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 ConstantSDNode *CN = BV->getConstantSplatValue();
660 // BuildVectors can truncate their operands. Ignore that case here.
661 if (CN && CN->getValueType(0) == N.getValueType().getScalarType())
668 SDValue DAGCombiner::ReassociateOps(unsigned Opc, SDLoc DL,
669 SDValue N0, SDValue N1) {
670 EVT VT = N0.getValueType();
671 if (N0.getOpcode() == Opc) {
672 if (SDNode *L = isConstantBuildVectorOrConstantInt(N0.getOperand(1))) {
673 if (SDNode *R = isConstantBuildVectorOrConstantInt(N1)) {
674 // reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2))
675 SDValue OpNode = DAG.FoldConstantArithmetic(Opc, VT, L, R);
676 if (!OpNode.getNode())
678 return DAG.getNode(Opc, DL, VT, N0.getOperand(0), OpNode);
680 if (N0.hasOneUse()) {
681 // reassoc. (op (op x, c1), y) -> (op (op x, y), c1) iff x+c1 has one
683 SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N0.getOperand(0), N1);
684 if (!OpNode.getNode())
686 AddToWorkList(OpNode.getNode());
687 return DAG.getNode(Opc, DL, VT, OpNode, N0.getOperand(1));
692 if (N1.getOpcode() == Opc) {
693 if (SDNode *R = isConstantBuildVectorOrConstantInt(N1.getOperand(1))) {
694 if (SDNode *L = isConstantBuildVectorOrConstantInt(N0)) {
695 // reassoc. (op c2, (op x, c1)) -> (op x, (op c1, c2))
696 SDValue OpNode = DAG.FoldConstantArithmetic(Opc, VT, R, L);
697 if (!OpNode.getNode())
699 return DAG.getNode(Opc, DL, VT, N1.getOperand(0), OpNode);
701 if (N1.hasOneUse()) {
702 // reassoc. (op y, (op x, c1)) -> (op (op x, y), c1) iff x+c1 has one
704 SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N1.getOperand(0), N0);
705 if (!OpNode.getNode())
707 AddToWorkList(OpNode.getNode());
708 return DAG.getNode(Opc, DL, VT, OpNode, N1.getOperand(1));
716 SDValue DAGCombiner::CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
718 assert(N->getNumValues() == NumTo && "Broken CombineTo call!");
720 DEBUG(dbgs() << "\nReplacing.1 ";
722 dbgs() << "\nWith: ";
723 To[0].getNode()->dump(&DAG);
724 dbgs() << " and " << NumTo-1 << " other values\n";
725 for (unsigned i = 0, e = NumTo; i != e; ++i)
726 assert((!To[i].getNode() ||
727 N->getValueType(i) == To[i].getValueType()) &&
728 "Cannot combine value to value of different type!"));
729 WorkListRemover DeadNodes(*this);
730 DAG.ReplaceAllUsesWith(N, To);
732 // Push the new nodes and any users onto the worklist
733 for (unsigned i = 0, e = NumTo; i != e; ++i) {
734 if (To[i].getNode()) {
735 AddToWorkList(To[i].getNode());
736 AddUsersToWorkList(To[i].getNode());
741 // Finally, if the node is now dead, remove it from the graph. The node
742 // may not be dead if the replacement process recursively simplified to
743 // something else needing this node.
744 if (N->use_empty()) {
745 // Nodes can be reintroduced into the worklist. Make sure we do not
746 // process a node that has been replaced.
747 removeFromWorkList(N);
749 // Finally, since the node is now dead, remove it from the graph.
752 return SDValue(N, 0);
756 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
757 // Replace all uses. If any nodes become isomorphic to other nodes and
758 // are deleted, make sure to remove them from our worklist.
759 WorkListRemover DeadNodes(*this);
760 DAG.ReplaceAllUsesOfValueWith(TLO.Old, TLO.New);
762 // Push the new node and any (possibly new) users onto the worklist.
763 AddToWorkList(TLO.New.getNode());
764 AddUsersToWorkList(TLO.New.getNode());
766 // Finally, if the node is now dead, remove it from the graph. The node
767 // may not be dead if the replacement process recursively simplified to
768 // something else needing this node.
769 if (TLO.Old.getNode()->use_empty()) {
770 removeFromWorkList(TLO.Old.getNode());
772 // If the operands of this node are only used by the node, they will now
773 // be dead. Make sure to visit them first to delete dead nodes early.
774 for (unsigned i = 0, e = TLO.Old.getNode()->getNumOperands(); i != e; ++i)
775 if (TLO.Old.getNode()->getOperand(i).getNode()->hasOneUse())
776 AddToWorkList(TLO.Old.getNode()->getOperand(i).getNode());
778 DAG.DeleteNode(TLO.Old.getNode());
782 /// SimplifyDemandedBits - Check the specified integer node value to see if
783 /// it can be simplified or if things it uses can be simplified by bit
784 /// propagation. If so, return true.
785 bool DAGCombiner::SimplifyDemandedBits(SDValue Op, const APInt &Demanded) {
786 TargetLowering::TargetLoweringOpt TLO(DAG, LegalTypes, LegalOperations);
787 APInt KnownZero, KnownOne;
788 if (!TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
792 AddToWorkList(Op.getNode());
794 // Replace the old value with the new one.
796 DEBUG(dbgs() << "\nReplacing.2 ";
797 TLO.Old.getNode()->dump(&DAG);
798 dbgs() << "\nWith: ";
799 TLO.New.getNode()->dump(&DAG);
802 CommitTargetLoweringOpt(TLO);
806 void DAGCombiner::ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad) {
808 EVT VT = Load->getValueType(0);
809 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, VT, SDValue(ExtLoad, 0));
811 DEBUG(dbgs() << "\nReplacing.9 ";
813 dbgs() << "\nWith: ";
814 Trunc.getNode()->dump(&DAG);
816 WorkListRemover DeadNodes(*this);
817 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 0), Trunc);
818 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), SDValue(ExtLoad, 1));
819 removeFromWorkList(Load);
820 DAG.DeleteNode(Load);
821 AddToWorkList(Trunc.getNode());
824 SDValue DAGCombiner::PromoteOperand(SDValue Op, EVT PVT, bool &Replace) {
827 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Op)) {
828 EVT MemVT = LD->getMemoryVT();
829 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
830 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT) ? ISD::ZEXTLOAD
832 : LD->getExtensionType();
834 return DAG.getExtLoad(ExtType, dl, PVT,
835 LD->getChain(), LD->getBasePtr(),
836 MemVT, LD->getMemOperand());
839 unsigned Opc = Op.getOpcode();
842 case ISD::AssertSext:
843 return DAG.getNode(ISD::AssertSext, dl, PVT,
844 SExtPromoteOperand(Op.getOperand(0), PVT),
846 case ISD::AssertZext:
847 return DAG.getNode(ISD::AssertZext, dl, PVT,
848 ZExtPromoteOperand(Op.getOperand(0), PVT),
850 case ISD::Constant: {
852 Op.getValueType().isByteSized() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
853 return DAG.getNode(ExtOpc, dl, PVT, Op);
857 if (!TLI.isOperationLegal(ISD::ANY_EXTEND, PVT))
859 return DAG.getNode(ISD::ANY_EXTEND, dl, PVT, Op);
862 SDValue DAGCombiner::SExtPromoteOperand(SDValue Op, EVT PVT) {
863 if (!TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, PVT))
865 EVT OldVT = Op.getValueType();
867 bool Replace = false;
868 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
869 if (!NewOp.getNode())
871 AddToWorkList(NewOp.getNode());
874 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
875 return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, NewOp.getValueType(), NewOp,
876 DAG.getValueType(OldVT));
879 SDValue DAGCombiner::ZExtPromoteOperand(SDValue Op, EVT PVT) {
880 EVT OldVT = Op.getValueType();
882 bool Replace = false;
883 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
884 if (!NewOp.getNode())
886 AddToWorkList(NewOp.getNode());
889 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
890 return DAG.getZeroExtendInReg(NewOp, dl, OldVT);
893 /// PromoteIntBinOp - Promote the specified integer binary operation if the
894 /// target indicates it is beneficial. e.g. On x86, it's usually better to
895 /// promote i16 operations to i32 since i16 instructions are longer.
896 SDValue DAGCombiner::PromoteIntBinOp(SDValue Op) {
897 if (!LegalOperations)
900 EVT VT = Op.getValueType();
901 if (VT.isVector() || !VT.isInteger())
904 // If operation type is 'undesirable', e.g. i16 on x86, consider
906 unsigned Opc = Op.getOpcode();
907 if (TLI.isTypeDesirableForOp(Opc, VT))
911 // Consult target whether it is a good idea to promote this operation and
912 // what's the right type to promote it to.
913 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
914 assert(PVT != VT && "Don't know what type to promote to!");
916 bool Replace0 = false;
917 SDValue N0 = Op.getOperand(0);
918 SDValue NN0 = PromoteOperand(N0, PVT, Replace0);
922 bool Replace1 = false;
923 SDValue N1 = Op.getOperand(1);
928 NN1 = PromoteOperand(N1, PVT, Replace1);
933 AddToWorkList(NN0.getNode());
935 AddToWorkList(NN1.getNode());
938 ReplaceLoadWithPromotedLoad(N0.getNode(), NN0.getNode());
940 ReplaceLoadWithPromotedLoad(N1.getNode(), NN1.getNode());
942 DEBUG(dbgs() << "\nPromoting ";
943 Op.getNode()->dump(&DAG));
945 return DAG.getNode(ISD::TRUNCATE, dl, VT,
946 DAG.getNode(Opc, dl, PVT, NN0, NN1));
951 /// PromoteIntShiftOp - Promote the specified integer shift operation if the
952 /// target indicates it is beneficial. e.g. On x86, it's usually better to
953 /// promote i16 operations to i32 since i16 instructions are longer.
954 SDValue DAGCombiner::PromoteIntShiftOp(SDValue Op) {
955 if (!LegalOperations)
958 EVT VT = Op.getValueType();
959 if (VT.isVector() || !VT.isInteger())
962 // If operation type is 'undesirable', e.g. i16 on x86, consider
964 unsigned Opc = Op.getOpcode();
965 if (TLI.isTypeDesirableForOp(Opc, VT))
969 // Consult target whether it is a good idea to promote this operation and
970 // what's the right type to promote it to.
971 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
972 assert(PVT != VT && "Don't know what type to promote to!");
974 bool Replace = false;
975 SDValue N0 = Op.getOperand(0);
977 N0 = SExtPromoteOperand(Op.getOperand(0), PVT);
978 else if (Opc == ISD::SRL)
979 N0 = ZExtPromoteOperand(Op.getOperand(0), PVT);
981 N0 = PromoteOperand(N0, PVT, Replace);
985 AddToWorkList(N0.getNode());
987 ReplaceLoadWithPromotedLoad(Op.getOperand(0).getNode(), N0.getNode());
989 DEBUG(dbgs() << "\nPromoting ";
990 Op.getNode()->dump(&DAG));
992 return DAG.getNode(ISD::TRUNCATE, dl, VT,
993 DAG.getNode(Opc, dl, PVT, N0, Op.getOperand(1)));
998 SDValue DAGCombiner::PromoteExtend(SDValue Op) {
999 if (!LegalOperations)
1002 EVT VT = Op.getValueType();
1003 if (VT.isVector() || !VT.isInteger())
1006 // If operation type is 'undesirable', e.g. i16 on x86, consider
1008 unsigned Opc = Op.getOpcode();
1009 if (TLI.isTypeDesirableForOp(Opc, VT))
1013 // Consult target whether it is a good idea to promote this operation and
1014 // what's the right type to promote it to.
1015 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1016 assert(PVT != VT && "Don't know what type to promote to!");
1017 // fold (aext (aext x)) -> (aext x)
1018 // fold (aext (zext x)) -> (zext x)
1019 // fold (aext (sext x)) -> (sext x)
1020 DEBUG(dbgs() << "\nPromoting ";
1021 Op.getNode()->dump(&DAG));
1022 return DAG.getNode(Op.getOpcode(), SDLoc(Op), VT, Op.getOperand(0));
1027 bool DAGCombiner::PromoteLoad(SDValue Op) {
1028 if (!LegalOperations)
1031 EVT VT = Op.getValueType();
1032 if (VT.isVector() || !VT.isInteger())
1035 // If operation type is 'undesirable', e.g. i16 on x86, consider
1037 unsigned Opc = Op.getOpcode();
1038 if (TLI.isTypeDesirableForOp(Opc, VT))
1042 // Consult target whether it is a good idea to promote this operation and
1043 // what's the right type to promote it to.
1044 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1045 assert(PVT != VT && "Don't know what type to promote to!");
1048 SDNode *N = Op.getNode();
1049 LoadSDNode *LD = cast<LoadSDNode>(N);
1050 EVT MemVT = LD->getMemoryVT();
1051 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
1052 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT) ? ISD::ZEXTLOAD
1054 : LD->getExtensionType();
1055 SDValue NewLD = DAG.getExtLoad(ExtType, dl, PVT,
1056 LD->getChain(), LD->getBasePtr(),
1057 MemVT, LD->getMemOperand());
1058 SDValue Result = DAG.getNode(ISD::TRUNCATE, dl, VT, NewLD);
1060 DEBUG(dbgs() << "\nPromoting ";
1063 Result.getNode()->dump(&DAG);
1065 WorkListRemover DeadNodes(*this);
1066 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result);
1067 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), NewLD.getValue(1));
1068 removeFromWorkList(N);
1070 AddToWorkList(Result.getNode());
1077 //===----------------------------------------------------------------------===//
1078 // Main DAG Combiner implementation
1079 //===----------------------------------------------------------------------===//
1081 void DAGCombiner::Run(CombineLevel AtLevel) {
1082 // set the instance variables, so that the various visit routines may use it.
1084 LegalOperations = Level >= AfterLegalizeVectorOps;
1085 LegalTypes = Level >= AfterLegalizeTypes;
1087 // Add all the dag nodes to the worklist.
1088 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
1089 E = DAG.allnodes_end(); I != E; ++I)
1092 // Create a dummy node (which is not added to allnodes), that adds a reference
1093 // to the root node, preventing it from being deleted, and tracking any
1094 // changes of the root.
1095 HandleSDNode Dummy(DAG.getRoot());
1097 // The root of the dag may dangle to deleted nodes until the dag combiner is
1098 // done. Set it to null to avoid confusion.
1099 DAG.setRoot(SDValue());
1101 // while the worklist isn't empty, find a node and
1102 // try and combine it.
1103 while (!WorkListContents.empty()) {
1105 // The WorkListOrder holds the SDNodes in order, but it may contain
1107 // In order to avoid a linear scan, we use a set (O(log N)) to hold what the
1108 // worklist *should* contain, and check the node we want to visit is should
1109 // actually be visited.
1111 N = WorkListOrder.pop_back_val();
1112 } while (!WorkListContents.erase(N));
1114 // If N has no uses, it is dead. Make sure to revisit all N's operands once
1115 // N is deleted from the DAG, since they too may now be dead or may have a
1116 // reduced number of uses, allowing other xforms.
1117 if (N->use_empty() && N != &Dummy) {
1118 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1119 AddToWorkList(N->getOperand(i).getNode());
1125 SDValue RV = combine(N);
1132 // If we get back the same node we passed in, rather than a new node or
1133 // zero, we know that the node must have defined multiple values and
1134 // CombineTo was used. Since CombineTo takes care of the worklist
1135 // mechanics for us, we have no work to do in this case.
1136 if (RV.getNode() == N)
1139 assert(N->getOpcode() != ISD::DELETED_NODE &&
1140 RV.getNode()->getOpcode() != ISD::DELETED_NODE &&
1141 "Node was deleted but visit returned new node!");
1143 DEBUG(dbgs() << "\nReplacing.3 ";
1145 dbgs() << "\nWith: ";
1146 RV.getNode()->dump(&DAG);
1149 // Transfer debug value.
1150 DAG.TransferDbgValues(SDValue(N, 0), RV);
1151 WorkListRemover DeadNodes(*this);
1152 if (N->getNumValues() == RV.getNode()->getNumValues())
1153 DAG.ReplaceAllUsesWith(N, RV.getNode());
1155 assert(N->getValueType(0) == RV.getValueType() &&
1156 N->getNumValues() == 1 && "Type mismatch");
1158 DAG.ReplaceAllUsesWith(N, &OpV);
1161 // Push the new node and any users onto the worklist
1162 AddToWorkList(RV.getNode());
1163 AddUsersToWorkList(RV.getNode());
1165 // Add any uses of the old node to the worklist in case this node is the
1166 // last one that uses them. They may become dead after this node is
1168 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1169 AddToWorkList(N->getOperand(i).getNode());
1171 // Finally, if the node is now dead, remove it from the graph. The node
1172 // may not be dead if the replacement process recursively simplified to
1173 // something else needing this node.
1174 if (N->use_empty()) {
1175 // Nodes can be reintroduced into the worklist. Make sure we do not
1176 // process a node that has been replaced.
1177 removeFromWorkList(N);
1179 // Finally, since the node is now dead, remove it from the graph.
1184 // If the root changed (e.g. it was a dead load, update the root).
1185 DAG.setRoot(Dummy.getValue());
1186 DAG.RemoveDeadNodes();
1189 SDValue DAGCombiner::visit(SDNode *N) {
1190 switch (N->getOpcode()) {
1192 case ISD::TokenFactor: return visitTokenFactor(N);
1193 case ISD::MERGE_VALUES: return visitMERGE_VALUES(N);
1194 case ISD::ADD: return visitADD(N);
1195 case ISD::SUB: return visitSUB(N);
1196 case ISD::ADDC: return visitADDC(N);
1197 case ISD::SUBC: return visitSUBC(N);
1198 case ISD::ADDE: return visitADDE(N);
1199 case ISD::SUBE: return visitSUBE(N);
1200 case ISD::MUL: return visitMUL(N);
1201 case ISD::SDIV: return visitSDIV(N);
1202 case ISD::UDIV: return visitUDIV(N);
1203 case ISD::SREM: return visitSREM(N);
1204 case ISD::UREM: return visitUREM(N);
1205 case ISD::MULHU: return visitMULHU(N);
1206 case ISD::MULHS: return visitMULHS(N);
1207 case ISD::SMUL_LOHI: return visitSMUL_LOHI(N);
1208 case ISD::UMUL_LOHI: return visitUMUL_LOHI(N);
1209 case ISD::SMULO: return visitSMULO(N);
1210 case ISD::UMULO: return visitUMULO(N);
1211 case ISD::SDIVREM: return visitSDIVREM(N);
1212 case ISD::UDIVREM: return visitUDIVREM(N);
1213 case ISD::AND: return visitAND(N);
1214 case ISD::OR: return visitOR(N);
1215 case ISD::XOR: return visitXOR(N);
1216 case ISD::SHL: return visitSHL(N);
1217 case ISD::SRA: return visitSRA(N);
1218 case ISD::SRL: return visitSRL(N);
1220 case ISD::ROTL: return visitRotate(N);
1221 case ISD::CTLZ: return visitCTLZ(N);
1222 case ISD::CTLZ_ZERO_UNDEF: return visitCTLZ_ZERO_UNDEF(N);
1223 case ISD::CTTZ: return visitCTTZ(N);
1224 case ISD::CTTZ_ZERO_UNDEF: return visitCTTZ_ZERO_UNDEF(N);
1225 case ISD::CTPOP: return visitCTPOP(N);
1226 case ISD::SELECT: return visitSELECT(N);
1227 case ISD::VSELECT: return visitVSELECT(N);
1228 case ISD::SELECT_CC: return visitSELECT_CC(N);
1229 case ISD::SETCC: return visitSETCC(N);
1230 case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N);
1231 case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N);
1232 case ISD::ANY_EXTEND: return visitANY_EXTEND(N);
1233 case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N);
1234 case ISD::TRUNCATE: return visitTRUNCATE(N);
1235 case ISD::BITCAST: return visitBITCAST(N);
1236 case ISD::BUILD_PAIR: return visitBUILD_PAIR(N);
1237 case ISD::FADD: return visitFADD(N);
1238 case ISD::FSUB: return visitFSUB(N);
1239 case ISD::FMUL: return visitFMUL(N);
1240 case ISD::FMA: return visitFMA(N);
1241 case ISD::FDIV: return visitFDIV(N);
1242 case ISD::FREM: return visitFREM(N);
1243 case ISD::FCOPYSIGN: return visitFCOPYSIGN(N);
1244 case ISD::SINT_TO_FP: return visitSINT_TO_FP(N);
1245 case ISD::UINT_TO_FP: return visitUINT_TO_FP(N);
1246 case ISD::FP_TO_SINT: return visitFP_TO_SINT(N);
1247 case ISD::FP_TO_UINT: return visitFP_TO_UINT(N);
1248 case ISD::FP_ROUND: return visitFP_ROUND(N);
1249 case ISD::FP_ROUND_INREG: return visitFP_ROUND_INREG(N);
1250 case ISD::FP_EXTEND: return visitFP_EXTEND(N);
1251 case ISD::FNEG: return visitFNEG(N);
1252 case ISD::FABS: return visitFABS(N);
1253 case ISD::FFLOOR: return visitFFLOOR(N);
1254 case ISD::FCEIL: return visitFCEIL(N);
1255 case ISD::FTRUNC: return visitFTRUNC(N);
1256 case ISD::BRCOND: return visitBRCOND(N);
1257 case ISD::BR_CC: return visitBR_CC(N);
1258 case ISD::LOAD: return visitLOAD(N);
1259 case ISD::STORE: return visitSTORE(N);
1260 case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N);
1261 case ISD::EXTRACT_VECTOR_ELT: return visitEXTRACT_VECTOR_ELT(N);
1262 case ISD::BUILD_VECTOR: return visitBUILD_VECTOR(N);
1263 case ISD::CONCAT_VECTORS: return visitCONCAT_VECTORS(N);
1264 case ISD::EXTRACT_SUBVECTOR: return visitEXTRACT_SUBVECTOR(N);
1265 case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N);
1266 case ISD::INSERT_SUBVECTOR: return visitINSERT_SUBVECTOR(N);
1271 SDValue DAGCombiner::combine(SDNode *N) {
1272 SDValue RV = visit(N);
1274 // If nothing happened, try a target-specific DAG combine.
1275 if (!RV.getNode()) {
1276 assert(N->getOpcode() != ISD::DELETED_NODE &&
1277 "Node was deleted but visit returned NULL!");
1279 if (N->getOpcode() >= ISD::BUILTIN_OP_END ||
1280 TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode())) {
1282 // Expose the DAG combiner to the target combiner impls.
1283 TargetLowering::DAGCombinerInfo
1284 DagCombineInfo(DAG, Level, false, this);
1286 RV = TLI.PerformDAGCombine(N, DagCombineInfo);
1290 // If nothing happened still, try promoting the operation.
1291 if (!RV.getNode()) {
1292 switch (N->getOpcode()) {
1300 RV = PromoteIntBinOp(SDValue(N, 0));
1305 RV = PromoteIntShiftOp(SDValue(N, 0));
1307 case ISD::SIGN_EXTEND:
1308 case ISD::ZERO_EXTEND:
1309 case ISD::ANY_EXTEND:
1310 RV = PromoteExtend(SDValue(N, 0));
1313 if (PromoteLoad(SDValue(N, 0)))
1319 // If N is a commutative binary node, try commuting it to enable more
1321 if (!RV.getNode() && SelectionDAG::isCommutativeBinOp(N->getOpcode()) &&
1322 N->getNumValues() == 1) {
1323 SDValue N0 = N->getOperand(0);
1324 SDValue N1 = N->getOperand(1);
1326 // Constant operands are canonicalized to RHS.
1327 if (isa<ConstantSDNode>(N0) || !isa<ConstantSDNode>(N1)) {
1328 SDValue Ops[] = { N1, N0 };
1329 SDNode *CSENode = DAG.getNodeIfExists(N->getOpcode(), N->getVTList(),
1332 return SDValue(CSENode, 0);
1339 /// getInputChainForNode - Given a node, return its input chain if it has one,
1340 /// otherwise return a null sd operand.
1341 static SDValue getInputChainForNode(SDNode *N) {
1342 if (unsigned NumOps = N->getNumOperands()) {
1343 if (N->getOperand(0).getValueType() == MVT::Other)
1344 return N->getOperand(0);
1345 if (N->getOperand(NumOps-1).getValueType() == MVT::Other)
1346 return N->getOperand(NumOps-1);
1347 for (unsigned i = 1; i < NumOps-1; ++i)
1348 if (N->getOperand(i).getValueType() == MVT::Other)
1349 return N->getOperand(i);
1354 SDValue DAGCombiner::visitTokenFactor(SDNode *N) {
1355 // If N has two operands, where one has an input chain equal to the other,
1356 // the 'other' chain is redundant.
1357 if (N->getNumOperands() == 2) {
1358 if (getInputChainForNode(N->getOperand(0).getNode()) == N->getOperand(1))
1359 return N->getOperand(0);
1360 if (getInputChainForNode(N->getOperand(1).getNode()) == N->getOperand(0))
1361 return N->getOperand(1);
1364 SmallVector<SDNode *, 8> TFs; // List of token factors to visit.
1365 SmallVector<SDValue, 8> Ops; // Ops for replacing token factor.
1366 SmallPtrSet<SDNode*, 16> SeenOps;
1367 bool Changed = false; // If we should replace this token factor.
1369 // Start out with this token factor.
1372 // Iterate through token factors. The TFs grows when new token factors are
1374 for (unsigned i = 0; i < TFs.size(); ++i) {
1375 SDNode *TF = TFs[i];
1377 // Check each of the operands.
1378 for (unsigned i = 0, ie = TF->getNumOperands(); i != ie; ++i) {
1379 SDValue Op = TF->getOperand(i);
1381 switch (Op.getOpcode()) {
1382 case ISD::EntryToken:
1383 // Entry tokens don't need to be added to the list. They are
1388 case ISD::TokenFactor:
1389 if (Op.hasOneUse() &&
1390 std::find(TFs.begin(), TFs.end(), Op.getNode()) == TFs.end()) {
1391 // Queue up for processing.
1392 TFs.push_back(Op.getNode());
1393 // Clean up in case the token factor is removed.
1394 AddToWorkList(Op.getNode());
1401 // Only add if it isn't already in the list.
1402 if (SeenOps.insert(Op.getNode()))
1413 // If we've change things around then replace token factor.
1416 // The entry token is the only possible outcome.
1417 Result = DAG.getEntryNode();
1419 // New and improved token factor.
1420 Result = DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, Ops);
1423 // Don't add users to work list.
1424 return CombineTo(N, Result, false);
1430 /// MERGE_VALUES can always be eliminated.
1431 SDValue DAGCombiner::visitMERGE_VALUES(SDNode *N) {
1432 WorkListRemover DeadNodes(*this);
1433 // Replacing results may cause a different MERGE_VALUES to suddenly
1434 // be CSE'd with N, and carry its uses with it. Iterate until no
1435 // uses remain, to ensure that the node can be safely deleted.
1436 // First add the users of this node to the work list so that they
1437 // can be tried again once they have new operands.
1438 AddUsersToWorkList(N);
1440 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1441 DAG.ReplaceAllUsesOfValueWith(SDValue(N, i), N->getOperand(i));
1442 } while (!N->use_empty());
1443 removeFromWorkList(N);
1445 return SDValue(N, 0); // Return N so it doesn't get rechecked!
1449 SDValue combineShlAddConstant(SDLoc DL, SDValue N0, SDValue N1,
1450 SelectionDAG &DAG) {
1451 EVT VT = N0.getValueType();
1452 SDValue N00 = N0.getOperand(0);
1453 SDValue N01 = N0.getOperand(1);
1454 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N01);
1456 if (N01C && N00.getOpcode() == ISD::ADD && N00.getNode()->hasOneUse() &&
1457 isa<ConstantSDNode>(N00.getOperand(1))) {
1458 // fold (add (shl (add x, c1), c2), ) -> (add (add (shl x, c2), c1<<c2), )
1459 N0 = DAG.getNode(ISD::ADD, SDLoc(N0), VT,
1460 DAG.getNode(ISD::SHL, SDLoc(N00), VT,
1461 N00.getOperand(0), N01),
1462 DAG.getNode(ISD::SHL, SDLoc(N01), VT,
1463 N00.getOperand(1), N01));
1464 return DAG.getNode(ISD::ADD, DL, VT, N0, N1);
1470 SDValue DAGCombiner::visitADD(SDNode *N) {
1471 SDValue N0 = N->getOperand(0);
1472 SDValue N1 = N->getOperand(1);
1473 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1474 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1475 EVT VT = N0.getValueType();
1478 if (VT.isVector()) {
1479 SDValue FoldedVOp = SimplifyVBinOp(N);
1480 if (FoldedVOp.getNode()) return FoldedVOp;
1482 // fold (add x, 0) -> x, vector edition
1483 if (ISD::isBuildVectorAllZeros(N1.getNode()))
1485 if (ISD::isBuildVectorAllZeros(N0.getNode()))
1489 // fold (add x, undef) -> undef
1490 if (N0.getOpcode() == ISD::UNDEF)
1492 if (N1.getOpcode() == ISD::UNDEF)
1494 // fold (add c1, c2) -> c1+c2
1496 return DAG.FoldConstantArithmetic(ISD::ADD, VT, N0C, N1C);
1497 // canonicalize constant to RHS
1499 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N1, N0);
1500 // fold (add x, 0) -> x
1501 if (N1C && N1C->isNullValue())
1503 // fold (add Sym, c) -> Sym+c
1504 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1505 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA) && N1C &&
1506 GA->getOpcode() == ISD::GlobalAddress)
1507 return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT,
1509 (uint64_t)N1C->getSExtValue());
1510 // fold ((c1-A)+c2) -> (c1+c2)-A
1511 if (N1C && N0.getOpcode() == ISD::SUB)
1512 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getOperand(0)))
1513 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1514 DAG.getConstant(N1C->getAPIntValue()+
1515 N0C->getAPIntValue(), VT),
1518 SDValue RADD = ReassociateOps(ISD::ADD, SDLoc(N), N0, N1);
1521 // fold ((0-A) + B) -> B-A
1522 if (N0.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N0.getOperand(0)) &&
1523 cast<ConstantSDNode>(N0.getOperand(0))->isNullValue())
1524 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1, N0.getOperand(1));
1525 // fold (A + (0-B)) -> A-B
1526 if (N1.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N1.getOperand(0)) &&
1527 cast<ConstantSDNode>(N1.getOperand(0))->isNullValue())
1528 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, N1.getOperand(1));
1529 // fold (A+(B-A)) -> B
1530 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1))
1531 return N1.getOperand(0);
1532 // fold ((B-A)+A) -> B
1533 if (N0.getOpcode() == ISD::SUB && N1 == N0.getOperand(1))
1534 return N0.getOperand(0);
1535 // fold (A+(B-(A+C))) to (B-C)
1536 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1537 N0 == N1.getOperand(1).getOperand(0))
1538 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1.getOperand(0),
1539 N1.getOperand(1).getOperand(1));
1540 // fold (A+(B-(C+A))) to (B-C)
1541 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1542 N0 == N1.getOperand(1).getOperand(1))
1543 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1.getOperand(0),
1544 N1.getOperand(1).getOperand(0));
1545 // fold (A+((B-A)+or-C)) to (B+or-C)
1546 if ((N1.getOpcode() == ISD::SUB || N1.getOpcode() == ISD::ADD) &&
1547 N1.getOperand(0).getOpcode() == ISD::SUB &&
1548 N0 == N1.getOperand(0).getOperand(1))
1549 return DAG.getNode(N1.getOpcode(), SDLoc(N), VT,
1550 N1.getOperand(0).getOperand(0), N1.getOperand(1));
1552 // fold (A-B)+(C-D) to (A+C)-(B+D) when A or C is constant
1553 if (N0.getOpcode() == ISD::SUB && N1.getOpcode() == ISD::SUB) {
1554 SDValue N00 = N0.getOperand(0);
1555 SDValue N01 = N0.getOperand(1);
1556 SDValue N10 = N1.getOperand(0);
1557 SDValue N11 = N1.getOperand(1);
1559 if (isa<ConstantSDNode>(N00) || isa<ConstantSDNode>(N10))
1560 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1561 DAG.getNode(ISD::ADD, SDLoc(N0), VT, N00, N10),
1562 DAG.getNode(ISD::ADD, SDLoc(N1), VT, N01, N11));
1565 if (!VT.isVector() && SimplifyDemandedBits(SDValue(N, 0)))
1566 return SDValue(N, 0);
1568 // fold (a+b) -> (a|b) iff a and b share no bits.
1569 if (VT.isInteger() && !VT.isVector()) {
1570 APInt LHSZero, LHSOne;
1571 APInt RHSZero, RHSOne;
1572 DAG.computeKnownBits(N0, LHSZero, LHSOne);
1574 if (LHSZero.getBoolValue()) {
1575 DAG.computeKnownBits(N1, RHSZero, RHSOne);
1577 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1578 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1579 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero){
1580 if (!LegalOperations || TLI.isOperationLegal(ISD::OR, VT))
1581 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1);
1586 // fold (add (shl (add x, c1), c2), ) -> (add (add (shl x, c2), c1<<c2), )
1587 if (N0.getOpcode() == ISD::SHL && N0.getNode()->hasOneUse()) {
1588 SDValue Result = combineShlAddConstant(SDLoc(N), N0, N1, DAG);
1589 if (Result.getNode()) return Result;
1591 if (N1.getOpcode() == ISD::SHL && N1.getNode()->hasOneUse()) {
1592 SDValue Result = combineShlAddConstant(SDLoc(N), N1, N0, DAG);
1593 if (Result.getNode()) return Result;
1596 // fold (add x, shl(0 - y, n)) -> sub(x, shl(y, n))
1597 if (N1.getOpcode() == ISD::SHL &&
1598 N1.getOperand(0).getOpcode() == ISD::SUB)
1599 if (ConstantSDNode *C =
1600 dyn_cast<ConstantSDNode>(N1.getOperand(0).getOperand(0)))
1601 if (C->getAPIntValue() == 0)
1602 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N0,
1603 DAG.getNode(ISD::SHL, SDLoc(N), VT,
1604 N1.getOperand(0).getOperand(1),
1606 if (N0.getOpcode() == ISD::SHL &&
1607 N0.getOperand(0).getOpcode() == ISD::SUB)
1608 if (ConstantSDNode *C =
1609 dyn_cast<ConstantSDNode>(N0.getOperand(0).getOperand(0)))
1610 if (C->getAPIntValue() == 0)
1611 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1,
1612 DAG.getNode(ISD::SHL, SDLoc(N), VT,
1613 N0.getOperand(0).getOperand(1),
1616 if (N1.getOpcode() == ISD::AND) {
1617 SDValue AndOp0 = N1.getOperand(0);
1618 ConstantSDNode *AndOp1 = dyn_cast<ConstantSDNode>(N1->getOperand(1));
1619 unsigned NumSignBits = DAG.ComputeNumSignBits(AndOp0);
1620 unsigned DestBits = VT.getScalarType().getSizeInBits();
1622 // (add z, (and (sbbl x, x), 1)) -> (sub z, (sbbl x, x))
1623 // and similar xforms where the inner op is either ~0 or 0.
1624 if (NumSignBits == DestBits && AndOp1 && AndOp1->isOne()) {
1626 return DAG.getNode(ISD::SUB, DL, VT, N->getOperand(0), AndOp0);
1630 // add (sext i1), X -> sub X, (zext i1)
1631 if (N0.getOpcode() == ISD::SIGN_EXTEND &&
1632 N0.getOperand(0).getValueType() == MVT::i1 &&
1633 !TLI.isOperationLegal(ISD::SIGN_EXTEND, MVT::i1)) {
1635 SDValue ZExt = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0));
1636 return DAG.getNode(ISD::SUB, DL, VT, N1, ZExt);
1642 SDValue DAGCombiner::visitADDC(SDNode *N) {
1643 SDValue N0 = N->getOperand(0);
1644 SDValue N1 = N->getOperand(1);
1645 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1646 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1647 EVT VT = N0.getValueType();
1649 // If the flag result is dead, turn this into an ADD.
1650 if (!N->hasAnyUseOfValue(1))
1651 return CombineTo(N, DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N1),
1652 DAG.getNode(ISD::CARRY_FALSE,
1653 SDLoc(N), MVT::Glue));
1655 // canonicalize constant to RHS.
1657 return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N1, N0);
1659 // fold (addc x, 0) -> x + no carry out
1660 if (N1C && N1C->isNullValue())
1661 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE,
1662 SDLoc(N), MVT::Glue));
1664 // fold (addc a, b) -> (or a, b), CARRY_FALSE iff a and b share no bits.
1665 APInt LHSZero, LHSOne;
1666 APInt RHSZero, RHSOne;
1667 DAG.computeKnownBits(N0, LHSZero, LHSOne);
1669 if (LHSZero.getBoolValue()) {
1670 DAG.computeKnownBits(N1, RHSZero, RHSOne);
1672 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1673 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1674 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero)
1675 return CombineTo(N, DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1),
1676 DAG.getNode(ISD::CARRY_FALSE,
1677 SDLoc(N), MVT::Glue));
1683 SDValue DAGCombiner::visitADDE(SDNode *N) {
1684 SDValue N0 = N->getOperand(0);
1685 SDValue N1 = N->getOperand(1);
1686 SDValue CarryIn = N->getOperand(2);
1687 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1688 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1690 // canonicalize constant to RHS
1692 return DAG.getNode(ISD::ADDE, SDLoc(N), N->getVTList(),
1695 // fold (adde x, y, false) -> (addc x, y)
1696 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1697 return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N0, N1);
1702 // Since it may not be valid to emit a fold to zero for vector initializers
1703 // check if we can before folding.
1704 static SDValue tryFoldToZero(SDLoc DL, const TargetLowering &TLI, EVT VT,
1706 bool LegalOperations, bool LegalTypes) {
1708 return DAG.getConstant(0, VT);
1709 if (!LegalOperations || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
1710 return DAG.getConstant(0, VT);
1714 SDValue DAGCombiner::visitSUB(SDNode *N) {
1715 SDValue N0 = N->getOperand(0);
1716 SDValue N1 = N->getOperand(1);
1717 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1718 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1719 ConstantSDNode *N1C1 = N1.getOpcode() != ISD::ADD ? nullptr :
1720 dyn_cast<ConstantSDNode>(N1.getOperand(1).getNode());
1721 EVT VT = N0.getValueType();
1724 if (VT.isVector()) {
1725 SDValue FoldedVOp = SimplifyVBinOp(N);
1726 if (FoldedVOp.getNode()) return FoldedVOp;
1728 // fold (sub x, 0) -> x, vector edition
1729 if (ISD::isBuildVectorAllZeros(N1.getNode()))
1733 // fold (sub x, x) -> 0
1734 // FIXME: Refactor this and xor and other similar operations together.
1736 return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations, LegalTypes);
1737 // fold (sub c1, c2) -> c1-c2
1739 return DAG.FoldConstantArithmetic(ISD::SUB, VT, N0C, N1C);
1740 // fold (sub x, c) -> (add x, -c)
1742 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0,
1743 DAG.getConstant(-N1C->getAPIntValue(), VT));
1744 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1)
1745 if (N0C && N0C->isAllOnesValue())
1746 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0);
1747 // fold A-(A-B) -> B
1748 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(0))
1749 return N1.getOperand(1);
1750 // fold (A+B)-A -> B
1751 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1)
1752 return N0.getOperand(1);
1753 // fold (A+B)-B -> A
1754 if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1)
1755 return N0.getOperand(0);
1756 // fold C2-(A+C1) -> (C2-C1)-A
1757 if (N1.getOpcode() == ISD::ADD && N0C && N1C1) {
1758 SDValue NewC = DAG.getConstant(N0C->getAPIntValue() - N1C1->getAPIntValue(),
1760 return DAG.getNode(ISD::SUB, SDLoc(N), VT, NewC,
1763 // fold ((A+(B+or-C))-B) -> A+or-C
1764 if (N0.getOpcode() == ISD::ADD &&
1765 (N0.getOperand(1).getOpcode() == ISD::SUB ||
1766 N0.getOperand(1).getOpcode() == ISD::ADD) &&
1767 N0.getOperand(1).getOperand(0) == N1)
1768 return DAG.getNode(N0.getOperand(1).getOpcode(), SDLoc(N), VT,
1769 N0.getOperand(0), N0.getOperand(1).getOperand(1));
1770 // fold ((A+(C+B))-B) -> A+C
1771 if (N0.getOpcode() == ISD::ADD &&
1772 N0.getOperand(1).getOpcode() == ISD::ADD &&
1773 N0.getOperand(1).getOperand(1) == N1)
1774 return DAG.getNode(ISD::ADD, SDLoc(N), VT,
1775 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1776 // fold ((A-(B-C))-C) -> A-B
1777 if (N0.getOpcode() == ISD::SUB &&
1778 N0.getOperand(1).getOpcode() == ISD::SUB &&
1779 N0.getOperand(1).getOperand(1) == N1)
1780 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1781 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1783 // If either operand of a sub is undef, the result is undef
1784 if (N0.getOpcode() == ISD::UNDEF)
1786 if (N1.getOpcode() == ISD::UNDEF)
1789 // If the relocation model supports it, consider symbol offsets.
1790 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1791 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA)) {
1792 // fold (sub Sym, c) -> Sym-c
1793 if (N1C && GA->getOpcode() == ISD::GlobalAddress)
1794 return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT,
1796 (uint64_t)N1C->getSExtValue());
1797 // fold (sub Sym+c1, Sym+c2) -> c1-c2
1798 if (GlobalAddressSDNode *GB = dyn_cast<GlobalAddressSDNode>(N1))
1799 if (GA->getGlobal() == GB->getGlobal())
1800 return DAG.getConstant((uint64_t)GA->getOffset() - GB->getOffset(),
1807 SDValue DAGCombiner::visitSUBC(SDNode *N) {
1808 SDValue N0 = N->getOperand(0);
1809 SDValue N1 = N->getOperand(1);
1810 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1811 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1812 EVT VT = N0.getValueType();
1814 // If the flag result is dead, turn this into an SUB.
1815 if (!N->hasAnyUseOfValue(1))
1816 return CombineTo(N, DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, N1),
1817 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1820 // fold (subc x, x) -> 0 + no borrow
1822 return CombineTo(N, DAG.getConstant(0, VT),
1823 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1826 // fold (subc x, 0) -> x + no borrow
1827 if (N1C && N1C->isNullValue())
1828 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1831 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1) + no borrow
1832 if (N0C && N0C->isAllOnesValue())
1833 return CombineTo(N, DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0),
1834 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1840 SDValue DAGCombiner::visitSUBE(SDNode *N) {
1841 SDValue N0 = N->getOperand(0);
1842 SDValue N1 = N->getOperand(1);
1843 SDValue CarryIn = N->getOperand(2);
1845 // fold (sube x, y, false) -> (subc x, y)
1846 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1847 return DAG.getNode(ISD::SUBC, SDLoc(N), N->getVTList(), N0, N1);
1852 SDValue DAGCombiner::visitMUL(SDNode *N) {
1853 SDValue N0 = N->getOperand(0);
1854 SDValue N1 = N->getOperand(1);
1855 EVT VT = N0.getValueType();
1857 // fold (mul x, undef) -> 0
1858 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
1859 return DAG.getConstant(0, VT);
1861 bool N0IsConst = false;
1862 bool N1IsConst = false;
1863 APInt ConstValue0, ConstValue1;
1865 if (VT.isVector()) {
1866 SDValue FoldedVOp = SimplifyVBinOp(N);
1867 if (FoldedVOp.getNode()) return FoldedVOp;
1869 N0IsConst = isConstantSplatVector(N0.getNode(), ConstValue0);
1870 N1IsConst = isConstantSplatVector(N1.getNode(), ConstValue1);
1872 N0IsConst = dyn_cast<ConstantSDNode>(N0) != nullptr;
1873 ConstValue0 = N0IsConst ? (dyn_cast<ConstantSDNode>(N0))->getAPIntValue()
1875 N1IsConst = dyn_cast<ConstantSDNode>(N1) != nullptr;
1876 ConstValue1 = N1IsConst ? (dyn_cast<ConstantSDNode>(N1))->getAPIntValue()
1880 // fold (mul c1, c2) -> c1*c2
1881 if (N0IsConst && N1IsConst)
1882 return DAG.FoldConstantArithmetic(ISD::MUL, VT, N0.getNode(), N1.getNode());
1884 // canonicalize constant to RHS
1885 if (N0IsConst && !N1IsConst)
1886 return DAG.getNode(ISD::MUL, SDLoc(N), VT, N1, N0);
1887 // fold (mul x, 0) -> 0
1888 if (N1IsConst && ConstValue1 == 0)
1890 // We require a splat of the entire scalar bit width for non-contiguous
1893 ConstValue1.getBitWidth() == VT.getScalarType().getSizeInBits();
1894 // fold (mul x, 1) -> x
1895 if (N1IsConst && ConstValue1 == 1 && IsFullSplat)
1897 // fold (mul x, -1) -> 0-x
1898 if (N1IsConst && ConstValue1.isAllOnesValue())
1899 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1900 DAG.getConstant(0, VT), N0);
1901 // fold (mul x, (1 << c)) -> x << c
1902 if (N1IsConst && ConstValue1.isPowerOf2() && IsFullSplat)
1903 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0,
1904 DAG.getConstant(ConstValue1.logBase2(),
1905 getShiftAmountTy(N0.getValueType())));
1906 // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c
1907 if (N1IsConst && (-ConstValue1).isPowerOf2() && IsFullSplat) {
1908 unsigned Log2Val = (-ConstValue1).logBase2();
1909 // FIXME: If the input is something that is easily negated (e.g. a
1910 // single-use add), we should put the negate there.
1911 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1912 DAG.getConstant(0, VT),
1913 DAG.getNode(ISD::SHL, SDLoc(N), VT, N0,
1914 DAG.getConstant(Log2Val,
1915 getShiftAmountTy(N0.getValueType()))));
1919 // (mul (shl X, c1), c2) -> (mul X, c2 << c1)
1920 if (N1IsConst && N0.getOpcode() == ISD::SHL &&
1921 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
1922 isa<ConstantSDNode>(N0.getOperand(1)))) {
1923 SDValue C3 = DAG.getNode(ISD::SHL, SDLoc(N), VT,
1924 N1, N0.getOperand(1));
1925 AddToWorkList(C3.getNode());
1926 return DAG.getNode(ISD::MUL, SDLoc(N), VT,
1927 N0.getOperand(0), C3);
1930 // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one
1933 SDValue Sh(nullptr,0), Y(nullptr,0);
1934 // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)).
1935 if (N0.getOpcode() == ISD::SHL &&
1936 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
1937 isa<ConstantSDNode>(N0.getOperand(1))) &&
1938 N0.getNode()->hasOneUse()) {
1940 } else if (N1.getOpcode() == ISD::SHL &&
1941 isa<ConstantSDNode>(N1.getOperand(1)) &&
1942 N1.getNode()->hasOneUse()) {
1947 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
1948 Sh.getOperand(0), Y);
1949 return DAG.getNode(ISD::SHL, SDLoc(N), VT,
1950 Mul, Sh.getOperand(1));
1954 // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2)
1955 if (N1IsConst && N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() &&
1956 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
1957 isa<ConstantSDNode>(N0.getOperand(1))))
1958 return DAG.getNode(ISD::ADD, SDLoc(N), VT,
1959 DAG.getNode(ISD::MUL, SDLoc(N0), VT,
1960 N0.getOperand(0), N1),
1961 DAG.getNode(ISD::MUL, SDLoc(N1), VT,
1962 N0.getOperand(1), N1));
1965 SDValue RMUL = ReassociateOps(ISD::MUL, SDLoc(N), N0, N1);
1972 SDValue DAGCombiner::visitSDIV(SDNode *N) {
1973 SDValue N0 = N->getOperand(0);
1974 SDValue N1 = N->getOperand(1);
1975 ConstantSDNode *N0C = isConstOrConstSplat(N0);
1976 ConstantSDNode *N1C = isConstOrConstSplat(N1);
1977 EVT VT = N->getValueType(0);
1980 if (VT.isVector()) {
1981 SDValue FoldedVOp = SimplifyVBinOp(N);
1982 if (FoldedVOp.getNode()) return FoldedVOp;
1985 // fold (sdiv c1, c2) -> c1/c2
1986 if (N0C && N1C && !N1C->isNullValue())
1987 return DAG.FoldConstantArithmetic(ISD::SDIV, VT, N0C, N1C);
1988 // fold (sdiv X, 1) -> X
1989 if (N1C && N1C->getAPIntValue() == 1LL)
1991 // fold (sdiv X, -1) -> 0-X
1992 if (N1C && N1C->isAllOnesValue())
1993 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1994 DAG.getConstant(0, VT), N0);
1995 // If we know the sign bits of both operands are zero, strength reduce to a
1996 // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2
1997 if (!VT.isVector()) {
1998 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
1999 return DAG.getNode(ISD::UDIV, SDLoc(N), N1.getValueType(),
2003 // fold (sdiv X, pow2) -> simple ops after legalize
2004 if (N1C && !N1C->isNullValue() && (N1C->getAPIntValue().isPowerOf2() ||
2005 (-N1C->getAPIntValue()).isPowerOf2())) {
2006 // If dividing by powers of two is cheap, then don't perform the following
2008 if (TLI.isPow2DivCheap())
2011 unsigned lg2 = N1C->getAPIntValue().countTrailingZeros();
2013 // Splat the sign bit into the register
2015 DAG.getNode(ISD::SRA, SDLoc(N), VT, N0,
2016 DAG.getConstant(VT.getScalarSizeInBits() - 1,
2017 getShiftAmountTy(N0.getValueType())));
2018 AddToWorkList(SGN.getNode());
2020 // Add (N0 < 0) ? abs2 - 1 : 0;
2022 DAG.getNode(ISD::SRL, SDLoc(N), VT, SGN,
2023 DAG.getConstant(VT.getScalarSizeInBits() - lg2,
2024 getShiftAmountTy(SGN.getValueType())));
2025 SDValue ADD = DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, SRL);
2026 AddToWorkList(SRL.getNode());
2027 AddToWorkList(ADD.getNode()); // Divide by pow2
2028 SDValue SRA = DAG.getNode(ISD::SRA, SDLoc(N), VT, ADD,
2029 DAG.getConstant(lg2, getShiftAmountTy(ADD.getValueType())));
2031 // If we're dividing by a positive value, we're done. Otherwise, we must
2032 // negate the result.
2033 if (N1C->getAPIntValue().isNonNegative())
2036 AddToWorkList(SRA.getNode());
2037 return DAG.getNode(ISD::SUB, SDLoc(N), VT, DAG.getConstant(0, VT), SRA);
2040 // if integer divide is expensive and we satisfy the requirements, emit an
2041 // alternate sequence.
2042 if (N1C && !TLI.isIntDivCheap()) {
2043 SDValue Op = BuildSDIV(N);
2044 if (Op.getNode()) return Op;
2048 if (N0.getOpcode() == ISD::UNDEF)
2049 return DAG.getConstant(0, VT);
2050 // X / undef -> undef
2051 if (N1.getOpcode() == ISD::UNDEF)
2057 SDValue DAGCombiner::visitUDIV(SDNode *N) {
2058 SDValue N0 = N->getOperand(0);
2059 SDValue N1 = N->getOperand(1);
2060 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2061 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2062 EVT VT = N->getValueType(0);
2065 if (VT.isVector()) {
2066 SDValue FoldedVOp = SimplifyVBinOp(N);
2067 if (FoldedVOp.getNode()) return FoldedVOp;
2070 // fold (udiv c1, c2) -> c1/c2
2071 if (N0C && N1C && !N1C->isNullValue())
2072 return DAG.FoldConstantArithmetic(ISD::UDIV, VT, N0C, N1C);
2073 // fold (udiv x, (1 << c)) -> x >>u c
2074 if (N1C && N1C->getAPIntValue().isPowerOf2())
2075 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0,
2076 DAG.getConstant(N1C->getAPIntValue().logBase2(),
2077 getShiftAmountTy(N0.getValueType())));
2078 // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2
2079 if (N1.getOpcode() == ISD::SHL) {
2080 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
2081 if (SHC->getAPIntValue().isPowerOf2()) {
2082 EVT ADDVT = N1.getOperand(1).getValueType();
2083 SDValue Add = DAG.getNode(ISD::ADD, SDLoc(N), ADDVT,
2085 DAG.getConstant(SHC->getAPIntValue()
2088 AddToWorkList(Add.getNode());
2089 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, Add);
2093 // fold (udiv x, c) -> alternate
2094 if (N1C && !TLI.isIntDivCheap()) {
2095 SDValue Op = BuildUDIV(N);
2096 if (Op.getNode()) return Op;
2100 if (N0.getOpcode() == ISD::UNDEF)
2101 return DAG.getConstant(0, VT);
2102 // X / undef -> undef
2103 if (N1.getOpcode() == ISD::UNDEF)
2109 SDValue DAGCombiner::visitSREM(SDNode *N) {
2110 SDValue N0 = N->getOperand(0);
2111 SDValue N1 = N->getOperand(1);
2112 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2113 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2114 EVT VT = N->getValueType(0);
2116 // fold (srem c1, c2) -> c1%c2
2117 if (N0C && N1C && !N1C->isNullValue())
2118 return DAG.FoldConstantArithmetic(ISD::SREM, VT, N0C, N1C);
2119 // If we know the sign bits of both operands are zero, strength reduce to a
2120 // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15
2121 if (!VT.isVector()) {
2122 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
2123 return DAG.getNode(ISD::UREM, SDLoc(N), VT, N0, N1);
2126 // If X/C can be simplified by the division-by-constant logic, lower
2127 // X%C to the equivalent of X-X/C*C.
2128 if (N1C && !N1C->isNullValue()) {
2129 SDValue Div = DAG.getNode(ISD::SDIV, SDLoc(N), VT, N0, N1);
2130 AddToWorkList(Div.getNode());
2131 SDValue OptimizedDiv = combine(Div.getNode());
2132 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2133 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2135 SDValue Sub = DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, Mul);
2136 AddToWorkList(Mul.getNode());
2142 if (N0.getOpcode() == ISD::UNDEF)
2143 return DAG.getConstant(0, VT);
2144 // X % undef -> undef
2145 if (N1.getOpcode() == ISD::UNDEF)
2151 SDValue DAGCombiner::visitUREM(SDNode *N) {
2152 SDValue N0 = N->getOperand(0);
2153 SDValue N1 = N->getOperand(1);
2154 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2155 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2156 EVT VT = N->getValueType(0);
2158 // fold (urem c1, c2) -> c1%c2
2159 if (N0C && N1C && !N1C->isNullValue())
2160 return DAG.FoldConstantArithmetic(ISD::UREM, VT, N0C, N1C);
2161 // fold (urem x, pow2) -> (and x, pow2-1)
2162 if (N1C && !N1C->isNullValue() && N1C->getAPIntValue().isPowerOf2())
2163 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0,
2164 DAG.getConstant(N1C->getAPIntValue()-1,VT));
2165 // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1))
2166 if (N1.getOpcode() == ISD::SHL) {
2167 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
2168 if (SHC->getAPIntValue().isPowerOf2()) {
2170 DAG.getNode(ISD::ADD, SDLoc(N), VT, N1,
2171 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()),
2173 AddToWorkList(Add.getNode());
2174 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, Add);
2179 // If X/C can be simplified by the division-by-constant logic, lower
2180 // X%C to the equivalent of X-X/C*C.
2181 if (N1C && !N1C->isNullValue()) {
2182 SDValue Div = DAG.getNode(ISD::UDIV, SDLoc(N), VT, N0, N1);
2183 AddToWorkList(Div.getNode());
2184 SDValue OptimizedDiv = combine(Div.getNode());
2185 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2186 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2188 SDValue Sub = DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, Mul);
2189 AddToWorkList(Mul.getNode());
2195 if (N0.getOpcode() == ISD::UNDEF)
2196 return DAG.getConstant(0, VT);
2197 // X % undef -> undef
2198 if (N1.getOpcode() == ISD::UNDEF)
2204 SDValue DAGCombiner::visitMULHS(SDNode *N) {
2205 SDValue N0 = N->getOperand(0);
2206 SDValue N1 = N->getOperand(1);
2207 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2208 EVT VT = N->getValueType(0);
2211 // fold (mulhs x, 0) -> 0
2212 if (N1C && N1C->isNullValue())
2214 // fold (mulhs x, 1) -> (sra x, size(x)-1)
2215 if (N1C && N1C->getAPIntValue() == 1)
2216 return DAG.getNode(ISD::SRA, SDLoc(N), N0.getValueType(), N0,
2217 DAG.getConstant(N0.getValueType().getSizeInBits() - 1,
2218 getShiftAmountTy(N0.getValueType())));
2219 // fold (mulhs x, undef) -> 0
2220 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2221 return DAG.getConstant(0, VT);
2223 // If the type twice as wide is legal, transform the mulhs to a wider multiply
2225 if (VT.isSimple() && !VT.isVector()) {
2226 MVT Simple = VT.getSimpleVT();
2227 unsigned SimpleSize = Simple.getSizeInBits();
2228 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2229 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2230 N0 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N0);
2231 N1 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N1);
2232 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2233 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2234 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2235 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2242 SDValue DAGCombiner::visitMULHU(SDNode *N) {
2243 SDValue N0 = N->getOperand(0);
2244 SDValue N1 = N->getOperand(1);
2245 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2246 EVT VT = N->getValueType(0);
2249 // fold (mulhu x, 0) -> 0
2250 if (N1C && N1C->isNullValue())
2252 // fold (mulhu x, 1) -> 0
2253 if (N1C && N1C->getAPIntValue() == 1)
2254 return DAG.getConstant(0, N0.getValueType());
2255 // fold (mulhu x, undef) -> 0
2256 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2257 return DAG.getConstant(0, VT);
2259 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2261 if (VT.isSimple() && !VT.isVector()) {
2262 MVT Simple = VT.getSimpleVT();
2263 unsigned SimpleSize = Simple.getSizeInBits();
2264 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2265 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2266 N0 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N0);
2267 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N1);
2268 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2269 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2270 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2271 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2278 /// SimplifyNodeWithTwoResults - Perform optimizations common to nodes that
2279 /// compute two values. LoOp and HiOp give the opcodes for the two computations
2280 /// that are being performed. Return true if a simplification was made.
2282 SDValue DAGCombiner::SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
2284 // If the high half is not needed, just compute the low half.
2285 bool HiExists = N->hasAnyUseOfValue(1);
2287 (!LegalOperations ||
2288 TLI.isOperationLegalOrCustom(LoOp, N->getValueType(0)))) {
2289 SDValue Res = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0),
2290 ArrayRef<SDUse>(N->op_begin(), N->op_end()));
2291 return CombineTo(N, Res, Res);
2294 // If the low half is not needed, just compute the high half.
2295 bool LoExists = N->hasAnyUseOfValue(0);
2297 (!LegalOperations ||
2298 TLI.isOperationLegal(HiOp, N->getValueType(1)))) {
2299 SDValue Res = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1),
2300 ArrayRef<SDUse>(N->op_begin(), N->op_end()));
2301 return CombineTo(N, Res, Res);
2304 // If both halves are used, return as it is.
2305 if (LoExists && HiExists)
2308 // If the two computed results can be simplified separately, separate them.
2310 SDValue Lo = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0),
2311 ArrayRef<SDUse>(N->op_begin(), N->op_end()));
2312 AddToWorkList(Lo.getNode());
2313 SDValue LoOpt = combine(Lo.getNode());
2314 if (LoOpt.getNode() && LoOpt.getNode() != Lo.getNode() &&
2315 (!LegalOperations ||
2316 TLI.isOperationLegal(LoOpt.getOpcode(), LoOpt.getValueType())))
2317 return CombineTo(N, LoOpt, LoOpt);
2321 SDValue Hi = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1),
2322 ArrayRef<SDUse>(N->op_begin(), N->op_end()));
2323 AddToWorkList(Hi.getNode());
2324 SDValue HiOpt = combine(Hi.getNode());
2325 if (HiOpt.getNode() && HiOpt != Hi &&
2326 (!LegalOperations ||
2327 TLI.isOperationLegal(HiOpt.getOpcode(), HiOpt.getValueType())))
2328 return CombineTo(N, HiOpt, HiOpt);
2334 SDValue DAGCombiner::visitSMUL_LOHI(SDNode *N) {
2335 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHS);
2336 if (Res.getNode()) return Res;
2338 EVT VT = N->getValueType(0);
2341 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2343 if (VT.isSimple() && !VT.isVector()) {
2344 MVT Simple = VT.getSimpleVT();
2345 unsigned SimpleSize = Simple.getSizeInBits();
2346 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2347 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2348 SDValue Lo = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(0));
2349 SDValue Hi = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(1));
2350 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2351 // Compute the high part as N1.
2352 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2353 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2354 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2355 // Compute the low part as N0.
2356 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2357 return CombineTo(N, Lo, Hi);
2364 SDValue DAGCombiner::visitUMUL_LOHI(SDNode *N) {
2365 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHU);
2366 if (Res.getNode()) return Res;
2368 EVT VT = N->getValueType(0);
2371 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2373 if (VT.isSimple() && !VT.isVector()) {
2374 MVT Simple = VT.getSimpleVT();
2375 unsigned SimpleSize = Simple.getSizeInBits();
2376 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2377 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2378 SDValue Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(0));
2379 SDValue Hi = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(1));
2380 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2381 // Compute the high part as N1.
2382 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2383 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2384 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2385 // Compute the low part as N0.
2386 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2387 return CombineTo(N, Lo, Hi);
2394 SDValue DAGCombiner::visitSMULO(SDNode *N) {
2395 // (smulo x, 2) -> (saddo x, x)
2396 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2397 if (C2->getAPIntValue() == 2)
2398 return DAG.getNode(ISD::SADDO, SDLoc(N), N->getVTList(),
2399 N->getOperand(0), N->getOperand(0));
2404 SDValue DAGCombiner::visitUMULO(SDNode *N) {
2405 // (umulo x, 2) -> (uaddo x, x)
2406 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2407 if (C2->getAPIntValue() == 2)
2408 return DAG.getNode(ISD::UADDO, SDLoc(N), N->getVTList(),
2409 N->getOperand(0), N->getOperand(0));
2414 SDValue DAGCombiner::visitSDIVREM(SDNode *N) {
2415 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::SDIV, ISD::SREM);
2416 if (Res.getNode()) return Res;
2421 SDValue DAGCombiner::visitUDIVREM(SDNode *N) {
2422 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::UDIV, ISD::UREM);
2423 if (Res.getNode()) return Res;
2428 /// SimplifyBinOpWithSameOpcodeHands - If this is a binary operator with
2429 /// two operands of the same opcode, try to simplify it.
2430 SDValue DAGCombiner::SimplifyBinOpWithSameOpcodeHands(SDNode *N) {
2431 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
2432 EVT VT = N0.getValueType();
2433 assert(N0.getOpcode() == N1.getOpcode() && "Bad input!");
2435 // Bail early if none of these transforms apply.
2436 if (N0.getNode()->getNumOperands() == 0) return SDValue();
2438 // For each of OP in AND/OR/XOR:
2439 // fold (OP (zext x), (zext y)) -> (zext (OP x, y))
2440 // fold (OP (sext x), (sext y)) -> (sext (OP x, y))
2441 // fold (OP (aext x), (aext y)) -> (aext (OP x, y))
2442 // fold (OP (trunc x), (trunc y)) -> (trunc (OP x, y)) (if trunc isn't free)
2444 // do not sink logical op inside of a vector extend, since it may combine
2446 EVT Op0VT = N0.getOperand(0).getValueType();
2447 if ((N0.getOpcode() == ISD::ZERO_EXTEND ||
2448 N0.getOpcode() == ISD::SIGN_EXTEND ||
2449 // Avoid infinite looping with PromoteIntBinOp.
2450 (N0.getOpcode() == ISD::ANY_EXTEND &&
2451 (!LegalTypes || TLI.isTypeDesirableForOp(N->getOpcode(), Op0VT))) ||
2452 (N0.getOpcode() == ISD::TRUNCATE &&
2453 (!TLI.isZExtFree(VT, Op0VT) ||
2454 !TLI.isTruncateFree(Op0VT, VT)) &&
2455 TLI.isTypeLegal(Op0VT))) &&
2457 Op0VT == N1.getOperand(0).getValueType() &&
2458 (!LegalOperations || TLI.isOperationLegal(N->getOpcode(), Op0VT))) {
2459 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0),
2460 N0.getOperand(0).getValueType(),
2461 N0.getOperand(0), N1.getOperand(0));
2462 AddToWorkList(ORNode.getNode());
2463 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, ORNode);
2466 // For each of OP in SHL/SRL/SRA/AND...
2467 // fold (and (OP x, z), (OP y, z)) -> (OP (and x, y), z)
2468 // fold (or (OP x, z), (OP y, z)) -> (OP (or x, y), z)
2469 // fold (xor (OP x, z), (OP y, z)) -> (OP (xor x, y), z)
2470 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL ||
2471 N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::AND) &&
2472 N0.getOperand(1) == N1.getOperand(1)) {
2473 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0),
2474 N0.getOperand(0).getValueType(),
2475 N0.getOperand(0), N1.getOperand(0));
2476 AddToWorkList(ORNode.getNode());
2477 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
2478 ORNode, N0.getOperand(1));
2481 // Simplify xor/and/or (bitcast(A), bitcast(B)) -> bitcast(op (A,B))
2482 // Only perform this optimization after type legalization and before
2483 // LegalizeVectorOprs. LegalizeVectorOprs promotes vector operations by
2484 // adding bitcasts. For example (xor v4i32) is promoted to (v2i64), and
2485 // we don't want to undo this promotion.
2486 // We also handle SCALAR_TO_VECTOR because xor/or/and operations are cheaper
2488 if ((N0.getOpcode() == ISD::BITCAST ||
2489 N0.getOpcode() == ISD::SCALAR_TO_VECTOR) &&
2490 Level == AfterLegalizeTypes) {
2491 SDValue In0 = N0.getOperand(0);
2492 SDValue In1 = N1.getOperand(0);
2493 EVT In0Ty = In0.getValueType();
2494 EVT In1Ty = In1.getValueType();
2496 // If both incoming values are integers, and the original types are the
2498 if (In0Ty.isInteger() && In1Ty.isInteger() && In0Ty == In1Ty) {
2499 SDValue Op = DAG.getNode(N->getOpcode(), DL, In0Ty, In0, In1);
2500 SDValue BC = DAG.getNode(N0.getOpcode(), DL, VT, Op);
2501 AddToWorkList(Op.getNode());
2506 // Xor/and/or are indifferent to the swizzle operation (shuffle of one value).
2507 // Simplify xor/and/or (shuff(A), shuff(B)) -> shuff(op (A,B))
2508 // If both shuffles use the same mask, and both shuffle within a single
2509 // vector, then it is worthwhile to move the swizzle after the operation.
2510 // The type-legalizer generates this pattern when loading illegal
2511 // vector types from memory. In many cases this allows additional shuffle
2513 // There are other cases where moving the shuffle after the xor/and/or
2514 // is profitable even if shuffles don't perform a swizzle.
2515 // If both shuffles use the same mask, and both shuffles have the same first
2516 // or second operand, then it might still be profitable to move the shuffle
2517 // after the xor/and/or operation.
2518 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG) {
2519 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(N0);
2520 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(N1);
2522 assert(N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType() &&
2523 "Inputs to shuffles are not the same type");
2525 // Check that both shuffles use the same mask. The masks are known to be of
2526 // the same length because the result vector type is the same.
2527 // Check also that shuffles have only one use to avoid introducing extra
2529 if (SVN0->hasOneUse() && SVN1->hasOneUse() &&
2530 SVN0->getMask().equals(SVN1->getMask())) {
2531 SDValue ShOp = N0->getOperand(1);
2533 // Don't try to fold this node if it requires introducing a
2534 // build vector of all zeros that might be illegal at this stage.
2535 if (N->getOpcode() == ISD::XOR && ShOp.getOpcode() != ISD::UNDEF) {
2537 ShOp = DAG.getConstant(0, VT);
2542 // (AND (shuf (A, C), shuf (B, C)) -> shuf (AND (A, B), C)
2543 // (OR (shuf (A, C), shuf (B, C)) -> shuf (OR (A, B), C)
2544 // (XOR (shuf (A, C), shuf (B, C)) -> shuf (XOR (A, B), V_0)
2545 if (N0.getOperand(1) == N1.getOperand(1) && ShOp.getNode()) {
2546 SDValue NewNode = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
2547 N0->getOperand(0), N1->getOperand(0));
2548 AddToWorkList(NewNode.getNode());
2549 return DAG.getVectorShuffle(VT, SDLoc(N), NewNode, ShOp,
2550 &SVN0->getMask()[0]);
2553 // Don't try to fold this node if it requires introducing a
2554 // build vector of all zeros that might be illegal at this stage.
2555 ShOp = N0->getOperand(0);
2556 if (N->getOpcode() == ISD::XOR && ShOp.getOpcode() != ISD::UNDEF) {
2558 ShOp = DAG.getConstant(0, VT);
2563 // (AND (shuf (C, A), shuf (C, B)) -> shuf (C, AND (A, B))
2564 // (OR (shuf (C, A), shuf (C, B)) -> shuf (C, OR (A, B))
2565 // (XOR (shuf (C, A), shuf (C, B)) -> shuf (V_0, XOR (A, B))
2566 if (N0->getOperand(0) == N1->getOperand(0) && ShOp.getNode()) {
2567 SDValue NewNode = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
2568 N0->getOperand(1), N1->getOperand(1));
2569 AddToWorkList(NewNode.getNode());
2570 return DAG.getVectorShuffle(VT, SDLoc(N), ShOp, NewNode,
2571 &SVN0->getMask()[0]);
2579 SDValue DAGCombiner::visitAND(SDNode *N) {
2580 SDValue N0 = N->getOperand(0);
2581 SDValue N1 = N->getOperand(1);
2582 SDValue LL, LR, RL, RR, CC0, CC1;
2583 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2584 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2585 EVT VT = N1.getValueType();
2586 unsigned BitWidth = VT.getScalarType().getSizeInBits();
2589 if (VT.isVector()) {
2590 SDValue FoldedVOp = SimplifyVBinOp(N);
2591 if (FoldedVOp.getNode()) return FoldedVOp;
2593 // fold (and x, 0) -> 0, vector edition
2594 if (ISD::isBuildVectorAllZeros(N0.getNode()))
2596 if (ISD::isBuildVectorAllZeros(N1.getNode()))
2599 // fold (and x, -1) -> x, vector edition
2600 if (ISD::isBuildVectorAllOnes(N0.getNode()))
2602 if (ISD::isBuildVectorAllOnes(N1.getNode()))
2606 // fold (and x, undef) -> 0
2607 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2608 return DAG.getConstant(0, VT);
2609 // fold (and c1, c2) -> c1&c2
2611 return DAG.FoldConstantArithmetic(ISD::AND, VT, N0C, N1C);
2612 // canonicalize constant to RHS
2614 return DAG.getNode(ISD::AND, SDLoc(N), VT, N1, N0);
2615 // fold (and x, -1) -> x
2616 if (N1C && N1C->isAllOnesValue())
2618 // if (and x, c) is known to be zero, return 0
2619 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
2620 APInt::getAllOnesValue(BitWidth)))
2621 return DAG.getConstant(0, VT);
2623 SDValue RAND = ReassociateOps(ISD::AND, SDLoc(N), N0, N1);
2626 // fold (and (or x, C), D) -> D if (C & D) == D
2627 if (N1C && N0.getOpcode() == ISD::OR)
2628 if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
2629 if ((ORI->getAPIntValue() & N1C->getAPIntValue()) == N1C->getAPIntValue())
2631 // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits.
2632 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
2633 SDValue N0Op0 = N0.getOperand(0);
2634 APInt Mask = ~N1C->getAPIntValue();
2635 Mask = Mask.trunc(N0Op0.getValueSizeInBits());
2636 if (DAG.MaskedValueIsZero(N0Op0, Mask)) {
2637 SDValue Zext = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N),
2638 N0.getValueType(), N0Op0);
2640 // Replace uses of the AND with uses of the Zero extend node.
2643 // We actually want to replace all uses of the any_extend with the
2644 // zero_extend, to avoid duplicating things. This will later cause this
2645 // AND to be folded.
2646 CombineTo(N0.getNode(), Zext);
2647 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2650 // similarly fold (and (X (load ([non_ext|any_ext|zero_ext] V))), c) ->
2651 // (X (load ([non_ext|zero_ext] V))) if 'and' only clears top bits which must
2652 // already be zero by virtue of the width of the base type of the load.
2654 // the 'X' node here can either be nothing or an extract_vector_elt to catch
2656 if ((N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
2657 N0.getOperand(0).getOpcode() == ISD::LOAD) ||
2658 N0.getOpcode() == ISD::LOAD) {
2659 LoadSDNode *Load = cast<LoadSDNode>( (N0.getOpcode() == ISD::LOAD) ?
2660 N0 : N0.getOperand(0) );
2662 // Get the constant (if applicable) the zero'th operand is being ANDed with.
2663 // This can be a pure constant or a vector splat, in which case we treat the
2664 // vector as a scalar and use the splat value.
2665 APInt Constant = APInt::getNullValue(1);
2666 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
2667 Constant = C->getAPIntValue();
2668 } else if (BuildVectorSDNode *Vector = dyn_cast<BuildVectorSDNode>(N1)) {
2669 APInt SplatValue, SplatUndef;
2670 unsigned SplatBitSize;
2672 bool IsSplat = Vector->isConstantSplat(SplatValue, SplatUndef,
2673 SplatBitSize, HasAnyUndefs);
2675 // Undef bits can contribute to a possible optimisation if set, so
2677 SplatValue |= SplatUndef;
2679 // The splat value may be something like "0x00FFFFFF", which means 0 for
2680 // the first vector value and FF for the rest, repeating. We need a mask
2681 // that will apply equally to all members of the vector, so AND all the
2682 // lanes of the constant together.
2683 EVT VT = Vector->getValueType(0);
2684 unsigned BitWidth = VT.getVectorElementType().getSizeInBits();
2686 // If the splat value has been compressed to a bitlength lower
2687 // than the size of the vector lane, we need to re-expand it to
2689 if (BitWidth > SplatBitSize)
2690 for (SplatValue = SplatValue.zextOrTrunc(BitWidth);
2691 SplatBitSize < BitWidth;
2692 SplatBitSize = SplatBitSize * 2)
2693 SplatValue |= SplatValue.shl(SplatBitSize);
2695 Constant = APInt::getAllOnesValue(BitWidth);
2696 for (unsigned i = 0, n = SplatBitSize/BitWidth; i < n; ++i)
2697 Constant &= SplatValue.lshr(i*BitWidth).zextOrTrunc(BitWidth);
2701 // If we want to change an EXTLOAD to a ZEXTLOAD, ensure a ZEXTLOAD is
2702 // actually legal and isn't going to get expanded, else this is a false
2704 bool CanZextLoadProfitably = TLI.isLoadExtLegal(ISD::ZEXTLOAD,
2705 Load->getMemoryVT());
2707 // Resize the constant to the same size as the original memory access before
2708 // extension. If it is still the AllOnesValue then this AND is completely
2711 Constant.zextOrTrunc(Load->getMemoryVT().getScalarType().getSizeInBits());
2714 switch (Load->getExtensionType()) {
2715 default: B = false; break;
2716 case ISD::EXTLOAD: B = CanZextLoadProfitably; break;
2718 case ISD::NON_EXTLOAD: B = true; break;
2721 if (B && Constant.isAllOnesValue()) {
2722 // If the load type was an EXTLOAD, convert to ZEXTLOAD in order to
2723 // preserve semantics once we get rid of the AND.
2724 SDValue NewLoad(Load, 0);
2725 if (Load->getExtensionType() == ISD::EXTLOAD) {
2726 NewLoad = DAG.getLoad(Load->getAddressingMode(), ISD::ZEXTLOAD,
2727 Load->getValueType(0), SDLoc(Load),
2728 Load->getChain(), Load->getBasePtr(),
2729 Load->getOffset(), Load->getMemoryVT(),
2730 Load->getMemOperand());
2731 // Replace uses of the EXTLOAD with the new ZEXTLOAD.
2732 if (Load->getNumValues() == 3) {
2733 // PRE/POST_INC loads have 3 values.
2734 SDValue To[] = { NewLoad.getValue(0), NewLoad.getValue(1),
2735 NewLoad.getValue(2) };
2736 CombineTo(Load, To, 3, true);
2738 CombineTo(Load, NewLoad.getValue(0), NewLoad.getValue(1));
2742 // Fold the AND away, taking care not to fold to the old load node if we
2744 CombineTo(N, (N0.getNode() == Load) ? NewLoad : N0);
2746 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2749 // fold (and (setcc x), (setcc y)) -> (setcc (and x, y))
2750 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
2751 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
2752 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
2754 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
2755 LL.getValueType().isInteger()) {
2756 // fold (and (seteq X, 0), (seteq Y, 0)) -> (seteq (or X, Y), 0)
2757 if (cast<ConstantSDNode>(LR)->isNullValue() && Op1 == ISD::SETEQ) {
2758 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(N0),
2759 LR.getValueType(), LL, RL);
2760 AddToWorkList(ORNode.getNode());
2761 return DAG.getSetCC(SDLoc(N), VT, ORNode, LR, Op1);
2763 // fold (and (seteq X, -1), (seteq Y, -1)) -> (seteq (and X, Y), -1)
2764 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETEQ) {
2765 SDValue ANDNode = DAG.getNode(ISD::AND, SDLoc(N0),
2766 LR.getValueType(), LL, RL);
2767 AddToWorkList(ANDNode.getNode());
2768 return DAG.getSetCC(SDLoc(N), VT, ANDNode, LR, Op1);
2770 // fold (and (setgt X, -1), (setgt Y, -1)) -> (setgt (or X, Y), -1)
2771 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETGT) {
2772 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(N0),
2773 LR.getValueType(), LL, RL);
2774 AddToWorkList(ORNode.getNode());
2775 return DAG.getSetCC(SDLoc(N), VT, ORNode, LR, Op1);
2778 // Simplify (and (setne X, 0), (setne X, -1)) -> (setuge (add X, 1), 2)
2779 if (LL == RL && isa<ConstantSDNode>(LR) && isa<ConstantSDNode>(RR) &&
2780 Op0 == Op1 && LL.getValueType().isInteger() &&
2781 Op0 == ISD::SETNE && ((cast<ConstantSDNode>(LR)->isNullValue() &&
2782 cast<ConstantSDNode>(RR)->isAllOnesValue()) ||
2783 (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
2784 cast<ConstantSDNode>(RR)->isNullValue()))) {
2785 SDValue ADDNode = DAG.getNode(ISD::ADD, SDLoc(N0), LL.getValueType(),
2786 LL, DAG.getConstant(1, LL.getValueType()));
2787 AddToWorkList(ADDNode.getNode());
2788 return DAG.getSetCC(SDLoc(N), VT, ADDNode,
2789 DAG.getConstant(2, LL.getValueType()), ISD::SETUGE);
2791 // canonicalize equivalent to ll == rl
2792 if (LL == RR && LR == RL) {
2793 Op1 = ISD::getSetCCSwappedOperands(Op1);
2796 if (LL == RL && LR == RR) {
2797 bool isInteger = LL.getValueType().isInteger();
2798 ISD::CondCode Result = ISD::getSetCCAndOperation(Op0, Op1, isInteger);
2799 if (Result != ISD::SETCC_INVALID &&
2800 (!LegalOperations ||
2801 (TLI.isCondCodeLegal(Result, LL.getSimpleValueType()) &&
2802 TLI.isOperationLegal(ISD::SETCC,
2803 getSetCCResultType(N0.getSimpleValueType())))))
2804 return DAG.getSetCC(SDLoc(N), N0.getValueType(),
2809 // Simplify: (and (op x...), (op y...)) -> (op (and x, y))
2810 if (N0.getOpcode() == N1.getOpcode()) {
2811 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
2812 if (Tmp.getNode()) return Tmp;
2815 // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1)
2816 // fold (and (sra)) -> (and (srl)) when possible.
2817 if (!VT.isVector() &&
2818 SimplifyDemandedBits(SDValue(N, 0)))
2819 return SDValue(N, 0);
2821 // fold (zext_inreg (extload x)) -> (zextload x)
2822 if (ISD::isEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode())) {
2823 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2824 EVT MemVT = LN0->getMemoryVT();
2825 // If we zero all the possible extended bits, then we can turn this into
2826 // a zextload if we are running before legalize or the operation is legal.
2827 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
2828 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
2829 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
2830 ((!LegalOperations && !LN0->isVolatile()) ||
2831 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
2832 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT,
2833 LN0->getChain(), LN0->getBasePtr(),
2834 MemVT, LN0->getMemOperand());
2836 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
2837 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2840 // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use
2841 if (ISD::isSEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
2843 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2844 EVT MemVT = LN0->getMemoryVT();
2845 // If we zero all the possible extended bits, then we can turn this into
2846 // a zextload if we are running before legalize or the operation is legal.
2847 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
2848 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
2849 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
2850 ((!LegalOperations && !LN0->isVolatile()) ||
2851 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
2852 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT,
2853 LN0->getChain(), LN0->getBasePtr(),
2854 MemVT, LN0->getMemOperand());
2856 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
2857 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2861 // fold (and (load x), 255) -> (zextload x, i8)
2862 // fold (and (extload x, i16), 255) -> (zextload x, i8)
2863 // fold (and (any_ext (extload x, i16)), 255) -> (zextload x, i8)
2864 if (N1C && (N0.getOpcode() == ISD::LOAD ||
2865 (N0.getOpcode() == ISD::ANY_EXTEND &&
2866 N0.getOperand(0).getOpcode() == ISD::LOAD))) {
2867 bool HasAnyExt = N0.getOpcode() == ISD::ANY_EXTEND;
2868 LoadSDNode *LN0 = HasAnyExt
2869 ? cast<LoadSDNode>(N0.getOperand(0))
2870 : cast<LoadSDNode>(N0);
2871 if (LN0->getExtensionType() != ISD::SEXTLOAD &&
2872 LN0->isUnindexed() && N0.hasOneUse() && SDValue(LN0, 0).hasOneUse()) {
2873 uint32_t ActiveBits = N1C->getAPIntValue().getActiveBits();
2874 if (ActiveBits > 0 && APIntOps::isMask(ActiveBits, N1C->getAPIntValue())){
2875 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits);
2876 EVT LoadedVT = LN0->getMemoryVT();
2878 if (ExtVT == LoadedVT &&
2879 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) {
2880 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
2883 DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy,
2884 LN0->getChain(), LN0->getBasePtr(), ExtVT,
2885 LN0->getMemOperand());
2887 CombineTo(LN0, NewLoad, NewLoad.getValue(1));
2888 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2891 // Do not change the width of a volatile load.
2892 // Do not generate loads of non-round integer types since these can
2893 // be expensive (and would be wrong if the type is not byte sized).
2894 if (!LN0->isVolatile() && LoadedVT.bitsGT(ExtVT) && ExtVT.isRound() &&
2895 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) {
2896 EVT PtrType = LN0->getOperand(1).getValueType();
2898 unsigned Alignment = LN0->getAlignment();
2899 SDValue NewPtr = LN0->getBasePtr();
2901 // For big endian targets, we need to add an offset to the pointer
2902 // to load the correct bytes. For little endian systems, we merely
2903 // need to read fewer bytes from the same pointer.
2904 if (TLI.isBigEndian()) {
2905 unsigned LVTStoreBytes = LoadedVT.getStoreSize();
2906 unsigned EVTStoreBytes = ExtVT.getStoreSize();
2907 unsigned PtrOff = LVTStoreBytes - EVTStoreBytes;
2908 NewPtr = DAG.getNode(ISD::ADD, SDLoc(LN0), PtrType,
2909 NewPtr, DAG.getConstant(PtrOff, PtrType));
2910 Alignment = MinAlign(Alignment, PtrOff);
2913 AddToWorkList(NewPtr.getNode());
2915 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
2917 DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy,
2918 LN0->getChain(), NewPtr,
2919 LN0->getPointerInfo(),
2920 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
2921 Alignment, LN0->getTBAAInfo());
2923 CombineTo(LN0, Load, Load.getValue(1));
2924 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2930 if (N0.getOpcode() == ISD::ADD && N1.getOpcode() == ISD::SRL &&
2931 VT.getSizeInBits() <= 64) {
2932 if (ConstantSDNode *ADDI = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
2933 APInt ADDC = ADDI->getAPIntValue();
2934 if (!TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
2935 // Look for (and (add x, c1), (lshr y, c2)). If C1 wasn't a legal
2936 // immediate for an add, but it is legal if its top c2 bits are set,
2937 // transform the ADD so the immediate doesn't need to be materialized
2939 if (ConstantSDNode *SRLI = dyn_cast<ConstantSDNode>(N1.getOperand(1))) {
2940 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(),
2941 SRLI->getZExtValue());
2942 if (DAG.MaskedValueIsZero(N0.getOperand(1), Mask)) {
2944 if (TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
2946 DAG.getNode(ISD::ADD, SDLoc(N0), VT,
2947 N0.getOperand(0), DAG.getConstant(ADDC, VT));
2948 CombineTo(N0.getNode(), NewAdd);
2949 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2957 // fold (and (or (srl N, 8), (shl N, 8)), 0xffff) -> (srl (bswap N), const)
2958 if (N1C && N1C->getAPIntValue() == 0xffff && N0.getOpcode() == ISD::OR) {
2959 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
2960 N0.getOperand(1), false);
2961 if (BSwap.getNode())
2968 /// MatchBSwapHWord - Match (a >> 8) | (a << 8) as (bswap a) >> 16
2970 SDValue DAGCombiner::MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
2971 bool DemandHighBits) {
2972 if (!LegalOperations)
2975 EVT VT = N->getValueType(0);
2976 if (VT != MVT::i64 && VT != MVT::i32 && VT != MVT::i16)
2978 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
2981 // Recognize (and (shl a, 8), 0xff), (and (srl a, 8), 0xff00)
2982 bool LookPassAnd0 = false;
2983 bool LookPassAnd1 = false;
2984 if (N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode() == ISD::SRL)
2986 if (N1.getOpcode() == ISD::AND && N1.getOperand(0).getOpcode() == ISD::SHL)
2988 if (N0.getOpcode() == ISD::AND) {
2989 if (!N0.getNode()->hasOneUse())
2991 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
2992 if (!N01C || N01C->getZExtValue() != 0xFF00)
2994 N0 = N0.getOperand(0);
2995 LookPassAnd0 = true;
2998 if (N1.getOpcode() == ISD::AND) {
2999 if (!N1.getNode()->hasOneUse())
3001 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
3002 if (!N11C || N11C->getZExtValue() != 0xFF)
3004 N1 = N1.getOperand(0);
3005 LookPassAnd1 = true;
3008 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
3010 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
3012 if (!N0.getNode()->hasOneUse() ||
3013 !N1.getNode()->hasOneUse())
3016 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3017 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
3020 if (N01C->getZExtValue() != 8 || N11C->getZExtValue() != 8)
3023 // Look for (shl (and a, 0xff), 8), (srl (and a, 0xff00), 8)
3024 SDValue N00 = N0->getOperand(0);
3025 if (!LookPassAnd0 && N00.getOpcode() == ISD::AND) {
3026 if (!N00.getNode()->hasOneUse())
3028 ConstantSDNode *N001C = dyn_cast<ConstantSDNode>(N00.getOperand(1));
3029 if (!N001C || N001C->getZExtValue() != 0xFF)
3031 N00 = N00.getOperand(0);
3032 LookPassAnd0 = true;
3035 SDValue N10 = N1->getOperand(0);
3036 if (!LookPassAnd1 && N10.getOpcode() == ISD::AND) {
3037 if (!N10.getNode()->hasOneUse())
3039 ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N10.getOperand(1));
3040 if (!N101C || N101C->getZExtValue() != 0xFF00)
3042 N10 = N10.getOperand(0);
3043 LookPassAnd1 = true;
3049 // Make sure everything beyond the low halfword gets set to zero since the SRL
3050 // 16 will clear the top bits.
3051 unsigned OpSizeInBits = VT.getSizeInBits();
3052 if (DemandHighBits && OpSizeInBits > 16) {
3053 // If the left-shift isn't masked out then the only way this is a bswap is
3054 // if all bits beyond the low 8 are 0. In that case the entire pattern
3055 // reduces to a left shift anyway: leave it for other parts of the combiner.
3059 // However, if the right shift isn't masked out then it might be because
3060 // it's not needed. See if we can spot that too.
3061 if (!LookPassAnd1 &&
3062 !DAG.MaskedValueIsZero(
3063 N10, APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - 16)))
3067 SDValue Res = DAG.getNode(ISD::BSWAP, SDLoc(N), VT, N00);
3068 if (OpSizeInBits > 16)
3069 Res = DAG.getNode(ISD::SRL, SDLoc(N), VT, Res,
3070 DAG.getConstant(OpSizeInBits-16, getShiftAmountTy(VT)));
3074 /// isBSwapHWordElement - Return true if the specified node is an element
3075 /// that makes up a 32-bit packed halfword byteswap. i.e.
3076 /// ((x&0xff)<<8)|((x&0xff00)>>8)|((x&0x00ff0000)<<8)|((x&0xff000000)>>8)
3077 static bool isBSwapHWordElement(SDValue N, SmallVectorImpl<SDNode *> &Parts) {
3078 if (!N.getNode()->hasOneUse())
3081 unsigned Opc = N.getOpcode();
3082 if (Opc != ISD::AND && Opc != ISD::SHL && Opc != ISD::SRL)
3085 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3090 switch (N1C->getZExtValue()) {
3093 case 0xFF: Num = 0; break;
3094 case 0xFF00: Num = 1; break;
3095 case 0xFF0000: Num = 2; break;
3096 case 0xFF000000: Num = 3; break;
3099 // Look for (x & 0xff) << 8 as well as ((x << 8) & 0xff00).
3100 SDValue N0 = N.getOperand(0);
3101 if (Opc == ISD::AND) {
3102 if (Num == 0 || Num == 2) {
3104 // (x >> 8) & 0xff0000
3105 if (N0.getOpcode() != ISD::SRL)
3107 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3108 if (!C || C->getZExtValue() != 8)
3111 // (x << 8) & 0xff00
3112 // (x << 8) & 0xff000000
3113 if (N0.getOpcode() != ISD::SHL)
3115 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3116 if (!C || C->getZExtValue() != 8)
3119 } else if (Opc == ISD::SHL) {
3121 // (x & 0xff0000) << 8
3122 if (Num != 0 && Num != 2)
3124 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3125 if (!C || C->getZExtValue() != 8)
3127 } else { // Opc == ISD::SRL
3128 // (x & 0xff00) >> 8
3129 // (x & 0xff000000) >> 8
3130 if (Num != 1 && Num != 3)
3132 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3133 if (!C || C->getZExtValue() != 8)
3140 Parts[Num] = N0.getOperand(0).getNode();
3144 /// MatchBSwapHWord - Match a 32-bit packed halfword bswap. That is
3145 /// ((x&0xff)<<8)|((x&0xff00)>>8)|((x&0x00ff0000)<<8)|((x&0xff000000)>>8)
3146 /// => (rotl (bswap x), 16)
3147 SDValue DAGCombiner::MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1) {
3148 if (!LegalOperations)
3151 EVT VT = N->getValueType(0);
3154 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
3157 SmallVector<SDNode*,4> Parts(4, (SDNode*)nullptr);
3159 // (or (or (and), (and)), (or (and), (and)))
3160 // (or (or (or (and), (and)), (and)), (and))
3161 if (N0.getOpcode() != ISD::OR)
3163 SDValue N00 = N0.getOperand(0);
3164 SDValue N01 = N0.getOperand(1);
3166 if (N1.getOpcode() == ISD::OR &&
3167 N00.getNumOperands() == 2 && N01.getNumOperands() == 2) {
3168 // (or (or (and), (and)), (or (and), (and)))
3169 SDValue N000 = N00.getOperand(0);
3170 if (!isBSwapHWordElement(N000, Parts))
3173 SDValue N001 = N00.getOperand(1);
3174 if (!isBSwapHWordElement(N001, Parts))
3176 SDValue N010 = N01.getOperand(0);
3177 if (!isBSwapHWordElement(N010, Parts))
3179 SDValue N011 = N01.getOperand(1);
3180 if (!isBSwapHWordElement(N011, Parts))
3183 // (or (or (or (and), (and)), (and)), (and))
3184 if (!isBSwapHWordElement(N1, Parts))
3186 if (!isBSwapHWordElement(N01, Parts))
3188 if (N00.getOpcode() != ISD::OR)
3190 SDValue N000 = N00.getOperand(0);
3191 if (!isBSwapHWordElement(N000, Parts))
3193 SDValue N001 = N00.getOperand(1);
3194 if (!isBSwapHWordElement(N001, Parts))
3198 // Make sure the parts are all coming from the same node.
3199 if (Parts[0] != Parts[1] || Parts[0] != Parts[2] || Parts[0] != Parts[3])
3202 SDValue BSwap = DAG.getNode(ISD::BSWAP, SDLoc(N), VT,
3203 SDValue(Parts[0],0));
3205 // Result of the bswap should be rotated by 16. If it's not legal, then
3206 // do (x << 16) | (x >> 16).
3207 SDValue ShAmt = DAG.getConstant(16, getShiftAmountTy(VT));
3208 if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT))
3209 return DAG.getNode(ISD::ROTL, SDLoc(N), VT, BSwap, ShAmt);
3210 if (TLI.isOperationLegalOrCustom(ISD::ROTR, VT))
3211 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, BSwap, ShAmt);
3212 return DAG.getNode(ISD::OR, SDLoc(N), VT,
3213 DAG.getNode(ISD::SHL, SDLoc(N), VT, BSwap, ShAmt),
3214 DAG.getNode(ISD::SRL, SDLoc(N), VT, BSwap, ShAmt));
3217 SDValue DAGCombiner::visitOR(SDNode *N) {
3218 SDValue N0 = N->getOperand(0);
3219 SDValue N1 = N->getOperand(1);
3220 SDValue LL, LR, RL, RR, CC0, CC1;
3221 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3222 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3223 EVT VT = N1.getValueType();
3226 if (VT.isVector()) {
3227 SDValue FoldedVOp = SimplifyVBinOp(N);
3228 if (FoldedVOp.getNode()) return FoldedVOp;
3230 // fold (or x, 0) -> x, vector edition
3231 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3233 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3236 // fold (or x, -1) -> -1, vector edition
3237 if (ISD::isBuildVectorAllOnes(N0.getNode()))
3239 if (ISD::isBuildVectorAllOnes(N1.getNode()))
3242 // fold (or (shuf A, V_0, MA), (shuf B, V_0, MB)) -> (shuf A, B, Mask1)
3243 // fold (or (shuf A, V_0, MA), (shuf B, V_0, MB)) -> (shuf B, A, Mask2)
3244 // Do this only if the resulting shuffle is legal.
3245 if (isa<ShuffleVectorSDNode>(N0) &&
3246 isa<ShuffleVectorSDNode>(N1) &&
3247 N0->getOperand(1) == N1->getOperand(1) &&
3248 ISD::isBuildVectorAllZeros(N0.getOperand(1).getNode())) {
3249 bool CanFold = true;
3250 unsigned NumElts = VT.getVectorNumElements();
3251 const ShuffleVectorSDNode *SV0 = cast<ShuffleVectorSDNode>(N0);
3252 const ShuffleVectorSDNode *SV1 = cast<ShuffleVectorSDNode>(N1);
3253 // We construct two shuffle masks:
3254 // - Mask1 is a shuffle mask for a shuffle with N0 as the first operand
3255 // and N1 as the second operand.
3256 // - Mask2 is a shuffle mask for a shuffle with N1 as the first operand
3257 // and N0 as the second operand.
3258 // We do this because OR is commutable and therefore there might be
3259 // two ways to fold this node into a shuffle.
3260 SmallVector<int,4> Mask1;
3261 SmallVector<int,4> Mask2;
3263 for (unsigned i = 0; i != NumElts && CanFold; ++i) {
3264 int M0 = SV0->getMaskElt(i);
3265 int M1 = SV1->getMaskElt(i);
3267 // Both shuffle indexes are undef. Propagate Undef.
3268 if (M0 < 0 && M1 < 0) {
3269 Mask1.push_back(M0);
3270 Mask2.push_back(M0);
3274 if (M0 < 0 || M1 < 0 ||
3275 (M0 < (int)NumElts && M1 < (int)NumElts) ||
3276 (M0 >= (int)NumElts && M1 >= (int)NumElts)) {
3281 Mask1.push_back(M0 < (int)NumElts ? M0 : M1 + NumElts);
3282 Mask2.push_back(M1 < (int)NumElts ? M1 : M0 + NumElts);
3286 // Fold this sequence only if the resulting shuffle is 'legal'.
3287 if (TLI.isShuffleMaskLegal(Mask1, VT))
3288 return DAG.getVectorShuffle(VT, SDLoc(N), N0->getOperand(0),
3289 N1->getOperand(0), &Mask1[0]);
3290 if (TLI.isShuffleMaskLegal(Mask2, VT))
3291 return DAG.getVectorShuffle(VT, SDLoc(N), N1->getOperand(0),
3292 N0->getOperand(0), &Mask2[0]);
3297 // fold (or x, undef) -> -1
3298 if (!LegalOperations &&
3299 (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)) {
3300 EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT;
3301 return DAG.getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()), VT);
3303 // fold (or c1, c2) -> c1|c2
3305 return DAG.FoldConstantArithmetic(ISD::OR, VT, N0C, N1C);
3306 // canonicalize constant to RHS
3308 return DAG.getNode(ISD::OR, SDLoc(N), VT, N1, N0);
3309 // fold (or x, 0) -> x
3310 if (N1C && N1C->isNullValue())
3312 // fold (or x, -1) -> -1
3313 if (N1C && N1C->isAllOnesValue())
3315 // fold (or x, c) -> c iff (x & ~c) == 0
3316 if (N1C && DAG.MaskedValueIsZero(N0, ~N1C->getAPIntValue()))
3319 // Recognize halfword bswaps as (bswap + rotl 16) or (bswap + shl 16)
3320 SDValue BSwap = MatchBSwapHWord(N, N0, N1);
3321 if (BSwap.getNode())
3323 BSwap = MatchBSwapHWordLow(N, N0, N1);
3324 if (BSwap.getNode())
3328 SDValue ROR = ReassociateOps(ISD::OR, SDLoc(N), N0, N1);
3331 // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2)
3332 // iff (c1 & c2) == 0.
3333 if (N1C && N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
3334 isa<ConstantSDNode>(N0.getOperand(1))) {
3335 ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1));
3336 if ((C1->getAPIntValue() & N1C->getAPIntValue()) != 0) {
3337 SDValue COR = DAG.FoldConstantArithmetic(ISD::OR, VT, N1C, C1);
3340 return DAG.getNode(ISD::AND, SDLoc(N), VT,
3341 DAG.getNode(ISD::OR, SDLoc(N0), VT,
3342 N0.getOperand(0), N1), COR);
3345 // fold (or (setcc x), (setcc y)) -> (setcc (or x, y))
3346 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
3347 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
3348 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
3350 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
3351 LL.getValueType().isInteger()) {
3352 // fold (or (setne X, 0), (setne Y, 0)) -> (setne (or X, Y), 0)
3353 // fold (or (setlt X, 0), (setlt Y, 0)) -> (setne (or X, Y), 0)
3354 if (cast<ConstantSDNode>(LR)->isNullValue() &&
3355 (Op1 == ISD::SETNE || Op1 == ISD::SETLT)) {
3356 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(LR),
3357 LR.getValueType(), LL, RL);
3358 AddToWorkList(ORNode.getNode());
3359 return DAG.getSetCC(SDLoc(N), VT, ORNode, LR, Op1);
3361 // fold (or (setne X, -1), (setne Y, -1)) -> (setne (and X, Y), -1)
3362 // fold (or (setgt X, -1), (setgt Y -1)) -> (setgt (and X, Y), -1)
3363 if (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
3364 (Op1 == ISD::SETNE || Op1 == ISD::SETGT)) {
3365 SDValue ANDNode = DAG.getNode(ISD::AND, SDLoc(LR),
3366 LR.getValueType(), LL, RL);
3367 AddToWorkList(ANDNode.getNode());
3368 return DAG.getSetCC(SDLoc(N), VT, ANDNode, LR, Op1);
3371 // canonicalize equivalent to ll == rl
3372 if (LL == RR && LR == RL) {
3373 Op1 = ISD::getSetCCSwappedOperands(Op1);
3376 if (LL == RL && LR == RR) {
3377 bool isInteger = LL.getValueType().isInteger();
3378 ISD::CondCode Result = ISD::getSetCCOrOperation(Op0, Op1, isInteger);
3379 if (Result != ISD::SETCC_INVALID &&
3380 (!LegalOperations ||
3381 (TLI.isCondCodeLegal(Result, LL.getSimpleValueType()) &&
3382 TLI.isOperationLegal(ISD::SETCC,
3383 getSetCCResultType(N0.getValueType())))))
3384 return DAG.getSetCC(SDLoc(N), N0.getValueType(),
3389 // Simplify: (or (op x...), (op y...)) -> (op (or x, y))
3390 if (N0.getOpcode() == N1.getOpcode()) {
3391 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3392 if (Tmp.getNode()) return Tmp;
3395 // (or (and X, C1), (and Y, C2)) -> (and (or X, Y), C3) if possible.
3396 if (N0.getOpcode() == ISD::AND &&
3397 N1.getOpcode() == ISD::AND &&
3398 N0.getOperand(1).getOpcode() == ISD::Constant &&
3399 N1.getOperand(1).getOpcode() == ISD::Constant &&
3400 // Don't increase # computations.
3401 (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) {
3402 // We can only do this xform if we know that bits from X that are set in C2
3403 // but not in C1 are already zero. Likewise for Y.
3404 const APInt &LHSMask =
3405 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
3406 const APInt &RHSMask =
3407 cast<ConstantSDNode>(N1.getOperand(1))->getAPIntValue();
3409 if (DAG.MaskedValueIsZero(N0.getOperand(0), RHSMask&~LHSMask) &&
3410 DAG.MaskedValueIsZero(N1.getOperand(0), LHSMask&~RHSMask)) {
3411 SDValue X = DAG.getNode(ISD::OR, SDLoc(N0), VT,
3412 N0.getOperand(0), N1.getOperand(0));
3413 return DAG.getNode(ISD::AND, SDLoc(N), VT, X,
3414 DAG.getConstant(LHSMask | RHSMask, VT));
3418 // See if this is some rotate idiom.
3419 if (SDNode *Rot = MatchRotate(N0, N1, SDLoc(N)))
3420 return SDValue(Rot, 0);
3422 // Simplify the operands using demanded-bits information.
3423 if (!VT.isVector() &&
3424 SimplifyDemandedBits(SDValue(N, 0)))
3425 return SDValue(N, 0);
3430 /// MatchRotateHalf - Match "(X shl/srl V1) & V2" where V2 may not be present.
3431 static bool MatchRotateHalf(SDValue Op, SDValue &Shift, SDValue &Mask) {
3432 if (Op.getOpcode() == ISD::AND) {
3433 if (isa<ConstantSDNode>(Op.getOperand(1))) {
3434 Mask = Op.getOperand(1);
3435 Op = Op.getOperand(0);
3441 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SHL) {
3449 // Return true if we can prove that, whenever Neg and Pos are both in the
3450 // range [0, OpSize), Neg == (Pos == 0 ? 0 : OpSize - Pos). This means that
3451 // for two opposing shifts shift1 and shift2 and a value X with OpBits bits:
3453 // (or (shift1 X, Neg), (shift2 X, Pos))
3455 // reduces to a rotate in direction shift2 by Pos or (equivalently) a rotate
3456 // in direction shift1 by Neg. The range [0, OpSize) means that we only need
3457 // to consider shift amounts with defined behavior.
3458 static bool matchRotateSub(SDValue Pos, SDValue Neg, unsigned OpSize) {
3459 // If OpSize is a power of 2 then:
3461 // (a) (Pos == 0 ? 0 : OpSize - Pos) == (OpSize - Pos) & (OpSize - 1)
3462 // (b) Neg == Neg & (OpSize - 1) whenever Neg is in [0, OpSize).
3464 // So if OpSize is a power of 2 and Neg is (and Neg', OpSize-1), we check
3465 // for the stronger condition:
3467 // Neg & (OpSize - 1) == (OpSize - Pos) & (OpSize - 1) [A]
3469 // for all Neg and Pos. Since Neg & (OpSize - 1) == Neg' & (OpSize - 1)
3470 // we can just replace Neg with Neg' for the rest of the function.
3472 // In other cases we check for the even stronger condition:
3474 // Neg == OpSize - Pos [B]
3476 // for all Neg and Pos. Note that the (or ...) then invokes undefined
3477 // behavior if Pos == 0 (and consequently Neg == OpSize).
3479 // We could actually use [A] whenever OpSize is a power of 2, but the
3480 // only extra cases that it would match are those uninteresting ones
3481 // where Neg and Pos are never in range at the same time. E.g. for
3482 // OpSize == 32, using [A] would allow a Neg of the form (sub 64, Pos)
3483 // as well as (sub 32, Pos), but:
3485 // (or (shift1 X, (sub 64, Pos)), (shift2 X, Pos))
3487 // always invokes undefined behavior for 32-bit X.
3489 // Below, Mask == OpSize - 1 when using [A] and is all-ones otherwise.
3490 unsigned MaskLoBits = 0;
3491 if (Neg.getOpcode() == ISD::AND &&
3492 isPowerOf2_64(OpSize) &&
3493 Neg.getOperand(1).getOpcode() == ISD::Constant &&
3494 cast<ConstantSDNode>(Neg.getOperand(1))->getAPIntValue() == OpSize - 1) {
3495 Neg = Neg.getOperand(0);
3496 MaskLoBits = Log2_64(OpSize);
3499 // Check whether Neg has the form (sub NegC, NegOp1) for some NegC and NegOp1.
3500 if (Neg.getOpcode() != ISD::SUB)
3502 ConstantSDNode *NegC = dyn_cast<ConstantSDNode>(Neg.getOperand(0));
3505 SDValue NegOp1 = Neg.getOperand(1);
3507 // On the RHS of [A], if Pos is Pos' & (OpSize - 1), just replace Pos with
3508 // Pos'. The truncation is redundant for the purpose of the equality.
3510 Pos.getOpcode() == ISD::AND &&
3511 Pos.getOperand(1).getOpcode() == ISD::Constant &&
3512 cast<ConstantSDNode>(Pos.getOperand(1))->getAPIntValue() == OpSize - 1)
3513 Pos = Pos.getOperand(0);
3515 // The condition we need is now:
3517 // (NegC - NegOp1) & Mask == (OpSize - Pos) & Mask
3519 // If NegOp1 == Pos then we need:
3521 // OpSize & Mask == NegC & Mask
3523 // (because "x & Mask" is a truncation and distributes through subtraction).
3526 Width = NegC->getAPIntValue();
3527 // Check for cases where Pos has the form (add NegOp1, PosC) for some PosC.
3528 // Then the condition we want to prove becomes:
3530 // (NegC - NegOp1) & Mask == (OpSize - (NegOp1 + PosC)) & Mask
3532 // which, again because "x & Mask" is a truncation, becomes:
3534 // NegC & Mask == (OpSize - PosC) & Mask
3535 // OpSize & Mask == (NegC + PosC) & Mask
3536 else if (Pos.getOpcode() == ISD::ADD &&
3537 Pos.getOperand(0) == NegOp1 &&
3538 Pos.getOperand(1).getOpcode() == ISD::Constant)
3539 Width = (cast<ConstantSDNode>(Pos.getOperand(1))->getAPIntValue() +
3540 NegC->getAPIntValue());
3544 // Now we just need to check that OpSize & Mask == Width & Mask.
3546 // Opsize & Mask is 0 since Mask is Opsize - 1.
3547 return Width.getLoBits(MaskLoBits) == 0;
3548 return Width == OpSize;
3551 // A subroutine of MatchRotate used once we have found an OR of two opposite
3552 // shifts of Shifted. If Neg == <operand size> - Pos then the OR reduces
3553 // to both (PosOpcode Shifted, Pos) and (NegOpcode Shifted, Neg), with the
3554 // former being preferred if supported. InnerPos and InnerNeg are Pos and
3555 // Neg with outer conversions stripped away.
3556 SDNode *DAGCombiner::MatchRotatePosNeg(SDValue Shifted, SDValue Pos,
3557 SDValue Neg, SDValue InnerPos,
3558 SDValue InnerNeg, unsigned PosOpcode,
3559 unsigned NegOpcode, SDLoc DL) {
3560 // fold (or (shl x, (*ext y)),
3561 // (srl x, (*ext (sub 32, y)))) ->
3562 // (rotl x, y) or (rotr x, (sub 32, y))
3564 // fold (or (shl x, (*ext (sub 32, y))),
3565 // (srl x, (*ext y))) ->
3566 // (rotr x, y) or (rotl x, (sub 32, y))
3567 EVT VT = Shifted.getValueType();
3568 if (matchRotateSub(InnerPos, InnerNeg, VT.getSizeInBits())) {
3569 bool HasPos = TLI.isOperationLegalOrCustom(PosOpcode, VT);
3570 return DAG.getNode(HasPos ? PosOpcode : NegOpcode, DL, VT, Shifted,
3571 HasPos ? Pos : Neg).getNode();
3577 // MatchRotate - Handle an 'or' of two operands. If this is one of the many
3578 // idioms for rotate, and if the target supports rotation instructions, generate
3580 SDNode *DAGCombiner::MatchRotate(SDValue LHS, SDValue RHS, SDLoc DL) {
3581 // Must be a legal type. Expanded 'n promoted things won't work with rotates.
3582 EVT VT = LHS.getValueType();
3583 if (!TLI.isTypeLegal(VT)) return nullptr;
3585 // The target must have at least one rotate flavor.
3586 bool HasROTL = TLI.isOperationLegalOrCustom(ISD::ROTL, VT);
3587 bool HasROTR = TLI.isOperationLegalOrCustom(ISD::ROTR, VT);
3588 if (!HasROTL && !HasROTR) return nullptr;
3590 // Match "(X shl/srl V1) & V2" where V2 may not be present.
3591 SDValue LHSShift; // The shift.
3592 SDValue LHSMask; // AND value if any.
3593 if (!MatchRotateHalf(LHS, LHSShift, LHSMask))
3594 return nullptr; // Not part of a rotate.
3596 SDValue RHSShift; // The shift.
3597 SDValue RHSMask; // AND value if any.
3598 if (!MatchRotateHalf(RHS, RHSShift, RHSMask))
3599 return nullptr; // Not part of a rotate.
3601 if (LHSShift.getOperand(0) != RHSShift.getOperand(0))
3602 return nullptr; // Not shifting the same value.
3604 if (LHSShift.getOpcode() == RHSShift.getOpcode())
3605 return nullptr; // Shifts must disagree.
3607 // Canonicalize shl to left side in a shl/srl pair.
3608 if (RHSShift.getOpcode() == ISD::SHL) {
3609 std::swap(LHS, RHS);
3610 std::swap(LHSShift, RHSShift);
3611 std::swap(LHSMask , RHSMask );
3614 unsigned OpSizeInBits = VT.getSizeInBits();
3615 SDValue LHSShiftArg = LHSShift.getOperand(0);
3616 SDValue LHSShiftAmt = LHSShift.getOperand(1);
3617 SDValue RHSShiftArg = RHSShift.getOperand(0);
3618 SDValue RHSShiftAmt = RHSShift.getOperand(1);
3620 // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1)
3621 // fold (or (shl x, C1), (srl x, C2)) -> (rotr x, C2)
3622 if (LHSShiftAmt.getOpcode() == ISD::Constant &&
3623 RHSShiftAmt.getOpcode() == ISD::Constant) {
3624 uint64_t LShVal = cast<ConstantSDNode>(LHSShiftAmt)->getZExtValue();
3625 uint64_t RShVal = cast<ConstantSDNode>(RHSShiftAmt)->getZExtValue();
3626 if ((LShVal + RShVal) != OpSizeInBits)
3629 SDValue Rot = DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT,
3630 LHSShiftArg, HasROTL ? LHSShiftAmt : RHSShiftAmt);
3632 // If there is an AND of either shifted operand, apply it to the result.
3633 if (LHSMask.getNode() || RHSMask.getNode()) {
3634 APInt Mask = APInt::getAllOnesValue(OpSizeInBits);
3636 if (LHSMask.getNode()) {
3637 APInt RHSBits = APInt::getLowBitsSet(OpSizeInBits, LShVal);
3638 Mask &= cast<ConstantSDNode>(LHSMask)->getAPIntValue() | RHSBits;
3640 if (RHSMask.getNode()) {
3641 APInt LHSBits = APInt::getHighBitsSet(OpSizeInBits, RShVal);
3642 Mask &= cast<ConstantSDNode>(RHSMask)->getAPIntValue() | LHSBits;
3645 Rot = DAG.getNode(ISD::AND, DL, VT, Rot, DAG.getConstant(Mask, VT));
3648 return Rot.getNode();
3651 // If there is a mask here, and we have a variable shift, we can't be sure
3652 // that we're masking out the right stuff.
3653 if (LHSMask.getNode() || RHSMask.getNode())
3656 // If the shift amount is sign/zext/any-extended just peel it off.
3657 SDValue LExtOp0 = LHSShiftAmt;
3658 SDValue RExtOp0 = RHSShiftAmt;
3659 if ((LHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3660 LHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3661 LHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3662 LHSShiftAmt.getOpcode() == ISD::TRUNCATE) &&
3663 (RHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3664 RHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3665 RHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3666 RHSShiftAmt.getOpcode() == ISD::TRUNCATE)) {
3667 LExtOp0 = LHSShiftAmt.getOperand(0);
3668 RExtOp0 = RHSShiftAmt.getOperand(0);
3671 SDNode *TryL = MatchRotatePosNeg(LHSShiftArg, LHSShiftAmt, RHSShiftAmt,
3672 LExtOp0, RExtOp0, ISD::ROTL, ISD::ROTR, DL);
3676 SDNode *TryR = MatchRotatePosNeg(RHSShiftArg, RHSShiftAmt, LHSShiftAmt,
3677 RExtOp0, LExtOp0, ISD::ROTR, ISD::ROTL, DL);
3684 SDValue DAGCombiner::visitXOR(SDNode *N) {
3685 SDValue N0 = N->getOperand(0);
3686 SDValue N1 = N->getOperand(1);
3687 SDValue LHS, RHS, CC;
3688 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3689 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3690 EVT VT = N0.getValueType();
3693 if (VT.isVector()) {
3694 SDValue FoldedVOp = SimplifyVBinOp(N);
3695 if (FoldedVOp.getNode()) return FoldedVOp;
3697 // fold (xor x, 0) -> x, vector edition
3698 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3700 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3704 // fold (xor undef, undef) -> 0. This is a common idiom (misuse).
3705 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
3706 return DAG.getConstant(0, VT);
3707 // fold (xor x, undef) -> undef
3708 if (N0.getOpcode() == ISD::UNDEF)
3710 if (N1.getOpcode() == ISD::UNDEF)
3712 // fold (xor c1, c2) -> c1^c2
3714 return DAG.FoldConstantArithmetic(ISD::XOR, VT, N0C, N1C);
3715 // canonicalize constant to RHS
3717 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0);
3718 // fold (xor x, 0) -> x
3719 if (N1C && N1C->isNullValue())
3722 SDValue RXOR = ReassociateOps(ISD::XOR, SDLoc(N), N0, N1);
3726 // fold !(x cc y) -> (x !cc y)
3727 if (N1C && N1C->getAPIntValue() == 1 && isSetCCEquivalent(N0, LHS, RHS, CC)) {
3728 bool isInt = LHS.getValueType().isInteger();
3729 ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
3732 if (!LegalOperations ||
3733 TLI.isCondCodeLegal(NotCC, LHS.getSimpleValueType())) {
3734 switch (N0.getOpcode()) {
3736 llvm_unreachable("Unhandled SetCC Equivalent!");
3738 return DAG.getSetCC(SDLoc(N), VT, LHS, RHS, NotCC);
3739 case ISD::SELECT_CC:
3740 return DAG.getSelectCC(SDLoc(N), LHS, RHS, N0.getOperand(2),
3741 N0.getOperand(3), NotCC);
3746 // fold (not (zext (setcc x, y))) -> (zext (not (setcc x, y)))
3747 if (N1C && N1C->getAPIntValue() == 1 && N0.getOpcode() == ISD::ZERO_EXTEND &&
3748 N0.getNode()->hasOneUse() &&
3749 isSetCCEquivalent(N0.getOperand(0), LHS, RHS, CC)){
3750 SDValue V = N0.getOperand(0);
3751 V = DAG.getNode(ISD::XOR, SDLoc(N0), V.getValueType(), V,
3752 DAG.getConstant(1, V.getValueType()));
3753 AddToWorkList(V.getNode());
3754 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, V);
3757 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are setcc
3758 if (N1C && N1C->getAPIntValue() == 1 && VT == MVT::i1 &&
3759 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3760 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3761 if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) {
3762 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3763 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS
3764 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS
3765 AddToWorkList(LHS.getNode()); AddToWorkList(RHS.getNode());
3766 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS);
3769 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are constants
3770 if (N1C && N1C->isAllOnesValue() &&
3771 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3772 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3773 if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) {
3774 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3775 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS
3776 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS
3777 AddToWorkList(LHS.getNode()); AddToWorkList(RHS.getNode());
3778 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS);
3781 // fold (xor (and x, y), y) -> (and (not x), y)
3782 if (N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
3783 N0->getOperand(1) == N1) {
3784 SDValue X = N0->getOperand(0);
3785 SDValue NotX = DAG.getNOT(SDLoc(X), X, VT);
3786 AddToWorkList(NotX.getNode());
3787 return DAG.getNode(ISD::AND, SDLoc(N), VT, NotX, N1);
3789 // fold (xor (xor x, c1), c2) -> (xor x, (xor c1, c2))
3790 if (N1C && N0.getOpcode() == ISD::XOR) {
3791 ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0));
3792 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3794 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N0.getOperand(1),
3795 DAG.getConstant(N1C->getAPIntValue() ^
3796 N00C->getAPIntValue(), VT));
3798 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N0.getOperand(0),
3799 DAG.getConstant(N1C->getAPIntValue() ^
3800 N01C->getAPIntValue(), VT));
3802 // fold (xor x, x) -> 0
3804 return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations, LegalTypes);
3806 // Simplify: xor (op x...), (op y...) -> (op (xor x, y))
3807 if (N0.getOpcode() == N1.getOpcode()) {
3808 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3809 if (Tmp.getNode()) return Tmp;
3812 // Simplify the expression using non-local knowledge.
3813 if (!VT.isVector() &&
3814 SimplifyDemandedBits(SDValue(N, 0)))
3815 return SDValue(N, 0);
3820 /// visitShiftByConstant - Handle transforms common to the three shifts, when
3821 /// the shift amount is a constant.
3822 SDValue DAGCombiner::visitShiftByConstant(SDNode *N, ConstantSDNode *Amt) {
3823 // We can't and shouldn't fold opaque constants.
3824 if (Amt->isOpaque())
3827 SDNode *LHS = N->getOperand(0).getNode();
3828 if (!LHS->hasOneUse()) return SDValue();
3830 // We want to pull some binops through shifts, so that we have (and (shift))
3831 // instead of (shift (and)), likewise for add, or, xor, etc. This sort of
3832 // thing happens with address calculations, so it's important to canonicalize
3834 bool HighBitSet = false; // Can we transform this if the high bit is set?
3836 switch (LHS->getOpcode()) {
3837 default: return SDValue();
3840 HighBitSet = false; // We can only transform sra if the high bit is clear.
3843 HighBitSet = true; // We can only transform sra if the high bit is set.
3846 if (N->getOpcode() != ISD::SHL)
3847 return SDValue(); // only shl(add) not sr[al](add).
3848 HighBitSet = false; // We can only transform sra if the high bit is clear.
3852 // We require the RHS of the binop to be a constant and not opaque as well.
3853 ConstantSDNode *BinOpCst = dyn_cast<ConstantSDNode>(LHS->getOperand(1));
3854 if (!BinOpCst || BinOpCst->isOpaque()) return SDValue();
3856 // FIXME: disable this unless the input to the binop is a shift by a constant.
3857 // If it is not a shift, it pessimizes some common cases like:
3859 // void foo(int *X, int i) { X[i & 1235] = 1; }
3860 // int bar(int *X, int i) { return X[i & 255]; }
3861 SDNode *BinOpLHSVal = LHS->getOperand(0).getNode();
3862 if ((BinOpLHSVal->getOpcode() != ISD::SHL &&
3863 BinOpLHSVal->getOpcode() != ISD::SRA &&
3864 BinOpLHSVal->getOpcode() != ISD::SRL) ||
3865 !isa<ConstantSDNode>(BinOpLHSVal->getOperand(1)))
3868 EVT VT = N->getValueType(0);
3870 // If this is a signed shift right, and the high bit is modified by the
3871 // logical operation, do not perform the transformation. The highBitSet
3872 // boolean indicates the value of the high bit of the constant which would
3873 // cause it to be modified for this operation.
3874 if (N->getOpcode() == ISD::SRA) {
3875 bool BinOpRHSSignSet = BinOpCst->getAPIntValue().isNegative();
3876 if (BinOpRHSSignSet != HighBitSet)
3880 if (!TLI.isDesirableToCommuteWithShift(LHS))
3883 // Fold the constants, shifting the binop RHS by the shift amount.
3884 SDValue NewRHS = DAG.getNode(N->getOpcode(), SDLoc(LHS->getOperand(1)),
3886 LHS->getOperand(1), N->getOperand(1));
3887 assert(isa<ConstantSDNode>(NewRHS) && "Folding was not successful!");
3889 // Create the new shift.
3890 SDValue NewShift = DAG.getNode(N->getOpcode(),
3891 SDLoc(LHS->getOperand(0)),
3892 VT, LHS->getOperand(0), N->getOperand(1));
3894 // Create the new binop.
3895 return DAG.getNode(LHS->getOpcode(), SDLoc(N), VT, NewShift, NewRHS);
3898 SDValue DAGCombiner::distributeTruncateThroughAnd(SDNode *N) {
3899 assert(N->getOpcode() == ISD::TRUNCATE);
3900 assert(N->getOperand(0).getOpcode() == ISD::AND);
3902 // (truncate:TruncVT (and N00, N01C)) -> (and (truncate:TruncVT N00), TruncC)
3903 if (N->hasOneUse() && N->getOperand(0).hasOneUse()) {
3904 SDValue N01 = N->getOperand(0).getOperand(1);
3906 if (ConstantSDNode *N01C = isConstOrConstSplat(N01)) {
3907 EVT TruncVT = N->getValueType(0);
3908 SDValue N00 = N->getOperand(0).getOperand(0);
3909 APInt TruncC = N01C->getAPIntValue();
3910 TruncC = TruncC.trunc(TruncVT.getScalarSizeInBits());
3912 return DAG.getNode(ISD::AND, SDLoc(N), TruncVT,
3913 DAG.getNode(ISD::TRUNCATE, SDLoc(N), TruncVT, N00),
3914 DAG.getConstant(TruncC, TruncVT));
3921 SDValue DAGCombiner::visitRotate(SDNode *N) {
3922 // fold (rot* x, (trunc (and y, c))) -> (rot* x, (and (trunc y), (trunc c))).
3923 if (N->getOperand(1).getOpcode() == ISD::TRUNCATE &&
3924 N->getOperand(1).getOperand(0).getOpcode() == ISD::AND) {
3925 SDValue NewOp1 = distributeTruncateThroughAnd(N->getOperand(1).getNode());
3926 if (NewOp1.getNode())
3927 return DAG.getNode(N->getOpcode(), SDLoc(N), N->getValueType(0),
3928 N->getOperand(0), NewOp1);
3933 SDValue DAGCombiner::visitSHL(SDNode *N) {
3934 SDValue N0 = N->getOperand(0);
3935 SDValue N1 = N->getOperand(1);
3936 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3937 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3938 EVT VT = N0.getValueType();
3939 unsigned OpSizeInBits = VT.getScalarSizeInBits();
3942 if (VT.isVector()) {
3943 SDValue FoldedVOp = SimplifyVBinOp(N);
3944 if (FoldedVOp.getNode()) return FoldedVOp;
3946 BuildVectorSDNode *N1CV = dyn_cast<BuildVectorSDNode>(N1);
3947 // If setcc produces all-one true value then:
3948 // (shl (and (setcc) N01CV) N1CV) -> (and (setcc) N01CV<<N1CV)
3949 if (N1CV && N1CV->isConstant()) {
3950 if (N0.getOpcode() == ISD::AND &&
3951 TLI.getBooleanContents(true) ==
3952 TargetLowering::ZeroOrNegativeOneBooleanContent) {
3953 SDValue N00 = N0->getOperand(0);
3954 SDValue N01 = N0->getOperand(1);
3955 BuildVectorSDNode *N01CV = dyn_cast<BuildVectorSDNode>(N01);
3957 if (N01CV && N01CV->isConstant() && N00.getOpcode() == ISD::SETCC) {
3958 SDValue C = DAG.FoldConstantArithmetic(ISD::SHL, VT, N01CV, N1CV);
3960 return DAG.getNode(ISD::AND, SDLoc(N), VT, N00, C);
3963 N1C = isConstOrConstSplat(N1);
3968 // fold (shl c1, c2) -> c1<<c2
3970 return DAG.FoldConstantArithmetic(ISD::SHL, VT, N0C, N1C);
3971 // fold (shl 0, x) -> 0
3972 if (N0C && N0C->isNullValue())
3974 // fold (shl x, c >= size(x)) -> undef
3975 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
3976 return DAG.getUNDEF(VT);
3977 // fold (shl x, 0) -> x
3978 if (N1C && N1C->isNullValue())
3980 // fold (shl undef, x) -> 0
3981 if (N0.getOpcode() == ISD::UNDEF)
3982 return DAG.getConstant(0, VT);
3983 // if (shl x, c) is known to be zero, return 0
3984 if (DAG.MaskedValueIsZero(SDValue(N, 0),
3985 APInt::getAllOnesValue(OpSizeInBits)))
3986 return DAG.getConstant(0, VT);
3987 // fold (shl x, (trunc (and y, c))) -> (shl x, (and (trunc y), (trunc c))).
3988 if (N1.getOpcode() == ISD::TRUNCATE &&
3989 N1.getOperand(0).getOpcode() == ISD::AND) {
3990 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
3991 if (NewOp1.getNode())
3992 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0, NewOp1);
3995 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
3996 return SDValue(N, 0);
3998 // fold (shl (shl x, c1), c2) -> 0 or (shl x, (add c1, c2))
3999 if (N1C && N0.getOpcode() == ISD::SHL) {
4000 if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
4001 uint64_t c1 = N0C1->getZExtValue();
4002 uint64_t c2 = N1C->getZExtValue();
4003 if (c1 + c2 >= OpSizeInBits)
4004 return DAG.getConstant(0, VT);
4005 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0),
4006 DAG.getConstant(c1 + c2, N1.getValueType()));
4010 // fold (shl (ext (shl x, c1)), c2) -> (ext (shl x, (add c1, c2)))
4011 // For this to be valid, the second form must not preserve any of the bits
4012 // that are shifted out by the inner shift in the first form. This means
4013 // the outer shift size must be >= the number of bits added by the ext.
4014 // As a corollary, we don't care what kind of ext it is.
4015 if (N1C && (N0.getOpcode() == ISD::ZERO_EXTEND ||
4016 N0.getOpcode() == ISD::ANY_EXTEND ||
4017 N0.getOpcode() == ISD::SIGN_EXTEND) &&
4018 N0.getOperand(0).getOpcode() == ISD::SHL) {
4019 SDValue N0Op0 = N0.getOperand(0);
4020 if (ConstantSDNode *N0Op0C1 = isConstOrConstSplat(N0Op0.getOperand(1))) {
4021 uint64_t c1 = N0Op0C1->getZExtValue();
4022 uint64_t c2 = N1C->getZExtValue();
4023 EVT InnerShiftVT = N0Op0.getValueType();
4024 uint64_t InnerShiftSize = InnerShiftVT.getScalarSizeInBits();
4025 if (c2 >= OpSizeInBits - InnerShiftSize) {
4026 if (c1 + c2 >= OpSizeInBits)
4027 return DAG.getConstant(0, VT);
4028 return DAG.getNode(ISD::SHL, SDLoc(N0), VT,
4029 DAG.getNode(N0.getOpcode(), SDLoc(N0), VT,
4030 N0Op0->getOperand(0)),
4031 DAG.getConstant(c1 + c2, N1.getValueType()));
4036 // fold (shl (zext (srl x, C)), C) -> (zext (shl (srl x, C), C))
4037 // Only fold this if the inner zext has no other uses to avoid increasing
4038 // the total number of instructions.
4039 if (N1C && N0.getOpcode() == ISD::ZERO_EXTEND && N0.hasOneUse() &&
4040 N0.getOperand(0).getOpcode() == ISD::SRL) {
4041 SDValue N0Op0 = N0.getOperand(0);
4042 if (ConstantSDNode *N0Op0C1 = isConstOrConstSplat(N0Op0.getOperand(1))) {
4043 uint64_t c1 = N0Op0C1->getZExtValue();
4044 if (c1 < VT.getScalarSizeInBits()) {
4045 uint64_t c2 = N1C->getZExtValue();
4047 SDValue NewOp0 = N0.getOperand(0);
4048 EVT CountVT = NewOp0.getOperand(1).getValueType();
4049 SDValue NewSHL = DAG.getNode(ISD::SHL, SDLoc(N), NewOp0.getValueType(),
4050 NewOp0, DAG.getConstant(c2, CountVT));
4051 AddToWorkList(NewSHL.getNode());
4052 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N0), VT, NewSHL);
4058 // fold (shl (srl x, c1), c2) -> (and (shl x, (sub c2, c1), MASK) or
4059 // (and (srl x, (sub c1, c2), MASK)
4060 // Only fold this if the inner shift has no other uses -- if it does, folding
4061 // this will increase the total number of instructions.
4062 if (N1C && N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
4063 if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
4064 uint64_t c1 = N0C1->getZExtValue();
4065 if (c1 < OpSizeInBits) {
4066 uint64_t c2 = N1C->getZExtValue();
4067 APInt Mask = APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - c1);
4070 Mask = Mask.shl(c2 - c1);
4071 Shift = DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0),
4072 DAG.getConstant(c2 - c1, N1.getValueType()));
4074 Mask = Mask.lshr(c1 - c2);
4075 Shift = DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0),
4076 DAG.getConstant(c1 - c2, N1.getValueType()));
4078 return DAG.getNode(ISD::AND, SDLoc(N0), VT, Shift,
4079 DAG.getConstant(Mask, VT));
4083 // fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1))
4084 if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1)) {
4085 unsigned BitSize = VT.getScalarSizeInBits();
4086 SDValue HiBitsMask =
4087 DAG.getConstant(APInt::getHighBitsSet(BitSize,
4088 BitSize - N1C->getZExtValue()), VT);
4089 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0.getOperand(0),
4094 SDValue NewSHL = visitShiftByConstant(N, N1C);
4095 if (NewSHL.getNode())
4102 SDValue DAGCombiner::visitSRA(SDNode *N) {
4103 SDValue N0 = N->getOperand(0);
4104 SDValue N1 = N->getOperand(1);
4105 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4106 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4107 EVT VT = N0.getValueType();
4108 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
4111 if (VT.isVector()) {
4112 SDValue FoldedVOp = SimplifyVBinOp(N);
4113 if (FoldedVOp.getNode()) return FoldedVOp;
4115 N1C = isConstOrConstSplat(N1);
4118 // fold (sra c1, c2) -> (sra c1, c2)
4120 return DAG.FoldConstantArithmetic(ISD::SRA, VT, N0C, N1C);
4121 // fold (sra 0, x) -> 0
4122 if (N0C && N0C->isNullValue())
4124 // fold (sra -1, x) -> -1
4125 if (N0C && N0C->isAllOnesValue())
4127 // fold (sra x, (setge c, size(x))) -> undef
4128 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4129 return DAG.getUNDEF(VT);
4130 // fold (sra x, 0) -> x
4131 if (N1C && N1C->isNullValue())
4133 // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports
4135 if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) {
4136 unsigned LowBits = OpSizeInBits - (unsigned)N1C->getZExtValue();
4137 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), LowBits);
4139 ExtVT = EVT::getVectorVT(*DAG.getContext(),
4140 ExtVT, VT.getVectorNumElements());
4141 if ((!LegalOperations ||
4142 TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, ExtVT)))
4143 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
4144 N0.getOperand(0), DAG.getValueType(ExtVT));
4147 // fold (sra (sra x, c1), c2) -> (sra x, (add c1, c2))
4148 if (N1C && N0.getOpcode() == ISD::SRA) {
4149 if (ConstantSDNode *C1 = isConstOrConstSplat(N0.getOperand(1))) {
4150 unsigned Sum = N1C->getZExtValue() + C1->getZExtValue();
4151 if (Sum >= OpSizeInBits)
4152 Sum = OpSizeInBits - 1;
4153 return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0.getOperand(0),
4154 DAG.getConstant(Sum, N1.getValueType()));
4158 // fold (sra (shl X, m), (sub result_size, n))
4159 // -> (sign_extend (trunc (shl X, (sub (sub result_size, n), m)))) for
4160 // result_size - n != m.
4161 // If truncate is free for the target sext(shl) is likely to result in better
4163 if (N0.getOpcode() == ISD::SHL && N1C) {
4164 // Get the two constanst of the shifts, CN0 = m, CN = n.
4165 const ConstantSDNode *N01C = isConstOrConstSplat(N0.getOperand(1));
4167 LLVMContext &Ctx = *DAG.getContext();
4168 // Determine what the truncate's result bitsize and type would be.
4169 EVT TruncVT = EVT::getIntegerVT(Ctx, OpSizeInBits - N1C->getZExtValue());
4172 TruncVT = EVT::getVectorVT(Ctx, TruncVT, VT.getVectorNumElements());
4174 // Determine the residual right-shift amount.
4175 signed ShiftAmt = N1C->getZExtValue() - N01C->getZExtValue();
4177 // If the shift is not a no-op (in which case this should be just a sign
4178 // extend already), the truncated to type is legal, sign_extend is legal
4179 // on that type, and the truncate to that type is both legal and free,
4180 // perform the transform.
4181 if ((ShiftAmt > 0) &&
4182 TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND, TruncVT) &&
4183 TLI.isOperationLegalOrCustom(ISD::TRUNCATE, VT) &&
4184 TLI.isTruncateFree(VT, TruncVT)) {
4186 SDValue Amt = DAG.getConstant(ShiftAmt,
4187 getShiftAmountTy(N0.getOperand(0).getValueType()));
4188 SDValue Shift = DAG.getNode(ISD::SRL, SDLoc(N0), VT,
4189 N0.getOperand(0), Amt);
4190 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), TruncVT,
4192 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N),
4193 N->getValueType(0), Trunc);
4198 // fold (sra x, (trunc (and y, c))) -> (sra x, (and (trunc y), (trunc c))).
4199 if (N1.getOpcode() == ISD::TRUNCATE &&
4200 N1.getOperand(0).getOpcode() == ISD::AND) {
4201 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4202 if (NewOp1.getNode())
4203 return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0, NewOp1);
4206 // fold (sra (trunc (srl x, c1)), c2) -> (trunc (sra x, c1 + c2))
4207 // if c1 is equal to the number of bits the trunc removes
4208 if (N0.getOpcode() == ISD::TRUNCATE &&
4209 (N0.getOperand(0).getOpcode() == ISD::SRL ||
4210 N0.getOperand(0).getOpcode() == ISD::SRA) &&
4211 N0.getOperand(0).hasOneUse() &&
4212 N0.getOperand(0).getOperand(1).hasOneUse() &&
4214 SDValue N0Op0 = N0.getOperand(0);
4215 if (ConstantSDNode *LargeShift = isConstOrConstSplat(N0Op0.getOperand(1))) {
4216 unsigned LargeShiftVal = LargeShift->getZExtValue();
4217 EVT LargeVT = N0Op0.getValueType();
4219 if (LargeVT.getScalarSizeInBits() - OpSizeInBits == LargeShiftVal) {
4221 DAG.getConstant(LargeShiftVal + N1C->getZExtValue(),
4222 getShiftAmountTy(N0Op0.getOperand(0).getValueType()));
4223 SDValue SRA = DAG.getNode(ISD::SRA, SDLoc(N), LargeVT,
4224 N0Op0.getOperand(0), Amt);
4225 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, SRA);
4230 // Simplify, based on bits shifted out of the LHS.
4231 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4232 return SDValue(N, 0);
4235 // If the sign bit is known to be zero, switch this to a SRL.
4236 if (DAG.SignBitIsZero(N0))
4237 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, N1);
4240 SDValue NewSRA = visitShiftByConstant(N, N1C);
4241 if (NewSRA.getNode())
4248 SDValue DAGCombiner::visitSRL(SDNode *N) {
4249 SDValue N0 = N->getOperand(0);
4250 SDValue N1 = N->getOperand(1);
4251 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4252 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4253 EVT VT = N0.getValueType();
4254 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
4257 if (VT.isVector()) {
4258 SDValue FoldedVOp = SimplifyVBinOp(N);
4259 if (FoldedVOp.getNode()) return FoldedVOp;
4261 N1C = isConstOrConstSplat(N1);
4264 // fold (srl c1, c2) -> c1 >>u c2
4266 return DAG.FoldConstantArithmetic(ISD::SRL, VT, N0C, N1C);
4267 // fold (srl 0, x) -> 0
4268 if (N0C && N0C->isNullValue())
4270 // fold (srl x, c >= size(x)) -> undef
4271 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4272 return DAG.getUNDEF(VT);
4273 // fold (srl x, 0) -> x
4274 if (N1C && N1C->isNullValue())
4276 // if (srl x, c) is known to be zero, return 0
4277 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
4278 APInt::getAllOnesValue(OpSizeInBits)))
4279 return DAG.getConstant(0, VT);
4281 // fold (srl (srl x, c1), c2) -> 0 or (srl x, (add c1, c2))
4282 if (N1C && N0.getOpcode() == ISD::SRL) {
4283 if (ConstantSDNode *N01C = isConstOrConstSplat(N0.getOperand(1))) {
4284 uint64_t c1 = N01C->getZExtValue();
4285 uint64_t c2 = N1C->getZExtValue();
4286 if (c1 + c2 >= OpSizeInBits)
4287 return DAG.getConstant(0, VT);
4288 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0),
4289 DAG.getConstant(c1 + c2, N1.getValueType()));
4293 // fold (srl (trunc (srl x, c1)), c2) -> 0 or (trunc (srl x, (add c1, c2)))
4294 if (N1C && N0.getOpcode() == ISD::TRUNCATE &&
4295 N0.getOperand(0).getOpcode() == ISD::SRL &&
4296 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
4298 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
4299 uint64_t c2 = N1C->getZExtValue();
4300 EVT InnerShiftVT = N0.getOperand(0).getValueType();
4301 EVT ShiftCountVT = N0.getOperand(0)->getOperand(1).getValueType();
4302 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
4303 // This is only valid if the OpSizeInBits + c1 = size of inner shift.
4304 if (c1 + OpSizeInBits == InnerShiftSize) {
4305 if (c1 + c2 >= InnerShiftSize)
4306 return DAG.getConstant(0, VT);
4307 return DAG.getNode(ISD::TRUNCATE, SDLoc(N0), VT,
4308 DAG.getNode(ISD::SRL, SDLoc(N0), InnerShiftVT,
4309 N0.getOperand(0)->getOperand(0),
4310 DAG.getConstant(c1 + c2, ShiftCountVT)));
4314 // fold (srl (shl x, c), c) -> (and x, cst2)
4315 if (N1C && N0.getOpcode() == ISD::SHL && N0.getOperand(1) == N1) {
4316 unsigned BitSize = N0.getScalarValueSizeInBits();
4317 if (BitSize <= 64) {
4318 uint64_t ShAmt = N1C->getZExtValue() + 64 - BitSize;
4319 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0.getOperand(0),
4320 DAG.getConstant(~0ULL >> ShAmt, VT));
4324 // fold (srl (anyextend x), c) -> (and (anyextend (srl x, c)), mask)
4325 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
4326 // Shifting in all undef bits?
4327 EVT SmallVT = N0.getOperand(0).getValueType();
4328 unsigned BitSize = SmallVT.getScalarSizeInBits();
4329 if (N1C->getZExtValue() >= BitSize)
4330 return DAG.getUNDEF(VT);
4332 if (!LegalTypes || TLI.isTypeDesirableForOp(ISD::SRL, SmallVT)) {
4333 uint64_t ShiftAmt = N1C->getZExtValue();
4334 SDValue SmallShift = DAG.getNode(ISD::SRL, SDLoc(N0), SmallVT,
4336 DAG.getConstant(ShiftAmt, getShiftAmountTy(SmallVT)));
4337 AddToWorkList(SmallShift.getNode());
4338 APInt Mask = APInt::getAllOnesValue(OpSizeInBits).lshr(ShiftAmt);
4339 return DAG.getNode(ISD::AND, SDLoc(N), VT,
4340 DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, SmallShift),
4341 DAG.getConstant(Mask, VT));
4345 // fold (srl (sra X, Y), 31) -> (srl X, 31). This srl only looks at the sign
4346 // bit, which is unmodified by sra.
4347 if (N1C && N1C->getZExtValue() + 1 == OpSizeInBits) {
4348 if (N0.getOpcode() == ISD::SRA)
4349 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0), N1);
4352 // fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit).
4353 if (N1C && N0.getOpcode() == ISD::CTLZ &&
4354 N1C->getAPIntValue() == Log2_32(OpSizeInBits)) {
4355 APInt KnownZero, KnownOne;
4356 DAG.computeKnownBits(N0.getOperand(0), KnownZero, KnownOne);
4358 // If any of the input bits are KnownOne, then the input couldn't be all
4359 // zeros, thus the result of the srl will always be zero.
4360 if (KnownOne.getBoolValue()) return DAG.getConstant(0, VT);
4362 // If all of the bits input the to ctlz node are known to be zero, then
4363 // the result of the ctlz is "32" and the result of the shift is one.
4364 APInt UnknownBits = ~KnownZero;
4365 if (UnknownBits == 0) return DAG.getConstant(1, VT);
4367 // Otherwise, check to see if there is exactly one bit input to the ctlz.
4368 if ((UnknownBits & (UnknownBits - 1)) == 0) {
4369 // Okay, we know that only that the single bit specified by UnknownBits
4370 // could be set on input to the CTLZ node. If this bit is set, the SRL
4371 // will return 0, if it is clear, it returns 1. Change the CTLZ/SRL pair
4372 // to an SRL/XOR pair, which is likely to simplify more.
4373 unsigned ShAmt = UnknownBits.countTrailingZeros();
4374 SDValue Op = N0.getOperand(0);
4377 Op = DAG.getNode(ISD::SRL, SDLoc(N0), VT, Op,
4378 DAG.getConstant(ShAmt, getShiftAmountTy(Op.getValueType())));
4379 AddToWorkList(Op.getNode());
4382 return DAG.getNode(ISD::XOR, SDLoc(N), VT,
4383 Op, DAG.getConstant(1, VT));
4387 // fold (srl x, (trunc (and y, c))) -> (srl x, (and (trunc y), (trunc c))).
4388 if (N1.getOpcode() == ISD::TRUNCATE &&
4389 N1.getOperand(0).getOpcode() == ISD::AND) {
4390 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4391 if (NewOp1.getNode())
4392 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, NewOp1);
4395 // fold operands of srl based on knowledge that the low bits are not
4397 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4398 return SDValue(N, 0);
4401 SDValue NewSRL = visitShiftByConstant(N, N1C);
4402 if (NewSRL.getNode())
4406 // Attempt to convert a srl of a load into a narrower zero-extending load.
4407 SDValue NarrowLoad = ReduceLoadWidth(N);
4408 if (NarrowLoad.getNode())
4411 // Here is a common situation. We want to optimize:
4414 // %b = and i32 %a, 2
4415 // %c = srl i32 %b, 1
4416 // brcond i32 %c ...
4422 // %c = setcc eq %b, 0
4425 // However when after the source operand of SRL is optimized into AND, the SRL
4426 // itself may not be optimized further. Look for it and add the BRCOND into
4428 if (N->hasOneUse()) {
4429 SDNode *Use = *N->use_begin();
4430 if (Use->getOpcode() == ISD::BRCOND)
4432 else if (Use->getOpcode() == ISD::TRUNCATE && Use->hasOneUse()) {
4433 // Also look pass the truncate.
4434 Use = *Use->use_begin();
4435 if (Use->getOpcode() == ISD::BRCOND)
4443 SDValue DAGCombiner::visitCTLZ(SDNode *N) {
4444 SDValue N0 = N->getOperand(0);
4445 EVT VT = N->getValueType(0);
4447 // fold (ctlz c1) -> c2
4448 if (isa<ConstantSDNode>(N0))
4449 return DAG.getNode(ISD::CTLZ, SDLoc(N), VT, N0);
4453 SDValue DAGCombiner::visitCTLZ_ZERO_UNDEF(SDNode *N) {
4454 SDValue N0 = N->getOperand(0);
4455 EVT VT = N->getValueType(0);
4457 // fold (ctlz_zero_undef c1) -> c2
4458 if (isa<ConstantSDNode>(N0))
4459 return DAG.getNode(ISD::CTLZ_ZERO_UNDEF, SDLoc(N), VT, N0);
4463 SDValue DAGCombiner::visitCTTZ(SDNode *N) {
4464 SDValue N0 = N->getOperand(0);
4465 EVT VT = N->getValueType(0);
4467 // fold (cttz c1) -> c2
4468 if (isa<ConstantSDNode>(N0))
4469 return DAG.getNode(ISD::CTTZ, SDLoc(N), VT, N0);
4473 SDValue DAGCombiner::visitCTTZ_ZERO_UNDEF(SDNode *N) {
4474 SDValue N0 = N->getOperand(0);
4475 EVT VT = N->getValueType(0);
4477 // fold (cttz_zero_undef c1) -> c2
4478 if (isa<ConstantSDNode>(N0))
4479 return DAG.getNode(ISD::CTTZ_ZERO_UNDEF, SDLoc(N), VT, N0);
4483 SDValue DAGCombiner::visitCTPOP(SDNode *N) {
4484 SDValue N0 = N->getOperand(0);
4485 EVT VT = N->getValueType(0);
4487 // fold (ctpop c1) -> c2
4488 if (isa<ConstantSDNode>(N0))
4489 return DAG.getNode(ISD::CTPOP, SDLoc(N), VT, N0);
4493 SDValue DAGCombiner::visitSELECT(SDNode *N) {
4494 SDValue N0 = N->getOperand(0);
4495 SDValue N1 = N->getOperand(1);
4496 SDValue N2 = N->getOperand(2);
4497 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4498 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4499 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
4500 EVT VT = N->getValueType(0);
4501 EVT VT0 = N0.getValueType();
4503 // fold (select C, X, X) -> X
4506 // fold (select true, X, Y) -> X
4507 if (N0C && !N0C->isNullValue())
4509 // fold (select false, X, Y) -> Y
4510 if (N0C && N0C->isNullValue())
4512 // fold (select C, 1, X) -> (or C, X)
4513 if (VT == MVT::i1 && N1C && N1C->getAPIntValue() == 1)
4514 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N2);
4515 // fold (select C, 0, 1) -> (xor C, 1)
4516 if (VT.isInteger() &&
4519 TLI.getBooleanContents(false) ==
4520 TargetLowering::ZeroOrOneBooleanContent)) &&
4521 N1C && N2C && N1C->isNullValue() && N2C->getAPIntValue() == 1) {
4524 return DAG.getNode(ISD::XOR, SDLoc(N), VT0,
4525 N0, DAG.getConstant(1, VT0));
4526 XORNode = DAG.getNode(ISD::XOR, SDLoc(N0), VT0,
4527 N0, DAG.getConstant(1, VT0));
4528 AddToWorkList(XORNode.getNode());
4530 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, XORNode);
4531 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, XORNode);
4533 // fold (select C, 0, X) -> (and (not C), X)
4534 if (VT == VT0 && VT == MVT::i1 && N1C && N1C->isNullValue()) {
4535 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT);
4536 AddToWorkList(NOTNode.getNode());
4537 return DAG.getNode(ISD::AND, SDLoc(N), VT, NOTNode, N2);
4539 // fold (select C, X, 1) -> (or (not C), X)
4540 if (VT == VT0 && VT == MVT::i1 && N2C && N2C->getAPIntValue() == 1) {
4541 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT);
4542 AddToWorkList(NOTNode.getNode());
4543 return DAG.getNode(ISD::OR, SDLoc(N), VT, NOTNode, N1);
4545 // fold (select C, X, 0) -> (and C, X)
4546 if (VT == MVT::i1 && N2C && N2C->isNullValue())
4547 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, N1);
4548 // fold (select X, X, Y) -> (or X, Y)
4549 // fold (select X, 1, Y) -> (or X, Y)
4550 if (VT == MVT::i1 && (N0 == N1 || (N1C && N1C->getAPIntValue() == 1)))
4551 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N2);
4552 // fold (select X, Y, X) -> (and X, Y)
4553 // fold (select X, Y, 0) -> (and X, Y)
4554 if (VT == MVT::i1 && (N0 == N2 || (N2C && N2C->getAPIntValue() == 0)))
4555 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, N1);
4557 // If we can fold this based on the true/false value, do so.
4558 if (SimplifySelectOps(N, N1, N2))
4559 return SDValue(N, 0); // Don't revisit N.
4561 // fold selects based on a setcc into other things, such as min/max/abs
4562 if (N0.getOpcode() == ISD::SETCC) {
4564 // Check against MVT::Other for SELECT_CC, which is a workaround for targets
4565 // having to say they don't support SELECT_CC on every type the DAG knows
4566 // about, since there is no way to mark an opcode illegal at all value types
4567 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, MVT::Other) &&
4568 TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT))
4569 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT,
4570 N0.getOperand(0), N0.getOperand(1),
4571 N1, N2, N0.getOperand(2));
4572 return SimplifySelect(SDLoc(N), N0, N1, N2);
4579 std::pair<SDValue, SDValue> SplitVSETCC(const SDNode *N, SelectionDAG &DAG) {
4582 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0));
4584 // Split the inputs.
4585 SDValue Lo, Hi, LL, LH, RL, RH;
4586 std::tie(LL, LH) = DAG.SplitVectorOperand(N, 0);
4587 std::tie(RL, RH) = DAG.SplitVectorOperand(N, 1);
4589 Lo = DAG.getNode(N->getOpcode(), DL, LoVT, LL, RL, N->getOperand(2));
4590 Hi = DAG.getNode(N->getOpcode(), DL, HiVT, LH, RH, N->getOperand(2));
4592 return std::make_pair(Lo, Hi);
4595 // This function assumes all the vselect's arguments are CONCAT_VECTOR
4596 // nodes and that the condition is a BV of ConstantSDNodes (or undefs).
4597 static SDValue ConvertSelectToConcatVector(SDNode *N, SelectionDAG &DAG) {
4599 SDValue Cond = N->getOperand(0);
4600 SDValue LHS = N->getOperand(1);
4601 SDValue RHS = N->getOperand(2);
4602 MVT VT = N->getSimpleValueType(0);
4603 int NumElems = VT.getVectorNumElements();
4604 assert(LHS.getOpcode() == ISD::CONCAT_VECTORS &&
4605 RHS.getOpcode() == ISD::CONCAT_VECTORS &&
4606 Cond.getOpcode() == ISD::BUILD_VECTOR);
4608 // We're sure we have an even number of elements due to the
4609 // concat_vectors we have as arguments to vselect.
4610 // Skip BV elements until we find one that's not an UNDEF
4611 // After we find an UNDEF element, keep looping until we get to half the
4612 // length of the BV and see if all the non-undef nodes are the same.
4613 ConstantSDNode *BottomHalf = nullptr;
4614 for (int i = 0; i < NumElems / 2; ++i) {
4615 if (Cond->getOperand(i)->getOpcode() == ISD::UNDEF)
4618 if (BottomHalf == nullptr)
4619 BottomHalf = cast<ConstantSDNode>(Cond.getOperand(i));
4620 else if (Cond->getOperand(i).getNode() != BottomHalf)
4624 // Do the same for the second half of the BuildVector
4625 ConstantSDNode *TopHalf = nullptr;
4626 for (int i = NumElems / 2; i < NumElems; ++i) {
4627 if (Cond->getOperand(i)->getOpcode() == ISD::UNDEF)
4630 if (TopHalf == nullptr)
4631 TopHalf = cast<ConstantSDNode>(Cond.getOperand(i));
4632 else if (Cond->getOperand(i).getNode() != TopHalf)
4636 assert(TopHalf && BottomHalf &&
4637 "One half of the selector was all UNDEFs and the other was all the "
4638 "same value. This should have been addressed before this function.");
4640 ISD::CONCAT_VECTORS, dl, VT,
4641 BottomHalf->isNullValue() ? RHS->getOperand(0) : LHS->getOperand(0),
4642 TopHalf->isNullValue() ? RHS->getOperand(1) : LHS->getOperand(1));
4645 SDValue DAGCombiner::visitVSELECT(SDNode *N) {
4646 SDValue N0 = N->getOperand(0);
4647 SDValue N1 = N->getOperand(1);
4648 SDValue N2 = N->getOperand(2);
4651 // Canonicalize integer abs.
4652 // vselect (setg[te] X, 0), X, -X ->
4653 // vselect (setgt X, -1), X, -X ->
4654 // vselect (setl[te] X, 0), -X, X ->
4655 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
4656 if (N0.getOpcode() == ISD::SETCC) {
4657 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
4658 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
4660 bool RHSIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
4662 if (((RHSIsAllZeros && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
4663 (ISD::isBuildVectorAllOnes(RHS.getNode()) && CC == ISD::SETGT)) &&
4664 N1 == LHS && N2.getOpcode() == ISD::SUB && N1 == N2.getOperand(1))
4665 isAbs = ISD::isBuildVectorAllZeros(N2.getOperand(0).getNode());
4666 else if ((RHSIsAllZeros && (CC == ISD::SETLT || CC == ISD::SETLE)) &&
4667 N2 == LHS && N1.getOpcode() == ISD::SUB && N2 == N1.getOperand(1))
4668 isAbs = ISD::isBuildVectorAllZeros(N1.getOperand(0).getNode());
4671 EVT VT = LHS.getValueType();
4672 SDValue Shift = DAG.getNode(
4673 ISD::SRA, DL, VT, LHS,
4674 DAG.getConstant(VT.getScalarType().getSizeInBits() - 1, VT));
4675 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, LHS, Shift);
4676 AddToWorkList(Shift.getNode());
4677 AddToWorkList(Add.getNode());
4678 return DAG.getNode(ISD::XOR, DL, VT, Add, Shift);
4682 // If the VSELECT result requires splitting and the mask is provided by a
4683 // SETCC, then split both nodes and its operands before legalization. This
4684 // prevents the type legalizer from unrolling SETCC into scalar comparisons
4685 // and enables future optimizations (e.g. min/max pattern matching on X86).
4686 if (N0.getOpcode() == ISD::SETCC) {
4687 EVT VT = N->getValueType(0);
4689 // Check if any splitting is required.
4690 if (TLI.getTypeAction(*DAG.getContext(), VT) !=
4691 TargetLowering::TypeSplitVector)
4694 SDValue Lo, Hi, CCLo, CCHi, LL, LH, RL, RH;
4695 std::tie(CCLo, CCHi) = SplitVSETCC(N0.getNode(), DAG);
4696 std::tie(LL, LH) = DAG.SplitVectorOperand(N, 1);
4697 std::tie(RL, RH) = DAG.SplitVectorOperand(N, 2);
4699 Lo = DAG.getNode(N->getOpcode(), DL, LL.getValueType(), CCLo, LL, RL);
4700 Hi = DAG.getNode(N->getOpcode(), DL, LH.getValueType(), CCHi, LH, RH);
4702 // Add the new VSELECT nodes to the work list in case they need to be split
4704 AddToWorkList(Lo.getNode());
4705 AddToWorkList(Hi.getNode());
4707 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
4710 // Fold (vselect (build_vector all_ones), N1, N2) -> N1
4711 if (ISD::isBuildVectorAllOnes(N0.getNode()))
4713 // Fold (vselect (build_vector all_zeros), N1, N2) -> N2
4714 if (ISD::isBuildVectorAllZeros(N0.getNode()))
4717 // The ConvertSelectToConcatVector function is assuming both the above
4718 // checks for (vselect (build_vector all{ones,zeros) ...) have been made
4720 if (N1.getOpcode() == ISD::CONCAT_VECTORS &&
4721 N2.getOpcode() == ISD::CONCAT_VECTORS &&
4722 ISD::isBuildVectorOfConstantSDNodes(N0.getNode())) {
4723 SDValue CV = ConvertSelectToConcatVector(N, DAG);
4731 SDValue DAGCombiner::visitSELECT_CC(SDNode *N) {
4732 SDValue N0 = N->getOperand(0);
4733 SDValue N1 = N->getOperand(1);
4734 SDValue N2 = N->getOperand(2);
4735 SDValue N3 = N->getOperand(3);
4736 SDValue N4 = N->getOperand(4);
4737 ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get();
4739 // fold select_cc lhs, rhs, x, x, cc -> x
4743 // Determine if the condition we're dealing with is constant
4744 SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()),
4745 N0, N1, CC, SDLoc(N), false);
4746 if (SCC.getNode()) {
4747 AddToWorkList(SCC.getNode());
4749 if (ConstantSDNode *SCCC = dyn_cast<ConstantSDNode>(SCC.getNode())) {
4750 if (!SCCC->isNullValue())
4751 return N2; // cond always true -> true val
4753 return N3; // cond always false -> false val
4756 // Fold to a simpler select_cc
4757 if (SCC.getOpcode() == ISD::SETCC)
4758 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), N2.getValueType(),
4759 SCC.getOperand(0), SCC.getOperand(1), N2, N3,
4763 // If we can fold this based on the true/false value, do so.
4764 if (SimplifySelectOps(N, N2, N3))
4765 return SDValue(N, 0); // Don't revisit N.
4767 // fold select_cc into other things, such as min/max/abs
4768 return SimplifySelectCC(SDLoc(N), N0, N1, N2, N3, CC);
4771 SDValue DAGCombiner::visitSETCC(SDNode *N) {
4772 return SimplifySetCC(N->getValueType(0), N->getOperand(0), N->getOperand(1),
4773 cast<CondCodeSDNode>(N->getOperand(2))->get(),
4777 // tryToFoldExtendOfConstant - Try to fold a sext/zext/aext
4778 // dag node into a ConstantSDNode or a build_vector of constants.
4779 // This function is called by the DAGCombiner when visiting sext/zext/aext
4780 // dag nodes (see for example method DAGCombiner::visitSIGN_EXTEND).
4781 // Vector extends are not folded if operations are legal; this is to
4782 // avoid introducing illegal build_vector dag nodes.
4783 static SDNode *tryToFoldExtendOfConstant(SDNode *N, const TargetLowering &TLI,
4784 SelectionDAG &DAG, bool LegalTypes,
4785 bool LegalOperations) {
4786 unsigned Opcode = N->getOpcode();
4787 SDValue N0 = N->getOperand(0);
4788 EVT VT = N->getValueType(0);
4790 assert((Opcode == ISD::SIGN_EXTEND || Opcode == ISD::ZERO_EXTEND ||
4791 Opcode == ISD::ANY_EXTEND) && "Expected EXTEND dag node in input!");
4793 // fold (sext c1) -> c1
4794 // fold (zext c1) -> c1
4795 // fold (aext c1) -> c1
4796 if (isa<ConstantSDNode>(N0))
4797 return DAG.getNode(Opcode, SDLoc(N), VT, N0).getNode();
4799 // fold (sext (build_vector AllConstants) -> (build_vector AllConstants)
4800 // fold (zext (build_vector AllConstants) -> (build_vector AllConstants)
4801 // fold (aext (build_vector AllConstants) -> (build_vector AllConstants)
4802 EVT SVT = VT.getScalarType();
4803 if (!(VT.isVector() &&
4804 (!LegalTypes || (!LegalOperations && TLI.isTypeLegal(SVT))) &&
4805 ISD::isBuildVectorOfConstantSDNodes(N0.getNode())))
4808 // We can fold this node into a build_vector.
4809 unsigned VTBits = SVT.getSizeInBits();
4810 unsigned EVTBits = N0->getValueType(0).getScalarType().getSizeInBits();
4811 unsigned ShAmt = VTBits - EVTBits;
4812 SmallVector<SDValue, 8> Elts;
4813 unsigned NumElts = N0->getNumOperands();
4816 for (unsigned i=0; i != NumElts; ++i) {
4817 SDValue Op = N0->getOperand(i);
4818 if (Op->getOpcode() == ISD::UNDEF) {
4819 Elts.push_back(DAG.getUNDEF(SVT));
4823 ConstantSDNode *CurrentND = cast<ConstantSDNode>(Op);
4824 const APInt &C = APInt(VTBits, CurrentND->getAPIntValue().getZExtValue());
4825 if (Opcode == ISD::SIGN_EXTEND)
4826 Elts.push_back(DAG.getConstant(C.shl(ShAmt).ashr(ShAmt).getZExtValue(),
4829 Elts.push_back(DAG.getConstant(C.shl(ShAmt).lshr(ShAmt).getZExtValue(),
4833 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Elts).getNode();
4836 // ExtendUsesToFormExtLoad - Trying to extend uses of a load to enable this:
4837 // "fold ({s|z|a}ext (load x)) -> ({s|z|a}ext (truncate ({s|z|a}extload x)))"
4838 // transformation. Returns true if extension are possible and the above
4839 // mentioned transformation is profitable.
4840 static bool ExtendUsesToFormExtLoad(SDNode *N, SDValue N0,
4842 SmallVectorImpl<SDNode *> &ExtendNodes,
4843 const TargetLowering &TLI) {
4844 bool HasCopyToRegUses = false;
4845 bool isTruncFree = TLI.isTruncateFree(N->getValueType(0), N0.getValueType());
4846 for (SDNode::use_iterator UI = N0.getNode()->use_begin(),
4847 UE = N0.getNode()->use_end();
4852 if (UI.getUse().getResNo() != N0.getResNo())
4854 // FIXME: Only extend SETCC N, N and SETCC N, c for now.
4855 if (ExtOpc != ISD::ANY_EXTEND && User->getOpcode() == ISD::SETCC) {
4856 ISD::CondCode CC = cast<CondCodeSDNode>(User->getOperand(2))->get();
4857 if (ExtOpc == ISD::ZERO_EXTEND && ISD::isSignedIntSetCC(CC))
4858 // Sign bits will be lost after a zext.
4861 for (unsigned i = 0; i != 2; ++i) {
4862 SDValue UseOp = User->getOperand(i);
4865 if (!isa<ConstantSDNode>(UseOp))
4870 ExtendNodes.push_back(User);
4873 // If truncates aren't free and there are users we can't
4874 // extend, it isn't worthwhile.
4877 // Remember if this value is live-out.
4878 if (User->getOpcode() == ISD::CopyToReg)
4879 HasCopyToRegUses = true;
4882 if (HasCopyToRegUses) {
4883 bool BothLiveOut = false;
4884 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
4886 SDUse &Use = UI.getUse();
4887 if (Use.getResNo() == 0 && Use.getUser()->getOpcode() == ISD::CopyToReg) {
4893 // Both unextended and extended values are live out. There had better be
4894 // a good reason for the transformation.
4895 return ExtendNodes.size();
4900 void DAGCombiner::ExtendSetCCUses(const SmallVectorImpl<SDNode *> &SetCCs,
4901 SDValue Trunc, SDValue ExtLoad, SDLoc DL,
4902 ISD::NodeType ExtType) {
4903 // Extend SetCC uses if necessary.
4904 for (unsigned i = 0, e = SetCCs.size(); i != e; ++i) {
4905 SDNode *SetCC = SetCCs[i];
4906 SmallVector<SDValue, 4> Ops;
4908 for (unsigned j = 0; j != 2; ++j) {
4909 SDValue SOp = SetCC->getOperand(j);
4911 Ops.push_back(ExtLoad);
4913 Ops.push_back(DAG.getNode(ExtType, DL, ExtLoad->getValueType(0), SOp));
4916 Ops.push_back(SetCC->getOperand(2));
4917 CombineTo(SetCC, DAG.getNode(ISD::SETCC, DL, SetCC->getValueType(0), Ops));
4921 SDValue DAGCombiner::visitSIGN_EXTEND(SDNode *N) {
4922 SDValue N0 = N->getOperand(0);
4923 EVT VT = N->getValueType(0);
4925 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
4927 return SDValue(Res, 0);
4929 // fold (sext (sext x)) -> (sext x)
4930 // fold (sext (aext x)) -> (sext x)
4931 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
4932 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT,
4935 if (N0.getOpcode() == ISD::TRUNCATE) {
4936 // fold (sext (truncate (load x))) -> (sext (smaller load x))
4937 // fold (sext (truncate (srl (load x), c))) -> (sext (smaller load (x+c/n)))
4938 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
4939 if (NarrowLoad.getNode()) {
4940 SDNode* oye = N0.getNode()->getOperand(0).getNode();
4941 if (NarrowLoad.getNode() != N0.getNode()) {
4942 CombineTo(N0.getNode(), NarrowLoad);
4943 // CombineTo deleted the truncate, if needed, but not what's under it.
4946 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4949 // See if the value being truncated is already sign extended. If so, just
4950 // eliminate the trunc/sext pair.
4951 SDValue Op = N0.getOperand(0);
4952 unsigned OpBits = Op.getValueType().getScalarType().getSizeInBits();
4953 unsigned MidBits = N0.getValueType().getScalarType().getSizeInBits();
4954 unsigned DestBits = VT.getScalarType().getSizeInBits();
4955 unsigned NumSignBits = DAG.ComputeNumSignBits(Op);
4957 if (OpBits == DestBits) {
4958 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
4959 // bits, it is already ready.
4960 if (NumSignBits > DestBits-MidBits)
4962 } else if (OpBits < DestBits) {
4963 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
4964 // bits, just sext from i32.
4965 if (NumSignBits > OpBits-MidBits)
4966 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, Op);
4968 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
4969 // bits, just truncate to i32.
4970 if (NumSignBits > OpBits-MidBits)
4971 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
4974 // fold (sext (truncate x)) -> (sextinreg x).
4975 if (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG,
4976 N0.getValueType())) {
4977 if (OpBits < DestBits)
4978 Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N0), VT, Op);
4979 else if (OpBits > DestBits)
4980 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), VT, Op);
4981 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, Op,
4982 DAG.getValueType(N0.getValueType()));
4986 // fold (sext (load x)) -> (sext (truncate (sextload x)))
4987 // None of the supported targets knows how to perform load and sign extend
4988 // on vectors in one instruction. We only perform this transformation on
4990 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
4991 ISD::isUNINDEXEDLoad(N0.getNode()) &&
4992 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
4993 TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()))) {
4994 bool DoXform = true;
4995 SmallVector<SDNode*, 4> SetCCs;
4996 if (!N0.hasOneUse())
4997 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::SIGN_EXTEND, SetCCs, TLI);
4999 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5000 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5002 LN0->getBasePtr(), N0.getValueType(),
5003 LN0->getMemOperand());
5004 CombineTo(N, ExtLoad);
5005 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5006 N0.getValueType(), ExtLoad);
5007 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
5008 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5010 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5014 // fold (sext (sextload x)) -> (sext (truncate (sextload x)))
5015 // fold (sext ( extload x)) -> (sext (truncate (sextload x)))
5016 if ((ISD::isSEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
5017 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
5018 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5019 EVT MemVT = LN0->getMemoryVT();
5020 if ((!LegalOperations && !LN0->isVolatile()) ||
5021 TLI.isLoadExtLegal(ISD::SEXTLOAD, MemVT)) {
5022 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5024 LN0->getBasePtr(), MemVT,
5025 LN0->getMemOperand());
5026 CombineTo(N, ExtLoad);
5027 CombineTo(N0.getNode(),
5028 DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5029 N0.getValueType(), ExtLoad),
5030 ExtLoad.getValue(1));
5031 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5035 // fold (sext (and/or/xor (load x), cst)) ->
5036 // (and/or/xor (sextload x), (sext cst))
5037 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
5038 N0.getOpcode() == ISD::XOR) &&
5039 isa<LoadSDNode>(N0.getOperand(0)) &&
5040 N0.getOperand(1).getOpcode() == ISD::Constant &&
5041 TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()) &&
5042 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
5043 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
5044 if (LN0->getExtensionType() != ISD::ZEXTLOAD && LN0->isUnindexed()) {
5045 bool DoXform = true;
5046 SmallVector<SDNode*, 4> SetCCs;
5047 if (!N0.hasOneUse())
5048 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::SIGN_EXTEND,
5051 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(LN0), VT,
5052 LN0->getChain(), LN0->getBasePtr(),
5054 LN0->getMemOperand());
5055 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5056 Mask = Mask.sext(VT.getSizeInBits());
5057 SDValue And = DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
5058 ExtLoad, DAG.getConstant(Mask, VT));
5059 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
5060 SDLoc(N0.getOperand(0)),
5061 N0.getOperand(0).getValueType(), ExtLoad);
5063 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
5064 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5066 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5071 if (N0.getOpcode() == ISD::SETCC) {
5072 // sext(setcc) -> sext_in_reg(vsetcc) for vectors.
5073 // Only do this before legalize for now.
5074 if (VT.isVector() && !LegalOperations &&
5075 TLI.getBooleanContents(true) ==
5076 TargetLowering::ZeroOrNegativeOneBooleanContent) {
5077 EVT N0VT = N0.getOperand(0).getValueType();
5078 // On some architectures (such as SSE/NEON/etc) the SETCC result type is
5079 // of the same size as the compared operands. Only optimize sext(setcc())
5080 // if this is the case.
5081 EVT SVT = getSetCCResultType(N0VT);
5083 // We know that the # elements of the results is the same as the
5084 // # elements of the compare (and the # elements of the compare result
5085 // for that matter). Check to see that they are the same size. If so,
5086 // we know that the element size of the sext'd result matches the
5087 // element size of the compare operands.
5088 if (VT.getSizeInBits() == SVT.getSizeInBits())
5089 return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
5091 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5093 // If the desired elements are smaller or larger than the source
5094 // elements we can use a matching integer vector type and then
5095 // truncate/sign extend
5096 EVT MatchingVectorType = N0VT.changeVectorElementTypeToInteger();
5097 if (SVT == MatchingVectorType) {
5098 SDValue VsetCC = DAG.getSetCC(SDLoc(N), MatchingVectorType,
5099 N0.getOperand(0), N0.getOperand(1),
5100 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5101 return DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT);
5105 // sext(setcc x, y, cc) -> (select (setcc x, y, cc), -1, 0)
5106 unsigned ElementWidth = VT.getScalarType().getSizeInBits();
5108 DAG.getConstant(APInt::getAllOnesValue(ElementWidth), VT);
5110 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
5111 NegOne, DAG.getConstant(0, VT),
5112 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
5113 if (SCC.getNode()) return SCC;
5115 if (!VT.isVector()) {
5116 EVT SetCCVT = getSetCCResultType(N0.getOperand(0).getValueType());
5117 if (!LegalOperations || TLI.isOperationLegal(ISD::SETCC, SetCCVT)) {
5119 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
5120 SDValue SetCC = DAG.getSetCC(DL,
5122 N0.getOperand(0), N0.getOperand(1), CC);
5123 EVT SelectVT = getSetCCResultType(VT);
5124 return DAG.getSelect(DL, VT,
5125 DAG.getSExtOrTrunc(SetCC, DL, SelectVT),
5126 NegOne, DAG.getConstant(0, VT));
5132 // fold (sext x) -> (zext x) if the sign bit is known zero.
5133 if ((!LegalOperations || TLI.isOperationLegal(ISD::ZERO_EXTEND, VT)) &&
5134 DAG.SignBitIsZero(N0))
5135 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, N0);
5140 // isTruncateOf - If N is a truncate of some other value, return true, record
5141 // the value being truncated in Op and which of Op's bits are zero in KnownZero.
5142 // This function computes KnownZero to avoid a duplicated call to
5143 // computeKnownBits in the caller.
5144 static bool isTruncateOf(SelectionDAG &DAG, SDValue N, SDValue &Op,
5147 if (N->getOpcode() == ISD::TRUNCATE) {
5148 Op = N->getOperand(0);
5149 DAG.computeKnownBits(Op, KnownZero, KnownOne);
5153 if (N->getOpcode() != ISD::SETCC || N->getValueType(0) != MVT::i1 ||
5154 cast<CondCodeSDNode>(N->getOperand(2))->get() != ISD::SETNE)
5157 SDValue Op0 = N->getOperand(0);
5158 SDValue Op1 = N->getOperand(1);
5159 assert(Op0.getValueType() == Op1.getValueType());
5161 ConstantSDNode *COp0 = dyn_cast<ConstantSDNode>(Op0);
5162 ConstantSDNode *COp1 = dyn_cast<ConstantSDNode>(Op1);
5163 if (COp0 && COp0->isNullValue())
5165 else if (COp1 && COp1->isNullValue())
5170 DAG.computeKnownBits(Op, KnownZero, KnownOne);
5172 if (!(KnownZero | APInt(Op.getValueSizeInBits(), 1)).isAllOnesValue())
5178 SDValue DAGCombiner::visitZERO_EXTEND(SDNode *N) {
5179 SDValue N0 = N->getOperand(0);
5180 EVT VT = N->getValueType(0);
5182 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
5184 return SDValue(Res, 0);
5186 // fold (zext (zext x)) -> (zext x)
5187 // fold (zext (aext x)) -> (zext x)
5188 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
5189 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT,
5192 // fold (zext (truncate x)) -> (zext x) or
5193 // (zext (truncate x)) -> (truncate x)
5194 // This is valid when the truncated bits of x are already zero.
5195 // FIXME: We should extend this to work for vectors too.
5198 if (!VT.isVector() && isTruncateOf(DAG, N0, Op, KnownZero)) {
5199 APInt TruncatedBits =
5200 (Op.getValueSizeInBits() == N0.getValueSizeInBits()) ?
5201 APInt(Op.getValueSizeInBits(), 0) :
5202 APInt::getBitsSet(Op.getValueSizeInBits(),
5203 N0.getValueSizeInBits(),
5204 std::min(Op.getValueSizeInBits(),
5205 VT.getSizeInBits()));
5206 if (TruncatedBits == (KnownZero & TruncatedBits)) {
5207 if (VT.bitsGT(Op.getValueType()))
5208 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, Op);
5209 if (VT.bitsLT(Op.getValueType()))
5210 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
5216 // fold (zext (truncate (load x))) -> (zext (smaller load x))
5217 // fold (zext (truncate (srl (load x), c))) -> (zext (small load (x+c/n)))
5218 if (N0.getOpcode() == ISD::TRUNCATE) {
5219 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5220 if (NarrowLoad.getNode()) {
5221 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5222 if (NarrowLoad.getNode() != N0.getNode()) {
5223 CombineTo(N0.getNode(), NarrowLoad);
5224 // CombineTo deleted the truncate, if needed, but not what's under it.
5227 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5231 // fold (zext (truncate x)) -> (and x, mask)
5232 if (N0.getOpcode() == ISD::TRUNCATE &&
5233 (!LegalOperations || TLI.isOperationLegal(ISD::AND, VT))) {
5235 // fold (zext (truncate (load x))) -> (zext (smaller load x))
5236 // fold (zext (truncate (srl (load x), c))) -> (zext (smaller load (x+c/n)))
5237 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5238 if (NarrowLoad.getNode()) {
5239 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5240 if (NarrowLoad.getNode() != N0.getNode()) {
5241 CombineTo(N0.getNode(), NarrowLoad);
5242 // CombineTo deleted the truncate, if needed, but not what's under it.
5245 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5248 SDValue Op = N0.getOperand(0);
5249 if (Op.getValueType().bitsLT(VT)) {
5250 Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, Op);
5251 AddToWorkList(Op.getNode());
5252 } else if (Op.getValueType().bitsGT(VT)) {
5253 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
5254 AddToWorkList(Op.getNode());
5256 return DAG.getZeroExtendInReg(Op, SDLoc(N),
5257 N0.getValueType().getScalarType());
5260 // Fold (zext (and (trunc x), cst)) -> (and x, cst),
5261 // if either of the casts is not free.
5262 if (N0.getOpcode() == ISD::AND &&
5263 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
5264 N0.getOperand(1).getOpcode() == ISD::Constant &&
5265 (!TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
5266 N0.getValueType()) ||
5267 !TLI.isZExtFree(N0.getValueType(), VT))) {
5268 SDValue X = N0.getOperand(0).getOperand(0);
5269 if (X.getValueType().bitsLT(VT)) {
5270 X = DAG.getNode(ISD::ANY_EXTEND, SDLoc(X), VT, X);
5271 } else if (X.getValueType().bitsGT(VT)) {
5272 X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X);
5274 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5275 Mask = Mask.zext(VT.getSizeInBits());
5276 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5277 X, DAG.getConstant(Mask, VT));
5280 // fold (zext (load x)) -> (zext (truncate (zextload x)))
5281 // None of the supported targets knows how to perform load and vector_zext
5282 // on vectors in one instruction. We only perform this transformation on
5284 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
5285 ISD::isUNINDEXEDLoad(N0.getNode()) &&
5286 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5287 TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()))) {
5288 bool DoXform = true;
5289 SmallVector<SDNode*, 4> SetCCs;
5290 if (!N0.hasOneUse())
5291 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ZERO_EXTEND, SetCCs, TLI);
5293 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5294 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT,
5296 LN0->getBasePtr(), N0.getValueType(),
5297 LN0->getMemOperand());
5298 CombineTo(N, ExtLoad);
5299 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5300 N0.getValueType(), ExtLoad);
5301 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
5303 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5305 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5309 // fold (zext (and/or/xor (load x), cst)) ->
5310 // (and/or/xor (zextload x), (zext cst))
5311 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
5312 N0.getOpcode() == ISD::XOR) &&
5313 isa<LoadSDNode>(N0.getOperand(0)) &&
5314 N0.getOperand(1).getOpcode() == ISD::Constant &&
5315 TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()) &&
5316 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
5317 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
5318 if (LN0->getExtensionType() != ISD::SEXTLOAD && LN0->isUnindexed()) {
5319 bool DoXform = true;
5320 SmallVector<SDNode*, 4> SetCCs;
5321 if (!N0.hasOneUse())
5322 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::ZERO_EXTEND,
5325 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), VT,
5326 LN0->getChain(), LN0->getBasePtr(),
5328 LN0->getMemOperand());
5329 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5330 Mask = Mask.zext(VT.getSizeInBits());
5331 SDValue And = DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
5332 ExtLoad, DAG.getConstant(Mask, VT));
5333 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
5334 SDLoc(N0.getOperand(0)),
5335 N0.getOperand(0).getValueType(), ExtLoad);
5337 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
5338 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5340 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5345 // fold (zext (zextload x)) -> (zext (truncate (zextload x)))
5346 // fold (zext ( extload x)) -> (zext (truncate (zextload x)))
5347 if ((ISD::isZEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
5348 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
5349 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5350 EVT MemVT = LN0->getMemoryVT();
5351 if ((!LegalOperations && !LN0->isVolatile()) ||
5352 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT)) {
5353 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT,
5355 LN0->getBasePtr(), MemVT,
5356 LN0->getMemOperand());
5357 CombineTo(N, ExtLoad);
5358 CombineTo(N0.getNode(),
5359 DAG.getNode(ISD::TRUNCATE, SDLoc(N0), N0.getValueType(),
5361 ExtLoad.getValue(1));
5362 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5366 if (N0.getOpcode() == ISD::SETCC) {
5367 if (!LegalOperations && VT.isVector() &&
5368 N0.getValueType().getVectorElementType() == MVT::i1) {
5369 EVT N0VT = N0.getOperand(0).getValueType();
5370 if (getSetCCResultType(N0VT) == N0.getValueType())
5373 // zext(setcc) -> (and (vsetcc), (1, 1, ...) for vectors.
5374 // Only do this before legalize for now.
5375 EVT EltVT = VT.getVectorElementType();
5376 SmallVector<SDValue,8> OneOps(VT.getVectorNumElements(),
5377 DAG.getConstant(1, EltVT));
5378 if (VT.getSizeInBits() == N0VT.getSizeInBits())
5379 // We know that the # elements of the results is the same as the
5380 // # elements of the compare (and the # elements of the compare result
5381 // for that matter). Check to see that they are the same size. If so,
5382 // we know that the element size of the sext'd result matches the
5383 // element size of the compare operands.
5384 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5385 DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
5387 cast<CondCodeSDNode>(N0.getOperand(2))->get()),
5388 DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT,
5391 // If the desired elements are smaller or larger than the source
5392 // elements we can use a matching integer vector type and then
5393 // truncate/sign extend
5394 EVT MatchingElementType =
5395 EVT::getIntegerVT(*DAG.getContext(),
5396 N0VT.getScalarType().getSizeInBits());
5397 EVT MatchingVectorType =
5398 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
5399 N0VT.getVectorNumElements());
5401 DAG.getSetCC(SDLoc(N), MatchingVectorType, N0.getOperand(0),
5403 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5404 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5405 DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT),
5406 DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, OneOps));
5409 // zext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
5411 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
5412 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
5413 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
5414 if (SCC.getNode()) return SCC;
5417 // (zext (shl (zext x), cst)) -> (shl (zext x), cst)
5418 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL) &&
5419 isa<ConstantSDNode>(N0.getOperand(1)) &&
5420 N0.getOperand(0).getOpcode() == ISD::ZERO_EXTEND &&
5422 SDValue ShAmt = N0.getOperand(1);
5423 unsigned ShAmtVal = cast<ConstantSDNode>(ShAmt)->getZExtValue();
5424 if (N0.getOpcode() == ISD::SHL) {
5425 SDValue InnerZExt = N0.getOperand(0);
5426 // If the original shl may be shifting out bits, do not perform this
5428 unsigned KnownZeroBits = InnerZExt.getValueType().getSizeInBits() -
5429 InnerZExt.getOperand(0).getValueType().getSizeInBits();
5430 if (ShAmtVal > KnownZeroBits)
5436 // Ensure that the shift amount is wide enough for the shifted value.
5437 if (VT.getSizeInBits() >= 256)
5438 ShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, ShAmt);
5440 return DAG.getNode(N0.getOpcode(), DL, VT,
5441 DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0)),
5448 SDValue DAGCombiner::visitANY_EXTEND(SDNode *N) {
5449 SDValue N0 = N->getOperand(0);
5450 EVT VT = N->getValueType(0);
5452 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
5454 return SDValue(Res, 0);
5456 // fold (aext (aext x)) -> (aext x)
5457 // fold (aext (zext x)) -> (zext x)
5458 // fold (aext (sext x)) -> (sext x)
5459 if (N0.getOpcode() == ISD::ANY_EXTEND ||
5460 N0.getOpcode() == ISD::ZERO_EXTEND ||
5461 N0.getOpcode() == ISD::SIGN_EXTEND)
5462 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, N0.getOperand(0));
5464 // fold (aext (truncate (load x))) -> (aext (smaller load x))
5465 // fold (aext (truncate (srl (load x), c))) -> (aext (small load (x+c/n)))
5466 if (N0.getOpcode() == ISD::TRUNCATE) {
5467 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5468 if (NarrowLoad.getNode()) {
5469 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5470 if (NarrowLoad.getNode() != N0.getNode()) {
5471 CombineTo(N0.getNode(), NarrowLoad);
5472 // CombineTo deleted the truncate, if needed, but not what's under it.
5475 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5479 // fold (aext (truncate x))
5480 if (N0.getOpcode() == ISD::TRUNCATE) {
5481 SDValue TruncOp = N0.getOperand(0);
5482 if (TruncOp.getValueType() == VT)
5483 return TruncOp; // x iff x size == zext size.
5484 if (TruncOp.getValueType().bitsGT(VT))
5485 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, TruncOp);
5486 return DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, TruncOp);
5489 // Fold (aext (and (trunc x), cst)) -> (and x, cst)
5490 // if the trunc is not free.
5491 if (N0.getOpcode() == ISD::AND &&
5492 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
5493 N0.getOperand(1).getOpcode() == ISD::Constant &&
5494 !TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
5495 N0.getValueType())) {
5496 SDValue X = N0.getOperand(0).getOperand(0);
5497 if (X.getValueType().bitsLT(VT)) {
5498 X = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, X);
5499 } else if (X.getValueType().bitsGT(VT)) {
5500 X = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, X);
5502 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5503 Mask = Mask.zext(VT.getSizeInBits());
5504 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5505 X, DAG.getConstant(Mask, VT));
5508 // fold (aext (load x)) -> (aext (truncate (extload x)))
5509 // None of the supported targets knows how to perform load and any_ext
5510 // on vectors in one instruction. We only perform this transformation on
5512 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
5513 ISD::isUNINDEXEDLoad(N0.getNode()) &&
5514 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5515 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType()))) {
5516 bool DoXform = true;
5517 SmallVector<SDNode*, 4> SetCCs;
5518 if (!N0.hasOneUse())
5519 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ANY_EXTEND, SetCCs, TLI);
5521 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5522 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT,
5524 LN0->getBasePtr(), N0.getValueType(),
5525 LN0->getMemOperand());
5526 CombineTo(N, ExtLoad);
5527 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5528 N0.getValueType(), ExtLoad);
5529 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
5530 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5532 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5536 // fold (aext (zextload x)) -> (aext (truncate (zextload x)))
5537 // fold (aext (sextload x)) -> (aext (truncate (sextload x)))
5538 // fold (aext ( extload x)) -> (aext (truncate (extload x)))
5539 if (N0.getOpcode() == ISD::LOAD &&
5540 !ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
5542 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5543 ISD::LoadExtType ExtType = LN0->getExtensionType();
5544 EVT MemVT = LN0->getMemoryVT();
5545 if (!LegalOperations || TLI.isLoadExtLegal(ExtType, MemVT)) {
5546 SDValue ExtLoad = DAG.getExtLoad(ExtType, SDLoc(N),
5547 VT, LN0->getChain(), LN0->getBasePtr(),
5548 MemVT, LN0->getMemOperand());
5549 CombineTo(N, ExtLoad);
5550 CombineTo(N0.getNode(),
5551 DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5552 N0.getValueType(), ExtLoad),
5553 ExtLoad.getValue(1));
5554 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5558 if (N0.getOpcode() == ISD::SETCC) {
5560 // aext(setcc) -> vsetcc
5561 // aext(setcc) -> truncate(vsetcc)
5562 // aext(setcc) -> aext(vsetcc)
5563 // Only do this before legalize for now.
5564 if (VT.isVector() && !LegalOperations) {
5565 EVT N0VT = N0.getOperand(0).getValueType();
5566 // We know that the # elements of the results is the same as the
5567 // # elements of the compare (and the # elements of the compare result
5568 // for that matter). Check to see that they are the same size. If so,
5569 // we know that the element size of the sext'd result matches the
5570 // element size of the compare operands.
5571 if (VT.getSizeInBits() == N0VT.getSizeInBits())
5572 return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
5574 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5575 // If the desired elements are smaller or larger than the source
5576 // elements we can use a matching integer vector type and then
5577 // truncate/any extend
5579 EVT MatchingVectorType = N0VT.changeVectorElementTypeToInteger();
5581 DAG.getSetCC(SDLoc(N), MatchingVectorType, N0.getOperand(0),
5583 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5584 return DAG.getAnyExtOrTrunc(VsetCC, SDLoc(N), VT);
5588 // aext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
5590 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
5591 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
5592 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
5600 /// GetDemandedBits - See if the specified operand can be simplified with the
5601 /// knowledge that only the bits specified by Mask are used. If so, return the
5602 /// simpler operand, otherwise return a null SDValue.
5603 SDValue DAGCombiner::GetDemandedBits(SDValue V, const APInt &Mask) {
5604 switch (V.getOpcode()) {
5606 case ISD::Constant: {
5607 const ConstantSDNode *CV = cast<ConstantSDNode>(V.getNode());
5608 assert(CV && "Const value should be ConstSDNode.");
5609 const APInt &CVal = CV->getAPIntValue();
5610 APInt NewVal = CVal & Mask;
5612 return DAG.getConstant(NewVal, V.getValueType());
5617 // If the LHS or RHS don't contribute bits to the or, drop them.
5618 if (DAG.MaskedValueIsZero(V.getOperand(0), Mask))
5619 return V.getOperand(1);
5620 if (DAG.MaskedValueIsZero(V.getOperand(1), Mask))
5621 return V.getOperand(0);
5624 // Only look at single-use SRLs.
5625 if (!V.getNode()->hasOneUse())
5627 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(V.getOperand(1))) {
5628 // See if we can recursively simplify the LHS.
5629 unsigned Amt = RHSC->getZExtValue();
5631 // Watch out for shift count overflow though.
5632 if (Amt >= Mask.getBitWidth()) break;
5633 APInt NewMask = Mask << Amt;
5634 SDValue SimplifyLHS = GetDemandedBits(V.getOperand(0), NewMask);
5635 if (SimplifyLHS.getNode())
5636 return DAG.getNode(ISD::SRL, SDLoc(V), V.getValueType(),
5637 SimplifyLHS, V.getOperand(1));
5643 /// ReduceLoadWidth - If the result of a wider load is shifted to right of N
5644 /// bits and then truncated to a narrower type and where N is a multiple
5645 /// of number of bits of the narrower type, transform it to a narrower load
5646 /// from address + N / num of bits of new type. If the result is to be
5647 /// extended, also fold the extension to form a extending load.
5648 SDValue DAGCombiner::ReduceLoadWidth(SDNode *N) {
5649 unsigned Opc = N->getOpcode();
5651 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
5652 SDValue N0 = N->getOperand(0);
5653 EVT VT = N->getValueType(0);
5656 // This transformation isn't valid for vector loads.
5660 // Special case: SIGN_EXTEND_INREG is basically truncating to ExtVT then
5662 if (Opc == ISD::SIGN_EXTEND_INREG) {
5663 ExtType = ISD::SEXTLOAD;
5664 ExtVT = cast<VTSDNode>(N->getOperand(1))->getVT();
5665 } else if (Opc == ISD::SRL) {
5666 // Another special-case: SRL is basically zero-extending a narrower value.
5667 ExtType = ISD::ZEXTLOAD;
5669 ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1));
5670 if (!N01) return SDValue();
5671 ExtVT = EVT::getIntegerVT(*DAG.getContext(),
5672 VT.getSizeInBits() - N01->getZExtValue());
5674 if (LegalOperations && !TLI.isLoadExtLegal(ExtType, ExtVT))
5677 unsigned EVTBits = ExtVT.getSizeInBits();
5679 // Do not generate loads of non-round integer types since these can
5680 // be expensive (and would be wrong if the type is not byte sized).
5681 if (!ExtVT.isRound())
5685 if (N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
5686 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
5687 ShAmt = N01->getZExtValue();
5688 // Is the shift amount a multiple of size of VT?
5689 if ((ShAmt & (EVTBits-1)) == 0) {
5690 N0 = N0.getOperand(0);
5691 // Is the load width a multiple of size of VT?
5692 if ((N0.getValueType().getSizeInBits() & (EVTBits-1)) != 0)
5696 // At this point, we must have a load or else we can't do the transform.
5697 if (!isa<LoadSDNode>(N0)) return SDValue();
5699 // Because a SRL must be assumed to *need* to zero-extend the high bits
5700 // (as opposed to anyext the high bits), we can't combine the zextload
5701 // lowering of SRL and an sextload.
5702 if (cast<LoadSDNode>(N0)->getExtensionType() == ISD::SEXTLOAD)
5705 // If the shift amount is larger than the input type then we're not
5706 // accessing any of the loaded bytes. If the load was a zextload/extload
5707 // then the result of the shift+trunc is zero/undef (handled elsewhere).
5708 if (ShAmt >= cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits())
5713 // If the load is shifted left (and the result isn't shifted back right),
5714 // we can fold the truncate through the shift.
5715 unsigned ShLeftAmt = 0;
5716 if (ShAmt == 0 && N0.getOpcode() == ISD::SHL && N0.hasOneUse() &&
5717 ExtVT == VT && TLI.isNarrowingProfitable(N0.getValueType(), VT)) {
5718 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
5719 ShLeftAmt = N01->getZExtValue();
5720 N0 = N0.getOperand(0);
5724 // If we haven't found a load, we can't narrow it. Don't transform one with
5725 // multiple uses, this would require adding a new load.
5726 if (!isa<LoadSDNode>(N0) || !N0.hasOneUse())
5729 // Don't change the width of a volatile load.
5730 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5731 if (LN0->isVolatile())
5734 // Verify that we are actually reducing a load width here.
5735 if (LN0->getMemoryVT().getSizeInBits() < EVTBits)
5738 // For the transform to be legal, the load must produce only two values
5739 // (the value loaded and the chain). Don't transform a pre-increment
5740 // load, for example, which produces an extra value. Otherwise the
5741 // transformation is not equivalent, and the downstream logic to replace
5742 // uses gets things wrong.
5743 if (LN0->getNumValues() > 2)
5746 // If the load that we're shrinking is an extload and we're not just
5747 // discarding the extension we can't simply shrink the load. Bail.
5748 // TODO: It would be possible to merge the extensions in some cases.
5749 if (LN0->getExtensionType() != ISD::NON_EXTLOAD &&
5750 LN0->getMemoryVT().getSizeInBits() < ExtVT.getSizeInBits() + ShAmt)
5753 EVT PtrType = N0.getOperand(1).getValueType();
5755 if (PtrType == MVT::Untyped || PtrType.isExtended())
5756 // It's not possible to generate a constant of extended or untyped type.
5759 // For big endian targets, we need to adjust the offset to the pointer to
5760 // load the correct bytes.
5761 if (TLI.isBigEndian()) {
5762 unsigned LVTStoreBits = LN0->getMemoryVT().getStoreSizeInBits();
5763 unsigned EVTStoreBits = ExtVT.getStoreSizeInBits();
5764 ShAmt = LVTStoreBits - EVTStoreBits - ShAmt;
5767 uint64_t PtrOff = ShAmt / 8;
5768 unsigned NewAlign = MinAlign(LN0->getAlignment(), PtrOff);
5769 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(LN0),
5770 PtrType, LN0->getBasePtr(),
5771 DAG.getConstant(PtrOff, PtrType));
5772 AddToWorkList(NewPtr.getNode());
5775 if (ExtType == ISD::NON_EXTLOAD)
5776 Load = DAG.getLoad(VT, SDLoc(N0), LN0->getChain(), NewPtr,
5777 LN0->getPointerInfo().getWithOffset(PtrOff),
5778 LN0->isVolatile(), LN0->isNonTemporal(),
5779 LN0->isInvariant(), NewAlign, LN0->getTBAAInfo());
5781 Load = DAG.getExtLoad(ExtType, SDLoc(N0), VT, LN0->getChain(),NewPtr,
5782 LN0->getPointerInfo().getWithOffset(PtrOff),
5783 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
5784 NewAlign, LN0->getTBAAInfo());
5786 // Replace the old load's chain with the new load's chain.
5787 WorkListRemover DeadNodes(*this);
5788 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1));
5790 // Shift the result left, if we've swallowed a left shift.
5791 SDValue Result = Load;
5792 if (ShLeftAmt != 0) {
5793 EVT ShImmTy = getShiftAmountTy(Result.getValueType());
5794 if (!isUIntN(ShImmTy.getSizeInBits(), ShLeftAmt))
5796 // If the shift amount is as large as the result size (but, presumably,
5797 // no larger than the source) then the useful bits of the result are
5798 // zero; we can't simply return the shortened shift, because the result
5799 // of that operation is undefined.
5800 if (ShLeftAmt >= VT.getSizeInBits())
5801 Result = DAG.getConstant(0, VT);
5803 Result = DAG.getNode(ISD::SHL, SDLoc(N0), VT,
5804 Result, DAG.getConstant(ShLeftAmt, ShImmTy));
5807 // Return the new loaded value.
5811 SDValue DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
5812 SDValue N0 = N->getOperand(0);
5813 SDValue N1 = N->getOperand(1);
5814 EVT VT = N->getValueType(0);
5815 EVT EVT = cast<VTSDNode>(N1)->getVT();
5816 unsigned VTBits = VT.getScalarType().getSizeInBits();
5817 unsigned EVTBits = EVT.getScalarType().getSizeInBits();
5819 // fold (sext_in_reg c1) -> c1
5820 if (isa<ConstantSDNode>(N0) || N0.getOpcode() == ISD::UNDEF)
5821 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, N0, N1);
5823 // If the input is already sign extended, just drop the extension.
5824 if (DAG.ComputeNumSignBits(N0) >= VTBits-EVTBits+1)
5827 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2
5828 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
5829 EVT.bitsLT(cast<VTSDNode>(N0.getOperand(1))->getVT()))
5830 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
5831 N0.getOperand(0), N1);
5833 // fold (sext_in_reg (sext x)) -> (sext x)
5834 // fold (sext_in_reg (aext x)) -> (sext x)
5835 // if x is small enough.
5836 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) {
5837 SDValue N00 = N0.getOperand(0);
5838 if (N00.getValueType().getScalarType().getSizeInBits() <= EVTBits &&
5839 (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND, VT)))
5840 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, N00, N1);
5843 // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is known zero.
5844 if (DAG.MaskedValueIsZero(N0, APInt::getBitsSet(VTBits, EVTBits-1, EVTBits)))
5845 return DAG.getZeroExtendInReg(N0, SDLoc(N), EVT);
5847 // fold operands of sext_in_reg based on knowledge that the top bits are not
5849 if (SimplifyDemandedBits(SDValue(N, 0)))
5850 return SDValue(N, 0);
5852 // fold (sext_in_reg (load x)) -> (smaller sextload x)
5853 // fold (sext_in_reg (srl (load x), c)) -> (smaller sextload (x+c/evtbits))
5854 SDValue NarrowLoad = ReduceLoadWidth(N);
5855 if (NarrowLoad.getNode())
5858 // fold (sext_in_reg (srl X, 24), i8) -> (sra X, 24)
5859 // fold (sext_in_reg (srl X, 23), i8) -> (sra X, 23) iff possible.
5860 // We already fold "(sext_in_reg (srl X, 25), i8) -> srl X, 25" above.
5861 if (N0.getOpcode() == ISD::SRL) {
5862 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
5863 if (ShAmt->getZExtValue()+EVTBits <= VTBits) {
5864 // We can turn this into an SRA iff the input to the SRL is already sign
5866 unsigned InSignBits = DAG.ComputeNumSignBits(N0.getOperand(0));
5867 if (VTBits-(ShAmt->getZExtValue()+EVTBits) < InSignBits)
5868 return DAG.getNode(ISD::SRA, SDLoc(N), VT,
5869 N0.getOperand(0), N0.getOperand(1));
5873 // fold (sext_inreg (extload x)) -> (sextload x)
5874 if (ISD::isEXTLoad(N0.getNode()) &&
5875 ISD::isUNINDEXEDLoad(N0.getNode()) &&
5876 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
5877 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5878 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) {
5879 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5880 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5882 LN0->getBasePtr(), EVT,
5883 LN0->getMemOperand());
5884 CombineTo(N, ExtLoad);
5885 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
5886 AddToWorkList(ExtLoad.getNode());
5887 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5889 // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use
5890 if (ISD::isZEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
5892 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
5893 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5894 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) {
5895 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5896 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5898 LN0->getBasePtr(), EVT,
5899 LN0->getMemOperand());
5900 CombineTo(N, ExtLoad);
5901 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
5902 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5905 // Form (sext_inreg (bswap >> 16)) or (sext_inreg (rotl (bswap) 16))
5906 if (EVTBits <= 16 && N0.getOpcode() == ISD::OR) {
5907 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
5908 N0.getOperand(1), false);
5909 if (BSwap.getNode())
5910 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
5914 // Fold a sext_inreg of a build_vector of ConstantSDNodes or undefs
5915 // into a build_vector.
5916 if (ISD::isBuildVectorOfConstantSDNodes(N0.getNode())) {
5917 SmallVector<SDValue, 8> Elts;
5918 unsigned NumElts = N0->getNumOperands();
5919 unsigned ShAmt = VTBits - EVTBits;
5921 for (unsigned i = 0; i != NumElts; ++i) {
5922 SDValue Op = N0->getOperand(i);
5923 if (Op->getOpcode() == ISD::UNDEF) {
5928 ConstantSDNode *CurrentND = cast<ConstantSDNode>(Op);
5929 const APInt &C = APInt(VTBits, CurrentND->getAPIntValue().getZExtValue());
5930 Elts.push_back(DAG.getConstant(C.shl(ShAmt).ashr(ShAmt).getZExtValue(),
5931 Op.getValueType()));
5934 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Elts);
5940 SDValue DAGCombiner::visitTRUNCATE(SDNode *N) {
5941 SDValue N0 = N->getOperand(0);
5942 EVT VT = N->getValueType(0);
5943 bool isLE = TLI.isLittleEndian();
5946 if (N0.getValueType() == N->getValueType(0))
5948 // fold (truncate c1) -> c1
5949 if (isa<ConstantSDNode>(N0))
5950 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0);
5951 // fold (truncate (truncate x)) -> (truncate x)
5952 if (N0.getOpcode() == ISD::TRUNCATE)
5953 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0));
5954 // fold (truncate (ext x)) -> (ext x) or (truncate x) or x
5955 if (N0.getOpcode() == ISD::ZERO_EXTEND ||
5956 N0.getOpcode() == ISD::SIGN_EXTEND ||
5957 N0.getOpcode() == ISD::ANY_EXTEND) {
5958 if (N0.getOperand(0).getValueType().bitsLT(VT))
5959 // if the source is smaller than the dest, we still need an extend
5960 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
5962 if (N0.getOperand(0).getValueType().bitsGT(VT))
5963 // if the source is larger than the dest, than we just need the truncate
5964 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0));
5965 // if the source and dest are the same type, we can drop both the extend
5966 // and the truncate.
5967 return N0.getOperand(0);
5970 // Fold extract-and-trunc into a narrow extract. For example:
5971 // i64 x = EXTRACT_VECTOR_ELT(v2i64 val, i32 1)
5972 // i32 y = TRUNCATE(i64 x)
5974 // v16i8 b = BITCAST (v2i64 val)
5975 // i8 x = EXTRACT_VECTOR_ELT(v16i8 b, i32 8)
5977 // Note: We only run this optimization after type legalization (which often
5978 // creates this pattern) and before operation legalization after which
5979 // we need to be more careful about the vector instructions that we generate.
5980 if (N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5981 LegalTypes && !LegalOperations && N0->hasOneUse() && VT != MVT::i1) {
5983 EVT VecTy = N0.getOperand(0).getValueType();
5984 EVT ExTy = N0.getValueType();
5985 EVT TrTy = N->getValueType(0);
5987 unsigned NumElem = VecTy.getVectorNumElements();
5988 unsigned SizeRatio = ExTy.getSizeInBits()/TrTy.getSizeInBits();
5990 EVT NVT = EVT::getVectorVT(*DAG.getContext(), TrTy, SizeRatio * NumElem);
5991 assert(NVT.getSizeInBits() == VecTy.getSizeInBits() && "Invalid Size");
5993 SDValue EltNo = N0->getOperand(1);
5994 if (isa<ConstantSDNode>(EltNo) && isTypeLegal(NVT)) {
5995 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
5996 EVT IndexTy = TLI.getVectorIdxTy();
5997 int Index = isLE ? (Elt*SizeRatio) : (Elt*SizeRatio + (SizeRatio-1));
5999 SDValue V = DAG.getNode(ISD::BITCAST, SDLoc(N),
6000 NVT, N0.getOperand(0));
6002 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
6004 DAG.getConstant(Index, IndexTy));
6008 // Fold a series of buildvector, bitcast, and truncate if possible.
6010 // (2xi32 trunc (bitcast ((4xi32)buildvector x, x, y, y) 2xi64)) to
6011 // (2xi32 (buildvector x, y)).
6012 if (Level == AfterLegalizeVectorOps && VT.isVector() &&
6013 N0.getOpcode() == ISD::BITCAST && N0.hasOneUse() &&
6014 N0.getOperand(0).getOpcode() == ISD::BUILD_VECTOR &&
6015 N0.getOperand(0).hasOneUse()) {
6017 SDValue BuildVect = N0.getOperand(0);
6018 EVT BuildVectEltTy = BuildVect.getValueType().getVectorElementType();
6019 EVT TruncVecEltTy = VT.getVectorElementType();
6021 // Check that the element types match.
6022 if (BuildVectEltTy == TruncVecEltTy) {
6023 // Now we only need to compute the offset of the truncated elements.
6024 unsigned BuildVecNumElts = BuildVect.getNumOperands();
6025 unsigned TruncVecNumElts = VT.getVectorNumElements();
6026 unsigned TruncEltOffset = BuildVecNumElts / TruncVecNumElts;
6028 assert((BuildVecNumElts % TruncVecNumElts) == 0 &&
6029 "Invalid number of elements");
6031 SmallVector<SDValue, 8> Opnds;
6032 for (unsigned i = 0, e = BuildVecNumElts; i != e; i += TruncEltOffset)
6033 Opnds.push_back(BuildVect.getOperand(i));
6035 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Opnds);
6039 // See if we can simplify the input to this truncate through knowledge that
6040 // only the low bits are being used.
6041 // For example "trunc (or (shl x, 8), y)" // -> trunc y
6042 // Currently we only perform this optimization on scalars because vectors
6043 // may have different active low bits.
6044 if (!VT.isVector()) {
6046 GetDemandedBits(N0, APInt::getLowBitsSet(N0.getValueSizeInBits(),
6047 VT.getSizeInBits()));
6048 if (Shorter.getNode())
6049 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Shorter);
6051 // fold (truncate (load x)) -> (smaller load x)
6052 // fold (truncate (srl (load x), c)) -> (smaller load (x+c/evtbits))
6053 if (!LegalTypes || TLI.isTypeDesirableForOp(N0.getOpcode(), VT)) {
6054 SDValue Reduced = ReduceLoadWidth(N);
6055 if (Reduced.getNode())
6057 // Handle the case where the load remains an extending load even
6058 // after truncation.
6059 if (N0.hasOneUse() && ISD::isUNINDEXEDLoad(N0.getNode())) {
6060 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6061 if (!LN0->isVolatile() &&
6062 LN0->getMemoryVT().getStoreSizeInBits() < VT.getSizeInBits()) {
6063 SDValue NewLoad = DAG.getExtLoad(LN0->getExtensionType(), SDLoc(LN0),
6064 VT, LN0->getChain(), LN0->getBasePtr(),
6066 LN0->getMemOperand());
6067 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLoad.getValue(1));
6072 // fold (trunc (concat ... x ...)) -> (concat ..., (trunc x), ...)),
6073 // where ... are all 'undef'.
6074 if (N0.getOpcode() == ISD::CONCAT_VECTORS && !LegalTypes) {
6075 SmallVector<EVT, 8> VTs;
6078 unsigned NumDefs = 0;
6080 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) {
6081 SDValue X = N0.getOperand(i);
6082 if (X.getOpcode() != ISD::UNDEF) {
6087 // Stop if more than one members are non-undef.
6090 VTs.push_back(EVT::getVectorVT(*DAG.getContext(),
6091 VT.getVectorElementType(),
6092 X.getValueType().getVectorNumElements()));
6096 return DAG.getUNDEF(VT);
6099 assert(V.getNode() && "The single defined operand is empty!");
6100 SmallVector<SDValue, 8> Opnds;
6101 for (unsigned i = 0, e = VTs.size(); i != e; ++i) {
6103 Opnds.push_back(DAG.getUNDEF(VTs[i]));
6106 SDValue NV = DAG.getNode(ISD::TRUNCATE, SDLoc(V), VTs[i], V);
6107 AddToWorkList(NV.getNode());
6108 Opnds.push_back(NV);
6110 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Opnds);
6114 // Simplify the operands using demanded-bits information.
6115 if (!VT.isVector() &&
6116 SimplifyDemandedBits(SDValue(N, 0)))
6117 return SDValue(N, 0);
6122 static SDNode *getBuildPairElt(SDNode *N, unsigned i) {
6123 SDValue Elt = N->getOperand(i);
6124 if (Elt.getOpcode() != ISD::MERGE_VALUES)
6125 return Elt.getNode();
6126 return Elt.getOperand(Elt.getResNo()).getNode();
6129 /// CombineConsecutiveLoads - build_pair (load, load) -> load
6130 /// if load locations are consecutive.
6131 SDValue DAGCombiner::CombineConsecutiveLoads(SDNode *N, EVT VT) {
6132 assert(N->getOpcode() == ISD::BUILD_PAIR);
6134 LoadSDNode *LD1 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 0));
6135 LoadSDNode *LD2 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 1));
6136 if (!LD1 || !LD2 || !ISD::isNON_EXTLoad(LD1) || !LD1->hasOneUse() ||
6137 LD1->getAddressSpace() != LD2->getAddressSpace())
6139 EVT LD1VT = LD1->getValueType(0);
6141 if (ISD::isNON_EXTLoad(LD2) &&
6143 // If both are volatile this would reduce the number of volatile loads.
6144 // If one is volatile it might be ok, but play conservative and bail out.
6145 !LD1->isVolatile() &&
6146 !LD2->isVolatile() &&
6147 DAG.isConsecutiveLoad(LD2, LD1, LD1VT.getSizeInBits()/8, 1)) {
6148 unsigned Align = LD1->getAlignment();
6149 unsigned NewAlign = TLI.getDataLayout()->
6150 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
6152 if (NewAlign <= Align &&
6153 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)))
6154 return DAG.getLoad(VT, SDLoc(N), LD1->getChain(),
6155 LD1->getBasePtr(), LD1->getPointerInfo(),
6156 false, false, false, Align);
6162 SDValue DAGCombiner::visitBITCAST(SDNode *N) {
6163 SDValue N0 = N->getOperand(0);
6164 EVT VT = N->getValueType(0);
6166 // If the input is a BUILD_VECTOR with all constant elements, fold this now.
6167 // Only do this before legalize, since afterward the target may be depending
6168 // on the bitconvert.
6169 // First check to see if this is all constant.
6171 N0.getOpcode() == ISD::BUILD_VECTOR && N0.getNode()->hasOneUse() &&
6173 bool isSimple = cast<BuildVectorSDNode>(N0)->isConstant();
6175 EVT DestEltVT = N->getValueType(0).getVectorElementType();
6176 assert(!DestEltVT.isVector() &&
6177 "Element type of vector ValueType must not be vector!");
6179 return ConstantFoldBITCASTofBUILD_VECTOR(N0.getNode(), DestEltVT);
6182 // If the input is a constant, let getNode fold it.
6183 if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) {
6184 SDValue Res = DAG.getNode(ISD::BITCAST, SDLoc(N), VT, N0);
6185 if (Res.getNode() != N) {
6186 if (!LegalOperations ||
6187 TLI.isOperationLegal(Res.getNode()->getOpcode(), VT))
6190 // Folding it resulted in an illegal node, and it's too late to
6191 // do that. Clean up the old node and forego the transformation.
6192 // Ideally this won't happen very often, because instcombine
6193 // and the earlier dagcombine runs (where illegal nodes are
6194 // permitted) should have folded most of them already.
6195 DAG.DeleteNode(Res.getNode());
6199 // (conv (conv x, t1), t2) -> (conv x, t2)
6200 if (N0.getOpcode() == ISD::BITCAST)
6201 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT,
6204 // fold (conv (load x)) -> (load (conv*)x)
6205 // If the resultant load doesn't need a higher alignment than the original!
6206 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
6207 // Do not change the width of a volatile load.
6208 !cast<LoadSDNode>(N0)->isVolatile() &&
6209 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)) &&
6210 TLI.isLoadBitCastBeneficial(N0.getValueType(), VT)) {
6211 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6212 unsigned Align = TLI.getDataLayout()->
6213 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
6214 unsigned OrigAlign = LN0->getAlignment();
6216 if (Align <= OrigAlign) {
6217 SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(),
6218 LN0->getBasePtr(), LN0->getPointerInfo(),
6219 LN0->isVolatile(), LN0->isNonTemporal(),
6220 LN0->isInvariant(), OrigAlign,
6221 LN0->getTBAAInfo());
6223 CombineTo(N0.getNode(),
6224 DAG.getNode(ISD::BITCAST, SDLoc(N0),
6225 N0.getValueType(), Load),
6231 // fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit)
6232 // fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit))
6233 // This often reduces constant pool loads.
6234 if (((N0.getOpcode() == ISD::FNEG && !TLI.isFNegFree(N0.getValueType())) ||
6235 (N0.getOpcode() == ISD::FABS && !TLI.isFAbsFree(N0.getValueType()))) &&
6236 N0.getNode()->hasOneUse() && VT.isInteger() &&
6237 !VT.isVector() && !N0.getValueType().isVector()) {
6238 SDValue NewConv = DAG.getNode(ISD::BITCAST, SDLoc(N0), VT,
6240 AddToWorkList(NewConv.getNode());
6242 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
6243 if (N0.getOpcode() == ISD::FNEG)
6244 return DAG.getNode(ISD::XOR, SDLoc(N), VT,
6245 NewConv, DAG.getConstant(SignBit, VT));
6246 assert(N0.getOpcode() == ISD::FABS);
6247 return DAG.getNode(ISD::AND, SDLoc(N), VT,
6248 NewConv, DAG.getConstant(~SignBit, VT));
6251 // fold (bitconvert (fcopysign cst, x)) ->
6252 // (or (and (bitconvert x), sign), (and cst, (not sign)))
6253 // Note that we don't handle (copysign x, cst) because this can always be
6254 // folded to an fneg or fabs.
6255 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse() &&
6256 isa<ConstantFPSDNode>(N0.getOperand(0)) &&
6257 VT.isInteger() && !VT.isVector()) {
6258 unsigned OrigXWidth = N0.getOperand(1).getValueType().getSizeInBits();
6259 EVT IntXVT = EVT::getIntegerVT(*DAG.getContext(), OrigXWidth);
6260 if (isTypeLegal(IntXVT)) {
6261 SDValue X = DAG.getNode(ISD::BITCAST, SDLoc(N0),
6262 IntXVT, N0.getOperand(1));
6263 AddToWorkList(X.getNode());
6265 // If X has a different width than the result/lhs, sext it or truncate it.
6266 unsigned VTWidth = VT.getSizeInBits();
6267 if (OrigXWidth < VTWidth) {
6268 X = DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, X);
6269 AddToWorkList(X.getNode());
6270 } else if (OrigXWidth > VTWidth) {
6271 // To get the sign bit in the right place, we have to shift it right
6272 // before truncating.
6273 X = DAG.getNode(ISD::SRL, SDLoc(X),
6274 X.getValueType(), X,
6275 DAG.getConstant(OrigXWidth-VTWidth, X.getValueType()));
6276 AddToWorkList(X.getNode());
6277 X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X);
6278 AddToWorkList(X.getNode());
6281 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
6282 X = DAG.getNode(ISD::AND, SDLoc(X), VT,
6283 X, DAG.getConstant(SignBit, VT));
6284 AddToWorkList(X.getNode());
6286 SDValue Cst = DAG.getNode(ISD::BITCAST, SDLoc(N0),
6287 VT, N0.getOperand(0));
6288 Cst = DAG.getNode(ISD::AND, SDLoc(Cst), VT,
6289 Cst, DAG.getConstant(~SignBit, VT));
6290 AddToWorkList(Cst.getNode());
6292 return DAG.getNode(ISD::OR, SDLoc(N), VT, X, Cst);
6296 // bitconvert(build_pair(ld, ld)) -> ld iff load locations are consecutive.
6297 if (N0.getOpcode() == ISD::BUILD_PAIR) {
6298 SDValue CombineLD = CombineConsecutiveLoads(N0.getNode(), VT);
6299 if (CombineLD.getNode())
6306 SDValue DAGCombiner::visitBUILD_PAIR(SDNode *N) {
6307 EVT VT = N->getValueType(0);
6308 return CombineConsecutiveLoads(N, VT);
6311 /// ConstantFoldBITCASTofBUILD_VECTOR - We know that BV is a build_vector
6312 /// node with Constant, ConstantFP or Undef operands. DstEltVT indicates the
6313 /// destination element value type.
6314 SDValue DAGCombiner::
6315 ConstantFoldBITCASTofBUILD_VECTOR(SDNode *BV, EVT DstEltVT) {
6316 EVT SrcEltVT = BV->getValueType(0).getVectorElementType();
6318 // If this is already the right type, we're done.
6319 if (SrcEltVT == DstEltVT) return SDValue(BV, 0);
6321 unsigned SrcBitSize = SrcEltVT.getSizeInBits();
6322 unsigned DstBitSize = DstEltVT.getSizeInBits();
6324 // If this is a conversion of N elements of one type to N elements of another
6325 // type, convert each element. This handles FP<->INT cases.
6326 if (SrcBitSize == DstBitSize) {
6327 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
6328 BV->getValueType(0).getVectorNumElements());
6330 // Due to the FP element handling below calling this routine recursively,
6331 // we can end up with a scalar-to-vector node here.
6332 if (BV->getOpcode() == ISD::SCALAR_TO_VECTOR)
6333 return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(BV), VT,
6334 DAG.getNode(ISD::BITCAST, SDLoc(BV),
6335 DstEltVT, BV->getOperand(0)));
6337 SmallVector<SDValue, 8> Ops;
6338 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
6339 SDValue Op = BV->getOperand(i);
6340 // If the vector element type is not legal, the BUILD_VECTOR operands
6341 // are promoted and implicitly truncated. Make that explicit here.
6342 if (Op.getValueType() != SrcEltVT)
6343 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(BV), SrcEltVT, Op);
6344 Ops.push_back(DAG.getNode(ISD::BITCAST, SDLoc(BV),
6346 AddToWorkList(Ops.back().getNode());
6348 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
6351 // Otherwise, we're growing or shrinking the elements. To avoid having to
6352 // handle annoying details of growing/shrinking FP values, we convert them to
6354 if (SrcEltVT.isFloatingPoint()) {
6355 // Convert the input float vector to a int vector where the elements are the
6357 assert((SrcEltVT == MVT::f32 || SrcEltVT == MVT::f64) && "Unknown FP VT!");
6358 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), SrcEltVT.getSizeInBits());
6359 BV = ConstantFoldBITCASTofBUILD_VECTOR(BV, IntVT).getNode();
6363 // Now we know the input is an integer vector. If the output is a FP type,
6364 // convert to integer first, then to FP of the right size.
6365 if (DstEltVT.isFloatingPoint()) {
6366 assert((DstEltVT == MVT::f32 || DstEltVT == MVT::f64) && "Unknown FP VT!");
6367 EVT TmpVT = EVT::getIntegerVT(*DAG.getContext(), DstEltVT.getSizeInBits());
6368 SDNode *Tmp = ConstantFoldBITCASTofBUILD_VECTOR(BV, TmpVT).getNode();
6370 // Next, convert to FP elements of the same size.
6371 return ConstantFoldBITCASTofBUILD_VECTOR(Tmp, DstEltVT);
6374 // Okay, we know the src/dst types are both integers of differing types.
6375 // Handling growing first.
6376 assert(SrcEltVT.isInteger() && DstEltVT.isInteger());
6377 if (SrcBitSize < DstBitSize) {
6378 unsigned NumInputsPerOutput = DstBitSize/SrcBitSize;
6380 SmallVector<SDValue, 8> Ops;
6381 for (unsigned i = 0, e = BV->getNumOperands(); i != e;
6382 i += NumInputsPerOutput) {
6383 bool isLE = TLI.isLittleEndian();
6384 APInt NewBits = APInt(DstBitSize, 0);
6385 bool EltIsUndef = true;
6386 for (unsigned j = 0; j != NumInputsPerOutput; ++j) {
6387 // Shift the previously computed bits over.
6388 NewBits <<= SrcBitSize;
6389 SDValue Op = BV->getOperand(i+ (isLE ? (NumInputsPerOutput-j-1) : j));
6390 if (Op.getOpcode() == ISD::UNDEF) continue;
6393 NewBits |= cast<ConstantSDNode>(Op)->getAPIntValue().
6394 zextOrTrunc(SrcBitSize).zext(DstBitSize);
6398 Ops.push_back(DAG.getUNDEF(DstEltVT));
6400 Ops.push_back(DAG.getConstant(NewBits, DstEltVT));
6403 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, Ops.size());
6404 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
6407 // Finally, this must be the case where we are shrinking elements: each input
6408 // turns into multiple outputs.
6409 bool isS2V = ISD::isScalarToVector(BV);
6410 unsigned NumOutputsPerInput = SrcBitSize/DstBitSize;
6411 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
6412 NumOutputsPerInput*BV->getNumOperands());
6413 SmallVector<SDValue, 8> Ops;
6415 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
6416 if (BV->getOperand(i).getOpcode() == ISD::UNDEF) {
6417 for (unsigned j = 0; j != NumOutputsPerInput; ++j)
6418 Ops.push_back(DAG.getUNDEF(DstEltVT));
6422 APInt OpVal = cast<ConstantSDNode>(BV->getOperand(i))->
6423 getAPIntValue().zextOrTrunc(SrcBitSize);
6425 for (unsigned j = 0; j != NumOutputsPerInput; ++j) {
6426 APInt ThisVal = OpVal.trunc(DstBitSize);
6427 Ops.push_back(DAG.getConstant(ThisVal, DstEltVT));
6428 if (isS2V && i == 0 && j == 0 && ThisVal.zext(SrcBitSize) == OpVal)
6429 // Simply turn this into a SCALAR_TO_VECTOR of the new type.
6430 return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(BV), VT,
6432 OpVal = OpVal.lshr(DstBitSize);
6435 // For big endian targets, swap the order of the pieces of each element.
6436 if (TLI.isBigEndian())
6437 std::reverse(Ops.end()-NumOutputsPerInput, Ops.end());
6440 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
6443 SDValue DAGCombiner::visitFADD(SDNode *N) {
6444 SDValue N0 = N->getOperand(0);
6445 SDValue N1 = N->getOperand(1);
6446 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6447 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6448 EVT VT = N->getValueType(0);
6451 if (VT.isVector()) {
6452 SDValue FoldedVOp = SimplifyVBinOp(N);
6453 if (FoldedVOp.getNode()) return FoldedVOp;
6456 // fold (fadd c1, c2) -> c1 + c2
6458 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N1);
6459 // canonicalize constant to RHS
6460 if (N0CFP && !N1CFP)
6461 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N1, N0);
6462 // fold (fadd A, 0) -> A
6463 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
6464 N1CFP->getValueAPF().isZero())
6466 // fold (fadd A, (fneg B)) -> (fsub A, B)
6467 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
6468 isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options) == 2)
6469 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N0,
6470 GetNegatedExpression(N1, DAG, LegalOperations));
6471 // fold (fadd (fneg A), B) -> (fsub B, A)
6472 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
6473 isNegatibleForFree(N0, LegalOperations, TLI, &DAG.getTarget().Options) == 2)
6474 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N1,
6475 GetNegatedExpression(N0, DAG, LegalOperations));
6477 // If allowed, fold (fadd (fadd x, c1), c2) -> (fadd x, (fadd c1, c2))
6478 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
6479 N0.getOpcode() == ISD::FADD && N0.getNode()->hasOneUse() &&
6480 isa<ConstantFPSDNode>(N0.getOperand(1)))
6481 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0.getOperand(0),
6482 DAG.getNode(ISD::FADD, SDLoc(N), VT,
6483 N0.getOperand(1), N1));
6485 // No FP constant should be created after legalization as Instruction
6486 // Selection pass has hard time in dealing with FP constant.
6488 // We don't need test this condition for transformation like following, as
6489 // the DAG being transformed implies it is legal to take FP constant as
6492 // (fadd (fmul c, x), x) -> (fmul c+1, x)
6494 bool AllowNewFpConst = (Level < AfterLegalizeDAG);
6496 // If allow, fold (fadd (fneg x), x) -> 0.0
6497 if (AllowNewFpConst && DAG.getTarget().Options.UnsafeFPMath &&
6498 N0.getOpcode() == ISD::FNEG && N0.getOperand(0) == N1)
6499 return DAG.getConstantFP(0.0, VT);
6501 // If allow, fold (fadd x, (fneg x)) -> 0.0
6502 if (AllowNewFpConst && DAG.getTarget().Options.UnsafeFPMath &&
6503 N1.getOpcode() == ISD::FNEG && N1.getOperand(0) == N0)
6504 return DAG.getConstantFP(0.0, VT);
6506 // In unsafe math mode, we can fold chains of FADD's of the same value
6507 // into multiplications. This transform is not safe in general because
6508 // we are reducing the number of rounding steps.
6509 if (DAG.getTarget().Options.UnsafeFPMath &&
6510 TLI.isOperationLegalOrCustom(ISD::FMUL, VT) &&
6512 if (N0.getOpcode() == ISD::FMUL) {
6513 ConstantFPSDNode *CFP00 = dyn_cast<ConstantFPSDNode>(N0.getOperand(0));
6514 ConstantFPSDNode *CFP01 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
6516 // (fadd (fmul c, x), x) -> (fmul x, c+1)
6517 if (CFP00 && !CFP01 && N0.getOperand(1) == N1) {
6518 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6520 DAG.getConstantFP(1.0, VT));
6521 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6525 // (fadd (fmul x, c), x) -> (fmul x, c+1)
6526 if (CFP01 && !CFP00 && N0.getOperand(0) == N1) {
6527 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6529 DAG.getConstantFP(1.0, VT));
6530 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6534 // (fadd (fmul c, x), (fadd x, x)) -> (fmul x, c+2)
6535 if (CFP00 && !CFP01 && N1.getOpcode() == ISD::FADD &&
6536 N1.getOperand(0) == N1.getOperand(1) &&
6537 N0.getOperand(1) == N1.getOperand(0)) {
6538 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6540 DAG.getConstantFP(2.0, VT));
6541 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6542 N0.getOperand(1), NewCFP);
6545 // (fadd (fmul x, c), (fadd x, x)) -> (fmul x, c+2)
6546 if (CFP01 && !CFP00 && N1.getOpcode() == ISD::FADD &&
6547 N1.getOperand(0) == N1.getOperand(1) &&
6548 N0.getOperand(0) == N1.getOperand(0)) {
6549 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6551 DAG.getConstantFP(2.0, VT));
6552 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6553 N0.getOperand(0), NewCFP);
6557 if (N1.getOpcode() == ISD::FMUL) {
6558 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0));
6559 ConstantFPSDNode *CFP11 = dyn_cast<ConstantFPSDNode>(N1.getOperand(1));
6561 // (fadd x, (fmul c, x)) -> (fmul x, c+1)
6562 if (CFP10 && !CFP11 && N1.getOperand(1) == N0) {
6563 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6565 DAG.getConstantFP(1.0, VT));
6566 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6570 // (fadd x, (fmul x, c)) -> (fmul x, c+1)
6571 if (CFP11 && !CFP10 && N1.getOperand(0) == N0) {
6572 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6574 DAG.getConstantFP(1.0, VT));
6575 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6580 // (fadd (fadd x, x), (fmul c, x)) -> (fmul x, c+2)
6581 if (CFP10 && !CFP11 && N0.getOpcode() == ISD::FADD &&
6582 N0.getOperand(0) == N0.getOperand(1) &&
6583 N1.getOperand(1) == N0.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 N1.getOperand(1), NewCFP);
6591 // (fadd (fadd x, x), (fmul x, c)) -> (fmul x, c+2)
6592 if (CFP11 && !CFP10 && N0.getOpcode() == ISD::FADD &&
6593 N0.getOperand(0) == N0.getOperand(1) &&
6594 N1.getOperand(0) == N0.getOperand(0)) {
6595 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6597 DAG.getConstantFP(2.0, VT));
6598 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6599 N1.getOperand(0), NewCFP);
6603 if (N0.getOpcode() == ISD::FADD && AllowNewFpConst) {
6604 ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N0.getOperand(0));
6605 // (fadd (fadd x, x), x) -> (fmul x, 3.0)
6606 if (!CFP && N0.getOperand(0) == N0.getOperand(1) &&
6607 (N0.getOperand(0) == N1))
6608 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6609 N1, DAG.getConstantFP(3.0, VT));
6612 if (N1.getOpcode() == ISD::FADD && AllowNewFpConst) {
6613 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0));
6614 // (fadd x, (fadd x, x)) -> (fmul x, 3.0)
6615 if (!CFP10 && N1.getOperand(0) == N1.getOperand(1) &&
6616 N1.getOperand(0) == N0)
6617 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6618 N0, DAG.getConstantFP(3.0, VT));
6621 // (fadd (fadd x, x), (fadd x, x)) -> (fmul x, 4.0)
6622 if (AllowNewFpConst &&
6623 N0.getOpcode() == ISD::FADD && N1.getOpcode() == ISD::FADD &&
6624 N0.getOperand(0) == N0.getOperand(1) &&
6625 N1.getOperand(0) == N1.getOperand(1) &&
6626 N0.getOperand(0) == N1.getOperand(0))
6627 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6629 DAG.getConstantFP(4.0, VT));
6632 // FADD -> FMA combines:
6633 if ((DAG.getTarget().Options.AllowFPOpFusion == FPOpFusion::Fast ||
6634 DAG.getTarget().Options.UnsafeFPMath) &&
6635 DAG.getTarget().getTargetLowering()->isFMAFasterThanFMulAndFAdd(VT) &&
6636 (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FMA, VT))) {
6638 // fold (fadd (fmul x, y), z) -> (fma x, y, z)
6639 if (N0.getOpcode() == ISD::FMUL && N0->hasOneUse())
6640 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
6641 N0.getOperand(0), N0.getOperand(1), N1);
6643 // fold (fadd x, (fmul y, z)) -> (fma y, z, x)
6644 // Note: Commutes FADD operands.
6645 if (N1.getOpcode() == ISD::FMUL && N1->hasOneUse())
6646 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
6647 N1.getOperand(0), N1.getOperand(1), N0);
6653 SDValue DAGCombiner::visitFSUB(SDNode *N) {
6654 SDValue N0 = N->getOperand(0);
6655 SDValue N1 = N->getOperand(1);
6656 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6657 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6658 EVT VT = N->getValueType(0);
6662 if (VT.isVector()) {
6663 SDValue FoldedVOp = SimplifyVBinOp(N);
6664 if (FoldedVOp.getNode()) return FoldedVOp;
6667 // fold (fsub c1, c2) -> c1-c2
6669 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N0, N1);
6670 // fold (fsub A, 0) -> A
6671 if (DAG.getTarget().Options.UnsafeFPMath &&
6672 N1CFP && N1CFP->getValueAPF().isZero())
6674 // fold (fsub 0, B) -> -B
6675 if (DAG.getTarget().Options.UnsafeFPMath &&
6676 N0CFP && N0CFP->getValueAPF().isZero()) {
6677 if (isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options))
6678 return GetNegatedExpression(N1, DAG, LegalOperations);
6679 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
6680 return DAG.getNode(ISD::FNEG, dl, VT, N1);
6682 // fold (fsub A, (fneg B)) -> (fadd A, B)
6683 if (isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options))
6684 return DAG.getNode(ISD::FADD, dl, VT, N0,
6685 GetNegatedExpression(N1, DAG, LegalOperations));
6687 // If 'unsafe math' is enabled, fold
6688 // (fsub x, x) -> 0.0 &
6689 // (fsub x, (fadd x, y)) -> (fneg y) &
6690 // (fsub x, (fadd y, x)) -> (fneg y)
6691 if (DAG.getTarget().Options.UnsafeFPMath) {
6693 return DAG.getConstantFP(0.0f, VT);
6695 if (N1.getOpcode() == ISD::FADD) {
6696 SDValue N10 = N1->getOperand(0);
6697 SDValue N11 = N1->getOperand(1);
6699 if (N10 == N0 && isNegatibleForFree(N11, LegalOperations, TLI,
6700 &DAG.getTarget().Options))
6701 return GetNegatedExpression(N11, DAG, LegalOperations);
6703 if (N11 == N0 && isNegatibleForFree(N10, LegalOperations, TLI,
6704 &DAG.getTarget().Options))
6705 return GetNegatedExpression(N10, DAG, LegalOperations);
6709 // FSUB -> FMA combines:
6710 if ((DAG.getTarget().Options.AllowFPOpFusion == FPOpFusion::Fast ||
6711 DAG.getTarget().Options.UnsafeFPMath) &&
6712 DAG.getTarget().getTargetLowering()->isFMAFasterThanFMulAndFAdd(VT) &&
6713 (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FMA, VT))) {
6715 // fold (fsub (fmul x, y), z) -> (fma x, y, (fneg z))
6716 if (N0.getOpcode() == ISD::FMUL && N0->hasOneUse())
6717 return DAG.getNode(ISD::FMA, dl, VT,
6718 N0.getOperand(0), N0.getOperand(1),
6719 DAG.getNode(ISD::FNEG, dl, VT, N1));
6721 // fold (fsub x, (fmul y, z)) -> (fma (fneg y), z, x)
6722 // Note: Commutes FSUB operands.
6723 if (N1.getOpcode() == ISD::FMUL && N1->hasOneUse())
6724 return DAG.getNode(ISD::FMA, dl, VT,
6725 DAG.getNode(ISD::FNEG, dl, VT,
6727 N1.getOperand(1), N0);
6729 // fold (fsub (fneg (fmul, x, y)), z) -> (fma (fneg x), y, (fneg z))
6730 if (N0.getOpcode() == ISD::FNEG &&
6731 N0.getOperand(0).getOpcode() == ISD::FMUL &&
6732 N0->hasOneUse() && N0.getOperand(0).hasOneUse()) {
6733 SDValue N00 = N0.getOperand(0).getOperand(0);
6734 SDValue N01 = N0.getOperand(0).getOperand(1);
6735 return DAG.getNode(ISD::FMA, dl, VT,
6736 DAG.getNode(ISD::FNEG, dl, VT, N00), N01,
6737 DAG.getNode(ISD::FNEG, dl, VT, N1));
6744 SDValue DAGCombiner::visitFMUL(SDNode *N) {
6745 SDValue N0 = N->getOperand(0);
6746 SDValue N1 = N->getOperand(1);
6747 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6748 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6749 EVT VT = N->getValueType(0);
6750 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6753 if (VT.isVector()) {
6754 SDValue FoldedVOp = SimplifyVBinOp(N);
6755 if (FoldedVOp.getNode()) return FoldedVOp;
6758 // fold (fmul c1, c2) -> c1*c2
6760 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0, N1);
6761 // canonicalize constant to RHS
6762 if (N0CFP && !N1CFP)
6763 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N1, N0);
6764 // fold (fmul A, 0) -> 0
6765 if (DAG.getTarget().Options.UnsafeFPMath &&
6766 N1CFP && N1CFP->getValueAPF().isZero())
6768 // fold (fmul A, 0) -> 0, vector edition.
6769 if (DAG.getTarget().Options.UnsafeFPMath &&
6770 ISD::isBuildVectorAllZeros(N1.getNode()))
6772 // fold (fmul A, 1.0) -> A
6773 if (N1CFP && N1CFP->isExactlyValue(1.0))
6775 // fold (fmul X, 2.0) -> (fadd X, X)
6776 if (N1CFP && N1CFP->isExactlyValue(+2.0))
6777 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N0);
6778 // fold (fmul X, -1.0) -> (fneg X)
6779 if (N1CFP && N1CFP->isExactlyValue(-1.0))
6780 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
6781 return DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0);
6783 // fold (fmul (fneg X), (fneg Y)) -> (fmul X, Y)
6784 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI,
6785 &DAG.getTarget().Options)) {
6786 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI,
6787 &DAG.getTarget().Options)) {
6788 // Both can be negated for free, check to see if at least one is cheaper
6790 if (LHSNeg == 2 || RHSNeg == 2)
6791 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6792 GetNegatedExpression(N0, DAG, LegalOperations),
6793 GetNegatedExpression(N1, DAG, LegalOperations));
6797 // If allowed, fold (fmul (fmul x, c1), c2) -> (fmul x, (fmul c1, c2))
6798 if (DAG.getTarget().Options.UnsafeFPMath &&
6799 N1CFP && N0.getOpcode() == ISD::FMUL &&
6800 N0.getNode()->hasOneUse() && isa<ConstantFPSDNode>(N0.getOperand(1)))
6801 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0.getOperand(0),
6802 DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6803 N0.getOperand(1), N1));
6808 SDValue DAGCombiner::visitFMA(SDNode *N) {
6809 SDValue N0 = N->getOperand(0);
6810 SDValue N1 = N->getOperand(1);
6811 SDValue N2 = N->getOperand(2);
6812 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6813 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6814 EVT VT = N->getValueType(0);
6817 if (DAG.getTarget().Options.UnsafeFPMath) {
6818 if (N0CFP && N0CFP->isZero())
6820 if (N1CFP && N1CFP->isZero())
6823 if (N0CFP && N0CFP->isExactlyValue(1.0))
6824 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N1, N2);
6825 if (N1CFP && N1CFP->isExactlyValue(1.0))
6826 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N2);
6828 // Canonicalize (fma c, x, y) -> (fma x, c, y)
6829 if (N0CFP && !N1CFP)
6830 return DAG.getNode(ISD::FMA, SDLoc(N), VT, N1, N0, N2);
6832 // (fma x, c1, (fmul x, c2)) -> (fmul x, c1+c2)
6833 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
6834 N2.getOpcode() == ISD::FMUL &&
6835 N0 == N2.getOperand(0) &&
6836 N2.getOperand(1).getOpcode() == ISD::ConstantFP) {
6837 return DAG.getNode(ISD::FMUL, dl, VT, N0,
6838 DAG.getNode(ISD::FADD, dl, VT, N1, N2.getOperand(1)));
6842 // (fma (fmul x, c1), c2, y) -> (fma x, c1*c2, y)
6843 if (DAG.getTarget().Options.UnsafeFPMath &&
6844 N0.getOpcode() == ISD::FMUL && N1CFP &&
6845 N0.getOperand(1).getOpcode() == ISD::ConstantFP) {
6846 return DAG.getNode(ISD::FMA, dl, VT,
6848 DAG.getNode(ISD::FMUL, dl, VT, N1, N0.getOperand(1)),
6852 // (fma x, 1, y) -> (fadd x, y)
6853 // (fma x, -1, y) -> (fadd (fneg x), y)
6855 if (N1CFP->isExactlyValue(1.0))
6856 return DAG.getNode(ISD::FADD, dl, VT, N0, N2);
6858 if (N1CFP->isExactlyValue(-1.0) &&
6859 (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))) {
6860 SDValue RHSNeg = DAG.getNode(ISD::FNEG, dl, VT, N0);
6861 AddToWorkList(RHSNeg.getNode());
6862 return DAG.getNode(ISD::FADD, dl, VT, N2, RHSNeg);
6866 // (fma x, c, x) -> (fmul x, (c+1))
6867 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP && N0 == N2)
6868 return DAG.getNode(ISD::FMUL, dl, VT, N0,
6869 DAG.getNode(ISD::FADD, dl, VT,
6870 N1, DAG.getConstantFP(1.0, VT)));
6872 // (fma x, c, (fneg x)) -> (fmul x, (c-1))
6873 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
6874 N2.getOpcode() == ISD::FNEG && N2.getOperand(0) == N0)
6875 return DAG.getNode(ISD::FMUL, dl, VT, N0,
6876 DAG.getNode(ISD::FADD, dl, VT,
6877 N1, DAG.getConstantFP(-1.0, VT)));
6883 SDValue DAGCombiner::visitFDIV(SDNode *N) {
6884 SDValue N0 = N->getOperand(0);
6885 SDValue N1 = N->getOperand(1);
6886 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6887 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6888 EVT VT = N->getValueType(0);
6889 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6892 if (VT.isVector()) {
6893 SDValue FoldedVOp = SimplifyVBinOp(N);
6894 if (FoldedVOp.getNode()) return FoldedVOp;
6897 // fold (fdiv c1, c2) -> c1/c2
6899 return DAG.getNode(ISD::FDIV, SDLoc(N), VT, N0, N1);
6901 // fold (fdiv X, c2) -> fmul X, 1/c2 if losing precision is acceptable.
6902 if (N1CFP && DAG.getTarget().Options.UnsafeFPMath) {
6903 // Compute the reciprocal 1.0 / c2.
6904 APFloat N1APF = N1CFP->getValueAPF();
6905 APFloat Recip(N1APF.getSemantics(), 1); // 1.0
6906 APFloat::opStatus st = Recip.divide(N1APF, APFloat::rmNearestTiesToEven);
6907 // Only do the transform if the reciprocal is a legal fp immediate that
6908 // isn't too nasty (eg NaN, denormal, ...).
6909 if ((st == APFloat::opOK || st == APFloat::opInexact) && // Not too nasty
6910 (!LegalOperations ||
6911 // FIXME: custom lowering of ConstantFP might fail (see e.g. ARM
6912 // backend)... we should handle this gracefully after Legalize.
6913 // TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT) ||
6914 TLI.isOperationLegal(llvm::ISD::ConstantFP, VT) ||
6915 TLI.isFPImmLegal(Recip, VT)))
6916 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0,
6917 DAG.getConstantFP(Recip, VT));
6920 // (fdiv (fneg X), (fneg Y)) -> (fdiv X, Y)
6921 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI,
6922 &DAG.getTarget().Options)) {
6923 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI,
6924 &DAG.getTarget().Options)) {
6925 // Both can be negated for free, check to see if at least one is cheaper
6927 if (LHSNeg == 2 || RHSNeg == 2)
6928 return DAG.getNode(ISD::FDIV, SDLoc(N), VT,
6929 GetNegatedExpression(N0, DAG, LegalOperations),
6930 GetNegatedExpression(N1, DAG, LegalOperations));
6937 SDValue DAGCombiner::visitFREM(SDNode *N) {
6938 SDValue N0 = N->getOperand(0);
6939 SDValue N1 = N->getOperand(1);
6940 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6941 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6942 EVT VT = N->getValueType(0);
6944 // fold (frem c1, c2) -> fmod(c1,c2)
6946 return DAG.getNode(ISD::FREM, SDLoc(N), VT, N0, N1);
6951 SDValue DAGCombiner::visitFCOPYSIGN(SDNode *N) {
6952 SDValue N0 = N->getOperand(0);
6953 SDValue N1 = N->getOperand(1);
6954 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6955 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6956 EVT VT = N->getValueType(0);
6958 if (N0CFP && N1CFP) // Constant fold
6959 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT, N0, N1);
6962 const APFloat& V = N1CFP->getValueAPF();
6963 // copysign(x, c1) -> fabs(x) iff ispos(c1)
6964 // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1)
6965 if (!V.isNegative()) {
6966 if (!LegalOperations || TLI.isOperationLegal(ISD::FABS, VT))
6967 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
6969 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
6970 return DAG.getNode(ISD::FNEG, SDLoc(N), VT,
6971 DAG.getNode(ISD::FABS, SDLoc(N0), VT, N0));
6975 // copysign(fabs(x), y) -> copysign(x, y)
6976 // copysign(fneg(x), y) -> copysign(x, y)
6977 // copysign(copysign(x,z), y) -> copysign(x, y)
6978 if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG ||
6979 N0.getOpcode() == ISD::FCOPYSIGN)
6980 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
6981 N0.getOperand(0), N1);
6983 // copysign(x, abs(y)) -> abs(x)
6984 if (N1.getOpcode() == ISD::FABS)
6985 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
6987 // copysign(x, copysign(y,z)) -> copysign(x, z)
6988 if (N1.getOpcode() == ISD::FCOPYSIGN)
6989 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
6990 N0, N1.getOperand(1));
6992 // copysign(x, fp_extend(y)) -> copysign(x, y)
6993 // copysign(x, fp_round(y)) -> copysign(x, y)
6994 if (N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND)
6995 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
6996 N0, N1.getOperand(0));
7001 SDValue DAGCombiner::visitSINT_TO_FP(SDNode *N) {
7002 SDValue N0 = N->getOperand(0);
7003 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
7004 EVT VT = N->getValueType(0);
7005 EVT OpVT = N0.getValueType();
7007 // fold (sint_to_fp c1) -> c1fp
7009 // ...but only if the target supports immediate floating-point values
7010 (!LegalOperations ||
7011 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
7012 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0);
7014 // If the input is a legal type, and SINT_TO_FP is not legal on this target,
7015 // but UINT_TO_FP is legal on this target, try to convert.
7016 if (!TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT) &&
7017 TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT)) {
7018 // If the sign bit is known to be zero, we can change this to UINT_TO_FP.
7019 if (DAG.SignBitIsZero(N0))
7020 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0);
7023 // The next optimizations are desirable only if SELECT_CC can be lowered.
7024 // Check against MVT::Other for SELECT_CC, which is a workaround for targets
7025 // having to say they don't support SELECT_CC on every type the DAG knows
7026 // about, since there is no way to mark an opcode illegal at all value types
7027 // (See also visitSELECT)
7028 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, MVT::Other)) {
7029 // fold (sint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
7030 if (N0.getOpcode() == ISD::SETCC && N0.getValueType() == MVT::i1 &&
7032 (!LegalOperations ||
7033 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
7035 { N0.getOperand(0), N0.getOperand(1),
7036 DAG.getConstantFP(-1.0, VT) , DAG.getConstantFP(0.0, VT),
7038 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops);
7041 // fold (sint_to_fp (zext (setcc x, y, cc))) ->
7042 // (select_cc x, y, 1.0, 0.0,, cc)
7043 if (N0.getOpcode() == ISD::ZERO_EXTEND &&
7044 N0.getOperand(0).getOpcode() == ISD::SETCC &&!VT.isVector() &&
7045 (!LegalOperations ||
7046 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
7048 { N0.getOperand(0).getOperand(0), N0.getOperand(0).getOperand(1),
7049 DAG.getConstantFP(1.0, VT) , DAG.getConstantFP(0.0, VT),
7050 N0.getOperand(0).getOperand(2) };
7051 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops);
7058 SDValue DAGCombiner::visitUINT_TO_FP(SDNode *N) {
7059 SDValue N0 = N->getOperand(0);
7060 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
7061 EVT VT = N->getValueType(0);
7062 EVT OpVT = N0.getValueType();
7064 // fold (uint_to_fp c1) -> c1fp
7066 // ...but only if the target supports immediate floating-point values
7067 (!LegalOperations ||
7068 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
7069 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0);
7071 // If the input is a legal type, and UINT_TO_FP is not legal on this target,
7072 // but SINT_TO_FP is legal on this target, try to convert.
7073 if (!TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT) &&
7074 TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT)) {
7075 // If the sign bit is known to be zero, we can change this to SINT_TO_FP.
7076 if (DAG.SignBitIsZero(N0))
7077 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0);
7080 // The next optimizations are desirable only if SELECT_CC can be lowered.
7081 // Check against MVT::Other for SELECT_CC, which is a workaround for targets
7082 // having to say they don't support SELECT_CC on every type the DAG knows
7083 // about, since there is no way to mark an opcode illegal at all value types
7084 // (See also visitSELECT)
7085 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, MVT::Other)) {
7086 // fold (uint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
7088 if (N0.getOpcode() == ISD::SETCC && !VT.isVector() &&
7089 (!LegalOperations ||
7090 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
7092 { N0.getOperand(0), N0.getOperand(1),
7093 DAG.getConstantFP(1.0, VT), DAG.getConstantFP(0.0, VT),
7095 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops);
7102 SDValue DAGCombiner::visitFP_TO_SINT(SDNode *N) {
7103 SDValue N0 = N->getOperand(0);
7104 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7105 EVT VT = N->getValueType(0);
7107 // fold (fp_to_sint c1fp) -> c1
7109 return DAG.getNode(ISD::FP_TO_SINT, SDLoc(N), VT, N0);
7114 SDValue DAGCombiner::visitFP_TO_UINT(SDNode *N) {
7115 SDValue N0 = N->getOperand(0);
7116 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7117 EVT VT = N->getValueType(0);
7119 // fold (fp_to_uint c1fp) -> c1
7121 return DAG.getNode(ISD::FP_TO_UINT, SDLoc(N), VT, N0);
7126 SDValue DAGCombiner::visitFP_ROUND(SDNode *N) {
7127 SDValue N0 = N->getOperand(0);
7128 SDValue N1 = N->getOperand(1);
7129 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7130 EVT VT = N->getValueType(0);
7132 // fold (fp_round c1fp) -> c1fp
7134 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, N0, N1);
7136 // fold (fp_round (fp_extend x)) -> x
7137 if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType())
7138 return N0.getOperand(0);
7140 // fold (fp_round (fp_round x)) -> (fp_round x)
7141 if (N0.getOpcode() == ISD::FP_ROUND) {
7142 // This is a value preserving truncation if both round's are.
7143 bool IsTrunc = N->getConstantOperandVal(1) == 1 &&
7144 N0.getNode()->getConstantOperandVal(1) == 1;
7145 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, N0.getOperand(0),
7146 DAG.getIntPtrConstant(IsTrunc));
7149 // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y)
7150 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse()) {
7151 SDValue Tmp = DAG.getNode(ISD::FP_ROUND, SDLoc(N0), VT,
7152 N0.getOperand(0), N1);
7153 AddToWorkList(Tmp.getNode());
7154 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
7155 Tmp, N0.getOperand(1));
7161 SDValue DAGCombiner::visitFP_ROUND_INREG(SDNode *N) {
7162 SDValue N0 = N->getOperand(0);
7163 EVT VT = N->getValueType(0);
7164 EVT EVT = cast<VTSDNode>(N->getOperand(1))->getVT();
7165 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7167 // fold (fp_round_inreg c1fp) -> c1fp
7168 if (N0CFP && isTypeLegal(EVT)) {
7169 SDValue Round = DAG.getConstantFP(*N0CFP->getConstantFPValue(), EVT);
7170 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, Round);
7176 SDValue DAGCombiner::visitFP_EXTEND(SDNode *N) {
7177 SDValue N0 = N->getOperand(0);
7178 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7179 EVT VT = N->getValueType(0);
7181 // If this is fp_round(fpextend), don't fold it, allow ourselves to be folded.
7182 if (N->hasOneUse() &&
7183 N->use_begin()->getOpcode() == ISD::FP_ROUND)
7186 // fold (fp_extend c1fp) -> c1fp
7188 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, N0);
7190 // Turn fp_extend(fp_round(X, 1)) -> x since the fp_round doesn't affect the
7192 if (N0.getOpcode() == ISD::FP_ROUND
7193 && N0.getNode()->getConstantOperandVal(1) == 1) {
7194 SDValue In = N0.getOperand(0);
7195 if (In.getValueType() == VT) return In;
7196 if (VT.bitsLT(In.getValueType()))
7197 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT,
7198 In, N0.getOperand(1));
7199 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, In);
7202 // fold (fpext (load x)) -> (fpext (fptrunc (extload x)))
7203 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
7204 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
7205 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType()))) {
7206 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
7207 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT,
7209 LN0->getBasePtr(), N0.getValueType(),
7210 LN0->getMemOperand());
7211 CombineTo(N, ExtLoad);
7212 CombineTo(N0.getNode(),
7213 DAG.getNode(ISD::FP_ROUND, SDLoc(N0),
7214 N0.getValueType(), ExtLoad, DAG.getIntPtrConstant(1)),
7215 ExtLoad.getValue(1));
7216 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7222 SDValue DAGCombiner::visitFNEG(SDNode *N) {
7223 SDValue N0 = N->getOperand(0);
7224 EVT VT = N->getValueType(0);
7226 if (VT.isVector()) {
7227 SDValue FoldedVOp = SimplifyVUnaryOp(N);
7228 if (FoldedVOp.getNode()) return FoldedVOp;
7231 if (isNegatibleForFree(N0, LegalOperations, DAG.getTargetLoweringInfo(),
7232 &DAG.getTarget().Options))
7233 return GetNegatedExpression(N0, DAG, LegalOperations);
7235 // Transform fneg(bitconvert(x)) -> bitconvert(x^sign) to avoid loading
7236 // constant pool values.
7237 if (!TLI.isFNegFree(VT) && N0.getOpcode() == ISD::BITCAST &&
7239 N0.getNode()->hasOneUse() &&
7240 N0.getOperand(0).getValueType().isInteger()) {
7241 SDValue Int = N0.getOperand(0);
7242 EVT IntVT = Int.getValueType();
7243 if (IntVT.isInteger() && !IntVT.isVector()) {
7244 Int = DAG.getNode(ISD::XOR, SDLoc(N0), IntVT, Int,
7245 DAG.getConstant(APInt::getSignBit(IntVT.getSizeInBits()), IntVT));
7246 AddToWorkList(Int.getNode());
7247 return DAG.getNode(ISD::BITCAST, SDLoc(N),
7252 // (fneg (fmul c, x)) -> (fmul -c, x)
7253 if (N0.getOpcode() == ISD::FMUL) {
7254 ConstantFPSDNode *CFP1 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
7256 APFloat CVal = CFP1->getValueAPF();
7258 if (Level >= AfterLegalizeDAG &&
7259 (TLI.isFPImmLegal(CVal, N->getValueType(0)) ||
7260 TLI.isOperationLegal(ISD::ConstantFP, N->getValueType(0))))
7262 ISD::FMUL, SDLoc(N), VT, N0.getOperand(0),
7263 DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0.getOperand(1)));
7270 SDValue DAGCombiner::visitFCEIL(SDNode *N) {
7271 SDValue N0 = N->getOperand(0);
7272 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7273 EVT VT = N->getValueType(0);
7275 // fold (fceil c1) -> fceil(c1)
7277 return DAG.getNode(ISD::FCEIL, SDLoc(N), VT, N0);
7282 SDValue DAGCombiner::visitFTRUNC(SDNode *N) {
7283 SDValue N0 = N->getOperand(0);
7284 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7285 EVT VT = N->getValueType(0);
7287 // fold (ftrunc c1) -> ftrunc(c1)
7289 return DAG.getNode(ISD::FTRUNC, SDLoc(N), VT, N0);
7294 SDValue DAGCombiner::visitFFLOOR(SDNode *N) {
7295 SDValue N0 = N->getOperand(0);
7296 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7297 EVT VT = N->getValueType(0);
7299 // fold (ffloor c1) -> ffloor(c1)
7301 return DAG.getNode(ISD::FFLOOR, SDLoc(N), VT, N0);
7306 SDValue DAGCombiner::visitFABS(SDNode *N) {
7307 SDValue N0 = N->getOperand(0);
7308 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7309 EVT VT = N->getValueType(0);
7311 if (VT.isVector()) {
7312 SDValue FoldedVOp = SimplifyVUnaryOp(N);
7313 if (FoldedVOp.getNode()) return FoldedVOp;
7316 // fold (fabs c1) -> fabs(c1)
7318 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
7319 // fold (fabs (fabs x)) -> (fabs x)
7320 if (N0.getOpcode() == ISD::FABS)
7321 return N->getOperand(0);
7322 // fold (fabs (fneg x)) -> (fabs x)
7323 // fold (fabs (fcopysign x, y)) -> (fabs x)
7324 if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN)
7325 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0.getOperand(0));
7327 // Transform fabs(bitconvert(x)) -> bitconvert(x&~sign) to avoid loading
7328 // constant pool values.
7329 if (!TLI.isFAbsFree(VT) &&
7330 N0.getOpcode() == ISD::BITCAST && N0.getNode()->hasOneUse() &&
7331 N0.getOperand(0).getValueType().isInteger() &&
7332 !N0.getOperand(0).getValueType().isVector()) {
7333 SDValue Int = N0.getOperand(0);
7334 EVT IntVT = Int.getValueType();
7335 if (IntVT.isInteger() && !IntVT.isVector()) {
7336 Int = DAG.getNode(ISD::AND, SDLoc(N0), IntVT, Int,
7337 DAG.getConstant(~APInt::getSignBit(IntVT.getSizeInBits()), IntVT));
7338 AddToWorkList(Int.getNode());
7339 return DAG.getNode(ISD::BITCAST, SDLoc(N),
7340 N->getValueType(0), Int);
7347 SDValue DAGCombiner::visitBRCOND(SDNode *N) {
7348 SDValue Chain = N->getOperand(0);
7349 SDValue N1 = N->getOperand(1);
7350 SDValue N2 = N->getOperand(2);
7352 // If N is a constant we could fold this into a fallthrough or unconditional
7353 // branch. However that doesn't happen very often in normal code, because
7354 // Instcombine/SimplifyCFG should have handled the available opportunities.
7355 // If we did this folding here, it would be necessary to update the
7356 // MachineBasicBlock CFG, which is awkward.
7358 // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal
7360 if (N1.getOpcode() == ISD::SETCC &&
7361 TLI.isOperationLegalOrCustom(ISD::BR_CC,
7362 N1.getOperand(0).getValueType())) {
7363 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other,
7364 Chain, N1.getOperand(2),
7365 N1.getOperand(0), N1.getOperand(1), N2);
7368 if ((N1.hasOneUse() && N1.getOpcode() == ISD::SRL) ||
7369 ((N1.getOpcode() == ISD::TRUNCATE && N1.hasOneUse()) &&
7370 (N1.getOperand(0).hasOneUse() &&
7371 N1.getOperand(0).getOpcode() == ISD::SRL))) {
7372 SDNode *Trunc = nullptr;
7373 if (N1.getOpcode() == ISD::TRUNCATE) {
7374 // Look pass the truncate.
7375 Trunc = N1.getNode();
7376 N1 = N1.getOperand(0);
7379 // Match this pattern so that we can generate simpler code:
7382 // %b = and i32 %a, 2
7383 // %c = srl i32 %b, 1
7384 // brcond i32 %c ...
7389 // %b = and i32 %a, 2
7390 // %c = setcc eq %b, 0
7393 // This applies only when the AND constant value has one bit set and the
7394 // SRL constant is equal to the log2 of the AND constant. The back-end is
7395 // smart enough to convert the result into a TEST/JMP sequence.
7396 SDValue Op0 = N1.getOperand(0);
7397 SDValue Op1 = N1.getOperand(1);
7399 if (Op0.getOpcode() == ISD::AND &&
7400 Op1.getOpcode() == ISD::Constant) {
7401 SDValue AndOp1 = Op0.getOperand(1);
7403 if (AndOp1.getOpcode() == ISD::Constant) {
7404 const APInt &AndConst = cast<ConstantSDNode>(AndOp1)->getAPIntValue();
7406 if (AndConst.isPowerOf2() &&
7407 cast<ConstantSDNode>(Op1)->getAPIntValue()==AndConst.logBase2()) {
7409 DAG.getSetCC(SDLoc(N),
7410 getSetCCResultType(Op0.getValueType()),
7411 Op0, DAG.getConstant(0, Op0.getValueType()),
7414 SDValue NewBRCond = DAG.getNode(ISD::BRCOND, SDLoc(N),
7415 MVT::Other, Chain, SetCC, N2);
7416 // Don't add the new BRCond into the worklist or else SimplifySelectCC
7417 // will convert it back to (X & C1) >> C2.
7418 CombineTo(N, NewBRCond, false);
7419 // Truncate is dead.
7421 removeFromWorkList(Trunc);
7422 DAG.DeleteNode(Trunc);
7424 // Replace the uses of SRL with SETCC
7425 WorkListRemover DeadNodes(*this);
7426 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
7427 removeFromWorkList(N1.getNode());
7428 DAG.DeleteNode(N1.getNode());
7429 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7435 // Restore N1 if the above transformation doesn't match.
7436 N1 = N->getOperand(1);
7439 // Transform br(xor(x, y)) -> br(x != y)
7440 // Transform br(xor(xor(x,y), 1)) -> br (x == y)
7441 if (N1.hasOneUse() && N1.getOpcode() == ISD::XOR) {
7442 SDNode *TheXor = N1.getNode();
7443 SDValue Op0 = TheXor->getOperand(0);
7444 SDValue Op1 = TheXor->getOperand(1);
7445 if (Op0.getOpcode() == Op1.getOpcode()) {
7446 // Avoid missing important xor optimizations.
7447 SDValue Tmp = visitXOR(TheXor);
7448 if (Tmp.getNode()) {
7449 if (Tmp.getNode() != TheXor) {
7450 DEBUG(dbgs() << "\nReplacing.8 ";
7452 dbgs() << "\nWith: ";
7453 Tmp.getNode()->dump(&DAG);
7455 WorkListRemover DeadNodes(*this);
7456 DAG.ReplaceAllUsesOfValueWith(N1, Tmp);
7457 removeFromWorkList(TheXor);
7458 DAG.DeleteNode(TheXor);
7459 return DAG.getNode(ISD::BRCOND, SDLoc(N),
7460 MVT::Other, Chain, Tmp, N2);
7463 // visitXOR has changed XOR's operands or replaced the XOR completely,
7465 return SDValue(N, 0);
7469 if (Op0.getOpcode() != ISD::SETCC && Op1.getOpcode() != ISD::SETCC) {
7471 if (ConstantSDNode *RHSCI = dyn_cast<ConstantSDNode>(Op0))
7472 if (RHSCI->getAPIntValue() == 1 && Op0.hasOneUse() &&
7473 Op0.getOpcode() == ISD::XOR) {
7474 TheXor = Op0.getNode();
7478 EVT SetCCVT = N1.getValueType();
7480 SetCCVT = getSetCCResultType(SetCCVT);
7481 SDValue SetCC = DAG.getSetCC(SDLoc(TheXor),
7484 Equal ? ISD::SETEQ : ISD::SETNE);
7485 // Replace the uses of XOR with SETCC
7486 WorkListRemover DeadNodes(*this);
7487 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
7488 removeFromWorkList(N1.getNode());
7489 DAG.DeleteNode(N1.getNode());
7490 return DAG.getNode(ISD::BRCOND, SDLoc(N),
7491 MVT::Other, Chain, SetCC, N2);
7498 // Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB.
7500 SDValue DAGCombiner::visitBR_CC(SDNode *N) {
7501 CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1));
7502 SDValue CondLHS = N->getOperand(2), CondRHS = N->getOperand(3);
7504 // If N is a constant we could fold this into a fallthrough or unconditional
7505 // branch. However that doesn't happen very often in normal code, because
7506 // Instcombine/SimplifyCFG should have handled the available opportunities.
7507 // If we did this folding here, it would be necessary to update the
7508 // MachineBasicBlock CFG, which is awkward.
7510 // Use SimplifySetCC to simplify SETCC's.
7511 SDValue Simp = SimplifySetCC(getSetCCResultType(CondLHS.getValueType()),
7512 CondLHS, CondRHS, CC->get(), SDLoc(N),
7514 if (Simp.getNode()) AddToWorkList(Simp.getNode());
7516 // fold to a simpler setcc
7517 if (Simp.getNode() && Simp.getOpcode() == ISD::SETCC)
7518 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other,
7519 N->getOperand(0), Simp.getOperand(2),
7520 Simp.getOperand(0), Simp.getOperand(1),
7526 /// canFoldInAddressingMode - Return true if 'Use' is a load or a store that
7527 /// uses N as its base pointer and that N may be folded in the load / store
7528 /// addressing mode.
7529 static bool canFoldInAddressingMode(SDNode *N, SDNode *Use,
7531 const TargetLowering &TLI) {
7533 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Use)) {
7534 if (LD->isIndexed() || LD->getBasePtr().getNode() != N)
7536 VT = Use->getValueType(0);
7537 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(Use)) {
7538 if (ST->isIndexed() || ST->getBasePtr().getNode() != N)
7540 VT = ST->getValue().getValueType();
7544 TargetLowering::AddrMode AM;
7545 if (N->getOpcode() == ISD::ADD) {
7546 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
7549 AM.BaseOffs = Offset->getSExtValue();
7553 } else if (N->getOpcode() == ISD::SUB) {
7554 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
7557 AM.BaseOffs = -Offset->getSExtValue();
7564 return TLI.isLegalAddressingMode(AM, VT.getTypeForEVT(*DAG.getContext()));
7567 /// CombineToPreIndexedLoadStore - Try turning a load / store into a
7568 /// pre-indexed load / store when the base pointer is an add or subtract
7569 /// and it has other uses besides the load / store. After the
7570 /// transformation, the new indexed load / store has effectively folded
7571 /// the add / subtract in and all of its other uses are redirected to the
7572 /// new load / store.
7573 bool DAGCombiner::CombineToPreIndexedLoadStore(SDNode *N) {
7574 if (Level < AfterLegalizeDAG)
7580 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7581 if (LD->isIndexed())
7583 VT = LD->getMemoryVT();
7584 if (!TLI.isIndexedLoadLegal(ISD::PRE_INC, VT) &&
7585 !TLI.isIndexedLoadLegal(ISD::PRE_DEC, VT))
7587 Ptr = LD->getBasePtr();
7588 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7589 if (ST->isIndexed())
7591 VT = ST->getMemoryVT();
7592 if (!TLI.isIndexedStoreLegal(ISD::PRE_INC, VT) &&
7593 !TLI.isIndexedStoreLegal(ISD::PRE_DEC, VT))
7595 Ptr = ST->getBasePtr();
7601 // If the pointer is not an add/sub, or if it doesn't have multiple uses, bail
7602 // out. There is no reason to make this a preinc/predec.
7603 if ((Ptr.getOpcode() != ISD::ADD && Ptr.getOpcode() != ISD::SUB) ||
7604 Ptr.getNode()->hasOneUse())
7607 // Ask the target to do addressing mode selection.
7610 ISD::MemIndexedMode AM = ISD::UNINDEXED;
7611 if (!TLI.getPreIndexedAddressParts(N, BasePtr, Offset, AM, DAG))
7614 // Backends without true r+i pre-indexed forms may need to pass a
7615 // constant base with a variable offset so that constant coercion
7616 // will work with the patterns in canonical form.
7617 bool Swapped = false;
7618 if (isa<ConstantSDNode>(BasePtr)) {
7619 std::swap(BasePtr, Offset);
7623 // Don't create a indexed load / store with zero offset.
7624 if (isa<ConstantSDNode>(Offset) &&
7625 cast<ConstantSDNode>(Offset)->isNullValue())
7628 // Try turning it into a pre-indexed load / store except when:
7629 // 1) The new base ptr is a frame index.
7630 // 2) If N is a store and the new base ptr is either the same as or is a
7631 // predecessor of the value being stored.
7632 // 3) Another use of old base ptr is a predecessor of N. If ptr is folded
7633 // that would create a cycle.
7634 // 4) All uses are load / store ops that use it as old base ptr.
7636 // Check #1. Preinc'ing a frame index would require copying the stack pointer
7637 // (plus the implicit offset) to a register to preinc anyway.
7638 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
7643 SDValue Val = cast<StoreSDNode>(N)->getValue();
7644 if (Val == BasePtr || BasePtr.getNode()->isPredecessorOf(Val.getNode()))
7648 // If the offset is a constant, there may be other adds of constants that
7649 // can be folded with this one. We should do this to avoid having to keep
7650 // a copy of the original base pointer.
7651 SmallVector<SDNode *, 16> OtherUses;
7652 if (isa<ConstantSDNode>(Offset))
7653 for (SDNode *Use : BasePtr.getNode()->uses()) {
7654 if (Use == Ptr.getNode())
7657 if (Use->isPredecessorOf(N))
7660 if (Use->getOpcode() != ISD::ADD && Use->getOpcode() != ISD::SUB) {
7665 SDValue Op0 = Use->getOperand(0), Op1 = Use->getOperand(1);
7666 if (Op1.getNode() == BasePtr.getNode())
7667 std::swap(Op0, Op1);
7668 assert(Op0.getNode() == BasePtr.getNode() &&
7669 "Use of ADD/SUB but not an operand");
7671 if (!isa<ConstantSDNode>(Op1)) {
7676 // FIXME: In some cases, we can be smarter about this.
7677 if (Op1.getValueType() != Offset.getValueType()) {
7682 OtherUses.push_back(Use);
7686 std::swap(BasePtr, Offset);
7688 // Now check for #3 and #4.
7689 bool RealUse = false;
7691 // Caches for hasPredecessorHelper
7692 SmallPtrSet<const SDNode *, 32> Visited;
7693 SmallVector<const SDNode *, 16> Worklist;
7695 for (SDNode *Use : Ptr.getNode()->uses()) {
7698 if (N->hasPredecessorHelper(Use, Visited, Worklist))
7701 // If Ptr may be folded in addressing mode of other use, then it's
7702 // not profitable to do this transformation.
7703 if (!canFoldInAddressingMode(Ptr.getNode(), Use, DAG, TLI))
7712 Result = DAG.getIndexedLoad(SDValue(N,0), SDLoc(N),
7713 BasePtr, Offset, AM);
7715 Result = DAG.getIndexedStore(SDValue(N,0), SDLoc(N),
7716 BasePtr, Offset, AM);
7719 DEBUG(dbgs() << "\nReplacing.4 ";
7721 dbgs() << "\nWith: ";
7722 Result.getNode()->dump(&DAG);
7724 WorkListRemover DeadNodes(*this);
7726 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
7727 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
7729 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
7732 // Finally, since the node is now dead, remove it from the graph.
7736 std::swap(BasePtr, Offset);
7738 // Replace other uses of BasePtr that can be updated to use Ptr
7739 for (unsigned i = 0, e = OtherUses.size(); i != e; ++i) {
7740 unsigned OffsetIdx = 1;
7741 if (OtherUses[i]->getOperand(OffsetIdx).getNode() == BasePtr.getNode())
7743 assert(OtherUses[i]->getOperand(!OffsetIdx).getNode() ==
7744 BasePtr.getNode() && "Expected BasePtr operand");
7746 // We need to replace ptr0 in the following expression:
7747 // x0 * offset0 + y0 * ptr0 = t0
7749 // x1 * offset1 + y1 * ptr0 = t1 (the indexed load/store)
7751 // where x0, x1, y0 and y1 in {-1, 1} are given by the types of the
7752 // indexed load/store and the expresion that needs to be re-written.
7754 // Therefore, we have:
7755 // t0 = (x0 * offset0 - x1 * y0 * y1 *offset1) + (y0 * y1) * t1
7757 ConstantSDNode *CN =
7758 cast<ConstantSDNode>(OtherUses[i]->getOperand(OffsetIdx));
7760 APInt Offset0 = CN->getAPIntValue();
7761 APInt Offset1 = cast<ConstantSDNode>(Offset)->getAPIntValue();
7763 X0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 1) ? -1 : 1;
7764 Y0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 0) ? -1 : 1;
7765 X1 = (AM == ISD::PRE_DEC && !Swapped) ? -1 : 1;
7766 Y1 = (AM == ISD::PRE_DEC && Swapped) ? -1 : 1;
7768 unsigned Opcode = (Y0 * Y1 < 0) ? ISD::SUB : ISD::ADD;
7770 APInt CNV = Offset0;
7771 if (X0 < 0) CNV = -CNV;
7772 if (X1 * Y0 * Y1 < 0) CNV = CNV + Offset1;
7773 else CNV = CNV - Offset1;
7775 // We can now generate the new expression.
7776 SDValue NewOp1 = DAG.getConstant(CNV, CN->getValueType(0));
7777 SDValue NewOp2 = Result.getValue(isLoad ? 1 : 0);
7779 SDValue NewUse = DAG.getNode(Opcode,
7780 SDLoc(OtherUses[i]),
7781 OtherUses[i]->getValueType(0), NewOp1, NewOp2);
7782 DAG.ReplaceAllUsesOfValueWith(SDValue(OtherUses[i], 0), NewUse);
7783 removeFromWorkList(OtherUses[i]);
7784 DAG.DeleteNode(OtherUses[i]);
7787 // Replace the uses of Ptr with uses of the updated base value.
7788 DAG.ReplaceAllUsesOfValueWith(Ptr, Result.getValue(isLoad ? 1 : 0));
7789 removeFromWorkList(Ptr.getNode());
7790 DAG.DeleteNode(Ptr.getNode());
7795 /// CombineToPostIndexedLoadStore - Try to combine a load / store with a
7796 /// add / sub of the base pointer node into a post-indexed load / store.
7797 /// The transformation folded the add / subtract into the new indexed
7798 /// load / store effectively and all of its uses are redirected to the
7799 /// new load / store.
7800 bool DAGCombiner::CombineToPostIndexedLoadStore(SDNode *N) {
7801 if (Level < AfterLegalizeDAG)
7807 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7808 if (LD->isIndexed())
7810 VT = LD->getMemoryVT();
7811 if (!TLI.isIndexedLoadLegal(ISD::POST_INC, VT) &&
7812 !TLI.isIndexedLoadLegal(ISD::POST_DEC, VT))
7814 Ptr = LD->getBasePtr();
7815 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7816 if (ST->isIndexed())
7818 VT = ST->getMemoryVT();
7819 if (!TLI.isIndexedStoreLegal(ISD::POST_INC, VT) &&
7820 !TLI.isIndexedStoreLegal(ISD::POST_DEC, VT))
7822 Ptr = ST->getBasePtr();
7828 if (Ptr.getNode()->hasOneUse())
7831 for (SDNode *Op : Ptr.getNode()->uses()) {
7833 (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB))
7838 ISD::MemIndexedMode AM = ISD::UNINDEXED;
7839 if (TLI.getPostIndexedAddressParts(N, Op, BasePtr, Offset, AM, DAG)) {
7840 // Don't create a indexed load / store with zero offset.
7841 if (isa<ConstantSDNode>(Offset) &&
7842 cast<ConstantSDNode>(Offset)->isNullValue())
7845 // Try turning it into a post-indexed load / store except when
7846 // 1) All uses are load / store ops that use it as base ptr (and
7847 // it may be folded as addressing mmode).
7848 // 2) Op must be independent of N, i.e. Op is neither a predecessor
7849 // nor a successor of N. Otherwise, if Op is folded that would
7852 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
7856 bool TryNext = false;
7857 for (SDNode *Use : BasePtr.getNode()->uses()) {
7858 if (Use == Ptr.getNode())
7861 // If all the uses are load / store addresses, then don't do the
7863 if (Use->getOpcode() == ISD::ADD || Use->getOpcode() == ISD::SUB){
7864 bool RealUse = false;
7865 for (SDNode *UseUse : Use->uses()) {
7866 if (!canFoldInAddressingMode(Use, UseUse, DAG, TLI))
7881 if (!Op->isPredecessorOf(N) && !N->isPredecessorOf(Op)) {
7882 SDValue Result = isLoad
7883 ? DAG.getIndexedLoad(SDValue(N,0), SDLoc(N),
7884 BasePtr, Offset, AM)
7885 : DAG.getIndexedStore(SDValue(N,0), SDLoc(N),
7886 BasePtr, Offset, AM);
7889 DEBUG(dbgs() << "\nReplacing.5 ";
7891 dbgs() << "\nWith: ";
7892 Result.getNode()->dump(&DAG);
7894 WorkListRemover DeadNodes(*this);
7896 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
7897 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
7899 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
7902 // Finally, since the node is now dead, remove it from the graph.
7905 // Replace the uses of Use with uses of the updated base value.
7906 DAG.ReplaceAllUsesOfValueWith(SDValue(Op, 0),
7907 Result.getValue(isLoad ? 1 : 0));
7908 removeFromWorkList(Op);
7918 SDValue DAGCombiner::visitLOAD(SDNode *N) {
7919 LoadSDNode *LD = cast<LoadSDNode>(N);
7920 SDValue Chain = LD->getChain();
7921 SDValue Ptr = LD->getBasePtr();
7923 // If load is not volatile and there are no uses of the loaded value (and
7924 // the updated indexed value in case of indexed loads), change uses of the
7925 // chain value into uses of the chain input (i.e. delete the dead load).
7926 if (!LD->isVolatile()) {
7927 if (N->getValueType(1) == MVT::Other) {
7929 if (!N->hasAnyUseOfValue(0)) {
7930 // It's not safe to use the two value CombineTo variant here. e.g.
7931 // v1, chain2 = load chain1, loc
7932 // v2, chain3 = load chain2, loc
7934 // Now we replace use of chain2 with chain1. This makes the second load
7935 // isomorphic to the one we are deleting, and thus makes this load live.
7936 DEBUG(dbgs() << "\nReplacing.6 ";
7938 dbgs() << "\nWith chain: ";
7939 Chain.getNode()->dump(&DAG);
7941 WorkListRemover DeadNodes(*this);
7942 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
7944 if (N->use_empty()) {
7945 removeFromWorkList(N);
7949 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7953 assert(N->getValueType(2) == MVT::Other && "Malformed indexed loads?");
7954 if (!N->hasAnyUseOfValue(0) && !N->hasAnyUseOfValue(1)) {
7955 SDValue Undef = DAG.getUNDEF(N->getValueType(0));
7956 DEBUG(dbgs() << "\nReplacing.7 ";
7958 dbgs() << "\nWith: ";
7959 Undef.getNode()->dump(&DAG);
7960 dbgs() << " and 2 other values\n");
7961 WorkListRemover DeadNodes(*this);
7962 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Undef);
7963 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1),
7964 DAG.getUNDEF(N->getValueType(1)));
7965 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 2), Chain);
7966 removeFromWorkList(N);
7968 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7973 // If this load is directly stored, replace the load value with the stored
7975 // TODO: Handle store large -> read small portion.
7976 // TODO: Handle TRUNCSTORE/LOADEXT
7977 if (ISD::isNormalLoad(N) && !LD->isVolatile()) {
7978 if (ISD::isNON_TRUNCStore(Chain.getNode())) {
7979 StoreSDNode *PrevST = cast<StoreSDNode>(Chain);
7980 if (PrevST->getBasePtr() == Ptr &&
7981 PrevST->getValue().getValueType() == N->getValueType(0))
7982 return CombineTo(N, Chain.getOperand(1), Chain);
7986 // Try to infer better alignment information than the load already has.
7987 if (OptLevel != CodeGenOpt::None && LD->isUnindexed()) {
7988 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
7989 if (Align > LD->getMemOperand()->getBaseAlignment()) {
7991 DAG.getExtLoad(LD->getExtensionType(), SDLoc(N),
7992 LD->getValueType(0),
7993 Chain, Ptr, LD->getPointerInfo(),
7995 LD->isVolatile(), LD->isNonTemporal(), Align,
7997 return CombineTo(N, NewLoad, SDValue(NewLoad.getNode(), 1), true);
8002 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA :
8003 TLI.getTargetMachine().getSubtarget<TargetSubtargetInfo>().useAA();
8005 if (CombinerAAOnlyFunc.getNumOccurrences() &&
8006 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
8009 if (UseAA && LD->isUnindexed()) {
8010 // Walk up chain skipping non-aliasing memory nodes.
8011 SDValue BetterChain = FindBetterChain(N, Chain);
8013 // If there is a better chain.
8014 if (Chain != BetterChain) {
8017 // Replace the chain to void dependency.
8018 if (LD->getExtensionType() == ISD::NON_EXTLOAD) {
8019 ReplLoad = DAG.getLoad(N->getValueType(0), SDLoc(LD),
8020 BetterChain, Ptr, LD->getMemOperand());
8022 ReplLoad = DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD),
8023 LD->getValueType(0),
8024 BetterChain, Ptr, LD->getMemoryVT(),
8025 LD->getMemOperand());
8028 // Create token factor to keep old chain connected.
8029 SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N),
8030 MVT::Other, Chain, ReplLoad.getValue(1));
8032 // Make sure the new and old chains are cleaned up.
8033 AddToWorkList(Token.getNode());
8035 // Replace uses with load result and token factor. Don't add users
8037 return CombineTo(N, ReplLoad.getValue(0), Token, false);
8041 // Try transforming N to an indexed load.
8042 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
8043 return SDValue(N, 0);
8045 // Try to slice up N to more direct loads if the slices are mapped to
8046 // different register banks or pairing can take place.
8048 return SDValue(N, 0);
8054 /// \brief Helper structure used to slice a load in smaller loads.
8055 /// Basically a slice is obtained from the following sequence:
8056 /// Origin = load Ty1, Base
8057 /// Shift = srl Ty1 Origin, CstTy Amount
8058 /// Inst = trunc Shift to Ty2
8060 /// Then, it will be rewriten into:
8061 /// Slice = load SliceTy, Base + SliceOffset
8062 /// [Inst = zext Slice to Ty2], only if SliceTy <> Ty2
8064 /// SliceTy is deduced from the number of bits that are actually used to
8066 struct LoadedSlice {
8067 /// \brief Helper structure used to compute the cost of a slice.
8069 /// Are we optimizing for code size.
8074 unsigned CrossRegisterBanksCopies;
8078 Cost(bool ForCodeSize = false)
8079 : ForCodeSize(ForCodeSize), Loads(0), Truncates(0),
8080 CrossRegisterBanksCopies(0), ZExts(0), Shift(0) {}
8082 /// \brief Get the cost of one isolated slice.
8083 Cost(const LoadedSlice &LS, bool ForCodeSize = false)
8084 : ForCodeSize(ForCodeSize), Loads(1), Truncates(0),
8085 CrossRegisterBanksCopies(0), ZExts(0), Shift(0) {
8086 EVT TruncType = LS.Inst->getValueType(0);
8087 EVT LoadedType = LS.getLoadedType();
8088 if (TruncType != LoadedType &&
8089 !LS.DAG->getTargetLoweringInfo().isZExtFree(LoadedType, TruncType))
8093 /// \brief Account for slicing gain in the current cost.
8094 /// Slicing provide a few gains like removing a shift or a
8095 /// truncate. This method allows to grow the cost of the original
8096 /// load with the gain from this slice.
8097 void addSliceGain(const LoadedSlice &LS) {
8098 // Each slice saves a truncate.
8099 const TargetLowering &TLI = LS.DAG->getTargetLoweringInfo();
8100 if (!TLI.isTruncateFree(LS.Inst->getValueType(0),
8101 LS.Inst->getOperand(0).getValueType()))
8103 // If there is a shift amount, this slice gets rid of it.
8106 // If this slice can merge a cross register bank copy, account for it.
8107 if (LS.canMergeExpensiveCrossRegisterBankCopy())
8108 ++CrossRegisterBanksCopies;
8111 Cost &operator+=(const Cost &RHS) {
8113 Truncates += RHS.Truncates;
8114 CrossRegisterBanksCopies += RHS.CrossRegisterBanksCopies;
8120 bool operator==(const Cost &RHS) const {
8121 return Loads == RHS.Loads && Truncates == RHS.Truncates &&
8122 CrossRegisterBanksCopies == RHS.CrossRegisterBanksCopies &&
8123 ZExts == RHS.ZExts && Shift == RHS.Shift;
8126 bool operator!=(const Cost &RHS) const { return !(*this == RHS); }
8128 bool operator<(const Cost &RHS) const {
8129 // Assume cross register banks copies are as expensive as loads.
8130 // FIXME: Do we want some more target hooks?
8131 unsigned ExpensiveOpsLHS = Loads + CrossRegisterBanksCopies;
8132 unsigned ExpensiveOpsRHS = RHS.Loads + RHS.CrossRegisterBanksCopies;
8133 // Unless we are optimizing for code size, consider the
8134 // expensive operation first.
8135 if (!ForCodeSize && ExpensiveOpsLHS != ExpensiveOpsRHS)
8136 return ExpensiveOpsLHS < ExpensiveOpsRHS;
8137 return (Truncates + ZExts + Shift + ExpensiveOpsLHS) <
8138 (RHS.Truncates + RHS.ZExts + RHS.Shift + ExpensiveOpsRHS);
8141 bool operator>(const Cost &RHS) const { return RHS < *this; }
8143 bool operator<=(const Cost &RHS) const { return !(RHS < *this); }
8145 bool operator>=(const Cost &RHS) const { return !(*this < RHS); }
8147 // The last instruction that represent the slice. This should be a
8148 // truncate instruction.
8150 // The original load instruction.
8152 // The right shift amount in bits from the original load.
8154 // The DAG from which Origin came from.
8155 // This is used to get some contextual information about legal types, etc.
8158 LoadedSlice(SDNode *Inst = nullptr, LoadSDNode *Origin = nullptr,
8159 unsigned Shift = 0, SelectionDAG *DAG = nullptr)
8160 : Inst(Inst), Origin(Origin), Shift(Shift), DAG(DAG) {}
8162 LoadedSlice(const LoadedSlice &LS)
8163 : Inst(LS.Inst), Origin(LS.Origin), Shift(LS.Shift), DAG(LS.DAG) {}
8165 /// \brief Get the bits used in a chunk of bits \p BitWidth large.
8166 /// \return Result is \p BitWidth and has used bits set to 1 and
8167 /// not used bits set to 0.
8168 APInt getUsedBits() const {
8169 // Reproduce the trunc(lshr) sequence:
8170 // - Start from the truncated value.
8171 // - Zero extend to the desired bit width.
8173 assert(Origin && "No original load to compare against.");
8174 unsigned BitWidth = Origin->getValueSizeInBits(0);
8175 assert(Inst && "This slice is not bound to an instruction");
8176 assert(Inst->getValueSizeInBits(0) <= BitWidth &&
8177 "Extracted slice is bigger than the whole type!");
8178 APInt UsedBits(Inst->getValueSizeInBits(0), 0);
8179 UsedBits.setAllBits();
8180 UsedBits = UsedBits.zext(BitWidth);
8185 /// \brief Get the size of the slice to be loaded in bytes.
8186 unsigned getLoadedSize() const {
8187 unsigned SliceSize = getUsedBits().countPopulation();
8188 assert(!(SliceSize & 0x7) && "Size is not a multiple of a byte.");
8189 return SliceSize / 8;
8192 /// \brief Get the type that will be loaded for this slice.
8193 /// Note: This may not be the final type for the slice.
8194 EVT getLoadedType() const {
8195 assert(DAG && "Missing context");
8196 LLVMContext &Ctxt = *DAG->getContext();
8197 return EVT::getIntegerVT(Ctxt, getLoadedSize() * 8);
8200 /// \brief Get the alignment of the load used for this slice.
8201 unsigned getAlignment() const {
8202 unsigned Alignment = Origin->getAlignment();
8203 unsigned Offset = getOffsetFromBase();
8205 Alignment = MinAlign(Alignment, Alignment + Offset);
8209 /// \brief Check if this slice can be rewritten with legal operations.
8210 bool isLegal() const {
8211 // An invalid slice is not legal.
8212 if (!Origin || !Inst || !DAG)
8215 // Offsets are for indexed load only, we do not handle that.
8216 if (Origin->getOffset().getOpcode() != ISD::UNDEF)
8219 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
8221 // Check that the type is legal.
8222 EVT SliceType = getLoadedType();
8223 if (!TLI.isTypeLegal(SliceType))
8226 // Check that the load is legal for this type.
8227 if (!TLI.isOperationLegal(ISD::LOAD, SliceType))
8230 // Check that the offset can be computed.
8231 // 1. Check its type.
8232 EVT PtrType = Origin->getBasePtr().getValueType();
8233 if (PtrType == MVT::Untyped || PtrType.isExtended())
8236 // 2. Check that it fits in the immediate.
8237 if (!TLI.isLegalAddImmediate(getOffsetFromBase()))
8240 // 3. Check that the computation is legal.
8241 if (!TLI.isOperationLegal(ISD::ADD, PtrType))
8244 // Check that the zext is legal if it needs one.
8245 EVT TruncateType = Inst->getValueType(0);
8246 if (TruncateType != SliceType &&
8247 !TLI.isOperationLegal(ISD::ZERO_EXTEND, TruncateType))
8253 /// \brief Get the offset in bytes of this slice in the original chunk of
8255 /// \pre DAG != nullptr.
8256 uint64_t getOffsetFromBase() const {
8257 assert(DAG && "Missing context.");
8259 DAG->getTargetLoweringInfo().getDataLayout()->isBigEndian();
8260 assert(!(Shift & 0x7) && "Shifts not aligned on Bytes are not supported.");
8261 uint64_t Offset = Shift / 8;
8262 unsigned TySizeInBytes = Origin->getValueSizeInBits(0) / 8;
8263 assert(!(Origin->getValueSizeInBits(0) & 0x7) &&
8264 "The size of the original loaded type is not a multiple of a"
8266 // If Offset is bigger than TySizeInBytes, it means we are loading all
8267 // zeros. This should have been optimized before in the process.
8268 assert(TySizeInBytes > Offset &&
8269 "Invalid shift amount for given loaded size");
8271 Offset = TySizeInBytes - Offset - getLoadedSize();
8275 /// \brief Generate the sequence of instructions to load the slice
8276 /// represented by this object and redirect the uses of this slice to
8277 /// this new sequence of instructions.
8278 /// \pre this->Inst && this->Origin are valid Instructions and this
8279 /// object passed the legal check: LoadedSlice::isLegal returned true.
8280 /// \return The last instruction of the sequence used to load the slice.
8281 SDValue loadSlice() const {
8282 assert(Inst && Origin && "Unable to replace a non-existing slice.");
8283 const SDValue &OldBaseAddr = Origin->getBasePtr();
8284 SDValue BaseAddr = OldBaseAddr;
8285 // Get the offset in that chunk of bytes w.r.t. the endianess.
8286 int64_t Offset = static_cast<int64_t>(getOffsetFromBase());
8287 assert(Offset >= 0 && "Offset too big to fit in int64_t!");
8289 // BaseAddr = BaseAddr + Offset.
8290 EVT ArithType = BaseAddr.getValueType();
8291 BaseAddr = DAG->getNode(ISD::ADD, SDLoc(Origin), ArithType, BaseAddr,
8292 DAG->getConstant(Offset, ArithType));
8295 // Create the type of the loaded slice according to its size.
8296 EVT SliceType = getLoadedType();
8298 // Create the load for the slice.
8299 SDValue LastInst = DAG->getLoad(
8300 SliceType, SDLoc(Origin), Origin->getChain(), BaseAddr,
8301 Origin->getPointerInfo().getWithOffset(Offset), Origin->isVolatile(),
8302 Origin->isNonTemporal(), Origin->isInvariant(), getAlignment());
8303 // If the final type is not the same as the loaded type, this means that
8304 // we have to pad with zero. Create a zero extend for that.
8305 EVT FinalType = Inst->getValueType(0);
8306 if (SliceType != FinalType)
8308 DAG->getNode(ISD::ZERO_EXTEND, SDLoc(LastInst), FinalType, LastInst);
8312 /// \brief Check if this slice can be merged with an expensive cross register
8313 /// bank copy. E.g.,
8315 /// f = bitcast i32 i to float
8316 bool canMergeExpensiveCrossRegisterBankCopy() const {
8317 if (!Inst || !Inst->hasOneUse())
8319 SDNode *Use = *Inst->use_begin();
8320 if (Use->getOpcode() != ISD::BITCAST)
8322 assert(DAG && "Missing context");
8323 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
8324 EVT ResVT = Use->getValueType(0);
8325 const TargetRegisterClass *ResRC = TLI.getRegClassFor(ResVT.getSimpleVT());
8326 const TargetRegisterClass *ArgRC =
8327 TLI.getRegClassFor(Use->getOperand(0).getValueType().getSimpleVT());
8328 if (ArgRC == ResRC || !TLI.isOperationLegal(ISD::LOAD, ResVT))
8331 // At this point, we know that we perform a cross-register-bank copy.
8332 // Check if it is expensive.
8333 const TargetRegisterInfo *TRI = TLI.getTargetMachine().getRegisterInfo();
8334 // Assume bitcasts are cheap, unless both register classes do not
8335 // explicitly share a common sub class.
8336 if (!TRI || TRI->getCommonSubClass(ArgRC, ResRC))
8339 // Check if it will be merged with the load.
8340 // 1. Check the alignment constraint.
8341 unsigned RequiredAlignment = TLI.getDataLayout()->getABITypeAlignment(
8342 ResVT.getTypeForEVT(*DAG->getContext()));
8344 if (RequiredAlignment > getAlignment())
8347 // 2. Check that the load is a legal operation for that type.
8348 if (!TLI.isOperationLegal(ISD::LOAD, ResVT))
8351 // 3. Check that we do not have a zext in the way.
8352 if (Inst->getValueType(0) != getLoadedType())
8360 /// \brief Check that all bits set in \p UsedBits form a dense region, i.e.,
8361 /// \p UsedBits looks like 0..0 1..1 0..0.
8362 static bool areUsedBitsDense(const APInt &UsedBits) {
8363 // If all the bits are one, this is dense!
8364 if (UsedBits.isAllOnesValue())
8367 // Get rid of the unused bits on the right.
8368 APInt NarrowedUsedBits = UsedBits.lshr(UsedBits.countTrailingZeros());
8369 // Get rid of the unused bits on the left.
8370 if (NarrowedUsedBits.countLeadingZeros())
8371 NarrowedUsedBits = NarrowedUsedBits.trunc(NarrowedUsedBits.getActiveBits());
8372 // Check that the chunk of bits is completely used.
8373 return NarrowedUsedBits.isAllOnesValue();
8376 /// \brief Check whether or not \p First and \p Second are next to each other
8377 /// in memory. This means that there is no hole between the bits loaded
8378 /// by \p First and the bits loaded by \p Second.
8379 static bool areSlicesNextToEachOther(const LoadedSlice &First,
8380 const LoadedSlice &Second) {
8381 assert(First.Origin == Second.Origin && First.Origin &&
8382 "Unable to match different memory origins.");
8383 APInt UsedBits = First.getUsedBits();
8384 assert((UsedBits & Second.getUsedBits()) == 0 &&
8385 "Slices are not supposed to overlap.");
8386 UsedBits |= Second.getUsedBits();
8387 return areUsedBitsDense(UsedBits);
8390 /// \brief Adjust the \p GlobalLSCost according to the target
8391 /// paring capabilities and the layout of the slices.
8392 /// \pre \p GlobalLSCost should account for at least as many loads as
8393 /// there is in the slices in \p LoadedSlices.
8394 static void adjustCostForPairing(SmallVectorImpl<LoadedSlice> &LoadedSlices,
8395 LoadedSlice::Cost &GlobalLSCost) {
8396 unsigned NumberOfSlices = LoadedSlices.size();
8397 // If there is less than 2 elements, no pairing is possible.
8398 if (NumberOfSlices < 2)
8401 // Sort the slices so that elements that are likely to be next to each
8402 // other in memory are next to each other in the list.
8403 std::sort(LoadedSlices.begin(), LoadedSlices.end(),
8404 [](const LoadedSlice &LHS, const LoadedSlice &RHS) {
8405 assert(LHS.Origin == RHS.Origin && "Different bases not implemented.");
8406 return LHS.getOffsetFromBase() < RHS.getOffsetFromBase();
8408 const TargetLowering &TLI = LoadedSlices[0].DAG->getTargetLoweringInfo();
8409 // First (resp. Second) is the first (resp. Second) potentially candidate
8410 // to be placed in a paired load.
8411 const LoadedSlice *First = nullptr;
8412 const LoadedSlice *Second = nullptr;
8413 for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice,
8414 // Set the beginning of the pair.
8417 Second = &LoadedSlices[CurrSlice];
8419 // If First is NULL, it means we start a new pair.
8420 // Get to the next slice.
8424 EVT LoadedType = First->getLoadedType();
8426 // If the types of the slices are different, we cannot pair them.
8427 if (LoadedType != Second->getLoadedType())
8430 // Check if the target supplies paired loads for this type.
8431 unsigned RequiredAlignment = 0;
8432 if (!TLI.hasPairedLoad(LoadedType, RequiredAlignment)) {
8433 // move to the next pair, this type is hopeless.
8437 // Check if we meet the alignment requirement.
8438 if (RequiredAlignment > First->getAlignment())
8441 // Check that both loads are next to each other in memory.
8442 if (!areSlicesNextToEachOther(*First, *Second))
8445 assert(GlobalLSCost.Loads > 0 && "We save more loads than we created!");
8446 --GlobalLSCost.Loads;
8447 // Move to the next pair.
8452 /// \brief Check the profitability of all involved LoadedSlice.
8453 /// Currently, it is considered profitable if there is exactly two
8454 /// involved slices (1) which are (2) next to each other in memory, and
8455 /// whose cost (\see LoadedSlice::Cost) is smaller than the original load (3).
8457 /// Note: The order of the elements in \p LoadedSlices may be modified, but not
8458 /// the elements themselves.
8460 /// FIXME: When the cost model will be mature enough, we can relax
8461 /// constraints (1) and (2).
8462 static bool isSlicingProfitable(SmallVectorImpl<LoadedSlice> &LoadedSlices,
8463 const APInt &UsedBits, bool ForCodeSize) {
8464 unsigned NumberOfSlices = LoadedSlices.size();
8465 if (StressLoadSlicing)
8466 return NumberOfSlices > 1;
8469 if (NumberOfSlices != 2)
8473 if (!areUsedBitsDense(UsedBits))
8477 LoadedSlice::Cost OrigCost(ForCodeSize), GlobalSlicingCost(ForCodeSize);
8478 // The original code has one big load.
8480 for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice) {
8481 const LoadedSlice &LS = LoadedSlices[CurrSlice];
8482 // Accumulate the cost of all the slices.
8483 LoadedSlice::Cost SliceCost(LS, ForCodeSize);
8484 GlobalSlicingCost += SliceCost;
8486 // Account as cost in the original configuration the gain obtained
8487 // with the current slices.
8488 OrigCost.addSliceGain(LS);
8491 // If the target supports paired load, adjust the cost accordingly.
8492 adjustCostForPairing(LoadedSlices, GlobalSlicingCost);
8493 return OrigCost > GlobalSlicingCost;
8496 /// \brief If the given load, \p LI, is used only by trunc or trunc(lshr)
8497 /// operations, split it in the various pieces being extracted.
8499 /// This sort of thing is introduced by SROA.
8500 /// This slicing takes care not to insert overlapping loads.
8501 /// \pre LI is a simple load (i.e., not an atomic or volatile load).
8502 bool DAGCombiner::SliceUpLoad(SDNode *N) {
8503 if (Level < AfterLegalizeDAG)
8506 LoadSDNode *LD = cast<LoadSDNode>(N);
8507 if (LD->isVolatile() || !ISD::isNormalLoad(LD) ||
8508 !LD->getValueType(0).isInteger())
8511 // Keep track of already used bits to detect overlapping values.
8512 // In that case, we will just abort the transformation.
8513 APInt UsedBits(LD->getValueSizeInBits(0), 0);
8515 SmallVector<LoadedSlice, 4> LoadedSlices;
8517 // Check if this load is used as several smaller chunks of bits.
8518 // Basically, look for uses in trunc or trunc(lshr) and record a new chain
8519 // of computation for each trunc.
8520 for (SDNode::use_iterator UI = LD->use_begin(), UIEnd = LD->use_end();
8521 UI != UIEnd; ++UI) {
8522 // Skip the uses of the chain.
8523 if (UI.getUse().getResNo() != 0)
8529 // Check if this is a trunc(lshr).
8530 if (User->getOpcode() == ISD::SRL && User->hasOneUse() &&
8531 isa<ConstantSDNode>(User->getOperand(1))) {
8532 Shift = cast<ConstantSDNode>(User->getOperand(1))->getZExtValue();
8533 User = *User->use_begin();
8536 // At this point, User is a Truncate, iff we encountered, trunc or
8538 if (User->getOpcode() != ISD::TRUNCATE)
8541 // The width of the type must be a power of 2 and greater than 8-bits.
8542 // Otherwise the load cannot be represented in LLVM IR.
8543 // Moreover, if we shifted with a non-8-bits multiple, the slice
8544 // will be across several bytes. We do not support that.
8545 unsigned Width = User->getValueSizeInBits(0);
8546 if (Width < 8 || !isPowerOf2_32(Width) || (Shift & 0x7))
8549 // Build the slice for this chain of computations.
8550 LoadedSlice LS(User, LD, Shift, &DAG);
8551 APInt CurrentUsedBits = LS.getUsedBits();
8553 // Check if this slice overlaps with another.
8554 if ((CurrentUsedBits & UsedBits) != 0)
8556 // Update the bits used globally.
8557 UsedBits |= CurrentUsedBits;
8559 // Check if the new slice would be legal.
8563 // Record the slice.
8564 LoadedSlices.push_back(LS);
8567 // Abort slicing if it does not seem to be profitable.
8568 if (!isSlicingProfitable(LoadedSlices, UsedBits, ForCodeSize))
8573 // Rewrite each chain to use an independent load.
8574 // By construction, each chain can be represented by a unique load.
8576 // Prepare the argument for the new token factor for all the slices.
8577 SmallVector<SDValue, 8> ArgChains;
8578 for (SmallVectorImpl<LoadedSlice>::const_iterator
8579 LSIt = LoadedSlices.begin(),
8580 LSItEnd = LoadedSlices.end();
8581 LSIt != LSItEnd; ++LSIt) {
8582 SDValue SliceInst = LSIt->loadSlice();
8583 CombineTo(LSIt->Inst, SliceInst, true);
8584 if (SliceInst.getNode()->getOpcode() != ISD::LOAD)
8585 SliceInst = SliceInst.getOperand(0);
8586 assert(SliceInst->getOpcode() == ISD::LOAD &&
8587 "It takes more than a zext to get to the loaded slice!!");
8588 ArgChains.push_back(SliceInst.getValue(1));
8591 SDValue Chain = DAG.getNode(ISD::TokenFactor, SDLoc(LD), MVT::Other,
8593 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
8597 /// CheckForMaskedLoad - Check to see if V is (and load (ptr), imm), where the
8598 /// load is having specific bytes cleared out. If so, return the byte size
8599 /// being masked out and the shift amount.
8600 static std::pair<unsigned, unsigned>
8601 CheckForMaskedLoad(SDValue V, SDValue Ptr, SDValue Chain) {
8602 std::pair<unsigned, unsigned> Result(0, 0);
8604 // Check for the structure we're looking for.
8605 if (V->getOpcode() != ISD::AND ||
8606 !isa<ConstantSDNode>(V->getOperand(1)) ||
8607 !ISD::isNormalLoad(V->getOperand(0).getNode()))
8610 // Check the chain and pointer.
8611 LoadSDNode *LD = cast<LoadSDNode>(V->getOperand(0));
8612 if (LD->getBasePtr() != Ptr) return Result; // Not from same pointer.
8614 // The store should be chained directly to the load or be an operand of a
8616 if (LD == Chain.getNode())
8618 else if (Chain->getOpcode() != ISD::TokenFactor)
8619 return Result; // Fail.
8622 for (unsigned i = 0, e = Chain->getNumOperands(); i != e; ++i)
8623 if (Chain->getOperand(i).getNode() == LD) {
8627 if (!isOk) return Result;
8630 // This only handles simple types.
8631 if (V.getValueType() != MVT::i16 &&
8632 V.getValueType() != MVT::i32 &&
8633 V.getValueType() != MVT::i64)
8636 // Check the constant mask. Invert it so that the bits being masked out are
8637 // 0 and the bits being kept are 1. Use getSExtValue so that leading bits
8638 // follow the sign bit for uniformity.
8639 uint64_t NotMask = ~cast<ConstantSDNode>(V->getOperand(1))->getSExtValue();
8640 unsigned NotMaskLZ = countLeadingZeros(NotMask);
8641 if (NotMaskLZ & 7) return Result; // Must be multiple of a byte.
8642 unsigned NotMaskTZ = countTrailingZeros(NotMask);
8643 if (NotMaskTZ & 7) return Result; // Must be multiple of a byte.
8644 if (NotMaskLZ == 64) return Result; // All zero mask.
8646 // See if we have a continuous run of bits. If so, we have 0*1+0*
8647 if (CountTrailingOnes_64(NotMask >> NotMaskTZ)+NotMaskTZ+NotMaskLZ != 64)
8650 // Adjust NotMaskLZ down to be from the actual size of the int instead of i64.
8651 if (V.getValueType() != MVT::i64 && NotMaskLZ)
8652 NotMaskLZ -= 64-V.getValueSizeInBits();
8654 unsigned MaskedBytes = (V.getValueSizeInBits()-NotMaskLZ-NotMaskTZ)/8;
8655 switch (MaskedBytes) {
8659 default: return Result; // All one mask, or 5-byte mask.
8662 // Verify that the first bit starts at a multiple of mask so that the access
8663 // is aligned the same as the access width.
8664 if (NotMaskTZ && NotMaskTZ/8 % MaskedBytes) return Result;
8666 Result.first = MaskedBytes;
8667 Result.second = NotMaskTZ/8;
8672 /// ShrinkLoadReplaceStoreWithStore - Check to see if IVal is something that
8673 /// provides a value as specified by MaskInfo. If so, replace the specified
8674 /// store with a narrower store of truncated IVal.
8676 ShrinkLoadReplaceStoreWithStore(const std::pair<unsigned, unsigned> &MaskInfo,
8677 SDValue IVal, StoreSDNode *St,
8679 unsigned NumBytes = MaskInfo.first;
8680 unsigned ByteShift = MaskInfo.second;
8681 SelectionDAG &DAG = DC->getDAG();
8683 // Check to see if IVal is all zeros in the part being masked in by the 'or'
8684 // that uses this. If not, this is not a replacement.
8685 APInt Mask = ~APInt::getBitsSet(IVal.getValueSizeInBits(),
8686 ByteShift*8, (ByteShift+NumBytes)*8);
8687 if (!DAG.MaskedValueIsZero(IVal, Mask)) return nullptr;
8689 // Check that it is legal on the target to do this. It is legal if the new
8690 // VT we're shrinking to (i8/i16/i32) is legal or we're still before type
8692 MVT VT = MVT::getIntegerVT(NumBytes*8);
8693 if (!DC->isTypeLegal(VT))
8696 // Okay, we can do this! Replace the 'St' store with a store of IVal that is
8697 // shifted by ByteShift and truncated down to NumBytes.
8699 IVal = DAG.getNode(ISD::SRL, SDLoc(IVal), IVal.getValueType(), IVal,
8700 DAG.getConstant(ByteShift*8,
8701 DC->getShiftAmountTy(IVal.getValueType())));
8703 // Figure out the offset for the store and the alignment of the access.
8705 unsigned NewAlign = St->getAlignment();
8707 if (DAG.getTargetLoweringInfo().isLittleEndian())
8708 StOffset = ByteShift;
8710 StOffset = IVal.getValueType().getStoreSize() - ByteShift - NumBytes;
8712 SDValue Ptr = St->getBasePtr();
8714 Ptr = DAG.getNode(ISD::ADD, SDLoc(IVal), Ptr.getValueType(),
8715 Ptr, DAG.getConstant(StOffset, Ptr.getValueType()));
8716 NewAlign = MinAlign(NewAlign, StOffset);
8719 // Truncate down to the new size.
8720 IVal = DAG.getNode(ISD::TRUNCATE, SDLoc(IVal), VT, IVal);
8723 return DAG.getStore(St->getChain(), SDLoc(St), IVal, Ptr,
8724 St->getPointerInfo().getWithOffset(StOffset),
8725 false, false, NewAlign).getNode();
8729 /// ReduceLoadOpStoreWidth - Look for sequence of load / op / store where op is
8730 /// one of 'or', 'xor', and 'and' of immediates. If 'op' is only touching some
8731 /// of the loaded bits, try narrowing the load and store if it would end up
8732 /// being a win for performance or code size.
8733 SDValue DAGCombiner::ReduceLoadOpStoreWidth(SDNode *N) {
8734 StoreSDNode *ST = cast<StoreSDNode>(N);
8735 if (ST->isVolatile())
8738 SDValue Chain = ST->getChain();
8739 SDValue Value = ST->getValue();
8740 SDValue Ptr = ST->getBasePtr();
8741 EVT VT = Value.getValueType();
8743 if (ST->isTruncatingStore() || VT.isVector() || !Value.hasOneUse())
8746 unsigned Opc = Value.getOpcode();
8748 // If this is "store (or X, Y), P" and X is "(and (load P), cst)", where cst
8749 // is a byte mask indicating a consecutive number of bytes, check to see if
8750 // Y is known to provide just those bytes. If so, we try to replace the
8751 // load + replace + store sequence with a single (narrower) store, which makes
8753 if (Opc == ISD::OR) {
8754 std::pair<unsigned, unsigned> MaskedLoad;
8755 MaskedLoad = CheckForMaskedLoad(Value.getOperand(0), Ptr, Chain);
8756 if (MaskedLoad.first)
8757 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
8758 Value.getOperand(1), ST,this))
8759 return SDValue(NewST, 0);
8761 // Or is commutative, so try swapping X and Y.
8762 MaskedLoad = CheckForMaskedLoad(Value.getOperand(1), Ptr, Chain);
8763 if (MaskedLoad.first)
8764 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
8765 Value.getOperand(0), ST,this))
8766 return SDValue(NewST, 0);
8769 if ((Opc != ISD::OR && Opc != ISD::XOR && Opc != ISD::AND) ||
8770 Value.getOperand(1).getOpcode() != ISD::Constant)
8773 SDValue N0 = Value.getOperand(0);
8774 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
8775 Chain == SDValue(N0.getNode(), 1)) {
8776 LoadSDNode *LD = cast<LoadSDNode>(N0);
8777 if (LD->getBasePtr() != Ptr ||
8778 LD->getPointerInfo().getAddrSpace() !=
8779 ST->getPointerInfo().getAddrSpace())
8782 // Find the type to narrow it the load / op / store to.
8783 SDValue N1 = Value.getOperand(1);
8784 unsigned BitWidth = N1.getValueSizeInBits();
8785 APInt Imm = cast<ConstantSDNode>(N1)->getAPIntValue();
8786 if (Opc == ISD::AND)
8787 Imm ^= APInt::getAllOnesValue(BitWidth);
8788 if (Imm == 0 || Imm.isAllOnesValue())
8790 unsigned ShAmt = Imm.countTrailingZeros();
8791 unsigned MSB = BitWidth - Imm.countLeadingZeros() - 1;
8792 unsigned NewBW = NextPowerOf2(MSB - ShAmt);
8793 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
8794 while (NewBW < BitWidth &&
8795 !(TLI.isOperationLegalOrCustom(Opc, NewVT) &&
8796 TLI.isNarrowingProfitable(VT, NewVT))) {
8797 NewBW = NextPowerOf2(NewBW);
8798 NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
8800 if (NewBW >= BitWidth)
8803 // If the lsb changed does not start at the type bitwidth boundary,
8804 // start at the previous one.
8806 ShAmt = (((ShAmt + NewBW - 1) / NewBW) * NewBW) - NewBW;
8807 APInt Mask = APInt::getBitsSet(BitWidth, ShAmt,
8808 std::min(BitWidth, ShAmt + NewBW));
8809 if ((Imm & Mask) == Imm) {
8810 APInt NewImm = (Imm & Mask).lshr(ShAmt).trunc(NewBW);
8811 if (Opc == ISD::AND)
8812 NewImm ^= APInt::getAllOnesValue(NewBW);
8813 uint64_t PtrOff = ShAmt / 8;
8814 // For big endian targets, we need to adjust the offset to the pointer to
8815 // load the correct bytes.
8816 if (TLI.isBigEndian())
8817 PtrOff = (BitWidth + 7 - NewBW) / 8 - PtrOff;
8819 unsigned NewAlign = MinAlign(LD->getAlignment(), PtrOff);
8820 Type *NewVTTy = NewVT.getTypeForEVT(*DAG.getContext());
8821 if (NewAlign < TLI.getDataLayout()->getABITypeAlignment(NewVTTy))
8824 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(LD),
8825 Ptr.getValueType(), Ptr,
8826 DAG.getConstant(PtrOff, Ptr.getValueType()));
8827 SDValue NewLD = DAG.getLoad(NewVT, SDLoc(N0),
8828 LD->getChain(), NewPtr,
8829 LD->getPointerInfo().getWithOffset(PtrOff),
8830 LD->isVolatile(), LD->isNonTemporal(),
8831 LD->isInvariant(), NewAlign,
8833 SDValue NewVal = DAG.getNode(Opc, SDLoc(Value), NewVT, NewLD,
8834 DAG.getConstant(NewImm, NewVT));
8835 SDValue NewST = DAG.getStore(Chain, SDLoc(N),
8837 ST->getPointerInfo().getWithOffset(PtrOff),
8838 false, false, NewAlign);
8840 AddToWorkList(NewPtr.getNode());
8841 AddToWorkList(NewLD.getNode());
8842 AddToWorkList(NewVal.getNode());
8843 WorkListRemover DeadNodes(*this);
8844 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLD.getValue(1));
8853 /// TransformFPLoadStorePair - For a given floating point load / store pair,
8854 /// if the load value isn't used by any other operations, then consider
8855 /// transforming the pair to integer load / store operations if the target
8856 /// deems the transformation profitable.
8857 SDValue DAGCombiner::TransformFPLoadStorePair(SDNode *N) {
8858 StoreSDNode *ST = cast<StoreSDNode>(N);
8859 SDValue Chain = ST->getChain();
8860 SDValue Value = ST->getValue();
8861 if (ISD::isNormalStore(ST) && ISD::isNormalLoad(Value.getNode()) &&
8862 Value.hasOneUse() &&
8863 Chain == SDValue(Value.getNode(), 1)) {
8864 LoadSDNode *LD = cast<LoadSDNode>(Value);
8865 EVT VT = LD->getMemoryVT();
8866 if (!VT.isFloatingPoint() ||
8867 VT != ST->getMemoryVT() ||
8868 LD->isNonTemporal() ||
8869 ST->isNonTemporal() ||
8870 LD->getPointerInfo().getAddrSpace() != 0 ||
8871 ST->getPointerInfo().getAddrSpace() != 0)
8874 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
8875 if (!TLI.isOperationLegal(ISD::LOAD, IntVT) ||
8876 !TLI.isOperationLegal(ISD::STORE, IntVT) ||
8877 !TLI.isDesirableToTransformToIntegerOp(ISD::LOAD, VT) ||
8878 !TLI.isDesirableToTransformToIntegerOp(ISD::STORE, VT))
8881 unsigned LDAlign = LD->getAlignment();
8882 unsigned STAlign = ST->getAlignment();
8883 Type *IntVTTy = IntVT.getTypeForEVT(*DAG.getContext());
8884 unsigned ABIAlign = TLI.getDataLayout()->getABITypeAlignment(IntVTTy);
8885 if (LDAlign < ABIAlign || STAlign < ABIAlign)
8888 SDValue NewLD = DAG.getLoad(IntVT, SDLoc(Value),
8889 LD->getChain(), LD->getBasePtr(),
8890 LD->getPointerInfo(),
8891 false, false, false, LDAlign);
8893 SDValue NewST = DAG.getStore(NewLD.getValue(1), SDLoc(N),
8894 NewLD, ST->getBasePtr(),
8895 ST->getPointerInfo(),
8896 false, false, STAlign);
8898 AddToWorkList(NewLD.getNode());
8899 AddToWorkList(NewST.getNode());
8900 WorkListRemover DeadNodes(*this);
8901 DAG.ReplaceAllUsesOfValueWith(Value.getValue(1), NewLD.getValue(1));
8909 /// Helper struct to parse and store a memory address as base + index + offset.
8910 /// We ignore sign extensions when it is safe to do so.
8911 /// The following two expressions are not equivalent. To differentiate we need
8912 /// to store whether there was a sign extension involved in the index
8914 /// (load (i64 add (i64 copyfromreg %c)
8915 /// (i64 signextend (add (i8 load %index)
8919 /// (load (i64 add (i64 copyfromreg %c)
8920 /// (i64 signextend (i32 add (i32 signextend (i8 load %index))
8922 struct BaseIndexOffset {
8926 bool IsIndexSignExt;
8928 BaseIndexOffset() : Offset(0), IsIndexSignExt(false) {}
8930 BaseIndexOffset(SDValue Base, SDValue Index, int64_t Offset,
8931 bool IsIndexSignExt) :
8932 Base(Base), Index(Index), Offset(Offset), IsIndexSignExt(IsIndexSignExt) {}
8934 bool equalBaseIndex(const BaseIndexOffset &Other) {
8935 return Other.Base == Base && Other.Index == Index &&
8936 Other.IsIndexSignExt == IsIndexSignExt;
8939 /// Parses tree in Ptr for base, index, offset addresses.
8940 static BaseIndexOffset match(SDValue Ptr) {
8941 bool IsIndexSignExt = false;
8943 // We only can pattern match BASE + INDEX + OFFSET. If Ptr is not an ADD
8944 // instruction, then it could be just the BASE or everything else we don't
8945 // know how to handle. Just use Ptr as BASE and give up.
8946 if (Ptr->getOpcode() != ISD::ADD)
8947 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
8949 // We know that we have at least an ADD instruction. Try to pattern match
8950 // the simple case of BASE + OFFSET.
8951 if (isa<ConstantSDNode>(Ptr->getOperand(1))) {
8952 int64_t Offset = cast<ConstantSDNode>(Ptr->getOperand(1))->getSExtValue();
8953 return BaseIndexOffset(Ptr->getOperand(0), SDValue(), Offset,
8957 // Inside a loop the current BASE pointer is calculated using an ADD and a
8958 // MUL instruction. In this case Ptr is the actual BASE pointer.
8959 // (i64 add (i64 %array_ptr)
8960 // (i64 mul (i64 %induction_var)
8961 // (i64 %element_size)))
8962 if (Ptr->getOperand(1)->getOpcode() == ISD::MUL)
8963 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
8965 // Look at Base + Index + Offset cases.
8966 SDValue Base = Ptr->getOperand(0);
8967 SDValue IndexOffset = Ptr->getOperand(1);
8969 // Skip signextends.
8970 if (IndexOffset->getOpcode() == ISD::SIGN_EXTEND) {
8971 IndexOffset = IndexOffset->getOperand(0);
8972 IsIndexSignExt = true;
8975 // Either the case of Base + Index (no offset) or something else.
8976 if (IndexOffset->getOpcode() != ISD::ADD)
8977 return BaseIndexOffset(Base, IndexOffset, 0, IsIndexSignExt);
8979 // Now we have the case of Base + Index + offset.
8980 SDValue Index = IndexOffset->getOperand(0);
8981 SDValue Offset = IndexOffset->getOperand(1);
8983 if (!isa<ConstantSDNode>(Offset))
8984 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
8986 // Ignore signextends.
8987 if (Index->getOpcode() == ISD::SIGN_EXTEND) {
8988 Index = Index->getOperand(0);
8989 IsIndexSignExt = true;
8990 } else IsIndexSignExt = false;
8992 int64_t Off = cast<ConstantSDNode>(Offset)->getSExtValue();
8993 return BaseIndexOffset(Base, Index, Off, IsIndexSignExt);
8997 /// Holds a pointer to an LSBaseSDNode as well as information on where it
8998 /// is located in a sequence of memory operations connected by a chain.
9000 MemOpLink (LSBaseSDNode *N, int64_t Offset, unsigned Seq):
9001 MemNode(N), OffsetFromBase(Offset), SequenceNum(Seq) { }
9002 // Ptr to the mem node.
9003 LSBaseSDNode *MemNode;
9004 // Offset from the base ptr.
9005 int64_t OffsetFromBase;
9006 // What is the sequence number of this mem node.
9007 // Lowest mem operand in the DAG starts at zero.
9008 unsigned SequenceNum;
9011 bool DAGCombiner::MergeConsecutiveStores(StoreSDNode* St) {
9012 EVT MemVT = St->getMemoryVT();
9013 int64_t ElementSizeBytes = MemVT.getSizeInBits()/8;
9014 bool NoVectors = DAG.getMachineFunction().getFunction()->getAttributes().
9015 hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat);
9017 // Don't merge vectors into wider inputs.
9018 if (MemVT.isVector() || !MemVT.isSimple())
9021 // Perform an early exit check. Do not bother looking at stored values that
9022 // are not constants or loads.
9023 SDValue StoredVal = St->getValue();
9024 bool IsLoadSrc = isa<LoadSDNode>(StoredVal);
9025 if (!isa<ConstantSDNode>(StoredVal) && !isa<ConstantFPSDNode>(StoredVal) &&
9029 // Only look at ends of store sequences.
9030 SDValue Chain = SDValue(St, 1);
9031 if (Chain->hasOneUse() && Chain->use_begin()->getOpcode() == ISD::STORE)
9034 // This holds the base pointer, index, and the offset in bytes from the base
9036 BaseIndexOffset BasePtr = BaseIndexOffset::match(St->getBasePtr());
9038 // We must have a base and an offset.
9039 if (!BasePtr.Base.getNode())
9042 // Do not handle stores to undef base pointers.
9043 if (BasePtr.Base.getOpcode() == ISD::UNDEF)
9046 // Save the LoadSDNodes that we find in the chain.
9047 // We need to make sure that these nodes do not interfere with
9048 // any of the store nodes.
9049 SmallVector<LSBaseSDNode*, 8> AliasLoadNodes;
9051 // Save the StoreSDNodes that we find in the chain.
9052 SmallVector<MemOpLink, 8> StoreNodes;
9054 // Walk up the chain and look for nodes with offsets from the same
9055 // base pointer. Stop when reaching an instruction with a different kind
9056 // or instruction which has a different base pointer.
9058 StoreSDNode *Index = St;
9060 // If the chain has more than one use, then we can't reorder the mem ops.
9061 if (Index != St && !SDValue(Index, 1)->hasOneUse())
9064 // Find the base pointer and offset for this memory node.
9065 BaseIndexOffset Ptr = BaseIndexOffset::match(Index->getBasePtr());
9067 // Check that the base pointer is the same as the original one.
9068 if (!Ptr.equalBaseIndex(BasePtr))
9071 // Check that the alignment is the same.
9072 if (Index->getAlignment() != St->getAlignment())
9075 // The memory operands must not be volatile.
9076 if (Index->isVolatile() || Index->isIndexed())
9080 if (StoreSDNode *St = dyn_cast<StoreSDNode>(Index))
9081 if (St->isTruncatingStore())
9084 // The stored memory type must be the same.
9085 if (Index->getMemoryVT() != MemVT)
9088 // We do not allow unaligned stores because we want to prevent overriding
9090 if (Index->getAlignment()*8 != MemVT.getSizeInBits())
9093 // We found a potential memory operand to merge.
9094 StoreNodes.push_back(MemOpLink(Index, Ptr.Offset, Seq++));
9096 // Find the next memory operand in the chain. If the next operand in the
9097 // chain is a store then move up and continue the scan with the next
9098 // memory operand. If the next operand is a load save it and use alias
9099 // information to check if it interferes with anything.
9100 SDNode *NextInChain = Index->getChain().getNode();
9102 if (StoreSDNode *STn = dyn_cast<StoreSDNode>(NextInChain)) {
9103 // We found a store node. Use it for the next iteration.
9106 } else if (LoadSDNode *Ldn = dyn_cast<LoadSDNode>(NextInChain)) {
9107 if (Ldn->isVolatile()) {
9112 // Save the load node for later. Continue the scan.
9113 AliasLoadNodes.push_back(Ldn);
9114 NextInChain = Ldn->getChain().getNode();
9123 // Check if there is anything to merge.
9124 if (StoreNodes.size() < 2)
9127 // Sort the memory operands according to their distance from the base pointer.
9128 std::sort(StoreNodes.begin(), StoreNodes.end(),
9129 [](MemOpLink LHS, MemOpLink RHS) {
9130 return LHS.OffsetFromBase < RHS.OffsetFromBase ||
9131 (LHS.OffsetFromBase == RHS.OffsetFromBase &&
9132 LHS.SequenceNum > RHS.SequenceNum);
9135 // Scan the memory operations on the chain and find the first non-consecutive
9136 // store memory address.
9137 unsigned LastConsecutiveStore = 0;
9138 int64_t StartAddress = StoreNodes[0].OffsetFromBase;
9139 for (unsigned i = 0, e = StoreNodes.size(); i < e; ++i) {
9141 // Check that the addresses are consecutive starting from the second
9142 // element in the list of stores.
9144 int64_t CurrAddress = StoreNodes[i].OffsetFromBase;
9145 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
9150 // Check if this store interferes with any of the loads that we found.
9151 for (unsigned ld = 0, lde = AliasLoadNodes.size(); ld < lde; ++ld)
9152 if (isAlias(AliasLoadNodes[ld], StoreNodes[i].MemNode)) {
9156 // We found a load that alias with this store. Stop the sequence.
9160 // Mark this node as useful.
9161 LastConsecutiveStore = i;
9164 // The node with the lowest store address.
9165 LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode;
9167 // Store the constants into memory as one consecutive store.
9169 unsigned LastLegalType = 0;
9170 unsigned LastLegalVectorType = 0;
9171 bool NonZero = false;
9172 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
9173 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
9174 SDValue StoredVal = St->getValue();
9176 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(StoredVal)) {
9177 NonZero |= !C->isNullValue();
9178 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(StoredVal)) {
9179 NonZero |= !C->getConstantFPValue()->isNullValue();
9185 // Find a legal type for the constant store.
9186 unsigned StoreBW = (i+1) * ElementSizeBytes * 8;
9187 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
9188 if (TLI.isTypeLegal(StoreTy))
9189 LastLegalType = i+1;
9190 // Or check whether a truncstore is legal.
9191 else if (TLI.getTypeAction(*DAG.getContext(), StoreTy) ==
9192 TargetLowering::TypePromoteInteger) {
9193 EVT LegalizedStoredValueTy =
9194 TLI.getTypeToTransformTo(*DAG.getContext(), StoredVal.getValueType());
9195 if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy))
9196 LastLegalType = i+1;
9199 // Find a legal type for the vector store.
9200 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
9201 if (TLI.isTypeLegal(Ty))
9202 LastLegalVectorType = i + 1;
9205 // We only use vectors if the constant is known to be zero and the
9206 // function is not marked with the noimplicitfloat attribute.
9207 if (NonZero || NoVectors)
9208 LastLegalVectorType = 0;
9210 // Check if we found a legal integer type to store.
9211 if (LastLegalType == 0 && LastLegalVectorType == 0)
9214 bool UseVector = (LastLegalVectorType > LastLegalType) && !NoVectors;
9215 unsigned NumElem = UseVector ? LastLegalVectorType : LastLegalType;
9217 // Make sure we have something to merge.
9221 unsigned EarliestNodeUsed = 0;
9222 for (unsigned i=0; i < NumElem; ++i) {
9223 // Find a chain for the new wide-store operand. Notice that some
9224 // of the store nodes that we found may not be selected for inclusion
9225 // in the wide store. The chain we use needs to be the chain of the
9226 // earliest store node which is *used* and replaced by the wide store.
9227 if (StoreNodes[i].SequenceNum > StoreNodes[EarliestNodeUsed].SequenceNum)
9228 EarliestNodeUsed = i;
9231 // The earliest Node in the DAG.
9232 LSBaseSDNode *EarliestOp = StoreNodes[EarliestNodeUsed].MemNode;
9233 SDLoc DL(StoreNodes[0].MemNode);
9237 // Find a legal type for the vector store.
9238 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
9239 assert(TLI.isTypeLegal(Ty) && "Illegal vector store");
9240 StoredVal = DAG.getConstant(0, Ty);
9242 unsigned StoreBW = NumElem * ElementSizeBytes * 8;
9243 APInt StoreInt(StoreBW, 0);
9245 // Construct a single integer constant which is made of the smaller
9247 bool IsLE = TLI.isLittleEndian();
9248 for (unsigned i = 0; i < NumElem ; ++i) {
9249 unsigned Idx = IsLE ?(NumElem - 1 - i) : i;
9250 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[Idx].MemNode);
9251 SDValue Val = St->getValue();
9252 StoreInt<<=ElementSizeBytes*8;
9253 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val)) {
9254 StoreInt|=C->getAPIntValue().zext(StoreBW);
9255 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Val)) {
9256 StoreInt|= C->getValueAPF().bitcastToAPInt().zext(StoreBW);
9258 assert(false && "Invalid constant element type");
9262 // Create the new Load and Store operations.
9263 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
9264 StoredVal = DAG.getConstant(StoreInt, StoreTy);
9267 SDValue NewStore = DAG.getStore(EarliestOp->getChain(), DL, StoredVal,
9268 FirstInChain->getBasePtr(),
9269 FirstInChain->getPointerInfo(),
9271 FirstInChain->getAlignment());
9273 // Replace the first store with the new store
9274 CombineTo(EarliestOp, NewStore);
9275 // Erase all other stores.
9276 for (unsigned i = 0; i < NumElem ; ++i) {
9277 if (StoreNodes[i].MemNode == EarliestOp)
9279 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
9280 // ReplaceAllUsesWith will replace all uses that existed when it was
9281 // called, but graph optimizations may cause new ones to appear. For
9282 // example, the case in pr14333 looks like
9284 // St's chain -> St -> another store -> X
9286 // And the only difference from St to the other store is the chain.
9287 // When we change it's chain to be St's chain they become identical,
9288 // get CSEed and the net result is that X is now a use of St.
9289 // Since we know that St is redundant, just iterate.
9290 while (!St->use_empty())
9291 DAG.ReplaceAllUsesWith(SDValue(St, 0), St->getChain());
9292 removeFromWorkList(St);
9299 // Below we handle the case of multiple consecutive stores that
9300 // come from multiple consecutive loads. We merge them into a single
9301 // wide load and a single wide store.
9303 // Look for load nodes which are used by the stored values.
9304 SmallVector<MemOpLink, 8> LoadNodes;
9306 // Find acceptable loads. Loads need to have the same chain (token factor),
9307 // must not be zext, volatile, indexed, and they must be consecutive.
9308 BaseIndexOffset LdBasePtr;
9309 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
9310 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
9311 LoadSDNode *Ld = dyn_cast<LoadSDNode>(St->getValue());
9314 // Loads must only have one use.
9315 if (!Ld->hasNUsesOfValue(1, 0))
9318 // Check that the alignment is the same as the stores.
9319 if (Ld->getAlignment() != St->getAlignment())
9322 // The memory operands must not be volatile.
9323 if (Ld->isVolatile() || Ld->isIndexed())
9326 // We do not accept ext loads.
9327 if (Ld->getExtensionType() != ISD::NON_EXTLOAD)
9330 // The stored memory type must be the same.
9331 if (Ld->getMemoryVT() != MemVT)
9334 BaseIndexOffset LdPtr = BaseIndexOffset::match(Ld->getBasePtr());
9335 // If this is not the first ptr that we check.
9336 if (LdBasePtr.Base.getNode()) {
9337 // The base ptr must be the same.
9338 if (!LdPtr.equalBaseIndex(LdBasePtr))
9341 // Check that all other base pointers are the same as this one.
9345 // We found a potential memory operand to merge.
9346 LoadNodes.push_back(MemOpLink(Ld, LdPtr.Offset, 0));
9349 if (LoadNodes.size() < 2)
9352 // Scan the memory operations on the chain and find the first non-consecutive
9353 // load memory address. These variables hold the index in the store node
9355 unsigned LastConsecutiveLoad = 0;
9356 // This variable refers to the size and not index in the array.
9357 unsigned LastLegalVectorType = 0;
9358 unsigned LastLegalIntegerType = 0;
9359 StartAddress = LoadNodes[0].OffsetFromBase;
9360 SDValue FirstChain = LoadNodes[0].MemNode->getChain();
9361 for (unsigned i = 1; i < LoadNodes.size(); ++i) {
9362 // All loads much share the same chain.
9363 if (LoadNodes[i].MemNode->getChain() != FirstChain)
9366 int64_t CurrAddress = LoadNodes[i].OffsetFromBase;
9367 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
9369 LastConsecutiveLoad = i;
9371 // Find a legal type for the vector store.
9372 EVT StoreTy = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
9373 if (TLI.isTypeLegal(StoreTy))
9374 LastLegalVectorType = i + 1;
9376 // Find a legal type for the integer store.
9377 unsigned StoreBW = (i+1) * ElementSizeBytes * 8;
9378 StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
9379 if (TLI.isTypeLegal(StoreTy))
9380 LastLegalIntegerType = i + 1;
9381 // Or check whether a truncstore and extload is legal.
9382 else if (TLI.getTypeAction(*DAG.getContext(), StoreTy) ==
9383 TargetLowering::TypePromoteInteger) {
9384 EVT LegalizedStoredValueTy =
9385 TLI.getTypeToTransformTo(*DAG.getContext(), StoreTy);
9386 if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy) &&
9387 TLI.isLoadExtLegal(ISD::ZEXTLOAD, StoreTy) &&
9388 TLI.isLoadExtLegal(ISD::SEXTLOAD, StoreTy) &&
9389 TLI.isLoadExtLegal(ISD::EXTLOAD, StoreTy))
9390 LastLegalIntegerType = i+1;
9394 // Only use vector types if the vector type is larger than the integer type.
9395 // If they are the same, use integers.
9396 bool UseVectorTy = LastLegalVectorType > LastLegalIntegerType && !NoVectors;
9397 unsigned LastLegalType = std::max(LastLegalVectorType, LastLegalIntegerType);
9399 // We add +1 here because the LastXXX variables refer to location while
9400 // the NumElem refers to array/index size.
9401 unsigned NumElem = std::min(LastConsecutiveStore, LastConsecutiveLoad) + 1;
9402 NumElem = std::min(LastLegalType, NumElem);
9407 // The earliest Node in the DAG.
9408 unsigned EarliestNodeUsed = 0;
9409 LSBaseSDNode *EarliestOp = StoreNodes[EarliestNodeUsed].MemNode;
9410 for (unsigned i=1; i<NumElem; ++i) {
9411 // Find a chain for the new wide-store operand. Notice that some
9412 // of the store nodes that we found may not be selected for inclusion
9413 // in the wide store. The chain we use needs to be the chain of the
9414 // earliest store node which is *used* and replaced by the wide store.
9415 if (StoreNodes[i].SequenceNum > StoreNodes[EarliestNodeUsed].SequenceNum)
9416 EarliestNodeUsed = i;
9419 // Find if it is better to use vectors or integers to load and store
9423 JointMemOpVT = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
9425 unsigned StoreBW = NumElem * ElementSizeBytes * 8;
9426 JointMemOpVT = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
9429 SDLoc LoadDL(LoadNodes[0].MemNode);
9430 SDLoc StoreDL(StoreNodes[0].MemNode);
9432 LoadSDNode *FirstLoad = cast<LoadSDNode>(LoadNodes[0].MemNode);
9433 SDValue NewLoad = DAG.getLoad(JointMemOpVT, LoadDL,
9434 FirstLoad->getChain(),
9435 FirstLoad->getBasePtr(),
9436 FirstLoad->getPointerInfo(),
9437 false, false, false,
9438 FirstLoad->getAlignment());
9440 SDValue NewStore = DAG.getStore(EarliestOp->getChain(), StoreDL, NewLoad,
9441 FirstInChain->getBasePtr(),
9442 FirstInChain->getPointerInfo(), false, false,
9443 FirstInChain->getAlignment());
9445 // Replace one of the loads with the new load.
9446 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[0].MemNode);
9447 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1),
9448 SDValue(NewLoad.getNode(), 1));
9450 // Remove the rest of the load chains.
9451 for (unsigned i = 1; i < NumElem ; ++i) {
9452 // Replace all chain users of the old load nodes with the chain of the new
9454 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[i].MemNode);
9455 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Ld->getChain());
9458 // Replace the first store with the new store.
9459 CombineTo(EarliestOp, NewStore);
9460 // Erase all other stores.
9461 for (unsigned i = 0; i < NumElem ; ++i) {
9462 // Remove all Store nodes.
9463 if (StoreNodes[i].MemNode == EarliestOp)
9465 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
9466 DAG.ReplaceAllUsesOfValueWith(SDValue(St, 0), St->getChain());
9467 removeFromWorkList(St);
9474 SDValue DAGCombiner::visitSTORE(SDNode *N) {
9475 StoreSDNode *ST = cast<StoreSDNode>(N);
9476 SDValue Chain = ST->getChain();
9477 SDValue Value = ST->getValue();
9478 SDValue Ptr = ST->getBasePtr();
9480 // If this is a store of a bit convert, store the input value if the
9481 // resultant store does not need a higher alignment than the original.
9482 if (Value.getOpcode() == ISD::BITCAST && !ST->isTruncatingStore() &&
9483 ST->isUnindexed()) {
9484 unsigned OrigAlign = ST->getAlignment();
9485 EVT SVT = Value.getOperand(0).getValueType();
9486 unsigned Align = TLI.getDataLayout()->
9487 getABITypeAlignment(SVT.getTypeForEVT(*DAG.getContext()));
9488 if (Align <= OrigAlign &&
9489 ((!LegalOperations && !ST->isVolatile()) ||
9490 TLI.isOperationLegalOrCustom(ISD::STORE, SVT)))
9491 return DAG.getStore(Chain, SDLoc(N), Value.getOperand(0),
9492 Ptr, ST->getPointerInfo(), ST->isVolatile(),
9493 ST->isNonTemporal(), OrigAlign,
9497 // Turn 'store undef, Ptr' -> nothing.
9498 if (Value.getOpcode() == ISD::UNDEF && ST->isUnindexed())
9501 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
9502 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Value)) {
9503 // NOTE: If the original store is volatile, this transform must not increase
9504 // the number of stores. For example, on x86-32 an f64 can be stored in one
9505 // processor operation but an i64 (which is not legal) requires two. So the
9506 // transform should not be done in this case.
9507 if (Value.getOpcode() != ISD::TargetConstantFP) {
9509 switch (CFP->getSimpleValueType(0).SimpleTy) {
9510 default: llvm_unreachable("Unknown FP type");
9511 case MVT::f16: // We don't do this for these yet.
9517 if ((isTypeLegal(MVT::i32) && !LegalOperations && !ST->isVolatile()) ||
9518 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
9519 Tmp = DAG.getConstant((uint32_t)CFP->getValueAPF().
9520 bitcastToAPInt().getZExtValue(), MVT::i32);
9521 return DAG.getStore(Chain, SDLoc(N), Tmp,
9522 Ptr, ST->getMemOperand());
9526 if ((TLI.isTypeLegal(MVT::i64) && !LegalOperations &&
9527 !ST->isVolatile()) ||
9528 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i64)) {
9529 Tmp = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
9530 getZExtValue(), MVT::i64);
9531 return DAG.getStore(Chain, SDLoc(N), Tmp,
9532 Ptr, ST->getMemOperand());
9535 if (!ST->isVolatile() &&
9536 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
9537 // Many FP stores are not made apparent until after legalize, e.g. for
9538 // argument passing. Since this is so common, custom legalize the
9539 // 64-bit integer store into two 32-bit stores.
9540 uint64_t Val = CFP->getValueAPF().bitcastToAPInt().getZExtValue();
9541 SDValue Lo = DAG.getConstant(Val & 0xFFFFFFFF, MVT::i32);
9542 SDValue Hi = DAG.getConstant(Val >> 32, MVT::i32);
9543 if (TLI.isBigEndian()) std::swap(Lo, Hi);
9545 unsigned Alignment = ST->getAlignment();
9546 bool isVolatile = ST->isVolatile();
9547 bool isNonTemporal = ST->isNonTemporal();
9548 const MDNode *TBAAInfo = ST->getTBAAInfo();
9550 SDValue St0 = DAG.getStore(Chain, SDLoc(ST), Lo,
9551 Ptr, ST->getPointerInfo(),
9552 isVolatile, isNonTemporal,
9553 ST->getAlignment(), TBAAInfo);
9554 Ptr = DAG.getNode(ISD::ADD, SDLoc(N), Ptr.getValueType(), Ptr,
9555 DAG.getConstant(4, Ptr.getValueType()));
9556 Alignment = MinAlign(Alignment, 4U);
9557 SDValue St1 = DAG.getStore(Chain, SDLoc(ST), Hi,
9558 Ptr, ST->getPointerInfo().getWithOffset(4),
9559 isVolatile, isNonTemporal,
9560 Alignment, TBAAInfo);
9561 return DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other,
9570 // Try to infer better alignment information than the store already has.
9571 if (OptLevel != CodeGenOpt::None && ST->isUnindexed()) {
9572 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
9573 if (Align > ST->getAlignment())
9574 return DAG.getTruncStore(Chain, SDLoc(N), Value,
9575 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
9576 ST->isVolatile(), ST->isNonTemporal(), Align,
9581 // Try transforming a pair floating point load / store ops to integer
9582 // load / store ops.
9583 SDValue NewST = TransformFPLoadStorePair(N);
9584 if (NewST.getNode())
9587 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA :
9588 TLI.getTargetMachine().getSubtarget<TargetSubtargetInfo>().useAA();
9590 if (CombinerAAOnlyFunc.getNumOccurrences() &&
9591 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
9594 if (UseAA && ST->isUnindexed()) {
9595 // Walk up chain skipping non-aliasing memory nodes.
9596 SDValue BetterChain = FindBetterChain(N, Chain);
9598 // If there is a better chain.
9599 if (Chain != BetterChain) {
9602 // Replace the chain to avoid dependency.
9603 if (ST->isTruncatingStore()) {
9604 ReplStore = DAG.getTruncStore(BetterChain, SDLoc(N), Value, Ptr,
9605 ST->getMemoryVT(), ST->getMemOperand());
9607 ReplStore = DAG.getStore(BetterChain, SDLoc(N), Value, Ptr,
9608 ST->getMemOperand());
9611 // Create token to keep both nodes around.
9612 SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N),
9613 MVT::Other, Chain, ReplStore);
9615 // Make sure the new and old chains are cleaned up.
9616 AddToWorkList(Token.getNode());
9618 // Don't add users to work list.
9619 return CombineTo(N, Token, false);
9623 // Try transforming N to an indexed store.
9624 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
9625 return SDValue(N, 0);
9627 // FIXME: is there such a thing as a truncating indexed store?
9628 if (ST->isTruncatingStore() && ST->isUnindexed() &&
9629 Value.getValueType().isInteger()) {
9630 // See if we can simplify the input to this truncstore with knowledge that
9631 // only the low bits are being used. For example:
9632 // "truncstore (or (shl x, 8), y), i8" -> "truncstore y, i8"
9634 GetDemandedBits(Value,
9635 APInt::getLowBitsSet(
9636 Value.getValueType().getScalarType().getSizeInBits(),
9637 ST->getMemoryVT().getScalarType().getSizeInBits()));
9638 AddToWorkList(Value.getNode());
9639 if (Shorter.getNode())
9640 return DAG.getTruncStore(Chain, SDLoc(N), Shorter,
9641 Ptr, ST->getMemoryVT(), ST->getMemOperand());
9643 // Otherwise, see if we can simplify the operation with
9644 // SimplifyDemandedBits, which only works if the value has a single use.
9645 if (SimplifyDemandedBits(Value,
9646 APInt::getLowBitsSet(
9647 Value.getValueType().getScalarType().getSizeInBits(),
9648 ST->getMemoryVT().getScalarType().getSizeInBits())))
9649 return SDValue(N, 0);
9652 // If this is a load followed by a store to the same location, then the store
9654 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Value)) {
9655 if (Ld->getBasePtr() == Ptr && ST->getMemoryVT() == Ld->getMemoryVT() &&
9656 ST->isUnindexed() && !ST->isVolatile() &&
9657 // There can't be any side effects between the load and store, such as
9659 Chain.reachesChainWithoutSideEffects(SDValue(Ld, 1))) {
9660 // The store is dead, remove it.
9665 // If this is an FP_ROUND or TRUNC followed by a store, fold this into a
9666 // truncating store. We can do this even if this is already a truncstore.
9667 if ((Value.getOpcode() == ISD::FP_ROUND || Value.getOpcode() == ISD::TRUNCATE)
9668 && Value.getNode()->hasOneUse() && ST->isUnindexed() &&
9669 TLI.isTruncStoreLegal(Value.getOperand(0).getValueType(),
9670 ST->getMemoryVT())) {
9671 return DAG.getTruncStore(Chain, SDLoc(N), Value.getOperand(0),
9672 Ptr, ST->getMemoryVT(), ST->getMemOperand());
9675 // Only perform this optimization before the types are legal, because we
9676 // don't want to perform this optimization on every DAGCombine invocation.
9678 bool EverChanged = false;
9681 // There can be multiple store sequences on the same chain.
9682 // Keep trying to merge store sequences until we are unable to do so
9683 // or until we merge the last store on the chain.
9684 bool Changed = MergeConsecutiveStores(ST);
9685 EverChanged |= Changed;
9686 if (!Changed) break;
9687 } while (ST->getOpcode() != ISD::DELETED_NODE);
9690 return SDValue(N, 0);
9693 return ReduceLoadOpStoreWidth(N);
9696 SDValue DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) {
9697 SDValue InVec = N->getOperand(0);
9698 SDValue InVal = N->getOperand(1);
9699 SDValue EltNo = N->getOperand(2);
9702 // If the inserted element is an UNDEF, just use the input vector.
9703 if (InVal.getOpcode() == ISD::UNDEF)
9706 EVT VT = InVec.getValueType();
9708 // If we can't generate a legal BUILD_VECTOR, exit
9709 if (LegalOperations && !TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
9712 // Check that we know which element is being inserted
9713 if (!isa<ConstantSDNode>(EltNo))
9715 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
9717 // Check that the operand is a BUILD_VECTOR (or UNDEF, which can essentially
9718 // be converted to a BUILD_VECTOR). Fill in the Ops vector with the
9720 SmallVector<SDValue, 8> Ops;
9721 // Do not combine these two vectors if the output vector will not replace
9722 // the input vector.
9723 if (InVec.getOpcode() == ISD::BUILD_VECTOR && InVec.hasOneUse()) {
9724 Ops.append(InVec.getNode()->op_begin(),
9725 InVec.getNode()->op_end());
9726 } else if (InVec.getOpcode() == ISD::UNDEF) {
9727 unsigned NElts = VT.getVectorNumElements();
9728 Ops.append(NElts, DAG.getUNDEF(InVal.getValueType()));
9733 // Insert the element
9734 if (Elt < Ops.size()) {
9735 // All the operands of BUILD_VECTOR must have the same type;
9736 // we enforce that here.
9737 EVT OpVT = Ops[0].getValueType();
9738 if (InVal.getValueType() != OpVT)
9739 InVal = OpVT.bitsGT(InVal.getValueType()) ?
9740 DAG.getNode(ISD::ANY_EXTEND, dl, OpVT, InVal) :
9741 DAG.getNode(ISD::TRUNCATE, dl, OpVT, InVal);
9745 // Return the new vector
9746 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
9749 SDValue DAGCombiner::ReplaceExtractVectorEltOfLoadWithNarrowedLoad(
9750 SDNode *EVE, EVT InVecVT, SDValue EltNo, LoadSDNode *OriginalLoad) {
9751 EVT ResultVT = EVE->getValueType(0);
9752 EVT VecEltVT = InVecVT.getVectorElementType();
9753 unsigned Align = OriginalLoad->getAlignment();
9754 unsigned NewAlign = TLI.getDataLayout()->getABITypeAlignment(
9755 VecEltVT.getTypeForEVT(*DAG.getContext()));
9757 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, VecEltVT))
9762 SDValue NewPtr = OriginalLoad->getBasePtr();
9764 EVT PtrType = NewPtr.getValueType();
9765 MachinePointerInfo MPI;
9766 if (auto *ConstEltNo = dyn_cast<ConstantSDNode>(EltNo)) {
9767 int Elt = ConstEltNo->getZExtValue();
9768 unsigned PtrOff = VecEltVT.getSizeInBits() * Elt / 8;
9769 if (TLI.isBigEndian())
9770 PtrOff = InVecVT.getSizeInBits() / 8 - PtrOff;
9771 Offset = DAG.getConstant(PtrOff, PtrType);
9772 MPI = OriginalLoad->getPointerInfo().getWithOffset(PtrOff);
9774 Offset = DAG.getNode(
9775 ISD::MUL, SDLoc(EVE), EltNo.getValueType(), EltNo,
9776 DAG.getConstant(VecEltVT.getStoreSize(), EltNo.getValueType()));
9777 if (TLI.isBigEndian())
9778 Offset = DAG.getNode(
9779 ISD::SUB, SDLoc(EVE), EltNo.getValueType(),
9780 DAG.getConstant(InVecVT.getStoreSize(), EltNo.getValueType()), Offset);
9781 MPI = OriginalLoad->getPointerInfo();
9783 NewPtr = DAG.getNode(ISD::ADD, SDLoc(EVE), PtrType, NewPtr, Offset);
9785 // The replacement we need to do here is a little tricky: we need to
9786 // replace an extractelement of a load with a load.
9787 // Use ReplaceAllUsesOfValuesWith to do the replacement.
9788 // Note that this replacement assumes that the extractvalue is the only
9789 // use of the load; that's okay because we don't want to perform this
9790 // transformation in other cases anyway.
9793 if (ResultVT.bitsGT(VecEltVT)) {
9794 // If the result type of vextract is wider than the load, then issue an
9795 // extending load instead.
9796 ISD::LoadExtType ExtType = TLI.isLoadExtLegal(ISD::ZEXTLOAD, VecEltVT)
9799 Load = DAG.getExtLoad(ExtType, SDLoc(EVE), ResultVT, OriginalLoad->getChain(),
9800 NewPtr, MPI, VecEltVT, OriginalLoad->isVolatile(),
9801 OriginalLoad->isNonTemporal(), Align,
9802 OriginalLoad->getTBAAInfo());
9803 Chain = Load.getValue(1);
9806 VecEltVT, SDLoc(EVE), OriginalLoad->getChain(), NewPtr, MPI,
9807 OriginalLoad->isVolatile(), OriginalLoad->isNonTemporal(),
9808 OriginalLoad->isInvariant(), Align, OriginalLoad->getTBAAInfo());
9809 Chain = Load.getValue(1);
9810 if (ResultVT.bitsLT(VecEltVT))
9811 Load = DAG.getNode(ISD::TRUNCATE, SDLoc(EVE), ResultVT, Load);
9813 Load = DAG.getNode(ISD::BITCAST, SDLoc(EVE), ResultVT, Load);
9815 WorkListRemover DeadNodes(*this);
9816 SDValue From[] = { SDValue(EVE, 0), SDValue(OriginalLoad, 1) };
9817 SDValue To[] = { Load, Chain };
9818 DAG.ReplaceAllUsesOfValuesWith(From, To, 2);
9819 // Since we're explicitly calling ReplaceAllUses, add the new node to the
9820 // worklist explicitly as well.
9821 AddToWorkList(Load.getNode());
9822 AddUsersToWorkList(Load.getNode()); // Add users too
9823 // Make sure to revisit this node to clean it up; it will usually be dead.
9826 return SDValue(EVE, 0);
9829 SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) {
9830 // (vextract (scalar_to_vector val, 0) -> val
9831 SDValue InVec = N->getOperand(0);
9832 EVT VT = InVec.getValueType();
9833 EVT NVT = N->getValueType(0);
9835 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
9836 // Check if the result type doesn't match the inserted element type. A
9837 // SCALAR_TO_VECTOR may truncate the inserted element and the
9838 // EXTRACT_VECTOR_ELT may widen the extracted vector.
9839 SDValue InOp = InVec.getOperand(0);
9840 if (InOp.getValueType() != NVT) {
9841 assert(InOp.getValueType().isInteger() && NVT.isInteger());
9842 return DAG.getSExtOrTrunc(InOp, SDLoc(InVec), NVT);
9847 SDValue EltNo = N->getOperand(1);
9848 bool ConstEltNo = isa<ConstantSDNode>(EltNo);
9850 // Transform: (EXTRACT_VECTOR_ELT( VECTOR_SHUFFLE )) -> EXTRACT_VECTOR_ELT.
9851 // We only perform this optimization before the op legalization phase because
9852 // we may introduce new vector instructions which are not backed by TD
9853 // patterns. For example on AVX, extracting elements from a wide vector
9854 // without using extract_subvector. However, if we can find an underlying
9855 // scalar value, then we can always use that.
9856 if (InVec.getOpcode() == ISD::VECTOR_SHUFFLE
9858 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
9859 int NumElem = VT.getVectorNumElements();
9860 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(InVec);
9861 // Find the new index to extract from.
9862 int OrigElt = SVOp->getMaskElt(Elt);
9864 // Extracting an undef index is undef.
9866 return DAG.getUNDEF(NVT);
9868 // Select the right vector half to extract from.
9870 if (OrigElt < NumElem) {
9871 SVInVec = InVec->getOperand(0);
9873 SVInVec = InVec->getOperand(1);
9877 if (SVInVec.getOpcode() == ISD::BUILD_VECTOR) {
9878 SDValue InOp = SVInVec.getOperand(OrigElt);
9879 if (InOp.getValueType() != NVT) {
9880 assert(InOp.getValueType().isInteger() && NVT.isInteger());
9881 InOp = DAG.getSExtOrTrunc(InOp, SDLoc(SVInVec), NVT);
9887 // FIXME: We should handle recursing on other vector shuffles and
9888 // scalar_to_vector here as well.
9890 if (!LegalOperations) {
9891 EVT IndexTy = TLI.getVectorIdxTy();
9892 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(N), NVT,
9893 SVInVec, DAG.getConstant(OrigElt, IndexTy));
9897 bool BCNumEltsChanged = false;
9898 EVT ExtVT = VT.getVectorElementType();
9901 // If the result of load has to be truncated, then it's not necessarily
9903 if (NVT.bitsLT(LVT) && !TLI.isTruncateFree(LVT, NVT))
9906 if (InVec.getOpcode() == ISD::BITCAST) {
9907 // Don't duplicate a load with other uses.
9908 if (!InVec.hasOneUse())
9911 EVT BCVT = InVec.getOperand(0).getValueType();
9912 if (!BCVT.isVector() || ExtVT.bitsGT(BCVT.getVectorElementType()))
9914 if (VT.getVectorNumElements() != BCVT.getVectorNumElements())
9915 BCNumEltsChanged = true;
9916 InVec = InVec.getOperand(0);
9917 ExtVT = BCVT.getVectorElementType();
9920 // (vextract (vN[if]M load $addr), i) -> ([if]M load $addr + i * size)
9921 if (!LegalOperations && !ConstEltNo && InVec.hasOneUse() &&
9922 ISD::isNormalLoad(InVec.getNode())) {
9923 SDValue Index = N->getOperand(1);
9924 if (LoadSDNode *OrigLoad = dyn_cast<LoadSDNode>(InVec))
9925 return ReplaceExtractVectorEltOfLoadWithNarrowedLoad(N, VT, Index,
9929 // Perform only after legalization to ensure build_vector / vector_shuffle
9930 // optimizations have already been done.
9931 if (!LegalOperations) return SDValue();
9933 // (vextract (v4f32 load $addr), c) -> (f32 load $addr+c*size)
9934 // (vextract (v4f32 s2v (f32 load $addr)), c) -> (f32 load $addr+c*size)
9935 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), 0) -> (f32 load $addr)
9938 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
9940 LoadSDNode *LN0 = nullptr;
9941 const ShuffleVectorSDNode *SVN = nullptr;
9942 if (ISD::isNormalLoad(InVec.getNode())) {
9943 LN0 = cast<LoadSDNode>(InVec);
9944 } else if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR &&
9945 InVec.getOperand(0).getValueType() == ExtVT &&
9946 ISD::isNormalLoad(InVec.getOperand(0).getNode())) {
9947 // Don't duplicate a load with other uses.
9948 if (!InVec.hasOneUse())
9951 LN0 = cast<LoadSDNode>(InVec.getOperand(0));
9952 } else if ((SVN = dyn_cast<ShuffleVectorSDNode>(InVec))) {
9953 // (vextract (vector_shuffle (load $addr), v2, <1, u, u, u>), 1)
9955 // (load $addr+1*size)
9957 // Don't duplicate a load with other uses.
9958 if (!InVec.hasOneUse())
9961 // If the bit convert changed the number of elements, it is unsafe
9962 // to examine the mask.
9963 if (BCNumEltsChanged)
9966 // Select the input vector, guarding against out of range extract vector.
9967 unsigned NumElems = VT.getVectorNumElements();
9968 int Idx = (Elt > (int)NumElems) ? -1 : SVN->getMaskElt(Elt);
9969 InVec = (Idx < (int)NumElems) ? InVec.getOperand(0) : InVec.getOperand(1);
9971 if (InVec.getOpcode() == ISD::BITCAST) {
9972 // Don't duplicate a load with other uses.
9973 if (!InVec.hasOneUse())
9976 InVec = InVec.getOperand(0);
9978 if (ISD::isNormalLoad(InVec.getNode())) {
9979 LN0 = cast<LoadSDNode>(InVec);
9980 Elt = (Idx < (int)NumElems) ? Idx : Idx - (int)NumElems;
9981 EltNo = DAG.getConstant(Elt, EltNo.getValueType());
9985 // Make sure we found a non-volatile load and the extractelement is
9987 if (!LN0 || !LN0->hasNUsesOfValue(1,0) || LN0->isVolatile())
9990 // If Idx was -1 above, Elt is going to be -1, so just return undef.
9992 return DAG.getUNDEF(LVT);
9994 return ReplaceExtractVectorEltOfLoadWithNarrowedLoad(N, VT, EltNo, LN0);
10000 // Simplify (build_vec (ext )) to (bitcast (build_vec ))
10001 SDValue DAGCombiner::reduceBuildVecExtToExtBuildVec(SDNode *N) {
10002 // We perform this optimization post type-legalization because
10003 // the type-legalizer often scalarizes integer-promoted vectors.
10004 // Performing this optimization before may create bit-casts which
10005 // will be type-legalized to complex code sequences.
10006 // We perform this optimization only before the operation legalizer because we
10007 // may introduce illegal operations.
10008 if (Level != AfterLegalizeVectorOps && Level != AfterLegalizeTypes)
10011 unsigned NumInScalars = N->getNumOperands();
10013 EVT VT = N->getValueType(0);
10015 // Check to see if this is a BUILD_VECTOR of a bunch of values
10016 // which come from any_extend or zero_extend nodes. If so, we can create
10017 // a new BUILD_VECTOR using bit-casts which may enable other BUILD_VECTOR
10018 // optimizations. We do not handle sign-extend because we can't fill the sign
10020 EVT SourceType = MVT::Other;
10021 bool AllAnyExt = true;
10023 for (unsigned i = 0; i != NumInScalars; ++i) {
10024 SDValue In = N->getOperand(i);
10025 // Ignore undef inputs.
10026 if (In.getOpcode() == ISD::UNDEF) continue;
10028 bool AnyExt = In.getOpcode() == ISD::ANY_EXTEND;
10029 bool ZeroExt = In.getOpcode() == ISD::ZERO_EXTEND;
10031 // Abort if the element is not an extension.
10032 if (!ZeroExt && !AnyExt) {
10033 SourceType = MVT::Other;
10037 // The input is a ZeroExt or AnyExt. Check the original type.
10038 EVT InTy = In.getOperand(0).getValueType();
10040 // Check that all of the widened source types are the same.
10041 if (SourceType == MVT::Other)
10044 else if (InTy != SourceType) {
10045 // Multiple income types. Abort.
10046 SourceType = MVT::Other;
10050 // Check if all of the extends are ANY_EXTENDs.
10051 AllAnyExt &= AnyExt;
10054 // In order to have valid types, all of the inputs must be extended from the
10055 // same source type and all of the inputs must be any or zero extend.
10056 // Scalar sizes must be a power of two.
10057 EVT OutScalarTy = VT.getScalarType();
10058 bool ValidTypes = SourceType != MVT::Other &&
10059 isPowerOf2_32(OutScalarTy.getSizeInBits()) &&
10060 isPowerOf2_32(SourceType.getSizeInBits());
10062 // Create a new simpler BUILD_VECTOR sequence which other optimizations can
10063 // turn into a single shuffle instruction.
10067 bool isLE = TLI.isLittleEndian();
10068 unsigned ElemRatio = OutScalarTy.getSizeInBits()/SourceType.getSizeInBits();
10069 assert(ElemRatio > 1 && "Invalid element size ratio");
10070 SDValue Filler = AllAnyExt ? DAG.getUNDEF(SourceType):
10071 DAG.getConstant(0, SourceType);
10073 unsigned NewBVElems = ElemRatio * VT.getVectorNumElements();
10074 SmallVector<SDValue, 8> Ops(NewBVElems, Filler);
10076 // Populate the new build_vector
10077 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
10078 SDValue Cast = N->getOperand(i);
10079 assert((Cast.getOpcode() == ISD::ANY_EXTEND ||
10080 Cast.getOpcode() == ISD::ZERO_EXTEND ||
10081 Cast.getOpcode() == ISD::UNDEF) && "Invalid cast opcode");
10083 if (Cast.getOpcode() == ISD::UNDEF)
10084 In = DAG.getUNDEF(SourceType);
10086 In = Cast->getOperand(0);
10087 unsigned Index = isLE ? (i * ElemRatio) :
10088 (i * ElemRatio + (ElemRatio - 1));
10090 assert(Index < Ops.size() && "Invalid index");
10094 // The type of the new BUILD_VECTOR node.
10095 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), SourceType, NewBVElems);
10096 assert(VecVT.getSizeInBits() == VT.getSizeInBits() &&
10097 "Invalid vector size");
10098 // Check if the new vector type is legal.
10099 if (!isTypeLegal(VecVT)) return SDValue();
10101 // Make the new BUILD_VECTOR.
10102 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
10104 // The new BUILD_VECTOR node has the potential to be further optimized.
10105 AddToWorkList(BV.getNode());
10106 // Bitcast to the desired type.
10107 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
10110 SDValue DAGCombiner::reduceBuildVecConvertToConvertBuildVec(SDNode *N) {
10111 EVT VT = N->getValueType(0);
10113 unsigned NumInScalars = N->getNumOperands();
10116 EVT SrcVT = MVT::Other;
10117 unsigned Opcode = ISD::DELETED_NODE;
10118 unsigned NumDefs = 0;
10120 for (unsigned i = 0; i != NumInScalars; ++i) {
10121 SDValue In = N->getOperand(i);
10122 unsigned Opc = In.getOpcode();
10124 if (Opc == ISD::UNDEF)
10127 // If all scalar values are floats and converted from integers.
10128 if (Opcode == ISD::DELETED_NODE &&
10129 (Opc == ISD::UINT_TO_FP || Opc == ISD::SINT_TO_FP)) {
10136 EVT InVT = In.getOperand(0).getValueType();
10138 // If all scalar values are typed differently, bail out. It's chosen to
10139 // simplify BUILD_VECTOR of integer types.
10140 if (SrcVT == MVT::Other)
10147 // If the vector has just one element defined, it's not worth to fold it into
10148 // a vectorized one.
10152 assert((Opcode == ISD::UINT_TO_FP || Opcode == ISD::SINT_TO_FP)
10153 && "Should only handle conversion from integer to float.");
10154 assert(SrcVT != MVT::Other && "Cannot determine source type!");
10156 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SrcVT, NumInScalars);
10158 if (!TLI.isOperationLegalOrCustom(Opcode, NVT))
10161 SmallVector<SDValue, 8> Opnds;
10162 for (unsigned i = 0; i != NumInScalars; ++i) {
10163 SDValue In = N->getOperand(i);
10165 if (In.getOpcode() == ISD::UNDEF)
10166 Opnds.push_back(DAG.getUNDEF(SrcVT));
10168 Opnds.push_back(In.getOperand(0));
10170 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NVT, Opnds);
10171 AddToWorkList(BV.getNode());
10173 return DAG.getNode(Opcode, dl, VT, BV);
10176 SDValue DAGCombiner::visitBUILD_VECTOR(SDNode *N) {
10177 unsigned NumInScalars = N->getNumOperands();
10179 EVT VT = N->getValueType(0);
10181 // A vector built entirely of undefs is undef.
10182 if (ISD::allOperandsUndef(N))
10183 return DAG.getUNDEF(VT);
10185 SDValue V = reduceBuildVecExtToExtBuildVec(N);
10189 V = reduceBuildVecConvertToConvertBuildVec(N);
10193 // Check to see if this is a BUILD_VECTOR of a bunch of EXTRACT_VECTOR_ELT
10194 // operations. If so, and if the EXTRACT_VECTOR_ELT vector inputs come from
10195 // at most two distinct vectors, turn this into a shuffle node.
10197 // May only combine to shuffle after legalize if shuffle is legal.
10198 if (LegalOperations &&
10199 !TLI.isOperationLegalOrCustom(ISD::VECTOR_SHUFFLE, VT))
10202 SDValue VecIn1, VecIn2;
10203 for (unsigned i = 0; i != NumInScalars; ++i) {
10204 // Ignore undef inputs.
10205 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
10207 // If this input is something other than a EXTRACT_VECTOR_ELT with a
10208 // constant index, bail out.
10209 if (N->getOperand(i).getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
10210 !isa<ConstantSDNode>(N->getOperand(i).getOperand(1))) {
10211 VecIn1 = VecIn2 = SDValue(nullptr, 0);
10215 // We allow up to two distinct input vectors.
10216 SDValue ExtractedFromVec = N->getOperand(i).getOperand(0);
10217 if (ExtractedFromVec == VecIn1 || ExtractedFromVec == VecIn2)
10220 if (!VecIn1.getNode()) {
10221 VecIn1 = ExtractedFromVec;
10222 } else if (!VecIn2.getNode()) {
10223 VecIn2 = ExtractedFromVec;
10225 // Too many inputs.
10226 VecIn1 = VecIn2 = SDValue(nullptr, 0);
10231 // If everything is good, we can make a shuffle operation.
10232 if (VecIn1.getNode()) {
10233 SmallVector<int, 8> Mask;
10234 for (unsigned i = 0; i != NumInScalars; ++i) {
10235 if (N->getOperand(i).getOpcode() == ISD::UNDEF) {
10236 Mask.push_back(-1);
10240 // If extracting from the first vector, just use the index directly.
10241 SDValue Extract = N->getOperand(i);
10242 SDValue ExtVal = Extract.getOperand(1);
10243 if (Extract.getOperand(0) == VecIn1) {
10244 unsigned ExtIndex = cast<ConstantSDNode>(ExtVal)->getZExtValue();
10245 if (ExtIndex > VT.getVectorNumElements())
10248 Mask.push_back(ExtIndex);
10252 // Otherwise, use InIdx + VecSize
10253 unsigned Idx = cast<ConstantSDNode>(ExtVal)->getZExtValue();
10254 Mask.push_back(Idx+NumInScalars);
10257 // We can't generate a shuffle node with mismatched input and output types.
10258 // Attempt to transform a single input vector to the correct type.
10259 if ((VT != VecIn1.getValueType())) {
10260 // We don't support shuffeling between TWO values of different types.
10261 if (VecIn2.getNode())
10264 // We only support widening of vectors which are half the size of the
10265 // output registers. For example XMM->YMM widening on X86 with AVX.
10266 if (VecIn1.getValueType().getSizeInBits()*2 != VT.getSizeInBits())
10269 // If the input vector type has a different base type to the output
10270 // vector type, bail out.
10271 if (VecIn1.getValueType().getVectorElementType() !=
10272 VT.getVectorElementType())
10275 // Widen the input vector by adding undef values.
10276 VecIn1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
10277 VecIn1, DAG.getUNDEF(VecIn1.getValueType()));
10280 // If VecIn2 is unused then change it to undef.
10281 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
10283 // Check that we were able to transform all incoming values to the same
10285 if (VecIn2.getValueType() != VecIn1.getValueType() ||
10286 VecIn1.getValueType() != VT)
10289 // Only type-legal BUILD_VECTOR nodes are converted to shuffle nodes.
10290 if (!isTypeLegal(VT))
10293 // Return the new VECTOR_SHUFFLE node.
10297 return DAG.getVectorShuffle(VT, dl, Ops[0], Ops[1], &Mask[0]);
10303 SDValue DAGCombiner::visitCONCAT_VECTORS(SDNode *N) {
10304 // TODO: Check to see if this is a CONCAT_VECTORS of a bunch of
10305 // EXTRACT_SUBVECTOR operations. If so, and if the EXTRACT_SUBVECTOR vector
10306 // inputs come from at most two distinct vectors, turn this into a shuffle
10309 // If we only have one input vector, we don't need to do any concatenation.
10310 if (N->getNumOperands() == 1)
10311 return N->getOperand(0);
10313 // Check if all of the operands are undefs.
10314 EVT VT = N->getValueType(0);
10315 if (ISD::allOperandsUndef(N))
10316 return DAG.getUNDEF(VT);
10318 // Optimize concat_vectors where one of the vectors is undef.
10319 if (N->getNumOperands() == 2 &&
10320 N->getOperand(1)->getOpcode() == ISD::UNDEF) {
10321 SDValue In = N->getOperand(0);
10322 assert(In.getValueType().isVector() && "Must concat vectors");
10324 // Transform: concat_vectors(scalar, undef) -> scalar_to_vector(sclr).
10325 if (In->getOpcode() == ISD::BITCAST &&
10326 !In->getOperand(0)->getValueType(0).isVector()) {
10327 SDValue Scalar = In->getOperand(0);
10328 EVT SclTy = Scalar->getValueType(0);
10330 if (!SclTy.isFloatingPoint() && !SclTy.isInteger())
10333 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SclTy,
10334 VT.getSizeInBits() / SclTy.getSizeInBits());
10335 if (!TLI.isTypeLegal(NVT) || !TLI.isTypeLegal(Scalar.getValueType()))
10338 SDLoc dl = SDLoc(N);
10339 SDValue Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, NVT, Scalar);
10340 return DAG.getNode(ISD::BITCAST, dl, VT, Res);
10344 // fold (concat_vectors (BUILD_VECTOR A, B, ...), (BUILD_VECTOR C, D, ...))
10345 // -> (BUILD_VECTOR A, B, ..., C, D, ...)
10346 if (N->getNumOperands() == 2 &&
10347 N->getOperand(0).getOpcode() == ISD::BUILD_VECTOR &&
10348 N->getOperand(1).getOpcode() == ISD::BUILD_VECTOR) {
10349 EVT VT = N->getValueType(0);
10350 SDValue N0 = N->getOperand(0);
10351 SDValue N1 = N->getOperand(1);
10352 SmallVector<SDValue, 8> Opnds;
10353 unsigned BuildVecNumElts = N0.getNumOperands();
10355 for (unsigned i = 0; i != BuildVecNumElts; ++i)
10356 Opnds.push_back(N0.getOperand(i));
10357 for (unsigned i = 0; i != BuildVecNumElts; ++i)
10358 Opnds.push_back(N1.getOperand(i));
10360 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Opnds);
10363 // Type legalization of vectors and DAG canonicalization of SHUFFLE_VECTOR
10364 // nodes often generate nop CONCAT_VECTOR nodes.
10365 // Scan the CONCAT_VECTOR operands and look for a CONCAT operations that
10366 // place the incoming vectors at the exact same location.
10367 SDValue SingleSource = SDValue();
10368 unsigned PartNumElem = N->getOperand(0).getValueType().getVectorNumElements();
10370 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
10371 SDValue Op = N->getOperand(i);
10373 if (Op.getOpcode() == ISD::UNDEF)
10376 // Check if this is the identity extract:
10377 if (Op.getOpcode() != ISD::EXTRACT_SUBVECTOR)
10380 // Find the single incoming vector for the extract_subvector.
10381 if (SingleSource.getNode()) {
10382 if (Op.getOperand(0) != SingleSource)
10385 SingleSource = Op.getOperand(0);
10387 // Check the source type is the same as the type of the result.
10388 // If not, this concat may extend the vector, so we can not
10389 // optimize it away.
10390 if (SingleSource.getValueType() != N->getValueType(0))
10394 unsigned IdentityIndex = i * PartNumElem;
10395 ConstantSDNode *CS = dyn_cast<ConstantSDNode>(Op.getOperand(1));
10396 // The extract index must be constant.
10400 // Check that we are reading from the identity index.
10401 if (CS->getZExtValue() != IdentityIndex)
10405 if (SingleSource.getNode())
10406 return SingleSource;
10411 SDValue DAGCombiner::visitEXTRACT_SUBVECTOR(SDNode* N) {
10412 EVT NVT = N->getValueType(0);
10413 SDValue V = N->getOperand(0);
10415 if (V->getOpcode() == ISD::CONCAT_VECTORS) {
10417 // (extract_subvec (concat V1, V2, ...), i)
10420 // Only operand 0 is checked as 'concat' assumes all inputs of the same
10422 if (V->getOperand(0).getValueType() != NVT)
10424 unsigned Idx = dyn_cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
10425 unsigned NumElems = NVT.getVectorNumElements();
10426 assert((Idx % NumElems) == 0 &&
10427 "IDX in concat is not a multiple of the result vector length.");
10428 return V->getOperand(Idx / NumElems);
10432 if (V->getOpcode() == ISD::BITCAST)
10433 V = V.getOperand(0);
10435 if (V->getOpcode() == ISD::INSERT_SUBVECTOR) {
10437 // Handle only simple case where vector being inserted and vector
10438 // being extracted are of same type, and are half size of larger vectors.
10439 EVT BigVT = V->getOperand(0).getValueType();
10440 EVT SmallVT = V->getOperand(1).getValueType();
10441 if (!NVT.bitsEq(SmallVT) || NVT.getSizeInBits()*2 != BigVT.getSizeInBits())
10444 // Only handle cases where both indexes are constants with the same type.
10445 ConstantSDNode *ExtIdx = dyn_cast<ConstantSDNode>(N->getOperand(1));
10446 ConstantSDNode *InsIdx = dyn_cast<ConstantSDNode>(V->getOperand(2));
10448 if (InsIdx && ExtIdx &&
10449 InsIdx->getValueType(0).getSizeInBits() <= 64 &&
10450 ExtIdx->getValueType(0).getSizeInBits() <= 64) {
10452 // (extract_subvec (insert_subvec V1, V2, InsIdx), ExtIdx)
10454 // indices are equal or bit offsets are equal => V1
10455 // otherwise => (extract_subvec V1, ExtIdx)
10456 if (InsIdx->getZExtValue() * SmallVT.getScalarType().getSizeInBits() ==
10457 ExtIdx->getZExtValue() * NVT.getScalarType().getSizeInBits())
10458 return DAG.getNode(ISD::BITCAST, dl, NVT, V->getOperand(1));
10459 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, NVT,
10460 DAG.getNode(ISD::BITCAST, dl,
10461 N->getOperand(0).getValueType(),
10462 V->getOperand(0)), N->getOperand(1));
10469 // Tries to turn a shuffle of two CONCAT_VECTORS into a single concat.
10470 static SDValue partitionShuffleOfConcats(SDNode *N, SelectionDAG &DAG) {
10471 EVT VT = N->getValueType(0);
10472 unsigned NumElts = VT.getVectorNumElements();
10474 SDValue N0 = N->getOperand(0);
10475 SDValue N1 = N->getOperand(1);
10476 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
10478 SmallVector<SDValue, 4> Ops;
10479 EVT ConcatVT = N0.getOperand(0).getValueType();
10480 unsigned NumElemsPerConcat = ConcatVT.getVectorNumElements();
10481 unsigned NumConcats = NumElts / NumElemsPerConcat;
10483 // Look at every vector that's inserted. We're looking for exact
10484 // subvector-sized copies from a concatenated vector
10485 for (unsigned I = 0; I != NumConcats; ++I) {
10486 // Make sure we're dealing with a copy.
10487 unsigned Begin = I * NumElemsPerConcat;
10488 bool AllUndef = true, NoUndef = true;
10489 for (unsigned J = Begin; J != Begin + NumElemsPerConcat; ++J) {
10490 if (SVN->getMaskElt(J) >= 0)
10497 if (SVN->getMaskElt(Begin) % NumElemsPerConcat != 0)
10500 for (unsigned J = 1; J != NumElemsPerConcat; ++J)
10501 if (SVN->getMaskElt(Begin + J - 1) + 1 != SVN->getMaskElt(Begin + J))
10504 unsigned FirstElt = SVN->getMaskElt(Begin) / NumElemsPerConcat;
10505 if (FirstElt < N0.getNumOperands())
10506 Ops.push_back(N0.getOperand(FirstElt));
10508 Ops.push_back(N1.getOperand(FirstElt - N0.getNumOperands()));
10510 } else if (AllUndef) {
10511 Ops.push_back(DAG.getUNDEF(N0.getOperand(0).getValueType()));
10512 } else { // Mixed with general masks and undefs, can't do optimization.
10517 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Ops);
10520 SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
10521 EVT VT = N->getValueType(0);
10522 unsigned NumElts = VT.getVectorNumElements();
10524 SDValue N0 = N->getOperand(0);
10525 SDValue N1 = N->getOperand(1);
10527 assert(N0.getValueType() == VT && "Vector shuffle must be normalized in DAG");
10529 // Canonicalize shuffle undef, undef -> undef
10530 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
10531 return DAG.getUNDEF(VT);
10533 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
10535 // Canonicalize shuffle v, v -> v, undef
10537 SmallVector<int, 8> NewMask;
10538 for (unsigned i = 0; i != NumElts; ++i) {
10539 int Idx = SVN->getMaskElt(i);
10540 if (Idx >= (int)NumElts) Idx -= NumElts;
10541 NewMask.push_back(Idx);
10543 return DAG.getVectorShuffle(VT, SDLoc(N), N0, DAG.getUNDEF(VT),
10547 // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask.
10548 if (N0.getOpcode() == ISD::UNDEF) {
10549 SmallVector<int, 8> NewMask;
10550 for (unsigned i = 0; i != NumElts; ++i) {
10551 int Idx = SVN->getMaskElt(i);
10553 if (Idx >= (int)NumElts)
10556 Idx = -1; // remove reference to lhs
10558 NewMask.push_back(Idx);
10560 return DAG.getVectorShuffle(VT, SDLoc(N), N1, DAG.getUNDEF(VT),
10564 // Remove references to rhs if it is undef
10565 if (N1.getOpcode() == ISD::UNDEF) {
10566 bool Changed = false;
10567 SmallVector<int, 8> NewMask;
10568 for (unsigned i = 0; i != NumElts; ++i) {
10569 int Idx = SVN->getMaskElt(i);
10570 if (Idx >= (int)NumElts) {
10574 NewMask.push_back(Idx);
10577 return DAG.getVectorShuffle(VT, SDLoc(N), N0, N1, &NewMask[0]);
10580 // If it is a splat, check if the argument vector is another splat or a
10581 // build_vector with all scalar elements the same.
10582 if (SVN->isSplat() && SVN->getSplatIndex() < (int)NumElts) {
10583 SDNode *V = N0.getNode();
10585 // If this is a bit convert that changes the element type of the vector but
10586 // not the number of vector elements, look through it. Be careful not to
10587 // look though conversions that change things like v4f32 to v2f64.
10588 if (V->getOpcode() == ISD::BITCAST) {
10589 SDValue ConvInput = V->getOperand(0);
10590 if (ConvInput.getValueType().isVector() &&
10591 ConvInput.getValueType().getVectorNumElements() == NumElts)
10592 V = ConvInput.getNode();
10595 if (V->getOpcode() == ISD::BUILD_VECTOR) {
10596 assert(V->getNumOperands() == NumElts &&
10597 "BUILD_VECTOR has wrong number of operands");
10599 bool AllSame = true;
10600 for (unsigned i = 0; i != NumElts; ++i) {
10601 if (V->getOperand(i).getOpcode() != ISD::UNDEF) {
10602 Base = V->getOperand(i);
10606 // Splat of <u, u, u, u>, return <u, u, u, u>
10607 if (!Base.getNode())
10609 for (unsigned i = 0; i != NumElts; ++i) {
10610 if (V->getOperand(i) != Base) {
10615 // Splat of <x, x, x, x>, return <x, x, x, x>
10621 if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
10622 Level < AfterLegalizeVectorOps &&
10623 (N1.getOpcode() == ISD::UNDEF ||
10624 (N1.getOpcode() == ISD::CONCAT_VECTORS &&
10625 N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()))) {
10626 SDValue V = partitionShuffleOfConcats(N, DAG);
10632 // If this shuffle node is simply a swizzle of another shuffle node,
10633 // and it reverses the swizzle of the previous shuffle then we can
10634 // optimize shuffle(shuffle(x, undef), undef) -> x.
10635 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
10636 N1.getOpcode() == ISD::UNDEF) {
10638 ShuffleVectorSDNode *OtherSV = cast<ShuffleVectorSDNode>(N0);
10640 // Shuffle nodes can only reverse shuffles with a single non-undef value.
10641 if (N0.getOperand(1).getOpcode() != ISD::UNDEF)
10644 // The incoming shuffle must be of the same type as the result of the
10645 // current shuffle.
10646 assert(OtherSV->getOperand(0).getValueType() == VT &&
10647 "Shuffle types don't match");
10649 for (unsigned i = 0; i != NumElts; ++i) {
10650 int Idx = SVN->getMaskElt(i);
10651 assert(Idx < (int)NumElts && "Index references undef operand");
10652 // Next, this index comes from the first value, which is the incoming
10653 // shuffle. Adopt the incoming index.
10655 Idx = OtherSV->getMaskElt(Idx);
10657 // The combined shuffle must map each index to itself.
10658 if (Idx >= 0 && (unsigned)Idx != i)
10662 return OtherSV->getOperand(0);
10668 SDValue DAGCombiner::visitINSERT_SUBVECTOR(SDNode *N) {
10669 SDValue N0 = N->getOperand(0);
10670 SDValue N2 = N->getOperand(2);
10672 // If the input vector is a concatenation, and the insert replaces
10673 // one of the halves, we can optimize into a single concat_vectors.
10674 if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
10675 N0->getNumOperands() == 2 && N2.getOpcode() == ISD::Constant) {
10676 APInt InsIdx = cast<ConstantSDNode>(N2)->getAPIntValue();
10677 EVT VT = N->getValueType(0);
10679 // Lower half: fold (insert_subvector (concat_vectors X, Y), Z) ->
10680 // (concat_vectors Z, Y)
10682 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
10683 N->getOperand(1), N0.getOperand(1));
10685 // Upper half: fold (insert_subvector (concat_vectors X, Y), Z) ->
10686 // (concat_vectors X, Z)
10687 if (InsIdx == VT.getVectorNumElements()/2)
10688 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
10689 N0.getOperand(0), N->getOperand(1));
10695 /// XformToShuffleWithZero - Returns a vector_shuffle if it able to transform
10696 /// an AND to a vector_shuffle with the destination vector and a zero vector.
10697 /// e.g. AND V, <0xffffffff, 0, 0xffffffff, 0>. ==>
10698 /// vector_shuffle V, Zero, <0, 4, 2, 4>
10699 SDValue DAGCombiner::XformToShuffleWithZero(SDNode *N) {
10700 EVT VT = N->getValueType(0);
10702 SDValue LHS = N->getOperand(0);
10703 SDValue RHS = N->getOperand(1);
10704 if (N->getOpcode() == ISD::AND) {
10705 if (RHS.getOpcode() == ISD::BITCAST)
10706 RHS = RHS.getOperand(0);
10707 if (RHS.getOpcode() == ISD::BUILD_VECTOR) {
10708 SmallVector<int, 8> Indices;
10709 unsigned NumElts = RHS.getNumOperands();
10710 for (unsigned i = 0; i != NumElts; ++i) {
10711 SDValue Elt = RHS.getOperand(i);
10712 if (!isa<ConstantSDNode>(Elt))
10715 if (cast<ConstantSDNode>(Elt)->isAllOnesValue())
10716 Indices.push_back(i);
10717 else if (cast<ConstantSDNode>(Elt)->isNullValue())
10718 Indices.push_back(NumElts);
10723 // Let's see if the target supports this vector_shuffle.
10724 EVT RVT = RHS.getValueType();
10725 if (!TLI.isVectorClearMaskLegal(Indices, RVT))
10728 // Return the new VECTOR_SHUFFLE node.
10729 EVT EltVT = RVT.getVectorElementType();
10730 SmallVector<SDValue,8> ZeroOps(RVT.getVectorNumElements(),
10731 DAG.getConstant(0, EltVT));
10732 SDValue Zero = DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), RVT, ZeroOps);
10733 LHS = DAG.getNode(ISD::BITCAST, dl, RVT, LHS);
10734 SDValue Shuf = DAG.getVectorShuffle(RVT, dl, LHS, Zero, &Indices[0]);
10735 return DAG.getNode(ISD::BITCAST, dl, VT, Shuf);
10742 /// SimplifyVBinOp - Visit a binary vector operation, like ADD.
10743 SDValue DAGCombiner::SimplifyVBinOp(SDNode *N) {
10744 assert(N->getValueType(0).isVector() &&
10745 "SimplifyVBinOp only works on vectors!");
10747 SDValue LHS = N->getOperand(0);
10748 SDValue RHS = N->getOperand(1);
10749 SDValue Shuffle = XformToShuffleWithZero(N);
10750 if (Shuffle.getNode()) return Shuffle;
10752 // If the LHS and RHS are BUILD_VECTOR nodes, see if we can constant fold
10754 if (LHS.getOpcode() == ISD::BUILD_VECTOR &&
10755 RHS.getOpcode() == ISD::BUILD_VECTOR) {
10756 // Check if both vectors are constants. If not bail out.
10757 if (!(cast<BuildVectorSDNode>(LHS)->isConstant() &&
10758 cast<BuildVectorSDNode>(RHS)->isConstant()))
10761 SmallVector<SDValue, 8> Ops;
10762 for (unsigned i = 0, e = LHS.getNumOperands(); i != e; ++i) {
10763 SDValue LHSOp = LHS.getOperand(i);
10764 SDValue RHSOp = RHS.getOperand(i);
10766 // Can't fold divide by zero.
10767 if (N->getOpcode() == ISD::SDIV || N->getOpcode() == ISD::UDIV ||
10768 N->getOpcode() == ISD::FDIV) {
10769 if ((RHSOp.getOpcode() == ISD::Constant &&
10770 cast<ConstantSDNode>(RHSOp.getNode())->isNullValue()) ||
10771 (RHSOp.getOpcode() == ISD::ConstantFP &&
10772 cast<ConstantFPSDNode>(RHSOp.getNode())->getValueAPF().isZero()))
10776 EVT VT = LHSOp.getValueType();
10777 EVT RVT = RHSOp.getValueType();
10779 // Integer BUILD_VECTOR operands may have types larger than the element
10780 // size (e.g., when the element type is not legal). Prior to type
10781 // legalization, the types may not match between the two BUILD_VECTORS.
10782 // Truncate one of the operands to make them match.
10783 if (RVT.getSizeInBits() > VT.getSizeInBits()) {
10784 RHSOp = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, RHSOp);
10786 LHSOp = DAG.getNode(ISD::TRUNCATE, SDLoc(N), RVT, LHSOp);
10790 SDValue FoldOp = DAG.getNode(N->getOpcode(), SDLoc(LHS), VT,
10792 if (FoldOp.getOpcode() != ISD::UNDEF &&
10793 FoldOp.getOpcode() != ISD::Constant &&
10794 FoldOp.getOpcode() != ISD::ConstantFP)
10796 Ops.push_back(FoldOp);
10797 AddToWorkList(FoldOp.getNode());
10800 if (Ops.size() == LHS.getNumOperands())
10801 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), LHS.getValueType(), Ops);
10804 // Type legalization might introduce new shuffles in the DAG.
10805 // Fold (VBinOp (shuffle (A, Undef, Mask)), (shuffle (B, Undef, Mask)))
10806 // -> (shuffle (VBinOp (A, B)), Undef, Mask).
10807 if (LegalTypes && isa<ShuffleVectorSDNode>(LHS) &&
10808 isa<ShuffleVectorSDNode>(RHS) && LHS.hasOneUse() && RHS.hasOneUse() &&
10809 LHS.getOperand(1).getOpcode() == ISD::UNDEF &&
10810 RHS.getOperand(1).getOpcode() == ISD::UNDEF) {
10811 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(LHS);
10812 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(RHS);
10814 if (SVN0->getMask().equals(SVN1->getMask())) {
10815 EVT VT = N->getValueType(0);
10816 SDValue UndefVector = LHS.getOperand(1);
10817 SDValue NewBinOp = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
10818 LHS.getOperand(0), RHS.getOperand(0));
10819 AddUsersToWorkList(N);
10820 return DAG.getVectorShuffle(VT, SDLoc(N), NewBinOp, UndefVector,
10821 &SVN0->getMask()[0]);
10828 /// SimplifyVUnaryOp - Visit a binary vector operation, like FABS/FNEG.
10829 SDValue DAGCombiner::SimplifyVUnaryOp(SDNode *N) {
10830 assert(N->getValueType(0).isVector() &&
10831 "SimplifyVUnaryOp only works on vectors!");
10833 SDValue N0 = N->getOperand(0);
10835 if (N0.getOpcode() != ISD::BUILD_VECTOR)
10838 // Operand is a BUILD_VECTOR node, see if we can constant fold it.
10839 SmallVector<SDValue, 8> Ops;
10840 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) {
10841 SDValue Op = N0.getOperand(i);
10842 if (Op.getOpcode() != ISD::UNDEF &&
10843 Op.getOpcode() != ISD::ConstantFP)
10845 EVT EltVT = Op.getValueType();
10846 SDValue FoldOp = DAG.getNode(N->getOpcode(), SDLoc(N0), EltVT, Op);
10847 if (FoldOp.getOpcode() != ISD::UNDEF &&
10848 FoldOp.getOpcode() != ISD::ConstantFP)
10850 Ops.push_back(FoldOp);
10851 AddToWorkList(FoldOp.getNode());
10854 if (Ops.size() != N0.getNumOperands())
10857 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), N0.getValueType(), Ops);
10860 SDValue DAGCombiner::SimplifySelect(SDLoc DL, SDValue N0,
10861 SDValue N1, SDValue N2){
10862 assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!");
10864 SDValue SCC = SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1), N1, N2,
10865 cast<CondCodeSDNode>(N0.getOperand(2))->get());
10867 // If we got a simplified select_cc node back from SimplifySelectCC, then
10868 // break it down into a new SETCC node, and a new SELECT node, and then return
10869 // the SELECT node, since we were called with a SELECT node.
10870 if (SCC.getNode()) {
10871 // Check to see if we got a select_cc back (to turn into setcc/select).
10872 // Otherwise, just return whatever node we got back, like fabs.
10873 if (SCC.getOpcode() == ISD::SELECT_CC) {
10874 SDValue SETCC = DAG.getNode(ISD::SETCC, SDLoc(N0),
10876 SCC.getOperand(0), SCC.getOperand(1),
10877 SCC.getOperand(4));
10878 AddToWorkList(SETCC.getNode());
10879 return DAG.getSelect(SDLoc(SCC), SCC.getValueType(),
10880 SCC.getOperand(2), SCC.getOperand(3), SETCC);
10888 /// SimplifySelectOps - Given a SELECT or a SELECT_CC node, where LHS and RHS
10889 /// are the two values being selected between, see if we can simplify the
10890 /// select. Callers of this should assume that TheSelect is deleted if this
10891 /// returns true. As such, they should return the appropriate thing (e.g. the
10892 /// node) back to the top-level of the DAG combiner loop to avoid it being
10894 bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDValue LHS,
10897 // Cannot simplify select with vector condition
10898 if (TheSelect->getOperand(0).getValueType().isVector()) return false;
10900 // If this is a select from two identical things, try to pull the operation
10901 // through the select.
10902 if (LHS.getOpcode() != RHS.getOpcode() ||
10903 !LHS.hasOneUse() || !RHS.hasOneUse())
10906 // If this is a load and the token chain is identical, replace the select
10907 // of two loads with a load through a select of the address to load from.
10908 // This triggers in things like "select bool X, 10.0, 123.0" after the FP
10909 // constants have been dropped into the constant pool.
10910 if (LHS.getOpcode() == ISD::LOAD) {
10911 LoadSDNode *LLD = cast<LoadSDNode>(LHS);
10912 LoadSDNode *RLD = cast<LoadSDNode>(RHS);
10914 // Token chains must be identical.
10915 if (LHS.getOperand(0) != RHS.getOperand(0) ||
10916 // Do not let this transformation reduce the number of volatile loads.
10917 LLD->isVolatile() || RLD->isVolatile() ||
10918 // If this is an EXTLOAD, the VT's must match.
10919 LLD->getMemoryVT() != RLD->getMemoryVT() ||
10920 // If this is an EXTLOAD, the kind of extension must match.
10921 (LLD->getExtensionType() != RLD->getExtensionType() &&
10922 // The only exception is if one of the extensions is anyext.
10923 LLD->getExtensionType() != ISD::EXTLOAD &&
10924 RLD->getExtensionType() != ISD::EXTLOAD) ||
10925 // FIXME: this discards src value information. This is
10926 // over-conservative. It would be beneficial to be able to remember
10927 // both potential memory locations. Since we are discarding
10928 // src value info, don't do the transformation if the memory
10929 // locations are not in the default address space.
10930 LLD->getPointerInfo().getAddrSpace() != 0 ||
10931 RLD->getPointerInfo().getAddrSpace() != 0 ||
10932 !TLI.isOperationLegalOrCustom(TheSelect->getOpcode(),
10933 LLD->getBasePtr().getValueType()))
10936 // Check that the select condition doesn't reach either load. If so,
10937 // folding this will induce a cycle into the DAG. If not, this is safe to
10938 // xform, so create a select of the addresses.
10940 if (TheSelect->getOpcode() == ISD::SELECT) {
10941 SDNode *CondNode = TheSelect->getOperand(0).getNode();
10942 if ((LLD->hasAnyUseOfValue(1) && LLD->isPredecessorOf(CondNode)) ||
10943 (RLD->hasAnyUseOfValue(1) && RLD->isPredecessorOf(CondNode)))
10945 // The loads must not depend on one another.
10946 if (LLD->isPredecessorOf(RLD) ||
10947 RLD->isPredecessorOf(LLD))
10949 Addr = DAG.getSelect(SDLoc(TheSelect),
10950 LLD->getBasePtr().getValueType(),
10951 TheSelect->getOperand(0), LLD->getBasePtr(),
10952 RLD->getBasePtr());
10953 } else { // Otherwise SELECT_CC
10954 SDNode *CondLHS = TheSelect->getOperand(0).getNode();
10955 SDNode *CondRHS = TheSelect->getOperand(1).getNode();
10957 if ((LLD->hasAnyUseOfValue(1) &&
10958 (LLD->isPredecessorOf(CondLHS) || LLD->isPredecessorOf(CondRHS))) ||
10959 (RLD->hasAnyUseOfValue(1) &&
10960 (RLD->isPredecessorOf(CondLHS) || RLD->isPredecessorOf(CondRHS))))
10963 Addr = DAG.getNode(ISD::SELECT_CC, SDLoc(TheSelect),
10964 LLD->getBasePtr().getValueType(),
10965 TheSelect->getOperand(0),
10966 TheSelect->getOperand(1),
10967 LLD->getBasePtr(), RLD->getBasePtr(),
10968 TheSelect->getOperand(4));
10972 if (LLD->getExtensionType() == ISD::NON_EXTLOAD) {
10973 Load = DAG.getLoad(TheSelect->getValueType(0),
10975 // FIXME: Discards pointer and TBAA info.
10976 LLD->getChain(), Addr, MachinePointerInfo(),
10977 LLD->isVolatile(), LLD->isNonTemporal(),
10978 LLD->isInvariant(), LLD->getAlignment());
10980 Load = DAG.getExtLoad(LLD->getExtensionType() == ISD::EXTLOAD ?
10981 RLD->getExtensionType() : LLD->getExtensionType(),
10983 TheSelect->getValueType(0),
10984 // FIXME: Discards pointer and TBAA info.
10985 LLD->getChain(), Addr, MachinePointerInfo(),
10986 LLD->getMemoryVT(), LLD->isVolatile(),
10987 LLD->isNonTemporal(), LLD->getAlignment());
10990 // Users of the select now use the result of the load.
10991 CombineTo(TheSelect, Load);
10993 // Users of the old loads now use the new load's chain. We know the
10994 // old-load value is dead now.
10995 CombineTo(LHS.getNode(), Load.getValue(0), Load.getValue(1));
10996 CombineTo(RHS.getNode(), Load.getValue(0), Load.getValue(1));
11003 /// SimplifySelectCC - Simplify an expression of the form (N0 cond N1) ? N2 : N3
11004 /// where 'cond' is the comparison specified by CC.
11005 SDValue DAGCombiner::SimplifySelectCC(SDLoc DL, SDValue N0, SDValue N1,
11006 SDValue N2, SDValue N3,
11007 ISD::CondCode CC, bool NotExtCompare) {
11008 // (x ? y : y) -> y.
11009 if (N2 == N3) return N2;
11011 EVT VT = N2.getValueType();
11012 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
11013 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
11014 ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N3.getNode());
11016 // Determine if the condition we're dealing with is constant
11017 SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()),
11018 N0, N1, CC, DL, false);
11019 if (SCC.getNode()) AddToWorkList(SCC.getNode());
11020 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode());
11022 // fold select_cc true, x, y -> x
11023 if (SCCC && !SCCC->isNullValue())
11025 // fold select_cc false, x, y -> y
11026 if (SCCC && SCCC->isNullValue())
11029 // Check to see if we can simplify the select into an fabs node
11030 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) {
11031 // Allow either -0.0 or 0.0
11032 if (CFP->getValueAPF().isZero()) {
11033 // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs
11034 if ((CC == ISD::SETGE || CC == ISD::SETGT) &&
11035 N0 == N2 && N3.getOpcode() == ISD::FNEG &&
11036 N2 == N3.getOperand(0))
11037 return DAG.getNode(ISD::FABS, DL, VT, N0);
11039 // select (setl[te] X, +/-0.0), fneg(X), X -> fabs
11040 if ((CC == ISD::SETLT || CC == ISD::SETLE) &&
11041 N0 == N3 && N2.getOpcode() == ISD::FNEG &&
11042 N2.getOperand(0) == N3)
11043 return DAG.getNode(ISD::FABS, DL, VT, N3);
11047 // Turn "(a cond b) ? 1.0f : 2.0f" into "load (tmp + ((a cond b) ? 0 : 4)"
11048 // where "tmp" is a constant pool entry containing an array with 1.0 and 2.0
11049 // in it. This is a win when the constant is not otherwise available because
11050 // it replaces two constant pool loads with one. We only do this if the FP
11051 // type is known to be legal, because if it isn't, then we are before legalize
11052 // types an we want the other legalization to happen first (e.g. to avoid
11053 // messing with soft float) and if the ConstantFP is not legal, because if
11054 // it is legal, we may not need to store the FP constant in a constant pool.
11055 if (ConstantFPSDNode *TV = dyn_cast<ConstantFPSDNode>(N2))
11056 if (ConstantFPSDNode *FV = dyn_cast<ConstantFPSDNode>(N3)) {
11057 if (TLI.isTypeLegal(N2.getValueType()) &&
11058 (TLI.getOperationAction(ISD::ConstantFP, N2.getValueType()) !=
11059 TargetLowering::Legal &&
11060 !TLI.isFPImmLegal(TV->getValueAPF(), TV->getValueType(0)) &&
11061 !TLI.isFPImmLegal(FV->getValueAPF(), FV->getValueType(0))) &&
11062 // If both constants have multiple uses, then we won't need to do an
11063 // extra load, they are likely around in registers for other users.
11064 (TV->hasOneUse() || FV->hasOneUse())) {
11065 Constant *Elts[] = {
11066 const_cast<ConstantFP*>(FV->getConstantFPValue()),
11067 const_cast<ConstantFP*>(TV->getConstantFPValue())
11069 Type *FPTy = Elts[0]->getType();
11070 const DataLayout &TD = *TLI.getDataLayout();
11072 // Create a ConstantArray of the two constants.
11073 Constant *CA = ConstantArray::get(ArrayType::get(FPTy, 2), Elts);
11074 SDValue CPIdx = DAG.getConstantPool(CA, TLI.getPointerTy(),
11075 TD.getPrefTypeAlignment(FPTy));
11076 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
11078 // Get the offsets to the 0 and 1 element of the array so that we can
11079 // select between them.
11080 SDValue Zero = DAG.getIntPtrConstant(0);
11081 unsigned EltSize = (unsigned)TD.getTypeAllocSize(Elts[0]->getType());
11082 SDValue One = DAG.getIntPtrConstant(EltSize);
11084 SDValue Cond = DAG.getSetCC(DL,
11085 getSetCCResultType(N0.getValueType()),
11087 AddToWorkList(Cond.getNode());
11088 SDValue CstOffset = DAG.getSelect(DL, Zero.getValueType(),
11090 AddToWorkList(CstOffset.getNode());
11091 CPIdx = DAG.getNode(ISD::ADD, DL, CPIdx.getValueType(), CPIdx,
11093 AddToWorkList(CPIdx.getNode());
11094 return DAG.getLoad(TV->getValueType(0), DL, DAG.getEntryNode(), CPIdx,
11095 MachinePointerInfo::getConstantPool(), false,
11096 false, false, Alignment);
11101 // Check to see if we can perform the "gzip trick", transforming
11102 // (select_cc setlt X, 0, A, 0) -> (and (sra X, (sub size(X), 1), A)
11103 if (N1C && N3C && N3C->isNullValue() && CC == ISD::SETLT &&
11104 (N1C->isNullValue() || // (a < 0) ? b : 0
11105 (N1C->getAPIntValue() == 1 && N0 == N2))) { // (a < 1) ? a : 0
11106 EVT XType = N0.getValueType();
11107 EVT AType = N2.getValueType();
11108 if (XType.bitsGE(AType)) {
11109 // and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a
11110 // single-bit constant.
11111 if (N2C && ((N2C->getAPIntValue() & (N2C->getAPIntValue()-1)) == 0)) {
11112 unsigned ShCtV = N2C->getAPIntValue().logBase2();
11113 ShCtV = XType.getSizeInBits()-ShCtV-1;
11114 SDValue ShCt = DAG.getConstant(ShCtV,
11115 getShiftAmountTy(N0.getValueType()));
11116 SDValue Shift = DAG.getNode(ISD::SRL, SDLoc(N0),
11118 AddToWorkList(Shift.getNode());
11120 if (XType.bitsGT(AType)) {
11121 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
11122 AddToWorkList(Shift.getNode());
11125 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
11128 SDValue Shift = DAG.getNode(ISD::SRA, SDLoc(N0),
11130 DAG.getConstant(XType.getSizeInBits()-1,
11131 getShiftAmountTy(N0.getValueType())));
11132 AddToWorkList(Shift.getNode());
11134 if (XType.bitsGT(AType)) {
11135 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
11136 AddToWorkList(Shift.getNode());
11139 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
11143 // fold (select_cc seteq (and x, y), 0, 0, A) -> (and (shr (shl x)) A)
11144 // where y is has a single bit set.
11145 // A plaintext description would be, we can turn the SELECT_CC into an AND
11146 // when the condition can be materialized as an all-ones register. Any
11147 // single bit-test can be materialized as an all-ones register with
11148 // shift-left and shift-right-arith.
11149 if (CC == ISD::SETEQ && N0->getOpcode() == ISD::AND &&
11150 N0->getValueType(0) == VT &&
11151 N1C && N1C->isNullValue() &&
11152 N2C && N2C->isNullValue()) {
11153 SDValue AndLHS = N0->getOperand(0);
11154 ConstantSDNode *ConstAndRHS = dyn_cast<ConstantSDNode>(N0->getOperand(1));
11155 if (ConstAndRHS && ConstAndRHS->getAPIntValue().countPopulation() == 1) {
11156 // Shift the tested bit over the sign bit.
11157 APInt AndMask = ConstAndRHS->getAPIntValue();
11159 DAG.getConstant(AndMask.countLeadingZeros(),
11160 getShiftAmountTy(AndLHS.getValueType()));
11161 SDValue Shl = DAG.getNode(ISD::SHL, SDLoc(N0), VT, AndLHS, ShlAmt);
11163 // Now arithmetic right shift it all the way over, so the result is either
11164 // all-ones, or zero.
11166 DAG.getConstant(AndMask.getBitWidth()-1,
11167 getShiftAmountTy(Shl.getValueType()));
11168 SDValue Shr = DAG.getNode(ISD::SRA, SDLoc(N0), VT, Shl, ShrAmt);
11170 return DAG.getNode(ISD::AND, DL, VT, Shr, N3);
11174 // fold select C, 16, 0 -> shl C, 4
11175 if (N2C && N3C && N3C->isNullValue() && N2C->getAPIntValue().isPowerOf2() &&
11176 TLI.getBooleanContents(N0.getValueType().isVector()) ==
11177 TargetLowering::ZeroOrOneBooleanContent) {
11179 // If the caller doesn't want us to simplify this into a zext of a compare,
11181 if (NotExtCompare && N2C->getAPIntValue() == 1)
11184 // Get a SetCC of the condition
11185 // NOTE: Don't create a SETCC if it's not legal on this target.
11186 if (!LegalOperations ||
11187 TLI.isOperationLegal(ISD::SETCC,
11188 LegalTypes ? getSetCCResultType(N0.getValueType()) : MVT::i1)) {
11190 // cast from setcc result type to select result type
11192 SCC = DAG.getSetCC(DL, getSetCCResultType(N0.getValueType()),
11194 if (N2.getValueType().bitsLT(SCC.getValueType()))
11195 Temp = DAG.getZeroExtendInReg(SCC, SDLoc(N2),
11196 N2.getValueType());
11198 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2),
11199 N2.getValueType(), SCC);
11201 SCC = DAG.getSetCC(SDLoc(N0), MVT::i1, N0, N1, CC);
11202 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2),
11203 N2.getValueType(), SCC);
11206 AddToWorkList(SCC.getNode());
11207 AddToWorkList(Temp.getNode());
11209 if (N2C->getAPIntValue() == 1)
11212 // shl setcc result by log2 n2c
11213 return DAG.getNode(
11214 ISD::SHL, DL, N2.getValueType(), Temp,
11215 DAG.getConstant(N2C->getAPIntValue().logBase2(),
11216 getShiftAmountTy(Temp.getValueType())));
11220 // Check to see if this is the equivalent of setcc
11221 // FIXME: Turn all of these into setcc if setcc if setcc is legal
11222 // otherwise, go ahead with the folds.
11223 if (0 && N3C && N3C->isNullValue() && N2C && (N2C->getAPIntValue() == 1ULL)) {
11224 EVT XType = N0.getValueType();
11225 if (!LegalOperations ||
11226 TLI.isOperationLegal(ISD::SETCC, getSetCCResultType(XType))) {
11227 SDValue Res = DAG.getSetCC(DL, getSetCCResultType(XType), N0, N1, CC);
11228 if (Res.getValueType() != VT)
11229 Res = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Res);
11233 // fold (seteq X, 0) -> (srl (ctlz X, log2(size(X))))
11234 if (N1C && N1C->isNullValue() && CC == ISD::SETEQ &&
11235 (!LegalOperations ||
11236 TLI.isOperationLegal(ISD::CTLZ, XType))) {
11237 SDValue Ctlz = DAG.getNode(ISD::CTLZ, SDLoc(N0), XType, N0);
11238 return DAG.getNode(ISD::SRL, DL, XType, Ctlz,
11239 DAG.getConstant(Log2_32(XType.getSizeInBits()),
11240 getShiftAmountTy(Ctlz.getValueType())));
11242 // fold (setgt X, 0) -> (srl (and (-X, ~X), size(X)-1))
11243 if (N1C && N1C->isNullValue() && CC == ISD::SETGT) {
11244 SDValue NegN0 = DAG.getNode(ISD::SUB, SDLoc(N0),
11245 XType, DAG.getConstant(0, XType), N0);
11246 SDValue NotN0 = DAG.getNOT(SDLoc(N0), N0, XType);
11247 return DAG.getNode(ISD::SRL, DL, XType,
11248 DAG.getNode(ISD::AND, DL, XType, NegN0, NotN0),
11249 DAG.getConstant(XType.getSizeInBits()-1,
11250 getShiftAmountTy(XType)));
11252 // fold (setgt X, -1) -> (xor (srl (X, size(X)-1), 1))
11253 if (N1C && N1C->isAllOnesValue() && CC == ISD::SETGT) {
11254 SDValue Sign = DAG.getNode(ISD::SRL, SDLoc(N0), XType, N0,
11255 DAG.getConstant(XType.getSizeInBits()-1,
11256 getShiftAmountTy(N0.getValueType())));
11257 return DAG.getNode(ISD::XOR, DL, XType, Sign, DAG.getConstant(1, XType));
11261 // Check to see if this is an integer abs.
11262 // select_cc setg[te] X, 0, X, -X ->
11263 // select_cc setgt X, -1, X, -X ->
11264 // select_cc setl[te] X, 0, -X, X ->
11265 // select_cc setlt X, 1, -X, X ->
11266 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
11268 ConstantSDNode *SubC = nullptr;
11269 if (((N1C->isNullValue() && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
11270 (N1C->isAllOnesValue() && CC == ISD::SETGT)) &&
11271 N0 == N2 && N3.getOpcode() == ISD::SUB && N0 == N3.getOperand(1))
11272 SubC = dyn_cast<ConstantSDNode>(N3.getOperand(0));
11273 else if (((N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE)) ||
11274 (N1C->isOne() && CC == ISD::SETLT)) &&
11275 N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1))
11276 SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0));
11278 EVT XType = N0.getValueType();
11279 if (SubC && SubC->isNullValue() && XType.isInteger()) {
11280 SDValue Shift = DAG.getNode(ISD::SRA, SDLoc(N0), XType,
11282 DAG.getConstant(XType.getSizeInBits()-1,
11283 getShiftAmountTy(N0.getValueType())));
11284 SDValue Add = DAG.getNode(ISD::ADD, SDLoc(N0),
11286 AddToWorkList(Shift.getNode());
11287 AddToWorkList(Add.getNode());
11288 return DAG.getNode(ISD::XOR, DL, XType, Add, Shift);
11295 /// SimplifySetCC - This is a stub for TargetLowering::SimplifySetCC.
11296 SDValue DAGCombiner::SimplifySetCC(EVT VT, SDValue N0,
11297 SDValue N1, ISD::CondCode Cond,
11298 SDLoc DL, bool foldBooleans) {
11299 TargetLowering::DAGCombinerInfo
11300 DagCombineInfo(DAG, Level, false, this);
11301 return TLI.SimplifySetCC(VT, N0, N1, Cond, foldBooleans, DagCombineInfo, DL);
11304 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
11305 /// return a DAG expression to select that will generate the same value by
11306 /// multiplying by a magic number. See:
11307 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
11308 SDValue DAGCombiner::BuildSDIV(SDNode *N) {
11309 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
11313 // Avoid division by zero.
11314 if (!C->getAPIntValue())
11317 std::vector<SDNode*> Built;
11319 TLI.BuildSDIV(N, C->getAPIntValue(), DAG, LegalOperations, &Built);
11321 for (SDNode *N : Built)
11326 /// BuildUDIV - Given an ISD::UDIV node expressing a divide by constant,
11327 /// return a DAG expression to select that will generate the same value by
11328 /// multiplying by a magic number. See:
11329 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
11330 SDValue DAGCombiner::BuildUDIV(SDNode *N) {
11331 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
11335 // Avoid division by zero.
11336 if (!C->getAPIntValue())
11339 std::vector<SDNode*> Built;
11341 TLI.BuildUDIV(N, C->getAPIntValue(), DAG, LegalOperations, &Built);
11343 for (SDNode *N : Built)
11348 /// FindBaseOffset - Return true if base is a frame index, which is known not
11349 // to alias with anything but itself. Provides base object and offset as
11351 static bool FindBaseOffset(SDValue Ptr, SDValue &Base, int64_t &Offset,
11352 const GlobalValue *&GV, const void *&CV) {
11353 // Assume it is a primitive operation.
11354 Base = Ptr; Offset = 0; GV = nullptr; CV = nullptr;
11356 // If it's an adding a simple constant then integrate the offset.
11357 if (Base.getOpcode() == ISD::ADD) {
11358 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Base.getOperand(1))) {
11359 Base = Base.getOperand(0);
11360 Offset += C->getZExtValue();
11364 // Return the underlying GlobalValue, and update the Offset. Return false
11365 // for GlobalAddressSDNode since the same GlobalAddress may be represented
11366 // by multiple nodes with different offsets.
11367 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Base)) {
11368 GV = G->getGlobal();
11369 Offset += G->getOffset();
11373 // Return the underlying Constant value, and update the Offset. Return false
11374 // for ConstantSDNodes since the same constant pool entry may be represented
11375 // by multiple nodes with different offsets.
11376 if (ConstantPoolSDNode *C = dyn_cast<ConstantPoolSDNode>(Base)) {
11377 CV = C->isMachineConstantPoolEntry() ? (const void *)C->getMachineCPVal()
11378 : (const void *)C->getConstVal();
11379 Offset += C->getOffset();
11382 // If it's any of the following then it can't alias with anything but itself.
11383 return isa<FrameIndexSDNode>(Base);
11386 /// isAlias - Return true if there is any possibility that the two addresses
11388 bool DAGCombiner::isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) const {
11389 // If they are the same then they must be aliases.
11390 if (Op0->getBasePtr() == Op1->getBasePtr()) return true;
11392 // If they are both volatile then they cannot be reordered.
11393 if (Op0->isVolatile() && Op1->isVolatile()) return true;
11395 // Gather base node and offset information.
11396 SDValue Base1, Base2;
11397 int64_t Offset1, Offset2;
11398 const GlobalValue *GV1, *GV2;
11399 const void *CV1, *CV2;
11400 bool isFrameIndex1 = FindBaseOffset(Op0->getBasePtr(),
11401 Base1, Offset1, GV1, CV1);
11402 bool isFrameIndex2 = FindBaseOffset(Op1->getBasePtr(),
11403 Base2, Offset2, GV2, CV2);
11405 // If they have a same base address then check to see if they overlap.
11406 if (Base1 == Base2 || (GV1 && (GV1 == GV2)) || (CV1 && (CV1 == CV2)))
11407 return !((Offset1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= Offset2 ||
11408 (Offset2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= Offset1);
11410 // It is possible for different frame indices to alias each other, mostly
11411 // when tail call optimization reuses return address slots for arguments.
11412 // To catch this case, look up the actual index of frame indices to compute
11413 // the real alias relationship.
11414 if (isFrameIndex1 && isFrameIndex2) {
11415 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
11416 Offset1 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base1)->getIndex());
11417 Offset2 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base2)->getIndex());
11418 return !((Offset1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= Offset2 ||
11419 (Offset2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= Offset1);
11422 // Otherwise, if we know what the bases are, and they aren't identical, then
11423 // we know they cannot alias.
11424 if ((isFrameIndex1 || CV1 || GV1) && (isFrameIndex2 || CV2 || GV2))
11427 // If we know required SrcValue1 and SrcValue2 have relatively large alignment
11428 // compared to the size and offset of the access, we may be able to prove they
11429 // do not alias. This check is conservative for now to catch cases created by
11430 // splitting vector types.
11431 if ((Op0->getOriginalAlignment() == Op1->getOriginalAlignment()) &&
11432 (Op0->getSrcValueOffset() != Op1->getSrcValueOffset()) &&
11433 (Op0->getMemoryVT().getSizeInBits() >> 3 ==
11434 Op1->getMemoryVT().getSizeInBits() >> 3) &&
11435 (Op0->getOriginalAlignment() > Op0->getMemoryVT().getSizeInBits()) >> 3) {
11436 int64_t OffAlign1 = Op0->getSrcValueOffset() % Op0->getOriginalAlignment();
11437 int64_t OffAlign2 = Op1->getSrcValueOffset() % Op1->getOriginalAlignment();
11439 // There is no overlap between these relatively aligned accesses of similar
11440 // size, return no alias.
11441 if ((OffAlign1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= OffAlign2 ||
11442 (OffAlign2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= OffAlign1)
11446 bool UseAA = CombinerGlobalAA.getNumOccurrences() > 0 ? CombinerGlobalAA :
11447 TLI.getTargetMachine().getSubtarget<TargetSubtargetInfo>().useAA();
11449 if (CombinerAAOnlyFunc.getNumOccurrences() &&
11450 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
11454 Op0->getMemOperand()->getValue() && Op1->getMemOperand()->getValue()) {
11455 // Use alias analysis information.
11456 int64_t MinOffset = std::min(Op0->getSrcValueOffset(),
11457 Op1->getSrcValueOffset());
11458 int64_t Overlap1 = (Op0->getMemoryVT().getSizeInBits() >> 3) +
11459 Op0->getSrcValueOffset() - MinOffset;
11460 int64_t Overlap2 = (Op1->getMemoryVT().getSizeInBits() >> 3) +
11461 Op1->getSrcValueOffset() - MinOffset;
11462 AliasAnalysis::AliasResult AAResult =
11463 AA.alias(AliasAnalysis::Location(Op0->getMemOperand()->getValue(),
11465 UseTBAA ? Op0->getTBAAInfo() : nullptr),
11466 AliasAnalysis::Location(Op1->getMemOperand()->getValue(),
11468 UseTBAA ? Op1->getTBAAInfo() : nullptr));
11469 if (AAResult == AliasAnalysis::NoAlias)
11473 // Otherwise we have to assume they alias.
11477 /// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes,
11478 /// looking for aliasing nodes and adding them to the Aliases vector.
11479 void DAGCombiner::GatherAllAliases(SDNode *N, SDValue OriginalChain,
11480 SmallVectorImpl<SDValue> &Aliases) {
11481 SmallVector<SDValue, 8> Chains; // List of chains to visit.
11482 SmallPtrSet<SDNode *, 16> Visited; // Visited node set.
11484 // Get alias information for node.
11485 bool IsLoad = isa<LoadSDNode>(N) && !cast<LSBaseSDNode>(N)->isVolatile();
11488 Chains.push_back(OriginalChain);
11489 unsigned Depth = 0;
11491 // Look at each chain and determine if it is an alias. If so, add it to the
11492 // aliases list. If not, then continue up the chain looking for the next
11494 while (!Chains.empty()) {
11495 SDValue Chain = Chains.back();
11498 // For TokenFactor nodes, look at each operand and only continue up the
11499 // chain until we find two aliases. If we've seen two aliases, assume we'll
11500 // find more and revert to original chain since the xform is unlikely to be
11503 // FIXME: The depth check could be made to return the last non-aliasing
11504 // chain we found before we hit a tokenfactor rather than the original
11506 if (Depth > 6 || Aliases.size() == 2) {
11508 Aliases.push_back(OriginalChain);
11512 // Don't bother if we've been before.
11513 if (!Visited.insert(Chain.getNode()))
11516 switch (Chain.getOpcode()) {
11517 case ISD::EntryToken:
11518 // Entry token is ideal chain operand, but handled in FindBetterChain.
11523 // Get alias information for Chain.
11524 bool IsOpLoad = isa<LoadSDNode>(Chain.getNode()) &&
11525 !cast<LSBaseSDNode>(Chain.getNode())->isVolatile();
11527 // If chain is alias then stop here.
11528 if (!(IsLoad && IsOpLoad) &&
11529 isAlias(cast<LSBaseSDNode>(N), cast<LSBaseSDNode>(Chain.getNode()))) {
11530 Aliases.push_back(Chain);
11532 // Look further up the chain.
11533 Chains.push_back(Chain.getOperand(0));
11539 case ISD::TokenFactor:
11540 // We have to check each of the operands of the token factor for "small"
11541 // token factors, so we queue them up. Adding the operands to the queue
11542 // (stack) in reverse order maintains the original order and increases the
11543 // likelihood that getNode will find a matching token factor (CSE.)
11544 if (Chain.getNumOperands() > 16) {
11545 Aliases.push_back(Chain);
11548 for (unsigned n = Chain.getNumOperands(); n;)
11549 Chains.push_back(Chain.getOperand(--n));
11554 // For all other instructions we will just have to take what we can get.
11555 Aliases.push_back(Chain);
11560 // We need to be careful here to also search for aliases through the
11561 // value operand of a store, etc. Consider the following situation:
11563 // L1 = load Token1, %52
11564 // S1 = store Token1, L1, %51
11565 // L2 = load Token1, %52+8
11566 // S2 = store Token1, L2, %51+8
11567 // Token2 = Token(S1, S2)
11568 // L3 = load Token2, %53
11569 // S3 = store Token2, L3, %52
11570 // L4 = load Token2, %53+8
11571 // S4 = store Token2, L4, %52+8
11572 // If we search for aliases of S3 (which loads address %52), and we look
11573 // only through the chain, then we'll miss the trivial dependence on L1
11574 // (which also loads from %52). We then might change all loads and
11575 // stores to use Token1 as their chain operand, which could result in
11576 // copying %53 into %52 before copying %52 into %51 (which should
11579 // The problem is, however, that searching for such data dependencies
11580 // can become expensive, and the cost is not directly related to the
11581 // chain depth. Instead, we'll rule out such configurations here by
11582 // insisting that we've visited all chain users (except for users
11583 // of the original chain, which is not necessary). When doing this,
11584 // we need to look through nodes we don't care about (otherwise, things
11585 // like register copies will interfere with trivial cases).
11587 SmallVector<const SDNode *, 16> Worklist;
11588 for (SmallPtrSet<SDNode *, 16>::iterator I = Visited.begin(),
11589 IE = Visited.end(); I != IE; ++I)
11590 if (*I != OriginalChain.getNode())
11591 Worklist.push_back(*I);
11593 while (!Worklist.empty()) {
11594 const SDNode *M = Worklist.pop_back_val();
11596 // We have already visited M, and want to make sure we've visited any uses
11597 // of M that we care about. For uses that we've not visisted, and don't
11598 // care about, queue them to the worklist.
11600 for (SDNode::use_iterator UI = M->use_begin(),
11601 UIE = M->use_end(); UI != UIE; ++UI)
11602 if (UI.getUse().getValueType() == MVT::Other && Visited.insert(*UI)) {
11603 if (isa<MemIntrinsicSDNode>(*UI) || isa<MemSDNode>(*UI)) {
11604 // We've not visited this use, and we care about it (it could have an
11605 // ordering dependency with the original node).
11607 Aliases.push_back(OriginalChain);
11611 // We've not visited this use, but we don't care about it. Mark it as
11612 // visited and enqueue it to the worklist.
11613 Worklist.push_back(*UI);
11618 /// FindBetterChain - Walk up chain skipping non-aliasing memory nodes, looking
11619 /// for a better chain (aliasing node.)
11620 SDValue DAGCombiner::FindBetterChain(SDNode *N, SDValue OldChain) {
11621 SmallVector<SDValue, 8> Aliases; // Ops for replacing token factor.
11623 // Accumulate all the aliases to this node.
11624 GatherAllAliases(N, OldChain, Aliases);
11626 // If no operands then chain to entry token.
11627 if (Aliases.size() == 0)
11628 return DAG.getEntryNode();
11630 // If a single operand then chain to it. We don't need to revisit it.
11631 if (Aliases.size() == 1)
11634 // Construct a custom tailored token factor.
11635 return DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, Aliases);
11638 // SelectionDAG::Combine - This is the entry point for the file.
11640 void SelectionDAG::Combine(CombineLevel Level, AliasAnalysis &AA,
11641 CodeGenOpt::Level OptLevel) {
11642 /// run - This is the main entry point to this class.
11644 DAGCombiner(*this, AA, OptLevel).Run(Level);