vector operation, like FABS/FNEG.
11000 SDValue DAGCombiner::SimplifyVUnaryOp(SDNode *N) {
11001 assert(N->getValueType(0).isVector() &&
11002 "SimplifyVUnaryOp only works on vectors!");
11004 SDValue N0 = N->getOperand(0);
11006 if (N0.getOpcode() != ISD::BUILD_VECTOR)
11009 // Operand is a BUILD_VECTOR node, see if we can constant fold it.
11010 SmallVector<SDValue, 8> Ops;
11011 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) {
11012 SDValue Op = N0.getOperand(i);
11013 if (Op.getOpcode() != ISD::UNDEF &&
11014 Op.getOpcode() != ISD::ConstantFP)
11016 EVT EltVT = Op.getValueType();
11017 SDValue FoldOp = DAG.getNode(N->getOpcode(), SDLoc(N0), EltVT, Op);
11018 if (FoldOp.getOpcode() != ISD::UNDEF &&
11019 FoldOp.getOpcode() != ISD::ConstantFP)
11021 Ops.push_back(FoldOp);
11022 AddToWorkList(FoldOp.getNode());
11025 if (Ops.size() != N0.getNumOperands())
11028 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), N0.getValueType(), Ops);
11031 SDValue DAGCombiner::SimplifySelect(SDLoc DL, SDValue N0,
11032 SDValue N1, SDValue N2){
11033 assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!");
11035 SDValue SCC = SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1), N1, N2,
11036 cast<CondCodeSDNode>(N0.getOperand(2))->get());
11038 // If we got a simplified select_cc node back from SimplifySelectCC, then
11039 // break it down into a new SETCC node, and a new SELECT node, and then return
11040 // the SELECT node, since we were called with a SELECT node.
11041 if (SCC.getNode()) {
11042 // Check to see if we got a select_cc back (to turn into setcc/select).
11043 // Otherwise, just return whatever node we got back, like fabs.
11044 if (SCC.getOpcode() == ISD::SELECT_CC) {
11045 SDValue SETCC = DAG.getNode(ISD::SETCC, SDLoc(N0),
11047 SCC.getOperand(0), SCC.getOperand(1),
11048 SCC.getOperand(4));
11049 AddToWorkList(SETCC.getNode());
11050 return DAG.getSelect(SDLoc(SCC), SCC.getValueType(),
11051 SCC.getOperand(2), SCC.getOperand(3), SETCC);
11059 /// SimplifySelectOps - Given a SELECT or a SELECT_CC node, where LHS and RHS
11060 /// are the two values being selected between, see if we can simplify the
11061 /// select. Callers of this should assume that TheSelect is deleted if this
11062 /// returns true. As such, they should return the appropriate thing (e.g. the
11063 /// node) back to the top-level of the DAG combiner loop to avoid it being
11065 bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDValue LHS,
11068 // Cannot simplify select with vector condition
11069 if (TheSelect->getOperand(0).getValueType().isVector()) return false;
11071 // If this is a select from two identical things, try to pull the operation
11072 // through the select.
11073 if (LHS.getOpcode() != RHS.getOpcode() ||
11074 !LHS.hasOneUse() || !RHS.hasOneUse())
11077 // If this is a load and the token chain is identical, replace the select
11078 // of two loads with a load through a select of the address to load from.
11079 // This triggers in things like "select bool X, 10.0, 123.0" after the FP
11080 // constants have been dropped into the constant pool.
11081 if (LHS.getOpcode() == ISD::LOAD) {
11082 LoadSDNode *LLD = cast<LoadSDNode>(LHS);
11083 LoadSDNode *RLD = cast<LoadSDNode>(RHS);
11085 // Token chains must be identical.
11086 if (LHS.getOperand(0) != RHS.getOperand(0) ||
11087 // Do not let this transformation reduce the number of volatile loads.
11088 LLD->isVolatile() || RLD->isVolatile() ||
11089 // If this is an EXTLOAD, the VT's must match.
11090 LLD->getMemoryVT() != RLD->getMemoryVT() ||
11091 // If this is an EXTLOAD, the kind of extension must match.
11092 (LLD->getExtensionType() != RLD->getExtensionType() &&
11093 // The only exception is if one of the extensions is anyext.
11094 LLD->getExtensionType() != ISD::EXTLOAD &&
11095 RLD->getExtensionType() != ISD::EXTLOAD) ||
11096 // FIXME: this discards src value information. This is
11097 // over-conservative. It would be beneficial to be able to remember
11098 // both potential memory locations. Since we are discarding
11099 // src value info, don't do the transformation if the memory
11100 // locations are not in the default address space.
11101 LLD->getPointerInfo().getAddrSpace() != 0 ||
11102 RLD->getPointerInfo().getAddrSpace() != 0 ||
11103 !TLI.isOperationLegalOrCustom(TheSelect->getOpcode(),
11104 LLD->getBasePtr().getValueType()))
11107 // Check that the select condition doesn't reach either load. If so,
11108 // folding this will induce a cycle into the DAG. If not, this is safe to
11109 // xform, so create a select of the addresses.
11111 if (TheSelect->getOpcode() == ISD::SELECT) {
11112 SDNode *CondNode = TheSelect->getOperand(0).getNode();
11113 if ((LLD->hasAnyUseOfValue(1) && LLD->isPredecessorOf(CondNode)) ||
11114 (RLD->hasAnyUseOfValue(1) && RLD->isPredecessorOf(CondNode)))
11116 // The loads must not depend on one another.
11117 if (LLD->isPredecessorOf(RLD) ||
11118 RLD->isPredecessorOf(LLD))
11120 Addr = DAG.getSelect(SDLoc(TheSelect),
11121 LLD->getBasePtr().getValueType(),
11122 TheSelect->getOperand(0), LLD->getBasePtr(),
11123 RLD->getBasePtr());
11124 } else { // Otherwise SELECT_CC
11125 SDNode *CondLHS = TheSelect->getOperand(0).getNode();
11126 SDNode *CondRHS = TheSelect->getOperand(1).getNode();
11128 if ((LLD->hasAnyUseOfValue(1) &&
11129 (LLD->isPredecessorOf(CondLHS) || LLD->isPredecessorOf(CondRHS))) ||
11130 (RLD->hasAnyUseOfValue(1) &&
11131 (RLD->isPredecessorOf(CondLHS) || RLD->isPredecessorOf(CondRHS))))
11134 Addr = DAG.getNode(ISD::SELECT_CC, SDLoc(TheSelect),
11135 LLD->getBasePtr().getValueType(),
11136 TheSelect->getOperand(0),
11137 TheSelect->getOperand(1),
11138 LLD->getBasePtr(), RLD->getBasePtr(),
11139 TheSelect->getOperand(4));
11143 if (LLD->getExtensionType() == ISD::NON_EXTLOAD) {
11144 Load = DAG.getLoad(TheSelect->getValueType(0),
11146 // FIXME: Discards pointer and TBAA info.
11147 LLD->getChain(), Addr, MachinePointerInfo(),
11148 LLD->isVolatile(), LLD->isNonTemporal(),
11149 LLD->isInvariant(), LLD->getAlignment());
11151 Load = DAG.getExtLoad(LLD->getExtensionType() == ISD::EXTLOAD ?
11152 RLD->getExtensionType() : LLD->getExtensionType(),
11154 TheSelect->getValueType(0),
11155 // FIXME: Discards pointer and TBAA info.
11156 LLD->getChain(), Addr, MachinePointerInfo(),
11157 LLD->getMemoryVT(), LLD->isVolatile(),
11158 LLD->isNonTemporal(), LLD->getAlignment());
11161 // Users of the select now use the result of the load.
11162 CombineTo(TheSelect, Load);
11164 // Users of the old loads now use the new load's chain. We know the
11165 // old-load value is dead now.
11166 CombineTo(LHS.getNode(), Load.getValue(0), Load.getValue(1));
11167 CombineTo(RHS.getNode(), Load.getValue(0), Load.getValue(1));
11174 /// SimplifySelectCC - Simplify an expression of the form (N0 cond N1) ? N2 : N3
11175 /// where 'cond' is the comparison specified by CC.
11176 SDValue DAGCombiner::SimplifySelectCC(SDLoc DL, SDValue N0, SDValue N1,
11177 SDValue N2, SDValue N3,
11178 ISD::CondCode CC, bool NotExtCompare) {
11179 // (x ? y : y) -> y.
11180 if (N2 == N3) return N2;
11182 EVT VT = N2.getValueType();
11183 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
11184 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
11185 ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N3.getNode());
11187 // Determine if the condition we're dealing with is constant
11188 SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()),
11189 N0, N1, CC, DL, false);
11190 if (SCC.getNode()) AddToWorkList(SCC.getNode());
11191 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode());
11193 // fold select_cc true, x, y -> x
11194 if (SCCC && !SCCC->isNullValue())
11196 // fold select_cc false, x, y -> y
11197 if (SCCC && SCCC->isNullValue())
11200 // Check to see if we can simplify the select into an fabs node
11201 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) {
11202 // Allow either -0.0 or 0.0
11203 if (CFP->getValueAPF().isZero()) {
11204 // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs
11205 if ((CC == ISD::SETGE || CC == ISD::SETGT) &&
11206 N0 == N2 && N3.getOpcode() == ISD::FNEG &&
11207 N2 == N3.getOperand(0))
11208 return DAG.getNode(ISD::FABS, DL, VT, N0);
11210 // select (setl[te] X, +/-0.0), fneg(X), X -> fabs
11211 if ((CC == ISD::SETLT || CC == ISD::SETLE) &&
11212 N0 == N3 && N2.getOpcode() == ISD::FNEG &&
11213 N2.getOperand(0) == N3)
11214 return DAG.getNode(ISD::FABS, DL, VT, N3);
11218 // Turn "(a cond b) ? 1.0f : 2.0f" into "load (tmp + ((a cond b) ? 0 : 4)"
11219 // where "tmp" is a constant pool entry containing an array with 1.0 and 2.0
11220 // in it. This is a win when the constant is not otherwise available because
11221 // it replaces two constant pool loads with one. We only do this if the FP
11222 // type is known to be legal, because if it isn't, then we are before legalize
11223 // types an we want the other legalization to happen first (e.g. to avoid
11224 // messing with soft float) and if the ConstantFP is not legal, because if
11225 // it is legal, we may not need to store the FP constant in a constant pool.
11226 if (ConstantFPSDNode *TV = dyn_cast<ConstantFPSDNode>(N2))
11227 if (ConstantFPSDNode *FV = dyn_cast<ConstantFPSDNode>(N3)) {
11228 if (TLI.isTypeLegal(N2.getValueType()) &&
11229 (TLI.getOperationAction(ISD::ConstantFP, N2.getValueType()) !=
11230 TargetLowering::Legal &&
11231 !TLI.isFPImmLegal(TV->getValueAPF(), TV->getValueType(0)) &&
11232 !TLI.isFPImmLegal(FV->getValueAPF(), FV->getValueType(0))) &&
11233 // If both constants have multiple uses, then we won't need to do an
11234 // extra load, they are likely around in registers for other users.
11235 (TV->hasOneUse() || FV->hasOneUse())) {
11236 Constant *Elts[] = {
11237 const_cast<ConstantFP*>(FV->getConstantFPValue()),
11238 const_cast<ConstantFP*>(TV->getConstantFPValue())
11240 Type *FPTy = Elts[0]->getType();
11241 const DataLayout &TD = *TLI.getDataLayout();
11243 // Create a ConstantArray of the two constants.
11244 Constant *CA = ConstantArray::get(ArrayType::get(FPTy, 2), Elts);
11245 SDValue CPIdx = DAG.getConstantPool(CA, TLI.getPointerTy(),
11246 TD.getPrefTypeAlignment(FPTy));
11247 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
11249 // Get the offsets to the 0 and 1 element of the array so that we can
11250 // select between them.
11251 SDValue Zero = DAG.getIntPtrConstant(0);
11252 unsigned EltSize = (unsigned)TD.getTypeAllocSize(Elts[0]->getType());
11253 SDValue One = DAG.getIntPtrConstant(EltSize);
11255 SDValue Cond = DAG.getSetCC(DL,
11256 getSetCCResultType(N0.getValueType()),
11258 AddToWorkList(Cond.getNode());
11259 SDValue CstOffset = DAG.getSelect(DL, Zero.getValueType(),
11261 AddToWorkList(CstOffset.getNode());
11262 CPIdx = DAG.getNode(ISD::ADD, DL, CPIdx.getValueType(), CPIdx,
11264 AddToWorkList(CPIdx.getNode());
11265 return DAG.getLoad(TV->getValueType(0), DL, DAG.getEntryNode(), CPIdx,
11266 MachinePointerInfo::getConstantPool(), false,
11267 false, false, Alignment);
11272 // Check to see if we can perform the "gzip trick", transforming
11273 // (select_cc setlt X, 0, A, 0) -> (and (sra X, (sub size(X), 1), A)
11274 if (N1C && N3C && N3C->isNullValue() && CC == ISD::SETLT &&
11275 (N1C->isNullValue() || // (a < 0) ? b : 0
11276 (N1C->getAPIntValue() == 1 && N0 == N2))) { // (a < 1) ? a : 0
11277 EVT XType = N0.getValueType();
11278 EVT AType = N2.getValueType();
11279 if (XType.bitsGE(AType)) {
11280 // and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a
11281 // single-bit constant.
11282 if (N2C && ((N2C->getAPIntValue() & (N2C->getAPIntValue()-1)) == 0)) {
11283 unsigned ShCtV = N2C->getAPIntValue().logBase2();
11284 ShCtV = XType.getSizeInBits()-ShCtV-1;
11285 SDValue ShCt = DAG.getConstant(ShCtV,
11286 getShiftAmountTy(N0.getValueType()));
11287 SDValue Shift = DAG.getNode(ISD::SRL, SDLoc(N0),
11289 AddToWorkList(Shift.getNode());
11291 if (XType.bitsGT(AType)) {
11292 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
11293 AddToWorkList(Shift.getNode());
11296 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
11299 SDValue Shift = DAG.getNode(ISD::SRA, SDLoc(N0),
11301 DAG.getConstant(XType.getSizeInBits()-1,
11302 getShiftAmountTy(N0.getValueType())));
11303 AddToWorkList(Shift.getNode());
11305 if (XType.bitsGT(AType)) {
11306 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
11307 AddToWorkList(Shift.getNode());
11310 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
11314 // fold (select_cc seteq (and x, y), 0, 0, A) -> (and (shr (shl x)) A)
11315 // where y is has a single bit set.
11316 // A plaintext description would be, we can turn the SELECT_CC into an AND
11317 // when the condition can be materialized as an all-ones register. Any
11318 // single bit-test can be materialized as an all-ones register with
11319 // shift-left and shift-right-arith.
11320 if (CC == ISD::SETEQ && N0->getOpcode() == ISD::AND &&
11321 N0->getValueType(0) == VT &&
11322 N1C && N1C->isNullValue() &&
11323 N2C && N2C->isNullValue()) {
11324 SDValue AndLHS = N0->getOperand(0);
11325 ConstantSDNode *ConstAndRHS = dyn_cast<ConstantSDNode>(N0->getOperand(1));
11326 if (ConstAndRHS && ConstAndRHS->getAPIntValue().countPopulation() == 1) {
11327 // Shift the tested bit over the sign bit.
11328 APInt AndMask = ConstAndRHS->getAPIntValue();
11330 DAG.getConstant(AndMask.countLeadingZeros(),
11331 getShiftAmountTy(AndLHS.getValueType()));
11332 SDValue Shl = DAG.getNode(ISD::SHL, SDLoc(N0), VT, AndLHS, ShlAmt);
11334 // Now arithmetic right shift it all the way over, so the result is either
11335 // all-ones, or zero.
11337 DAG.getConstant(AndMask.getBitWidth()-1,
11338 getShiftAmountTy(Shl.getValueType()));
11339 SDValue Shr = DAG.getNode(ISD::SRA, SDLoc(N0), VT, Shl, ShrAmt);
11341 return DAG.getNode(ISD::AND, DL, VT, Shr, N3);
11345 // fold select C, 16, 0 -> shl C, 4
11346 if (N2C && N3C && N3C->isNullValue() && N2C->getAPIntValue().isPowerOf2() &&
11347 TLI.getBooleanContents(N0.getValueType()) ==
11348 TargetLowering::ZeroOrOneBooleanContent) {
11350 // If the caller doesn't want us to simplify this into a zext of a compare,
11352 if (NotExtCompare && N2C->getAPIntValue() == 1)
11355 // Get a SetCC of the condition
11356 // NOTE: Don't create a SETCC if it's not legal on this target.
11357 if (!LegalOperations ||
11358 TLI.isOperationLegal(ISD::SETCC,
11359 LegalTypes ? getSetCCResultType(N0.getValueType()) : MVT::i1)) {
11361 // cast from setcc result type to select result type
11363 SCC = DAG.getSetCC(DL, getSetCCResultType(N0.getValueType()),
11365 if (N2.getValueType().bitsLT(SCC.getValueType()))
11366 Temp = DAG.getZeroExtendInReg(SCC, SDLoc(N2),
11367 N2.getValueType());
11369 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2),
11370 N2.getValueType(), SCC);
11372 SCC = DAG.getSetCC(SDLoc(N0), MVT::i1, N0, N1, CC);
11373 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2),
11374 N2.getValueType(), SCC);
11377 AddToWorkList(SCC.getNode());
11378 AddToWorkList(Temp.getNode());
11380 if (N2C->getAPIntValue() == 1)
11383 // shl setcc result by log2 n2c
11384 return DAG.getNode(
11385 ISD::SHL, DL, N2.getValueType(), Temp,
11386 DAG.getConstant(N2C->getAPIntValue().logBase2(),
11387 getShiftAmountTy(Temp.getValueType())));
11391 // Check to see if this is the equivalent of setcc
11392 // FIXME: Turn all of these into setcc if setcc if setcc is legal
11393 // otherwise, go ahead with the folds.
11394 if (0 && N3C && N3C->isNullValue() && N2C && (N2C->getAPIntValue() == 1ULL)) {
11395 EVT XType = N0.getValueType();
11396 if (!LegalOperations ||
11397 TLI.isOperationLegal(ISD::SETCC, getSetCCResultType(XType))) {
11398 SDValue Res = DAG.getSetCC(DL, getSetCCResultType(XType), N0, N1, CC);
11399 if (Res.getValueType() != VT)
11400 Res = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Res);
11404 // fold (seteq X, 0) -> (srl (ctlz X, log2(size(X))))
11405 if (N1C && N1C->isNullValue() && CC == ISD::SETEQ &&
11406 (!LegalOperations ||
11407 TLI.isOperationLegal(ISD::CTLZ, XType))) {
11408 SDValue Ctlz = DAG.getNode(ISD::CTLZ, SDLoc(N0), XType, N0);
11409 return DAG.getNode(ISD::SRL, DL, XType, Ctlz,
11410 DAG.getConstant(Log2_32(XType.getSizeInBits()),
11411 getShiftAmountTy(Ctlz.getValueType())));
11413 // fold (setgt X, 0) -> (srl (and (-X, ~X), size(X)-1))
11414 if (N1C && N1C->isNullValue() && CC == ISD::SETGT) {
11415 SDValue NegN0 = DAG.getNode(ISD::SUB, SDLoc(N0),
11416 XType, DAG.getConstant(0, XType), N0);
11417 SDValue NotN0 = DAG.getNOT(SDLoc(N0), N0, XType);
11418 return DAG.getNode(ISD::SRL, DL, XType,
11419 DAG.getNode(ISD::AND, DL, XType, NegN0, NotN0),
11420 DAG.getConstant(XType.getSizeInBits()-1,
11421 getShiftAmountTy(XType)));
11423 // fold (setgt X, -1) -> (xor (srl (X, size(X)-1), 1))
11424 if (N1C && N1C->isAllOnesValue() && CC == ISD::SETGT) {
11425 SDValue Sign = DAG.getNode(ISD::SRL, SDLoc(N0), XType, N0,
11426 DAG.getConstant(XType.getSizeInBits()-1,
11427 getShiftAmountTy(N0.getValueType())));
11428 return DAG.getNode(ISD::XOR, DL, XType, Sign, DAG.getConstant(1, XType));
11432 // Check to see if this is an integer abs.
11433 // select_cc setg[te] X, 0, X, -X ->
11434 // select_cc setgt X, -1, X, -X ->
11435 // select_cc setl[te] X, 0, -X, X ->
11436 // select_cc setlt X, 1, -X, X ->
11437 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
11439 ConstantSDNode *SubC = nullptr;
11440 if (((N1C->isNullValue() && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
11441 (N1C->isAllOnesValue() && CC == ISD::SETGT)) &&
11442 N0 == N2 && N3.getOpcode() == ISD::SUB && N0 == N3.getOperand(1))
11443 SubC = dyn_cast<ConstantSDNode>(N3.getOperand(0));
11444 else if (((N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE)) ||
11445 (N1C->isOne() && CC == ISD::SETLT)) &&
11446 N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1))
11447 SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0));
11449 EVT XType = N0.getValueType();
11450 if (SubC && SubC->isNullValue() && XType.isInteger()) {
11451 SDValue Shift = DAG.getNode(ISD::SRA, SDLoc(N0), XType,
11453 DAG.getConstant(XType.getSizeInBits()-1,
11454 getShiftAmountTy(N0.getValueType())));
11455 SDValue Add = DAG.getNode(ISD::ADD, SDLoc(N0),
11457 AddToWorkList(Shift.getNode());
11458 AddToWorkList(Add.getNode());
11459 return DAG.getNode(ISD::XOR, DL, XType, Add, Shift);
11466 /// SimplifySetCC - This is a stub for TargetLowering::SimplifySetCC.
11467 SDValue DAGCombiner::SimplifySetCC(EVT VT, SDValue N0,
11468 SDValue N1, ISD::CondCode Cond,
11469 SDLoc DL, bool foldBooleans) {
11470 TargetLowering::DAGCombinerInfo
11471 DagCombineInfo(DAG, Level, false, this);
11472 return TLI.SimplifySetCC(VT, N0, N1, Cond, foldBooleans, DagCombineInfo, DL);
11475 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
11476 /// return a DAG expression to select that will generate the same value by
11477 /// multiplying by a magic number. See:
11478 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
11479 SDValue DAGCombiner::BuildSDIV(SDNode *N) {
11480 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
11484 // Avoid division by zero.
11485 if (!C->getAPIntValue())
11488 std::vector<SDNode*> Built;
11490 TLI.BuildSDIV(N, C->getAPIntValue(), DAG, LegalOperations, &Built);
11492 for (SDNode *N : Built)
11497 /// BuildUDIV - Given an ISD::UDIV node expressing a divide by constant,
11498 /// return a DAG expression to select that will generate the same value by
11499 /// multiplying by a magic number. See:
11500 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
11501 SDValue DAGCombiner::BuildUDIV(SDNode *N) {
11502 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
11506 // Avoid division by zero.
11507 if (!C->getAPIntValue())
11510 std::vector<SDNode*> Built;
11512 TLI.BuildUDIV(N, C->getAPIntValue(), DAG, LegalOperations, &Built);
11514 for (SDNode *N : Built)
11519 /// FindBaseOffset - Return true if base is a frame index, which is known not
11520 // to alias with anything but itself. Provides base object and offset as
11522 static bool FindBaseOffset(SDValue Ptr, SDValue &Base, int64_t &Offset,
11523 const GlobalValue *&GV, const void *&CV) {
11524 // Assume it is a primitive operation.
11525 Base = Ptr; Offset = 0; GV = nullptr; CV = nullptr;
11527 // If it's an adding a simple constant then integrate the offset.
11528 if (Base.getOpcode() == ISD::ADD) {
11529 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Base.getOperand(1))) {
11530 Base = Base.getOperand(0);
11531 Offset += C->getZExtValue();
11535 // Return the underlying GlobalValue, and update the Offset. Return false
11536 // for GlobalAddressSDNode since the same GlobalAddress may be represented
11537 // by multiple nodes with different offsets.
11538 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Base)) {
11539 GV = G->getGlobal();
11540 Offset += G->getOffset();
11544 // Return the underlying Constant value, and update the Offset. Return false
11545 // for ConstantSDNodes since the same constant pool entry may be represented
11546 // by multiple nodes with different offsets.
11547 if (ConstantPoolSDNode *C = dyn_cast<ConstantPoolSDNode>(Base)) {
11548 CV = C->isMachineConstantPoolEntry() ? (const void *)C->getMachineCPVal()
11549 : (const void *)C->getConstVal();
11550 Offset += C->getOffset();
11553 // If it's any of the following then it can't alias with anything but itself.
11554 return isa<FrameIndexSDNode>(Base);
11557 /// isAlias - Return true if there is any possibility that the two addresses
11559 bool DAGCombiner::isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) const {
11560 // If they are the same then they must be aliases.
11561 if (Op0->getBasePtr() == Op1->getBasePtr()) return true;
11563 // If they are both volatile then they cannot be reordered.
11564 if (Op0->isVolatile() && Op1->isVolatile()) return true;
11566 // Gather base node and offset information.
11567 SDValue Base1, Base2;
11568 int64_t Offset1, Offset2;
11569 const GlobalValue *GV1, *GV2;
11570 const void *CV1, *CV2;
11571 bool isFrameIndex1 = FindBaseOffset(Op0->getBasePtr(),
11572 Base1, Offset1, GV1, CV1);
11573 bool isFrameIndex2 = FindBaseOffset(Op1->getBasePtr(),
11574 Base2, Offset2, GV2, CV2);
11576 // If they have a same base address then check to see if they overlap.
11577 if (Base1 == Base2 || (GV1 && (GV1 == GV2)) || (CV1 && (CV1 == CV2)))
11578 return !((Offset1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= Offset2 ||
11579 (Offset2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= Offset1);
11581 // It is possible for different frame indices to alias each other, mostly
11582 // when tail call optimization reuses return address slots for arguments.
11583 // To catch this case, look up the actual index of frame indices to compute
11584 // the real alias relationship.
11585 if (isFrameIndex1 && isFrameIndex2) {
11586 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
11587 Offset1 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base1)->getIndex());
11588 Offset2 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base2)->getIndex());
11589 return !((Offset1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= Offset2 ||
11590 (Offset2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= Offset1);
11593 // Otherwise, if we know what the bases are, and they aren't identical, then
11594 // we know they cannot alias.
11595 if ((isFrameIndex1 || CV1 || GV1) && (isFrameIndex2 || CV2 || GV2))
11598 // If we know required SrcValue1 and SrcValue2 have relatively large alignment
11599 // compared to the size and offset of the access, we may be able to prove they
11600 // do not alias. This check is conservative for now to catch cases created by
11601 // splitting vector types.
11602 if ((Op0->getOriginalAlignment() == Op1->getOriginalAlignment()) &&
11603 (Op0->getSrcValueOffset() != Op1->getSrcValueOffset()) &&
11604 (Op0->getMemoryVT().getSizeInBits() >> 3 ==
11605 Op1->getMemoryVT().getSizeInBits() >> 3) &&
11606 (Op0->getOriginalAlignment() > Op0->getMemoryVT().getSizeInBits()) >> 3) {
11607 int64_t OffAlign1 = Op0->getSrcValueOffset() % Op0->getOriginalAlignment();
11608 int64_t OffAlign2 = Op1->getSrcValueOffset() % Op1->getOriginalAlignment();
11610 // There is no overlap between these relatively aligned accesses of similar
11611 // size, return no alias.
11612 if ((OffAlign1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= OffAlign2 ||
11613 (OffAlign2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= OffAlign1)
11617 bool UseAA = CombinerGlobalAA.getNumOccurrences() > 0 ? CombinerGlobalAA :
11618 TLI.getTargetMachine().getSubtarget<TargetSubtargetInfo>().useAA();
11620 if (CombinerAAOnlyFunc.getNumOccurrences() &&
11621 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
11625 Op0->getMemOperand()->getValue() && Op1->getMemOperand()->getValue()) {
11626 // Use alias analysis information.
11627 int64_t MinOffset = std::min(Op0->getSrcValueOffset(),
11628 Op1->getSrcValueOffset());
11629 int64_t Overlap1 = (Op0->getMemoryVT().getSizeInBits() >> 3) +
11630 Op0->getSrcValueOffset() - MinOffset;
11631 int64_t Overlap2 = (Op1->getMemoryVT().getSizeInBits() >> 3) +
11632 Op1->getSrcValueOffset() - MinOffset;
11633 AliasAnalysis::AliasResult AAResult =
11634 AA.alias(AliasAnalysis::Location(Op0->getMemOperand()->getValue(),
11636 UseTBAA ? Op0->getTBAAInfo() : nullptr),
11637 AliasAnalysis::Location(Op1->getMemOperand()->getValue(),
11639 UseTBAA ? Op1->getTBAAInfo() : nullptr));
11640 if (AAResult == AliasAnalysis::NoAlias)
11644 // Otherwise we have to assume they alias.
11648 /// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes,
11649 /// looking for aliasing nodes and adding them to the Aliases vector.
11650 void DAGCombiner::GatherAllAliases(SDNode *N, SDValue OriginalChain,
11651 SmallVectorImpl<SDValue> &Aliases) {
11652 SmallVector<SDValue, 8> Chains; // List of chains to visit.
11653 SmallPtrSet<SDNode *, 16> Visited; // Visited node set.
11655 // Get alias information for node.
11656 bool IsLoad = isa<LoadSDNode>(N) && !cast<LSBaseSDNode>(N)->isVolatile();
11659 Chains.push_back(OriginalChain);
11660 unsigned Depth = 0;
11662 // Look at each chain and determine if it is an alias. If so, add it to the
11663 // aliases list. If not, then continue up the chain looking for the next
11665 while (!Chains.empty()) {
11666 SDValue Chain = Chains.back();
11669 // For TokenFactor nodes, look at each operand and only continue up the
11670 // chain until we find two aliases. If we've seen two aliases, assume we'll
11671 // find more and revert to original chain since the xform is unlikely to be
11674 // FIXME: The depth check could be made to return the last non-aliasing
11675 // chain we found before we hit a tokenfactor rather than the original
11677 if (Depth > 6 || Aliases.size() == 2) {
11679 Aliases.push_back(OriginalChain);
11683 // Don't bother if we've been before.
11684 if (!Visited.insert(Chain.getNode()))
11687 switch (Chain.getOpcode()) {
11688 case ISD::EntryToken:
11689 // Entry token is ideal chain operand, but handled in FindBetterChain.
11694 // Get alias information for Chain.
11695 bool IsOpLoad = isa<LoadSDNode>(Chain.getNode()) &&
11696 !cast<LSBaseSDNode>(Chain.getNode())->isVolatile();
11698 // If chain is alias then stop here.
11699 if (!(IsLoad && IsOpLoad) &&
11700 isAlias(cast<LSBaseSDNode>(N), cast<LSBaseSDNode>(Chain.getNode()))) {
11701 Aliases.push_back(Chain);
11703 // Look further up the chain.
11704 Chains.push_back(Chain.getOperand(0));
11710 case ISD::TokenFactor:
11711 // We have to check each of the operands of the token factor for "small"
11712 // token factors, so we queue them up. Adding the operands to the queue
11713 // (stack) in reverse order maintains the original order and increases the
11714 // likelihood that getNode will find a matching token factor (CSE.)
11715 if (Chain.getNumOperands() > 16) {
11716 Aliases.push_back(Chain);
11719 for (unsigned n = Chain.getNumOperands(); n;)
11720 Chains.push_back(Chain.getOperand(--n));
11725 // For all other instructions we will just have to take what we can get.
11726 Aliases.push_back(Chain);
11731 // We need to be careful here to also search for aliases through the
11732 // value operand of a store, etc. Consider the following situation:
11734 // L1 = load Token1, %52
11735 // S1 = store Token1, L1, %51
11736 // L2 = load Token1, %52+8
11737 // S2 = store Token1, L2, %51+8
11738 // Token2 = Token(S1, S2)
11739 // L3 = load Token2, %53
11740 // S3 = store Token2, L3, %52
11741 // L4 = load Token2, %53+8
11742 // S4 = store Token2, L4, %52+8
11743 // If we search for aliases of S3 (which loads address %52), and we look
11744 // only through the chain, then we'll miss the trivial dependence on L1
11745 // (which also loads from %52). We then might change all loads and
11746 // stores to use Token1 as their chain operand, which could result in
11747 // copying %53 into %52 before copying %52 into %51 (which should
11750 // The problem is, however, that searching for such data dependencies
11751 // can become expensive, and the cost is not directly related to the
11752 // chain depth. Instead, we'll rule out such configurations here by
11753 // insisting that we've visited all chain users (except for users
11754 // of the original chain, which is not necessary). When doing this,
11755 // we need to look through nodes we don't care about (otherwise, things
11756 // like register copies will interfere with trivial cases).
11758 SmallVector<const SDNode *, 16> Worklist;
11759 for (SmallPtrSet<SDNode *, 16>::iterator I = Visited.begin(),
11760 IE = Visited.end(); I != IE; ++I)
11761 if (*I != OriginalChain.getNode())
11762 Worklist.push_back(*I);
11764 while (!Worklist.empty()) {
11765 const SDNode *M = Worklist.pop_back_val();
11767 // We have already visited M, and want to make sure we've visited any uses
11768 // of M that we care about. For uses that we've not visisted, and don't
11769 // care about, queue them to the worklist.
11771 for (SDNode::use_iterator UI = M->use_begin(),
11772 UIE = M->use_end(); UI != UIE; ++UI)
11773 if (UI.getUse().getValueType() == MVT::Other && Visited.insert(*UI)) {
11774 if (isa<MemIntrinsicSDNode>(*UI) || isa<MemSDNode>(*UI)) {
11775 // We've not visited this use, and we care about it (it could have an
11776 // ordering dependency with the original node).
11778 Aliases.push_back(OriginalChain);
11782 // We've not visited this use, but we don't care about it. Mark it as
11783 // visited and enqueue it to the worklist.
11784 Worklist.push_back(*UI);
11789 /// FindBetterChain - Walk up chain skipping non-aliasing memory nodes, looking
11790 /// for a better chain (aliasing node.)
11791 SDValue DAGCombiner::FindBetterChain(SDNode *N, SDValue OldChain) {
11792 SmallVector<SDValue, 8> Aliases; // Ops for replacing token factor.
11794 // Accumulate all the aliases to this node.
11795 GatherAllAliases(N, OldChain, Aliases);
11797 // If no operands then chain to entry token.
11798 if (Aliases.size() == 0)
11799 return DAG.getEntryNode();
11801 // If a single operand then chain to it. We don't need to revisit it.
11802 if (Aliases.size() == 1)
11805 // Construct a custom tailored token factor.
11806 return DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, Aliases);
11809 // SelectionDAG::Combine - This is the entry point for the file.
11811 void SelectionDAG::Combine(CombineLevel Level, AliasAnalysis &AA,
11812 CodeGenOpt::Level OptLevel) {
11813 /// run - This is the main entry point to this class.
11815 DAGCombiner(*this, AA, OptLevel).Run(Level);