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/SetVector.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/Analysis/AliasAnalysis.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineFunction.h"
26 #include "llvm/IR/DataLayout.h"
27 #include "llvm/IR/DerivedTypes.h"
28 #include "llvm/IR/Function.h"
29 #include "llvm/IR/LLVMContext.h"
30 #include "llvm/Support/CommandLine.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/ErrorHandling.h"
33 #include "llvm/Support/MathExtras.h"
34 #include "llvm/Support/raw_ostream.h"
35 #include "llvm/Target/TargetLowering.h"
36 #include "llvm/Target/TargetMachine.h"
37 #include "llvm/Target/TargetOptions.h"
38 #include "llvm/Target/TargetRegisterInfo.h"
39 #include "llvm/Target/TargetSubtargetInfo.h"
43 #define DEBUG_TYPE "dagcombine"
45 STATISTIC(NodesCombined , "Number of dag nodes combined");
46 STATISTIC(PreIndexedNodes , "Number of pre-indexed nodes created");
47 STATISTIC(PostIndexedNodes, "Number of post-indexed nodes created");
48 STATISTIC(OpsNarrowed , "Number of load/op/store narrowed");
49 STATISTIC(LdStFP2Int , "Number of fp load/store pairs transformed to int");
50 STATISTIC(SlicedLoads, "Number of load sliced");
54 CombinerAA("combiner-alias-analysis", cl::Hidden,
55 cl::desc("Enable DAG combiner alias-analysis heuristics"));
58 CombinerGlobalAA("combiner-global-alias-analysis", cl::Hidden,
59 cl::desc("Enable DAG combiner's use of IR alias analysis"));
62 UseTBAA("combiner-use-tbaa", cl::Hidden, cl::init(true),
63 cl::desc("Enable DAG combiner's use of TBAA"));
66 static cl::opt<std::string>
67 CombinerAAOnlyFunc("combiner-aa-only-func", cl::Hidden,
68 cl::desc("Only use DAG-combiner alias analysis in this"
72 /// Hidden option to stress test load slicing, i.e., when this option
73 /// is enabled, load slicing bypasses most of its profitability guards.
75 StressLoadSlicing("combiner-stress-load-slicing", cl::Hidden,
76 cl::desc("Bypass the profitability model of load "
80 //------------------------------ DAGCombiner ---------------------------------//
84 const TargetLowering &TLI;
86 CodeGenOpt::Level OptLevel;
91 /// \brief Worklist of all of the nodes that need to be simplified.
93 /// This must behave as a stack -- new nodes to process are pushed onto the
94 /// back and when processing we pop off of the back.
96 /// The worklist will not contain duplicates but may contain null entries
97 /// due to nodes being deleted from the underlying DAG.
98 SmallVector<SDNode *, 64> Worklist;
100 /// \brief Mapping from an SDNode to its position on the worklist.
102 /// This is used to find and remove nodes from the worklist (by nulling
103 /// them) when they are deleted from the underlying DAG. It relies on
104 /// stable indices of nodes within the worklist.
105 DenseMap<SDNode *, unsigned> WorklistMap;
107 // AA - Used for DAG load/store alias analysis.
110 /// AddUsersToWorklist - When an instruction is simplified, add all users of
111 /// the instruction to the work lists because they might get more simplified
114 void AddUsersToWorklist(SDNode *N) {
115 for (SDNode *Node : N->uses())
119 /// visit - call the node-specific routine that knows how to fold each
120 /// particular type of node.
121 SDValue visit(SDNode *N);
124 /// AddToWorklist - Add to the work list making sure its instance is at the
125 /// back (next to be processed.)
126 void AddToWorklist(SDNode *N) {
127 // Skip handle nodes as they can't usefully be combined and confuse the
128 // zero-use deletion strategy.
129 if (N->getOpcode() == ISD::HANDLENODE)
132 if (WorklistMap.insert(std::make_pair(N, Worklist.size())).second)
133 Worklist.push_back(N);
136 /// removeFromWorklist - remove all instances of N from the worklist.
138 void removeFromWorklist(SDNode *N) {
139 auto It = WorklistMap.find(N);
140 if (It == WorklistMap.end())
141 return; // Not in the worklist.
143 // Null out the entry rather than erasing it to avoid a linear operation.
144 Worklist[It->second] = nullptr;
145 WorklistMap.erase(It);
148 bool recursivelyDeleteUnusedNodes(SDNode *N);
150 SDValue CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
153 SDValue CombineTo(SDNode *N, SDValue Res, bool AddTo = true) {
154 return CombineTo(N, &Res, 1, AddTo);
157 SDValue CombineTo(SDNode *N, SDValue Res0, SDValue Res1,
159 SDValue To[] = { Res0, Res1 };
160 return CombineTo(N, To, 2, AddTo);
163 void CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO);
167 /// SimplifyDemandedBits - Check the specified integer node value to see if
168 /// it can be simplified or if things it uses can be simplified by bit
169 /// propagation. If so, return true.
170 bool SimplifyDemandedBits(SDValue Op) {
171 unsigned BitWidth = Op.getValueType().getScalarType().getSizeInBits();
172 APInt Demanded = APInt::getAllOnesValue(BitWidth);
173 return SimplifyDemandedBits(Op, Demanded);
176 bool SimplifyDemandedBits(SDValue Op, const APInt &Demanded);
178 bool CombineToPreIndexedLoadStore(SDNode *N);
179 bool CombineToPostIndexedLoadStore(SDNode *N);
180 bool SliceUpLoad(SDNode *N);
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 BuildSDIVPow2(SDNode *N);
308 SDValue BuildUDIV(SDNode *N);
309 SDValue MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
310 bool DemandHighBits = true);
311 SDValue MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1);
312 SDNode *MatchRotatePosNeg(SDValue Shifted, SDValue Pos, SDValue Neg,
313 SDValue InnerPos, SDValue InnerNeg,
314 unsigned PosOpcode, unsigned NegOpcode,
316 SDNode *MatchRotate(SDValue LHS, SDValue RHS, SDLoc DL);
317 SDValue ReduceLoadWidth(SDNode *N);
318 SDValue ReduceLoadOpStoreWidth(SDNode *N);
319 SDValue TransformFPLoadStorePair(SDNode *N);
320 SDValue reduceBuildVecExtToExtBuildVec(SDNode *N);
321 SDValue reduceBuildVecConvertToConvertBuildVec(SDNode *N);
323 SDValue GetDemandedBits(SDValue V, const APInt &Mask);
325 /// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes,
326 /// looking for aliasing nodes and adding them to the Aliases vector.
327 void GatherAllAliases(SDNode *N, SDValue OriginalChain,
328 SmallVectorImpl<SDValue> &Aliases);
330 /// isAlias - Return true if there is any possibility that the two addresses
332 bool isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) const;
334 /// FindBetterChain - Walk up chain skipping non-aliasing memory nodes,
335 /// looking for a better chain (aliasing node.)
336 SDValue FindBetterChain(SDNode *N, SDValue Chain);
338 /// Merge consecutive store operations into a wide store.
339 /// This optimization uses wide integers or vectors when possible.
340 /// \return True if some memory operations were changed.
341 bool MergeConsecutiveStores(StoreSDNode *N);
343 /// \brief Try to transform a truncation where C is a constant:
344 /// (trunc (and X, C)) -> (and (trunc X), (trunc C))
346 /// \p N needs to be a truncation and its first operand an AND. Other
347 /// requirements are checked by the function (e.g. that trunc is
348 /// single-use) and if missed an empty SDValue is returned.
349 SDValue distributeTruncateThroughAnd(SDNode *N);
352 DAGCombiner(SelectionDAG &D, AliasAnalysis &A, CodeGenOpt::Level OL)
353 : DAG(D), TLI(D.getTargetLoweringInfo()), Level(BeforeLegalizeTypes),
354 OptLevel(OL), LegalOperations(false), LegalTypes(false), AA(A) {
355 AttributeSet FnAttrs =
356 DAG.getMachineFunction().getFunction()->getAttributes();
358 FnAttrs.hasAttribute(AttributeSet::FunctionIndex,
359 Attribute::OptimizeForSize) ||
360 FnAttrs.hasAttribute(AttributeSet::FunctionIndex, Attribute::MinSize);
363 /// Run - runs the dag combiner on all nodes in the work list
364 void Run(CombineLevel AtLevel);
366 SelectionDAG &getDAG() const { return DAG; }
368 /// getShiftAmountTy - Returns a type large enough to hold any valid
369 /// shift amount - before type legalization these can be huge.
370 EVT getShiftAmountTy(EVT LHSTy) {
371 assert(LHSTy.isInteger() && "Shift amount is not an integer type!");
372 if (LHSTy.isVector())
374 return LegalTypes ? TLI.getScalarShiftAmountTy(LHSTy)
375 : TLI.getPointerTy();
378 /// isTypeLegal - This method returns true if we are running before type
379 /// legalization or if the specified VT is legal.
380 bool isTypeLegal(const EVT &VT) {
381 if (!LegalTypes) return true;
382 return TLI.isTypeLegal(VT);
385 /// getSetCCResultType - Convenience wrapper around
386 /// TargetLowering::getSetCCResultType
387 EVT getSetCCResultType(EVT VT) const {
388 return TLI.getSetCCResultType(*DAG.getContext(), VT);
395 /// WorklistRemover - This class is a DAGUpdateListener that removes any deleted
396 /// nodes from the worklist.
397 class WorklistRemover : public SelectionDAG::DAGUpdateListener {
400 explicit WorklistRemover(DAGCombiner &dc)
401 : SelectionDAG::DAGUpdateListener(dc.getDAG()), DC(dc) {}
403 void NodeDeleted(SDNode *N, SDNode *E) override {
404 DC.removeFromWorklist(N);
409 //===----------------------------------------------------------------------===//
410 // TargetLowering::DAGCombinerInfo implementation
411 //===----------------------------------------------------------------------===//
413 void TargetLowering::DAGCombinerInfo::AddToWorklist(SDNode *N) {
414 ((DAGCombiner*)DC)->AddToWorklist(N);
417 void TargetLowering::DAGCombinerInfo::RemoveFromWorklist(SDNode *N) {
418 ((DAGCombiner*)DC)->removeFromWorklist(N);
421 SDValue TargetLowering::DAGCombinerInfo::
422 CombineTo(SDNode *N, const std::vector<SDValue> &To, bool AddTo) {
423 return ((DAGCombiner*)DC)->CombineTo(N, &To[0], To.size(), AddTo);
426 SDValue TargetLowering::DAGCombinerInfo::
427 CombineTo(SDNode *N, SDValue Res, bool AddTo) {
428 return ((DAGCombiner*)DC)->CombineTo(N, Res, AddTo);
432 SDValue TargetLowering::DAGCombinerInfo::
433 CombineTo(SDNode *N, SDValue Res0, SDValue Res1, bool AddTo) {
434 return ((DAGCombiner*)DC)->CombineTo(N, Res0, Res1, AddTo);
437 void TargetLowering::DAGCombinerInfo::
438 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
439 return ((DAGCombiner*)DC)->CommitTargetLoweringOpt(TLO);
442 //===----------------------------------------------------------------------===//
444 //===----------------------------------------------------------------------===//
446 /// isNegatibleForFree - Return 1 if we can compute the negated form of the
447 /// specified expression for the same cost as the expression itself, or 2 if we
448 /// can compute the negated form more cheaply than the expression itself.
449 static char isNegatibleForFree(SDValue Op, bool LegalOperations,
450 const TargetLowering &TLI,
451 const TargetOptions *Options,
452 unsigned Depth = 0) {
453 // fneg is removable even if it has multiple uses.
454 if (Op.getOpcode() == ISD::FNEG) return 2;
456 // Don't allow anything with multiple uses.
457 if (!Op.hasOneUse()) return 0;
459 // Don't recurse exponentially.
460 if (Depth > 6) return 0;
462 switch (Op.getOpcode()) {
463 default: return false;
464 case ISD::ConstantFP:
465 // Don't invert constant FP values after legalize. The negated constant
466 // isn't necessarily legal.
467 return LegalOperations ? 0 : 1;
469 // FIXME: determine better conditions for this xform.
470 if (!Options->UnsafeFPMath) return 0;
472 // After operation legalization, it might not be legal to create new FSUBs.
473 if (LegalOperations &&
474 !TLI.isOperationLegalOrCustom(ISD::FSUB, Op.getValueType()))
477 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B)
478 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI,
481 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
482 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options,
485 // We can't turn -(A-B) into B-A when we honor signed zeros.
486 if (!Options->UnsafeFPMath) return 0;
488 // fold (fneg (fsub A, B)) -> (fsub B, A)
493 if (Options->HonorSignDependentRoundingFPMath()) return 0;
495 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y) or (fmul X, (fneg Y))
496 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI,
500 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options,
506 return isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI, Options,
511 /// GetNegatedExpression - If isNegatibleForFree returns true, this function
512 /// returns the newly negated expression.
513 static SDValue GetNegatedExpression(SDValue Op, SelectionDAG &DAG,
514 bool LegalOperations, unsigned Depth = 0) {
515 // fneg is removable even if it has multiple uses.
516 if (Op.getOpcode() == ISD::FNEG) return Op.getOperand(0);
518 // Don't allow anything with multiple uses.
519 assert(Op.hasOneUse() && "Unknown reuse!");
521 assert(Depth <= 6 && "GetNegatedExpression doesn't match isNegatibleForFree");
522 switch (Op.getOpcode()) {
523 default: llvm_unreachable("Unknown code");
524 case ISD::ConstantFP: {
525 APFloat V = cast<ConstantFPSDNode>(Op)->getValueAPF();
527 return DAG.getConstantFP(V, Op.getValueType());
530 // FIXME: determine better conditions for this xform.
531 assert(DAG.getTarget().Options.UnsafeFPMath);
533 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B)
534 if (isNegatibleForFree(Op.getOperand(0), LegalOperations,
535 DAG.getTargetLoweringInfo(),
536 &DAG.getTarget().Options, Depth+1))
537 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
538 GetNegatedExpression(Op.getOperand(0), DAG,
539 LegalOperations, Depth+1),
541 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
542 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
543 GetNegatedExpression(Op.getOperand(1), DAG,
544 LegalOperations, Depth+1),
547 // We can't turn -(A-B) into B-A when we honor signed zeros.
548 assert(DAG.getTarget().Options.UnsafeFPMath);
550 // fold (fneg (fsub 0, B)) -> B
551 if (ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(Op.getOperand(0)))
552 if (N0CFP->getValueAPF().isZero())
553 return Op.getOperand(1);
555 // fold (fneg (fsub A, B)) -> (fsub B, A)
556 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
557 Op.getOperand(1), Op.getOperand(0));
561 assert(!DAG.getTarget().Options.HonorSignDependentRoundingFPMath());
563 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y)
564 if (isNegatibleForFree(Op.getOperand(0), LegalOperations,
565 DAG.getTargetLoweringInfo(),
566 &DAG.getTarget().Options, Depth+1))
567 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
568 GetNegatedExpression(Op.getOperand(0), DAG,
569 LegalOperations, Depth+1),
572 // fold (fneg (fmul X, Y)) -> (fmul X, (fneg Y))
573 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
575 GetNegatedExpression(Op.getOperand(1), DAG,
576 LegalOperations, Depth+1));
580 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
581 GetNegatedExpression(Op.getOperand(0), DAG,
582 LegalOperations, Depth+1));
584 return DAG.getNode(ISD::FP_ROUND, SDLoc(Op), Op.getValueType(),
585 GetNegatedExpression(Op.getOperand(0), DAG,
586 LegalOperations, Depth+1),
591 // isSetCCEquivalent - Return true if this node is a setcc, or is a select_cc
592 // that selects between the target values used for true and false, making it
593 // equivalent to a setcc. Also, set the incoming LHS, RHS, and CC references to
594 // the appropriate nodes based on the type of node we are checking. This
595 // simplifies life a bit for the callers.
596 bool DAGCombiner::isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS,
598 if (N.getOpcode() == ISD::SETCC) {
599 LHS = N.getOperand(0);
600 RHS = N.getOperand(1);
601 CC = N.getOperand(2);
605 if (N.getOpcode() != ISD::SELECT_CC ||
606 !TLI.isConstTrueVal(N.getOperand(2).getNode()) ||
607 !TLI.isConstFalseVal(N.getOperand(3).getNode()))
610 LHS = N.getOperand(0);
611 RHS = N.getOperand(1);
612 CC = N.getOperand(4);
616 // isOneUseSetCC - Return true if this is a SetCC-equivalent operation with only
617 // one use. If this is true, it allows the users to invert the operation for
618 // free when it is profitable to do so.
619 bool DAGCombiner::isOneUseSetCC(SDValue N) const {
621 if (isSetCCEquivalent(N, N0, N1, N2) && N.getNode()->hasOneUse())
626 /// isConstantSplatVector - Returns true if N is a BUILD_VECTOR node whose
627 /// elements are all the same constant or undefined.
628 static bool isConstantSplatVector(SDNode *N, APInt& SplatValue) {
629 BuildVectorSDNode *C = dyn_cast<BuildVectorSDNode>(N);
634 unsigned SplatBitSize;
636 EVT EltVT = N->getValueType(0).getVectorElementType();
637 return (C->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
639 EltVT.getSizeInBits() >= SplatBitSize);
642 // \brief Returns the SDNode if it is a constant BuildVector or constant.
643 static SDNode *isConstantBuildVectorOrConstantInt(SDValue N) {
644 if (isa<ConstantSDNode>(N))
646 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N);
647 if(BV && BV->isConstant())
652 // \brief Returns the SDNode if it is a constant splat BuildVector or constant
654 static ConstantSDNode *isConstOrConstSplat(SDValue N) {
655 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N))
658 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N)) {
659 BitVector UndefElements;
660 ConstantSDNode *CN = BV->getConstantSplatNode(&UndefElements);
662 // BuildVectors can truncate their operands. Ignore that case here.
663 // FIXME: We blindly ignore splats which include undef which is overly
665 if (CN && UndefElements.none() &&
666 CN->getValueType(0) == N.getValueType().getScalarType())
673 SDValue DAGCombiner::ReassociateOps(unsigned Opc, SDLoc DL,
674 SDValue N0, SDValue N1) {
675 EVT VT = N0.getValueType();
676 if (N0.getOpcode() == Opc) {
677 if (SDNode *L = isConstantBuildVectorOrConstantInt(N0.getOperand(1))) {
678 if (SDNode *R = isConstantBuildVectorOrConstantInt(N1)) {
679 // reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2))
680 SDValue OpNode = DAG.FoldConstantArithmetic(Opc, VT, L, R);
681 if (!OpNode.getNode())
683 return DAG.getNode(Opc, DL, VT, N0.getOperand(0), OpNode);
685 if (N0.hasOneUse()) {
686 // reassoc. (op (op x, c1), y) -> (op (op x, y), c1) iff x+c1 has one
688 SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N0.getOperand(0), N1);
689 if (!OpNode.getNode())
691 AddToWorklist(OpNode.getNode());
692 return DAG.getNode(Opc, DL, VT, OpNode, N0.getOperand(1));
697 if (N1.getOpcode() == Opc) {
698 if (SDNode *R = isConstantBuildVectorOrConstantInt(N1.getOperand(1))) {
699 if (SDNode *L = isConstantBuildVectorOrConstantInt(N0)) {
700 // reassoc. (op c2, (op x, c1)) -> (op x, (op c1, c2))
701 SDValue OpNode = DAG.FoldConstantArithmetic(Opc, VT, R, L);
702 if (!OpNode.getNode())
704 return DAG.getNode(Opc, DL, VT, N1.getOperand(0), OpNode);
706 if (N1.hasOneUse()) {
707 // reassoc. (op y, (op x, c1)) -> (op (op x, y), c1) iff x+c1 has one
709 SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N1.getOperand(0), N0);
710 if (!OpNode.getNode())
712 AddToWorklist(OpNode.getNode());
713 return DAG.getNode(Opc, DL, VT, OpNode, N1.getOperand(1));
721 SDValue DAGCombiner::CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
723 assert(N->getNumValues() == NumTo && "Broken CombineTo call!");
725 DEBUG(dbgs() << "\nReplacing.1 ";
727 dbgs() << "\nWith: ";
728 To[0].getNode()->dump(&DAG);
729 dbgs() << " and " << NumTo-1 << " other values\n";
730 for (unsigned i = 0, e = NumTo; i != e; ++i)
731 assert((!To[i].getNode() ||
732 N->getValueType(i) == To[i].getValueType()) &&
733 "Cannot combine value to value of different type!"));
734 WorklistRemover DeadNodes(*this);
735 DAG.ReplaceAllUsesWith(N, To);
737 // Push the new nodes and any users onto the worklist
738 for (unsigned i = 0, e = NumTo; i != e; ++i) {
739 if (To[i].getNode()) {
740 AddToWorklist(To[i].getNode());
741 AddUsersToWorklist(To[i].getNode());
746 // Finally, if the node is now dead, remove it from the graph. The node
747 // may not be dead if the replacement process recursively simplified to
748 // something else needing this node.
749 if (N->use_empty()) {
750 // Nodes can be reintroduced into the worklist. Make sure we do not
751 // process a node that has been replaced.
752 removeFromWorklist(N);
754 // Finally, since the node is now dead, remove it from the graph.
757 return SDValue(N, 0);
761 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
762 // Replace all uses. If any nodes become isomorphic to other nodes and
763 // are deleted, make sure to remove them from our worklist.
764 WorklistRemover DeadNodes(*this);
765 DAG.ReplaceAllUsesOfValueWith(TLO.Old, TLO.New);
767 // Push the new node and any (possibly new) users onto the worklist.
768 AddToWorklist(TLO.New.getNode());
769 AddUsersToWorklist(TLO.New.getNode());
771 // Finally, if the node is now dead, remove it from the graph. The node
772 // may not be dead if the replacement process recursively simplified to
773 // something else needing this node.
774 if (TLO.Old.getNode()->use_empty()) {
775 removeFromWorklist(TLO.Old.getNode());
777 // If the operands of this node are only used by the node, they will now
778 // be dead. Make sure to visit them first to delete dead nodes early.
779 for (unsigned i = 0, e = TLO.Old.getNode()->getNumOperands(); i != e; ++i)
780 if (TLO.Old.getNode()->getOperand(i).getNode()->hasOneUse())
781 AddToWorklist(TLO.Old.getNode()->getOperand(i).getNode());
783 DAG.DeleteNode(TLO.Old.getNode());
787 /// SimplifyDemandedBits - Check the specified integer node value to see if
788 /// it can be simplified or if things it uses can be simplified by bit
789 /// propagation. If so, return true.
790 bool DAGCombiner::SimplifyDemandedBits(SDValue Op, const APInt &Demanded) {
791 TargetLowering::TargetLoweringOpt TLO(DAG, LegalTypes, LegalOperations);
792 APInt KnownZero, KnownOne;
793 if (!TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
797 AddToWorklist(Op.getNode());
799 // Replace the old value with the new one.
801 DEBUG(dbgs() << "\nReplacing.2 ";
802 TLO.Old.getNode()->dump(&DAG);
803 dbgs() << "\nWith: ";
804 TLO.New.getNode()->dump(&DAG);
807 CommitTargetLoweringOpt(TLO);
811 void DAGCombiner::ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad) {
813 EVT VT = Load->getValueType(0);
814 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, VT, SDValue(ExtLoad, 0));
816 DEBUG(dbgs() << "\nReplacing.9 ";
818 dbgs() << "\nWith: ";
819 Trunc.getNode()->dump(&DAG);
821 WorklistRemover DeadNodes(*this);
822 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 0), Trunc);
823 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), SDValue(ExtLoad, 1));
824 removeFromWorklist(Load);
825 DAG.DeleteNode(Load);
826 AddToWorklist(Trunc.getNode());
829 SDValue DAGCombiner::PromoteOperand(SDValue Op, EVT PVT, bool &Replace) {
832 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Op)) {
833 EVT MemVT = LD->getMemoryVT();
834 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
835 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT) ? ISD::ZEXTLOAD
837 : LD->getExtensionType();
839 return DAG.getExtLoad(ExtType, dl, PVT,
840 LD->getChain(), LD->getBasePtr(),
841 MemVT, LD->getMemOperand());
844 unsigned Opc = Op.getOpcode();
847 case ISD::AssertSext:
848 return DAG.getNode(ISD::AssertSext, dl, PVT,
849 SExtPromoteOperand(Op.getOperand(0), PVT),
851 case ISD::AssertZext:
852 return DAG.getNode(ISD::AssertZext, dl, PVT,
853 ZExtPromoteOperand(Op.getOperand(0), PVT),
855 case ISD::Constant: {
857 Op.getValueType().isByteSized() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
858 return DAG.getNode(ExtOpc, dl, PVT, Op);
862 if (!TLI.isOperationLegal(ISD::ANY_EXTEND, PVT))
864 return DAG.getNode(ISD::ANY_EXTEND, dl, PVT, Op);
867 SDValue DAGCombiner::SExtPromoteOperand(SDValue Op, EVT PVT) {
868 if (!TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, PVT))
870 EVT OldVT = Op.getValueType();
872 bool Replace = false;
873 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
874 if (!NewOp.getNode())
876 AddToWorklist(NewOp.getNode());
879 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
880 return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, NewOp.getValueType(), NewOp,
881 DAG.getValueType(OldVT));
884 SDValue DAGCombiner::ZExtPromoteOperand(SDValue Op, EVT PVT) {
885 EVT OldVT = Op.getValueType();
887 bool Replace = false;
888 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
889 if (!NewOp.getNode())
891 AddToWorklist(NewOp.getNode());
894 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
895 return DAG.getZeroExtendInReg(NewOp, dl, OldVT);
898 /// PromoteIntBinOp - Promote the specified integer binary operation if the
899 /// target indicates it is beneficial. e.g. On x86, it's usually better to
900 /// promote i16 operations to i32 since i16 instructions are longer.
901 SDValue DAGCombiner::PromoteIntBinOp(SDValue Op) {
902 if (!LegalOperations)
905 EVT VT = Op.getValueType();
906 if (VT.isVector() || !VT.isInteger())
909 // If operation type is 'undesirable', e.g. i16 on x86, consider
911 unsigned Opc = Op.getOpcode();
912 if (TLI.isTypeDesirableForOp(Opc, VT))
916 // Consult target whether it is a good idea to promote this operation and
917 // what's the right type to promote it to.
918 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
919 assert(PVT != VT && "Don't know what type to promote to!");
921 bool Replace0 = false;
922 SDValue N0 = Op.getOperand(0);
923 SDValue NN0 = PromoteOperand(N0, PVT, Replace0);
927 bool Replace1 = false;
928 SDValue N1 = Op.getOperand(1);
933 NN1 = PromoteOperand(N1, PVT, Replace1);
938 AddToWorklist(NN0.getNode());
940 AddToWorklist(NN1.getNode());
943 ReplaceLoadWithPromotedLoad(N0.getNode(), NN0.getNode());
945 ReplaceLoadWithPromotedLoad(N1.getNode(), NN1.getNode());
947 DEBUG(dbgs() << "\nPromoting ";
948 Op.getNode()->dump(&DAG));
950 return DAG.getNode(ISD::TRUNCATE, dl, VT,
951 DAG.getNode(Opc, dl, PVT, NN0, NN1));
956 /// PromoteIntShiftOp - Promote the specified integer shift operation if the
957 /// target indicates it is beneficial. e.g. On x86, it's usually better to
958 /// promote i16 operations to i32 since i16 instructions are longer.
959 SDValue DAGCombiner::PromoteIntShiftOp(SDValue Op) {
960 if (!LegalOperations)
963 EVT VT = Op.getValueType();
964 if (VT.isVector() || !VT.isInteger())
967 // If operation type is 'undesirable', e.g. i16 on x86, consider
969 unsigned Opc = Op.getOpcode();
970 if (TLI.isTypeDesirableForOp(Opc, VT))
974 // Consult target whether it is a good idea to promote this operation and
975 // what's the right type to promote it to.
976 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
977 assert(PVT != VT && "Don't know what type to promote to!");
979 bool Replace = false;
980 SDValue N0 = Op.getOperand(0);
982 N0 = SExtPromoteOperand(Op.getOperand(0), PVT);
983 else if (Opc == ISD::SRL)
984 N0 = ZExtPromoteOperand(Op.getOperand(0), PVT);
986 N0 = PromoteOperand(N0, PVT, Replace);
990 AddToWorklist(N0.getNode());
992 ReplaceLoadWithPromotedLoad(Op.getOperand(0).getNode(), N0.getNode());
994 DEBUG(dbgs() << "\nPromoting ";
995 Op.getNode()->dump(&DAG));
997 return DAG.getNode(ISD::TRUNCATE, dl, VT,
998 DAG.getNode(Opc, dl, PVT, N0, Op.getOperand(1)));
1003 SDValue DAGCombiner::PromoteExtend(SDValue Op) {
1004 if (!LegalOperations)
1007 EVT VT = Op.getValueType();
1008 if (VT.isVector() || !VT.isInteger())
1011 // If operation type is 'undesirable', e.g. i16 on x86, consider
1013 unsigned Opc = Op.getOpcode();
1014 if (TLI.isTypeDesirableForOp(Opc, VT))
1018 // Consult target whether it is a good idea to promote this operation and
1019 // what's the right type to promote it to.
1020 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1021 assert(PVT != VT && "Don't know what type to promote to!");
1022 // fold (aext (aext x)) -> (aext x)
1023 // fold (aext (zext x)) -> (zext x)
1024 // fold (aext (sext x)) -> (sext x)
1025 DEBUG(dbgs() << "\nPromoting ";
1026 Op.getNode()->dump(&DAG));
1027 return DAG.getNode(Op.getOpcode(), SDLoc(Op), VT, Op.getOperand(0));
1032 bool DAGCombiner::PromoteLoad(SDValue Op) {
1033 if (!LegalOperations)
1036 EVT VT = Op.getValueType();
1037 if (VT.isVector() || !VT.isInteger())
1040 // If operation type is 'undesirable', e.g. i16 on x86, consider
1042 unsigned Opc = Op.getOpcode();
1043 if (TLI.isTypeDesirableForOp(Opc, VT))
1047 // Consult target whether it is a good idea to promote this operation and
1048 // what's the right type to promote it to.
1049 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1050 assert(PVT != VT && "Don't know what type to promote to!");
1053 SDNode *N = Op.getNode();
1054 LoadSDNode *LD = cast<LoadSDNode>(N);
1055 EVT MemVT = LD->getMemoryVT();
1056 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
1057 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT) ? ISD::ZEXTLOAD
1059 : LD->getExtensionType();
1060 SDValue NewLD = DAG.getExtLoad(ExtType, dl, PVT,
1061 LD->getChain(), LD->getBasePtr(),
1062 MemVT, LD->getMemOperand());
1063 SDValue Result = DAG.getNode(ISD::TRUNCATE, dl, VT, NewLD);
1065 DEBUG(dbgs() << "\nPromoting ";
1068 Result.getNode()->dump(&DAG);
1070 WorklistRemover DeadNodes(*this);
1071 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result);
1072 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), NewLD.getValue(1));
1073 removeFromWorklist(N);
1075 AddToWorklist(Result.getNode());
1081 /// \brief Recursively delete a node which has no uses and any operands for
1082 /// which it is the only use.
1084 /// Note that this both deletes the nodes and removes them from the worklist.
1085 /// It also adds any nodes who have had a user deleted to the worklist as they
1086 /// may now have only one use and subject to other combines.
1087 bool DAGCombiner::recursivelyDeleteUnusedNodes(SDNode *N) {
1088 if (!N->use_empty())
1091 SmallSetVector<SDNode *, 16> Nodes;
1094 N = Nodes.pop_back_val();
1098 if (N->use_empty()) {
1099 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1100 Nodes.insert(N->getOperand(i).getNode());
1102 removeFromWorklist(N);
1107 } while (!Nodes.empty());
1111 //===----------------------------------------------------------------------===//
1112 // Main DAG Combiner implementation
1113 //===----------------------------------------------------------------------===//
1115 void DAGCombiner::Run(CombineLevel AtLevel) {
1116 // set the instance variables, so that the various visit routines may use it.
1118 LegalOperations = Level >= AfterLegalizeVectorOps;
1119 LegalTypes = Level >= AfterLegalizeTypes;
1121 // Add all the dag nodes to the worklist.
1122 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
1123 E = DAG.allnodes_end(); I != E; ++I)
1126 // Create a dummy node (which is not added to allnodes), that adds a reference
1127 // to the root node, preventing it from being deleted, and tracking any
1128 // changes of the root.
1129 HandleSDNode Dummy(DAG.getRoot());
1131 // The root of the dag may dangle to deleted nodes until the dag combiner is
1132 // done. Set it to null to avoid confusion.
1133 DAG.setRoot(SDValue());
1135 // while the worklist isn't empty, find a node and
1136 // try and combine it.
1137 while (!WorklistMap.empty()) {
1139 // The Worklist holds the SDNodes in order, but it may contain null entries.
1141 N = Worklist.pop_back_val();
1144 bool GoodWorklistEntry = WorklistMap.erase(N);
1145 (void)GoodWorklistEntry;
1146 assert(GoodWorklistEntry &&
1147 "Found a worklist entry without a corresponding map entry!");
1149 // If N has no uses, it is dead. Make sure to revisit all N's operands once
1150 // N is deleted from the DAG, since they too may now be dead or may have a
1151 // reduced number of uses, allowing other xforms.
1152 if (recursivelyDeleteUnusedNodes(N))
1155 WorklistRemover DeadNodes(*this);
1157 SDValue RV = combine(N);
1164 // If we get back the same node we passed in, rather than a new node or
1165 // zero, we know that the node must have defined multiple values and
1166 // CombineTo was used. Since CombineTo takes care of the worklist
1167 // mechanics for us, we have no work to do in this case.
1168 if (RV.getNode() == N)
1171 assert(N->getOpcode() != ISD::DELETED_NODE &&
1172 RV.getNode()->getOpcode() != ISD::DELETED_NODE &&
1173 "Node was deleted but visit returned new node!");
1175 DEBUG(dbgs() << "\nReplacing.3 ";
1177 dbgs() << "\nWith: ";
1178 RV.getNode()->dump(&DAG);
1181 // Transfer debug value.
1182 DAG.TransferDbgValues(SDValue(N, 0), RV);
1183 if (N->getNumValues() == RV.getNode()->getNumValues())
1184 DAG.ReplaceAllUsesWith(N, RV.getNode());
1186 assert(N->getValueType(0) == RV.getValueType() &&
1187 N->getNumValues() == 1 && "Type mismatch");
1189 DAG.ReplaceAllUsesWith(N, &OpV);
1192 // Push the new node and any users onto the worklist
1193 AddToWorklist(RV.getNode());
1194 AddUsersToWorklist(RV.getNode());
1196 // Finally, if the node is now dead, remove it from the graph. The node
1197 // may not be dead if the replacement process recursively simplified to
1198 // something else needing this node. This will also take care of adding any
1199 // operands which have lost a user to the worklist.
1200 recursivelyDeleteUnusedNodes(N);
1203 // If the root changed (e.g. it was a dead load, update the root).
1204 DAG.setRoot(Dummy.getValue());
1205 DAG.RemoveDeadNodes();
1208 SDValue DAGCombiner::visit(SDNode *N) {
1209 switch (N->getOpcode()) {
1211 case ISD::TokenFactor: return visitTokenFactor(N);
1212 case ISD::MERGE_VALUES: return visitMERGE_VALUES(N);
1213 case ISD::ADD: return visitADD(N);
1214 case ISD::SUB: return visitSUB(N);
1215 case ISD::ADDC: return visitADDC(N);
1216 case ISD::SUBC: return visitSUBC(N);
1217 case ISD::ADDE: return visitADDE(N);
1218 case ISD::SUBE: return visitSUBE(N);
1219 case ISD::MUL: return visitMUL(N);
1220 case ISD::SDIV: return visitSDIV(N);
1221 case ISD::UDIV: return visitUDIV(N);
1222 case ISD::SREM: return visitSREM(N);
1223 case ISD::UREM: return visitUREM(N);
1224 case ISD::MULHU: return visitMULHU(N);
1225 case ISD::MULHS: return visitMULHS(N);
1226 case ISD::SMUL_LOHI: return visitSMUL_LOHI(N);
1227 case ISD::UMUL_LOHI: return visitUMUL_LOHI(N);
1228 case ISD::SMULO: return visitSMULO(N);
1229 case ISD::UMULO: return visitUMULO(N);
1230 case ISD::SDIVREM: return visitSDIVREM(N);
1231 case ISD::UDIVREM: return visitUDIVREM(N);
1232 case ISD::AND: return visitAND(N);
1233 case ISD::OR: return visitOR(N);
1234 case ISD::XOR: return visitXOR(N);
1235 case ISD::SHL: return visitSHL(N);
1236 case ISD::SRA: return visitSRA(N);
1237 case ISD::SRL: return visitSRL(N);
1239 case ISD::ROTL: return visitRotate(N);
1240 case ISD::CTLZ: return visitCTLZ(N);
1241 case ISD::CTLZ_ZERO_UNDEF: return visitCTLZ_ZERO_UNDEF(N);
1242 case ISD::CTTZ: return visitCTTZ(N);
1243 case ISD::CTTZ_ZERO_UNDEF: return visitCTTZ_ZERO_UNDEF(N);
1244 case ISD::CTPOP: return visitCTPOP(N);
1245 case ISD::SELECT: return visitSELECT(N);
1246 case ISD::VSELECT: return visitVSELECT(N);
1247 case ISD::SELECT_CC: return visitSELECT_CC(N);
1248 case ISD::SETCC: return visitSETCC(N);
1249 case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N);
1250 case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N);
1251 case ISD::ANY_EXTEND: return visitANY_EXTEND(N);
1252 case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N);
1253 case ISD::TRUNCATE: return visitTRUNCATE(N);
1254 case ISD::BITCAST: return visitBITCAST(N);
1255 case ISD::BUILD_PAIR: return visitBUILD_PAIR(N);
1256 case ISD::FADD: return visitFADD(N);
1257 case ISD::FSUB: return visitFSUB(N);
1258 case ISD::FMUL: return visitFMUL(N);
1259 case ISD::FMA: return visitFMA(N);
1260 case ISD::FDIV: return visitFDIV(N);
1261 case ISD::FREM: return visitFREM(N);
1262 case ISD::FCOPYSIGN: return visitFCOPYSIGN(N);
1263 case ISD::SINT_TO_FP: return visitSINT_TO_FP(N);
1264 case ISD::UINT_TO_FP: return visitUINT_TO_FP(N);
1265 case ISD::FP_TO_SINT: return visitFP_TO_SINT(N);
1266 case ISD::FP_TO_UINT: return visitFP_TO_UINT(N);
1267 case ISD::FP_ROUND: return visitFP_ROUND(N);
1268 case ISD::FP_ROUND_INREG: return visitFP_ROUND_INREG(N);
1269 case ISD::FP_EXTEND: return visitFP_EXTEND(N);
1270 case ISD::FNEG: return visitFNEG(N);
1271 case ISD::FABS: return visitFABS(N);
1272 case ISD::FFLOOR: return visitFFLOOR(N);
1273 case ISD::FCEIL: return visitFCEIL(N);
1274 case ISD::FTRUNC: return visitFTRUNC(N);
1275 case ISD::BRCOND: return visitBRCOND(N);
1276 case ISD::BR_CC: return visitBR_CC(N);
1277 case ISD::LOAD: return visitLOAD(N);
1278 case ISD::STORE: return visitSTORE(N);
1279 case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N);
1280 case ISD::EXTRACT_VECTOR_ELT: return visitEXTRACT_VECTOR_ELT(N);
1281 case ISD::BUILD_VECTOR: return visitBUILD_VECTOR(N);
1282 case ISD::CONCAT_VECTORS: return visitCONCAT_VECTORS(N);
1283 case ISD::EXTRACT_SUBVECTOR: return visitEXTRACT_SUBVECTOR(N);
1284 case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N);
1285 case ISD::INSERT_SUBVECTOR: return visitINSERT_SUBVECTOR(N);
1290 SDValue DAGCombiner::combine(SDNode *N) {
1291 SDValue RV = visit(N);
1293 // If nothing happened, try a target-specific DAG combine.
1294 if (!RV.getNode()) {
1295 assert(N->getOpcode() != ISD::DELETED_NODE &&
1296 "Node was deleted but visit returned NULL!");
1298 if (N->getOpcode() >= ISD::BUILTIN_OP_END ||
1299 TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode())) {
1301 // Expose the DAG combiner to the target combiner impls.
1302 TargetLowering::DAGCombinerInfo
1303 DagCombineInfo(DAG, Level, false, this);
1305 RV = TLI.PerformDAGCombine(N, DagCombineInfo);
1309 // If nothing happened still, try promoting the operation.
1310 if (!RV.getNode()) {
1311 switch (N->getOpcode()) {
1319 RV = PromoteIntBinOp(SDValue(N, 0));
1324 RV = PromoteIntShiftOp(SDValue(N, 0));
1326 case ISD::SIGN_EXTEND:
1327 case ISD::ZERO_EXTEND:
1328 case ISD::ANY_EXTEND:
1329 RV = PromoteExtend(SDValue(N, 0));
1332 if (PromoteLoad(SDValue(N, 0)))
1338 // If N is a commutative binary node, try commuting it to enable more
1340 if (!RV.getNode() && SelectionDAG::isCommutativeBinOp(N->getOpcode()) &&
1341 N->getNumValues() == 1) {
1342 SDValue N0 = N->getOperand(0);
1343 SDValue N1 = N->getOperand(1);
1345 // Constant operands are canonicalized to RHS.
1346 if (isa<ConstantSDNode>(N0) || !isa<ConstantSDNode>(N1)) {
1347 SDValue Ops[] = {N1, N0};
1349 if (const BinaryWithFlagsSDNode *BinNode =
1350 dyn_cast<BinaryWithFlagsSDNode>(N)) {
1351 CSENode = DAG.getNodeIfExists(
1352 N->getOpcode(), N->getVTList(), Ops, BinNode->hasNoUnsignedWrap(),
1353 BinNode->hasNoSignedWrap(), BinNode->isExact());
1355 CSENode = DAG.getNodeIfExists(N->getOpcode(), N->getVTList(), Ops);
1358 return SDValue(CSENode, 0);
1365 /// getInputChainForNode - Given a node, return its input chain if it has one,
1366 /// otherwise return a null sd operand.
1367 static SDValue getInputChainForNode(SDNode *N) {
1368 if (unsigned NumOps = N->getNumOperands()) {
1369 if (N->getOperand(0).getValueType() == MVT::Other)
1370 return N->getOperand(0);
1371 if (N->getOperand(NumOps-1).getValueType() == MVT::Other)
1372 return N->getOperand(NumOps-1);
1373 for (unsigned i = 1; i < NumOps-1; ++i)
1374 if (N->getOperand(i).getValueType() == MVT::Other)
1375 return N->getOperand(i);
1380 SDValue DAGCombiner::visitTokenFactor(SDNode *N) {
1381 // If N has two operands, where one has an input chain equal to the other,
1382 // the 'other' chain is redundant.
1383 if (N->getNumOperands() == 2) {
1384 if (getInputChainForNode(N->getOperand(0).getNode()) == N->getOperand(1))
1385 return N->getOperand(0);
1386 if (getInputChainForNode(N->getOperand(1).getNode()) == N->getOperand(0))
1387 return N->getOperand(1);
1390 SmallVector<SDNode *, 8> TFs; // List of token factors to visit.
1391 SmallVector<SDValue, 8> Ops; // Ops for replacing token factor.
1392 SmallPtrSet<SDNode*, 16> SeenOps;
1393 bool Changed = false; // If we should replace this token factor.
1395 // Start out with this token factor.
1398 // Iterate through token factors. The TFs grows when new token factors are
1400 for (unsigned i = 0; i < TFs.size(); ++i) {
1401 SDNode *TF = TFs[i];
1403 // Check each of the operands.
1404 for (unsigned i = 0, ie = TF->getNumOperands(); i != ie; ++i) {
1405 SDValue Op = TF->getOperand(i);
1407 switch (Op.getOpcode()) {
1408 case ISD::EntryToken:
1409 // Entry tokens don't need to be added to the list. They are
1414 case ISD::TokenFactor:
1415 if (Op.hasOneUse() &&
1416 std::find(TFs.begin(), TFs.end(), Op.getNode()) == TFs.end()) {
1417 // Queue up for processing.
1418 TFs.push_back(Op.getNode());
1419 // Clean up in case the token factor is removed.
1420 AddToWorklist(Op.getNode());
1427 // Only add if it isn't already in the list.
1428 if (SeenOps.insert(Op.getNode()))
1439 // If we've change things around then replace token factor.
1442 // The entry token is the only possible outcome.
1443 Result = DAG.getEntryNode();
1445 // New and improved token factor.
1446 Result = DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, Ops);
1449 // Don't add users to work list.
1450 return CombineTo(N, Result, false);
1456 /// MERGE_VALUES can always be eliminated.
1457 SDValue DAGCombiner::visitMERGE_VALUES(SDNode *N) {
1458 WorklistRemover DeadNodes(*this);
1459 // Replacing results may cause a different MERGE_VALUES to suddenly
1460 // be CSE'd with N, and carry its uses with it. Iterate until no
1461 // uses remain, to ensure that the node can be safely deleted.
1462 // First add the users of this node to the work list so that they
1463 // can be tried again once they have new operands.
1464 AddUsersToWorklist(N);
1466 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1467 DAG.ReplaceAllUsesOfValueWith(SDValue(N, i), N->getOperand(i));
1468 } while (!N->use_empty());
1469 removeFromWorklist(N);
1471 return SDValue(N, 0); // Return N so it doesn't get rechecked!
1475 SDValue combineShlAddConstant(SDLoc DL, SDValue N0, SDValue N1,
1476 SelectionDAG &DAG) {
1477 EVT VT = N0.getValueType();
1478 SDValue N00 = N0.getOperand(0);
1479 SDValue N01 = N0.getOperand(1);
1480 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N01);
1482 if (N01C && N00.getOpcode() == ISD::ADD && N00.getNode()->hasOneUse() &&
1483 isa<ConstantSDNode>(N00.getOperand(1))) {
1484 // fold (add (shl (add x, c1), c2), ) -> (add (add (shl x, c2), c1<<c2), )
1485 N0 = DAG.getNode(ISD::ADD, SDLoc(N0), VT,
1486 DAG.getNode(ISD::SHL, SDLoc(N00), VT,
1487 N00.getOperand(0), N01),
1488 DAG.getNode(ISD::SHL, SDLoc(N01), VT,
1489 N00.getOperand(1), N01));
1490 return DAG.getNode(ISD::ADD, DL, VT, N0, N1);
1496 SDValue DAGCombiner::visitADD(SDNode *N) {
1497 SDValue N0 = N->getOperand(0);
1498 SDValue N1 = N->getOperand(1);
1499 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1500 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1501 EVT VT = N0.getValueType();
1504 if (VT.isVector()) {
1505 SDValue FoldedVOp = SimplifyVBinOp(N);
1506 if (FoldedVOp.getNode()) return FoldedVOp;
1508 // fold (add x, 0) -> x, vector edition
1509 if (ISD::isBuildVectorAllZeros(N1.getNode()))
1511 if (ISD::isBuildVectorAllZeros(N0.getNode()))
1515 // fold (add x, undef) -> undef
1516 if (N0.getOpcode() == ISD::UNDEF)
1518 if (N1.getOpcode() == ISD::UNDEF)
1520 // fold (add c1, c2) -> c1+c2
1522 return DAG.FoldConstantArithmetic(ISD::ADD, VT, N0C, N1C);
1523 // canonicalize constant to RHS
1525 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N1, N0);
1526 // fold (add x, 0) -> x
1527 if (N1C && N1C->isNullValue())
1529 // fold (add Sym, c) -> Sym+c
1530 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1531 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA) && N1C &&
1532 GA->getOpcode() == ISD::GlobalAddress)
1533 return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT,
1535 (uint64_t)N1C->getSExtValue());
1536 // fold ((c1-A)+c2) -> (c1+c2)-A
1537 if (N1C && N0.getOpcode() == ISD::SUB)
1538 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getOperand(0)))
1539 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1540 DAG.getConstant(N1C->getAPIntValue()+
1541 N0C->getAPIntValue(), VT),
1544 SDValue RADD = ReassociateOps(ISD::ADD, SDLoc(N), N0, N1);
1547 // fold ((0-A) + B) -> B-A
1548 if (N0.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N0.getOperand(0)) &&
1549 cast<ConstantSDNode>(N0.getOperand(0))->isNullValue())
1550 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1, N0.getOperand(1));
1551 // fold (A + (0-B)) -> A-B
1552 if (N1.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N1.getOperand(0)) &&
1553 cast<ConstantSDNode>(N1.getOperand(0))->isNullValue())
1554 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, N1.getOperand(1));
1555 // fold (A+(B-A)) -> B
1556 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1))
1557 return N1.getOperand(0);
1558 // fold ((B-A)+A) -> B
1559 if (N0.getOpcode() == ISD::SUB && N1 == N0.getOperand(1))
1560 return N0.getOperand(0);
1561 // fold (A+(B-(A+C))) to (B-C)
1562 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1563 N0 == N1.getOperand(1).getOperand(0))
1564 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1.getOperand(0),
1565 N1.getOperand(1).getOperand(1));
1566 // fold (A+(B-(C+A))) to (B-C)
1567 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1568 N0 == N1.getOperand(1).getOperand(1))
1569 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1.getOperand(0),
1570 N1.getOperand(1).getOperand(0));
1571 // fold (A+((B-A)+or-C)) to (B+or-C)
1572 if ((N1.getOpcode() == ISD::SUB || N1.getOpcode() == ISD::ADD) &&
1573 N1.getOperand(0).getOpcode() == ISD::SUB &&
1574 N0 == N1.getOperand(0).getOperand(1))
1575 return DAG.getNode(N1.getOpcode(), SDLoc(N), VT,
1576 N1.getOperand(0).getOperand(0), N1.getOperand(1));
1578 // fold (A-B)+(C-D) to (A+C)-(B+D) when A or C is constant
1579 if (N0.getOpcode() == ISD::SUB && N1.getOpcode() == ISD::SUB) {
1580 SDValue N00 = N0.getOperand(0);
1581 SDValue N01 = N0.getOperand(1);
1582 SDValue N10 = N1.getOperand(0);
1583 SDValue N11 = N1.getOperand(1);
1585 if (isa<ConstantSDNode>(N00) || isa<ConstantSDNode>(N10))
1586 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1587 DAG.getNode(ISD::ADD, SDLoc(N0), VT, N00, N10),
1588 DAG.getNode(ISD::ADD, SDLoc(N1), VT, N01, N11));
1591 if (!VT.isVector() && SimplifyDemandedBits(SDValue(N, 0)))
1592 return SDValue(N, 0);
1594 // fold (a+b) -> (a|b) iff a and b share no bits.
1595 if (VT.isInteger() && !VT.isVector()) {
1596 APInt LHSZero, LHSOne;
1597 APInt RHSZero, RHSOne;
1598 DAG.computeKnownBits(N0, LHSZero, LHSOne);
1600 if (LHSZero.getBoolValue()) {
1601 DAG.computeKnownBits(N1, RHSZero, RHSOne);
1603 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1604 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1605 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero){
1606 if (!LegalOperations || TLI.isOperationLegal(ISD::OR, VT))
1607 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1);
1612 // fold (add (shl (add x, c1), c2), ) -> (add (add (shl x, c2), c1<<c2), )
1613 if (N0.getOpcode() == ISD::SHL && N0.getNode()->hasOneUse()) {
1614 SDValue Result = combineShlAddConstant(SDLoc(N), N0, N1, DAG);
1615 if (Result.getNode()) return Result;
1617 if (N1.getOpcode() == ISD::SHL && N1.getNode()->hasOneUse()) {
1618 SDValue Result = combineShlAddConstant(SDLoc(N), N1, N0, DAG);
1619 if (Result.getNode()) return Result;
1622 // fold (add x, shl(0 - y, n)) -> sub(x, shl(y, n))
1623 if (N1.getOpcode() == ISD::SHL &&
1624 N1.getOperand(0).getOpcode() == ISD::SUB)
1625 if (ConstantSDNode *C =
1626 dyn_cast<ConstantSDNode>(N1.getOperand(0).getOperand(0)))
1627 if (C->getAPIntValue() == 0)
1628 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N0,
1629 DAG.getNode(ISD::SHL, SDLoc(N), VT,
1630 N1.getOperand(0).getOperand(1),
1632 if (N0.getOpcode() == ISD::SHL &&
1633 N0.getOperand(0).getOpcode() == ISD::SUB)
1634 if (ConstantSDNode *C =
1635 dyn_cast<ConstantSDNode>(N0.getOperand(0).getOperand(0)))
1636 if (C->getAPIntValue() == 0)
1637 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1,
1638 DAG.getNode(ISD::SHL, SDLoc(N), VT,
1639 N0.getOperand(0).getOperand(1),
1642 if (N1.getOpcode() == ISD::AND) {
1643 SDValue AndOp0 = N1.getOperand(0);
1644 ConstantSDNode *AndOp1 = dyn_cast<ConstantSDNode>(N1->getOperand(1));
1645 unsigned NumSignBits = DAG.ComputeNumSignBits(AndOp0);
1646 unsigned DestBits = VT.getScalarType().getSizeInBits();
1648 // (add z, (and (sbbl x, x), 1)) -> (sub z, (sbbl x, x))
1649 // and similar xforms where the inner op is either ~0 or 0.
1650 if (NumSignBits == DestBits && AndOp1 && AndOp1->isOne()) {
1652 return DAG.getNode(ISD::SUB, DL, VT, N->getOperand(0), AndOp0);
1656 // add (sext i1), X -> sub X, (zext i1)
1657 if (N0.getOpcode() == ISD::SIGN_EXTEND &&
1658 N0.getOperand(0).getValueType() == MVT::i1 &&
1659 !TLI.isOperationLegal(ISD::SIGN_EXTEND, MVT::i1)) {
1661 SDValue ZExt = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0));
1662 return DAG.getNode(ISD::SUB, DL, VT, N1, ZExt);
1668 SDValue DAGCombiner::visitADDC(SDNode *N) {
1669 SDValue N0 = N->getOperand(0);
1670 SDValue N1 = N->getOperand(1);
1671 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1672 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1673 EVT VT = N0.getValueType();
1675 // If the flag result is dead, turn this into an ADD.
1676 if (!N->hasAnyUseOfValue(1))
1677 return CombineTo(N, DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N1),
1678 DAG.getNode(ISD::CARRY_FALSE,
1679 SDLoc(N), MVT::Glue));
1681 // canonicalize constant to RHS.
1683 return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N1, N0);
1685 // fold (addc x, 0) -> x + no carry out
1686 if (N1C && N1C->isNullValue())
1687 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE,
1688 SDLoc(N), MVT::Glue));
1690 // fold (addc a, b) -> (or a, b), CARRY_FALSE iff a and b share no bits.
1691 APInt LHSZero, LHSOne;
1692 APInt RHSZero, RHSOne;
1693 DAG.computeKnownBits(N0, LHSZero, LHSOne);
1695 if (LHSZero.getBoolValue()) {
1696 DAG.computeKnownBits(N1, RHSZero, RHSOne);
1698 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1699 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1700 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero)
1701 return CombineTo(N, DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1),
1702 DAG.getNode(ISD::CARRY_FALSE,
1703 SDLoc(N), MVT::Glue));
1709 SDValue DAGCombiner::visitADDE(SDNode *N) {
1710 SDValue N0 = N->getOperand(0);
1711 SDValue N1 = N->getOperand(1);
1712 SDValue CarryIn = N->getOperand(2);
1713 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1714 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1716 // canonicalize constant to RHS
1718 return DAG.getNode(ISD::ADDE, SDLoc(N), N->getVTList(),
1721 // fold (adde x, y, false) -> (addc x, y)
1722 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1723 return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N0, N1);
1728 // Since it may not be valid to emit a fold to zero for vector initializers
1729 // check if we can before folding.
1730 static SDValue tryFoldToZero(SDLoc DL, const TargetLowering &TLI, EVT VT,
1732 bool LegalOperations, bool LegalTypes) {
1734 return DAG.getConstant(0, VT);
1735 if (!LegalOperations || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
1736 return DAG.getConstant(0, VT);
1740 SDValue DAGCombiner::visitSUB(SDNode *N) {
1741 SDValue N0 = N->getOperand(0);
1742 SDValue N1 = N->getOperand(1);
1743 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1744 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1745 ConstantSDNode *N1C1 = N1.getOpcode() != ISD::ADD ? nullptr :
1746 dyn_cast<ConstantSDNode>(N1.getOperand(1).getNode());
1747 EVT VT = N0.getValueType();
1750 if (VT.isVector()) {
1751 SDValue FoldedVOp = SimplifyVBinOp(N);
1752 if (FoldedVOp.getNode()) return FoldedVOp;
1754 // fold (sub x, 0) -> x, vector edition
1755 if (ISD::isBuildVectorAllZeros(N1.getNode()))
1759 // fold (sub x, x) -> 0
1760 // FIXME: Refactor this and xor and other similar operations together.
1762 return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations, LegalTypes);
1763 // fold (sub c1, c2) -> c1-c2
1765 return DAG.FoldConstantArithmetic(ISD::SUB, VT, N0C, N1C);
1766 // fold (sub x, c) -> (add x, -c)
1768 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0,
1769 DAG.getConstant(-N1C->getAPIntValue(), VT));
1770 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1)
1771 if (N0C && N0C->isAllOnesValue())
1772 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0);
1773 // fold A-(A-B) -> B
1774 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(0))
1775 return N1.getOperand(1);
1776 // fold (A+B)-A -> B
1777 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1)
1778 return N0.getOperand(1);
1779 // fold (A+B)-B -> A
1780 if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1)
1781 return N0.getOperand(0);
1782 // fold C2-(A+C1) -> (C2-C1)-A
1783 if (N1.getOpcode() == ISD::ADD && N0C && N1C1) {
1784 SDValue NewC = DAG.getConstant(N0C->getAPIntValue() - N1C1->getAPIntValue(),
1786 return DAG.getNode(ISD::SUB, SDLoc(N), VT, NewC,
1789 // fold ((A+(B+or-C))-B) -> A+or-C
1790 if (N0.getOpcode() == ISD::ADD &&
1791 (N0.getOperand(1).getOpcode() == ISD::SUB ||
1792 N0.getOperand(1).getOpcode() == ISD::ADD) &&
1793 N0.getOperand(1).getOperand(0) == N1)
1794 return DAG.getNode(N0.getOperand(1).getOpcode(), SDLoc(N), VT,
1795 N0.getOperand(0), N0.getOperand(1).getOperand(1));
1796 // fold ((A+(C+B))-B) -> A+C
1797 if (N0.getOpcode() == ISD::ADD &&
1798 N0.getOperand(1).getOpcode() == ISD::ADD &&
1799 N0.getOperand(1).getOperand(1) == N1)
1800 return DAG.getNode(ISD::ADD, SDLoc(N), VT,
1801 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1802 // fold ((A-(B-C))-C) -> A-B
1803 if (N0.getOpcode() == ISD::SUB &&
1804 N0.getOperand(1).getOpcode() == ISD::SUB &&
1805 N0.getOperand(1).getOperand(1) == N1)
1806 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1807 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1809 // If either operand of a sub is undef, the result is undef
1810 if (N0.getOpcode() == ISD::UNDEF)
1812 if (N1.getOpcode() == ISD::UNDEF)
1815 // If the relocation model supports it, consider symbol offsets.
1816 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1817 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA)) {
1818 // fold (sub Sym, c) -> Sym-c
1819 if (N1C && GA->getOpcode() == ISD::GlobalAddress)
1820 return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT,
1822 (uint64_t)N1C->getSExtValue());
1823 // fold (sub Sym+c1, Sym+c2) -> c1-c2
1824 if (GlobalAddressSDNode *GB = dyn_cast<GlobalAddressSDNode>(N1))
1825 if (GA->getGlobal() == GB->getGlobal())
1826 return DAG.getConstant((uint64_t)GA->getOffset() - GB->getOffset(),
1833 SDValue DAGCombiner::visitSUBC(SDNode *N) {
1834 SDValue N0 = N->getOperand(0);
1835 SDValue N1 = N->getOperand(1);
1836 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1837 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1838 EVT VT = N0.getValueType();
1840 // If the flag result is dead, turn this into an SUB.
1841 if (!N->hasAnyUseOfValue(1))
1842 return CombineTo(N, DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, N1),
1843 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1846 // fold (subc x, x) -> 0 + no borrow
1848 return CombineTo(N, DAG.getConstant(0, VT),
1849 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1852 // fold (subc x, 0) -> x + no borrow
1853 if (N1C && N1C->isNullValue())
1854 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1857 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1) + no borrow
1858 if (N0C && N0C->isAllOnesValue())
1859 return CombineTo(N, DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0),
1860 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1866 SDValue DAGCombiner::visitSUBE(SDNode *N) {
1867 SDValue N0 = N->getOperand(0);
1868 SDValue N1 = N->getOperand(1);
1869 SDValue CarryIn = N->getOperand(2);
1871 // fold (sube x, y, false) -> (subc x, y)
1872 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1873 return DAG.getNode(ISD::SUBC, SDLoc(N), N->getVTList(), N0, N1);
1878 SDValue DAGCombiner::visitMUL(SDNode *N) {
1879 SDValue N0 = N->getOperand(0);
1880 SDValue N1 = N->getOperand(1);
1881 EVT VT = N0.getValueType();
1883 // fold (mul x, undef) -> 0
1884 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
1885 return DAG.getConstant(0, VT);
1887 bool N0IsConst = false;
1888 bool N1IsConst = false;
1889 APInt ConstValue0, ConstValue1;
1891 if (VT.isVector()) {
1892 SDValue FoldedVOp = SimplifyVBinOp(N);
1893 if (FoldedVOp.getNode()) return FoldedVOp;
1895 N0IsConst = isConstantSplatVector(N0.getNode(), ConstValue0);
1896 N1IsConst = isConstantSplatVector(N1.getNode(), ConstValue1);
1898 N0IsConst = dyn_cast<ConstantSDNode>(N0) != nullptr;
1899 ConstValue0 = N0IsConst ? (dyn_cast<ConstantSDNode>(N0))->getAPIntValue()
1901 N1IsConst = dyn_cast<ConstantSDNode>(N1) != nullptr;
1902 ConstValue1 = N1IsConst ? (dyn_cast<ConstantSDNode>(N1))->getAPIntValue()
1906 // fold (mul c1, c2) -> c1*c2
1907 if (N0IsConst && N1IsConst)
1908 return DAG.FoldConstantArithmetic(ISD::MUL, VT, N0.getNode(), N1.getNode());
1910 // canonicalize constant to RHS
1911 if (N0IsConst && !N1IsConst)
1912 return DAG.getNode(ISD::MUL, SDLoc(N), VT, N1, N0);
1913 // fold (mul x, 0) -> 0
1914 if (N1IsConst && ConstValue1 == 0)
1916 // We require a splat of the entire scalar bit width for non-contiguous
1919 ConstValue1.getBitWidth() == VT.getScalarType().getSizeInBits();
1920 // fold (mul x, 1) -> x
1921 if (N1IsConst && ConstValue1 == 1 && IsFullSplat)
1923 // fold (mul x, -1) -> 0-x
1924 if (N1IsConst && ConstValue1.isAllOnesValue())
1925 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1926 DAG.getConstant(0, VT), N0);
1927 // fold (mul x, (1 << c)) -> x << c
1928 if (N1IsConst && ConstValue1.isPowerOf2() && IsFullSplat)
1929 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0,
1930 DAG.getConstant(ConstValue1.logBase2(),
1931 getShiftAmountTy(N0.getValueType())));
1932 // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c
1933 if (N1IsConst && (-ConstValue1).isPowerOf2() && IsFullSplat) {
1934 unsigned Log2Val = (-ConstValue1).logBase2();
1935 // FIXME: If the input is something that is easily negated (e.g. a
1936 // single-use add), we should put the negate there.
1937 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1938 DAG.getConstant(0, VT),
1939 DAG.getNode(ISD::SHL, SDLoc(N), VT, N0,
1940 DAG.getConstant(Log2Val,
1941 getShiftAmountTy(N0.getValueType()))));
1945 // (mul (shl X, c1), c2) -> (mul X, c2 << c1)
1946 if (N1IsConst && N0.getOpcode() == ISD::SHL &&
1947 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
1948 isa<ConstantSDNode>(N0.getOperand(1)))) {
1949 SDValue C3 = DAG.getNode(ISD::SHL, SDLoc(N), VT,
1950 N1, N0.getOperand(1));
1951 AddToWorklist(C3.getNode());
1952 return DAG.getNode(ISD::MUL, SDLoc(N), VT,
1953 N0.getOperand(0), C3);
1956 // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one
1959 SDValue Sh(nullptr,0), Y(nullptr,0);
1960 // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)).
1961 if (N0.getOpcode() == ISD::SHL &&
1962 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
1963 isa<ConstantSDNode>(N0.getOperand(1))) &&
1964 N0.getNode()->hasOneUse()) {
1966 } else if (N1.getOpcode() == ISD::SHL &&
1967 isa<ConstantSDNode>(N1.getOperand(1)) &&
1968 N1.getNode()->hasOneUse()) {
1973 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
1974 Sh.getOperand(0), Y);
1975 return DAG.getNode(ISD::SHL, SDLoc(N), VT,
1976 Mul, Sh.getOperand(1));
1980 // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2)
1981 if (N1IsConst && N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() &&
1982 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
1983 isa<ConstantSDNode>(N0.getOperand(1))))
1984 return DAG.getNode(ISD::ADD, SDLoc(N), VT,
1985 DAG.getNode(ISD::MUL, SDLoc(N0), VT,
1986 N0.getOperand(0), N1),
1987 DAG.getNode(ISD::MUL, SDLoc(N1), VT,
1988 N0.getOperand(1), N1));
1991 SDValue RMUL = ReassociateOps(ISD::MUL, SDLoc(N), N0, N1);
1998 SDValue DAGCombiner::visitSDIV(SDNode *N) {
1999 SDValue N0 = N->getOperand(0);
2000 SDValue N1 = N->getOperand(1);
2001 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2002 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2003 EVT VT = N->getValueType(0);
2006 if (VT.isVector()) {
2007 SDValue FoldedVOp = SimplifyVBinOp(N);
2008 if (FoldedVOp.getNode()) return FoldedVOp;
2011 // fold (sdiv c1, c2) -> c1/c2
2012 if (N0C && N1C && !N1C->isNullValue())
2013 return DAG.FoldConstantArithmetic(ISD::SDIV, VT, N0C, N1C);
2014 // fold (sdiv X, 1) -> X
2015 if (N1C && N1C->getAPIntValue() == 1LL)
2017 // fold (sdiv X, -1) -> 0-X
2018 if (N1C && N1C->isAllOnesValue())
2019 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
2020 DAG.getConstant(0, VT), N0);
2021 // If we know the sign bits of both operands are zero, strength reduce to a
2022 // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2
2023 if (!VT.isVector()) {
2024 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
2025 return DAG.getNode(ISD::UDIV, SDLoc(N), N1.getValueType(),
2029 // fold (sdiv X, pow2) -> simple ops after legalize
2030 if (N1C && !N1C->isNullValue() && (N1C->getAPIntValue().isPowerOf2() ||
2031 (-N1C->getAPIntValue()).isPowerOf2())) {
2032 // If dividing by powers of two is cheap, then don't perform the following
2034 if (TLI.isPow2DivCheap())
2037 // Target-specific implementation of sdiv x, pow2.
2038 SDValue Res = BuildSDIVPow2(N);
2042 unsigned lg2 = N1C->getAPIntValue().countTrailingZeros();
2044 // Splat the sign bit into the register
2046 DAG.getNode(ISD::SRA, SDLoc(N), VT, N0,
2047 DAG.getConstant(VT.getScalarSizeInBits() - 1,
2048 getShiftAmountTy(N0.getValueType())));
2049 AddToWorklist(SGN.getNode());
2051 // Add (N0 < 0) ? abs2 - 1 : 0;
2053 DAG.getNode(ISD::SRL, SDLoc(N), VT, SGN,
2054 DAG.getConstant(VT.getScalarSizeInBits() - lg2,
2055 getShiftAmountTy(SGN.getValueType())));
2056 SDValue ADD = DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, SRL);
2057 AddToWorklist(SRL.getNode());
2058 AddToWorklist(ADD.getNode()); // Divide by pow2
2059 SDValue SRA = DAG.getNode(ISD::SRA, SDLoc(N), VT, ADD,
2060 DAG.getConstant(lg2, getShiftAmountTy(ADD.getValueType())));
2062 // If we're dividing by a positive value, we're done. Otherwise, we must
2063 // negate the result.
2064 if (N1C->getAPIntValue().isNonNegative())
2067 AddToWorklist(SRA.getNode());
2068 return DAG.getNode(ISD::SUB, SDLoc(N), VT, DAG.getConstant(0, VT), SRA);
2071 // if integer divide is expensive and we satisfy the requirements, emit an
2072 // alternate sequence.
2073 if (N1C && !TLI.isIntDivCheap()) {
2074 SDValue Op = BuildSDIV(N);
2075 if (Op.getNode()) return Op;
2079 if (N0.getOpcode() == ISD::UNDEF)
2080 return DAG.getConstant(0, VT);
2081 // X / undef -> undef
2082 if (N1.getOpcode() == ISD::UNDEF)
2088 SDValue DAGCombiner::visitUDIV(SDNode *N) {
2089 SDValue N0 = N->getOperand(0);
2090 SDValue N1 = N->getOperand(1);
2091 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2092 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2093 EVT VT = N->getValueType(0);
2096 if (VT.isVector()) {
2097 SDValue FoldedVOp = SimplifyVBinOp(N);
2098 if (FoldedVOp.getNode()) return FoldedVOp;
2101 // fold (udiv c1, c2) -> c1/c2
2102 if (N0C && N1C && !N1C->isNullValue())
2103 return DAG.FoldConstantArithmetic(ISD::UDIV, VT, N0C, N1C);
2104 // fold (udiv x, (1 << c)) -> x >>u c
2105 if (N1C && N1C->getAPIntValue().isPowerOf2())
2106 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0,
2107 DAG.getConstant(N1C->getAPIntValue().logBase2(),
2108 getShiftAmountTy(N0.getValueType())));
2109 // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2
2110 if (N1.getOpcode() == ISD::SHL) {
2111 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
2112 if (SHC->getAPIntValue().isPowerOf2()) {
2113 EVT ADDVT = N1.getOperand(1).getValueType();
2114 SDValue Add = DAG.getNode(ISD::ADD, SDLoc(N), ADDVT,
2116 DAG.getConstant(SHC->getAPIntValue()
2119 AddToWorklist(Add.getNode());
2120 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, Add);
2124 // fold (udiv x, c) -> alternate
2125 if (N1C && !TLI.isIntDivCheap()) {
2126 SDValue Op = BuildUDIV(N);
2127 if (Op.getNode()) return Op;
2131 if (N0.getOpcode() == ISD::UNDEF)
2132 return DAG.getConstant(0, VT);
2133 // X / undef -> undef
2134 if (N1.getOpcode() == ISD::UNDEF)
2140 SDValue DAGCombiner::visitSREM(SDNode *N) {
2141 SDValue N0 = N->getOperand(0);
2142 SDValue N1 = N->getOperand(1);
2143 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2144 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2145 EVT VT = N->getValueType(0);
2147 // fold (srem c1, c2) -> c1%c2
2148 if (N0C && N1C && !N1C->isNullValue())
2149 return DAG.FoldConstantArithmetic(ISD::SREM, VT, N0C, N1C);
2150 // If we know the sign bits of both operands are zero, strength reduce to a
2151 // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15
2152 if (!VT.isVector()) {
2153 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
2154 return DAG.getNode(ISD::UREM, SDLoc(N), VT, N0, N1);
2157 // If X/C can be simplified by the division-by-constant logic, lower
2158 // X%C to the equivalent of X-X/C*C.
2159 if (N1C && !N1C->isNullValue()) {
2160 SDValue Div = DAG.getNode(ISD::SDIV, SDLoc(N), VT, N0, N1);
2161 AddToWorklist(Div.getNode());
2162 SDValue OptimizedDiv = combine(Div.getNode());
2163 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2164 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2166 SDValue Sub = DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, Mul);
2167 AddToWorklist(Mul.getNode());
2173 if (N0.getOpcode() == ISD::UNDEF)
2174 return DAG.getConstant(0, VT);
2175 // X % undef -> undef
2176 if (N1.getOpcode() == ISD::UNDEF)
2182 SDValue DAGCombiner::visitUREM(SDNode *N) {
2183 SDValue N0 = N->getOperand(0);
2184 SDValue N1 = N->getOperand(1);
2185 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2186 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2187 EVT VT = N->getValueType(0);
2189 // fold (urem c1, c2) -> c1%c2
2190 if (N0C && N1C && !N1C->isNullValue())
2191 return DAG.FoldConstantArithmetic(ISD::UREM, VT, N0C, N1C);
2192 // fold (urem x, pow2) -> (and x, pow2-1)
2193 if (N1C && !N1C->isNullValue() && N1C->getAPIntValue().isPowerOf2())
2194 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0,
2195 DAG.getConstant(N1C->getAPIntValue()-1,VT));
2196 // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1))
2197 if (N1.getOpcode() == ISD::SHL) {
2198 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
2199 if (SHC->getAPIntValue().isPowerOf2()) {
2201 DAG.getNode(ISD::ADD, SDLoc(N), VT, N1,
2202 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()),
2204 AddToWorklist(Add.getNode());
2205 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, Add);
2210 // If X/C can be simplified by the division-by-constant logic, lower
2211 // X%C to the equivalent of X-X/C*C.
2212 if (N1C && !N1C->isNullValue()) {
2213 SDValue Div = DAG.getNode(ISD::UDIV, SDLoc(N), VT, N0, N1);
2214 AddToWorklist(Div.getNode());
2215 SDValue OptimizedDiv = combine(Div.getNode());
2216 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2217 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2219 SDValue Sub = DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, Mul);
2220 AddToWorklist(Mul.getNode());
2226 if (N0.getOpcode() == ISD::UNDEF)
2227 return DAG.getConstant(0, VT);
2228 // X % undef -> undef
2229 if (N1.getOpcode() == ISD::UNDEF)
2235 SDValue DAGCombiner::visitMULHS(SDNode *N) {
2236 SDValue N0 = N->getOperand(0);
2237 SDValue N1 = N->getOperand(1);
2238 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2239 EVT VT = N->getValueType(0);
2242 // fold (mulhs x, 0) -> 0
2243 if (N1C && N1C->isNullValue())
2245 // fold (mulhs x, 1) -> (sra x, size(x)-1)
2246 if (N1C && N1C->getAPIntValue() == 1)
2247 return DAG.getNode(ISD::SRA, SDLoc(N), N0.getValueType(), N0,
2248 DAG.getConstant(N0.getValueType().getSizeInBits() - 1,
2249 getShiftAmountTy(N0.getValueType())));
2250 // fold (mulhs x, undef) -> 0
2251 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2252 return DAG.getConstant(0, VT);
2254 // If the type twice as wide is legal, transform the mulhs to a wider multiply
2256 if (VT.isSimple() && !VT.isVector()) {
2257 MVT Simple = VT.getSimpleVT();
2258 unsigned SimpleSize = Simple.getSizeInBits();
2259 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2260 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2261 N0 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N0);
2262 N1 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N1);
2263 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2264 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2265 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2266 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2273 SDValue DAGCombiner::visitMULHU(SDNode *N) {
2274 SDValue N0 = N->getOperand(0);
2275 SDValue N1 = N->getOperand(1);
2276 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2277 EVT VT = N->getValueType(0);
2280 // fold (mulhu x, 0) -> 0
2281 if (N1C && N1C->isNullValue())
2283 // fold (mulhu x, 1) -> 0
2284 if (N1C && N1C->getAPIntValue() == 1)
2285 return DAG.getConstant(0, N0.getValueType());
2286 // fold (mulhu x, undef) -> 0
2287 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2288 return DAG.getConstant(0, VT);
2290 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2292 if (VT.isSimple() && !VT.isVector()) {
2293 MVT Simple = VT.getSimpleVT();
2294 unsigned SimpleSize = Simple.getSizeInBits();
2295 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2296 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2297 N0 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N0);
2298 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N1);
2299 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2300 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2301 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2302 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2309 /// SimplifyNodeWithTwoResults - Perform optimizations common to nodes that
2310 /// compute two values. LoOp and HiOp give the opcodes for the two computations
2311 /// that are being performed. Return true if a simplification was made.
2313 SDValue DAGCombiner::SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
2315 // If the high half is not needed, just compute the low half.
2316 bool HiExists = N->hasAnyUseOfValue(1);
2318 (!LegalOperations ||
2319 TLI.isOperationLegalOrCustom(LoOp, N->getValueType(0)))) {
2320 SDValue Res = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0),
2321 ArrayRef<SDUse>(N->op_begin(), N->op_end()));
2322 return CombineTo(N, Res, Res);
2325 // If the low half is not needed, just compute the high half.
2326 bool LoExists = N->hasAnyUseOfValue(0);
2328 (!LegalOperations ||
2329 TLI.isOperationLegal(HiOp, N->getValueType(1)))) {
2330 SDValue Res = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1),
2331 ArrayRef<SDUse>(N->op_begin(), N->op_end()));
2332 return CombineTo(N, Res, Res);
2335 // If both halves are used, return as it is.
2336 if (LoExists && HiExists)
2339 // If the two computed results can be simplified separately, separate them.
2341 SDValue Lo = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0),
2342 ArrayRef<SDUse>(N->op_begin(), N->op_end()));
2343 AddToWorklist(Lo.getNode());
2344 SDValue LoOpt = combine(Lo.getNode());
2345 if (LoOpt.getNode() && LoOpt.getNode() != Lo.getNode() &&
2346 (!LegalOperations ||
2347 TLI.isOperationLegal(LoOpt.getOpcode(), LoOpt.getValueType())))
2348 return CombineTo(N, LoOpt, LoOpt);
2352 SDValue Hi = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1),
2353 ArrayRef<SDUse>(N->op_begin(), N->op_end()));
2354 AddToWorklist(Hi.getNode());
2355 SDValue HiOpt = combine(Hi.getNode());
2356 if (HiOpt.getNode() && HiOpt != Hi &&
2357 (!LegalOperations ||
2358 TLI.isOperationLegal(HiOpt.getOpcode(), HiOpt.getValueType())))
2359 return CombineTo(N, HiOpt, HiOpt);
2365 SDValue DAGCombiner::visitSMUL_LOHI(SDNode *N) {
2366 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHS);
2367 if (Res.getNode()) return Res;
2369 EVT VT = N->getValueType(0);
2372 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2374 if (VT.isSimple() && !VT.isVector()) {
2375 MVT Simple = VT.getSimpleVT();
2376 unsigned SimpleSize = Simple.getSizeInBits();
2377 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2378 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2379 SDValue Lo = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(0));
2380 SDValue Hi = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(1));
2381 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2382 // Compute the high part as N1.
2383 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2384 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2385 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2386 // Compute the low part as N0.
2387 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2388 return CombineTo(N, Lo, Hi);
2395 SDValue DAGCombiner::visitUMUL_LOHI(SDNode *N) {
2396 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHU);
2397 if (Res.getNode()) return Res;
2399 EVT VT = N->getValueType(0);
2402 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2404 if (VT.isSimple() && !VT.isVector()) {
2405 MVT Simple = VT.getSimpleVT();
2406 unsigned SimpleSize = Simple.getSizeInBits();
2407 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2408 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2409 SDValue Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(0));
2410 SDValue Hi = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(1));
2411 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2412 // Compute the high part as N1.
2413 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2414 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2415 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2416 // Compute the low part as N0.
2417 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2418 return CombineTo(N, Lo, Hi);
2425 SDValue DAGCombiner::visitSMULO(SDNode *N) {
2426 // (smulo x, 2) -> (saddo x, x)
2427 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2428 if (C2->getAPIntValue() == 2)
2429 return DAG.getNode(ISD::SADDO, SDLoc(N), N->getVTList(),
2430 N->getOperand(0), N->getOperand(0));
2435 SDValue DAGCombiner::visitUMULO(SDNode *N) {
2436 // (umulo x, 2) -> (uaddo x, x)
2437 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2438 if (C2->getAPIntValue() == 2)
2439 return DAG.getNode(ISD::UADDO, SDLoc(N), N->getVTList(),
2440 N->getOperand(0), N->getOperand(0));
2445 SDValue DAGCombiner::visitSDIVREM(SDNode *N) {
2446 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::SDIV, ISD::SREM);
2447 if (Res.getNode()) return Res;
2452 SDValue DAGCombiner::visitUDIVREM(SDNode *N) {
2453 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::UDIV, ISD::UREM);
2454 if (Res.getNode()) return Res;
2459 /// SimplifyBinOpWithSameOpcodeHands - If this is a binary operator with
2460 /// two operands of the same opcode, try to simplify it.
2461 SDValue DAGCombiner::SimplifyBinOpWithSameOpcodeHands(SDNode *N) {
2462 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
2463 EVT VT = N0.getValueType();
2464 assert(N0.getOpcode() == N1.getOpcode() && "Bad input!");
2466 // Bail early if none of these transforms apply.
2467 if (N0.getNode()->getNumOperands() == 0) return SDValue();
2469 // For each of OP in AND/OR/XOR:
2470 // fold (OP (zext x), (zext y)) -> (zext (OP x, y))
2471 // fold (OP (sext x), (sext y)) -> (sext (OP x, y))
2472 // fold (OP (aext x), (aext y)) -> (aext (OP x, y))
2473 // fold (OP (trunc x), (trunc y)) -> (trunc (OP x, y)) (if trunc isn't free)
2475 // do not sink logical op inside of a vector extend, since it may combine
2477 EVT Op0VT = N0.getOperand(0).getValueType();
2478 if ((N0.getOpcode() == ISD::ZERO_EXTEND ||
2479 N0.getOpcode() == ISD::SIGN_EXTEND ||
2480 // Avoid infinite looping with PromoteIntBinOp.
2481 (N0.getOpcode() == ISD::ANY_EXTEND &&
2482 (!LegalTypes || TLI.isTypeDesirableForOp(N->getOpcode(), Op0VT))) ||
2483 (N0.getOpcode() == ISD::TRUNCATE &&
2484 (!TLI.isZExtFree(VT, Op0VT) ||
2485 !TLI.isTruncateFree(Op0VT, VT)) &&
2486 TLI.isTypeLegal(Op0VT))) &&
2488 Op0VT == N1.getOperand(0).getValueType() &&
2489 (!LegalOperations || TLI.isOperationLegal(N->getOpcode(), Op0VT))) {
2490 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0),
2491 N0.getOperand(0).getValueType(),
2492 N0.getOperand(0), N1.getOperand(0));
2493 AddToWorklist(ORNode.getNode());
2494 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, ORNode);
2497 // For each of OP in SHL/SRL/SRA/AND...
2498 // fold (and (OP x, z), (OP y, z)) -> (OP (and x, y), z)
2499 // fold (or (OP x, z), (OP y, z)) -> (OP (or x, y), z)
2500 // fold (xor (OP x, z), (OP y, z)) -> (OP (xor x, y), z)
2501 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL ||
2502 N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::AND) &&
2503 N0.getOperand(1) == N1.getOperand(1)) {
2504 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0),
2505 N0.getOperand(0).getValueType(),
2506 N0.getOperand(0), N1.getOperand(0));
2507 AddToWorklist(ORNode.getNode());
2508 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
2509 ORNode, N0.getOperand(1));
2512 // Simplify xor/and/or (bitcast(A), bitcast(B)) -> bitcast(op (A,B))
2513 // Only perform this optimization after type legalization and before
2514 // LegalizeVectorOprs. LegalizeVectorOprs promotes vector operations by
2515 // adding bitcasts. For example (xor v4i32) is promoted to (v2i64), and
2516 // we don't want to undo this promotion.
2517 // We also handle SCALAR_TO_VECTOR because xor/or/and operations are cheaper
2519 if ((N0.getOpcode() == ISD::BITCAST ||
2520 N0.getOpcode() == ISD::SCALAR_TO_VECTOR) &&
2521 Level == AfterLegalizeTypes) {
2522 SDValue In0 = N0.getOperand(0);
2523 SDValue In1 = N1.getOperand(0);
2524 EVT In0Ty = In0.getValueType();
2525 EVT In1Ty = In1.getValueType();
2527 // If both incoming values are integers, and the original types are the
2529 if (In0Ty.isInteger() && In1Ty.isInteger() && In0Ty == In1Ty) {
2530 SDValue Op = DAG.getNode(N->getOpcode(), DL, In0Ty, In0, In1);
2531 SDValue BC = DAG.getNode(N0.getOpcode(), DL, VT, Op);
2532 AddToWorklist(Op.getNode());
2537 // Xor/and/or are indifferent to the swizzle operation (shuffle of one value).
2538 // Simplify xor/and/or (shuff(A), shuff(B)) -> shuff(op (A,B))
2539 // If both shuffles use the same mask, and both shuffle within a single
2540 // vector, then it is worthwhile to move the swizzle after the operation.
2541 // The type-legalizer generates this pattern when loading illegal
2542 // vector types from memory. In many cases this allows additional shuffle
2544 // There are other cases where moving the shuffle after the xor/and/or
2545 // is profitable even if shuffles don't perform a swizzle.
2546 // If both shuffles use the same mask, and both shuffles have the same first
2547 // or second operand, then it might still be profitable to move the shuffle
2548 // after the xor/and/or operation.
2549 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG) {
2550 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(N0);
2551 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(N1);
2553 assert(N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType() &&
2554 "Inputs to shuffles are not the same type");
2556 // Check that both shuffles use the same mask. The masks are known to be of
2557 // the same length because the result vector type is the same.
2558 // Check also that shuffles have only one use to avoid introducing extra
2560 if (SVN0->hasOneUse() && SVN1->hasOneUse() &&
2561 SVN0->getMask().equals(SVN1->getMask())) {
2562 SDValue ShOp = N0->getOperand(1);
2564 // Don't try to fold this node if it requires introducing a
2565 // build vector of all zeros that might be illegal at this stage.
2566 if (N->getOpcode() == ISD::XOR && ShOp.getOpcode() != ISD::UNDEF) {
2568 ShOp = DAG.getConstant(0, VT);
2573 // (AND (shuf (A, C), shuf (B, C)) -> shuf (AND (A, B), C)
2574 // (OR (shuf (A, C), shuf (B, C)) -> shuf (OR (A, B), C)
2575 // (XOR (shuf (A, C), shuf (B, C)) -> shuf (XOR (A, B), V_0)
2576 if (N0.getOperand(1) == N1.getOperand(1) && ShOp.getNode()) {
2577 SDValue NewNode = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
2578 N0->getOperand(0), N1->getOperand(0));
2579 AddToWorklist(NewNode.getNode());
2580 return DAG.getVectorShuffle(VT, SDLoc(N), NewNode, ShOp,
2581 &SVN0->getMask()[0]);
2584 // Don't try to fold this node if it requires introducing a
2585 // build vector of all zeros that might be illegal at this stage.
2586 ShOp = N0->getOperand(0);
2587 if (N->getOpcode() == ISD::XOR && ShOp.getOpcode() != ISD::UNDEF) {
2589 ShOp = DAG.getConstant(0, VT);
2594 // (AND (shuf (C, A), shuf (C, B)) -> shuf (C, AND (A, B))
2595 // (OR (shuf (C, A), shuf (C, B)) -> shuf (C, OR (A, B))
2596 // (XOR (shuf (C, A), shuf (C, B)) -> shuf (V_0, XOR (A, B))
2597 if (N0->getOperand(0) == N1->getOperand(0) && ShOp.getNode()) {
2598 SDValue NewNode = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
2599 N0->getOperand(1), N1->getOperand(1));
2600 AddToWorklist(NewNode.getNode());
2601 return DAG.getVectorShuffle(VT, SDLoc(N), ShOp, NewNode,
2602 &SVN0->getMask()[0]);
2610 SDValue DAGCombiner::visitAND(SDNode *N) {
2611 SDValue N0 = N->getOperand(0);
2612 SDValue N1 = N->getOperand(1);
2613 SDValue LL, LR, RL, RR, CC0, CC1;
2614 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2615 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2616 EVT VT = N1.getValueType();
2617 unsigned BitWidth = VT.getScalarType().getSizeInBits();
2620 if (VT.isVector()) {
2621 SDValue FoldedVOp = SimplifyVBinOp(N);
2622 if (FoldedVOp.getNode()) return FoldedVOp;
2624 // fold (and x, 0) -> 0, vector edition
2625 if (ISD::isBuildVectorAllZeros(N0.getNode()))
2627 if (ISD::isBuildVectorAllZeros(N1.getNode()))
2630 // fold (and x, -1) -> x, vector edition
2631 if (ISD::isBuildVectorAllOnes(N0.getNode()))
2633 if (ISD::isBuildVectorAllOnes(N1.getNode()))
2637 // fold (and x, undef) -> 0
2638 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2639 return DAG.getConstant(0, VT);
2640 // fold (and c1, c2) -> c1&c2
2642 return DAG.FoldConstantArithmetic(ISD::AND, VT, N0C, N1C);
2643 // canonicalize constant to RHS
2645 return DAG.getNode(ISD::AND, SDLoc(N), VT, N1, N0);
2646 // fold (and x, -1) -> x
2647 if (N1C && N1C->isAllOnesValue())
2649 // if (and x, c) is known to be zero, return 0
2650 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
2651 APInt::getAllOnesValue(BitWidth)))
2652 return DAG.getConstant(0, VT);
2654 SDValue RAND = ReassociateOps(ISD::AND, SDLoc(N), N0, N1);
2657 // fold (and (or x, C), D) -> D if (C & D) == D
2658 if (N1C && N0.getOpcode() == ISD::OR)
2659 if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
2660 if ((ORI->getAPIntValue() & N1C->getAPIntValue()) == N1C->getAPIntValue())
2662 // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits.
2663 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
2664 SDValue N0Op0 = N0.getOperand(0);
2665 APInt Mask = ~N1C->getAPIntValue();
2666 Mask = Mask.trunc(N0Op0.getValueSizeInBits());
2667 if (DAG.MaskedValueIsZero(N0Op0, Mask)) {
2668 SDValue Zext = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N),
2669 N0.getValueType(), N0Op0);
2671 // Replace uses of the AND with uses of the Zero extend node.
2674 // We actually want to replace all uses of the any_extend with the
2675 // zero_extend, to avoid duplicating things. This will later cause this
2676 // AND to be folded.
2677 CombineTo(N0.getNode(), Zext);
2678 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2681 // similarly fold (and (X (load ([non_ext|any_ext|zero_ext] V))), c) ->
2682 // (X (load ([non_ext|zero_ext] V))) if 'and' only clears top bits which must
2683 // already be zero by virtue of the width of the base type of the load.
2685 // the 'X' node here can either be nothing or an extract_vector_elt to catch
2687 if ((N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
2688 N0.getOperand(0).getOpcode() == ISD::LOAD) ||
2689 N0.getOpcode() == ISD::LOAD) {
2690 LoadSDNode *Load = cast<LoadSDNode>( (N0.getOpcode() == ISD::LOAD) ?
2691 N0 : N0.getOperand(0) );
2693 // Get the constant (if applicable) the zero'th operand is being ANDed with.
2694 // This can be a pure constant or a vector splat, in which case we treat the
2695 // vector as a scalar and use the splat value.
2696 APInt Constant = APInt::getNullValue(1);
2697 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
2698 Constant = C->getAPIntValue();
2699 } else if (BuildVectorSDNode *Vector = dyn_cast<BuildVectorSDNode>(N1)) {
2700 APInt SplatValue, SplatUndef;
2701 unsigned SplatBitSize;
2703 bool IsSplat = Vector->isConstantSplat(SplatValue, SplatUndef,
2704 SplatBitSize, HasAnyUndefs);
2706 // Undef bits can contribute to a possible optimisation if set, so
2708 SplatValue |= SplatUndef;
2710 // The splat value may be something like "0x00FFFFFF", which means 0 for
2711 // the first vector value and FF for the rest, repeating. We need a mask
2712 // that will apply equally to all members of the vector, so AND all the
2713 // lanes of the constant together.
2714 EVT VT = Vector->getValueType(0);
2715 unsigned BitWidth = VT.getVectorElementType().getSizeInBits();
2717 // If the splat value has been compressed to a bitlength lower
2718 // than the size of the vector lane, we need to re-expand it to
2720 if (BitWidth > SplatBitSize)
2721 for (SplatValue = SplatValue.zextOrTrunc(BitWidth);
2722 SplatBitSize < BitWidth;
2723 SplatBitSize = SplatBitSize * 2)
2724 SplatValue |= SplatValue.shl(SplatBitSize);
2726 Constant = APInt::getAllOnesValue(BitWidth);
2727 for (unsigned i = 0, n = SplatBitSize/BitWidth; i < n; ++i)
2728 Constant &= SplatValue.lshr(i*BitWidth).zextOrTrunc(BitWidth);
2732 // If we want to change an EXTLOAD to a ZEXTLOAD, ensure a ZEXTLOAD is
2733 // actually legal and isn't going to get expanded, else this is a false
2735 bool CanZextLoadProfitably = TLI.isLoadExtLegal(ISD::ZEXTLOAD,
2736 Load->getMemoryVT());
2738 // Resize the constant to the same size as the original memory access before
2739 // extension. If it is still the AllOnesValue then this AND is completely
2742 Constant.zextOrTrunc(Load->getMemoryVT().getScalarType().getSizeInBits());
2745 switch (Load->getExtensionType()) {
2746 default: B = false; break;
2747 case ISD::EXTLOAD: B = CanZextLoadProfitably; break;
2749 case ISD::NON_EXTLOAD: B = true; break;
2752 if (B && Constant.isAllOnesValue()) {
2753 // If the load type was an EXTLOAD, convert to ZEXTLOAD in order to
2754 // preserve semantics once we get rid of the AND.
2755 SDValue NewLoad(Load, 0);
2756 if (Load->getExtensionType() == ISD::EXTLOAD) {
2757 NewLoad = DAG.getLoad(Load->getAddressingMode(), ISD::ZEXTLOAD,
2758 Load->getValueType(0), SDLoc(Load),
2759 Load->getChain(), Load->getBasePtr(),
2760 Load->getOffset(), Load->getMemoryVT(),
2761 Load->getMemOperand());
2762 // Replace uses of the EXTLOAD with the new ZEXTLOAD.
2763 if (Load->getNumValues() == 3) {
2764 // PRE/POST_INC loads have 3 values.
2765 SDValue To[] = { NewLoad.getValue(0), NewLoad.getValue(1),
2766 NewLoad.getValue(2) };
2767 CombineTo(Load, To, 3, true);
2769 CombineTo(Load, NewLoad.getValue(0), NewLoad.getValue(1));
2773 // Fold the AND away, taking care not to fold to the old load node if we
2775 CombineTo(N, (N0.getNode() == Load) ? NewLoad : N0);
2777 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2780 // fold (and (setcc x), (setcc y)) -> (setcc (and x, y))
2781 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
2782 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
2783 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
2785 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
2786 LL.getValueType().isInteger()) {
2787 // fold (and (seteq X, 0), (seteq Y, 0)) -> (seteq (or X, Y), 0)
2788 if (cast<ConstantSDNode>(LR)->isNullValue() && Op1 == ISD::SETEQ) {
2789 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(N0),
2790 LR.getValueType(), LL, RL);
2791 AddToWorklist(ORNode.getNode());
2792 return DAG.getSetCC(SDLoc(N), VT, ORNode, LR, Op1);
2794 // fold (and (seteq X, -1), (seteq Y, -1)) -> (seteq (and X, Y), -1)
2795 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETEQ) {
2796 SDValue ANDNode = DAG.getNode(ISD::AND, SDLoc(N0),
2797 LR.getValueType(), LL, RL);
2798 AddToWorklist(ANDNode.getNode());
2799 return DAG.getSetCC(SDLoc(N), VT, ANDNode, LR, Op1);
2801 // fold (and (setgt X, -1), (setgt Y, -1)) -> (setgt (or X, Y), -1)
2802 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETGT) {
2803 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(N0),
2804 LR.getValueType(), LL, RL);
2805 AddToWorklist(ORNode.getNode());
2806 return DAG.getSetCC(SDLoc(N), VT, ORNode, LR, Op1);
2809 // Simplify (and (setne X, 0), (setne X, -1)) -> (setuge (add X, 1), 2)
2810 if (LL == RL && isa<ConstantSDNode>(LR) && isa<ConstantSDNode>(RR) &&
2811 Op0 == Op1 && LL.getValueType().isInteger() &&
2812 Op0 == ISD::SETNE && ((cast<ConstantSDNode>(LR)->isNullValue() &&
2813 cast<ConstantSDNode>(RR)->isAllOnesValue()) ||
2814 (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
2815 cast<ConstantSDNode>(RR)->isNullValue()))) {
2816 SDValue ADDNode = DAG.getNode(ISD::ADD, SDLoc(N0), LL.getValueType(),
2817 LL, DAG.getConstant(1, LL.getValueType()));
2818 AddToWorklist(ADDNode.getNode());
2819 return DAG.getSetCC(SDLoc(N), VT, ADDNode,
2820 DAG.getConstant(2, LL.getValueType()), ISD::SETUGE);
2822 // canonicalize equivalent to ll == rl
2823 if (LL == RR && LR == RL) {
2824 Op1 = ISD::getSetCCSwappedOperands(Op1);
2827 if (LL == RL && LR == RR) {
2828 bool isInteger = LL.getValueType().isInteger();
2829 ISD::CondCode Result = ISD::getSetCCAndOperation(Op0, Op1, isInteger);
2830 if (Result != ISD::SETCC_INVALID &&
2831 (!LegalOperations ||
2832 (TLI.isCondCodeLegal(Result, LL.getSimpleValueType()) &&
2833 TLI.isOperationLegal(ISD::SETCC,
2834 getSetCCResultType(N0.getSimpleValueType())))))
2835 return DAG.getSetCC(SDLoc(N), N0.getValueType(),
2840 // Simplify: (and (op x...), (op y...)) -> (op (and x, y))
2841 if (N0.getOpcode() == N1.getOpcode()) {
2842 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
2843 if (Tmp.getNode()) return Tmp;
2846 // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1)
2847 // fold (and (sra)) -> (and (srl)) when possible.
2848 if (!VT.isVector() &&
2849 SimplifyDemandedBits(SDValue(N, 0)))
2850 return SDValue(N, 0);
2852 // fold (zext_inreg (extload x)) -> (zextload x)
2853 if (ISD::isEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode())) {
2854 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2855 EVT MemVT = LN0->getMemoryVT();
2856 // If we zero all the possible extended bits, then we can turn this into
2857 // a zextload if we are running before legalize or the operation is legal.
2858 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
2859 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
2860 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
2861 ((!LegalOperations && !LN0->isVolatile()) ||
2862 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
2863 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT,
2864 LN0->getChain(), LN0->getBasePtr(),
2865 MemVT, LN0->getMemOperand());
2867 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
2868 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2871 // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use
2872 if (ISD::isSEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
2874 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2875 EVT MemVT = LN0->getMemoryVT();
2876 // If we zero all the possible extended bits, then we can turn this into
2877 // a zextload if we are running before legalize or the operation is legal.
2878 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
2879 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
2880 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
2881 ((!LegalOperations && !LN0->isVolatile()) ||
2882 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
2883 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT,
2884 LN0->getChain(), LN0->getBasePtr(),
2885 MemVT, LN0->getMemOperand());
2887 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
2888 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2892 // fold (and (load x), 255) -> (zextload x, i8)
2893 // fold (and (extload x, i16), 255) -> (zextload x, i8)
2894 // fold (and (any_ext (extload x, i16)), 255) -> (zextload x, i8)
2895 if (N1C && (N0.getOpcode() == ISD::LOAD ||
2896 (N0.getOpcode() == ISD::ANY_EXTEND &&
2897 N0.getOperand(0).getOpcode() == ISD::LOAD))) {
2898 bool HasAnyExt = N0.getOpcode() == ISD::ANY_EXTEND;
2899 LoadSDNode *LN0 = HasAnyExt
2900 ? cast<LoadSDNode>(N0.getOperand(0))
2901 : cast<LoadSDNode>(N0);
2902 if (LN0->getExtensionType() != ISD::SEXTLOAD &&
2903 LN0->isUnindexed() && N0.hasOneUse() && SDValue(LN0, 0).hasOneUse()) {
2904 uint32_t ActiveBits = N1C->getAPIntValue().getActiveBits();
2905 if (ActiveBits > 0 && APIntOps::isMask(ActiveBits, N1C->getAPIntValue())){
2906 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits);
2907 EVT LoadedVT = LN0->getMemoryVT();
2909 if (ExtVT == LoadedVT &&
2910 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) {
2911 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
2914 DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy,
2915 LN0->getChain(), LN0->getBasePtr(), ExtVT,
2916 LN0->getMemOperand());
2918 CombineTo(LN0, NewLoad, NewLoad.getValue(1));
2919 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2922 // Do not change the width of a volatile load.
2923 // Do not generate loads of non-round integer types since these can
2924 // be expensive (and would be wrong if the type is not byte sized).
2925 if (!LN0->isVolatile() && LoadedVT.bitsGT(ExtVT) && ExtVT.isRound() &&
2926 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) {
2927 EVT PtrType = LN0->getOperand(1).getValueType();
2929 unsigned Alignment = LN0->getAlignment();
2930 SDValue NewPtr = LN0->getBasePtr();
2932 // For big endian targets, we need to add an offset to the pointer
2933 // to load the correct bytes. For little endian systems, we merely
2934 // need to read fewer bytes from the same pointer.
2935 if (TLI.isBigEndian()) {
2936 unsigned LVTStoreBytes = LoadedVT.getStoreSize();
2937 unsigned EVTStoreBytes = ExtVT.getStoreSize();
2938 unsigned PtrOff = LVTStoreBytes - EVTStoreBytes;
2939 NewPtr = DAG.getNode(ISD::ADD, SDLoc(LN0), PtrType,
2940 NewPtr, DAG.getConstant(PtrOff, PtrType));
2941 Alignment = MinAlign(Alignment, PtrOff);
2944 AddToWorklist(NewPtr.getNode());
2946 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
2948 DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy,
2949 LN0->getChain(), NewPtr,
2950 LN0->getPointerInfo(),
2951 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
2952 Alignment, LN0->getAAInfo());
2954 CombineTo(LN0, Load, Load.getValue(1));
2955 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2961 if (N0.getOpcode() == ISD::ADD && N1.getOpcode() == ISD::SRL &&
2962 VT.getSizeInBits() <= 64) {
2963 if (ConstantSDNode *ADDI = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
2964 APInt ADDC = ADDI->getAPIntValue();
2965 if (!TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
2966 // Look for (and (add x, c1), (lshr y, c2)). If C1 wasn't a legal
2967 // immediate for an add, but it is legal if its top c2 bits are set,
2968 // transform the ADD so the immediate doesn't need to be materialized
2970 if (ConstantSDNode *SRLI = dyn_cast<ConstantSDNode>(N1.getOperand(1))) {
2971 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(),
2972 SRLI->getZExtValue());
2973 if (DAG.MaskedValueIsZero(N0.getOperand(1), Mask)) {
2975 if (TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
2977 DAG.getNode(ISD::ADD, SDLoc(N0), VT,
2978 N0.getOperand(0), DAG.getConstant(ADDC, VT));
2979 CombineTo(N0.getNode(), NewAdd);
2980 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2988 // fold (and (or (srl N, 8), (shl N, 8)), 0xffff) -> (srl (bswap N), const)
2989 if (N1C && N1C->getAPIntValue() == 0xffff && N0.getOpcode() == ISD::OR) {
2990 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
2991 N0.getOperand(1), false);
2992 if (BSwap.getNode())
2999 /// MatchBSwapHWord - Match (a >> 8) | (a << 8) as (bswap a) >> 16
3001 SDValue DAGCombiner::MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
3002 bool DemandHighBits) {
3003 if (!LegalOperations)
3006 EVT VT = N->getValueType(0);
3007 if (VT != MVT::i64 && VT != MVT::i32 && VT != MVT::i16)
3009 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
3012 // Recognize (and (shl a, 8), 0xff), (and (srl a, 8), 0xff00)
3013 bool LookPassAnd0 = false;
3014 bool LookPassAnd1 = false;
3015 if (N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode() == ISD::SRL)
3017 if (N1.getOpcode() == ISD::AND && N1.getOperand(0).getOpcode() == ISD::SHL)
3019 if (N0.getOpcode() == ISD::AND) {
3020 if (!N0.getNode()->hasOneUse())
3022 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3023 if (!N01C || N01C->getZExtValue() != 0xFF00)
3025 N0 = N0.getOperand(0);
3026 LookPassAnd0 = true;
3029 if (N1.getOpcode() == ISD::AND) {
3030 if (!N1.getNode()->hasOneUse())
3032 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
3033 if (!N11C || N11C->getZExtValue() != 0xFF)
3035 N1 = N1.getOperand(0);
3036 LookPassAnd1 = true;
3039 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
3041 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
3043 if (!N0.getNode()->hasOneUse() ||
3044 !N1.getNode()->hasOneUse())
3047 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3048 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
3051 if (N01C->getZExtValue() != 8 || N11C->getZExtValue() != 8)
3054 // Look for (shl (and a, 0xff), 8), (srl (and a, 0xff00), 8)
3055 SDValue N00 = N0->getOperand(0);
3056 if (!LookPassAnd0 && N00.getOpcode() == ISD::AND) {
3057 if (!N00.getNode()->hasOneUse())
3059 ConstantSDNode *N001C = dyn_cast<ConstantSDNode>(N00.getOperand(1));
3060 if (!N001C || N001C->getZExtValue() != 0xFF)
3062 N00 = N00.getOperand(0);
3063 LookPassAnd0 = true;
3066 SDValue N10 = N1->getOperand(0);
3067 if (!LookPassAnd1 && N10.getOpcode() == ISD::AND) {
3068 if (!N10.getNode()->hasOneUse())
3070 ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N10.getOperand(1));
3071 if (!N101C || N101C->getZExtValue() != 0xFF00)
3073 N10 = N10.getOperand(0);
3074 LookPassAnd1 = true;
3080 // Make sure everything beyond the low halfword gets set to zero since the SRL
3081 // 16 will clear the top bits.
3082 unsigned OpSizeInBits = VT.getSizeInBits();
3083 if (DemandHighBits && OpSizeInBits > 16) {
3084 // If the left-shift isn't masked out then the only way this is a bswap is
3085 // if all bits beyond the low 8 are 0. In that case the entire pattern
3086 // reduces to a left shift anyway: leave it for other parts of the combiner.
3090 // However, if the right shift isn't masked out then it might be because
3091 // it's not needed. See if we can spot that too.
3092 if (!LookPassAnd1 &&
3093 !DAG.MaskedValueIsZero(
3094 N10, APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - 16)))
3098 SDValue Res = DAG.getNode(ISD::BSWAP, SDLoc(N), VT, N00);
3099 if (OpSizeInBits > 16)
3100 Res = DAG.getNode(ISD::SRL, SDLoc(N), VT, Res,
3101 DAG.getConstant(OpSizeInBits-16, getShiftAmountTy(VT)));
3105 /// isBSwapHWordElement - Return true if the specified node is an element
3106 /// that makes up a 32-bit packed halfword byteswap. i.e.
3107 /// ((x&0xff)<<8)|((x&0xff00)>>8)|((x&0x00ff0000)<<8)|((x&0xff000000)>>8)
3108 static bool isBSwapHWordElement(SDValue N, SmallVectorImpl<SDNode *> &Parts) {
3109 if (!N.getNode()->hasOneUse())
3112 unsigned Opc = N.getOpcode();
3113 if (Opc != ISD::AND && Opc != ISD::SHL && Opc != ISD::SRL)
3116 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3121 switch (N1C->getZExtValue()) {
3124 case 0xFF: Num = 0; break;
3125 case 0xFF00: Num = 1; break;
3126 case 0xFF0000: Num = 2; break;
3127 case 0xFF000000: Num = 3; break;
3130 // Look for (x & 0xff) << 8 as well as ((x << 8) & 0xff00).
3131 SDValue N0 = N.getOperand(0);
3132 if (Opc == ISD::AND) {
3133 if (Num == 0 || Num == 2) {
3135 // (x >> 8) & 0xff0000
3136 if (N0.getOpcode() != ISD::SRL)
3138 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3139 if (!C || C->getZExtValue() != 8)
3142 // (x << 8) & 0xff00
3143 // (x << 8) & 0xff000000
3144 if (N0.getOpcode() != ISD::SHL)
3146 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3147 if (!C || C->getZExtValue() != 8)
3150 } else if (Opc == ISD::SHL) {
3152 // (x & 0xff0000) << 8
3153 if (Num != 0 && Num != 2)
3155 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3156 if (!C || C->getZExtValue() != 8)
3158 } else { // Opc == ISD::SRL
3159 // (x & 0xff00) >> 8
3160 // (x & 0xff000000) >> 8
3161 if (Num != 1 && Num != 3)
3163 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3164 if (!C || C->getZExtValue() != 8)
3171 Parts[Num] = N0.getOperand(0).getNode();
3175 /// MatchBSwapHWord - Match a 32-bit packed halfword bswap. That is
3176 /// ((x&0xff)<<8)|((x&0xff00)>>8)|((x&0x00ff0000)<<8)|((x&0xff000000)>>8)
3177 /// => (rotl (bswap x), 16)
3178 SDValue DAGCombiner::MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1) {
3179 if (!LegalOperations)
3182 EVT VT = N->getValueType(0);
3185 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
3188 SmallVector<SDNode*,4> Parts(4, (SDNode*)nullptr);
3190 // (or (or (and), (and)), (or (and), (and)))
3191 // (or (or (or (and), (and)), (and)), (and))
3192 if (N0.getOpcode() != ISD::OR)
3194 SDValue N00 = N0.getOperand(0);
3195 SDValue N01 = N0.getOperand(1);
3197 if (N1.getOpcode() == ISD::OR &&
3198 N00.getNumOperands() == 2 && N01.getNumOperands() == 2) {
3199 // (or (or (and), (and)), (or (and), (and)))
3200 SDValue N000 = N00.getOperand(0);
3201 if (!isBSwapHWordElement(N000, Parts))
3204 SDValue N001 = N00.getOperand(1);
3205 if (!isBSwapHWordElement(N001, Parts))
3207 SDValue N010 = N01.getOperand(0);
3208 if (!isBSwapHWordElement(N010, Parts))
3210 SDValue N011 = N01.getOperand(1);
3211 if (!isBSwapHWordElement(N011, Parts))
3214 // (or (or (or (and), (and)), (and)), (and))
3215 if (!isBSwapHWordElement(N1, Parts))
3217 if (!isBSwapHWordElement(N01, Parts))
3219 if (N00.getOpcode() != ISD::OR)
3221 SDValue N000 = N00.getOperand(0);
3222 if (!isBSwapHWordElement(N000, Parts))
3224 SDValue N001 = N00.getOperand(1);
3225 if (!isBSwapHWordElement(N001, Parts))
3229 // Make sure the parts are all coming from the same node.
3230 if (Parts[0] != Parts[1] || Parts[0] != Parts[2] || Parts[0] != Parts[3])
3233 SDValue BSwap = DAG.getNode(ISD::BSWAP, SDLoc(N), VT,
3234 SDValue(Parts[0],0));
3236 // Result of the bswap should be rotated by 16. If it's not legal, then
3237 // do (x << 16) | (x >> 16).
3238 SDValue ShAmt = DAG.getConstant(16, getShiftAmountTy(VT));
3239 if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT))
3240 return DAG.getNode(ISD::ROTL, SDLoc(N), VT, BSwap, ShAmt);
3241 if (TLI.isOperationLegalOrCustom(ISD::ROTR, VT))
3242 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, BSwap, ShAmt);
3243 return DAG.getNode(ISD::OR, SDLoc(N), VT,
3244 DAG.getNode(ISD::SHL, SDLoc(N), VT, BSwap, ShAmt),
3245 DAG.getNode(ISD::SRL, SDLoc(N), VT, BSwap, ShAmt));
3248 SDValue DAGCombiner::visitOR(SDNode *N) {
3249 SDValue N0 = N->getOperand(0);
3250 SDValue N1 = N->getOperand(1);
3251 SDValue LL, LR, RL, RR, CC0, CC1;
3252 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3253 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3254 EVT VT = N1.getValueType();
3257 if (VT.isVector()) {
3258 SDValue FoldedVOp = SimplifyVBinOp(N);
3259 if (FoldedVOp.getNode()) return FoldedVOp;
3261 // fold (or x, 0) -> x, vector edition
3262 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3264 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3267 // fold (or x, -1) -> -1, vector edition
3268 if (ISD::isBuildVectorAllOnes(N0.getNode()))
3270 if (ISD::isBuildVectorAllOnes(N1.getNode()))
3273 // fold (or (shuf A, V_0, MA), (shuf B, V_0, MB)) -> (shuf A, B, Mask1)
3274 // fold (or (shuf A, V_0, MA), (shuf B, V_0, MB)) -> (shuf B, A, Mask2)
3275 // Do this only if the resulting shuffle is legal.
3276 if (isa<ShuffleVectorSDNode>(N0) &&
3277 isa<ShuffleVectorSDNode>(N1) &&
3278 // Avoid folding a node with illegal type.
3279 TLI.isTypeLegal(VT) &&
3280 N0->getOperand(1) == N1->getOperand(1) &&
3281 ISD::isBuildVectorAllZeros(N0.getOperand(1).getNode())) {
3282 bool CanFold = true;
3283 unsigned NumElts = VT.getVectorNumElements();
3284 const ShuffleVectorSDNode *SV0 = cast<ShuffleVectorSDNode>(N0);
3285 const ShuffleVectorSDNode *SV1 = cast<ShuffleVectorSDNode>(N1);
3286 // We construct two shuffle masks:
3287 // - Mask1 is a shuffle mask for a shuffle with N0 as the first operand
3288 // and N1 as the second operand.
3289 // - Mask2 is a shuffle mask for a shuffle with N1 as the first operand
3290 // and N0 as the second operand.
3291 // We do this because OR is commutable and therefore there might be
3292 // two ways to fold this node into a shuffle.
3293 SmallVector<int,4> Mask1;
3294 SmallVector<int,4> Mask2;
3296 for (unsigned i = 0; i != NumElts && CanFold; ++i) {
3297 int M0 = SV0->getMaskElt(i);
3298 int M1 = SV1->getMaskElt(i);
3300 // Both shuffle indexes are undef. Propagate Undef.
3301 if (M0 < 0 && M1 < 0) {
3302 Mask1.push_back(M0);
3303 Mask2.push_back(M0);
3307 if (M0 < 0 || M1 < 0 ||
3308 (M0 < (int)NumElts && M1 < (int)NumElts) ||
3309 (M0 >= (int)NumElts && M1 >= (int)NumElts)) {
3314 Mask1.push_back(M0 < (int)NumElts ? M0 : M1 + NumElts);
3315 Mask2.push_back(M1 < (int)NumElts ? M1 : M0 + NumElts);
3319 // Fold this sequence only if the resulting shuffle is 'legal'.
3320 if (TLI.isShuffleMaskLegal(Mask1, VT))
3321 return DAG.getVectorShuffle(VT, SDLoc(N), N0->getOperand(0),
3322 N1->getOperand(0), &Mask1[0]);
3323 if (TLI.isShuffleMaskLegal(Mask2, VT))
3324 return DAG.getVectorShuffle(VT, SDLoc(N), N1->getOperand(0),
3325 N0->getOperand(0), &Mask2[0]);
3330 // fold (or x, undef) -> -1
3331 if (!LegalOperations &&
3332 (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)) {
3333 EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT;
3334 return DAG.getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()), VT);
3336 // fold (or c1, c2) -> c1|c2
3338 return DAG.FoldConstantArithmetic(ISD::OR, VT, N0C, N1C);
3339 // canonicalize constant to RHS
3341 return DAG.getNode(ISD::OR, SDLoc(N), VT, N1, N0);
3342 // fold (or x, 0) -> x
3343 if (N1C && N1C->isNullValue())
3345 // fold (or x, -1) -> -1
3346 if (N1C && N1C->isAllOnesValue())
3348 // fold (or x, c) -> c iff (x & ~c) == 0
3349 if (N1C && DAG.MaskedValueIsZero(N0, ~N1C->getAPIntValue()))
3352 // Recognize halfword bswaps as (bswap + rotl 16) or (bswap + shl 16)
3353 SDValue BSwap = MatchBSwapHWord(N, N0, N1);
3354 if (BSwap.getNode())
3356 BSwap = MatchBSwapHWordLow(N, N0, N1);
3357 if (BSwap.getNode())
3361 SDValue ROR = ReassociateOps(ISD::OR, SDLoc(N), N0, N1);
3364 // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2)
3365 // iff (c1 & c2) == 0.
3366 if (N1C && N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
3367 isa<ConstantSDNode>(N0.getOperand(1))) {
3368 ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1));
3369 if ((C1->getAPIntValue() & N1C->getAPIntValue()) != 0) {
3370 SDValue COR = DAG.FoldConstantArithmetic(ISD::OR, VT, N1C, C1);
3373 return DAG.getNode(ISD::AND, SDLoc(N), VT,
3374 DAG.getNode(ISD::OR, SDLoc(N0), VT,
3375 N0.getOperand(0), N1), COR);
3378 // fold (or (setcc x), (setcc y)) -> (setcc (or x, y))
3379 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
3380 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
3381 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
3383 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
3384 LL.getValueType().isInteger()) {
3385 // fold (or (setne X, 0), (setne Y, 0)) -> (setne (or X, Y), 0)
3386 // fold (or (setlt X, 0), (setlt Y, 0)) -> (setne (or X, Y), 0)
3387 if (cast<ConstantSDNode>(LR)->isNullValue() &&
3388 (Op1 == ISD::SETNE || Op1 == ISD::SETLT)) {
3389 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(LR),
3390 LR.getValueType(), LL, RL);
3391 AddToWorklist(ORNode.getNode());
3392 return DAG.getSetCC(SDLoc(N), VT, ORNode, LR, Op1);
3394 // fold (or (setne X, -1), (setne Y, -1)) -> (setne (and X, Y), -1)
3395 // fold (or (setgt X, -1), (setgt Y -1)) -> (setgt (and X, Y), -1)
3396 if (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
3397 (Op1 == ISD::SETNE || Op1 == ISD::SETGT)) {
3398 SDValue ANDNode = DAG.getNode(ISD::AND, SDLoc(LR),
3399 LR.getValueType(), LL, RL);
3400 AddToWorklist(ANDNode.getNode());
3401 return DAG.getSetCC(SDLoc(N), VT, ANDNode, LR, Op1);
3404 // canonicalize equivalent to ll == rl
3405 if (LL == RR && LR == RL) {
3406 Op1 = ISD::getSetCCSwappedOperands(Op1);
3409 if (LL == RL && LR == RR) {
3410 bool isInteger = LL.getValueType().isInteger();
3411 ISD::CondCode Result = ISD::getSetCCOrOperation(Op0, Op1, isInteger);
3412 if (Result != ISD::SETCC_INVALID &&
3413 (!LegalOperations ||
3414 (TLI.isCondCodeLegal(Result, LL.getSimpleValueType()) &&
3415 TLI.isOperationLegal(ISD::SETCC,
3416 getSetCCResultType(N0.getValueType())))))
3417 return DAG.getSetCC(SDLoc(N), N0.getValueType(),
3422 // Simplify: (or (op x...), (op y...)) -> (op (or x, y))
3423 if (N0.getOpcode() == N1.getOpcode()) {
3424 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3425 if (Tmp.getNode()) return Tmp;
3428 // (or (and X, C1), (and Y, C2)) -> (and (or X, Y), C3) if possible.
3429 if (N0.getOpcode() == ISD::AND &&
3430 N1.getOpcode() == ISD::AND &&
3431 N0.getOperand(1).getOpcode() == ISD::Constant &&
3432 N1.getOperand(1).getOpcode() == ISD::Constant &&
3433 // Don't increase # computations.
3434 (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) {
3435 // We can only do this xform if we know that bits from X that are set in C2
3436 // but not in C1 are already zero. Likewise for Y.
3437 const APInt &LHSMask =
3438 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
3439 const APInt &RHSMask =
3440 cast<ConstantSDNode>(N1.getOperand(1))->getAPIntValue();
3442 if (DAG.MaskedValueIsZero(N0.getOperand(0), RHSMask&~LHSMask) &&
3443 DAG.MaskedValueIsZero(N1.getOperand(0), LHSMask&~RHSMask)) {
3444 SDValue X = DAG.getNode(ISD::OR, SDLoc(N0), VT,
3445 N0.getOperand(0), N1.getOperand(0));
3446 return DAG.getNode(ISD::AND, SDLoc(N), VT, X,
3447 DAG.getConstant(LHSMask | RHSMask, VT));
3451 // See if this is some rotate idiom.
3452 if (SDNode *Rot = MatchRotate(N0, N1, SDLoc(N)))
3453 return SDValue(Rot, 0);
3455 // Simplify the operands using demanded-bits information.
3456 if (!VT.isVector() &&
3457 SimplifyDemandedBits(SDValue(N, 0)))
3458 return SDValue(N, 0);
3463 /// MatchRotateHalf - Match "(X shl/srl V1) & V2" where V2 may not be present.
3464 static bool MatchRotateHalf(SDValue Op, SDValue &Shift, SDValue &Mask) {
3465 if (Op.getOpcode() == ISD::AND) {
3466 if (isa<ConstantSDNode>(Op.getOperand(1))) {
3467 Mask = Op.getOperand(1);
3468 Op = Op.getOperand(0);
3474 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SHL) {
3482 // Return true if we can prove that, whenever Neg and Pos are both in the
3483 // range [0, OpSize), Neg == (Pos == 0 ? 0 : OpSize - Pos). This means that
3484 // for two opposing shifts shift1 and shift2 and a value X with OpBits bits:
3486 // (or (shift1 X, Neg), (shift2 X, Pos))
3488 // reduces to a rotate in direction shift2 by Pos or (equivalently) a rotate
3489 // in direction shift1 by Neg. The range [0, OpSize) means that we only need
3490 // to consider shift amounts with defined behavior.
3491 static bool matchRotateSub(SDValue Pos, SDValue Neg, unsigned OpSize) {
3492 // If OpSize is a power of 2 then:
3494 // (a) (Pos == 0 ? 0 : OpSize - Pos) == (OpSize - Pos) & (OpSize - 1)
3495 // (b) Neg == Neg & (OpSize - 1) whenever Neg is in [0, OpSize).
3497 // So if OpSize is a power of 2 and Neg is (and Neg', OpSize-1), we check
3498 // for the stronger condition:
3500 // Neg & (OpSize - 1) == (OpSize - Pos) & (OpSize - 1) [A]
3502 // for all Neg and Pos. Since Neg & (OpSize - 1) == Neg' & (OpSize - 1)
3503 // we can just replace Neg with Neg' for the rest of the function.
3505 // In other cases we check for the even stronger condition:
3507 // Neg == OpSize - Pos [B]
3509 // for all Neg and Pos. Note that the (or ...) then invokes undefined
3510 // behavior if Pos == 0 (and consequently Neg == OpSize).
3512 // We could actually use [A] whenever OpSize is a power of 2, but the
3513 // only extra cases that it would match are those uninteresting ones
3514 // where Neg and Pos are never in range at the same time. E.g. for
3515 // OpSize == 32, using [A] would allow a Neg of the form (sub 64, Pos)
3516 // as well as (sub 32, Pos), but:
3518 // (or (shift1 X, (sub 64, Pos)), (shift2 X, Pos))
3520 // always invokes undefined behavior for 32-bit X.
3522 // Below, Mask == OpSize - 1 when using [A] and is all-ones otherwise.
3523 unsigned MaskLoBits = 0;
3524 if (Neg.getOpcode() == ISD::AND &&
3525 isPowerOf2_64(OpSize) &&
3526 Neg.getOperand(1).getOpcode() == ISD::Constant &&
3527 cast<ConstantSDNode>(Neg.getOperand(1))->getAPIntValue() == OpSize - 1) {
3528 Neg = Neg.getOperand(0);
3529 MaskLoBits = Log2_64(OpSize);
3532 // Check whether Neg has the form (sub NegC, NegOp1) for some NegC and NegOp1.
3533 if (Neg.getOpcode() != ISD::SUB)
3535 ConstantSDNode *NegC = dyn_cast<ConstantSDNode>(Neg.getOperand(0));
3538 SDValue NegOp1 = Neg.getOperand(1);
3540 // On the RHS of [A], if Pos is Pos' & (OpSize - 1), just replace Pos with
3541 // Pos'. The truncation is redundant for the purpose of the equality.
3543 Pos.getOpcode() == ISD::AND &&
3544 Pos.getOperand(1).getOpcode() == ISD::Constant &&
3545 cast<ConstantSDNode>(Pos.getOperand(1))->getAPIntValue() == OpSize - 1)
3546 Pos = Pos.getOperand(0);
3548 // The condition we need is now:
3550 // (NegC - NegOp1) & Mask == (OpSize - Pos) & Mask
3552 // If NegOp1 == Pos then we need:
3554 // OpSize & Mask == NegC & Mask
3556 // (because "x & Mask" is a truncation and distributes through subtraction).
3559 Width = NegC->getAPIntValue();
3560 // Check for cases where Pos has the form (add NegOp1, PosC) for some PosC.
3561 // Then the condition we want to prove becomes:
3563 // (NegC - NegOp1) & Mask == (OpSize - (NegOp1 + PosC)) & Mask
3565 // which, again because "x & Mask" is a truncation, becomes:
3567 // NegC & Mask == (OpSize - PosC) & Mask
3568 // OpSize & Mask == (NegC + PosC) & Mask
3569 else if (Pos.getOpcode() == ISD::ADD &&
3570 Pos.getOperand(0) == NegOp1 &&
3571 Pos.getOperand(1).getOpcode() == ISD::Constant)
3572 Width = (cast<ConstantSDNode>(Pos.getOperand(1))->getAPIntValue() +
3573 NegC->getAPIntValue());
3577 // Now we just need to check that OpSize & Mask == Width & Mask.
3579 // Opsize & Mask is 0 since Mask is Opsize - 1.
3580 return Width.getLoBits(MaskLoBits) == 0;
3581 return Width == OpSize;
3584 // A subroutine of MatchRotate used once we have found an OR of two opposite
3585 // shifts of Shifted. If Neg == <operand size> - Pos then the OR reduces
3586 // to both (PosOpcode Shifted, Pos) and (NegOpcode Shifted, Neg), with the
3587 // former being preferred if supported. InnerPos and InnerNeg are Pos and
3588 // Neg with outer conversions stripped away.
3589 SDNode *DAGCombiner::MatchRotatePosNeg(SDValue Shifted, SDValue Pos,
3590 SDValue Neg, SDValue InnerPos,
3591 SDValue InnerNeg, unsigned PosOpcode,
3592 unsigned NegOpcode, SDLoc DL) {
3593 // fold (or (shl x, (*ext y)),
3594 // (srl x, (*ext (sub 32, y)))) ->
3595 // (rotl x, y) or (rotr x, (sub 32, y))
3597 // fold (or (shl x, (*ext (sub 32, y))),
3598 // (srl x, (*ext y))) ->
3599 // (rotr x, y) or (rotl x, (sub 32, y))
3600 EVT VT = Shifted.getValueType();
3601 if (matchRotateSub(InnerPos, InnerNeg, VT.getSizeInBits())) {
3602 bool HasPos = TLI.isOperationLegalOrCustom(PosOpcode, VT);
3603 return DAG.getNode(HasPos ? PosOpcode : NegOpcode, DL, VT, Shifted,
3604 HasPos ? Pos : Neg).getNode();
3610 // MatchRotate - Handle an 'or' of two operands. If this is one of the many
3611 // idioms for rotate, and if the target supports rotation instructions, generate
3613 SDNode *DAGCombiner::MatchRotate(SDValue LHS, SDValue RHS, SDLoc DL) {
3614 // Must be a legal type. Expanded 'n promoted things won't work with rotates.
3615 EVT VT = LHS.getValueType();
3616 if (!TLI.isTypeLegal(VT)) return nullptr;
3618 // The target must have at least one rotate flavor.
3619 bool HasROTL = TLI.isOperationLegalOrCustom(ISD::ROTL, VT);
3620 bool HasROTR = TLI.isOperationLegalOrCustom(ISD::ROTR, VT);
3621 if (!HasROTL && !HasROTR) return nullptr;
3623 // Match "(X shl/srl V1) & V2" where V2 may not be present.
3624 SDValue LHSShift; // The shift.
3625 SDValue LHSMask; // AND value if any.
3626 if (!MatchRotateHalf(LHS, LHSShift, LHSMask))
3627 return nullptr; // Not part of a rotate.
3629 SDValue RHSShift; // The shift.
3630 SDValue RHSMask; // AND value if any.
3631 if (!MatchRotateHalf(RHS, RHSShift, RHSMask))
3632 return nullptr; // Not part of a rotate.
3634 if (LHSShift.getOperand(0) != RHSShift.getOperand(0))
3635 return nullptr; // Not shifting the same value.
3637 if (LHSShift.getOpcode() == RHSShift.getOpcode())
3638 return nullptr; // Shifts must disagree.
3640 // Canonicalize shl to left side in a shl/srl pair.
3641 if (RHSShift.getOpcode() == ISD::SHL) {
3642 std::swap(LHS, RHS);
3643 std::swap(LHSShift, RHSShift);
3644 std::swap(LHSMask , RHSMask );
3647 unsigned OpSizeInBits = VT.getSizeInBits();
3648 SDValue LHSShiftArg = LHSShift.getOperand(0);
3649 SDValue LHSShiftAmt = LHSShift.getOperand(1);
3650 SDValue RHSShiftArg = RHSShift.getOperand(0);
3651 SDValue RHSShiftAmt = RHSShift.getOperand(1);
3653 // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1)
3654 // fold (or (shl x, C1), (srl x, C2)) -> (rotr x, C2)
3655 if (LHSShiftAmt.getOpcode() == ISD::Constant &&
3656 RHSShiftAmt.getOpcode() == ISD::Constant) {
3657 uint64_t LShVal = cast<ConstantSDNode>(LHSShiftAmt)->getZExtValue();
3658 uint64_t RShVal = cast<ConstantSDNode>(RHSShiftAmt)->getZExtValue();
3659 if ((LShVal + RShVal) != OpSizeInBits)
3662 SDValue Rot = DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT,
3663 LHSShiftArg, HasROTL ? LHSShiftAmt : RHSShiftAmt);
3665 // If there is an AND of either shifted operand, apply it to the result.
3666 if (LHSMask.getNode() || RHSMask.getNode()) {
3667 APInt Mask = APInt::getAllOnesValue(OpSizeInBits);
3669 if (LHSMask.getNode()) {
3670 APInt RHSBits = APInt::getLowBitsSet(OpSizeInBits, LShVal);
3671 Mask &= cast<ConstantSDNode>(LHSMask)->getAPIntValue() | RHSBits;
3673 if (RHSMask.getNode()) {
3674 APInt LHSBits = APInt::getHighBitsSet(OpSizeInBits, RShVal);
3675 Mask &= cast<ConstantSDNode>(RHSMask)->getAPIntValue() | LHSBits;
3678 Rot = DAG.getNode(ISD::AND, DL, VT, Rot, DAG.getConstant(Mask, VT));
3681 return Rot.getNode();
3684 // If there is a mask here, and we have a variable shift, we can't be sure
3685 // that we're masking out the right stuff.
3686 if (LHSMask.getNode() || RHSMask.getNode())
3689 // If the shift amount is sign/zext/any-extended just peel it off.
3690 SDValue LExtOp0 = LHSShiftAmt;
3691 SDValue RExtOp0 = RHSShiftAmt;
3692 if ((LHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3693 LHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3694 LHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3695 LHSShiftAmt.getOpcode() == ISD::TRUNCATE) &&
3696 (RHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3697 RHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3698 RHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3699 RHSShiftAmt.getOpcode() == ISD::TRUNCATE)) {
3700 LExtOp0 = LHSShiftAmt.getOperand(0);
3701 RExtOp0 = RHSShiftAmt.getOperand(0);
3704 SDNode *TryL = MatchRotatePosNeg(LHSShiftArg, LHSShiftAmt, RHSShiftAmt,
3705 LExtOp0, RExtOp0, ISD::ROTL, ISD::ROTR, DL);
3709 SDNode *TryR = MatchRotatePosNeg(RHSShiftArg, RHSShiftAmt, LHSShiftAmt,
3710 RExtOp0, LExtOp0, ISD::ROTR, ISD::ROTL, DL);
3717 SDValue DAGCombiner::visitXOR(SDNode *N) {
3718 SDValue N0 = N->getOperand(0);
3719 SDValue N1 = N->getOperand(1);
3720 SDValue LHS, RHS, CC;
3721 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3722 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3723 EVT VT = N0.getValueType();
3726 if (VT.isVector()) {
3727 SDValue FoldedVOp = SimplifyVBinOp(N);
3728 if (FoldedVOp.getNode()) return FoldedVOp;
3730 // fold (xor x, 0) -> x, vector edition
3731 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3733 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3737 // fold (xor undef, undef) -> 0. This is a common idiom (misuse).
3738 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
3739 return DAG.getConstant(0, VT);
3740 // fold (xor x, undef) -> undef
3741 if (N0.getOpcode() == ISD::UNDEF)
3743 if (N1.getOpcode() == ISD::UNDEF)
3745 // fold (xor c1, c2) -> c1^c2
3747 return DAG.FoldConstantArithmetic(ISD::XOR, VT, N0C, N1C);
3748 // canonicalize constant to RHS
3750 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0);
3751 // fold (xor x, 0) -> x
3752 if (N1C && N1C->isNullValue())
3755 SDValue RXOR = ReassociateOps(ISD::XOR, SDLoc(N), N0, N1);
3759 // fold !(x cc y) -> (x !cc y)
3760 if (N1C && N1C->getAPIntValue() == 1 && isSetCCEquivalent(N0, LHS, RHS, CC)) {
3761 bool isInt = LHS.getValueType().isInteger();
3762 ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
3765 if (!LegalOperations ||
3766 TLI.isCondCodeLegal(NotCC, LHS.getSimpleValueType())) {
3767 switch (N0.getOpcode()) {
3769 llvm_unreachable("Unhandled SetCC Equivalent!");
3771 return DAG.getSetCC(SDLoc(N), VT, LHS, RHS, NotCC);
3772 case ISD::SELECT_CC:
3773 return DAG.getSelectCC(SDLoc(N), LHS, RHS, N0.getOperand(2),
3774 N0.getOperand(3), NotCC);
3779 // fold (not (zext (setcc x, y))) -> (zext (not (setcc x, y)))
3780 if (N1C && N1C->getAPIntValue() == 1 && N0.getOpcode() == ISD::ZERO_EXTEND &&
3781 N0.getNode()->hasOneUse() &&
3782 isSetCCEquivalent(N0.getOperand(0), LHS, RHS, CC)){
3783 SDValue V = N0.getOperand(0);
3784 V = DAG.getNode(ISD::XOR, SDLoc(N0), V.getValueType(), V,
3785 DAG.getConstant(1, V.getValueType()));
3786 AddToWorklist(V.getNode());
3787 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, V);
3790 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are setcc
3791 if (N1C && N1C->getAPIntValue() == 1 && VT == MVT::i1 &&
3792 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3793 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3794 if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) {
3795 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3796 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS
3797 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS
3798 AddToWorklist(LHS.getNode()); AddToWorklist(RHS.getNode());
3799 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS);
3802 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are constants
3803 if (N1C && N1C->isAllOnesValue() &&
3804 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3805 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3806 if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) {
3807 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3808 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS
3809 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS
3810 AddToWorklist(LHS.getNode()); AddToWorklist(RHS.getNode());
3811 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS);
3814 // fold (xor (and x, y), y) -> (and (not x), y)
3815 if (N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
3816 N0->getOperand(1) == N1) {
3817 SDValue X = N0->getOperand(0);
3818 SDValue NotX = DAG.getNOT(SDLoc(X), X, VT);
3819 AddToWorklist(NotX.getNode());
3820 return DAG.getNode(ISD::AND, SDLoc(N), VT, NotX, N1);
3822 // fold (xor (xor x, c1), c2) -> (xor x, (xor c1, c2))
3823 if (N1C && N0.getOpcode() == ISD::XOR) {
3824 ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0));
3825 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3827 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N0.getOperand(1),
3828 DAG.getConstant(N1C->getAPIntValue() ^
3829 N00C->getAPIntValue(), VT));
3831 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N0.getOperand(0),
3832 DAG.getConstant(N1C->getAPIntValue() ^
3833 N01C->getAPIntValue(), VT));
3835 // fold (xor x, x) -> 0
3837 return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations, LegalTypes);
3839 // Simplify: xor (op x...), (op y...) -> (op (xor x, y))
3840 if (N0.getOpcode() == N1.getOpcode()) {
3841 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3842 if (Tmp.getNode()) return Tmp;
3845 // Simplify the expression using non-local knowledge.
3846 if (!VT.isVector() &&
3847 SimplifyDemandedBits(SDValue(N, 0)))
3848 return SDValue(N, 0);
3853 /// visitShiftByConstant - Handle transforms common to the three shifts, when
3854 /// the shift amount is a constant.
3855 SDValue DAGCombiner::visitShiftByConstant(SDNode *N, ConstantSDNode *Amt) {
3856 // We can't and shouldn't fold opaque constants.
3857 if (Amt->isOpaque())
3860 SDNode *LHS = N->getOperand(0).getNode();
3861 if (!LHS->hasOneUse()) return SDValue();
3863 // We want to pull some binops through shifts, so that we have (and (shift))
3864 // instead of (shift (and)), likewise for add, or, xor, etc. This sort of
3865 // thing happens with address calculations, so it's important to canonicalize
3867 bool HighBitSet = false; // Can we transform this if the high bit is set?
3869 switch (LHS->getOpcode()) {
3870 default: return SDValue();
3873 HighBitSet = false; // We can only transform sra if the high bit is clear.
3876 HighBitSet = true; // We can only transform sra if the high bit is set.
3879 if (N->getOpcode() != ISD::SHL)
3880 return SDValue(); // only shl(add) not sr[al](add).
3881 HighBitSet = false; // We can only transform sra if the high bit is clear.
3885 // We require the RHS of the binop to be a constant and not opaque as well.
3886 ConstantSDNode *BinOpCst = dyn_cast<ConstantSDNode>(LHS->getOperand(1));
3887 if (!BinOpCst || BinOpCst->isOpaque()) return SDValue();
3889 // FIXME: disable this unless the input to the binop is a shift by a constant.
3890 // If it is not a shift, it pessimizes some common cases like:
3892 // void foo(int *X, int i) { X[i & 1235] = 1; }
3893 // int bar(int *X, int i) { return X[i & 255]; }
3894 SDNode *BinOpLHSVal = LHS->getOperand(0).getNode();
3895 if ((BinOpLHSVal->getOpcode() != ISD::SHL &&
3896 BinOpLHSVal->getOpcode() != ISD::SRA &&
3897 BinOpLHSVal->getOpcode() != ISD::SRL) ||
3898 !isa<ConstantSDNode>(BinOpLHSVal->getOperand(1)))
3901 EVT VT = N->getValueType(0);
3903 // If this is a signed shift right, and the high bit is modified by the
3904 // logical operation, do not perform the transformation. The highBitSet
3905 // boolean indicates the value of the high bit of the constant which would
3906 // cause it to be modified for this operation.
3907 if (N->getOpcode() == ISD::SRA) {
3908 bool BinOpRHSSignSet = BinOpCst->getAPIntValue().isNegative();
3909 if (BinOpRHSSignSet != HighBitSet)
3913 if (!TLI.isDesirableToCommuteWithShift(LHS))
3916 // Fold the constants, shifting the binop RHS by the shift amount.
3917 SDValue NewRHS = DAG.getNode(N->getOpcode(), SDLoc(LHS->getOperand(1)),
3919 LHS->getOperand(1), N->getOperand(1));
3920 assert(isa<ConstantSDNode>(NewRHS) && "Folding was not successful!");
3922 // Create the new shift.
3923 SDValue NewShift = DAG.getNode(N->getOpcode(),
3924 SDLoc(LHS->getOperand(0)),
3925 VT, LHS->getOperand(0), N->getOperand(1));
3927 // Create the new binop.
3928 return DAG.getNode(LHS->getOpcode(), SDLoc(N), VT, NewShift, NewRHS);
3931 SDValue DAGCombiner::distributeTruncateThroughAnd(SDNode *N) {
3932 assert(N->getOpcode() == ISD::TRUNCATE);
3933 assert(N->getOperand(0).getOpcode() == ISD::AND);
3935 // (truncate:TruncVT (and N00, N01C)) -> (and (truncate:TruncVT N00), TruncC)
3936 if (N->hasOneUse() && N->getOperand(0).hasOneUse()) {
3937 SDValue N01 = N->getOperand(0).getOperand(1);
3939 if (ConstantSDNode *N01C = isConstOrConstSplat(N01)) {
3940 EVT TruncVT = N->getValueType(0);
3941 SDValue N00 = N->getOperand(0).getOperand(0);
3942 APInt TruncC = N01C->getAPIntValue();
3943 TruncC = TruncC.trunc(TruncVT.getScalarSizeInBits());
3945 return DAG.getNode(ISD::AND, SDLoc(N), TruncVT,
3946 DAG.getNode(ISD::TRUNCATE, SDLoc(N), TruncVT, N00),
3947 DAG.getConstant(TruncC, TruncVT));
3954 SDValue DAGCombiner::visitRotate(SDNode *N) {
3955 // fold (rot* x, (trunc (and y, c))) -> (rot* x, (and (trunc y), (trunc c))).
3956 if (N->getOperand(1).getOpcode() == ISD::TRUNCATE &&
3957 N->getOperand(1).getOperand(0).getOpcode() == ISD::AND) {
3958 SDValue NewOp1 = distributeTruncateThroughAnd(N->getOperand(1).getNode());
3959 if (NewOp1.getNode())
3960 return DAG.getNode(N->getOpcode(), SDLoc(N), N->getValueType(0),
3961 N->getOperand(0), NewOp1);
3966 SDValue DAGCombiner::visitSHL(SDNode *N) {
3967 SDValue N0 = N->getOperand(0);
3968 SDValue N1 = N->getOperand(1);
3969 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3970 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3971 EVT VT = N0.getValueType();
3972 unsigned OpSizeInBits = VT.getScalarSizeInBits();
3975 if (VT.isVector()) {
3976 SDValue FoldedVOp = SimplifyVBinOp(N);
3977 if (FoldedVOp.getNode()) return FoldedVOp;
3979 BuildVectorSDNode *N1CV = dyn_cast<BuildVectorSDNode>(N1);
3980 // If setcc produces all-one true value then:
3981 // (shl (and (setcc) N01CV) N1CV) -> (and (setcc) N01CV<<N1CV)
3982 if (N1CV && N1CV->isConstant()) {
3983 if (N0.getOpcode() == ISD::AND) {
3984 SDValue N00 = N0->getOperand(0);
3985 SDValue N01 = N0->getOperand(1);
3986 BuildVectorSDNode *N01CV = dyn_cast<BuildVectorSDNode>(N01);
3988 if (N01CV && N01CV->isConstant() && N00.getOpcode() == ISD::SETCC &&
3989 TLI.getBooleanContents(N00.getOperand(0).getValueType()) ==
3990 TargetLowering::ZeroOrNegativeOneBooleanContent) {
3991 SDValue C = DAG.FoldConstantArithmetic(ISD::SHL, VT, N01CV, N1CV);
3993 return DAG.getNode(ISD::AND, SDLoc(N), VT, N00, C);
3996 N1C = isConstOrConstSplat(N1);
4001 // fold (shl c1, c2) -> c1<<c2
4003 return DAG.FoldConstantArithmetic(ISD::SHL, VT, N0C, N1C);
4004 // fold (shl 0, x) -> 0
4005 if (N0C && N0C->isNullValue())
4007 // fold (shl x, c >= size(x)) -> undef
4008 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4009 return DAG.getUNDEF(VT);
4010 // fold (shl x, 0) -> x
4011 if (N1C && N1C->isNullValue())
4013 // fold (shl undef, x) -> 0
4014 if (N0.getOpcode() == ISD::UNDEF)
4015 return DAG.getConstant(0, VT);
4016 // if (shl x, c) is known to be zero, return 0
4017 if (DAG.MaskedValueIsZero(SDValue(N, 0),
4018 APInt::getAllOnesValue(OpSizeInBits)))
4019 return DAG.getConstant(0, VT);
4020 // fold (shl x, (trunc (and y, c))) -> (shl x, (and (trunc y), (trunc c))).
4021 if (N1.getOpcode() == ISD::TRUNCATE &&
4022 N1.getOperand(0).getOpcode() == ISD::AND) {
4023 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4024 if (NewOp1.getNode())
4025 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0, NewOp1);
4028 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4029 return SDValue(N, 0);
4031 // fold (shl (shl x, c1), c2) -> 0 or (shl x, (add c1, c2))
4032 if (N1C && N0.getOpcode() == ISD::SHL) {
4033 if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
4034 uint64_t c1 = N0C1->getZExtValue();
4035 uint64_t c2 = N1C->getZExtValue();
4036 if (c1 + c2 >= OpSizeInBits)
4037 return DAG.getConstant(0, VT);
4038 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0),
4039 DAG.getConstant(c1 + c2, N1.getValueType()));
4043 // fold (shl (ext (shl x, c1)), c2) -> (ext (shl x, (add c1, c2)))
4044 // For this to be valid, the second form must not preserve any of the bits
4045 // that are shifted out by the inner shift in the first form. This means
4046 // the outer shift size must be >= the number of bits added by the ext.
4047 // As a corollary, we don't care what kind of ext it is.
4048 if (N1C && (N0.getOpcode() == ISD::ZERO_EXTEND ||
4049 N0.getOpcode() == ISD::ANY_EXTEND ||
4050 N0.getOpcode() == ISD::SIGN_EXTEND) &&
4051 N0.getOperand(0).getOpcode() == ISD::SHL) {
4052 SDValue N0Op0 = N0.getOperand(0);
4053 if (ConstantSDNode *N0Op0C1 = isConstOrConstSplat(N0Op0.getOperand(1))) {
4054 uint64_t c1 = N0Op0C1->getZExtValue();
4055 uint64_t c2 = N1C->getZExtValue();
4056 EVT InnerShiftVT = N0Op0.getValueType();
4057 uint64_t InnerShiftSize = InnerShiftVT.getScalarSizeInBits();
4058 if (c2 >= OpSizeInBits - InnerShiftSize) {
4059 if (c1 + c2 >= OpSizeInBits)
4060 return DAG.getConstant(0, VT);
4061 return DAG.getNode(ISD::SHL, SDLoc(N0), VT,
4062 DAG.getNode(N0.getOpcode(), SDLoc(N0), VT,
4063 N0Op0->getOperand(0)),
4064 DAG.getConstant(c1 + c2, N1.getValueType()));
4069 // fold (shl (zext (srl x, C)), C) -> (zext (shl (srl x, C), C))
4070 // Only fold this if the inner zext has no other uses to avoid increasing
4071 // the total number of instructions.
4072 if (N1C && N0.getOpcode() == ISD::ZERO_EXTEND && N0.hasOneUse() &&
4073 N0.getOperand(0).getOpcode() == ISD::SRL) {
4074 SDValue N0Op0 = N0.getOperand(0);
4075 if (ConstantSDNode *N0Op0C1 = isConstOrConstSplat(N0Op0.getOperand(1))) {
4076 uint64_t c1 = N0Op0C1->getZExtValue();
4077 if (c1 < VT.getScalarSizeInBits()) {
4078 uint64_t c2 = N1C->getZExtValue();
4080 SDValue NewOp0 = N0.getOperand(0);
4081 EVT CountVT = NewOp0.getOperand(1).getValueType();
4082 SDValue NewSHL = DAG.getNode(ISD::SHL, SDLoc(N), NewOp0.getValueType(),
4083 NewOp0, DAG.getConstant(c2, CountVT));
4084 AddToWorklist(NewSHL.getNode());
4085 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N0), VT, NewSHL);
4091 // fold (shl (srl x, c1), c2) -> (and (shl x, (sub c2, c1), MASK) or
4092 // (and (srl x, (sub c1, c2), MASK)
4093 // Only fold this if the inner shift has no other uses -- if it does, folding
4094 // this will increase the total number of instructions.
4095 if (N1C && N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
4096 if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
4097 uint64_t c1 = N0C1->getZExtValue();
4098 if (c1 < OpSizeInBits) {
4099 uint64_t c2 = N1C->getZExtValue();
4100 APInt Mask = APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - c1);
4103 Mask = Mask.shl(c2 - c1);
4104 Shift = DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0),
4105 DAG.getConstant(c2 - c1, N1.getValueType()));
4107 Mask = Mask.lshr(c1 - c2);
4108 Shift = DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0),
4109 DAG.getConstant(c1 - c2, N1.getValueType()));
4111 return DAG.getNode(ISD::AND, SDLoc(N0), VT, Shift,
4112 DAG.getConstant(Mask, VT));
4116 // fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1))
4117 if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1)) {
4118 unsigned BitSize = VT.getScalarSizeInBits();
4119 SDValue HiBitsMask =
4120 DAG.getConstant(APInt::getHighBitsSet(BitSize,
4121 BitSize - N1C->getZExtValue()), VT);
4122 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0.getOperand(0),
4127 SDValue NewSHL = visitShiftByConstant(N, N1C);
4128 if (NewSHL.getNode())
4135 SDValue DAGCombiner::visitSRA(SDNode *N) {
4136 SDValue N0 = N->getOperand(0);
4137 SDValue N1 = N->getOperand(1);
4138 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4139 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4140 EVT VT = N0.getValueType();
4141 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
4144 if (VT.isVector()) {
4145 SDValue FoldedVOp = SimplifyVBinOp(N);
4146 if (FoldedVOp.getNode()) return FoldedVOp;
4148 N1C = isConstOrConstSplat(N1);
4151 // fold (sra c1, c2) -> (sra c1, c2)
4153 return DAG.FoldConstantArithmetic(ISD::SRA, VT, N0C, N1C);
4154 // fold (sra 0, x) -> 0
4155 if (N0C && N0C->isNullValue())
4157 // fold (sra -1, x) -> -1
4158 if (N0C && N0C->isAllOnesValue())
4160 // fold (sra x, (setge c, size(x))) -> undef
4161 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4162 return DAG.getUNDEF(VT);
4163 // fold (sra x, 0) -> x
4164 if (N1C && N1C->isNullValue())
4166 // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports
4168 if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) {
4169 unsigned LowBits = OpSizeInBits - (unsigned)N1C->getZExtValue();
4170 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), LowBits);
4172 ExtVT = EVT::getVectorVT(*DAG.getContext(),
4173 ExtVT, VT.getVectorNumElements());
4174 if ((!LegalOperations ||
4175 TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, ExtVT)))
4176 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
4177 N0.getOperand(0), DAG.getValueType(ExtVT));
4180 // fold (sra (sra x, c1), c2) -> (sra x, (add c1, c2))
4181 if (N1C && N0.getOpcode() == ISD::SRA) {
4182 if (ConstantSDNode *C1 = isConstOrConstSplat(N0.getOperand(1))) {
4183 unsigned Sum = N1C->getZExtValue() + C1->getZExtValue();
4184 if (Sum >= OpSizeInBits)
4185 Sum = OpSizeInBits - 1;
4186 return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0.getOperand(0),
4187 DAG.getConstant(Sum, N1.getValueType()));
4191 // fold (sra (shl X, m), (sub result_size, n))
4192 // -> (sign_extend (trunc (shl X, (sub (sub result_size, n), m)))) for
4193 // result_size - n != m.
4194 // If truncate is free for the target sext(shl) is likely to result in better
4196 if (N0.getOpcode() == ISD::SHL && N1C) {
4197 // Get the two constanst of the shifts, CN0 = m, CN = n.
4198 const ConstantSDNode *N01C = isConstOrConstSplat(N0.getOperand(1));
4200 LLVMContext &Ctx = *DAG.getContext();
4201 // Determine what the truncate's result bitsize and type would be.
4202 EVT TruncVT = EVT::getIntegerVT(Ctx, OpSizeInBits - N1C->getZExtValue());
4205 TruncVT = EVT::getVectorVT(Ctx, TruncVT, VT.getVectorNumElements());
4207 // Determine the residual right-shift amount.
4208 signed ShiftAmt = N1C->getZExtValue() - N01C->getZExtValue();
4210 // If the shift is not a no-op (in which case this should be just a sign
4211 // extend already), the truncated to type is legal, sign_extend is legal
4212 // on that type, and the truncate to that type is both legal and free,
4213 // perform the transform.
4214 if ((ShiftAmt > 0) &&
4215 TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND, TruncVT) &&
4216 TLI.isOperationLegalOrCustom(ISD::TRUNCATE, VT) &&
4217 TLI.isTruncateFree(VT, TruncVT)) {
4219 SDValue Amt = DAG.getConstant(ShiftAmt,
4220 getShiftAmountTy(N0.getOperand(0).getValueType()));
4221 SDValue Shift = DAG.getNode(ISD::SRL, SDLoc(N0), VT,
4222 N0.getOperand(0), Amt);
4223 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), TruncVT,
4225 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N),
4226 N->getValueType(0), Trunc);
4231 // fold (sra x, (trunc (and y, c))) -> (sra x, (and (trunc y), (trunc c))).
4232 if (N1.getOpcode() == ISD::TRUNCATE &&
4233 N1.getOperand(0).getOpcode() == ISD::AND) {
4234 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4235 if (NewOp1.getNode())
4236 return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0, NewOp1);
4239 // fold (sra (trunc (srl x, c1)), c2) -> (trunc (sra x, c1 + c2))
4240 // if c1 is equal to the number of bits the trunc removes
4241 if (N0.getOpcode() == ISD::TRUNCATE &&
4242 (N0.getOperand(0).getOpcode() == ISD::SRL ||
4243 N0.getOperand(0).getOpcode() == ISD::SRA) &&
4244 N0.getOperand(0).hasOneUse() &&
4245 N0.getOperand(0).getOperand(1).hasOneUse() &&
4247 SDValue N0Op0 = N0.getOperand(0);
4248 if (ConstantSDNode *LargeShift = isConstOrConstSplat(N0Op0.getOperand(1))) {
4249 unsigned LargeShiftVal = LargeShift->getZExtValue();
4250 EVT LargeVT = N0Op0.getValueType();
4252 if (LargeVT.getScalarSizeInBits() - OpSizeInBits == LargeShiftVal) {
4254 DAG.getConstant(LargeShiftVal + N1C->getZExtValue(),
4255 getShiftAmountTy(N0Op0.getOperand(0).getValueType()));
4256 SDValue SRA = DAG.getNode(ISD::SRA, SDLoc(N), LargeVT,
4257 N0Op0.getOperand(0), Amt);
4258 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, SRA);
4263 // Simplify, based on bits shifted out of the LHS.
4264 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4265 return SDValue(N, 0);
4268 // If the sign bit is known to be zero, switch this to a SRL.
4269 if (DAG.SignBitIsZero(N0))
4270 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, N1);
4273 SDValue NewSRA = visitShiftByConstant(N, N1C);
4274 if (NewSRA.getNode())
4281 SDValue DAGCombiner::visitSRL(SDNode *N) {
4282 SDValue N0 = N->getOperand(0);
4283 SDValue N1 = N->getOperand(1);
4284 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4285 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4286 EVT VT = N0.getValueType();
4287 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
4290 if (VT.isVector()) {
4291 SDValue FoldedVOp = SimplifyVBinOp(N);
4292 if (FoldedVOp.getNode()) return FoldedVOp;
4294 N1C = isConstOrConstSplat(N1);
4297 // fold (srl c1, c2) -> c1 >>u c2
4299 return DAG.FoldConstantArithmetic(ISD::SRL, VT, N0C, N1C);
4300 // fold (srl 0, x) -> 0
4301 if (N0C && N0C->isNullValue())
4303 // fold (srl x, c >= size(x)) -> undef
4304 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4305 return DAG.getUNDEF(VT);
4306 // fold (srl x, 0) -> x
4307 if (N1C && N1C->isNullValue())
4309 // if (srl x, c) is known to be zero, return 0
4310 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
4311 APInt::getAllOnesValue(OpSizeInBits)))
4312 return DAG.getConstant(0, VT);
4314 // fold (srl (srl x, c1), c2) -> 0 or (srl x, (add c1, c2))
4315 if (N1C && N0.getOpcode() == ISD::SRL) {
4316 if (ConstantSDNode *N01C = isConstOrConstSplat(N0.getOperand(1))) {
4317 uint64_t c1 = N01C->getZExtValue();
4318 uint64_t c2 = N1C->getZExtValue();
4319 if (c1 + c2 >= OpSizeInBits)
4320 return DAG.getConstant(0, VT);
4321 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0),
4322 DAG.getConstant(c1 + c2, N1.getValueType()));
4326 // fold (srl (trunc (srl x, c1)), c2) -> 0 or (trunc (srl x, (add c1, c2)))
4327 if (N1C && N0.getOpcode() == ISD::TRUNCATE &&
4328 N0.getOperand(0).getOpcode() == ISD::SRL &&
4329 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
4331 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
4332 uint64_t c2 = N1C->getZExtValue();
4333 EVT InnerShiftVT = N0.getOperand(0).getValueType();
4334 EVT ShiftCountVT = N0.getOperand(0)->getOperand(1).getValueType();
4335 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
4336 // This is only valid if the OpSizeInBits + c1 = size of inner shift.
4337 if (c1 + OpSizeInBits == InnerShiftSize) {
4338 if (c1 + c2 >= InnerShiftSize)
4339 return DAG.getConstant(0, VT);
4340 return DAG.getNode(ISD::TRUNCATE, SDLoc(N0), VT,
4341 DAG.getNode(ISD::SRL, SDLoc(N0), InnerShiftVT,
4342 N0.getOperand(0)->getOperand(0),
4343 DAG.getConstant(c1 + c2, ShiftCountVT)));
4347 // fold (srl (shl x, c), c) -> (and x, cst2)
4348 if (N1C && N0.getOpcode() == ISD::SHL && N0.getOperand(1) == N1) {
4349 unsigned BitSize = N0.getScalarValueSizeInBits();
4350 if (BitSize <= 64) {
4351 uint64_t ShAmt = N1C->getZExtValue() + 64 - BitSize;
4352 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0.getOperand(0),
4353 DAG.getConstant(~0ULL >> ShAmt, VT));
4357 // fold (srl (anyextend x), c) -> (and (anyextend (srl x, c)), mask)
4358 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
4359 // Shifting in all undef bits?
4360 EVT SmallVT = N0.getOperand(0).getValueType();
4361 unsigned BitSize = SmallVT.getScalarSizeInBits();
4362 if (N1C->getZExtValue() >= BitSize)
4363 return DAG.getUNDEF(VT);
4365 if (!LegalTypes || TLI.isTypeDesirableForOp(ISD::SRL, SmallVT)) {
4366 uint64_t ShiftAmt = N1C->getZExtValue();
4367 SDValue SmallShift = DAG.getNode(ISD::SRL, SDLoc(N0), SmallVT,
4369 DAG.getConstant(ShiftAmt, getShiftAmountTy(SmallVT)));
4370 AddToWorklist(SmallShift.getNode());
4371 APInt Mask = APInt::getAllOnesValue(OpSizeInBits).lshr(ShiftAmt);
4372 return DAG.getNode(ISD::AND, SDLoc(N), VT,
4373 DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, SmallShift),
4374 DAG.getConstant(Mask, VT));
4378 // fold (srl (sra X, Y), 31) -> (srl X, 31). This srl only looks at the sign
4379 // bit, which is unmodified by sra.
4380 if (N1C && N1C->getZExtValue() + 1 == OpSizeInBits) {
4381 if (N0.getOpcode() == ISD::SRA)
4382 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0), N1);
4385 // fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit).
4386 if (N1C && N0.getOpcode() == ISD::CTLZ &&
4387 N1C->getAPIntValue() == Log2_32(OpSizeInBits)) {
4388 APInt KnownZero, KnownOne;
4389 DAG.computeKnownBits(N0.getOperand(0), KnownZero, KnownOne);
4391 // If any of the input bits are KnownOne, then the input couldn't be all
4392 // zeros, thus the result of the srl will always be zero.
4393 if (KnownOne.getBoolValue()) return DAG.getConstant(0, VT);
4395 // If all of the bits input the to ctlz node are known to be zero, then
4396 // the result of the ctlz is "32" and the result of the shift is one.
4397 APInt UnknownBits = ~KnownZero;
4398 if (UnknownBits == 0) return DAG.getConstant(1, VT);
4400 // Otherwise, check to see if there is exactly one bit input to the ctlz.
4401 if ((UnknownBits & (UnknownBits - 1)) == 0) {
4402 // Okay, we know that only that the single bit specified by UnknownBits
4403 // could be set on input to the CTLZ node. If this bit is set, the SRL
4404 // will return 0, if it is clear, it returns 1. Change the CTLZ/SRL pair
4405 // to an SRL/XOR pair, which is likely to simplify more.
4406 unsigned ShAmt = UnknownBits.countTrailingZeros();
4407 SDValue Op = N0.getOperand(0);
4410 Op = DAG.getNode(ISD::SRL, SDLoc(N0), VT, Op,
4411 DAG.getConstant(ShAmt, getShiftAmountTy(Op.getValueType())));
4412 AddToWorklist(Op.getNode());
4415 return DAG.getNode(ISD::XOR, SDLoc(N), VT,
4416 Op, DAG.getConstant(1, VT));
4420 // fold (srl x, (trunc (and y, c))) -> (srl x, (and (trunc y), (trunc c))).
4421 if (N1.getOpcode() == ISD::TRUNCATE &&
4422 N1.getOperand(0).getOpcode() == ISD::AND) {
4423 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4424 if (NewOp1.getNode())
4425 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, NewOp1);
4428 // fold operands of srl based on knowledge that the low bits are not
4430 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4431 return SDValue(N, 0);
4434 SDValue NewSRL = visitShiftByConstant(N, N1C);
4435 if (NewSRL.getNode())
4439 // Attempt to convert a srl of a load into a narrower zero-extending load.
4440 SDValue NarrowLoad = ReduceLoadWidth(N);
4441 if (NarrowLoad.getNode())
4444 // Here is a common situation. We want to optimize:
4447 // %b = and i32 %a, 2
4448 // %c = srl i32 %b, 1
4449 // brcond i32 %c ...
4455 // %c = setcc eq %b, 0
4458 // However when after the source operand of SRL is optimized into AND, the SRL
4459 // itself may not be optimized further. Look for it and add the BRCOND into
4461 if (N->hasOneUse()) {
4462 SDNode *Use = *N->use_begin();
4463 if (Use->getOpcode() == ISD::BRCOND)
4465 else if (Use->getOpcode() == ISD::TRUNCATE && Use->hasOneUse()) {
4466 // Also look pass the truncate.
4467 Use = *Use->use_begin();
4468 if (Use->getOpcode() == ISD::BRCOND)
4476 SDValue DAGCombiner::visitCTLZ(SDNode *N) {
4477 SDValue N0 = N->getOperand(0);
4478 EVT VT = N->getValueType(0);
4480 // fold (ctlz c1) -> c2
4481 if (isa<ConstantSDNode>(N0))
4482 return DAG.getNode(ISD::CTLZ, SDLoc(N), VT, N0);
4486 SDValue DAGCombiner::visitCTLZ_ZERO_UNDEF(SDNode *N) {
4487 SDValue N0 = N->getOperand(0);
4488 EVT VT = N->getValueType(0);
4490 // fold (ctlz_zero_undef c1) -> c2
4491 if (isa<ConstantSDNode>(N0))
4492 return DAG.getNode(ISD::CTLZ_ZERO_UNDEF, SDLoc(N), VT, N0);
4496 SDValue DAGCombiner::visitCTTZ(SDNode *N) {
4497 SDValue N0 = N->getOperand(0);
4498 EVT VT = N->getValueType(0);
4500 // fold (cttz c1) -> c2
4501 if (isa<ConstantSDNode>(N0))
4502 return DAG.getNode(ISD::CTTZ, SDLoc(N), VT, N0);
4506 SDValue DAGCombiner::visitCTTZ_ZERO_UNDEF(SDNode *N) {
4507 SDValue N0 = N->getOperand(0);
4508 EVT VT = N->getValueType(0);
4510 // fold (cttz_zero_undef c1) -> c2
4511 if (isa<ConstantSDNode>(N0))
4512 return DAG.getNode(ISD::CTTZ_ZERO_UNDEF, SDLoc(N), VT, N0);
4516 SDValue DAGCombiner::visitCTPOP(SDNode *N) {
4517 SDValue N0 = N->getOperand(0);
4518 EVT VT = N->getValueType(0);
4520 // fold (ctpop c1) -> c2
4521 if (isa<ConstantSDNode>(N0))
4522 return DAG.getNode(ISD::CTPOP, SDLoc(N), VT, N0);
4526 SDValue DAGCombiner::visitSELECT(SDNode *N) {
4527 SDValue N0 = N->getOperand(0);
4528 SDValue N1 = N->getOperand(1);
4529 SDValue N2 = N->getOperand(2);
4530 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4531 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4532 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
4533 EVT VT = N->getValueType(0);
4534 EVT VT0 = N0.getValueType();
4536 // fold (select C, X, X) -> X
4539 // fold (select true, X, Y) -> X
4540 if (N0C && !N0C->isNullValue())
4542 // fold (select false, X, Y) -> Y
4543 if (N0C && N0C->isNullValue())
4545 // fold (select C, 1, X) -> (or C, X)
4546 if (VT == MVT::i1 && N1C && N1C->getAPIntValue() == 1)
4547 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N2);
4548 // fold (select C, 0, 1) -> (xor C, 1)
4549 // We can't do this reliably if integer based booleans have different contents
4550 // to floating point based booleans. This is because we can't tell whether we
4551 // have an integer-based boolean or a floating-point-based boolean unless we
4552 // can find the SETCC that produced it and inspect its operands. This is
4553 // fairly easy if C is the SETCC node, but it can potentially be
4554 // undiscoverable (or not reasonably discoverable). For example, it could be
4555 // in another basic block or it could require searching a complicated
4557 if (VT.isInteger() &&
4558 (VT0 == MVT::i1 || (VT0.isInteger() &&
4559 TLI.getBooleanContents(false, false) ==
4560 TLI.getBooleanContents(false, true) &&
4561 TLI.getBooleanContents(false, false) ==
4562 TargetLowering::ZeroOrOneBooleanContent)) &&
4563 N1C && N2C && N1C->isNullValue() && N2C->getAPIntValue() == 1) {
4566 return DAG.getNode(ISD::XOR, SDLoc(N), VT0,
4567 N0, DAG.getConstant(1, VT0));
4568 XORNode = DAG.getNode(ISD::XOR, SDLoc(N0), VT0,
4569 N0, DAG.getConstant(1, VT0));
4570 AddToWorklist(XORNode.getNode());
4572 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, XORNode);
4573 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, XORNode);
4575 // fold (select C, 0, X) -> (and (not C), X)
4576 if (VT == VT0 && VT == MVT::i1 && N1C && N1C->isNullValue()) {
4577 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT);
4578 AddToWorklist(NOTNode.getNode());
4579 return DAG.getNode(ISD::AND, SDLoc(N), VT, NOTNode, N2);
4581 // fold (select C, X, 1) -> (or (not C), X)
4582 if (VT == VT0 && VT == MVT::i1 && N2C && N2C->getAPIntValue() == 1) {
4583 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT);
4584 AddToWorklist(NOTNode.getNode());
4585 return DAG.getNode(ISD::OR, SDLoc(N), VT, NOTNode, N1);
4587 // fold (select C, X, 0) -> (and C, X)
4588 if (VT == MVT::i1 && N2C && N2C->isNullValue())
4589 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, N1);
4590 // fold (select X, X, Y) -> (or X, Y)
4591 // fold (select X, 1, Y) -> (or X, Y)
4592 if (VT == MVT::i1 && (N0 == N1 || (N1C && N1C->getAPIntValue() == 1)))
4593 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N2);
4594 // fold (select X, Y, X) -> (and X, Y)
4595 // fold (select X, Y, 0) -> (and X, Y)
4596 if (VT == MVT::i1 && (N0 == N2 || (N2C && N2C->getAPIntValue() == 0)))
4597 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, N1);
4599 // If we can fold this based on the true/false value, do so.
4600 if (SimplifySelectOps(N, N1, N2))
4601 return SDValue(N, 0); // Don't revisit N.
4603 // fold selects based on a setcc into other things, such as min/max/abs
4604 if (N0.getOpcode() == ISD::SETCC) {
4605 if ((!LegalOperations &&
4606 TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT)) ||
4607 TLI.isOperationLegal(ISD::SELECT_CC, VT))
4608 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT,
4609 N0.getOperand(0), N0.getOperand(1),
4610 N1, N2, N0.getOperand(2));
4611 return SimplifySelect(SDLoc(N), N0, N1, N2);
4618 std::pair<SDValue, SDValue> SplitVSETCC(const SDNode *N, SelectionDAG &DAG) {
4621 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0));
4623 // Split the inputs.
4624 SDValue Lo, Hi, LL, LH, RL, RH;
4625 std::tie(LL, LH) = DAG.SplitVectorOperand(N, 0);
4626 std::tie(RL, RH) = DAG.SplitVectorOperand(N, 1);
4628 Lo = DAG.getNode(N->getOpcode(), DL, LoVT, LL, RL, N->getOperand(2));
4629 Hi = DAG.getNode(N->getOpcode(), DL, HiVT, LH, RH, N->getOperand(2));
4631 return std::make_pair(Lo, Hi);
4634 // This function assumes all the vselect's arguments are CONCAT_VECTOR
4635 // nodes and that the condition is a BV of ConstantSDNodes (or undefs).
4636 static SDValue ConvertSelectToConcatVector(SDNode *N, SelectionDAG &DAG) {
4638 SDValue Cond = N->getOperand(0);
4639 SDValue LHS = N->getOperand(1);
4640 SDValue RHS = N->getOperand(2);
4641 MVT VT = N->getSimpleValueType(0);
4642 int NumElems = VT.getVectorNumElements();
4643 assert(LHS.getOpcode() == ISD::CONCAT_VECTORS &&
4644 RHS.getOpcode() == ISD::CONCAT_VECTORS &&
4645 Cond.getOpcode() == ISD::BUILD_VECTOR);
4647 // We're sure we have an even number of elements due to the
4648 // concat_vectors we have as arguments to vselect.
4649 // Skip BV elements until we find one that's not an UNDEF
4650 // After we find an UNDEF element, keep looping until we get to half the
4651 // length of the BV and see if all the non-undef nodes are the same.
4652 ConstantSDNode *BottomHalf = nullptr;
4653 for (int i = 0; i < NumElems / 2; ++i) {
4654 if (Cond->getOperand(i)->getOpcode() == ISD::UNDEF)
4657 if (BottomHalf == nullptr)
4658 BottomHalf = cast<ConstantSDNode>(Cond.getOperand(i));
4659 else if (Cond->getOperand(i).getNode() != BottomHalf)
4663 // Do the same for the second half of the BuildVector
4664 ConstantSDNode *TopHalf = nullptr;
4665 for (int i = NumElems / 2; i < NumElems; ++i) {
4666 if (Cond->getOperand(i)->getOpcode() == ISD::UNDEF)
4669 if (TopHalf == nullptr)
4670 TopHalf = cast<ConstantSDNode>(Cond.getOperand(i));
4671 else if (Cond->getOperand(i).getNode() != TopHalf)
4675 assert(TopHalf && BottomHalf &&
4676 "One half of the selector was all UNDEFs and the other was all the "
4677 "same value. This should have been addressed before this function.");
4679 ISD::CONCAT_VECTORS, dl, VT,
4680 BottomHalf->isNullValue() ? RHS->getOperand(0) : LHS->getOperand(0),
4681 TopHalf->isNullValue() ? RHS->getOperand(1) : LHS->getOperand(1));
4684 SDValue DAGCombiner::visitVSELECT(SDNode *N) {
4685 SDValue N0 = N->getOperand(0);
4686 SDValue N1 = N->getOperand(1);
4687 SDValue N2 = N->getOperand(2);
4690 // Canonicalize integer abs.
4691 // vselect (setg[te] X, 0), X, -X ->
4692 // vselect (setgt X, -1), X, -X ->
4693 // vselect (setl[te] X, 0), -X, X ->
4694 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
4695 if (N0.getOpcode() == ISD::SETCC) {
4696 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
4697 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
4699 bool RHSIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
4701 if (((RHSIsAllZeros && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
4702 (ISD::isBuildVectorAllOnes(RHS.getNode()) && CC == ISD::SETGT)) &&
4703 N1 == LHS && N2.getOpcode() == ISD::SUB && N1 == N2.getOperand(1))
4704 isAbs = ISD::isBuildVectorAllZeros(N2.getOperand(0).getNode());
4705 else if ((RHSIsAllZeros && (CC == ISD::SETLT || CC == ISD::SETLE)) &&
4706 N2 == LHS && N1.getOpcode() == ISD::SUB && N2 == N1.getOperand(1))
4707 isAbs = ISD::isBuildVectorAllZeros(N1.getOperand(0).getNode());
4710 EVT VT = LHS.getValueType();
4711 SDValue Shift = DAG.getNode(
4712 ISD::SRA, DL, VT, LHS,
4713 DAG.getConstant(VT.getScalarType().getSizeInBits() - 1, VT));
4714 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, LHS, Shift);
4715 AddToWorklist(Shift.getNode());
4716 AddToWorklist(Add.getNode());
4717 return DAG.getNode(ISD::XOR, DL, VT, Add, Shift);
4721 // If the VSELECT result requires splitting and the mask is provided by a
4722 // SETCC, then split both nodes and its operands before legalization. This
4723 // prevents the type legalizer from unrolling SETCC into scalar comparisons
4724 // and enables future optimizations (e.g. min/max pattern matching on X86).
4725 if (N0.getOpcode() == ISD::SETCC) {
4726 EVT VT = N->getValueType(0);
4728 // Check if any splitting is required.
4729 if (TLI.getTypeAction(*DAG.getContext(), VT) !=
4730 TargetLowering::TypeSplitVector)
4733 SDValue Lo, Hi, CCLo, CCHi, LL, LH, RL, RH;
4734 std::tie(CCLo, CCHi) = SplitVSETCC(N0.getNode(), DAG);
4735 std::tie(LL, LH) = DAG.SplitVectorOperand(N, 1);
4736 std::tie(RL, RH) = DAG.SplitVectorOperand(N, 2);
4738 Lo = DAG.getNode(N->getOpcode(), DL, LL.getValueType(), CCLo, LL, RL);
4739 Hi = DAG.getNode(N->getOpcode(), DL, LH.getValueType(), CCHi, LH, RH);
4741 // Add the new VSELECT nodes to the work list in case they need to be split
4743 AddToWorklist(Lo.getNode());
4744 AddToWorklist(Hi.getNode());
4746 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
4749 // Fold (vselect (build_vector all_ones), N1, N2) -> N1
4750 if (ISD::isBuildVectorAllOnes(N0.getNode()))
4752 // Fold (vselect (build_vector all_zeros), N1, N2) -> N2
4753 if (ISD::isBuildVectorAllZeros(N0.getNode()))
4756 // The ConvertSelectToConcatVector function is assuming both the above
4757 // checks for (vselect (build_vector all{ones,zeros) ...) have been made
4759 if (N1.getOpcode() == ISD::CONCAT_VECTORS &&
4760 N2.getOpcode() == ISD::CONCAT_VECTORS &&
4761 ISD::isBuildVectorOfConstantSDNodes(N0.getNode())) {
4762 SDValue CV = ConvertSelectToConcatVector(N, DAG);
4770 SDValue DAGCombiner::visitSELECT_CC(SDNode *N) {
4771 SDValue N0 = N->getOperand(0);
4772 SDValue N1 = N->getOperand(1);
4773 SDValue N2 = N->getOperand(2);
4774 SDValue N3 = N->getOperand(3);
4775 SDValue N4 = N->getOperand(4);
4776 ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get();
4778 // fold select_cc lhs, rhs, x, x, cc -> x
4782 // Determine if the condition we're dealing with is constant
4783 SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()),
4784 N0, N1, CC, SDLoc(N), false);
4785 if (SCC.getNode()) {
4786 AddToWorklist(SCC.getNode());
4788 if (ConstantSDNode *SCCC = dyn_cast<ConstantSDNode>(SCC.getNode())) {
4789 if (!SCCC->isNullValue())
4790 return N2; // cond always true -> true val
4792 return N3; // cond always false -> false val
4795 // Fold to a simpler select_cc
4796 if (SCC.getOpcode() == ISD::SETCC)
4797 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), N2.getValueType(),
4798 SCC.getOperand(0), SCC.getOperand(1), N2, N3,
4802 // If we can fold this based on the true/false value, do so.
4803 if (SimplifySelectOps(N, N2, N3))
4804 return SDValue(N, 0); // Don't revisit N.
4806 // fold select_cc into other things, such as min/max/abs
4807 return SimplifySelectCC(SDLoc(N), N0, N1, N2, N3, CC);
4810 SDValue DAGCombiner::visitSETCC(SDNode *N) {
4811 return SimplifySetCC(N->getValueType(0), N->getOperand(0), N->getOperand(1),
4812 cast<CondCodeSDNode>(N->getOperand(2))->get(),
4816 // tryToFoldExtendOfConstant - Try to fold a sext/zext/aext
4817 // dag node into a ConstantSDNode or a build_vector of constants.
4818 // This function is called by the DAGCombiner when visiting sext/zext/aext
4819 // dag nodes (see for example method DAGCombiner::visitSIGN_EXTEND).
4820 // Vector extends are not folded if operations are legal; this is to
4821 // avoid introducing illegal build_vector dag nodes.
4822 static SDNode *tryToFoldExtendOfConstant(SDNode *N, const TargetLowering &TLI,
4823 SelectionDAG &DAG, bool LegalTypes,
4824 bool LegalOperations) {
4825 unsigned Opcode = N->getOpcode();
4826 SDValue N0 = N->getOperand(0);
4827 EVT VT = N->getValueType(0);
4829 assert((Opcode == ISD::SIGN_EXTEND || Opcode == ISD::ZERO_EXTEND ||
4830 Opcode == ISD::ANY_EXTEND) && "Expected EXTEND dag node in input!");
4832 // fold (sext c1) -> c1
4833 // fold (zext c1) -> c1
4834 // fold (aext c1) -> c1
4835 if (isa<ConstantSDNode>(N0))
4836 return DAG.getNode(Opcode, SDLoc(N), VT, N0).getNode();
4838 // fold (sext (build_vector AllConstants) -> (build_vector AllConstants)
4839 // fold (zext (build_vector AllConstants) -> (build_vector AllConstants)
4840 // fold (aext (build_vector AllConstants) -> (build_vector AllConstants)
4841 EVT SVT = VT.getScalarType();
4842 if (!(VT.isVector() &&
4843 (!LegalTypes || (!LegalOperations && TLI.isTypeLegal(SVT))) &&
4844 ISD::isBuildVectorOfConstantSDNodes(N0.getNode())))
4847 // We can fold this node into a build_vector.
4848 unsigned VTBits = SVT.getSizeInBits();
4849 unsigned EVTBits = N0->getValueType(0).getScalarType().getSizeInBits();
4850 unsigned ShAmt = VTBits - EVTBits;
4851 SmallVector<SDValue, 8> Elts;
4852 unsigned NumElts = N0->getNumOperands();
4855 for (unsigned i=0; i != NumElts; ++i) {
4856 SDValue Op = N0->getOperand(i);
4857 if (Op->getOpcode() == ISD::UNDEF) {
4858 Elts.push_back(DAG.getUNDEF(SVT));
4862 ConstantSDNode *CurrentND = cast<ConstantSDNode>(Op);
4863 const APInt &C = APInt(VTBits, CurrentND->getAPIntValue().getZExtValue());
4864 if (Opcode == ISD::SIGN_EXTEND)
4865 Elts.push_back(DAG.getConstant(C.shl(ShAmt).ashr(ShAmt).getZExtValue(),
4868 Elts.push_back(DAG.getConstant(C.shl(ShAmt).lshr(ShAmt).getZExtValue(),
4872 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Elts).getNode();
4875 // ExtendUsesToFormExtLoad - Trying to extend uses of a load to enable this:
4876 // "fold ({s|z|a}ext (load x)) -> ({s|z|a}ext (truncate ({s|z|a}extload x)))"
4877 // transformation. Returns true if extension are possible and the above
4878 // mentioned transformation is profitable.
4879 static bool ExtendUsesToFormExtLoad(SDNode *N, SDValue N0,
4881 SmallVectorImpl<SDNode *> &ExtendNodes,
4882 const TargetLowering &TLI) {
4883 bool HasCopyToRegUses = false;
4884 bool isTruncFree = TLI.isTruncateFree(N->getValueType(0), N0.getValueType());
4885 for (SDNode::use_iterator UI = N0.getNode()->use_begin(),
4886 UE = N0.getNode()->use_end();
4891 if (UI.getUse().getResNo() != N0.getResNo())
4893 // FIXME: Only extend SETCC N, N and SETCC N, c for now.
4894 if (ExtOpc != ISD::ANY_EXTEND && User->getOpcode() == ISD::SETCC) {
4895 ISD::CondCode CC = cast<CondCodeSDNode>(User->getOperand(2))->get();
4896 if (ExtOpc == ISD::ZERO_EXTEND && ISD::isSignedIntSetCC(CC))
4897 // Sign bits will be lost after a zext.
4900 for (unsigned i = 0; i != 2; ++i) {
4901 SDValue UseOp = User->getOperand(i);
4904 if (!isa<ConstantSDNode>(UseOp))
4909 ExtendNodes.push_back(User);
4912 // If truncates aren't free and there are users we can't
4913 // extend, it isn't worthwhile.
4916 // Remember if this value is live-out.
4917 if (User->getOpcode() == ISD::CopyToReg)
4918 HasCopyToRegUses = true;
4921 if (HasCopyToRegUses) {
4922 bool BothLiveOut = false;
4923 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
4925 SDUse &Use = UI.getUse();
4926 if (Use.getResNo() == 0 && Use.getUser()->getOpcode() == ISD::CopyToReg) {
4932 // Both unextended and extended values are live out. There had better be
4933 // a good reason for the transformation.
4934 return ExtendNodes.size();
4939 void DAGCombiner::ExtendSetCCUses(const SmallVectorImpl<SDNode *> &SetCCs,
4940 SDValue Trunc, SDValue ExtLoad, SDLoc DL,
4941 ISD::NodeType ExtType) {
4942 // Extend SetCC uses if necessary.
4943 for (unsigned i = 0, e = SetCCs.size(); i != e; ++i) {
4944 SDNode *SetCC = SetCCs[i];
4945 SmallVector<SDValue, 4> Ops;
4947 for (unsigned j = 0; j != 2; ++j) {
4948 SDValue SOp = SetCC->getOperand(j);
4950 Ops.push_back(ExtLoad);
4952 Ops.push_back(DAG.getNode(ExtType, DL, ExtLoad->getValueType(0), SOp));
4955 Ops.push_back(SetCC->getOperand(2));
4956 CombineTo(SetCC, DAG.getNode(ISD::SETCC, DL, SetCC->getValueType(0), Ops));
4960 SDValue DAGCombiner::visitSIGN_EXTEND(SDNode *N) {
4961 SDValue N0 = N->getOperand(0);
4962 EVT VT = N->getValueType(0);
4964 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
4966 return SDValue(Res, 0);
4968 // fold (sext (sext x)) -> (sext x)
4969 // fold (sext (aext x)) -> (sext x)
4970 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
4971 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT,
4974 if (N0.getOpcode() == ISD::TRUNCATE) {
4975 // fold (sext (truncate (load x))) -> (sext (smaller load x))
4976 // fold (sext (truncate (srl (load x), c))) -> (sext (smaller load (x+c/n)))
4977 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
4978 if (NarrowLoad.getNode()) {
4979 SDNode* oye = N0.getNode()->getOperand(0).getNode();
4980 if (NarrowLoad.getNode() != N0.getNode()) {
4981 CombineTo(N0.getNode(), NarrowLoad);
4982 // CombineTo deleted the truncate, if needed, but not what's under it.
4985 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4988 // See if the value being truncated is already sign extended. If so, just
4989 // eliminate the trunc/sext pair.
4990 SDValue Op = N0.getOperand(0);
4991 unsigned OpBits = Op.getValueType().getScalarType().getSizeInBits();
4992 unsigned MidBits = N0.getValueType().getScalarType().getSizeInBits();
4993 unsigned DestBits = VT.getScalarType().getSizeInBits();
4994 unsigned NumSignBits = DAG.ComputeNumSignBits(Op);
4996 if (OpBits == DestBits) {
4997 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
4998 // bits, it is already ready.
4999 if (NumSignBits > DestBits-MidBits)
5001 } else if (OpBits < DestBits) {
5002 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
5003 // bits, just sext from i32.
5004 if (NumSignBits > OpBits-MidBits)
5005 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, Op);
5007 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
5008 // bits, just truncate to i32.
5009 if (NumSignBits > OpBits-MidBits)
5010 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
5013 // fold (sext (truncate x)) -> (sextinreg x).
5014 if (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG,
5015 N0.getValueType())) {
5016 if (OpBits < DestBits)
5017 Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N0), VT, Op);
5018 else if (OpBits > DestBits)
5019 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), VT, Op);
5020 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, Op,
5021 DAG.getValueType(N0.getValueType()));
5025 // fold (sext (load x)) -> (sext (truncate (sextload x)))
5026 // None of the supported targets knows how to perform load and sign extend
5027 // on vectors in one instruction. We only perform this transformation on
5029 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
5030 ISD::isUNINDEXEDLoad(N0.getNode()) &&
5031 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5032 TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()))) {
5033 bool DoXform = true;
5034 SmallVector<SDNode*, 4> SetCCs;
5035 if (!N0.hasOneUse())
5036 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::SIGN_EXTEND, SetCCs, TLI);
5038 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5039 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5041 LN0->getBasePtr(), N0.getValueType(),
5042 LN0->getMemOperand());
5043 CombineTo(N, ExtLoad);
5044 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5045 N0.getValueType(), ExtLoad);
5046 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
5047 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5049 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5053 // fold (sext (sextload x)) -> (sext (truncate (sextload x)))
5054 // fold (sext ( extload x)) -> (sext (truncate (sextload x)))
5055 if ((ISD::isSEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
5056 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
5057 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5058 EVT MemVT = LN0->getMemoryVT();
5059 if ((!LegalOperations && !LN0->isVolatile()) ||
5060 TLI.isLoadExtLegal(ISD::SEXTLOAD, MemVT)) {
5061 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5063 LN0->getBasePtr(), MemVT,
5064 LN0->getMemOperand());
5065 CombineTo(N, ExtLoad);
5066 CombineTo(N0.getNode(),
5067 DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5068 N0.getValueType(), ExtLoad),
5069 ExtLoad.getValue(1));
5070 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5074 // fold (sext (and/or/xor (load x), cst)) ->
5075 // (and/or/xor (sextload x), (sext cst))
5076 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
5077 N0.getOpcode() == ISD::XOR) &&
5078 isa<LoadSDNode>(N0.getOperand(0)) &&
5079 N0.getOperand(1).getOpcode() == ISD::Constant &&
5080 TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()) &&
5081 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
5082 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
5083 if (LN0->getExtensionType() != ISD::ZEXTLOAD && LN0->isUnindexed()) {
5084 bool DoXform = true;
5085 SmallVector<SDNode*, 4> SetCCs;
5086 if (!N0.hasOneUse())
5087 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::SIGN_EXTEND,
5090 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(LN0), VT,
5091 LN0->getChain(), LN0->getBasePtr(),
5093 LN0->getMemOperand());
5094 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5095 Mask = Mask.sext(VT.getSizeInBits());
5096 SDValue And = DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
5097 ExtLoad, DAG.getConstant(Mask, VT));
5098 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
5099 SDLoc(N0.getOperand(0)),
5100 N0.getOperand(0).getValueType(), ExtLoad);
5102 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
5103 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5105 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5110 if (N0.getOpcode() == ISD::SETCC) {
5111 EVT N0VT = N0.getOperand(0).getValueType();
5112 // sext(setcc) -> sext_in_reg(vsetcc) for vectors.
5113 // Only do this before legalize for now.
5114 if (VT.isVector() && !LegalOperations &&
5115 TLI.getBooleanContents(N0VT) ==
5116 TargetLowering::ZeroOrNegativeOneBooleanContent) {
5117 // On some architectures (such as SSE/NEON/etc) the SETCC result type is
5118 // of the same size as the compared operands. Only optimize sext(setcc())
5119 // if this is the case.
5120 EVT SVT = getSetCCResultType(N0VT);
5122 // We know that the # elements of the results is the same as the
5123 // # elements of the compare (and the # elements of the compare result
5124 // for that matter). Check to see that they are the same size. If so,
5125 // we know that the element size of the sext'd result matches the
5126 // element size of the compare operands.
5127 if (VT.getSizeInBits() == SVT.getSizeInBits())
5128 return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
5130 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5132 // If the desired elements are smaller or larger than the source
5133 // elements we can use a matching integer vector type and then
5134 // truncate/sign extend
5135 EVT MatchingVectorType = N0VT.changeVectorElementTypeToInteger();
5136 if (SVT == MatchingVectorType) {
5137 SDValue VsetCC = DAG.getSetCC(SDLoc(N), MatchingVectorType,
5138 N0.getOperand(0), N0.getOperand(1),
5139 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5140 return DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT);
5144 // sext(setcc x, y, cc) -> (select (setcc x, y, cc), -1, 0)
5145 unsigned ElementWidth = VT.getScalarType().getSizeInBits();
5147 DAG.getConstant(APInt::getAllOnesValue(ElementWidth), VT);
5149 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
5150 NegOne, DAG.getConstant(0, VT),
5151 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
5152 if (SCC.getNode()) return SCC;
5154 if (!VT.isVector()) {
5155 EVT SetCCVT = getSetCCResultType(N0.getOperand(0).getValueType());
5156 if (!LegalOperations || TLI.isOperationLegal(ISD::SETCC, SetCCVT)) {
5158 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
5159 SDValue SetCC = DAG.getSetCC(DL,
5161 N0.getOperand(0), N0.getOperand(1), CC);
5162 EVT SelectVT = getSetCCResultType(VT);
5163 return DAG.getSelect(DL, VT,
5164 DAG.getSExtOrTrunc(SetCC, DL, SelectVT),
5165 NegOne, DAG.getConstant(0, VT));
5171 // fold (sext x) -> (zext x) if the sign bit is known zero.
5172 if ((!LegalOperations || TLI.isOperationLegal(ISD::ZERO_EXTEND, VT)) &&
5173 DAG.SignBitIsZero(N0))
5174 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, N0);
5179 // isTruncateOf - If N is a truncate of some other value, return true, record
5180 // the value being truncated in Op and which of Op's bits are zero in KnownZero.
5181 // This function computes KnownZero to avoid a duplicated call to
5182 // computeKnownBits in the caller.
5183 static bool isTruncateOf(SelectionDAG &DAG, SDValue N, SDValue &Op,
5186 if (N->getOpcode() == ISD::TRUNCATE) {
5187 Op = N->getOperand(0);
5188 DAG.computeKnownBits(Op, KnownZero, KnownOne);
5192 if (N->getOpcode() != ISD::SETCC || N->getValueType(0) != MVT::i1 ||
5193 cast<CondCodeSDNode>(N->getOperand(2))->get() != ISD::SETNE)
5196 SDValue Op0 = N->getOperand(0);
5197 SDValue Op1 = N->getOperand(1);
5198 assert(Op0.getValueType() == Op1.getValueType());
5200 ConstantSDNode *COp0 = dyn_cast<ConstantSDNode>(Op0);
5201 ConstantSDNode *COp1 = dyn_cast<ConstantSDNode>(Op1);
5202 if (COp0 && COp0->isNullValue())
5204 else if (COp1 && COp1->isNullValue())
5209 DAG.computeKnownBits(Op, KnownZero, KnownOne);
5211 if (!(KnownZero | APInt(Op.getValueSizeInBits(), 1)).isAllOnesValue())
5217 SDValue DAGCombiner::visitZERO_EXTEND(SDNode *N) {
5218 SDValue N0 = N->getOperand(0);
5219 EVT VT = N->getValueType(0);
5221 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
5223 return SDValue(Res, 0);
5225 // fold (zext (zext x)) -> (zext x)
5226 // fold (zext (aext x)) -> (zext x)
5227 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
5228 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT,
5231 // fold (zext (truncate x)) -> (zext x) or
5232 // (zext (truncate x)) -> (truncate x)
5233 // This is valid when the truncated bits of x are already zero.
5234 // FIXME: We should extend this to work for vectors too.
5237 if (!VT.isVector() && isTruncateOf(DAG, N0, Op, KnownZero)) {
5238 APInt TruncatedBits =
5239 (Op.getValueSizeInBits() == N0.getValueSizeInBits()) ?
5240 APInt(Op.getValueSizeInBits(), 0) :
5241 APInt::getBitsSet(Op.getValueSizeInBits(),
5242 N0.getValueSizeInBits(),
5243 std::min(Op.getValueSizeInBits(),
5244 VT.getSizeInBits()));
5245 if (TruncatedBits == (KnownZero & TruncatedBits)) {
5246 if (VT.bitsGT(Op.getValueType()))
5247 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, Op);
5248 if (VT.bitsLT(Op.getValueType()))
5249 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
5255 // fold (zext (truncate (load x))) -> (zext (smaller load x))
5256 // fold (zext (truncate (srl (load x), c))) -> (zext (small load (x+c/n)))
5257 if (N0.getOpcode() == ISD::TRUNCATE) {
5258 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5259 if (NarrowLoad.getNode()) {
5260 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5261 if (NarrowLoad.getNode() != N0.getNode()) {
5262 CombineTo(N0.getNode(), NarrowLoad);
5263 // CombineTo deleted the truncate, if needed, but not what's under it.
5266 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5270 // fold (zext (truncate x)) -> (and x, mask)
5271 if (N0.getOpcode() == ISD::TRUNCATE &&
5272 (!LegalOperations || TLI.isOperationLegal(ISD::AND, VT))) {
5274 // fold (zext (truncate (load x))) -> (zext (smaller load x))
5275 // fold (zext (truncate (srl (load x), c))) -> (zext (smaller load (x+c/n)))
5276 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5277 if (NarrowLoad.getNode()) {
5278 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5279 if (NarrowLoad.getNode() != N0.getNode()) {
5280 CombineTo(N0.getNode(), NarrowLoad);
5281 // CombineTo deleted the truncate, if needed, but not what's under it.
5284 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5287 SDValue Op = N0.getOperand(0);
5288 if (Op.getValueType().bitsLT(VT)) {
5289 Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, Op);
5290 AddToWorklist(Op.getNode());
5291 } else if (Op.getValueType().bitsGT(VT)) {
5292 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
5293 AddToWorklist(Op.getNode());
5295 return DAG.getZeroExtendInReg(Op, SDLoc(N),
5296 N0.getValueType().getScalarType());
5299 // Fold (zext (and (trunc x), cst)) -> (and x, cst),
5300 // if either of the casts is not free.
5301 if (N0.getOpcode() == ISD::AND &&
5302 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
5303 N0.getOperand(1).getOpcode() == ISD::Constant &&
5304 (!TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
5305 N0.getValueType()) ||
5306 !TLI.isZExtFree(N0.getValueType(), VT))) {
5307 SDValue X = N0.getOperand(0).getOperand(0);
5308 if (X.getValueType().bitsLT(VT)) {
5309 X = DAG.getNode(ISD::ANY_EXTEND, SDLoc(X), VT, X);
5310 } else if (X.getValueType().bitsGT(VT)) {
5311 X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X);
5313 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5314 Mask = Mask.zext(VT.getSizeInBits());
5315 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5316 X, DAG.getConstant(Mask, VT));
5319 // fold (zext (load x)) -> (zext (truncate (zextload x)))
5320 // None of the supported targets knows how to perform load and vector_zext
5321 // on vectors in one instruction. We only perform this transformation on
5323 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
5324 ISD::isUNINDEXEDLoad(N0.getNode()) &&
5325 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5326 TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()))) {
5327 bool DoXform = true;
5328 SmallVector<SDNode*, 4> SetCCs;
5329 if (!N0.hasOneUse())
5330 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ZERO_EXTEND, SetCCs, TLI);
5332 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5333 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT,
5335 LN0->getBasePtr(), N0.getValueType(),
5336 LN0->getMemOperand());
5337 CombineTo(N, ExtLoad);
5338 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5339 N0.getValueType(), ExtLoad);
5340 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
5342 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5344 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5348 // fold (zext (and/or/xor (load x), cst)) ->
5349 // (and/or/xor (zextload x), (zext cst))
5350 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
5351 N0.getOpcode() == ISD::XOR) &&
5352 isa<LoadSDNode>(N0.getOperand(0)) &&
5353 N0.getOperand(1).getOpcode() == ISD::Constant &&
5354 TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()) &&
5355 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
5356 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
5357 if (LN0->getExtensionType() != ISD::SEXTLOAD && LN0->isUnindexed()) {
5358 bool DoXform = true;
5359 SmallVector<SDNode*, 4> SetCCs;
5360 if (!N0.hasOneUse())
5361 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::ZERO_EXTEND,
5364 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), VT,
5365 LN0->getChain(), LN0->getBasePtr(),
5367 LN0->getMemOperand());
5368 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5369 Mask = Mask.zext(VT.getSizeInBits());
5370 SDValue And = DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
5371 ExtLoad, DAG.getConstant(Mask, VT));
5372 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
5373 SDLoc(N0.getOperand(0)),
5374 N0.getOperand(0).getValueType(), ExtLoad);
5376 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
5377 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5379 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5384 // fold (zext (zextload x)) -> (zext (truncate (zextload x)))
5385 // fold (zext ( extload x)) -> (zext (truncate (zextload x)))
5386 if ((ISD::isZEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
5387 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
5388 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5389 EVT MemVT = LN0->getMemoryVT();
5390 if ((!LegalOperations && !LN0->isVolatile()) ||
5391 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT)) {
5392 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT,
5394 LN0->getBasePtr(), MemVT,
5395 LN0->getMemOperand());
5396 CombineTo(N, ExtLoad);
5397 CombineTo(N0.getNode(),
5398 DAG.getNode(ISD::TRUNCATE, SDLoc(N0), N0.getValueType(),
5400 ExtLoad.getValue(1));
5401 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5405 if (N0.getOpcode() == ISD::SETCC) {
5406 if (!LegalOperations && VT.isVector() &&
5407 N0.getValueType().getVectorElementType() == MVT::i1) {
5408 EVT N0VT = N0.getOperand(0).getValueType();
5409 if (getSetCCResultType(N0VT) == N0.getValueType())
5412 // zext(setcc) -> (and (vsetcc), (1, 1, ...) for vectors.
5413 // Only do this before legalize for now.
5414 EVT EltVT = VT.getVectorElementType();
5415 SmallVector<SDValue,8> OneOps(VT.getVectorNumElements(),
5416 DAG.getConstant(1, EltVT));
5417 if (VT.getSizeInBits() == N0VT.getSizeInBits())
5418 // We know that the # elements of the results is the same as the
5419 // # elements of the compare (and the # elements of the compare result
5420 // for that matter). Check to see that they are the same size. If so,
5421 // we know that the element size of the sext'd result matches the
5422 // element size of the compare operands.
5423 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5424 DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
5426 cast<CondCodeSDNode>(N0.getOperand(2))->get()),
5427 DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT,
5430 // If the desired elements are smaller or larger than the source
5431 // elements we can use a matching integer vector type and then
5432 // truncate/sign extend
5433 EVT MatchingElementType =
5434 EVT::getIntegerVT(*DAG.getContext(),
5435 N0VT.getScalarType().getSizeInBits());
5436 EVT MatchingVectorType =
5437 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
5438 N0VT.getVectorNumElements());
5440 DAG.getSetCC(SDLoc(N), MatchingVectorType, N0.getOperand(0),
5442 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5443 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5444 DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT),
5445 DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, OneOps));
5448 // zext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
5450 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
5451 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
5452 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
5453 if (SCC.getNode()) return SCC;
5456 // (zext (shl (zext x), cst)) -> (shl (zext x), cst)
5457 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL) &&
5458 isa<ConstantSDNode>(N0.getOperand(1)) &&
5459 N0.getOperand(0).getOpcode() == ISD::ZERO_EXTEND &&
5461 SDValue ShAmt = N0.getOperand(1);
5462 unsigned ShAmtVal = cast<ConstantSDNode>(ShAmt)->getZExtValue();
5463 if (N0.getOpcode() == ISD::SHL) {
5464 SDValue InnerZExt = N0.getOperand(0);
5465 // If the original shl may be shifting out bits, do not perform this
5467 unsigned KnownZeroBits = InnerZExt.getValueType().getSizeInBits() -
5468 InnerZExt.getOperand(0).getValueType().getSizeInBits();
5469 if (ShAmtVal > KnownZeroBits)
5475 // Ensure that the shift amount is wide enough for the shifted value.
5476 if (VT.getSizeInBits() >= 256)
5477 ShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, ShAmt);
5479 return DAG.getNode(N0.getOpcode(), DL, VT,
5480 DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0)),
5487 SDValue DAGCombiner::visitANY_EXTEND(SDNode *N) {
5488 SDValue N0 = N->getOperand(0);
5489 EVT VT = N->getValueType(0);
5491 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
5493 return SDValue(Res, 0);
5495 // fold (aext (aext x)) -> (aext x)
5496 // fold (aext (zext x)) -> (zext x)
5497 // fold (aext (sext x)) -> (sext x)
5498 if (N0.getOpcode() == ISD::ANY_EXTEND ||
5499 N0.getOpcode() == ISD::ZERO_EXTEND ||
5500 N0.getOpcode() == ISD::SIGN_EXTEND)
5501 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, N0.getOperand(0));
5503 // fold (aext (truncate (load x))) -> (aext (smaller load x))
5504 // fold (aext (truncate (srl (load x), c))) -> (aext (small load (x+c/n)))
5505 if (N0.getOpcode() == ISD::TRUNCATE) {
5506 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5507 if (NarrowLoad.getNode()) {
5508 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5509 if (NarrowLoad.getNode() != N0.getNode()) {
5510 CombineTo(N0.getNode(), NarrowLoad);
5511 // CombineTo deleted the truncate, if needed, but not what's under it.
5514 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5518 // fold (aext (truncate x))
5519 if (N0.getOpcode() == ISD::TRUNCATE) {
5520 SDValue TruncOp = N0.getOperand(0);
5521 if (TruncOp.getValueType() == VT)
5522 return TruncOp; // x iff x size == zext size.
5523 if (TruncOp.getValueType().bitsGT(VT))
5524 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, TruncOp);
5525 return DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, TruncOp);
5528 // Fold (aext (and (trunc x), cst)) -> (and x, cst)
5529 // if the trunc is not free.
5530 if (N0.getOpcode() == ISD::AND &&
5531 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
5532 N0.getOperand(1).getOpcode() == ISD::Constant &&
5533 !TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
5534 N0.getValueType())) {
5535 SDValue X = N0.getOperand(0).getOperand(0);
5536 if (X.getValueType().bitsLT(VT)) {
5537 X = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, X);
5538 } else if (X.getValueType().bitsGT(VT)) {
5539 X = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, X);
5541 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5542 Mask = Mask.zext(VT.getSizeInBits());
5543 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5544 X, DAG.getConstant(Mask, VT));
5547 // fold (aext (load x)) -> (aext (truncate (extload x)))
5548 // None of the supported targets knows how to perform load and any_ext
5549 // on vectors in one instruction. We only perform this transformation on
5551 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
5552 ISD::isUNINDEXEDLoad(N0.getNode()) &&
5553 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType())) {
5554 bool DoXform = true;
5555 SmallVector<SDNode*, 4> SetCCs;
5556 if (!N0.hasOneUse())
5557 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ANY_EXTEND, SetCCs, TLI);
5559 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5560 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT,
5562 LN0->getBasePtr(), N0.getValueType(),
5563 LN0->getMemOperand());
5564 CombineTo(N, ExtLoad);
5565 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5566 N0.getValueType(), ExtLoad);
5567 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
5568 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5570 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5574 // fold (aext (zextload x)) -> (aext (truncate (zextload x)))
5575 // fold (aext (sextload x)) -> (aext (truncate (sextload x)))
5576 // fold (aext ( extload x)) -> (aext (truncate (extload x)))
5577 if (N0.getOpcode() == ISD::LOAD &&
5578 !ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
5580 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5581 ISD::LoadExtType ExtType = LN0->getExtensionType();
5582 EVT MemVT = LN0->getMemoryVT();
5583 if (!LegalOperations || TLI.isLoadExtLegal(ExtType, MemVT)) {
5584 SDValue ExtLoad = DAG.getExtLoad(ExtType, SDLoc(N),
5585 VT, LN0->getChain(), LN0->getBasePtr(),
5586 MemVT, LN0->getMemOperand());
5587 CombineTo(N, ExtLoad);
5588 CombineTo(N0.getNode(),
5589 DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5590 N0.getValueType(), ExtLoad),
5591 ExtLoad.getValue(1));
5592 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5596 if (N0.getOpcode() == ISD::SETCC) {
5598 // aext(setcc) -> vsetcc
5599 // aext(setcc) -> truncate(vsetcc)
5600 // aext(setcc) -> aext(vsetcc)
5601 // Only do this before legalize for now.
5602 if (VT.isVector() && !LegalOperations) {
5603 EVT N0VT = N0.getOperand(0).getValueType();
5604 // We know that the # elements of the results is the same as the
5605 // # elements of the compare (and the # elements of the compare result
5606 // for that matter). Check to see that they are the same size. If so,
5607 // we know that the element size of the sext'd result matches the
5608 // element size of the compare operands.
5609 if (VT.getSizeInBits() == N0VT.getSizeInBits())
5610 return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
5612 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5613 // If the desired elements are smaller or larger than the source
5614 // elements we can use a matching integer vector type and then
5615 // truncate/any extend
5617 EVT MatchingVectorType = N0VT.changeVectorElementTypeToInteger();
5619 DAG.getSetCC(SDLoc(N), MatchingVectorType, N0.getOperand(0),
5621 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5622 return DAG.getAnyExtOrTrunc(VsetCC, SDLoc(N), VT);
5626 // aext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
5628 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
5629 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
5630 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
5638 /// GetDemandedBits - See if the specified operand can be simplified with the
5639 /// knowledge that only the bits specified by Mask are used. If so, return the
5640 /// simpler operand, otherwise return a null SDValue.
5641 SDValue DAGCombiner::GetDemandedBits(SDValue V, const APInt &Mask) {
5642 switch (V.getOpcode()) {
5644 case ISD::Constant: {
5645 const ConstantSDNode *CV = cast<ConstantSDNode>(V.getNode());
5646 assert(CV && "Const value should be ConstSDNode.");
5647 const APInt &CVal = CV->getAPIntValue();
5648 APInt NewVal = CVal & Mask;
5650 return DAG.getConstant(NewVal, V.getValueType());
5655 // If the LHS or RHS don't contribute bits to the or, drop them.
5656 if (DAG.MaskedValueIsZero(V.getOperand(0), Mask))
5657 return V.getOperand(1);
5658 if (DAG.MaskedValueIsZero(V.getOperand(1), Mask))
5659 return V.getOperand(0);
5662 // Only look at single-use SRLs.
5663 if (!V.getNode()->hasOneUse())
5665 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(V.getOperand(1))) {
5666 // See if we can recursively simplify the LHS.
5667 unsigned Amt = RHSC->getZExtValue();
5669 // Watch out for shift count overflow though.
5670 if (Amt >= Mask.getBitWidth()) break;
5671 APInt NewMask = Mask << Amt;
5672 SDValue SimplifyLHS = GetDemandedBits(V.getOperand(0), NewMask);
5673 if (SimplifyLHS.getNode())
5674 return DAG.getNode(ISD::SRL, SDLoc(V), V.getValueType(),
5675 SimplifyLHS, V.getOperand(1));
5681 /// ReduceLoadWidth - If the result of a wider load is shifted to right of N
5682 /// bits and then truncated to a narrower type and where N is a multiple
5683 /// of number of bits of the narrower type, transform it to a narrower load
5684 /// from address + N / num of bits of new type. If the result is to be
5685 /// extended, also fold the extension to form a extending load.
5686 SDValue DAGCombiner::ReduceLoadWidth(SDNode *N) {
5687 unsigned Opc = N->getOpcode();
5689 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
5690 SDValue N0 = N->getOperand(0);
5691 EVT VT = N->getValueType(0);
5694 // This transformation isn't valid for vector loads.
5698 // Special case: SIGN_EXTEND_INREG is basically truncating to ExtVT then
5700 if (Opc == ISD::SIGN_EXTEND_INREG) {
5701 ExtType = ISD::SEXTLOAD;
5702 ExtVT = cast<VTSDNode>(N->getOperand(1))->getVT();
5703 } else if (Opc == ISD::SRL) {
5704 // Another special-case: SRL is basically zero-extending a narrower value.
5705 ExtType = ISD::ZEXTLOAD;
5707 ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1));
5708 if (!N01) return SDValue();
5709 ExtVT = EVT::getIntegerVT(*DAG.getContext(),
5710 VT.getSizeInBits() - N01->getZExtValue());
5712 if (LegalOperations && !TLI.isLoadExtLegal(ExtType, ExtVT))
5715 unsigned EVTBits = ExtVT.getSizeInBits();
5717 // Do not generate loads of non-round integer types since these can
5718 // be expensive (and would be wrong if the type is not byte sized).
5719 if (!ExtVT.isRound())
5723 if (N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
5724 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
5725 ShAmt = N01->getZExtValue();
5726 // Is the shift amount a multiple of size of VT?
5727 if ((ShAmt & (EVTBits-1)) == 0) {
5728 N0 = N0.getOperand(0);
5729 // Is the load width a multiple of size of VT?
5730 if ((N0.getValueType().getSizeInBits() & (EVTBits-1)) != 0)
5734 // At this point, we must have a load or else we can't do the transform.
5735 if (!isa<LoadSDNode>(N0)) return SDValue();
5737 // Because a SRL must be assumed to *need* to zero-extend the high bits
5738 // (as opposed to anyext the high bits), we can't combine the zextload
5739 // lowering of SRL and an sextload.
5740 if (cast<LoadSDNode>(N0)->getExtensionType() == ISD::SEXTLOAD)
5743 // If the shift amount is larger than the input type then we're not
5744 // accessing any of the loaded bytes. If the load was a zextload/extload
5745 // then the result of the shift+trunc is zero/undef (handled elsewhere).
5746 if (ShAmt >= cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits())
5751 // If the load is shifted left (and the result isn't shifted back right),
5752 // we can fold the truncate through the shift.
5753 unsigned ShLeftAmt = 0;
5754 if (ShAmt == 0 && N0.getOpcode() == ISD::SHL && N0.hasOneUse() &&
5755 ExtVT == VT && TLI.isNarrowingProfitable(N0.getValueType(), VT)) {
5756 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
5757 ShLeftAmt = N01->getZExtValue();
5758 N0 = N0.getOperand(0);
5762 // If we haven't found a load, we can't narrow it. Don't transform one with
5763 // multiple uses, this would require adding a new load.
5764 if (!isa<LoadSDNode>(N0) || !N0.hasOneUse())
5767 // Don't change the width of a volatile load.
5768 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5769 if (LN0->isVolatile())
5772 // Verify that we are actually reducing a load width here.
5773 if (LN0->getMemoryVT().getSizeInBits() < EVTBits)
5776 // For the transform to be legal, the load must produce only two values
5777 // (the value loaded and the chain). Don't transform a pre-increment
5778 // load, for example, which produces an extra value. Otherwise the
5779 // transformation is not equivalent, and the downstream logic to replace
5780 // uses gets things wrong.
5781 if (LN0->getNumValues() > 2)
5784 // If the load that we're shrinking is an extload and we're not just
5785 // discarding the extension we can't simply shrink the load. Bail.
5786 // TODO: It would be possible to merge the extensions in some cases.
5787 if (LN0->getExtensionType() != ISD::NON_EXTLOAD &&
5788 LN0->getMemoryVT().getSizeInBits() < ExtVT.getSizeInBits() + ShAmt)
5791 EVT PtrType = N0.getOperand(1).getValueType();
5793 if (PtrType == MVT::Untyped || PtrType.isExtended())
5794 // It's not possible to generate a constant of extended or untyped type.
5797 // For big endian targets, we need to adjust the offset to the pointer to
5798 // load the correct bytes.
5799 if (TLI.isBigEndian()) {
5800 unsigned LVTStoreBits = LN0->getMemoryVT().getStoreSizeInBits();
5801 unsigned EVTStoreBits = ExtVT.getStoreSizeInBits();
5802 ShAmt = LVTStoreBits - EVTStoreBits - ShAmt;
5805 uint64_t PtrOff = ShAmt / 8;
5806 unsigned NewAlign = MinAlign(LN0->getAlignment(), PtrOff);
5807 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(LN0),
5808 PtrType, LN0->getBasePtr(),
5809 DAG.getConstant(PtrOff, PtrType));
5810 AddToWorklist(NewPtr.getNode());
5813 if (ExtType == ISD::NON_EXTLOAD)
5814 Load = DAG.getLoad(VT, SDLoc(N0), LN0->getChain(), NewPtr,
5815 LN0->getPointerInfo().getWithOffset(PtrOff),
5816 LN0->isVolatile(), LN0->isNonTemporal(),
5817 LN0->isInvariant(), NewAlign, LN0->getAAInfo());
5819 Load = DAG.getExtLoad(ExtType, SDLoc(N0), VT, LN0->getChain(),NewPtr,
5820 LN0->getPointerInfo().getWithOffset(PtrOff),
5821 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
5822 NewAlign, LN0->getAAInfo());
5824 // Replace the old load's chain with the new load's chain.
5825 WorklistRemover DeadNodes(*this);
5826 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1));
5828 // Shift the result left, if we've swallowed a left shift.
5829 SDValue Result = Load;
5830 if (ShLeftAmt != 0) {
5831 EVT ShImmTy = getShiftAmountTy(Result.getValueType());
5832 if (!isUIntN(ShImmTy.getSizeInBits(), ShLeftAmt))
5834 // If the shift amount is as large as the result size (but, presumably,
5835 // no larger than the source) then the useful bits of the result are
5836 // zero; we can't simply return the shortened shift, because the result
5837 // of that operation is undefined.
5838 if (ShLeftAmt >= VT.getSizeInBits())
5839 Result = DAG.getConstant(0, VT);
5841 Result = DAG.getNode(ISD::SHL, SDLoc(N0), VT,
5842 Result, DAG.getConstant(ShLeftAmt, ShImmTy));
5845 // Return the new loaded value.
5849 SDValue DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
5850 SDValue N0 = N->getOperand(0);
5851 SDValue N1 = N->getOperand(1);
5852 EVT VT = N->getValueType(0);
5853 EVT EVT = cast<VTSDNode>(N1)->getVT();
5854 unsigned VTBits = VT.getScalarType().getSizeInBits();
5855 unsigned EVTBits = EVT.getScalarType().getSizeInBits();
5857 // fold (sext_in_reg c1) -> c1
5858 if (isa<ConstantSDNode>(N0) || N0.getOpcode() == ISD::UNDEF)
5859 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, N0, N1);
5861 // If the input is already sign extended, just drop the extension.
5862 if (DAG.ComputeNumSignBits(N0) >= VTBits-EVTBits+1)
5865 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2
5866 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
5867 EVT.bitsLT(cast<VTSDNode>(N0.getOperand(1))->getVT()))
5868 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
5869 N0.getOperand(0), N1);
5871 // fold (sext_in_reg (sext x)) -> (sext x)
5872 // fold (sext_in_reg (aext x)) -> (sext x)
5873 // if x is small enough.
5874 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) {
5875 SDValue N00 = N0.getOperand(0);
5876 if (N00.getValueType().getScalarType().getSizeInBits() <= EVTBits &&
5877 (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND, VT)))
5878 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, N00, N1);
5881 // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is known zero.
5882 if (DAG.MaskedValueIsZero(N0, APInt::getBitsSet(VTBits, EVTBits-1, EVTBits)))
5883 return DAG.getZeroExtendInReg(N0, SDLoc(N), EVT);
5885 // fold operands of sext_in_reg based on knowledge that the top bits are not
5887 if (SimplifyDemandedBits(SDValue(N, 0)))
5888 return SDValue(N, 0);
5890 // fold (sext_in_reg (load x)) -> (smaller sextload x)
5891 // fold (sext_in_reg (srl (load x), c)) -> (smaller sextload (x+c/evtbits))
5892 SDValue NarrowLoad = ReduceLoadWidth(N);
5893 if (NarrowLoad.getNode())
5896 // fold (sext_in_reg (srl X, 24), i8) -> (sra X, 24)
5897 // fold (sext_in_reg (srl X, 23), i8) -> (sra X, 23) iff possible.
5898 // We already fold "(sext_in_reg (srl X, 25), i8) -> srl X, 25" above.
5899 if (N0.getOpcode() == ISD::SRL) {
5900 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
5901 if (ShAmt->getZExtValue()+EVTBits <= VTBits) {
5902 // We can turn this into an SRA iff the input to the SRL is already sign
5904 unsigned InSignBits = DAG.ComputeNumSignBits(N0.getOperand(0));
5905 if (VTBits-(ShAmt->getZExtValue()+EVTBits) < InSignBits)
5906 return DAG.getNode(ISD::SRA, SDLoc(N), VT,
5907 N0.getOperand(0), N0.getOperand(1));
5911 // fold (sext_inreg (extload x)) -> (sextload x)
5912 if (ISD::isEXTLoad(N0.getNode()) &&
5913 ISD::isUNINDEXEDLoad(N0.getNode()) &&
5914 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
5915 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5916 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) {
5917 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5918 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5920 LN0->getBasePtr(), EVT,
5921 LN0->getMemOperand());
5922 CombineTo(N, ExtLoad);
5923 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
5924 AddToWorklist(ExtLoad.getNode());
5925 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5927 // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use
5928 if (ISD::isZEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
5930 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
5931 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5932 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) {
5933 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5934 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5936 LN0->getBasePtr(), EVT,
5937 LN0->getMemOperand());
5938 CombineTo(N, ExtLoad);
5939 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
5940 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5943 // Form (sext_inreg (bswap >> 16)) or (sext_inreg (rotl (bswap) 16))
5944 if (EVTBits <= 16 && N0.getOpcode() == ISD::OR) {
5945 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
5946 N0.getOperand(1), false);
5947 if (BSwap.getNode())
5948 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
5952 // Fold a sext_inreg of a build_vector of ConstantSDNodes or undefs
5953 // into a build_vector.
5954 if (ISD::isBuildVectorOfConstantSDNodes(N0.getNode())) {
5955 SmallVector<SDValue, 8> Elts;
5956 unsigned NumElts = N0->getNumOperands();
5957 unsigned ShAmt = VTBits - EVTBits;
5959 for (unsigned i = 0; i != NumElts; ++i) {
5960 SDValue Op = N0->getOperand(i);
5961 if (Op->getOpcode() == ISD::UNDEF) {
5966 ConstantSDNode *CurrentND = cast<ConstantSDNode>(Op);
5967 const APInt &C = APInt(VTBits, CurrentND->getAPIntValue().getZExtValue());
5968 Elts.push_back(DAG.getConstant(C.shl(ShAmt).ashr(ShAmt).getZExtValue(),
5969 Op.getValueType()));
5972 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Elts);
5978 SDValue DAGCombiner::visitTRUNCATE(SDNode *N) {
5979 SDValue N0 = N->getOperand(0);
5980 EVT VT = N->getValueType(0);
5981 bool isLE = TLI.isLittleEndian();
5984 if (N0.getValueType() == N->getValueType(0))
5986 // fold (truncate c1) -> c1
5987 if (isa<ConstantSDNode>(N0))
5988 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0);
5989 // fold (truncate (truncate x)) -> (truncate x)
5990 if (N0.getOpcode() == ISD::TRUNCATE)
5991 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0));
5992 // fold (truncate (ext x)) -> (ext x) or (truncate x) or x
5993 if (N0.getOpcode() == ISD::ZERO_EXTEND ||
5994 N0.getOpcode() == ISD::SIGN_EXTEND ||
5995 N0.getOpcode() == ISD::ANY_EXTEND) {
5996 if (N0.getOperand(0).getValueType().bitsLT(VT))
5997 // if the source is smaller than the dest, we still need an extend
5998 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
6000 if (N0.getOperand(0).getValueType().bitsGT(VT))
6001 // if the source is larger than the dest, than we just need the truncate
6002 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0));
6003 // if the source and dest are the same type, we can drop both the extend
6004 // and the truncate.
6005 return N0.getOperand(0);
6008 // Fold extract-and-trunc into a narrow extract. For example:
6009 // i64 x = EXTRACT_VECTOR_ELT(v2i64 val, i32 1)
6010 // i32 y = TRUNCATE(i64 x)
6012 // v16i8 b = BITCAST (v2i64 val)
6013 // i8 x = EXTRACT_VECTOR_ELT(v16i8 b, i32 8)
6015 // Note: We only run this optimization after type legalization (which often
6016 // creates this pattern) and before operation legalization after which
6017 // we need to be more careful about the vector instructions that we generate.
6018 if (N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
6019 LegalTypes && !LegalOperations && N0->hasOneUse() && VT != MVT::i1) {
6021 EVT VecTy = N0.getOperand(0).getValueType();
6022 EVT ExTy = N0.getValueType();
6023 EVT TrTy = N->getValueType(0);
6025 unsigned NumElem = VecTy.getVectorNumElements();
6026 unsigned SizeRatio = ExTy.getSizeInBits()/TrTy.getSizeInBits();
6028 EVT NVT = EVT::getVectorVT(*DAG.getContext(), TrTy, SizeRatio * NumElem);
6029 assert(NVT.getSizeInBits() == VecTy.getSizeInBits() && "Invalid Size");
6031 SDValue EltNo = N0->getOperand(1);
6032 if (isa<ConstantSDNode>(EltNo) && isTypeLegal(NVT)) {
6033 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
6034 EVT IndexTy = TLI.getVectorIdxTy();
6035 int Index = isLE ? (Elt*SizeRatio) : (Elt*SizeRatio + (SizeRatio-1));
6037 SDValue V = DAG.getNode(ISD::BITCAST, SDLoc(N),
6038 NVT, N0.getOperand(0));
6040 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
6042 DAG.getConstant(Index, IndexTy));
6046 // trunc (select c, a, b) -> select c, (trunc a), (trunc b)
6047 if (N0.getOpcode() == ISD::SELECT) {
6048 EVT SrcVT = N0.getValueType();
6049 if ((!LegalOperations || TLI.isOperationLegal(ISD::SELECT, SrcVT)) &&
6050 TLI.isTruncateFree(SrcVT, VT)) {
6052 SDValue Cond = N0.getOperand(0);
6053 SDValue TruncOp0 = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(1));
6054 SDValue TruncOp1 = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(2));
6055 return DAG.getNode(ISD::SELECT, SDLoc(N), VT, Cond, TruncOp0, TruncOp1);
6059 // Fold a series of buildvector, bitcast, and truncate if possible.
6061 // (2xi32 trunc (bitcast ((4xi32)buildvector x, x, y, y) 2xi64)) to
6062 // (2xi32 (buildvector x, y)).
6063 if (Level == AfterLegalizeVectorOps && VT.isVector() &&
6064 N0.getOpcode() == ISD::BITCAST && N0.hasOneUse() &&
6065 N0.getOperand(0).getOpcode() == ISD::BUILD_VECTOR &&
6066 N0.getOperand(0).hasOneUse()) {
6068 SDValue BuildVect = N0.getOperand(0);
6069 EVT BuildVectEltTy = BuildVect.getValueType().getVectorElementType();
6070 EVT TruncVecEltTy = VT.getVectorElementType();
6072 // Check that the element types match.
6073 if (BuildVectEltTy == TruncVecEltTy) {
6074 // Now we only need to compute the offset of the truncated elements.
6075 unsigned BuildVecNumElts = BuildVect.getNumOperands();
6076 unsigned TruncVecNumElts = VT.getVectorNumElements();
6077 unsigned TruncEltOffset = BuildVecNumElts / TruncVecNumElts;
6079 assert((BuildVecNumElts % TruncVecNumElts) == 0 &&
6080 "Invalid number of elements");
6082 SmallVector<SDValue, 8> Opnds;
6083 for (unsigned i = 0, e = BuildVecNumElts; i != e; i += TruncEltOffset)
6084 Opnds.push_back(BuildVect.getOperand(i));
6086 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Opnds);
6090 // See if we can simplify the input to this truncate through knowledge that
6091 // only the low bits are being used.
6092 // For example "trunc (or (shl x, 8), y)" // -> trunc y
6093 // Currently we only perform this optimization on scalars because vectors
6094 // may have different active low bits.
6095 if (!VT.isVector()) {
6097 GetDemandedBits(N0, APInt::getLowBitsSet(N0.getValueSizeInBits(),
6098 VT.getSizeInBits()));
6099 if (Shorter.getNode())
6100 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Shorter);
6102 // fold (truncate (load x)) -> (smaller load x)
6103 // fold (truncate (srl (load x), c)) -> (smaller load (x+c/evtbits))
6104 if (!LegalTypes || TLI.isTypeDesirableForOp(N0.getOpcode(), VT)) {
6105 SDValue Reduced = ReduceLoadWidth(N);
6106 if (Reduced.getNode())
6108 // Handle the case where the load remains an extending load even
6109 // after truncation.
6110 if (N0.hasOneUse() && ISD::isUNINDEXEDLoad(N0.getNode())) {
6111 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6112 if (!LN0->isVolatile() &&
6113 LN0->getMemoryVT().getStoreSizeInBits() < VT.getSizeInBits()) {
6114 SDValue NewLoad = DAG.getExtLoad(LN0->getExtensionType(), SDLoc(LN0),
6115 VT, LN0->getChain(), LN0->getBasePtr(),
6117 LN0->getMemOperand());
6118 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLoad.getValue(1));
6123 // fold (trunc (concat ... x ...)) -> (concat ..., (trunc x), ...)),
6124 // where ... are all 'undef'.
6125 if (N0.getOpcode() == ISD::CONCAT_VECTORS && !LegalTypes) {
6126 SmallVector<EVT, 8> VTs;
6129 unsigned NumDefs = 0;
6131 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) {
6132 SDValue X = N0.getOperand(i);
6133 if (X.getOpcode() != ISD::UNDEF) {
6138 // Stop if more than one members are non-undef.
6141 VTs.push_back(EVT::getVectorVT(*DAG.getContext(),
6142 VT.getVectorElementType(),
6143 X.getValueType().getVectorNumElements()));
6147 return DAG.getUNDEF(VT);
6150 assert(V.getNode() && "The single defined operand is empty!");
6151 SmallVector<SDValue, 8> Opnds;
6152 for (unsigned i = 0, e = VTs.size(); i != e; ++i) {
6154 Opnds.push_back(DAG.getUNDEF(VTs[i]));
6157 SDValue NV = DAG.getNode(ISD::TRUNCATE, SDLoc(V), VTs[i], V);
6158 AddToWorklist(NV.getNode());
6159 Opnds.push_back(NV);
6161 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Opnds);
6165 // Simplify the operands using demanded-bits information.
6166 if (!VT.isVector() &&
6167 SimplifyDemandedBits(SDValue(N, 0)))
6168 return SDValue(N, 0);
6173 static SDNode *getBuildPairElt(SDNode *N, unsigned i) {
6174 SDValue Elt = N->getOperand(i);
6175 if (Elt.getOpcode() != ISD::MERGE_VALUES)
6176 return Elt.getNode();
6177 return Elt.getOperand(Elt.getResNo()).getNode();
6180 /// CombineConsecutiveLoads - build_pair (load, load) -> load
6181 /// if load locations are consecutive.
6182 SDValue DAGCombiner::CombineConsecutiveLoads(SDNode *N, EVT VT) {
6183 assert(N->getOpcode() == ISD::BUILD_PAIR);
6185 LoadSDNode *LD1 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 0));
6186 LoadSDNode *LD2 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 1));
6187 if (!LD1 || !LD2 || !ISD::isNON_EXTLoad(LD1) || !LD1->hasOneUse() ||
6188 LD1->getAddressSpace() != LD2->getAddressSpace())
6190 EVT LD1VT = LD1->getValueType(0);
6192 if (ISD::isNON_EXTLoad(LD2) &&
6194 // If both are volatile this would reduce the number of volatile loads.
6195 // If one is volatile it might be ok, but play conservative and bail out.
6196 !LD1->isVolatile() &&
6197 !LD2->isVolatile() &&
6198 DAG.isConsecutiveLoad(LD2, LD1, LD1VT.getSizeInBits()/8, 1)) {
6199 unsigned Align = LD1->getAlignment();
6200 unsigned NewAlign = TLI.getDataLayout()->
6201 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
6203 if (NewAlign <= Align &&
6204 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)))
6205 return DAG.getLoad(VT, SDLoc(N), LD1->getChain(),
6206 LD1->getBasePtr(), LD1->getPointerInfo(),
6207 false, false, false, Align);
6213 SDValue DAGCombiner::visitBITCAST(SDNode *N) {
6214 SDValue N0 = N->getOperand(0);
6215 EVT VT = N->getValueType(0);
6217 // If the input is a BUILD_VECTOR with all constant elements, fold this now.
6218 // Only do this before legalize, since afterward the target may be depending
6219 // on the bitconvert.
6220 // First check to see if this is all constant.
6222 N0.getOpcode() == ISD::BUILD_VECTOR && N0.getNode()->hasOneUse() &&
6224 bool isSimple = cast<BuildVectorSDNode>(N0)->isConstant();
6226 EVT DestEltVT = N->getValueType(0).getVectorElementType();
6227 assert(!DestEltVT.isVector() &&
6228 "Element type of vector ValueType must not be vector!");
6230 return ConstantFoldBITCASTofBUILD_VECTOR(N0.getNode(), DestEltVT);
6233 // If the input is a constant, let getNode fold it.
6234 if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) {
6235 SDValue Res = DAG.getNode(ISD::BITCAST, SDLoc(N), VT, N0);
6236 if (Res.getNode() != N) {
6237 if (!LegalOperations ||
6238 TLI.isOperationLegal(Res.getNode()->getOpcode(), VT))
6241 // Folding it resulted in an illegal node, and it's too late to
6242 // do that. Clean up the old node and forego the transformation.
6243 // Ideally this won't happen very often, because instcombine
6244 // and the earlier dagcombine runs (where illegal nodes are
6245 // permitted) should have folded most of them already.
6246 DAG.DeleteNode(Res.getNode());
6250 // (conv (conv x, t1), t2) -> (conv x, t2)
6251 if (N0.getOpcode() == ISD::BITCAST)
6252 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT,
6255 // fold (conv (load x)) -> (load (conv*)x)
6256 // If the resultant load doesn't need a higher alignment than the original!
6257 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
6258 // Do not change the width of a volatile load.
6259 !cast<LoadSDNode>(N0)->isVolatile() &&
6260 // Do not remove the cast if the types differ in endian layout.
6261 TLI.hasBigEndianPartOrdering(N0.getValueType()) ==
6262 TLI.hasBigEndianPartOrdering(VT) &&
6263 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)) &&
6264 TLI.isLoadBitCastBeneficial(N0.getValueType(), VT)) {
6265 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6266 unsigned Align = TLI.getDataLayout()->
6267 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
6268 unsigned OrigAlign = LN0->getAlignment();
6270 if (Align <= OrigAlign) {
6271 SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(),
6272 LN0->getBasePtr(), LN0->getPointerInfo(),
6273 LN0->isVolatile(), LN0->isNonTemporal(),
6274 LN0->isInvariant(), OrigAlign,
6277 CombineTo(N0.getNode(),
6278 DAG.getNode(ISD::BITCAST, SDLoc(N0),
6279 N0.getValueType(), Load),
6285 // fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit)
6286 // fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit))
6287 // This often reduces constant pool loads.
6288 if (((N0.getOpcode() == ISD::FNEG && !TLI.isFNegFree(N0.getValueType())) ||
6289 (N0.getOpcode() == ISD::FABS && !TLI.isFAbsFree(N0.getValueType()))) &&
6290 N0.getNode()->hasOneUse() && VT.isInteger() &&
6291 !VT.isVector() && !N0.getValueType().isVector()) {
6292 SDValue NewConv = DAG.getNode(ISD::BITCAST, SDLoc(N0), VT,
6294 AddToWorklist(NewConv.getNode());
6296 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
6297 if (N0.getOpcode() == ISD::FNEG)
6298 return DAG.getNode(ISD::XOR, SDLoc(N), VT,
6299 NewConv, DAG.getConstant(SignBit, VT));
6300 assert(N0.getOpcode() == ISD::FABS);
6301 return DAG.getNode(ISD::AND, SDLoc(N), VT,
6302 NewConv, DAG.getConstant(~SignBit, VT));
6305 // fold (bitconvert (fcopysign cst, x)) ->
6306 // (or (and (bitconvert x), sign), (and cst, (not sign)))
6307 // Note that we don't handle (copysign x, cst) because this can always be
6308 // folded to an fneg or fabs.
6309 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse() &&
6310 isa<ConstantFPSDNode>(N0.getOperand(0)) &&
6311 VT.isInteger() && !VT.isVector()) {
6312 unsigned OrigXWidth = N0.getOperand(1).getValueType().getSizeInBits();
6313 EVT IntXVT = EVT::getIntegerVT(*DAG.getContext(), OrigXWidth);
6314 if (isTypeLegal(IntXVT)) {
6315 SDValue X = DAG.getNode(ISD::BITCAST, SDLoc(N0),
6316 IntXVT, N0.getOperand(1));
6317 AddToWorklist(X.getNode());
6319 // If X has a different width than the result/lhs, sext it or truncate it.
6320 unsigned VTWidth = VT.getSizeInBits();
6321 if (OrigXWidth < VTWidth) {
6322 X = DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, X);
6323 AddToWorklist(X.getNode());
6324 } else if (OrigXWidth > VTWidth) {
6325 // To get the sign bit in the right place, we have to shift it right
6326 // before truncating.
6327 X = DAG.getNode(ISD::SRL, SDLoc(X),
6328 X.getValueType(), X,
6329 DAG.getConstant(OrigXWidth-VTWidth, X.getValueType()));
6330 AddToWorklist(X.getNode());
6331 X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X);
6332 AddToWorklist(X.getNode());
6335 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
6336 X = DAG.getNode(ISD::AND, SDLoc(X), VT,
6337 X, DAG.getConstant(SignBit, VT));
6338 AddToWorklist(X.getNode());
6340 SDValue Cst = DAG.getNode(ISD::BITCAST, SDLoc(N0),
6341 VT, N0.getOperand(0));
6342 Cst = DAG.getNode(ISD::AND, SDLoc(Cst), VT,
6343 Cst, DAG.getConstant(~SignBit, VT));
6344 AddToWorklist(Cst.getNode());
6346 return DAG.getNode(ISD::OR, SDLoc(N), VT, X, Cst);
6350 // bitconvert(build_pair(ld, ld)) -> ld iff load locations are consecutive.
6351 if (N0.getOpcode() == ISD::BUILD_PAIR) {
6352 SDValue CombineLD = CombineConsecutiveLoads(N0.getNode(), VT);
6353 if (CombineLD.getNode())
6360 SDValue DAGCombiner::visitBUILD_PAIR(SDNode *N) {
6361 EVT VT = N->getValueType(0);
6362 return CombineConsecutiveLoads(N, VT);
6365 /// ConstantFoldBITCASTofBUILD_VECTOR - We know that BV is a build_vector
6366 /// node with Constant, ConstantFP or Undef operands. DstEltVT indicates the
6367 /// destination element value type.
6368 SDValue DAGCombiner::
6369 ConstantFoldBITCASTofBUILD_VECTOR(SDNode *BV, EVT DstEltVT) {
6370 EVT SrcEltVT = BV->getValueType(0).getVectorElementType();
6372 // If this is already the right type, we're done.
6373 if (SrcEltVT == DstEltVT) return SDValue(BV, 0);
6375 unsigned SrcBitSize = SrcEltVT.getSizeInBits();
6376 unsigned DstBitSize = DstEltVT.getSizeInBits();
6378 // If this is a conversion of N elements of one type to N elements of another
6379 // type, convert each element. This handles FP<->INT cases.
6380 if (SrcBitSize == DstBitSize) {
6381 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
6382 BV->getValueType(0).getVectorNumElements());
6384 // Due to the FP element handling below calling this routine recursively,
6385 // we can end up with a scalar-to-vector node here.
6386 if (BV->getOpcode() == ISD::SCALAR_TO_VECTOR)
6387 return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(BV), VT,
6388 DAG.getNode(ISD::BITCAST, SDLoc(BV),
6389 DstEltVT, BV->getOperand(0)));
6391 SmallVector<SDValue, 8> Ops;
6392 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
6393 SDValue Op = BV->getOperand(i);
6394 // If the vector element type is not legal, the BUILD_VECTOR operands
6395 // are promoted and implicitly truncated. Make that explicit here.
6396 if (Op.getValueType() != SrcEltVT)
6397 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(BV), SrcEltVT, Op);
6398 Ops.push_back(DAG.getNode(ISD::BITCAST, SDLoc(BV),
6400 AddToWorklist(Ops.back().getNode());
6402 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
6405 // Otherwise, we're growing or shrinking the elements. To avoid having to
6406 // handle annoying details of growing/shrinking FP values, we convert them to
6408 if (SrcEltVT.isFloatingPoint()) {
6409 // Convert the input float vector to a int vector where the elements are the
6411 assert((SrcEltVT == MVT::f32 || SrcEltVT == MVT::f64) && "Unknown FP VT!");
6412 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), SrcEltVT.getSizeInBits());
6413 BV = ConstantFoldBITCASTofBUILD_VECTOR(BV, IntVT).getNode();
6417 // Now we know the input is an integer vector. If the output is a FP type,
6418 // convert to integer first, then to FP of the right size.
6419 if (DstEltVT.isFloatingPoint()) {
6420 assert((DstEltVT == MVT::f32 || DstEltVT == MVT::f64) && "Unknown FP VT!");
6421 EVT TmpVT = EVT::getIntegerVT(*DAG.getContext(), DstEltVT.getSizeInBits());
6422 SDNode *Tmp = ConstantFoldBITCASTofBUILD_VECTOR(BV, TmpVT).getNode();
6424 // Next, convert to FP elements of the same size.
6425 return ConstantFoldBITCASTofBUILD_VECTOR(Tmp, DstEltVT);
6428 // Okay, we know the src/dst types are both integers of differing types.
6429 // Handling growing first.
6430 assert(SrcEltVT.isInteger() && DstEltVT.isInteger());
6431 if (SrcBitSize < DstBitSize) {
6432 unsigned NumInputsPerOutput = DstBitSize/SrcBitSize;
6434 SmallVector<SDValue, 8> Ops;
6435 for (unsigned i = 0, e = BV->getNumOperands(); i != e;
6436 i += NumInputsPerOutput) {
6437 bool isLE = TLI.isLittleEndian();
6438 APInt NewBits = APInt(DstBitSize, 0);
6439 bool EltIsUndef = true;
6440 for (unsigned j = 0; j != NumInputsPerOutput; ++j) {
6441 // Shift the previously computed bits over.
6442 NewBits <<= SrcBitSize;
6443 SDValue Op = BV->getOperand(i+ (isLE ? (NumInputsPerOutput-j-1) : j));
6444 if (Op.getOpcode() == ISD::UNDEF) continue;
6447 NewBits |= cast<ConstantSDNode>(Op)->getAPIntValue().
6448 zextOrTrunc(SrcBitSize).zext(DstBitSize);
6452 Ops.push_back(DAG.getUNDEF(DstEltVT));
6454 Ops.push_back(DAG.getConstant(NewBits, DstEltVT));
6457 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, Ops.size());
6458 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
6461 // Finally, this must be the case where we are shrinking elements: each input
6462 // turns into multiple outputs.
6463 bool isS2V = ISD::isScalarToVector(BV);
6464 unsigned NumOutputsPerInput = SrcBitSize/DstBitSize;
6465 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
6466 NumOutputsPerInput*BV->getNumOperands());
6467 SmallVector<SDValue, 8> Ops;
6469 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
6470 if (BV->getOperand(i).getOpcode() == ISD::UNDEF) {
6471 for (unsigned j = 0; j != NumOutputsPerInput; ++j)
6472 Ops.push_back(DAG.getUNDEF(DstEltVT));
6476 APInt OpVal = cast<ConstantSDNode>(BV->getOperand(i))->
6477 getAPIntValue().zextOrTrunc(SrcBitSize);
6479 for (unsigned j = 0; j != NumOutputsPerInput; ++j) {
6480 APInt ThisVal = OpVal.trunc(DstBitSize);
6481 Ops.push_back(DAG.getConstant(ThisVal, DstEltVT));
6482 if (isS2V && i == 0 && j == 0 && ThisVal.zext(SrcBitSize) == OpVal)
6483 // Simply turn this into a SCALAR_TO_VECTOR of the new type.
6484 return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(BV), VT,
6486 OpVal = OpVal.lshr(DstBitSize);
6489 // For big endian targets, swap the order of the pieces of each element.
6490 if (TLI.isBigEndian())
6491 std::reverse(Ops.end()-NumOutputsPerInput, Ops.end());
6494 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
6497 SDValue DAGCombiner::visitFADD(SDNode *N) {
6498 SDValue N0 = N->getOperand(0);
6499 SDValue N1 = N->getOperand(1);
6500 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6501 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6502 EVT VT = N->getValueType(0);
6505 if (VT.isVector()) {
6506 SDValue FoldedVOp = SimplifyVBinOp(N);
6507 if (FoldedVOp.getNode()) return FoldedVOp;
6510 // fold (fadd c1, c2) -> c1 + c2
6512 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N1);
6513 // canonicalize constant to RHS
6514 if (N0CFP && !N1CFP)
6515 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N1, N0);
6516 // fold (fadd A, 0) -> A
6517 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
6518 N1CFP->getValueAPF().isZero())
6520 // fold (fadd A, (fneg B)) -> (fsub A, B)
6521 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
6522 isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options) == 2)
6523 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N0,
6524 GetNegatedExpression(N1, DAG, LegalOperations));
6525 // fold (fadd (fneg A), B) -> (fsub B, A)
6526 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
6527 isNegatibleForFree(N0, LegalOperations, TLI, &DAG.getTarget().Options) == 2)
6528 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N1,
6529 GetNegatedExpression(N0, DAG, LegalOperations));
6531 // If allowed, fold (fadd (fadd x, c1), c2) -> (fadd x, (fadd c1, c2))
6532 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
6533 N0.getOpcode() == ISD::FADD && N0.getNode()->hasOneUse() &&
6534 isa<ConstantFPSDNode>(N0.getOperand(1)))
6535 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0.getOperand(0),
6536 DAG.getNode(ISD::FADD, SDLoc(N), VT,
6537 N0.getOperand(1), N1));
6539 // No FP constant should be created after legalization as Instruction
6540 // Selection pass has hard time in dealing with FP constant.
6542 // We don't need test this condition for transformation like following, as
6543 // the DAG being transformed implies it is legal to take FP constant as
6546 // (fadd (fmul c, x), x) -> (fmul c+1, x)
6548 bool AllowNewFpConst = (Level < AfterLegalizeDAG);
6550 // If allow, fold (fadd (fneg x), x) -> 0.0
6551 if (AllowNewFpConst && DAG.getTarget().Options.UnsafeFPMath &&
6552 N0.getOpcode() == ISD::FNEG && N0.getOperand(0) == N1)
6553 return DAG.getConstantFP(0.0, VT);
6555 // If allow, fold (fadd x, (fneg x)) -> 0.0
6556 if (AllowNewFpConst && DAG.getTarget().Options.UnsafeFPMath &&
6557 N1.getOpcode() == ISD::FNEG && N1.getOperand(0) == N0)
6558 return DAG.getConstantFP(0.0, VT);
6560 // In unsafe math mode, we can fold chains of FADD's of the same value
6561 // into multiplications. This transform is not safe in general because
6562 // we are reducing the number of rounding steps.
6563 if (DAG.getTarget().Options.UnsafeFPMath &&
6564 TLI.isOperationLegalOrCustom(ISD::FMUL, VT) &&
6566 if (N0.getOpcode() == ISD::FMUL) {
6567 ConstantFPSDNode *CFP00 = dyn_cast<ConstantFPSDNode>(N0.getOperand(0));
6568 ConstantFPSDNode *CFP01 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
6570 // (fadd (fmul c, x), x) -> (fmul x, c+1)
6571 if (CFP00 && !CFP01 && N0.getOperand(1) == N1) {
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,
6579 // (fadd (fmul x, c), x) -> (fmul x, c+1)
6580 if (CFP01 && !CFP00 && N0.getOperand(0) == N1) {
6581 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6583 DAG.getConstantFP(1.0, VT));
6584 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6588 // (fadd (fmul c, x), (fadd x, x)) -> (fmul x, c+2)
6589 if (CFP00 && !CFP01 && N1.getOpcode() == ISD::FADD &&
6590 N1.getOperand(0) == N1.getOperand(1) &&
6591 N0.getOperand(1) == N1.getOperand(0)) {
6592 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6594 DAG.getConstantFP(2.0, VT));
6595 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6596 N0.getOperand(1), NewCFP);
6599 // (fadd (fmul x, c), (fadd x, x)) -> (fmul x, c+2)
6600 if (CFP01 && !CFP00 && N1.getOpcode() == ISD::FADD &&
6601 N1.getOperand(0) == N1.getOperand(1) &&
6602 N0.getOperand(0) == N1.getOperand(0)) {
6603 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6605 DAG.getConstantFP(2.0, VT));
6606 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6607 N0.getOperand(0), NewCFP);
6611 if (N1.getOpcode() == ISD::FMUL) {
6612 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0));
6613 ConstantFPSDNode *CFP11 = dyn_cast<ConstantFPSDNode>(N1.getOperand(1));
6615 // (fadd x, (fmul c, x)) -> (fmul x, c+1)
6616 if (CFP10 && !CFP11 && N1.getOperand(1) == N0) {
6617 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6619 DAG.getConstantFP(1.0, VT));
6620 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6624 // (fadd x, (fmul x, c)) -> (fmul x, c+1)
6625 if (CFP11 && !CFP10 && N1.getOperand(0) == N0) {
6626 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6628 DAG.getConstantFP(1.0, VT));
6629 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6634 // (fadd (fadd x, x), (fmul c, x)) -> (fmul x, c+2)
6635 if (CFP10 && !CFP11 && N0.getOpcode() == ISD::FADD &&
6636 N0.getOperand(0) == N0.getOperand(1) &&
6637 N1.getOperand(1) == N0.getOperand(0)) {
6638 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6640 DAG.getConstantFP(2.0, VT));
6641 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6642 N1.getOperand(1), NewCFP);
6645 // (fadd (fadd x, x), (fmul x, c)) -> (fmul x, c+2)
6646 if (CFP11 && !CFP10 && N0.getOpcode() == ISD::FADD &&
6647 N0.getOperand(0) == N0.getOperand(1) &&
6648 N1.getOperand(0) == N0.getOperand(0)) {
6649 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6651 DAG.getConstantFP(2.0, VT));
6652 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6653 N1.getOperand(0), NewCFP);
6657 if (N0.getOpcode() == ISD::FADD && AllowNewFpConst) {
6658 ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N0.getOperand(0));
6659 // (fadd (fadd x, x), x) -> (fmul x, 3.0)
6660 if (!CFP && N0.getOperand(0) == N0.getOperand(1) &&
6661 (N0.getOperand(0) == N1))
6662 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6663 N1, DAG.getConstantFP(3.0, VT));
6666 if (N1.getOpcode() == ISD::FADD && AllowNewFpConst) {
6667 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0));
6668 // (fadd x, (fadd x, x)) -> (fmul x, 3.0)
6669 if (!CFP10 && N1.getOperand(0) == N1.getOperand(1) &&
6670 N1.getOperand(0) == N0)
6671 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6672 N0, DAG.getConstantFP(3.0, VT));
6675 // (fadd (fadd x, x), (fadd x, x)) -> (fmul x, 4.0)
6676 if (AllowNewFpConst &&
6677 N0.getOpcode() == ISD::FADD && N1.getOpcode() == ISD::FADD &&
6678 N0.getOperand(0) == N0.getOperand(1) &&
6679 N1.getOperand(0) == N1.getOperand(1) &&
6680 N0.getOperand(0) == N1.getOperand(0))
6681 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6683 DAG.getConstantFP(4.0, VT));
6686 // FADD -> FMA combines:
6687 if ((DAG.getTarget().Options.AllowFPOpFusion == FPOpFusion::Fast ||
6688 DAG.getTarget().Options.UnsafeFPMath) &&
6689 DAG.getTarget().getTargetLowering()->isFMAFasterThanFMulAndFAdd(VT) &&
6690 (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FMA, VT))) {
6692 // fold (fadd (fmul x, y), z) -> (fma x, y, z)
6693 if (N0.getOpcode() == ISD::FMUL && N0->hasOneUse())
6694 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
6695 N0.getOperand(0), N0.getOperand(1), N1);
6697 // fold (fadd x, (fmul y, z)) -> (fma y, z, x)
6698 // Note: Commutes FADD operands.
6699 if (N1.getOpcode() == ISD::FMUL && N1->hasOneUse())
6700 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
6701 N1.getOperand(0), N1.getOperand(1), N0);
6707 SDValue DAGCombiner::visitFSUB(SDNode *N) {
6708 SDValue N0 = N->getOperand(0);
6709 SDValue N1 = N->getOperand(1);
6710 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6711 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6712 EVT VT = N->getValueType(0);
6716 if (VT.isVector()) {
6717 SDValue FoldedVOp = SimplifyVBinOp(N);
6718 if (FoldedVOp.getNode()) return FoldedVOp;
6721 // fold (fsub c1, c2) -> c1-c2
6723 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N0, N1);
6724 // fold (fsub A, 0) -> A
6725 if (DAG.getTarget().Options.UnsafeFPMath &&
6726 N1CFP && N1CFP->getValueAPF().isZero())
6728 // fold (fsub 0, B) -> -B
6729 if (DAG.getTarget().Options.UnsafeFPMath &&
6730 N0CFP && N0CFP->getValueAPF().isZero()) {
6731 if (isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options))
6732 return GetNegatedExpression(N1, DAG, LegalOperations);
6733 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
6734 return DAG.getNode(ISD::FNEG, dl, VT, N1);
6736 // fold (fsub A, (fneg B)) -> (fadd A, B)
6737 if (isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options))
6738 return DAG.getNode(ISD::FADD, dl, VT, N0,
6739 GetNegatedExpression(N1, DAG, LegalOperations));
6741 // If 'unsafe math' is enabled, fold
6742 // (fsub x, x) -> 0.0 &
6743 // (fsub x, (fadd x, y)) -> (fneg y) &
6744 // (fsub x, (fadd y, x)) -> (fneg y)
6745 if (DAG.getTarget().Options.UnsafeFPMath) {
6747 return DAG.getConstantFP(0.0f, VT);
6749 if (N1.getOpcode() == ISD::FADD) {
6750 SDValue N10 = N1->getOperand(0);
6751 SDValue N11 = N1->getOperand(1);
6753 if (N10 == N0 && isNegatibleForFree(N11, LegalOperations, TLI,
6754 &DAG.getTarget().Options))
6755 return GetNegatedExpression(N11, DAG, LegalOperations);
6757 if (N11 == N0 && isNegatibleForFree(N10, LegalOperations, TLI,
6758 &DAG.getTarget().Options))
6759 return GetNegatedExpression(N10, DAG, LegalOperations);
6763 // FSUB -> FMA combines:
6764 if ((DAG.getTarget().Options.AllowFPOpFusion == FPOpFusion::Fast ||
6765 DAG.getTarget().Options.UnsafeFPMath) &&
6766 DAG.getTarget().getTargetLowering()->isFMAFasterThanFMulAndFAdd(VT) &&
6767 (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FMA, VT))) {
6769 // fold (fsub (fmul x, y), z) -> (fma x, y, (fneg z))
6770 if (N0.getOpcode() == ISD::FMUL && N0->hasOneUse())
6771 return DAG.getNode(ISD::FMA, dl, VT,
6772 N0.getOperand(0), N0.getOperand(1),
6773 DAG.getNode(ISD::FNEG, dl, VT, N1));
6775 // fold (fsub x, (fmul y, z)) -> (fma (fneg y), z, x)
6776 // Note: Commutes FSUB operands.
6777 if (N1.getOpcode() == ISD::FMUL && N1->hasOneUse())
6778 return DAG.getNode(ISD::FMA, dl, VT,
6779 DAG.getNode(ISD::FNEG, dl, VT,
6781 N1.getOperand(1), N0);
6783 // fold (fsub (fneg (fmul, x, y)), z) -> (fma (fneg x), y, (fneg z))
6784 if (N0.getOpcode() == ISD::FNEG &&
6785 N0.getOperand(0).getOpcode() == ISD::FMUL &&
6786 N0->hasOneUse() && N0.getOperand(0).hasOneUse()) {
6787 SDValue N00 = N0.getOperand(0).getOperand(0);
6788 SDValue N01 = N0.getOperand(0).getOperand(1);
6789 return DAG.getNode(ISD::FMA, dl, VT,
6790 DAG.getNode(ISD::FNEG, dl, VT, N00), N01,
6791 DAG.getNode(ISD::FNEG, dl, VT, N1));
6798 SDValue DAGCombiner::visitFMUL(SDNode *N) {
6799 SDValue N0 = N->getOperand(0);
6800 SDValue N1 = N->getOperand(1);
6801 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6802 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6803 EVT VT = N->getValueType(0);
6804 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6807 if (VT.isVector()) {
6808 SDValue FoldedVOp = SimplifyVBinOp(N);
6809 if (FoldedVOp.getNode()) return FoldedVOp;
6812 // fold (fmul c1, c2) -> c1*c2
6814 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0, N1);
6815 // canonicalize constant to RHS
6816 if (N0CFP && !N1CFP)
6817 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N1, N0);
6818 // fold (fmul A, 0) -> 0
6819 if (DAG.getTarget().Options.UnsafeFPMath &&
6820 N1CFP && N1CFP->getValueAPF().isZero())
6822 // fold (fmul A, 0) -> 0, vector edition.
6823 if (DAG.getTarget().Options.UnsafeFPMath &&
6824 ISD::isBuildVectorAllZeros(N1.getNode()))
6826 // fold (fmul A, 1.0) -> A
6827 if (N1CFP && N1CFP->isExactlyValue(1.0))
6829 // fold (fmul X, 2.0) -> (fadd X, X)
6830 if (N1CFP && N1CFP->isExactlyValue(+2.0))
6831 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N0);
6832 // fold (fmul X, -1.0) -> (fneg X)
6833 if (N1CFP && N1CFP->isExactlyValue(-1.0))
6834 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
6835 return DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0);
6837 // fold (fmul (fneg X), (fneg Y)) -> (fmul X, Y)
6838 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI,
6839 &DAG.getTarget().Options)) {
6840 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI,
6841 &DAG.getTarget().Options)) {
6842 // Both can be negated for free, check to see if at least one is cheaper
6844 if (LHSNeg == 2 || RHSNeg == 2)
6845 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6846 GetNegatedExpression(N0, DAG, LegalOperations),
6847 GetNegatedExpression(N1, DAG, LegalOperations));
6851 // If allowed, fold (fmul (fmul x, c1), c2) -> (fmul x, (fmul c1, c2))
6852 if (DAG.getTarget().Options.UnsafeFPMath &&
6853 N1CFP && N0.getOpcode() == ISD::FMUL &&
6854 N0.getNode()->hasOneUse() && isa<ConstantFPSDNode>(N0.getOperand(1)))
6855 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0.getOperand(0),
6856 DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6857 N0.getOperand(1), N1));
6862 SDValue DAGCombiner::visitFMA(SDNode *N) {
6863 SDValue N0 = N->getOperand(0);
6864 SDValue N1 = N->getOperand(1);
6865 SDValue N2 = N->getOperand(2);
6866 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6867 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6868 EVT VT = N->getValueType(0);
6871 if (DAG.getTarget().Options.UnsafeFPMath) {
6872 if (N0CFP && N0CFP->isZero())
6874 if (N1CFP && N1CFP->isZero())
6877 if (N0CFP && N0CFP->isExactlyValue(1.0))
6878 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N1, N2);
6879 if (N1CFP && N1CFP->isExactlyValue(1.0))
6880 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N2);
6882 // Canonicalize (fma c, x, y) -> (fma x, c, y)
6883 if (N0CFP && !N1CFP)
6884 return DAG.getNode(ISD::FMA, SDLoc(N), VT, N1, N0, N2);
6886 // (fma x, c1, (fmul x, c2)) -> (fmul x, c1+c2)
6887 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
6888 N2.getOpcode() == ISD::FMUL &&
6889 N0 == N2.getOperand(0) &&
6890 N2.getOperand(1).getOpcode() == ISD::ConstantFP) {
6891 return DAG.getNode(ISD::FMUL, dl, VT, N0,
6892 DAG.getNode(ISD::FADD, dl, VT, N1, N2.getOperand(1)));
6896 // (fma (fmul x, c1), c2, y) -> (fma x, c1*c2, y)
6897 if (DAG.getTarget().Options.UnsafeFPMath &&
6898 N0.getOpcode() == ISD::FMUL && N1CFP &&
6899 N0.getOperand(1).getOpcode() == ISD::ConstantFP) {
6900 return DAG.getNode(ISD::FMA, dl, VT,
6902 DAG.getNode(ISD::FMUL, dl, VT, N1, N0.getOperand(1)),
6906 // (fma x, 1, y) -> (fadd x, y)
6907 // (fma x, -1, y) -> (fadd (fneg x), y)
6909 if (N1CFP->isExactlyValue(1.0))
6910 return DAG.getNode(ISD::FADD, dl, VT, N0, N2);
6912 if (N1CFP->isExactlyValue(-1.0) &&
6913 (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))) {
6914 SDValue RHSNeg = DAG.getNode(ISD::FNEG, dl, VT, N0);
6915 AddToWorklist(RHSNeg.getNode());
6916 return DAG.getNode(ISD::FADD, dl, VT, N2, RHSNeg);
6920 // (fma x, c, x) -> (fmul x, (c+1))
6921 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP && N0 == N2)
6922 return DAG.getNode(ISD::FMUL, dl, VT, N0,
6923 DAG.getNode(ISD::FADD, dl, VT,
6924 N1, DAG.getConstantFP(1.0, VT)));
6926 // (fma x, c, (fneg x)) -> (fmul x, (c-1))
6927 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
6928 N2.getOpcode() == ISD::FNEG && N2.getOperand(0) == N0)
6929 return DAG.getNode(ISD::FMUL, dl, VT, N0,
6930 DAG.getNode(ISD::FADD, dl, VT,
6931 N1, DAG.getConstantFP(-1.0, VT)));
6937 SDValue DAGCombiner::visitFDIV(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);
6943 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6946 if (VT.isVector()) {
6947 SDValue FoldedVOp = SimplifyVBinOp(N);
6948 if (FoldedVOp.getNode()) return FoldedVOp;
6951 // fold (fdiv c1, c2) -> c1/c2
6953 return DAG.getNode(ISD::FDIV, SDLoc(N), VT, N0, N1);
6955 // fold (fdiv X, c2) -> fmul X, 1/c2 if losing precision is acceptable.
6956 if (N1CFP && DAG.getTarget().Options.UnsafeFPMath) {
6957 // Compute the reciprocal 1.0 / c2.
6958 APFloat N1APF = N1CFP->getValueAPF();
6959 APFloat Recip(N1APF.getSemantics(), 1); // 1.0
6960 APFloat::opStatus st = Recip.divide(N1APF, APFloat::rmNearestTiesToEven);
6961 // Only do the transform if the reciprocal is a legal fp immediate that
6962 // isn't too nasty (eg NaN, denormal, ...).
6963 if ((st == APFloat::opOK || st == APFloat::opInexact) && // Not too nasty
6964 (!LegalOperations ||
6965 // FIXME: custom lowering of ConstantFP might fail (see e.g. ARM
6966 // backend)... we should handle this gracefully after Legalize.
6967 // TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT) ||
6968 TLI.isOperationLegal(llvm::ISD::ConstantFP, VT) ||
6969 TLI.isFPImmLegal(Recip, VT)))
6970 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0,
6971 DAG.getConstantFP(Recip, VT));
6974 // (fdiv (fneg X), (fneg Y)) -> (fdiv X, Y)
6975 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI,
6976 &DAG.getTarget().Options)) {
6977 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI,
6978 &DAG.getTarget().Options)) {
6979 // Both can be negated for free, check to see if at least one is cheaper
6981 if (LHSNeg == 2 || RHSNeg == 2)
6982 return DAG.getNode(ISD::FDIV, SDLoc(N), VT,
6983 GetNegatedExpression(N0, DAG, LegalOperations),
6984 GetNegatedExpression(N1, DAG, LegalOperations));
6991 SDValue DAGCombiner::visitFREM(SDNode *N) {
6992 SDValue N0 = N->getOperand(0);
6993 SDValue N1 = N->getOperand(1);
6994 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6995 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6996 EVT VT = N->getValueType(0);
6998 // fold (frem c1, c2) -> fmod(c1,c2)
7000 return DAG.getNode(ISD::FREM, SDLoc(N), VT, N0, N1);
7005 SDValue DAGCombiner::visitFCOPYSIGN(SDNode *N) {
7006 SDValue N0 = N->getOperand(0);
7007 SDValue N1 = N->getOperand(1);
7008 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7009 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
7010 EVT VT = N->getValueType(0);
7012 if (N0CFP && N1CFP) // Constant fold
7013 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT, N0, N1);
7016 const APFloat& V = N1CFP->getValueAPF();
7017 // copysign(x, c1) -> fabs(x) iff ispos(c1)
7018 // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1)
7019 if (!V.isNegative()) {
7020 if (!LegalOperations || TLI.isOperationLegal(ISD::FABS, VT))
7021 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
7023 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
7024 return DAG.getNode(ISD::FNEG, SDLoc(N), VT,
7025 DAG.getNode(ISD::FABS, SDLoc(N0), VT, N0));
7029 // copysign(fabs(x), y) -> copysign(x, y)
7030 // copysign(fneg(x), y) -> copysign(x, y)
7031 // copysign(copysign(x,z), y) -> copysign(x, y)
7032 if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG ||
7033 N0.getOpcode() == ISD::FCOPYSIGN)
7034 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
7035 N0.getOperand(0), N1);
7037 // copysign(x, abs(y)) -> abs(x)
7038 if (N1.getOpcode() == ISD::FABS)
7039 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
7041 // copysign(x, copysign(y,z)) -> copysign(x, z)
7042 if (N1.getOpcode() == ISD::FCOPYSIGN)
7043 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
7044 N0, N1.getOperand(1));
7046 // copysign(x, fp_extend(y)) -> copysign(x, y)
7047 // copysign(x, fp_round(y)) -> copysign(x, y)
7048 if (N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND)
7049 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
7050 N0, N1.getOperand(0));
7055 SDValue DAGCombiner::visitSINT_TO_FP(SDNode *N) {
7056 SDValue N0 = N->getOperand(0);
7057 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
7058 EVT VT = N->getValueType(0);
7059 EVT OpVT = N0.getValueType();
7061 // fold (sint_to_fp c1) -> c1fp
7063 // ...but only if the target supports immediate floating-point values
7064 (!LegalOperations ||
7065 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
7066 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0);
7068 // If the input is a legal type, and SINT_TO_FP is not legal on this target,
7069 // but UINT_TO_FP is legal on this target, try to convert.
7070 if (!TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT) &&
7071 TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT)) {
7072 // If the sign bit is known to be zero, we can change this to UINT_TO_FP.
7073 if (DAG.SignBitIsZero(N0))
7074 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0);
7077 // The next optimizations are desirable only if SELECT_CC can be lowered.
7078 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT) || !LegalOperations) {
7079 // fold (sint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
7080 if (N0.getOpcode() == ISD::SETCC && N0.getValueType() == MVT::i1 &&
7082 (!LegalOperations ||
7083 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
7085 { N0.getOperand(0), N0.getOperand(1),
7086 DAG.getConstantFP(-1.0, VT) , DAG.getConstantFP(0.0, VT),
7088 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops);
7091 // fold (sint_to_fp (zext (setcc x, y, cc))) ->
7092 // (select_cc x, y, 1.0, 0.0,, cc)
7093 if (N0.getOpcode() == ISD::ZERO_EXTEND &&
7094 N0.getOperand(0).getOpcode() == ISD::SETCC &&!VT.isVector() &&
7095 (!LegalOperations ||
7096 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
7098 { N0.getOperand(0).getOperand(0), N0.getOperand(0).getOperand(1),
7099 DAG.getConstantFP(1.0, VT) , DAG.getConstantFP(0.0, VT),
7100 N0.getOperand(0).getOperand(2) };
7101 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops);
7108 SDValue DAGCombiner::visitUINT_TO_FP(SDNode *N) {
7109 SDValue N0 = N->getOperand(0);
7110 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
7111 EVT VT = N->getValueType(0);
7112 EVT OpVT = N0.getValueType();
7114 // fold (uint_to_fp c1) -> c1fp
7116 // ...but only if the target supports immediate floating-point values
7117 (!LegalOperations ||
7118 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
7119 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0);
7121 // If the input is a legal type, and UINT_TO_FP is not legal on this target,
7122 // but SINT_TO_FP is legal on this target, try to convert.
7123 if (!TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT) &&
7124 TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT)) {
7125 // If the sign bit is known to be zero, we can change this to SINT_TO_FP.
7126 if (DAG.SignBitIsZero(N0))
7127 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0);
7130 // The next optimizations are desirable only if SELECT_CC can be lowered.
7131 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT) || !LegalOperations) {
7132 // fold (uint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
7134 if (N0.getOpcode() == ISD::SETCC && !VT.isVector() &&
7135 (!LegalOperations ||
7136 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
7138 { N0.getOperand(0), N0.getOperand(1),
7139 DAG.getConstantFP(1.0, VT), DAG.getConstantFP(0.0, VT),
7141 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops);
7148 SDValue DAGCombiner::visitFP_TO_SINT(SDNode *N) {
7149 SDValue N0 = N->getOperand(0);
7150 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7151 EVT VT = N->getValueType(0);
7153 // fold (fp_to_sint c1fp) -> c1
7155 return DAG.getNode(ISD::FP_TO_SINT, SDLoc(N), VT, N0);
7160 SDValue DAGCombiner::visitFP_TO_UINT(SDNode *N) {
7161 SDValue N0 = N->getOperand(0);
7162 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7163 EVT VT = N->getValueType(0);
7165 // fold (fp_to_uint c1fp) -> c1
7167 return DAG.getNode(ISD::FP_TO_UINT, SDLoc(N), VT, N0);
7172 SDValue DAGCombiner::visitFP_ROUND(SDNode *N) {
7173 SDValue N0 = N->getOperand(0);
7174 SDValue N1 = N->getOperand(1);
7175 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7176 EVT VT = N->getValueType(0);
7178 // fold (fp_round c1fp) -> c1fp
7180 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, N0, N1);
7182 // fold (fp_round (fp_extend x)) -> x
7183 if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType())
7184 return N0.getOperand(0);
7186 // fold (fp_round (fp_round x)) -> (fp_round x)
7187 if (N0.getOpcode() == ISD::FP_ROUND) {
7188 // This is a value preserving truncation if both round's are.
7189 bool IsTrunc = N->getConstantOperandVal(1) == 1 &&
7190 N0.getNode()->getConstantOperandVal(1) == 1;
7191 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, N0.getOperand(0),
7192 DAG.getIntPtrConstant(IsTrunc));
7195 // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y)
7196 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse()) {
7197 SDValue Tmp = DAG.getNode(ISD::FP_ROUND, SDLoc(N0), VT,
7198 N0.getOperand(0), N1);
7199 AddToWorklist(Tmp.getNode());
7200 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
7201 Tmp, N0.getOperand(1));
7207 SDValue DAGCombiner::visitFP_ROUND_INREG(SDNode *N) {
7208 SDValue N0 = N->getOperand(0);
7209 EVT VT = N->getValueType(0);
7210 EVT EVT = cast<VTSDNode>(N->getOperand(1))->getVT();
7211 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7213 // fold (fp_round_inreg c1fp) -> c1fp
7214 if (N0CFP && isTypeLegal(EVT)) {
7215 SDValue Round = DAG.getConstantFP(*N0CFP->getConstantFPValue(), EVT);
7216 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, Round);
7222 SDValue DAGCombiner::visitFP_EXTEND(SDNode *N) {
7223 SDValue N0 = N->getOperand(0);
7224 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7225 EVT VT = N->getValueType(0);
7227 // If this is fp_round(fpextend), don't fold it, allow ourselves to be folded.
7228 if (N->hasOneUse() &&
7229 N->use_begin()->getOpcode() == ISD::FP_ROUND)
7232 // fold (fp_extend c1fp) -> c1fp
7234 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, N0);
7236 // Turn fp_extend(fp_round(X, 1)) -> x since the fp_round doesn't affect the
7238 if (N0.getOpcode() == ISD::FP_ROUND
7239 && N0.getNode()->getConstantOperandVal(1) == 1) {
7240 SDValue In = N0.getOperand(0);
7241 if (In.getValueType() == VT) return In;
7242 if (VT.bitsLT(In.getValueType()))
7243 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT,
7244 In, N0.getOperand(1));
7245 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, In);
7248 // fold (fpext (load x)) -> (fpext (fptrunc (extload x)))
7249 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
7250 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType())) {
7251 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
7252 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT,
7254 LN0->getBasePtr(), N0.getValueType(),
7255 LN0->getMemOperand());
7256 CombineTo(N, ExtLoad);
7257 CombineTo(N0.getNode(),
7258 DAG.getNode(ISD::FP_ROUND, SDLoc(N0),
7259 N0.getValueType(), ExtLoad, DAG.getIntPtrConstant(1)),
7260 ExtLoad.getValue(1));
7261 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7267 SDValue DAGCombiner::visitFNEG(SDNode *N) {
7268 SDValue N0 = N->getOperand(0);
7269 EVT VT = N->getValueType(0);
7271 if (VT.isVector()) {
7272 SDValue FoldedVOp = SimplifyVUnaryOp(N);
7273 if (FoldedVOp.getNode()) return FoldedVOp;
7276 if (isNegatibleForFree(N0, LegalOperations, DAG.getTargetLoweringInfo(),
7277 &DAG.getTarget().Options))
7278 return GetNegatedExpression(N0, DAG, LegalOperations);
7280 // Transform fneg(bitconvert(x)) -> bitconvert(x^sign) to avoid loading
7281 // constant pool values.
7282 if (!TLI.isFNegFree(VT) && N0.getOpcode() == ISD::BITCAST &&
7284 N0.getNode()->hasOneUse() &&
7285 N0.getOperand(0).getValueType().isInteger()) {
7286 SDValue Int = N0.getOperand(0);
7287 EVT IntVT = Int.getValueType();
7288 if (IntVT.isInteger() && !IntVT.isVector()) {
7289 Int = DAG.getNode(ISD::XOR, SDLoc(N0), IntVT, Int,
7290 DAG.getConstant(APInt::getSignBit(IntVT.getSizeInBits()), IntVT));
7291 AddToWorklist(Int.getNode());
7292 return DAG.getNode(ISD::BITCAST, SDLoc(N),
7297 // (fneg (fmul c, x)) -> (fmul -c, x)
7298 if (N0.getOpcode() == ISD::FMUL) {
7299 ConstantFPSDNode *CFP1 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
7301 APFloat CVal = CFP1->getValueAPF();
7303 if (Level >= AfterLegalizeDAG &&
7304 (TLI.isFPImmLegal(CVal, N->getValueType(0)) ||
7305 TLI.isOperationLegal(ISD::ConstantFP, N->getValueType(0))))
7307 ISD::FMUL, SDLoc(N), VT, N0.getOperand(0),
7308 DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0.getOperand(1)));
7315 SDValue DAGCombiner::visitFCEIL(SDNode *N) {
7316 SDValue N0 = N->getOperand(0);
7317 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7318 EVT VT = N->getValueType(0);
7320 // fold (fceil c1) -> fceil(c1)
7322 return DAG.getNode(ISD::FCEIL, SDLoc(N), VT, N0);
7327 SDValue DAGCombiner::visitFTRUNC(SDNode *N) {
7328 SDValue N0 = N->getOperand(0);
7329 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7330 EVT VT = N->getValueType(0);
7332 // fold (ftrunc c1) -> ftrunc(c1)
7334 return DAG.getNode(ISD::FTRUNC, SDLoc(N), VT, N0);
7339 SDValue DAGCombiner::visitFFLOOR(SDNode *N) {
7340 SDValue N0 = N->getOperand(0);
7341 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7342 EVT VT = N->getValueType(0);
7344 // fold (ffloor c1) -> ffloor(c1)
7346 return DAG.getNode(ISD::FFLOOR, SDLoc(N), VT, N0);
7351 SDValue DAGCombiner::visitFABS(SDNode *N) {
7352 SDValue N0 = N->getOperand(0);
7353 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7354 EVT VT = N->getValueType(0);
7356 if (VT.isVector()) {
7357 SDValue FoldedVOp = SimplifyVUnaryOp(N);
7358 if (FoldedVOp.getNode()) return FoldedVOp;
7361 // fold (fabs c1) -> fabs(c1)
7363 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
7364 // fold (fabs (fabs x)) -> (fabs x)
7365 if (N0.getOpcode() == ISD::FABS)
7366 return N->getOperand(0);
7367 // fold (fabs (fneg x)) -> (fabs x)
7368 // fold (fabs (fcopysign x, y)) -> (fabs x)
7369 if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN)
7370 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0.getOperand(0));
7372 // Transform fabs(bitconvert(x)) -> bitconvert(x&~sign) to avoid loading
7373 // constant pool values.
7374 if (!TLI.isFAbsFree(VT) &&
7375 N0.getOpcode() == ISD::BITCAST && N0.getNode()->hasOneUse() &&
7376 N0.getOperand(0).getValueType().isInteger() &&
7377 !N0.getOperand(0).getValueType().isVector()) {
7378 SDValue Int = N0.getOperand(0);
7379 EVT IntVT = Int.getValueType();
7380 if (IntVT.isInteger() && !IntVT.isVector()) {
7381 Int = DAG.getNode(ISD::AND, SDLoc(N0), IntVT, Int,
7382 DAG.getConstant(~APInt::getSignBit(IntVT.getSizeInBits()), IntVT));
7383 AddToWorklist(Int.getNode());
7384 return DAG.getNode(ISD::BITCAST, SDLoc(N),
7385 N->getValueType(0), Int);
7392 SDValue DAGCombiner::visitBRCOND(SDNode *N) {
7393 SDValue Chain = N->getOperand(0);
7394 SDValue N1 = N->getOperand(1);
7395 SDValue N2 = N->getOperand(2);
7397 // If N is a constant we could fold this into a fallthrough or unconditional
7398 // branch. However that doesn't happen very often in normal code, because
7399 // Instcombine/SimplifyCFG should have handled the available opportunities.
7400 // If we did this folding here, it would be necessary to update the
7401 // MachineBasicBlock CFG, which is awkward.
7403 // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal
7405 if (N1.getOpcode() == ISD::SETCC &&
7406 TLI.isOperationLegalOrCustom(ISD::BR_CC,
7407 N1.getOperand(0).getValueType())) {
7408 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other,
7409 Chain, N1.getOperand(2),
7410 N1.getOperand(0), N1.getOperand(1), N2);
7413 if ((N1.hasOneUse() && N1.getOpcode() == ISD::SRL) ||
7414 ((N1.getOpcode() == ISD::TRUNCATE && N1.hasOneUse()) &&
7415 (N1.getOperand(0).hasOneUse() &&
7416 N1.getOperand(0).getOpcode() == ISD::SRL))) {
7417 SDNode *Trunc = nullptr;
7418 if (N1.getOpcode() == ISD::TRUNCATE) {
7419 // Look pass the truncate.
7420 Trunc = N1.getNode();
7421 N1 = N1.getOperand(0);
7424 // Match this pattern so that we can generate simpler code:
7427 // %b = and i32 %a, 2
7428 // %c = srl i32 %b, 1
7429 // brcond i32 %c ...
7434 // %b = and i32 %a, 2
7435 // %c = setcc eq %b, 0
7438 // This applies only when the AND constant value has one bit set and the
7439 // SRL constant is equal to the log2 of the AND constant. The back-end is
7440 // smart enough to convert the result into a TEST/JMP sequence.
7441 SDValue Op0 = N1.getOperand(0);
7442 SDValue Op1 = N1.getOperand(1);
7444 if (Op0.getOpcode() == ISD::AND &&
7445 Op1.getOpcode() == ISD::Constant) {
7446 SDValue AndOp1 = Op0.getOperand(1);
7448 if (AndOp1.getOpcode() == ISD::Constant) {
7449 const APInt &AndConst = cast<ConstantSDNode>(AndOp1)->getAPIntValue();
7451 if (AndConst.isPowerOf2() &&
7452 cast<ConstantSDNode>(Op1)->getAPIntValue()==AndConst.logBase2()) {
7454 DAG.getSetCC(SDLoc(N),
7455 getSetCCResultType(Op0.getValueType()),
7456 Op0, DAG.getConstant(0, Op0.getValueType()),
7459 SDValue NewBRCond = DAG.getNode(ISD::BRCOND, SDLoc(N),
7460 MVT::Other, Chain, SetCC, N2);
7461 // Don't add the new BRCond into the worklist or else SimplifySelectCC
7462 // will convert it back to (X & C1) >> C2.
7463 CombineTo(N, NewBRCond, false);
7464 // Truncate is dead.
7466 removeFromWorklist(Trunc);
7467 DAG.DeleteNode(Trunc);
7469 // Replace the uses of SRL with SETCC
7470 WorklistRemover DeadNodes(*this);
7471 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
7472 removeFromWorklist(N1.getNode());
7473 DAG.DeleteNode(N1.getNode());
7474 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7480 // Restore N1 if the above transformation doesn't match.
7481 N1 = N->getOperand(1);
7484 // Transform br(xor(x, y)) -> br(x != y)
7485 // Transform br(xor(xor(x,y), 1)) -> br (x == y)
7486 if (N1.hasOneUse() && N1.getOpcode() == ISD::XOR) {
7487 SDNode *TheXor = N1.getNode();
7488 SDValue Op0 = TheXor->getOperand(0);
7489 SDValue Op1 = TheXor->getOperand(1);
7490 if (Op0.getOpcode() == Op1.getOpcode()) {
7491 // Avoid missing important xor optimizations.
7492 SDValue Tmp = visitXOR(TheXor);
7493 if (Tmp.getNode()) {
7494 if (Tmp.getNode() != TheXor) {
7495 DEBUG(dbgs() << "\nReplacing.8 ";
7497 dbgs() << "\nWith: ";
7498 Tmp.getNode()->dump(&DAG);
7500 WorklistRemover DeadNodes(*this);
7501 DAG.ReplaceAllUsesOfValueWith(N1, Tmp);
7502 removeFromWorklist(TheXor);
7503 DAG.DeleteNode(TheXor);
7504 return DAG.getNode(ISD::BRCOND, SDLoc(N),
7505 MVT::Other, Chain, Tmp, N2);
7508 // visitXOR has changed XOR's operands or replaced the XOR completely,
7510 return SDValue(N, 0);
7514 if (Op0.getOpcode() != ISD::SETCC && Op1.getOpcode() != ISD::SETCC) {
7516 if (ConstantSDNode *RHSCI = dyn_cast<ConstantSDNode>(Op0))
7517 if (RHSCI->getAPIntValue() == 1 && Op0.hasOneUse() &&
7518 Op0.getOpcode() == ISD::XOR) {
7519 TheXor = Op0.getNode();
7523 EVT SetCCVT = N1.getValueType();
7525 SetCCVT = getSetCCResultType(SetCCVT);
7526 SDValue SetCC = DAG.getSetCC(SDLoc(TheXor),
7529 Equal ? ISD::SETEQ : ISD::SETNE);
7530 // Replace the uses of XOR with SETCC
7531 WorklistRemover DeadNodes(*this);
7532 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
7533 removeFromWorklist(N1.getNode());
7534 DAG.DeleteNode(N1.getNode());
7535 return DAG.getNode(ISD::BRCOND, SDLoc(N),
7536 MVT::Other, Chain, SetCC, N2);
7543 // Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB.
7545 SDValue DAGCombiner::visitBR_CC(SDNode *N) {
7546 CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1));
7547 SDValue CondLHS = N->getOperand(2), CondRHS = N->getOperand(3);
7549 // If N is a constant we could fold this into a fallthrough or unconditional
7550 // branch. However that doesn't happen very often in normal code, because
7551 // Instcombine/SimplifyCFG should have handled the available opportunities.
7552 // If we did this folding here, it would be necessary to update the
7553 // MachineBasicBlock CFG, which is awkward.
7555 // Use SimplifySetCC to simplify SETCC's.
7556 SDValue Simp = SimplifySetCC(getSetCCResultType(CondLHS.getValueType()),
7557 CondLHS, CondRHS, CC->get(), SDLoc(N),
7559 if (Simp.getNode()) AddToWorklist(Simp.getNode());
7561 // fold to a simpler setcc
7562 if (Simp.getNode() && Simp.getOpcode() == ISD::SETCC)
7563 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other,
7564 N->getOperand(0), Simp.getOperand(2),
7565 Simp.getOperand(0), Simp.getOperand(1),
7571 /// canFoldInAddressingMode - Return true if 'Use' is a load or a store that
7572 /// uses N as its base pointer and that N may be folded in the load / store
7573 /// addressing mode.
7574 static bool canFoldInAddressingMode(SDNode *N, SDNode *Use,
7576 const TargetLowering &TLI) {
7578 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Use)) {
7579 if (LD->isIndexed() || LD->getBasePtr().getNode() != N)
7581 VT = Use->getValueType(0);
7582 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(Use)) {
7583 if (ST->isIndexed() || ST->getBasePtr().getNode() != N)
7585 VT = ST->getValue().getValueType();
7589 TargetLowering::AddrMode AM;
7590 if (N->getOpcode() == ISD::ADD) {
7591 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
7594 AM.BaseOffs = Offset->getSExtValue();
7598 } else if (N->getOpcode() == ISD::SUB) {
7599 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
7602 AM.BaseOffs = -Offset->getSExtValue();
7609 return TLI.isLegalAddressingMode(AM, VT.getTypeForEVT(*DAG.getContext()));
7612 /// CombineToPreIndexedLoadStore - Try turning a load / store into a
7613 /// pre-indexed load / store when the base pointer is an add or subtract
7614 /// and it has other uses besides the load / store. After the
7615 /// transformation, the new indexed load / store has effectively folded
7616 /// the add / subtract in and all of its other uses are redirected to the
7617 /// new load / store.
7618 bool DAGCombiner::CombineToPreIndexedLoadStore(SDNode *N) {
7619 if (Level < AfterLegalizeDAG)
7625 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7626 if (LD->isIndexed())
7628 VT = LD->getMemoryVT();
7629 if (!TLI.isIndexedLoadLegal(ISD::PRE_INC, VT) &&
7630 !TLI.isIndexedLoadLegal(ISD::PRE_DEC, VT))
7632 Ptr = LD->getBasePtr();
7633 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7634 if (ST->isIndexed())
7636 VT = ST->getMemoryVT();
7637 if (!TLI.isIndexedStoreLegal(ISD::PRE_INC, VT) &&
7638 !TLI.isIndexedStoreLegal(ISD::PRE_DEC, VT))
7640 Ptr = ST->getBasePtr();
7646 // If the pointer is not an add/sub, or if it doesn't have multiple uses, bail
7647 // out. There is no reason to make this a preinc/predec.
7648 if ((Ptr.getOpcode() != ISD::ADD && Ptr.getOpcode() != ISD::SUB) ||
7649 Ptr.getNode()->hasOneUse())
7652 // Ask the target to do addressing mode selection.
7655 ISD::MemIndexedMode AM = ISD::UNINDEXED;
7656 if (!TLI.getPreIndexedAddressParts(N, BasePtr, Offset, AM, DAG))
7659 // Backends without true r+i pre-indexed forms may need to pass a
7660 // constant base with a variable offset so that constant coercion
7661 // will work with the patterns in canonical form.
7662 bool Swapped = false;
7663 if (isa<ConstantSDNode>(BasePtr)) {
7664 std::swap(BasePtr, Offset);
7668 // Don't create a indexed load / store with zero offset.
7669 if (isa<ConstantSDNode>(Offset) &&
7670 cast<ConstantSDNode>(Offset)->isNullValue())
7673 // Try turning it into a pre-indexed load / store except when:
7674 // 1) The new base ptr is a frame index.
7675 // 2) If N is a store and the new base ptr is either the same as or is a
7676 // predecessor of the value being stored.
7677 // 3) Another use of old base ptr is a predecessor of N. If ptr is folded
7678 // that would create a cycle.
7679 // 4) All uses are load / store ops that use it as old base ptr.
7681 // Check #1. Preinc'ing a frame index would require copying the stack pointer
7682 // (plus the implicit offset) to a register to preinc anyway.
7683 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
7688 SDValue Val = cast<StoreSDNode>(N)->getValue();
7689 if (Val == BasePtr || BasePtr.getNode()->isPredecessorOf(Val.getNode()))
7693 // If the offset is a constant, there may be other adds of constants that
7694 // can be folded with this one. We should do this to avoid having to keep
7695 // a copy of the original base pointer.
7696 SmallVector<SDNode *, 16> OtherUses;
7697 if (isa<ConstantSDNode>(Offset))
7698 for (SDNode *Use : BasePtr.getNode()->uses()) {
7699 if (Use == Ptr.getNode())
7702 if (Use->isPredecessorOf(N))
7705 if (Use->getOpcode() != ISD::ADD && Use->getOpcode() != ISD::SUB) {
7710 SDValue Op0 = Use->getOperand(0), Op1 = Use->getOperand(1);
7711 if (Op1.getNode() == BasePtr.getNode())
7712 std::swap(Op0, Op1);
7713 assert(Op0.getNode() == BasePtr.getNode() &&
7714 "Use of ADD/SUB but not an operand");
7716 if (!isa<ConstantSDNode>(Op1)) {
7721 // FIXME: In some cases, we can be smarter about this.
7722 if (Op1.getValueType() != Offset.getValueType()) {
7727 OtherUses.push_back(Use);
7731 std::swap(BasePtr, Offset);
7733 // Now check for #3 and #4.
7734 bool RealUse = false;
7736 // Caches for hasPredecessorHelper
7737 SmallPtrSet<const SDNode *, 32> Visited;
7738 SmallVector<const SDNode *, 16> Worklist;
7740 for (SDNode *Use : Ptr.getNode()->uses()) {
7743 if (N->hasPredecessorHelper(Use, Visited, Worklist))
7746 // If Ptr may be folded in addressing mode of other use, then it's
7747 // not profitable to do this transformation.
7748 if (!canFoldInAddressingMode(Ptr.getNode(), Use, DAG, TLI))
7757 Result = DAG.getIndexedLoad(SDValue(N,0), SDLoc(N),
7758 BasePtr, Offset, AM);
7760 Result = DAG.getIndexedStore(SDValue(N,0), SDLoc(N),
7761 BasePtr, Offset, AM);
7764 DEBUG(dbgs() << "\nReplacing.4 ";
7766 dbgs() << "\nWith: ";
7767 Result.getNode()->dump(&DAG);
7769 WorklistRemover DeadNodes(*this);
7771 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
7772 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
7774 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
7777 // Finally, since the node is now dead, remove it from the graph.
7781 std::swap(BasePtr, Offset);
7783 // Replace other uses of BasePtr that can be updated to use Ptr
7784 for (unsigned i = 0, e = OtherUses.size(); i != e; ++i) {
7785 unsigned OffsetIdx = 1;
7786 if (OtherUses[i]->getOperand(OffsetIdx).getNode() == BasePtr.getNode())
7788 assert(OtherUses[i]->getOperand(!OffsetIdx).getNode() ==
7789 BasePtr.getNode() && "Expected BasePtr operand");
7791 // We need to replace ptr0 in the following expression:
7792 // x0 * offset0 + y0 * ptr0 = t0
7794 // x1 * offset1 + y1 * ptr0 = t1 (the indexed load/store)
7796 // where x0, x1, y0 and y1 in {-1, 1} are given by the types of the
7797 // indexed load/store and the expresion that needs to be re-written.
7799 // Therefore, we have:
7800 // t0 = (x0 * offset0 - x1 * y0 * y1 *offset1) + (y0 * y1) * t1
7802 ConstantSDNode *CN =
7803 cast<ConstantSDNode>(OtherUses[i]->getOperand(OffsetIdx));
7805 APInt Offset0 = CN->getAPIntValue();
7806 APInt Offset1 = cast<ConstantSDNode>(Offset)->getAPIntValue();
7808 X0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 1) ? -1 : 1;
7809 Y0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 0) ? -1 : 1;
7810 X1 = (AM == ISD::PRE_DEC && !Swapped) ? -1 : 1;
7811 Y1 = (AM == ISD::PRE_DEC && Swapped) ? -1 : 1;
7813 unsigned Opcode = (Y0 * Y1 < 0) ? ISD::SUB : ISD::ADD;
7815 APInt CNV = Offset0;
7816 if (X0 < 0) CNV = -CNV;
7817 if (X1 * Y0 * Y1 < 0) CNV = CNV + Offset1;
7818 else CNV = CNV - Offset1;
7820 // We can now generate the new expression.
7821 SDValue NewOp1 = DAG.getConstant(CNV, CN->getValueType(0));
7822 SDValue NewOp2 = Result.getValue(isLoad ? 1 : 0);
7824 SDValue NewUse = DAG.getNode(Opcode,
7825 SDLoc(OtherUses[i]),
7826 OtherUses[i]->getValueType(0), NewOp1, NewOp2);
7827 DAG.ReplaceAllUsesOfValueWith(SDValue(OtherUses[i], 0), NewUse);
7828 removeFromWorklist(OtherUses[i]);
7829 DAG.DeleteNode(OtherUses[i]);
7832 // Replace the uses of Ptr with uses of the updated base value.
7833 DAG.ReplaceAllUsesOfValueWith(Ptr, Result.getValue(isLoad ? 1 : 0));
7834 removeFromWorklist(Ptr.getNode());
7835 DAG.DeleteNode(Ptr.getNode());
7840 /// CombineToPostIndexedLoadStore - Try to combine a load / store with a
7841 /// add / sub of the base pointer node into a post-indexed load / store.
7842 /// The transformation folded the add / subtract into the new indexed
7843 /// load / store effectively and all of its uses are redirected to the
7844 /// new load / store.
7845 bool DAGCombiner::CombineToPostIndexedLoadStore(SDNode *N) {
7846 if (Level < AfterLegalizeDAG)
7852 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7853 if (LD->isIndexed())
7855 VT = LD->getMemoryVT();
7856 if (!TLI.isIndexedLoadLegal(ISD::POST_INC, VT) &&
7857 !TLI.isIndexedLoadLegal(ISD::POST_DEC, VT))
7859 Ptr = LD->getBasePtr();
7860 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7861 if (ST->isIndexed())
7863 VT = ST->getMemoryVT();
7864 if (!TLI.isIndexedStoreLegal(ISD::POST_INC, VT) &&
7865 !TLI.isIndexedStoreLegal(ISD::POST_DEC, VT))
7867 Ptr = ST->getBasePtr();
7873 if (Ptr.getNode()->hasOneUse())
7876 for (SDNode *Op : Ptr.getNode()->uses()) {
7878 (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB))
7883 ISD::MemIndexedMode AM = ISD::UNINDEXED;
7884 if (TLI.getPostIndexedAddressParts(N, Op, BasePtr, Offset, AM, DAG)) {
7885 // Don't create a indexed load / store with zero offset.
7886 if (isa<ConstantSDNode>(Offset) &&
7887 cast<ConstantSDNode>(Offset)->isNullValue())
7890 // Try turning it into a post-indexed load / store except when
7891 // 1) All uses are load / store ops that use it as base ptr (and
7892 // it may be folded as addressing mmode).
7893 // 2) Op must be independent of N, i.e. Op is neither a predecessor
7894 // nor a successor of N. Otherwise, if Op is folded that would
7897 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
7901 bool TryNext = false;
7902 for (SDNode *Use : BasePtr.getNode()->uses()) {
7903 if (Use == Ptr.getNode())
7906 // If all the uses are load / store addresses, then don't do the
7908 if (Use->getOpcode() == ISD::ADD || Use->getOpcode() == ISD::SUB){
7909 bool RealUse = false;
7910 for (SDNode *UseUse : Use->uses()) {
7911 if (!canFoldInAddressingMode(Use, UseUse, DAG, TLI))
7926 if (!Op->isPredecessorOf(N) && !N->isPredecessorOf(Op)) {
7927 SDValue Result = isLoad
7928 ? DAG.getIndexedLoad(SDValue(N,0), SDLoc(N),
7929 BasePtr, Offset, AM)
7930 : DAG.getIndexedStore(SDValue(N,0), SDLoc(N),
7931 BasePtr, Offset, AM);
7934 DEBUG(dbgs() << "\nReplacing.5 ";
7936 dbgs() << "\nWith: ";
7937 Result.getNode()->dump(&DAG);
7939 WorklistRemover DeadNodes(*this);
7941 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
7942 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
7944 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
7947 // Finally, since the node is now dead, remove it from the graph.
7950 // Replace the uses of Use with uses of the updated base value.
7951 DAG.ReplaceAllUsesOfValueWith(SDValue(Op, 0),
7952 Result.getValue(isLoad ? 1 : 0));
7953 removeFromWorklist(Op);
7963 SDValue DAGCombiner::visitLOAD(SDNode *N) {
7964 LoadSDNode *LD = cast<LoadSDNode>(N);
7965 SDValue Chain = LD->getChain();
7966 SDValue Ptr = LD->getBasePtr();
7968 // If load is not volatile and there are no uses of the loaded value (and
7969 // the updated indexed value in case of indexed loads), change uses of the
7970 // chain value into uses of the chain input (i.e. delete the dead load).
7971 if (!LD->isVolatile()) {
7972 if (N->getValueType(1) == MVT::Other) {
7974 if (!N->hasAnyUseOfValue(0)) {
7975 // It's not safe to use the two value CombineTo variant here. e.g.
7976 // v1, chain2 = load chain1, loc
7977 // v2, chain3 = load chain2, loc
7979 // Now we replace use of chain2 with chain1. This makes the second load
7980 // isomorphic to the one we are deleting, and thus makes this load live.
7981 DEBUG(dbgs() << "\nReplacing.6 ";
7983 dbgs() << "\nWith chain: ";
7984 Chain.getNode()->dump(&DAG);
7986 WorklistRemover DeadNodes(*this);
7987 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
7989 if (N->use_empty()) {
7990 removeFromWorklist(N);
7994 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7998 assert(N->getValueType(2) == MVT::Other && "Malformed indexed loads?");
7999 if (!N->hasAnyUseOfValue(0) && !N->hasAnyUseOfValue(1)) {
8000 SDValue Undef = DAG.getUNDEF(N->getValueType(0));
8001 DEBUG(dbgs() << "\nReplacing.7 ";
8003 dbgs() << "\nWith: ";
8004 Undef.getNode()->dump(&DAG);
8005 dbgs() << " and 2 other values\n");
8006 WorklistRemover DeadNodes(*this);
8007 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Undef);
8008 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1),
8009 DAG.getUNDEF(N->getValueType(1)));
8010 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 2), Chain);
8011 removeFromWorklist(N);
8013 return SDValue(N, 0); // Return N so it doesn't get rechecked!
8018 // If this load is directly stored, replace the load value with the stored
8020 // TODO: Handle store large -> read small portion.
8021 // TODO: Handle TRUNCSTORE/LOADEXT
8022 if (ISD::isNormalLoad(N) && !LD->isVolatile()) {
8023 if (ISD::isNON_TRUNCStore(Chain.getNode())) {
8024 StoreSDNode *PrevST = cast<StoreSDNode>(Chain);
8025 if (PrevST->getBasePtr() == Ptr &&
8026 PrevST->getValue().getValueType() == N->getValueType(0))
8027 return CombineTo(N, Chain.getOperand(1), Chain);
8031 // Try to infer better alignment information than the load already has.
8032 if (OptLevel != CodeGenOpt::None && LD->isUnindexed()) {
8033 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
8034 if (Align > LD->getMemOperand()->getBaseAlignment()) {
8036 DAG.getExtLoad(LD->getExtensionType(), SDLoc(N),
8037 LD->getValueType(0),
8038 Chain, Ptr, LD->getPointerInfo(),
8040 LD->isVolatile(), LD->isNonTemporal(), Align,
8042 return CombineTo(N, NewLoad, SDValue(NewLoad.getNode(), 1), true);
8047 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA :
8048 TLI.getTargetMachine().getSubtarget<TargetSubtargetInfo>().useAA();
8050 if (CombinerAAOnlyFunc.getNumOccurrences() &&
8051 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
8054 if (UseAA && LD->isUnindexed()) {
8055 // Walk up chain skipping non-aliasing memory nodes.
8056 SDValue BetterChain = FindBetterChain(N, Chain);
8058 // If there is a better chain.
8059 if (Chain != BetterChain) {
8062 // Replace the chain to void dependency.
8063 if (LD->getExtensionType() == ISD::NON_EXTLOAD) {
8064 ReplLoad = DAG.getLoad(N->getValueType(0), SDLoc(LD),
8065 BetterChain, Ptr, LD->getMemOperand());
8067 ReplLoad = DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD),
8068 LD->getValueType(0),
8069 BetterChain, Ptr, LD->getMemoryVT(),
8070 LD->getMemOperand());
8073 // Create token factor to keep old chain connected.
8074 SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N),
8075 MVT::Other, Chain, ReplLoad.getValue(1));
8077 // Make sure the new and old chains are cleaned up.
8078 AddToWorklist(Token.getNode());
8080 // Replace uses with load result and token factor. Don't add users
8082 return CombineTo(N, ReplLoad.getValue(0), Token, false);
8086 // Try transforming N to an indexed load.
8087 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
8088 return SDValue(N, 0);
8090 // Try to slice up N to more direct loads if the slices are mapped to
8091 // different register banks or pairing can take place.
8093 return SDValue(N, 0);
8099 /// \brief Helper structure used to slice a load in smaller loads.
8100 /// Basically a slice is obtained from the following sequence:
8101 /// Origin = load Ty1, Base
8102 /// Shift = srl Ty1 Origin, CstTy Amount
8103 /// Inst = trunc Shift to Ty2
8105 /// Then, it will be rewriten into:
8106 /// Slice = load SliceTy, Base + SliceOffset
8107 /// [Inst = zext Slice to Ty2], only if SliceTy <> Ty2
8109 /// SliceTy is deduced from the number of bits that are actually used to
8111 struct LoadedSlice {
8112 /// \brief Helper structure used to compute the cost of a slice.
8114 /// Are we optimizing for code size.
8119 unsigned CrossRegisterBanksCopies;
8123 Cost(bool ForCodeSize = false)
8124 : ForCodeSize(ForCodeSize), Loads(0), Truncates(0),
8125 CrossRegisterBanksCopies(0), ZExts(0), Shift(0) {}
8127 /// \brief Get the cost of one isolated slice.
8128 Cost(const LoadedSlice &LS, bool ForCodeSize = false)
8129 : ForCodeSize(ForCodeSize), Loads(1), Truncates(0),
8130 CrossRegisterBanksCopies(0), ZExts(0), Shift(0) {
8131 EVT TruncType = LS.Inst->getValueType(0);
8132 EVT LoadedType = LS.getLoadedType();
8133 if (TruncType != LoadedType &&
8134 !LS.DAG->getTargetLoweringInfo().isZExtFree(LoadedType, TruncType))
8138 /// \brief Account for slicing gain in the current cost.
8139 /// Slicing provide a few gains like removing a shift or a
8140 /// truncate. This method allows to grow the cost of the original
8141 /// load with the gain from this slice.
8142 void addSliceGain(const LoadedSlice &LS) {
8143 // Each slice saves a truncate.
8144 const TargetLowering &TLI = LS.DAG->getTargetLoweringInfo();
8145 if (!TLI.isTruncateFree(LS.Inst->getValueType(0),
8146 LS.Inst->getOperand(0).getValueType()))
8148 // If there is a shift amount, this slice gets rid of it.
8151 // If this slice can merge a cross register bank copy, account for it.
8152 if (LS.canMergeExpensiveCrossRegisterBankCopy())
8153 ++CrossRegisterBanksCopies;
8156 Cost &operator+=(const Cost &RHS) {
8158 Truncates += RHS.Truncates;
8159 CrossRegisterBanksCopies += RHS.CrossRegisterBanksCopies;
8165 bool operator==(const Cost &RHS) const {
8166 return Loads == RHS.Loads && Truncates == RHS.Truncates &&
8167 CrossRegisterBanksCopies == RHS.CrossRegisterBanksCopies &&
8168 ZExts == RHS.ZExts && Shift == RHS.Shift;
8171 bool operator!=(const Cost &RHS) const { return !(*this == RHS); }
8173 bool operator<(const Cost &RHS) const {
8174 // Assume cross register banks copies are as expensive as loads.
8175 // FIXME: Do we want some more target hooks?
8176 unsigned ExpensiveOpsLHS = Loads + CrossRegisterBanksCopies;
8177 unsigned ExpensiveOpsRHS = RHS.Loads + RHS.CrossRegisterBanksCopies;
8178 // Unless we are optimizing for code size, consider the
8179 // expensive operation first.
8180 if (!ForCodeSize && ExpensiveOpsLHS != ExpensiveOpsRHS)
8181 return ExpensiveOpsLHS < ExpensiveOpsRHS;
8182 return (Truncates + ZExts + Shift + ExpensiveOpsLHS) <
8183 (RHS.Truncates + RHS.ZExts + RHS.Shift + ExpensiveOpsRHS);
8186 bool operator>(const Cost &RHS) const { return RHS < *this; }
8188 bool operator<=(const Cost &RHS) const { return !(RHS < *this); }
8190 bool operator>=(const Cost &RHS) const { return !(*this < RHS); }
8192 // The last instruction that represent the slice. This should be a
8193 // truncate instruction.
8195 // The original load instruction.
8197 // The right shift amount in bits from the original load.
8199 // The DAG from which Origin came from.
8200 // This is used to get some contextual information about legal types, etc.
8203 LoadedSlice(SDNode *Inst = nullptr, LoadSDNode *Origin = nullptr,
8204 unsigned Shift = 0, SelectionDAG *DAG = nullptr)
8205 : Inst(Inst), Origin(Origin), Shift(Shift), DAG(DAG) {}
8207 LoadedSlice(const LoadedSlice &LS)
8208 : Inst(LS.Inst), Origin(LS.Origin), Shift(LS.Shift), DAG(LS.DAG) {}
8210 /// \brief Get the bits used in a chunk of bits \p BitWidth large.
8211 /// \return Result is \p BitWidth and has used bits set to 1 and
8212 /// not used bits set to 0.
8213 APInt getUsedBits() const {
8214 // Reproduce the trunc(lshr) sequence:
8215 // - Start from the truncated value.
8216 // - Zero extend to the desired bit width.
8218 assert(Origin && "No original load to compare against.");
8219 unsigned BitWidth = Origin->getValueSizeInBits(0);
8220 assert(Inst && "This slice is not bound to an instruction");
8221 assert(Inst->getValueSizeInBits(0) <= BitWidth &&
8222 "Extracted slice is bigger than the whole type!");
8223 APInt UsedBits(Inst->getValueSizeInBits(0), 0);
8224 UsedBits.setAllBits();
8225 UsedBits = UsedBits.zext(BitWidth);
8230 /// \brief Get the size of the slice to be loaded in bytes.
8231 unsigned getLoadedSize() const {
8232 unsigned SliceSize = getUsedBits().countPopulation();
8233 assert(!(SliceSize & 0x7) && "Size is not a multiple of a byte.");
8234 return SliceSize / 8;
8237 /// \brief Get the type that will be loaded for this slice.
8238 /// Note: This may not be the final type for the slice.
8239 EVT getLoadedType() const {
8240 assert(DAG && "Missing context");
8241 LLVMContext &Ctxt = *DAG->getContext();
8242 return EVT::getIntegerVT(Ctxt, getLoadedSize() * 8);
8245 /// \brief Get the alignment of the load used for this slice.
8246 unsigned getAlignment() const {
8247 unsigned Alignment = Origin->getAlignment();
8248 unsigned Offset = getOffsetFromBase();
8250 Alignment = MinAlign(Alignment, Alignment + Offset);
8254 /// \brief Check if this slice can be rewritten with legal operations.
8255 bool isLegal() const {
8256 // An invalid slice is not legal.
8257 if (!Origin || !Inst || !DAG)
8260 // Offsets are for indexed load only, we do not handle that.
8261 if (Origin->getOffset().getOpcode() != ISD::UNDEF)
8264 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
8266 // Check that the type is legal.
8267 EVT SliceType = getLoadedType();
8268 if (!TLI.isTypeLegal(SliceType))
8271 // Check that the load is legal for this type.
8272 if (!TLI.isOperationLegal(ISD::LOAD, SliceType))
8275 // Check that the offset can be computed.
8276 // 1. Check its type.
8277 EVT PtrType = Origin->getBasePtr().getValueType();
8278 if (PtrType == MVT::Untyped || PtrType.isExtended())
8281 // 2. Check that it fits in the immediate.
8282 if (!TLI.isLegalAddImmediate(getOffsetFromBase()))
8285 // 3. Check that the computation is legal.
8286 if (!TLI.isOperationLegal(ISD::ADD, PtrType))
8289 // Check that the zext is legal if it needs one.
8290 EVT TruncateType = Inst->getValueType(0);
8291 if (TruncateType != SliceType &&
8292 !TLI.isOperationLegal(ISD::ZERO_EXTEND, TruncateType))
8298 /// \brief Get the offset in bytes of this slice in the original chunk of
8300 /// \pre DAG != nullptr.
8301 uint64_t getOffsetFromBase() const {
8302 assert(DAG && "Missing context.");
8304 DAG->getTargetLoweringInfo().getDataLayout()->isBigEndian();
8305 assert(!(Shift & 0x7) && "Shifts not aligned on Bytes are not supported.");
8306 uint64_t Offset = Shift / 8;
8307 unsigned TySizeInBytes = Origin->getValueSizeInBits(0) / 8;
8308 assert(!(Origin->getValueSizeInBits(0) & 0x7) &&
8309 "The size of the original loaded type is not a multiple of a"
8311 // If Offset is bigger than TySizeInBytes, it means we are loading all
8312 // zeros. This should have been optimized before in the process.
8313 assert(TySizeInBytes > Offset &&
8314 "Invalid shift amount for given loaded size");
8316 Offset = TySizeInBytes - Offset - getLoadedSize();
8320 /// \brief Generate the sequence of instructions to load the slice
8321 /// represented by this object and redirect the uses of this slice to
8322 /// this new sequence of instructions.
8323 /// \pre this->Inst && this->Origin are valid Instructions and this
8324 /// object passed the legal check: LoadedSlice::isLegal returned true.
8325 /// \return The last instruction of the sequence used to load the slice.
8326 SDValue loadSlice() const {
8327 assert(Inst && Origin && "Unable to replace a non-existing slice.");
8328 const SDValue &OldBaseAddr = Origin->getBasePtr();
8329 SDValue BaseAddr = OldBaseAddr;
8330 // Get the offset in that chunk of bytes w.r.t. the endianess.
8331 int64_t Offset = static_cast<int64_t>(getOffsetFromBase());
8332 assert(Offset >= 0 && "Offset too big to fit in int64_t!");
8334 // BaseAddr = BaseAddr + Offset.
8335 EVT ArithType = BaseAddr.getValueType();
8336 BaseAddr = DAG->getNode(ISD::ADD, SDLoc(Origin), ArithType, BaseAddr,
8337 DAG->getConstant(Offset, ArithType));
8340 // Create the type of the loaded slice according to its size.
8341 EVT SliceType = getLoadedType();
8343 // Create the load for the slice.
8344 SDValue LastInst = DAG->getLoad(
8345 SliceType, SDLoc(Origin), Origin->getChain(), BaseAddr,
8346 Origin->getPointerInfo().getWithOffset(Offset), Origin->isVolatile(),
8347 Origin->isNonTemporal(), Origin->isInvariant(), getAlignment());
8348 // If the final type is not the same as the loaded type, this means that
8349 // we have to pad with zero. Create a zero extend for that.
8350 EVT FinalType = Inst->getValueType(0);
8351 if (SliceType != FinalType)
8353 DAG->getNode(ISD::ZERO_EXTEND, SDLoc(LastInst), FinalType, LastInst);
8357 /// \brief Check if this slice can be merged with an expensive cross register
8358 /// bank copy. E.g.,
8360 /// f = bitcast i32 i to float
8361 bool canMergeExpensiveCrossRegisterBankCopy() const {
8362 if (!Inst || !Inst->hasOneUse())
8364 SDNode *Use = *Inst->use_begin();
8365 if (Use->getOpcode() != ISD::BITCAST)
8367 assert(DAG && "Missing context");
8368 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
8369 EVT ResVT = Use->getValueType(0);
8370 const TargetRegisterClass *ResRC = TLI.getRegClassFor(ResVT.getSimpleVT());
8371 const TargetRegisterClass *ArgRC =
8372 TLI.getRegClassFor(Use->getOperand(0).getValueType().getSimpleVT());
8373 if (ArgRC == ResRC || !TLI.isOperationLegal(ISD::LOAD, ResVT))
8376 // At this point, we know that we perform a cross-register-bank copy.
8377 // Check if it is expensive.
8378 const TargetRegisterInfo *TRI = TLI.getTargetMachine().getRegisterInfo();
8379 // Assume bitcasts are cheap, unless both register classes do not
8380 // explicitly share a common sub class.
8381 if (!TRI || TRI->getCommonSubClass(ArgRC, ResRC))
8384 // Check if it will be merged with the load.
8385 // 1. Check the alignment constraint.
8386 unsigned RequiredAlignment = TLI.getDataLayout()->getABITypeAlignment(
8387 ResVT.getTypeForEVT(*DAG->getContext()));
8389 if (RequiredAlignment > getAlignment())
8392 // 2. Check that the load is a legal operation for that type.
8393 if (!TLI.isOperationLegal(ISD::LOAD, ResVT))
8396 // 3. Check that we do not have a zext in the way.
8397 if (Inst->getValueType(0) != getLoadedType())
8405 /// \brief Check that all bits set in \p UsedBits form a dense region, i.e.,
8406 /// \p UsedBits looks like 0..0 1..1 0..0.
8407 static bool areUsedBitsDense(const APInt &UsedBits) {
8408 // If all the bits are one, this is dense!
8409 if (UsedBits.isAllOnesValue())
8412 // Get rid of the unused bits on the right.
8413 APInt NarrowedUsedBits = UsedBits.lshr(UsedBits.countTrailingZeros());
8414 // Get rid of the unused bits on the left.
8415 if (NarrowedUsedBits.countLeadingZeros())
8416 NarrowedUsedBits = NarrowedUsedBits.trunc(NarrowedUsedBits.getActiveBits());
8417 // Check that the chunk of bits is completely used.
8418 return NarrowedUsedBits.isAllOnesValue();
8421 /// \brief Check whether or not \p First and \p Second are next to each other
8422 /// in memory. This means that there is no hole between the bits loaded
8423 /// by \p First and the bits loaded by \p Second.
8424 static bool areSlicesNextToEachOther(const LoadedSlice &First,
8425 const LoadedSlice &Second) {
8426 assert(First.Origin == Second.Origin && First.Origin &&
8427 "Unable to match different memory origins.");
8428 APInt UsedBits = First.getUsedBits();
8429 assert((UsedBits & Second.getUsedBits()) == 0 &&
8430 "Slices are not supposed to overlap.");
8431 UsedBits |= Second.getUsedBits();
8432 return areUsedBitsDense(UsedBits);
8435 /// \brief Adjust the \p GlobalLSCost according to the target
8436 /// paring capabilities and the layout of the slices.
8437 /// \pre \p GlobalLSCost should account for at least as many loads as
8438 /// there is in the slices in \p LoadedSlices.
8439 static void adjustCostForPairing(SmallVectorImpl<LoadedSlice> &LoadedSlices,
8440 LoadedSlice::Cost &GlobalLSCost) {
8441 unsigned NumberOfSlices = LoadedSlices.size();
8442 // If there is less than 2 elements, no pairing is possible.
8443 if (NumberOfSlices < 2)
8446 // Sort the slices so that elements that are likely to be next to each
8447 // other in memory are next to each other in the list.
8448 std::sort(LoadedSlices.begin(), LoadedSlices.end(),
8449 [](const LoadedSlice &LHS, const LoadedSlice &RHS) {
8450 assert(LHS.Origin == RHS.Origin && "Different bases not implemented.");
8451 return LHS.getOffsetFromBase() < RHS.getOffsetFromBase();
8453 const TargetLowering &TLI = LoadedSlices[0].DAG->getTargetLoweringInfo();
8454 // First (resp. Second) is the first (resp. Second) potentially candidate
8455 // to be placed in a paired load.
8456 const LoadedSlice *First = nullptr;
8457 const LoadedSlice *Second = nullptr;
8458 for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice,
8459 // Set the beginning of the pair.
8462 Second = &LoadedSlices[CurrSlice];
8464 // If First is NULL, it means we start a new pair.
8465 // Get to the next slice.
8469 EVT LoadedType = First->getLoadedType();
8471 // If the types of the slices are different, we cannot pair them.
8472 if (LoadedType != Second->getLoadedType())
8475 // Check if the target supplies paired loads for this type.
8476 unsigned RequiredAlignment = 0;
8477 if (!TLI.hasPairedLoad(LoadedType, RequiredAlignment)) {
8478 // move to the next pair, this type is hopeless.
8482 // Check if we meet the alignment requirement.
8483 if (RequiredAlignment > First->getAlignment())
8486 // Check that both loads are next to each other in memory.
8487 if (!areSlicesNextToEachOther(*First, *Second))
8490 assert(GlobalLSCost.Loads > 0 && "We save more loads than we created!");
8491 --GlobalLSCost.Loads;
8492 // Move to the next pair.
8497 /// \brief Check the profitability of all involved LoadedSlice.
8498 /// Currently, it is considered profitable if there is exactly two
8499 /// involved slices (1) which are (2) next to each other in memory, and
8500 /// whose cost (\see LoadedSlice::Cost) is smaller than the original load (3).
8502 /// Note: The order of the elements in \p LoadedSlices may be modified, but not
8503 /// the elements themselves.
8505 /// FIXME: When the cost model will be mature enough, we can relax
8506 /// constraints (1) and (2).
8507 static bool isSlicingProfitable(SmallVectorImpl<LoadedSlice> &LoadedSlices,
8508 const APInt &UsedBits, bool ForCodeSize) {
8509 unsigned NumberOfSlices = LoadedSlices.size();
8510 if (StressLoadSlicing)
8511 return NumberOfSlices > 1;
8514 if (NumberOfSlices != 2)
8518 if (!areUsedBitsDense(UsedBits))
8522 LoadedSlice::Cost OrigCost(ForCodeSize), GlobalSlicingCost(ForCodeSize);
8523 // The original code has one big load.
8525 for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice) {
8526 const LoadedSlice &LS = LoadedSlices[CurrSlice];
8527 // Accumulate the cost of all the slices.
8528 LoadedSlice::Cost SliceCost(LS, ForCodeSize);
8529 GlobalSlicingCost += SliceCost;
8531 // Account as cost in the original configuration the gain obtained
8532 // with the current slices.
8533 OrigCost.addSliceGain(LS);
8536 // If the target supports paired load, adjust the cost accordingly.
8537 adjustCostForPairing(LoadedSlices, GlobalSlicingCost);
8538 return OrigCost > GlobalSlicingCost;
8541 /// \brief If the given load, \p LI, is used only by trunc or trunc(lshr)
8542 /// operations, split it in the various pieces being extracted.
8544 /// This sort of thing is introduced by SROA.
8545 /// This slicing takes care not to insert overlapping loads.
8546 /// \pre LI is a simple load (i.e., not an atomic or volatile load).
8547 bool DAGCombiner::SliceUpLoad(SDNode *N) {
8548 if (Level < AfterLegalizeDAG)
8551 LoadSDNode *LD = cast<LoadSDNode>(N);
8552 if (LD->isVolatile() || !ISD::isNormalLoad(LD) ||
8553 !LD->getValueType(0).isInteger())
8556 // Keep track of already used bits to detect overlapping values.
8557 // In that case, we will just abort the transformation.
8558 APInt UsedBits(LD->getValueSizeInBits(0), 0);
8560 SmallVector<LoadedSlice, 4> LoadedSlices;
8562 // Check if this load is used as several smaller chunks of bits.
8563 // Basically, look for uses in trunc or trunc(lshr) and record a new chain
8564 // of computation for each trunc.
8565 for (SDNode::use_iterator UI = LD->use_begin(), UIEnd = LD->use_end();
8566 UI != UIEnd; ++UI) {
8567 // Skip the uses of the chain.
8568 if (UI.getUse().getResNo() != 0)
8574 // Check if this is a trunc(lshr).
8575 if (User->getOpcode() == ISD::SRL && User->hasOneUse() &&
8576 isa<ConstantSDNode>(User->getOperand(1))) {
8577 Shift = cast<ConstantSDNode>(User->getOperand(1))->getZExtValue();
8578 User = *User->use_begin();
8581 // At this point, User is a Truncate, iff we encountered, trunc or
8583 if (User->getOpcode() != ISD::TRUNCATE)
8586 // The width of the type must be a power of 2 and greater than 8-bits.
8587 // Otherwise the load cannot be represented in LLVM IR.
8588 // Moreover, if we shifted with a non-8-bits multiple, the slice
8589 // will be across several bytes. We do not support that.
8590 unsigned Width = User->getValueSizeInBits(0);
8591 if (Width < 8 || !isPowerOf2_32(Width) || (Shift & 0x7))
8594 // Build the slice for this chain of computations.
8595 LoadedSlice LS(User, LD, Shift, &DAG);
8596 APInt CurrentUsedBits = LS.getUsedBits();
8598 // Check if this slice overlaps with another.
8599 if ((CurrentUsedBits & UsedBits) != 0)
8601 // Update the bits used globally.
8602 UsedBits |= CurrentUsedBits;
8604 // Check if the new slice would be legal.
8608 // Record the slice.
8609 LoadedSlices.push_back(LS);
8612 // Abort slicing if it does not seem to be profitable.
8613 if (!isSlicingProfitable(LoadedSlices, UsedBits, ForCodeSize))
8618 // Rewrite each chain to use an independent load.
8619 // By construction, each chain can be represented by a unique load.
8621 // Prepare the argument for the new token factor for all the slices.
8622 SmallVector<SDValue, 8> ArgChains;
8623 for (SmallVectorImpl<LoadedSlice>::const_iterator
8624 LSIt = LoadedSlices.begin(),
8625 LSItEnd = LoadedSlices.end();
8626 LSIt != LSItEnd; ++LSIt) {
8627 SDValue SliceInst = LSIt->loadSlice();
8628 CombineTo(LSIt->Inst, SliceInst, true);
8629 if (SliceInst.getNode()->getOpcode() != ISD::LOAD)
8630 SliceInst = SliceInst.getOperand(0);
8631 assert(SliceInst->getOpcode() == ISD::LOAD &&
8632 "It takes more than a zext to get to the loaded slice!!");
8633 ArgChains.push_back(SliceInst.getValue(1));
8636 SDValue Chain = DAG.getNode(ISD::TokenFactor, SDLoc(LD), MVT::Other,
8638 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
8642 /// CheckForMaskedLoad - Check to see if V is (and load (ptr), imm), where the
8643 /// load is having specific bytes cleared out. If so, return the byte size
8644 /// being masked out and the shift amount.
8645 static std::pair<unsigned, unsigned>
8646 CheckForMaskedLoad(SDValue V, SDValue Ptr, SDValue Chain) {
8647 std::pair<unsigned, unsigned> Result(0, 0);
8649 // Check for the structure we're looking for.
8650 if (V->getOpcode() != ISD::AND ||
8651 !isa<ConstantSDNode>(V->getOperand(1)) ||
8652 !ISD::isNormalLoad(V->getOperand(0).getNode()))
8655 // Check the chain and pointer.
8656 LoadSDNode *LD = cast<LoadSDNode>(V->getOperand(0));
8657 if (LD->getBasePtr() != Ptr) return Result; // Not from same pointer.
8659 // The store should be chained directly to the load or be an operand of a
8661 if (LD == Chain.getNode())
8663 else if (Chain->getOpcode() != ISD::TokenFactor)
8664 return Result; // Fail.
8667 for (unsigned i = 0, e = Chain->getNumOperands(); i != e; ++i)
8668 if (Chain->getOperand(i).getNode() == LD) {
8672 if (!isOk) return Result;
8675 // This only handles simple types.
8676 if (V.getValueType() != MVT::i16 &&
8677 V.getValueType() != MVT::i32 &&
8678 V.getValueType() != MVT::i64)
8681 // Check the constant mask. Invert it so that the bits being masked out are
8682 // 0 and the bits being kept are 1. Use getSExtValue so that leading bits
8683 // follow the sign bit for uniformity.
8684 uint64_t NotMask = ~cast<ConstantSDNode>(V->getOperand(1))->getSExtValue();
8685 unsigned NotMaskLZ = countLeadingZeros(NotMask);
8686 if (NotMaskLZ & 7) return Result; // Must be multiple of a byte.
8687 unsigned NotMaskTZ = countTrailingZeros(NotMask);
8688 if (NotMaskTZ & 7) return Result; // Must be multiple of a byte.
8689 if (NotMaskLZ == 64) return Result; // All zero mask.
8691 // See if we have a continuous run of bits. If so, we have 0*1+0*
8692 if (CountTrailingOnes_64(NotMask >> NotMaskTZ)+NotMaskTZ+NotMaskLZ != 64)
8695 // Adjust NotMaskLZ down to be from the actual size of the int instead of i64.
8696 if (V.getValueType() != MVT::i64 && NotMaskLZ)
8697 NotMaskLZ -= 64-V.getValueSizeInBits();
8699 unsigned MaskedBytes = (V.getValueSizeInBits()-NotMaskLZ-NotMaskTZ)/8;
8700 switch (MaskedBytes) {
8704 default: return Result; // All one mask, or 5-byte mask.
8707 // Verify that the first bit starts at a multiple of mask so that the access
8708 // is aligned the same as the access width.
8709 if (NotMaskTZ && NotMaskTZ/8 % MaskedBytes) return Result;
8711 Result.first = MaskedBytes;
8712 Result.second = NotMaskTZ/8;
8717 /// ShrinkLoadReplaceStoreWithStore - Check to see if IVal is something that
8718 /// provides a value as specified by MaskInfo. If so, replace the specified
8719 /// store with a narrower store of truncated IVal.
8721 ShrinkLoadReplaceStoreWithStore(const std::pair<unsigned, unsigned> &MaskInfo,
8722 SDValue IVal, StoreSDNode *St,
8724 unsigned NumBytes = MaskInfo.first;
8725 unsigned ByteShift = MaskInfo.second;
8726 SelectionDAG &DAG = DC->getDAG();
8728 // Check to see if IVal is all zeros in the part being masked in by the 'or'
8729 // that uses this. If not, this is not a replacement.
8730 APInt Mask = ~APInt::getBitsSet(IVal.getValueSizeInBits(),
8731 ByteShift*8, (ByteShift+NumBytes)*8);
8732 if (!DAG.MaskedValueIsZero(IVal, Mask)) return nullptr;
8734 // Check that it is legal on the target to do this. It is legal if the new
8735 // VT we're shrinking to (i8/i16/i32) is legal or we're still before type
8737 MVT VT = MVT::getIntegerVT(NumBytes*8);
8738 if (!DC->isTypeLegal(VT))
8741 // Okay, we can do this! Replace the 'St' store with a store of IVal that is
8742 // shifted by ByteShift and truncated down to NumBytes.
8744 IVal = DAG.getNode(ISD::SRL, SDLoc(IVal), IVal.getValueType(), IVal,
8745 DAG.getConstant(ByteShift*8,
8746 DC->getShiftAmountTy(IVal.getValueType())));
8748 // Figure out the offset for the store and the alignment of the access.
8750 unsigned NewAlign = St->getAlignment();
8752 if (DAG.getTargetLoweringInfo().isLittleEndian())
8753 StOffset = ByteShift;
8755 StOffset = IVal.getValueType().getStoreSize() - ByteShift - NumBytes;
8757 SDValue Ptr = St->getBasePtr();
8759 Ptr = DAG.getNode(ISD::ADD, SDLoc(IVal), Ptr.getValueType(),
8760 Ptr, DAG.getConstant(StOffset, Ptr.getValueType()));
8761 NewAlign = MinAlign(NewAlign, StOffset);
8764 // Truncate down to the new size.
8765 IVal = DAG.getNode(ISD::TRUNCATE, SDLoc(IVal), VT, IVal);
8768 return DAG.getStore(St->getChain(), SDLoc(St), IVal, Ptr,
8769 St->getPointerInfo().getWithOffset(StOffset),
8770 false, false, NewAlign).getNode();
8774 /// ReduceLoadOpStoreWidth - Look for sequence of load / op / store where op is
8775 /// one of 'or', 'xor', and 'and' of immediates. If 'op' is only touching some
8776 /// of the loaded bits, try narrowing the load and store if it would end up
8777 /// being a win for performance or code size.
8778 SDValue DAGCombiner::ReduceLoadOpStoreWidth(SDNode *N) {
8779 StoreSDNode *ST = cast<StoreSDNode>(N);
8780 if (ST->isVolatile())
8783 SDValue Chain = ST->getChain();
8784 SDValue Value = ST->getValue();
8785 SDValue Ptr = ST->getBasePtr();
8786 EVT VT = Value.getValueType();
8788 if (ST->isTruncatingStore() || VT.isVector() || !Value.hasOneUse())
8791 unsigned Opc = Value.getOpcode();
8793 // If this is "store (or X, Y), P" and X is "(and (load P), cst)", where cst
8794 // is a byte mask indicating a consecutive number of bytes, check to see if
8795 // Y is known to provide just those bytes. If so, we try to replace the
8796 // load + replace + store sequence with a single (narrower) store, which makes
8798 if (Opc == ISD::OR) {
8799 std::pair<unsigned, unsigned> MaskedLoad;
8800 MaskedLoad = CheckForMaskedLoad(Value.getOperand(0), Ptr, Chain);
8801 if (MaskedLoad.first)
8802 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
8803 Value.getOperand(1), ST,this))
8804 return SDValue(NewST, 0);
8806 // Or is commutative, so try swapping X and Y.
8807 MaskedLoad = CheckForMaskedLoad(Value.getOperand(1), Ptr, Chain);
8808 if (MaskedLoad.first)
8809 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
8810 Value.getOperand(0), ST,this))
8811 return SDValue(NewST, 0);
8814 if ((Opc != ISD::OR && Opc != ISD::XOR && Opc != ISD::AND) ||
8815 Value.getOperand(1).getOpcode() != ISD::Constant)
8818 SDValue N0 = Value.getOperand(0);
8819 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
8820 Chain == SDValue(N0.getNode(), 1)) {
8821 LoadSDNode *LD = cast<LoadSDNode>(N0);
8822 if (LD->getBasePtr() != Ptr ||
8823 LD->getPointerInfo().getAddrSpace() !=
8824 ST->getPointerInfo().getAddrSpace())
8827 // Find the type to narrow it the load / op / store to.
8828 SDValue N1 = Value.getOperand(1);
8829 unsigned BitWidth = N1.getValueSizeInBits();
8830 APInt Imm = cast<ConstantSDNode>(N1)->getAPIntValue();
8831 if (Opc == ISD::AND)
8832 Imm ^= APInt::getAllOnesValue(BitWidth);
8833 if (Imm == 0 || Imm.isAllOnesValue())
8835 unsigned ShAmt = Imm.countTrailingZeros();
8836 unsigned MSB = BitWidth - Imm.countLeadingZeros() - 1;
8837 unsigned NewBW = NextPowerOf2(MSB - ShAmt);
8838 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
8839 while (NewBW < BitWidth &&
8840 !(TLI.isOperationLegalOrCustom(Opc, NewVT) &&
8841 TLI.isNarrowingProfitable(VT, NewVT))) {
8842 NewBW = NextPowerOf2(NewBW);
8843 NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
8845 if (NewBW >= BitWidth)
8848 // If the lsb changed does not start at the type bitwidth boundary,
8849 // start at the previous one.
8851 ShAmt = (((ShAmt + NewBW - 1) / NewBW) * NewBW) - NewBW;
8852 APInt Mask = APInt::getBitsSet(BitWidth, ShAmt,
8853 std::min(BitWidth, ShAmt + NewBW));
8854 if ((Imm & Mask) == Imm) {
8855 APInt NewImm = (Imm & Mask).lshr(ShAmt).trunc(NewBW);
8856 if (Opc == ISD::AND)
8857 NewImm ^= APInt::getAllOnesValue(NewBW);
8858 uint64_t PtrOff = ShAmt / 8;
8859 // For big endian targets, we need to adjust the offset to the pointer to
8860 // load the correct bytes.
8861 if (TLI.isBigEndian())
8862 PtrOff = (BitWidth + 7 - NewBW) / 8 - PtrOff;
8864 unsigned NewAlign = MinAlign(LD->getAlignment(), PtrOff);
8865 Type *NewVTTy = NewVT.getTypeForEVT(*DAG.getContext());
8866 if (NewAlign < TLI.getDataLayout()->getABITypeAlignment(NewVTTy))
8869 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(LD),
8870 Ptr.getValueType(), Ptr,
8871 DAG.getConstant(PtrOff, Ptr.getValueType()));
8872 SDValue NewLD = DAG.getLoad(NewVT, SDLoc(N0),
8873 LD->getChain(), NewPtr,
8874 LD->getPointerInfo().getWithOffset(PtrOff),
8875 LD->isVolatile(), LD->isNonTemporal(),
8876 LD->isInvariant(), NewAlign,
8878 SDValue NewVal = DAG.getNode(Opc, SDLoc(Value), NewVT, NewLD,
8879 DAG.getConstant(NewImm, NewVT));
8880 SDValue NewST = DAG.getStore(Chain, SDLoc(N),
8882 ST->getPointerInfo().getWithOffset(PtrOff),
8883 false, false, NewAlign);
8885 AddToWorklist(NewPtr.getNode());
8886 AddToWorklist(NewLD.getNode());
8887 AddToWorklist(NewVal.getNode());
8888 WorklistRemover DeadNodes(*this);
8889 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLD.getValue(1));
8898 /// TransformFPLoadStorePair - For a given floating point load / store pair,
8899 /// if the load value isn't used by any other operations, then consider
8900 /// transforming the pair to integer load / store operations if the target
8901 /// deems the transformation profitable.
8902 SDValue DAGCombiner::TransformFPLoadStorePair(SDNode *N) {
8903 StoreSDNode *ST = cast<StoreSDNode>(N);
8904 SDValue Chain = ST->getChain();
8905 SDValue Value = ST->getValue();
8906 if (ISD::isNormalStore(ST) && ISD::isNormalLoad(Value.getNode()) &&
8907 Value.hasOneUse() &&
8908 Chain == SDValue(Value.getNode(), 1)) {
8909 LoadSDNode *LD = cast<LoadSDNode>(Value);
8910 EVT VT = LD->getMemoryVT();
8911 if (!VT.isFloatingPoint() ||
8912 VT != ST->getMemoryVT() ||
8913 LD->isNonTemporal() ||
8914 ST->isNonTemporal() ||
8915 LD->getPointerInfo().getAddrSpace() != 0 ||
8916 ST->getPointerInfo().getAddrSpace() != 0)
8919 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
8920 if (!TLI.isOperationLegal(ISD::LOAD, IntVT) ||
8921 !TLI.isOperationLegal(ISD::STORE, IntVT) ||
8922 !TLI.isDesirableToTransformToIntegerOp(ISD::LOAD, VT) ||
8923 !TLI.isDesirableToTransformToIntegerOp(ISD::STORE, VT))
8926 unsigned LDAlign = LD->getAlignment();
8927 unsigned STAlign = ST->getAlignment();
8928 Type *IntVTTy = IntVT.getTypeForEVT(*DAG.getContext());
8929 unsigned ABIAlign = TLI.getDataLayout()->getABITypeAlignment(IntVTTy);
8930 if (LDAlign < ABIAlign || STAlign < ABIAlign)
8933 SDValue NewLD = DAG.getLoad(IntVT, SDLoc(Value),
8934 LD->getChain(), LD->getBasePtr(),
8935 LD->getPointerInfo(),
8936 false, false, false, LDAlign);
8938 SDValue NewST = DAG.getStore(NewLD.getValue(1), SDLoc(N),
8939 NewLD, ST->getBasePtr(),
8940 ST->getPointerInfo(),
8941 false, false, STAlign);
8943 AddToWorklist(NewLD.getNode());
8944 AddToWorklist(NewST.getNode());
8945 WorklistRemover DeadNodes(*this);
8946 DAG.ReplaceAllUsesOfValueWith(Value.getValue(1), NewLD.getValue(1));
8954 /// Helper struct to parse and store a memory address as base + index + offset.
8955 /// We ignore sign extensions when it is safe to do so.
8956 /// The following two expressions are not equivalent. To differentiate we need
8957 /// to store whether there was a sign extension involved in the index
8959 /// (load (i64 add (i64 copyfromreg %c)
8960 /// (i64 signextend (add (i8 load %index)
8964 /// (load (i64 add (i64 copyfromreg %c)
8965 /// (i64 signextend (i32 add (i32 signextend (i8 load %index))
8967 struct BaseIndexOffset {
8971 bool IsIndexSignExt;
8973 BaseIndexOffset() : Offset(0), IsIndexSignExt(false) {}
8975 BaseIndexOffset(SDValue Base, SDValue Index, int64_t Offset,
8976 bool IsIndexSignExt) :
8977 Base(Base), Index(Index), Offset(Offset), IsIndexSignExt(IsIndexSignExt) {}
8979 bool equalBaseIndex(const BaseIndexOffset &Other) {
8980 return Other.Base == Base && Other.Index == Index &&
8981 Other.IsIndexSignExt == IsIndexSignExt;
8984 /// Parses tree in Ptr for base, index, offset addresses.
8985 static BaseIndexOffset match(SDValue Ptr) {
8986 bool IsIndexSignExt = false;
8988 // We only can pattern match BASE + INDEX + OFFSET. If Ptr is not an ADD
8989 // instruction, then it could be just the BASE or everything else we don't
8990 // know how to handle. Just use Ptr as BASE and give up.
8991 if (Ptr->getOpcode() != ISD::ADD)
8992 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
8994 // We know that we have at least an ADD instruction. Try to pattern match
8995 // the simple case of BASE + OFFSET.
8996 if (isa<ConstantSDNode>(Ptr->getOperand(1))) {
8997 int64_t Offset = cast<ConstantSDNode>(Ptr->getOperand(1))->getSExtValue();
8998 return BaseIndexOffset(Ptr->getOperand(0), SDValue(), Offset,
9002 // Inside a loop the current BASE pointer is calculated using an ADD and a
9003 // MUL instruction. In this case Ptr is the actual BASE pointer.
9004 // (i64 add (i64 %array_ptr)
9005 // (i64 mul (i64 %induction_var)
9006 // (i64 %element_size)))
9007 if (Ptr->getOperand(1)->getOpcode() == ISD::MUL)
9008 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
9010 // Look at Base + Index + Offset cases.
9011 SDValue Base = Ptr->getOperand(0);
9012 SDValue IndexOffset = Ptr->getOperand(1);
9014 // Skip signextends.
9015 if (IndexOffset->getOpcode() == ISD::SIGN_EXTEND) {
9016 IndexOffset = IndexOffset->getOperand(0);
9017 IsIndexSignExt = true;
9020 // Either the case of Base + Index (no offset) or something else.
9021 if (IndexOffset->getOpcode() != ISD::ADD)
9022 return BaseIndexOffset(Base, IndexOffset, 0, IsIndexSignExt);
9024 // Now we have the case of Base + Index + offset.
9025 SDValue Index = IndexOffset->getOperand(0);
9026 SDValue Offset = IndexOffset->getOperand(1);
9028 if (!isa<ConstantSDNode>(Offset))
9029 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
9031 // Ignore signextends.
9032 if (Index->getOpcode() == ISD::SIGN_EXTEND) {
9033 Index = Index->getOperand(0);
9034 IsIndexSignExt = true;
9035 } else IsIndexSignExt = false;
9037 int64_t Off = cast<ConstantSDNode>(Offset)->getSExtValue();
9038 return BaseIndexOffset(Base, Index, Off, IsIndexSignExt);
9042 /// Holds a pointer to an LSBaseSDNode as well as information on where it
9043 /// is located in a sequence of memory operations connected by a chain.
9045 MemOpLink (LSBaseSDNode *N, int64_t Offset, unsigned Seq):
9046 MemNode(N), OffsetFromBase(Offset), SequenceNum(Seq) { }
9047 // Ptr to the mem node.
9048 LSBaseSDNode *MemNode;
9049 // Offset from the base ptr.
9050 int64_t OffsetFromBase;
9051 // What is the sequence number of this mem node.
9052 // Lowest mem operand in the DAG starts at zero.
9053 unsigned SequenceNum;
9056 bool DAGCombiner::MergeConsecutiveStores(StoreSDNode* St) {
9057 EVT MemVT = St->getMemoryVT();
9058 int64_t ElementSizeBytes = MemVT.getSizeInBits()/8;
9059 bool NoVectors = DAG.getMachineFunction().getFunction()->getAttributes().
9060 hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat);
9062 // Don't merge vectors into wider inputs.
9063 if (MemVT.isVector() || !MemVT.isSimple())
9066 // Perform an early exit check. Do not bother looking at stored values that
9067 // are not constants or loads.
9068 SDValue StoredVal = St->getValue();
9069 bool IsLoadSrc = isa<LoadSDNode>(StoredVal);
9070 if (!isa<ConstantSDNode>(StoredVal) && !isa<ConstantFPSDNode>(StoredVal) &&
9074 // Only look at ends of store sequences.
9075 SDValue Chain = SDValue(St, 1);
9076 if (Chain->hasOneUse() && Chain->use_begin()->getOpcode() == ISD::STORE)
9079 // This holds the base pointer, index, and the offset in bytes from the base
9081 BaseIndexOffset BasePtr = BaseIndexOffset::match(St->getBasePtr());
9083 // We must have a base and an offset.
9084 if (!BasePtr.Base.getNode())
9087 // Do not handle stores to undef base pointers.
9088 if (BasePtr.Base.getOpcode() == ISD::UNDEF)
9091 // Save the LoadSDNodes that we find in the chain.
9092 // We need to make sure that these nodes do not interfere with
9093 // any of the store nodes.
9094 SmallVector<LSBaseSDNode*, 8> AliasLoadNodes;
9096 // Save the StoreSDNodes that we find in the chain.
9097 SmallVector<MemOpLink, 8> StoreNodes;
9099 // Walk up the chain and look for nodes with offsets from the same
9100 // base pointer. Stop when reaching an instruction with a different kind
9101 // or instruction which has a different base pointer.
9103 StoreSDNode *Index = St;
9105 // If the chain has more than one use, then we can't reorder the mem ops.
9106 if (Index != St && !SDValue(Index, 1)->hasOneUse())
9109 // Find the base pointer and offset for this memory node.
9110 BaseIndexOffset Ptr = BaseIndexOffset::match(Index->getBasePtr());
9112 // Check that the base pointer is the same as the original one.
9113 if (!Ptr.equalBaseIndex(BasePtr))
9116 // Check that the alignment is the same.
9117 if (Index->getAlignment() != St->getAlignment())
9120 // The memory operands must not be volatile.
9121 if (Index->isVolatile() || Index->isIndexed())
9125 if (StoreSDNode *St = dyn_cast<StoreSDNode>(Index))
9126 if (St->isTruncatingStore())
9129 // The stored memory type must be the same.
9130 if (Index->getMemoryVT() != MemVT)
9133 // We do not allow unaligned stores because we want to prevent overriding
9135 if (Index->getAlignment()*8 != MemVT.getSizeInBits())
9138 // We found a potential memory operand to merge.
9139 StoreNodes.push_back(MemOpLink(Index, Ptr.Offset, Seq++));
9141 // Find the next memory operand in the chain. If the next operand in the
9142 // chain is a store then move up and continue the scan with the next
9143 // memory operand. If the next operand is a load save it and use alias
9144 // information to check if it interferes with anything.
9145 SDNode *NextInChain = Index->getChain().getNode();
9147 if (StoreSDNode *STn = dyn_cast<StoreSDNode>(NextInChain)) {
9148 // We found a store node. Use it for the next iteration.
9151 } else if (LoadSDNode *Ldn = dyn_cast<LoadSDNode>(NextInChain)) {
9152 if (Ldn->isVolatile()) {
9157 // Save the load node for later. Continue the scan.
9158 AliasLoadNodes.push_back(Ldn);
9159 NextInChain = Ldn->getChain().getNode();
9168 // Check if there is anything to merge.
9169 if (StoreNodes.size() < 2)
9172 // Sort the memory operands according to their distance from the base pointer.
9173 std::sort(StoreNodes.begin(), StoreNodes.end(),
9174 [](MemOpLink LHS, MemOpLink RHS) {
9175 return LHS.OffsetFromBase < RHS.OffsetFromBase ||
9176 (LHS.OffsetFromBase == RHS.OffsetFromBase &&
9177 LHS.SequenceNum > RHS.SequenceNum);
9180 // Scan the memory operations on the chain and find the first non-consecutive
9181 // store memory address.
9182 unsigned LastConsecutiveStore = 0;
9183 int64_t StartAddress = StoreNodes[0].OffsetFromBase;
9184 for (unsigned i = 0, e = StoreNodes.size(); i < e; ++i) {
9186 // Check that the addresses are consecutive starting from the second
9187 // element in the list of stores.
9189 int64_t CurrAddress = StoreNodes[i].OffsetFromBase;
9190 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
9195 // Check if this store interferes with any of the loads that we found.
9196 for (unsigned ld = 0, lde = AliasLoadNodes.size(); ld < lde; ++ld)
9197 if (isAlias(AliasLoadNodes[ld], StoreNodes[i].MemNode)) {
9201 // We found a load that alias with this store. Stop the sequence.
9205 // Mark this node as useful.
9206 LastConsecutiveStore = i;
9209 // The node with the lowest store address.
9210 LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode;
9212 // Store the constants into memory as one consecutive store.
9214 unsigned LastLegalType = 0;
9215 unsigned LastLegalVectorType = 0;
9216 bool NonZero = false;
9217 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
9218 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
9219 SDValue StoredVal = St->getValue();
9221 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(StoredVal)) {
9222 NonZero |= !C->isNullValue();
9223 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(StoredVal)) {
9224 NonZero |= !C->getConstantFPValue()->isNullValue();
9230 // Find a legal type for the constant store.
9231 unsigned StoreBW = (i+1) * ElementSizeBytes * 8;
9232 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
9233 if (TLI.isTypeLegal(StoreTy))
9234 LastLegalType = i+1;
9235 // Or check whether a truncstore is legal.
9236 else if (TLI.getTypeAction(*DAG.getContext(), StoreTy) ==
9237 TargetLowering::TypePromoteInteger) {
9238 EVT LegalizedStoredValueTy =
9239 TLI.getTypeToTransformTo(*DAG.getContext(), StoredVal.getValueType());
9240 if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy))
9241 LastLegalType = i+1;
9244 // Find a legal type for the vector store.
9245 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
9246 if (TLI.isTypeLegal(Ty))
9247 LastLegalVectorType = i + 1;
9250 // We only use vectors if the constant is known to be zero and the
9251 // function is not marked with the noimplicitfloat attribute.
9252 if (NonZero || NoVectors)
9253 LastLegalVectorType = 0;
9255 // Check if we found a legal integer type to store.
9256 if (LastLegalType == 0 && LastLegalVectorType == 0)
9259 bool UseVector = (LastLegalVectorType > LastLegalType) && !NoVectors;
9260 unsigned NumElem = UseVector ? LastLegalVectorType : LastLegalType;
9262 // Make sure we have something to merge.
9266 unsigned EarliestNodeUsed = 0;
9267 for (unsigned i=0; i < NumElem; ++i) {
9268 // Find a chain for the new wide-store operand. Notice that some
9269 // of the store nodes that we found may not be selected for inclusion
9270 // in the wide store. The chain we use needs to be the chain of the
9271 // earliest store node which is *used* and replaced by the wide store.
9272 if (StoreNodes[i].SequenceNum > StoreNodes[EarliestNodeUsed].SequenceNum)
9273 EarliestNodeUsed = i;
9276 // The earliest Node in the DAG.
9277 LSBaseSDNode *EarliestOp = StoreNodes[EarliestNodeUsed].MemNode;
9278 SDLoc DL(StoreNodes[0].MemNode);
9282 // Find a legal type for the vector store.
9283 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
9284 assert(TLI.isTypeLegal(Ty) && "Illegal vector store");
9285 StoredVal = DAG.getConstant(0, Ty);
9287 unsigned StoreBW = NumElem * ElementSizeBytes * 8;
9288 APInt StoreInt(StoreBW, 0);
9290 // Construct a single integer constant which is made of the smaller
9292 bool IsLE = TLI.isLittleEndian();
9293 for (unsigned i = 0; i < NumElem ; ++i) {
9294 unsigned Idx = IsLE ?(NumElem - 1 - i) : i;
9295 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[Idx].MemNode);
9296 SDValue Val = St->getValue();
9297 StoreInt<<=ElementSizeBytes*8;
9298 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val)) {
9299 StoreInt|=C->getAPIntValue().zext(StoreBW);
9300 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Val)) {
9301 StoreInt|= C->getValueAPF().bitcastToAPInt().zext(StoreBW);
9303 assert(false && "Invalid constant element type");
9307 // Create the new Load and Store operations.
9308 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
9309 StoredVal = DAG.getConstant(StoreInt, StoreTy);
9312 SDValue NewStore = DAG.getStore(EarliestOp->getChain(), DL, StoredVal,
9313 FirstInChain->getBasePtr(),
9314 FirstInChain->getPointerInfo(),
9316 FirstInChain->getAlignment());
9318 // Replace the first store with the new store
9319 CombineTo(EarliestOp, NewStore);
9320 // Erase all other stores.
9321 for (unsigned i = 0; i < NumElem ; ++i) {
9322 if (StoreNodes[i].MemNode == EarliestOp)
9324 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
9325 // ReplaceAllUsesWith will replace all uses that existed when it was
9326 // called, but graph optimizations may cause new ones to appear. For
9327 // example, the case in pr14333 looks like
9329 // St's chain -> St -> another store -> X
9331 // And the only difference from St to the other store is the chain.
9332 // When we change it's chain to be St's chain they become identical,
9333 // get CSEed and the net result is that X is now a use of St.
9334 // Since we know that St is redundant, just iterate.
9335 while (!St->use_empty())
9336 DAG.ReplaceAllUsesWith(SDValue(St, 0), St->getChain());
9337 removeFromWorklist(St);
9344 // Below we handle the case of multiple consecutive stores that
9345 // come from multiple consecutive loads. We merge them into a single
9346 // wide load and a single wide store.
9348 // Look for load nodes which are used by the stored values.
9349 SmallVector<MemOpLink, 8> LoadNodes;
9351 // Find acceptable loads. Loads need to have the same chain (token factor),
9352 // must not be zext, volatile, indexed, and they must be consecutive.
9353 BaseIndexOffset LdBasePtr;
9354 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
9355 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
9356 LoadSDNode *Ld = dyn_cast<LoadSDNode>(St->getValue());
9359 // Loads must only have one use.
9360 if (!Ld->hasNUsesOfValue(1, 0))
9363 // Check that the alignment is the same as the stores.
9364 if (Ld->getAlignment() != St->getAlignment())
9367 // The memory operands must not be volatile.
9368 if (Ld->isVolatile() || Ld->isIndexed())
9371 // We do not accept ext loads.
9372 if (Ld->getExtensionType() != ISD::NON_EXTLOAD)
9375 // The stored memory type must be the same.
9376 if (Ld->getMemoryVT() != MemVT)
9379 BaseIndexOffset LdPtr = BaseIndexOffset::match(Ld->getBasePtr());
9380 // If this is not the first ptr that we check.
9381 if (LdBasePtr.Base.getNode()) {
9382 // The base ptr must be the same.
9383 if (!LdPtr.equalBaseIndex(LdBasePtr))
9386 // Check that all other base pointers are the same as this one.
9390 // We found a potential memory operand to merge.
9391 LoadNodes.push_back(MemOpLink(Ld, LdPtr.Offset, 0));
9394 if (LoadNodes.size() < 2)
9397 // Scan the memory operations on the chain and find the first non-consecutive
9398 // load memory address. These variables hold the index in the store node
9400 unsigned LastConsecutiveLoad = 0;
9401 // This variable refers to the size and not index in the array.
9402 unsigned LastLegalVectorType = 0;
9403 unsigned LastLegalIntegerType = 0;
9404 StartAddress = LoadNodes[0].OffsetFromBase;
9405 SDValue FirstChain = LoadNodes[0].MemNode->getChain();
9406 for (unsigned i = 1; i < LoadNodes.size(); ++i) {
9407 // All loads much share the same chain.
9408 if (LoadNodes[i].MemNode->getChain() != FirstChain)
9411 int64_t CurrAddress = LoadNodes[i].OffsetFromBase;
9412 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
9414 LastConsecutiveLoad = i;
9416 // Find a legal type for the vector store.
9417 EVT StoreTy = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
9418 if (TLI.isTypeLegal(StoreTy))
9419 LastLegalVectorType = i + 1;
9421 // Find a legal type for the integer store.
9422 unsigned StoreBW = (i+1) * ElementSizeBytes * 8;
9423 StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
9424 if (TLI.isTypeLegal(StoreTy))
9425 LastLegalIntegerType = i + 1;
9426 // Or check whether a truncstore and extload is legal.
9427 else if (TLI.getTypeAction(*DAG.getContext(), StoreTy) ==
9428 TargetLowering::TypePromoteInteger) {
9429 EVT LegalizedStoredValueTy =
9430 TLI.getTypeToTransformTo(*DAG.getContext(), StoreTy);
9431 if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy) &&
9432 TLI.isLoadExtLegal(ISD::ZEXTLOAD, StoreTy) &&
9433 TLI.isLoadExtLegal(ISD::SEXTLOAD, StoreTy) &&
9434 TLI.isLoadExtLegal(ISD::EXTLOAD, StoreTy))
9435 LastLegalIntegerType = i+1;
9439 // Only use vector types if the vector type is larger than the integer type.
9440 // If they are the same, use integers.
9441 bool UseVectorTy = LastLegalVectorType > LastLegalIntegerType && !NoVectors;
9442 unsigned LastLegalType = std::max(LastLegalVectorType, LastLegalIntegerType);
9444 // We add +1 here because the LastXXX variables refer to location while
9445 // the NumElem refers to array/index size.
9446 unsigned NumElem = std::min(LastConsecutiveStore, LastConsecutiveLoad) + 1;
9447 NumElem = std::min(LastLegalType, NumElem);
9452 // The earliest Node in the DAG.
9453 unsigned EarliestNodeUsed = 0;
9454 LSBaseSDNode *EarliestOp = StoreNodes[EarliestNodeUsed].MemNode;
9455 for (unsigned i=1; i<NumElem; ++i) {
9456 // Find a chain for the new wide-store operand. Notice that some
9457 // of the store nodes that we found may not be selected for inclusion
9458 // in the wide store. The chain we use needs to be the chain of the
9459 // earliest store node which is *used* and replaced by the wide store.
9460 if (StoreNodes[i].SequenceNum > StoreNodes[EarliestNodeUsed].SequenceNum)
9461 EarliestNodeUsed = i;
9464 // Find if it is better to use vectors or integers to load and store
9468 JointMemOpVT = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
9470 unsigned StoreBW = NumElem * ElementSizeBytes * 8;
9471 JointMemOpVT = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
9474 SDLoc LoadDL(LoadNodes[0].MemNode);
9475 SDLoc StoreDL(StoreNodes[0].MemNode);
9477 LoadSDNode *FirstLoad = cast<LoadSDNode>(LoadNodes[0].MemNode);
9478 SDValue NewLoad = DAG.getLoad(JointMemOpVT, LoadDL,
9479 FirstLoad->getChain(),
9480 FirstLoad->getBasePtr(),
9481 FirstLoad->getPointerInfo(),
9482 false, false, false,
9483 FirstLoad->getAlignment());
9485 SDValue NewStore = DAG.getStore(EarliestOp->getChain(), StoreDL, NewLoad,
9486 FirstInChain->getBasePtr(),
9487 FirstInChain->getPointerInfo(), false, false,
9488 FirstInChain->getAlignment());
9490 // Replace one of the loads with the new load.
9491 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[0].MemNode);
9492 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1),
9493 SDValue(NewLoad.getNode(), 1));
9495 // Remove the rest of the load chains.
9496 for (unsigned i = 1; i < NumElem ; ++i) {
9497 // Replace all chain users of the old load nodes with the chain of the new
9499 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[i].MemNode);
9500 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Ld->getChain());
9503 // Replace the first store with the new store.
9504 CombineTo(EarliestOp, NewStore);
9505 // Erase all other stores.
9506 for (unsigned i = 0; i < NumElem ; ++i) {
9507 // Remove all Store nodes.
9508 if (StoreNodes[i].MemNode == EarliestOp)
9510 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
9511 DAG.ReplaceAllUsesOfValueWith(SDValue(St, 0), St->getChain());
9512 removeFromWorklist(St);
9519 SDValue DAGCombiner::visitSTORE(SDNode *N) {
9520 StoreSDNode *ST = cast<StoreSDNode>(N);
9521 SDValue Chain = ST->getChain();
9522 SDValue Value = ST->getValue();
9523 SDValue Ptr = ST->getBasePtr();
9525 // If this is a store of a bit convert, store the input value if the
9526 // resultant store does not need a higher alignment than the original.
9527 if (Value.getOpcode() == ISD::BITCAST && !ST->isTruncatingStore() &&
9528 ST->isUnindexed()) {
9529 unsigned OrigAlign = ST->getAlignment();
9530 EVT SVT = Value.getOperand(0).getValueType();
9531 unsigned Align = TLI.getDataLayout()->
9532 getABITypeAlignment(SVT.getTypeForEVT(*DAG.getContext()));
9533 if (Align <= OrigAlign &&
9534 ((!LegalOperations && !ST->isVolatile()) ||
9535 TLI.isOperationLegalOrCustom(ISD::STORE, SVT)))
9536 return DAG.getStore(Chain, SDLoc(N), Value.getOperand(0),
9537 Ptr, ST->getPointerInfo(), ST->isVolatile(),
9538 ST->isNonTemporal(), OrigAlign,
9542 // Turn 'store undef, Ptr' -> nothing.
9543 if (Value.getOpcode() == ISD::UNDEF && ST->isUnindexed())
9546 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
9547 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Value)) {
9548 // NOTE: If the original store is volatile, this transform must not increase
9549 // the number of stores. For example, on x86-32 an f64 can be stored in one
9550 // processor operation but an i64 (which is not legal) requires two. So the
9551 // transform should not be done in this case.
9552 if (Value.getOpcode() != ISD::TargetConstantFP) {
9554 switch (CFP->getSimpleValueType(0).SimpleTy) {
9555 default: llvm_unreachable("Unknown FP type");
9556 case MVT::f16: // We don't do this for these yet.
9562 if ((isTypeLegal(MVT::i32) && !LegalOperations && !ST->isVolatile()) ||
9563 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
9564 Tmp = DAG.getConstant((uint32_t)CFP->getValueAPF().
9565 bitcastToAPInt().getZExtValue(), MVT::i32);
9566 return DAG.getStore(Chain, SDLoc(N), Tmp,
9567 Ptr, ST->getMemOperand());
9571 if ((TLI.isTypeLegal(MVT::i64) && !LegalOperations &&
9572 !ST->isVolatile()) ||
9573 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i64)) {
9574 Tmp = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
9575 getZExtValue(), MVT::i64);
9576 return DAG.getStore(Chain, SDLoc(N), Tmp,
9577 Ptr, ST->getMemOperand());
9580 if (!ST->isVolatile() &&
9581 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
9582 // Many FP stores are not made apparent until after legalize, e.g. for
9583 // argument passing. Since this is so common, custom legalize the
9584 // 64-bit integer store into two 32-bit stores.
9585 uint64_t Val = CFP->getValueAPF().bitcastToAPInt().getZExtValue();
9586 SDValue Lo = DAG.getConstant(Val & 0xFFFFFFFF, MVT::i32);
9587 SDValue Hi = DAG.getConstant(Val >> 32, MVT::i32);
9588 if (TLI.isBigEndian()) std::swap(Lo, Hi);
9590 unsigned Alignment = ST->getAlignment();
9591 bool isVolatile = ST->isVolatile();
9592 bool isNonTemporal = ST->isNonTemporal();
9593 AAMDNodes AAInfo = ST->getAAInfo();
9595 SDValue St0 = DAG.getStore(Chain, SDLoc(ST), Lo,
9596 Ptr, ST->getPointerInfo(),
9597 isVolatile, isNonTemporal,
9598 ST->getAlignment(), AAInfo);
9599 Ptr = DAG.getNode(ISD::ADD, SDLoc(N), Ptr.getValueType(), Ptr,
9600 DAG.getConstant(4, Ptr.getValueType()));
9601 Alignment = MinAlign(Alignment, 4U);
9602 SDValue St1 = DAG.getStore(Chain, SDLoc(ST), Hi,
9603 Ptr, ST->getPointerInfo().getWithOffset(4),
9604 isVolatile, isNonTemporal,
9606 return DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other,
9615 // Try to infer better alignment information than the store already has.
9616 if (OptLevel != CodeGenOpt::None && ST->isUnindexed()) {
9617 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
9618 if (Align > ST->getAlignment())
9619 return DAG.getTruncStore(Chain, SDLoc(N), Value,
9620 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
9621 ST->isVolatile(), ST->isNonTemporal(), Align,
9626 // Try transforming a pair floating point load / store ops to integer
9627 // load / store ops.
9628 SDValue NewST = TransformFPLoadStorePair(N);
9629 if (NewST.getNode())
9632 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA :
9633 TLI.getTargetMachine().getSubtarget<TargetSubtargetInfo>().useAA();
9635 if (CombinerAAOnlyFunc.getNumOccurrences() &&
9636 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
9639 if (UseAA && ST->isUnindexed()) {
9640 // Walk up chain skipping non-aliasing memory nodes.
9641 SDValue BetterChain = FindBetterChain(N, Chain);
9643 // If there is a better chain.
9644 if (Chain != BetterChain) {
9647 // Replace the chain to avoid dependency.
9648 if (ST->isTruncatingStore()) {
9649 ReplStore = DAG.getTruncStore(BetterChain, SDLoc(N), Value, Ptr,
9650 ST->getMemoryVT(), ST->getMemOperand());
9652 ReplStore = DAG.getStore(BetterChain, SDLoc(N), Value, Ptr,
9653 ST->getMemOperand());
9656 // Create token to keep both nodes around.
9657 SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N),
9658 MVT::Other, Chain, ReplStore);
9660 // Make sure the new and old chains are cleaned up.
9661 AddToWorklist(Token.getNode());
9663 // Don't add users to work list.
9664 return CombineTo(N, Token, false);
9668 // Try transforming N to an indexed store.
9669 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
9670 return SDValue(N, 0);
9672 // FIXME: is there such a thing as a truncating indexed store?
9673 if (ST->isTruncatingStore() && ST->isUnindexed() &&
9674 Value.getValueType().isInteger()) {
9675 // See if we can simplify the input to this truncstore with knowledge that
9676 // only the low bits are being used. For example:
9677 // "truncstore (or (shl x, 8), y), i8" -> "truncstore y, i8"
9679 GetDemandedBits(Value,
9680 APInt::getLowBitsSet(
9681 Value.getValueType().getScalarType().getSizeInBits(),
9682 ST->getMemoryVT().getScalarType().getSizeInBits()));
9683 AddToWorklist(Value.getNode());
9684 if (Shorter.getNode())
9685 return DAG.getTruncStore(Chain, SDLoc(N), Shorter,
9686 Ptr, ST->getMemoryVT(), ST->getMemOperand());
9688 // Otherwise, see if we can simplify the operation with
9689 // SimplifyDemandedBits, which only works if the value has a single use.
9690 if (SimplifyDemandedBits(Value,
9691 APInt::getLowBitsSet(
9692 Value.getValueType().getScalarType().getSizeInBits(),
9693 ST->getMemoryVT().getScalarType().getSizeInBits())))
9694 return SDValue(N, 0);
9697 // If this is a load followed by a store to the same location, then the store
9699 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Value)) {
9700 if (Ld->getBasePtr() == Ptr && ST->getMemoryVT() == Ld->getMemoryVT() &&
9701 ST->isUnindexed() && !ST->isVolatile() &&
9702 // There can't be any side effects between the load and store, such as
9704 Chain.reachesChainWithoutSideEffects(SDValue(Ld, 1))) {
9705 // The store is dead, remove it.
9710 // If this is an FP_ROUND or TRUNC followed by a store, fold this into a
9711 // truncating store. We can do this even if this is already a truncstore.
9712 if ((Value.getOpcode() == ISD::FP_ROUND || Value.getOpcode() == ISD::TRUNCATE)
9713 && Value.getNode()->hasOneUse() && ST->isUnindexed() &&
9714 TLI.isTruncStoreLegal(Value.getOperand(0).getValueType(),
9715 ST->getMemoryVT())) {
9716 return DAG.getTruncStore(Chain, SDLoc(N), Value.getOperand(0),
9717 Ptr, ST->getMemoryVT(), ST->getMemOperand());
9720 // Only perform this optimization before the types are legal, because we
9721 // don't want to perform this optimization on every DAGCombine invocation.
9723 bool EverChanged = false;
9726 // There can be multiple store sequences on the same chain.
9727 // Keep trying to merge store sequences until we are unable to do so
9728 // or until we merge the last store on the chain.
9729 bool Changed = MergeConsecutiveStores(ST);
9730 EverChanged |= Changed;
9731 if (!Changed) break;
9732 } while (ST->getOpcode() != ISD::DELETED_NODE);
9735 return SDValue(N, 0);
9738 return ReduceLoadOpStoreWidth(N);
9741 SDValue DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) {
9742 SDValue InVec = N->getOperand(0);
9743 SDValue InVal = N->getOperand(1);
9744 SDValue EltNo = N->getOperand(2);
9747 // If the inserted element is an UNDEF, just use the input vector.
9748 if (InVal.getOpcode() == ISD::UNDEF)
9751 EVT VT = InVec.getValueType();
9753 // If we can't generate a legal BUILD_VECTOR, exit
9754 if (LegalOperations && !TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
9757 // Check that we know which element is being inserted
9758 if (!isa<ConstantSDNode>(EltNo))
9760 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
9762 // Canonicalize insert_vector_elt dag nodes.
9764 // (insert_vector_elt (insert_vector_elt A, Idx0), Idx1)
9765 // -> (insert_vector_elt (insert_vector_elt A, Idx1), Idx0)
9767 // Do this only if the child insert_vector node has one use; also
9768 // do this only if indices are both constants and Idx1 < Idx0.
9769 if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT && InVec.hasOneUse()
9770 && isa<ConstantSDNode>(InVec.getOperand(2))) {
9772 cast<ConstantSDNode>(InVec.getOperand(2))->getZExtValue();
9773 if (Elt < OtherElt) {
9775 SDValue NewOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(N), VT,
9776 InVec.getOperand(0), InVal, EltNo);
9777 AddToWorklist(NewOp.getNode());
9778 return DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(InVec.getNode()),
9779 VT, NewOp, InVec.getOperand(1), InVec.getOperand(2));
9783 // Check that the operand is a BUILD_VECTOR (or UNDEF, which can essentially
9784 // be converted to a BUILD_VECTOR). Fill in the Ops vector with the
9786 SmallVector<SDValue, 8> Ops;
9787 // Do not combine these two vectors if the output vector will not replace
9788 // the input vector.
9789 if (InVec.getOpcode() == ISD::BUILD_VECTOR && InVec.hasOneUse()) {
9790 Ops.append(InVec.getNode()->op_begin(),
9791 InVec.getNode()->op_end());
9792 } else if (InVec.getOpcode() == ISD::UNDEF) {
9793 unsigned NElts = VT.getVectorNumElements();
9794 Ops.append(NElts, DAG.getUNDEF(InVal.getValueType()));
9799 // Insert the element
9800 if (Elt < Ops.size()) {
9801 // All the operands of BUILD_VECTOR must have the same type;
9802 // we enforce that here.
9803 EVT OpVT = Ops[0].getValueType();
9804 if (InVal.getValueType() != OpVT)
9805 InVal = OpVT.bitsGT(InVal.getValueType()) ?
9806 DAG.getNode(ISD::ANY_EXTEND, dl, OpVT, InVal) :
9807 DAG.getNode(ISD::TRUNCATE, dl, OpVT, InVal);
9811 // Return the new vector
9812 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
9815 SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) {
9816 // (vextract (scalar_to_vector val, 0) -> val
9817 SDValue InVec = N->getOperand(0);
9818 EVT VT = InVec.getValueType();
9819 EVT NVT = N->getValueType(0);
9821 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
9822 // Check if the result type doesn't match the inserted element type. A
9823 // SCALAR_TO_VECTOR may truncate the inserted element and the
9824 // EXTRACT_VECTOR_ELT may widen the extracted vector.
9825 SDValue InOp = InVec.getOperand(0);
9826 if (InOp.getValueType() != NVT) {
9827 assert(InOp.getValueType().isInteger() && NVT.isInteger());
9828 return DAG.getSExtOrTrunc(InOp, SDLoc(InVec), NVT);
9833 SDValue EltNo = N->getOperand(1);
9834 bool ConstEltNo = isa<ConstantSDNode>(EltNo);
9836 // Transform: (EXTRACT_VECTOR_ELT( VECTOR_SHUFFLE )) -> EXTRACT_VECTOR_ELT.
9837 // We only perform this optimization before the op legalization phase because
9838 // we may introduce new vector instructions which are not backed by TD
9839 // patterns. For example on AVX, extracting elements from a wide vector
9840 // without using extract_subvector. However, if we can find an underlying
9841 // scalar value, then we can always use that.
9842 if (InVec.getOpcode() == ISD::VECTOR_SHUFFLE
9844 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
9845 int NumElem = VT.getVectorNumElements();
9846 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(InVec);
9847 // Find the new index to extract from.
9848 int OrigElt = SVOp->getMaskElt(Elt);
9850 // Extracting an undef index is undef.
9852 return DAG.getUNDEF(NVT);
9854 // Select the right vector half to extract from.
9856 if (OrigElt < NumElem) {
9857 SVInVec = InVec->getOperand(0);
9859 SVInVec = InVec->getOperand(1);
9863 if (SVInVec.getOpcode() == ISD::BUILD_VECTOR) {
9864 SDValue InOp = SVInVec.getOperand(OrigElt);
9865 if (InOp.getValueType() != NVT) {
9866 assert(InOp.getValueType().isInteger() && NVT.isInteger());
9867 InOp = DAG.getSExtOrTrunc(InOp, SDLoc(SVInVec), NVT);
9873 // FIXME: We should handle recursing on other vector shuffles and
9874 // scalar_to_vector here as well.
9876 if (!LegalOperations) {
9877 EVT IndexTy = TLI.getVectorIdxTy();
9878 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(N), NVT,
9879 SVInVec, DAG.getConstant(OrigElt, IndexTy));
9883 // Perform only after legalization to ensure build_vector / vector_shuffle
9884 // optimizations have already been done.
9885 if (!LegalOperations) return SDValue();
9887 // (vextract (v4f32 load $addr), c) -> (f32 load $addr+c*size)
9888 // (vextract (v4f32 s2v (f32 load $addr)), c) -> (f32 load $addr+c*size)
9889 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), 0) -> (f32 load $addr)
9892 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
9893 bool NewLoad = false;
9894 bool BCNumEltsChanged = false;
9895 EVT ExtVT = VT.getVectorElementType();
9898 // If the result of load has to be truncated, then it's not necessarily
9900 if (NVT.bitsLT(LVT) && !TLI.isTruncateFree(LVT, NVT))
9903 if (InVec.getOpcode() == ISD::BITCAST) {
9904 // Don't duplicate a load with other uses.
9905 if (!InVec.hasOneUse())
9908 EVT BCVT = InVec.getOperand(0).getValueType();
9909 if (!BCVT.isVector() || ExtVT.bitsGT(BCVT.getVectorElementType()))
9911 if (VT.getVectorNumElements() != BCVT.getVectorNumElements())
9912 BCNumEltsChanged = true;
9913 InVec = InVec.getOperand(0);
9914 ExtVT = BCVT.getVectorElementType();
9918 LoadSDNode *LN0 = nullptr;
9919 const ShuffleVectorSDNode *SVN = nullptr;
9920 if (ISD::isNormalLoad(InVec.getNode())) {
9921 LN0 = cast<LoadSDNode>(InVec);
9922 } else if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR &&
9923 InVec.getOperand(0).getValueType() == ExtVT &&
9924 ISD::isNormalLoad(InVec.getOperand(0).getNode())) {
9925 // Don't duplicate a load with other uses.
9926 if (!InVec.hasOneUse())
9929 LN0 = cast<LoadSDNode>(InVec.getOperand(0));
9930 } else if ((SVN = dyn_cast<ShuffleVectorSDNode>(InVec))) {
9931 // (vextract (vector_shuffle (load $addr), v2, <1, u, u, u>), 1)
9933 // (load $addr+1*size)
9935 // Don't duplicate a load with other uses.
9936 if (!InVec.hasOneUse())
9939 // If the bit convert changed the number of elements, it is unsafe
9940 // to examine the mask.
9941 if (BCNumEltsChanged)
9944 // Select the input vector, guarding against out of range extract vector.
9945 unsigned NumElems = VT.getVectorNumElements();
9946 int Idx = (Elt > (int)NumElems) ? -1 : SVN->getMaskElt(Elt);
9947 InVec = (Idx < (int)NumElems) ? InVec.getOperand(0) : InVec.getOperand(1);
9949 if (InVec.getOpcode() == ISD::BITCAST) {
9950 // Don't duplicate a load with other uses.
9951 if (!InVec.hasOneUse())
9954 InVec = InVec.getOperand(0);
9956 if (ISD::isNormalLoad(InVec.getNode())) {
9957 LN0 = cast<LoadSDNode>(InVec);
9958 Elt = (Idx < (int)NumElems) ? Idx : Idx - (int)NumElems;
9962 // Make sure we found a non-volatile load and the extractelement is
9964 if (!LN0 || !LN0->hasNUsesOfValue(1,0) || LN0->isVolatile())
9967 // If Idx was -1 above, Elt is going to be -1, so just return undef.
9969 return DAG.getUNDEF(LVT);
9971 unsigned Align = LN0->getAlignment();
9973 // Check the resultant load doesn't need a higher alignment than the
9977 ->getABITypeAlignment(LVT.getTypeForEVT(*DAG.getContext()));
9979 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, LVT))
9985 SDValue NewPtr = LN0->getBasePtr();
9986 unsigned PtrOff = 0;
9989 PtrOff = LVT.getSizeInBits() * Elt / 8;
9990 EVT PtrType = NewPtr.getValueType();
9991 if (TLI.isBigEndian())
9992 PtrOff = VT.getSizeInBits() / 8 - PtrOff;
9993 NewPtr = DAG.getNode(ISD::ADD, SDLoc(N), PtrType, NewPtr,
9994 DAG.getConstant(PtrOff, PtrType));
9997 // The replacement we need to do here is a little tricky: we need to
9998 // replace an extractelement of a load with a load.
9999 // Use ReplaceAllUsesOfValuesWith to do the replacement.
10000 // Note that this replacement assumes that the extractvalue is the only
10001 // use of the load; that's okay because we don't want to perform this
10002 // transformation in other cases anyway.
10005 if (NVT.bitsGT(LVT)) {
10006 // If the result type of vextract is wider than the load, then issue an
10007 // extending load instead.
10008 ISD::LoadExtType ExtType = TLI.isLoadExtLegal(ISD::ZEXTLOAD, LVT)
10009 ? ISD::ZEXTLOAD : ISD::EXTLOAD;
10010 Load = DAG.getExtLoad(ExtType, SDLoc(N), NVT, LN0->getChain(),
10011 NewPtr, LN0->getPointerInfo().getWithOffset(PtrOff),
10012 LVT, LN0->isVolatile(), LN0->isNonTemporal(),
10013 Align, LN0->getAAInfo());
10014 Chain = Load.getValue(1);
10016 Load = DAG.getLoad(LVT, SDLoc(N), LN0->getChain(), NewPtr,
10017 LN0->getPointerInfo().getWithOffset(PtrOff),
10018 LN0->isVolatile(), LN0->isNonTemporal(),
10019 LN0->isInvariant(), Align, LN0->getAAInfo());
10020 Chain = Load.getValue(1);
10021 if (NVT.bitsLT(LVT))
10022 Load = DAG.getNode(ISD::TRUNCATE, SDLoc(N), NVT, Load);
10024 Load = DAG.getNode(ISD::BITCAST, SDLoc(N), NVT, Load);
10026 WorklistRemover DeadNodes(*this);
10027 SDValue From[] = { SDValue(N, 0), SDValue(LN0,1) };
10028 SDValue To[] = { Load, Chain };
10029 DAG.ReplaceAllUsesOfValuesWith(From, To, 2);
10030 // Since we're explcitly calling ReplaceAllUses, add the new node to the
10031 // worklist explicitly as well.
10032 AddToWorklist(Load.getNode());
10033 AddUsersToWorklist(Load.getNode()); // Add users too
10034 // Make sure to revisit this node to clean it up; it will usually be dead.
10036 return SDValue(N, 0);
10042 // Simplify (build_vec (ext )) to (bitcast (build_vec ))
10043 SDValue DAGCombiner::reduceBuildVecExtToExtBuildVec(SDNode *N) {
10044 // We perform this optimization post type-legalization because
10045 // the type-legalizer often scalarizes integer-promoted vectors.
10046 // Performing this optimization before may create bit-casts which
10047 // will be type-legalized to complex code sequences.
10048 // We perform this optimization only before the operation legalizer because we
10049 // may introduce illegal operations.
10050 if (Level != AfterLegalizeVectorOps && Level != AfterLegalizeTypes)
10053 unsigned NumInScalars = N->getNumOperands();
10055 EVT VT = N->getValueType(0);
10057 // Check to see if this is a BUILD_VECTOR of a bunch of values
10058 // which come from any_extend or zero_extend nodes. If so, we can create
10059 // a new BUILD_VECTOR using bit-casts which may enable other BUILD_VECTOR
10060 // optimizations. We do not handle sign-extend because we can't fill the sign
10062 EVT SourceType = MVT::Other;
10063 bool AllAnyExt = true;
10065 for (unsigned i = 0; i != NumInScalars; ++i) {
10066 SDValue In = N->getOperand(i);
10067 // Ignore undef inputs.
10068 if (In.getOpcode() == ISD::UNDEF) continue;
10070 bool AnyExt = In.getOpcode() == ISD::ANY_EXTEND;
10071 bool ZeroExt = In.getOpcode() == ISD::ZERO_EXTEND;
10073 // Abort if the element is not an extension.
10074 if (!ZeroExt && !AnyExt) {
10075 SourceType = MVT::Other;
10079 // The input is a ZeroExt or AnyExt. Check the original type.
10080 EVT InTy = In.getOperand(0).getValueType();
10082 // Check that all of the widened source types are the same.
10083 if (SourceType == MVT::Other)
10086 else if (InTy != SourceType) {
10087 // Multiple income types. Abort.
10088 SourceType = MVT::Other;
10092 // Check if all of the extends are ANY_EXTENDs.
10093 AllAnyExt &= AnyExt;
10096 // In order to have valid types, all of the inputs must be extended from the
10097 // same source type and all of the inputs must be any or zero extend.
10098 // Scalar sizes must be a power of two.
10099 EVT OutScalarTy = VT.getScalarType();
10100 bool ValidTypes = SourceType != MVT::Other &&
10101 isPowerOf2_32(OutScalarTy.getSizeInBits()) &&
10102 isPowerOf2_32(SourceType.getSizeInBits());
10104 // Create a new simpler BUILD_VECTOR sequence which other optimizations can
10105 // turn into a single shuffle instruction.
10109 bool isLE = TLI.isLittleEndian();
10110 unsigned ElemRatio = OutScalarTy.getSizeInBits()/SourceType.getSizeInBits();
10111 assert(ElemRatio > 1 && "Invalid element size ratio");
10112 SDValue Filler = AllAnyExt ? DAG.getUNDEF(SourceType):
10113 DAG.getConstant(0, SourceType);
10115 unsigned NewBVElems = ElemRatio * VT.getVectorNumElements();
10116 SmallVector<SDValue, 8> Ops(NewBVElems, Filler);
10118 // Populate the new build_vector
10119 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
10120 SDValue Cast = N->getOperand(i);
10121 assert((Cast.getOpcode() == ISD::ANY_EXTEND ||
10122 Cast.getOpcode() == ISD::ZERO_EXTEND ||
10123 Cast.getOpcode() == ISD::UNDEF) && "Invalid cast opcode");
10125 if (Cast.getOpcode() == ISD::UNDEF)
10126 In = DAG.getUNDEF(SourceType);
10128 In = Cast->getOperand(0);
10129 unsigned Index = isLE ? (i * ElemRatio) :
10130 (i * ElemRatio + (ElemRatio - 1));
10132 assert(Index < Ops.size() && "Invalid index");
10136 // The type of the new BUILD_VECTOR node.
10137 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), SourceType, NewBVElems);
10138 assert(VecVT.getSizeInBits() == VT.getSizeInBits() &&
10139 "Invalid vector size");
10140 // Check if the new vector type is legal.
10141 if (!isTypeLegal(VecVT)) return SDValue();
10143 // Make the new BUILD_VECTOR.
10144 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
10146 // The new BUILD_VECTOR node has the potential to be further optimized.
10147 AddToWorklist(BV.getNode());
10148 // Bitcast to the desired type.
10149 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
10152 SDValue DAGCombiner::reduceBuildVecConvertToConvertBuildVec(SDNode *N) {
10153 EVT VT = N->getValueType(0);
10155 unsigned NumInScalars = N->getNumOperands();
10158 EVT SrcVT = MVT::Other;
10159 unsigned Opcode = ISD::DELETED_NODE;
10160 unsigned NumDefs = 0;
10162 for (unsigned i = 0; i != NumInScalars; ++i) {
10163 SDValue In = N->getOperand(i);
10164 unsigned Opc = In.getOpcode();
10166 if (Opc == ISD::UNDEF)
10169 // If all scalar values are floats and converted from integers.
10170 if (Opcode == ISD::DELETED_NODE &&
10171 (Opc == ISD::UINT_TO_FP || Opc == ISD::SINT_TO_FP)) {
10178 EVT InVT = In.getOperand(0).getValueType();
10180 // If all scalar values are typed differently, bail out. It's chosen to
10181 // simplify BUILD_VECTOR of integer types.
10182 if (SrcVT == MVT::Other)
10189 // If the vector has just one element defined, it's not worth to fold it into
10190 // a vectorized one.
10194 assert((Opcode == ISD::UINT_TO_FP || Opcode == ISD::SINT_TO_FP)
10195 && "Should only handle conversion from integer to float.");
10196 assert(SrcVT != MVT::Other && "Cannot determine source type!");
10198 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SrcVT, NumInScalars);
10200 if (!TLI.isOperationLegalOrCustom(Opcode, NVT))
10203 SmallVector<SDValue, 8> Opnds;
10204 for (unsigned i = 0; i != NumInScalars; ++i) {
10205 SDValue In = N->getOperand(i);
10207 if (In.getOpcode() == ISD::UNDEF)
10208 Opnds.push_back(DAG.getUNDEF(SrcVT));
10210 Opnds.push_back(In.getOperand(0));
10212 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NVT, Opnds);
10213 AddToWorklist(BV.getNode());
10215 return DAG.getNode(Opcode, dl, VT, BV);
10218 SDValue DAGCombiner::visitBUILD_VECTOR(SDNode *N) {
10219 unsigned NumInScalars = N->getNumOperands();
10221 EVT VT = N->getValueType(0);
10223 // A vector built entirely of undefs is undef.
10224 if (ISD::allOperandsUndef(N))
10225 return DAG.getUNDEF(VT);
10227 SDValue V = reduceBuildVecExtToExtBuildVec(N);
10231 V = reduceBuildVecConvertToConvertBuildVec(N);
10235 // Check to see if this is a BUILD_VECTOR of a bunch of EXTRACT_VECTOR_ELT
10236 // operations. If so, and if the EXTRACT_VECTOR_ELT vector inputs come from
10237 // at most two distinct vectors, turn this into a shuffle node.
10239 // May only combine to shuffle after legalize if shuffle is legal.
10240 if (LegalOperations &&
10241 !TLI.isOperationLegalOrCustom(ISD::VECTOR_SHUFFLE, VT))
10244 SDValue VecIn1, VecIn2;
10245 for (unsigned i = 0; i != NumInScalars; ++i) {
10246 // Ignore undef inputs.
10247 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
10249 // If this input is something other than a EXTRACT_VECTOR_ELT with a
10250 // constant index, bail out.
10251 if (N->getOperand(i).getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
10252 !isa<ConstantSDNode>(N->getOperand(i).getOperand(1))) {
10253 VecIn1 = VecIn2 = SDValue(nullptr, 0);
10257 // We allow up to two distinct input vectors.
10258 SDValue ExtractedFromVec = N->getOperand(i).getOperand(0);
10259 if (ExtractedFromVec == VecIn1 || ExtractedFromVec == VecIn2)
10262 if (!VecIn1.getNode()) {
10263 VecIn1 = ExtractedFromVec;
10264 } else if (!VecIn2.getNode()) {
10265 VecIn2 = ExtractedFromVec;
10267 // Too many inputs.
10268 VecIn1 = VecIn2 = SDValue(nullptr, 0);
10273 // If everything is good, we can make a shuffle operation.
10274 if (VecIn1.getNode()) {
10275 SmallVector<int, 8> Mask;
10276 for (unsigned i = 0; i != NumInScalars; ++i) {
10277 if (N->getOperand(i).getOpcode() == ISD::UNDEF) {
10278 Mask.push_back(-1);
10282 // If extracting from the first vector, just use the index directly.
10283 SDValue Extract = N->getOperand(i);
10284 SDValue ExtVal = Extract.getOperand(1);
10285 if (Extract.getOperand(0) == VecIn1) {
10286 unsigned ExtIndex = cast<ConstantSDNode>(ExtVal)->getZExtValue();
10287 if (ExtIndex > VT.getVectorNumElements())
10290 Mask.push_back(ExtIndex);
10294 // Otherwise, use InIdx + VecSize
10295 unsigned Idx = cast<ConstantSDNode>(ExtVal)->getZExtValue();
10296 Mask.push_back(Idx+NumInScalars);
10299 // We can't generate a shuffle node with mismatched input and output types.
10300 // Attempt to transform a single input vector to the correct type.
10301 if ((VT != VecIn1.getValueType())) {
10302 // We don't support shuffeling between TWO values of different types.
10303 if (VecIn2.getNode())
10306 // We only support widening of vectors which are half the size of the
10307 // output registers. For example XMM->YMM widening on X86 with AVX.
10308 if (VecIn1.getValueType().getSizeInBits()*2 != VT.getSizeInBits())
10311 // If the input vector type has a different base type to the output
10312 // vector type, bail out.
10313 if (VecIn1.getValueType().getVectorElementType() !=
10314 VT.getVectorElementType())
10317 // Widen the input vector by adding undef values.
10318 VecIn1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
10319 VecIn1, DAG.getUNDEF(VecIn1.getValueType()));
10322 // If VecIn2 is unused then change it to undef.
10323 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
10325 // Check that we were able to transform all incoming values to the same
10327 if (VecIn2.getValueType() != VecIn1.getValueType() ||
10328 VecIn1.getValueType() != VT)
10331 // Only type-legal BUILD_VECTOR nodes are converted to shuffle nodes.
10332 if (!isTypeLegal(VT))
10335 // Return the new VECTOR_SHUFFLE node.
10339 return DAG.getVectorShuffle(VT, dl, Ops[0], Ops[1], &Mask[0]);
10345 SDValue DAGCombiner::visitCONCAT_VECTORS(SDNode *N) {
10346 // TODO: Check to see if this is a CONCAT_VECTORS of a bunch of
10347 // EXTRACT_SUBVECTOR operations. If so, and if the EXTRACT_SUBVECTOR vector
10348 // inputs come from at most two distinct vectors, turn this into a shuffle
10351 // If we only have one input vector, we don't need to do any concatenation.
10352 if (N->getNumOperands() == 1)
10353 return N->getOperand(0);
10355 // Check if all of the operands are undefs.
10356 EVT VT = N->getValueType(0);
10357 if (ISD::allOperandsUndef(N))
10358 return DAG.getUNDEF(VT);
10360 // Optimize concat_vectors where one of the vectors is undef.
10361 if (N->getNumOperands() == 2 &&
10362 N->getOperand(1)->getOpcode() == ISD::UNDEF) {
10363 SDValue In = N->getOperand(0);
10364 assert(In.getValueType().isVector() && "Must concat vectors");
10366 // Transform: concat_vectors(scalar, undef) -> scalar_to_vector(sclr).
10367 if (In->getOpcode() == ISD::BITCAST &&
10368 !In->getOperand(0)->getValueType(0).isVector()) {
10369 SDValue Scalar = In->getOperand(0);
10370 EVT SclTy = Scalar->getValueType(0);
10372 if (!SclTy.isFloatingPoint() && !SclTy.isInteger())
10375 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SclTy,
10376 VT.getSizeInBits() / SclTy.getSizeInBits());
10377 if (!TLI.isTypeLegal(NVT) || !TLI.isTypeLegal(Scalar.getValueType()))
10380 SDLoc dl = SDLoc(N);
10381 SDValue Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, NVT, Scalar);
10382 return DAG.getNode(ISD::BITCAST, dl, VT, Res);
10386 // fold (concat_vectors (BUILD_VECTOR A, B, ...), (BUILD_VECTOR C, D, ...))
10387 // -> (BUILD_VECTOR A, B, ..., C, D, ...)
10388 if (N->getNumOperands() == 2 &&
10389 N->getOperand(0).getOpcode() == ISD::BUILD_VECTOR &&
10390 N->getOperand(1).getOpcode() == ISD::BUILD_VECTOR) {
10391 EVT VT = N->getValueType(0);
10392 SDValue N0 = N->getOperand(0);
10393 SDValue N1 = N->getOperand(1);
10394 SmallVector<SDValue, 8> Opnds;
10395 unsigned BuildVecNumElts = N0.getNumOperands();
10397 EVT SclTy0 = N0.getOperand(0)->getValueType(0);
10398 EVT SclTy1 = N1.getOperand(0)->getValueType(0);
10399 if (SclTy0.isFloatingPoint()) {
10400 for (unsigned i = 0; i != BuildVecNumElts; ++i)
10401 Opnds.push_back(N0.getOperand(i));
10402 for (unsigned i = 0; i != BuildVecNumElts; ++i)
10403 Opnds.push_back(N1.getOperand(i));
10405 // If BUILD_VECTOR are from built from integer, they may have different
10406 // operand types. Get the smaller type and truncate all operands to it.
10407 EVT MinTy = SclTy0.bitsLE(SclTy1) ? SclTy0 : SclTy1;
10408 for (unsigned i = 0; i != BuildVecNumElts; ++i)
10409 Opnds.push_back(DAG.getNode(ISD::TRUNCATE, SDLoc(N), MinTy,
10410 N0.getOperand(i)));
10411 for (unsigned i = 0; i != BuildVecNumElts; ++i)
10412 Opnds.push_back(DAG.getNode(ISD::TRUNCATE, SDLoc(N), MinTy,
10413 N1.getOperand(i)));
10416 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Opnds);
10419 // Type legalization of vectors and DAG canonicalization of SHUFFLE_VECTOR
10420 // nodes often generate nop CONCAT_VECTOR nodes.
10421 // Scan the CONCAT_VECTOR operands and look for a CONCAT operations that
10422 // place the incoming vectors at the exact same location.
10423 SDValue SingleSource = SDValue();
10424 unsigned PartNumElem = N->getOperand(0).getValueType().getVectorNumElements();
10426 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
10427 SDValue Op = N->getOperand(i);
10429 if (Op.getOpcode() == ISD::UNDEF)
10432 // Check if this is the identity extract:
10433 if (Op.getOpcode() != ISD::EXTRACT_SUBVECTOR)
10436 // Find the single incoming vector for the extract_subvector.
10437 if (SingleSource.getNode()) {
10438 if (Op.getOperand(0) != SingleSource)
10441 SingleSource = Op.getOperand(0);
10443 // Check the source type is the same as the type of the result.
10444 // If not, this concat may extend the vector, so we can not
10445 // optimize it away.
10446 if (SingleSource.getValueType() != N->getValueType(0))
10450 unsigned IdentityIndex = i * PartNumElem;
10451 ConstantSDNode *CS = dyn_cast<ConstantSDNode>(Op.getOperand(1));
10452 // The extract index must be constant.
10456 // Check that we are reading from the identity index.
10457 if (CS->getZExtValue() != IdentityIndex)
10461 if (SingleSource.getNode())
10462 return SingleSource;
10467 SDValue DAGCombiner::visitEXTRACT_SUBVECTOR(SDNode* N) {
10468 EVT NVT = N->getValueType(0);
10469 SDValue V = N->getOperand(0);
10471 if (V->getOpcode() == ISD::CONCAT_VECTORS) {
10473 // (extract_subvec (concat V1, V2, ...), i)
10476 // Only operand 0 is checked as 'concat' assumes all inputs of the same
10478 if (V->getOperand(0).getValueType() != NVT)
10480 unsigned Idx = dyn_cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
10481 unsigned NumElems = NVT.getVectorNumElements();
10482 assert((Idx % NumElems) == 0 &&
10483 "IDX in concat is not a multiple of the result vector length.");
10484 return V->getOperand(Idx / NumElems);
10488 if (V->getOpcode() == ISD::BITCAST)
10489 V = V.getOperand(0);
10491 if (V->getOpcode() == ISD::INSERT_SUBVECTOR) {
10493 // Handle only simple case where vector being inserted and vector
10494 // being extracted are of same type, and are half size of larger vectors.
10495 EVT BigVT = V->getOperand(0).getValueType();
10496 EVT SmallVT = V->getOperand(1).getValueType();
10497 if (!NVT.bitsEq(SmallVT) || NVT.getSizeInBits()*2 != BigVT.getSizeInBits())
10500 // Only handle cases where both indexes are constants with the same type.
10501 ConstantSDNode *ExtIdx = dyn_cast<ConstantSDNode>(N->getOperand(1));
10502 ConstantSDNode *InsIdx = dyn_cast<ConstantSDNode>(V->getOperand(2));
10504 if (InsIdx && ExtIdx &&
10505 InsIdx->getValueType(0).getSizeInBits() <= 64 &&
10506 ExtIdx->getValueType(0).getSizeInBits() <= 64) {
10508 // (extract_subvec (insert_subvec V1, V2, InsIdx), ExtIdx)
10510 // indices are equal or bit offsets are equal => V1
10511 // otherwise => (extract_subvec V1, ExtIdx)
10512 if (InsIdx->getZExtValue() * SmallVT.getScalarType().getSizeInBits() ==
10513 ExtIdx->getZExtValue() * NVT.getScalarType().getSizeInBits())
10514 return DAG.getNode(ISD::BITCAST, dl, NVT, V->getOperand(1));
10515 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, NVT,
10516 DAG.getNode(ISD::BITCAST, dl,
10517 N->getOperand(0).getValueType(),
10518 V->getOperand(0)), N->getOperand(1));
10525 // Tries to turn a shuffle of two CONCAT_VECTORS into a single concat.
10526 static SDValue partitionShuffleOfConcats(SDNode *N, SelectionDAG &DAG) {
10527 EVT VT = N->getValueType(0);
10528 unsigned NumElts = VT.getVectorNumElements();
10530 SDValue N0 = N->getOperand(0);
10531 SDValue N1 = N->getOperand(1);
10532 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
10534 SmallVector<SDValue, 4> Ops;
10535 EVT ConcatVT = N0.getOperand(0).getValueType();
10536 unsigned NumElemsPerConcat = ConcatVT.getVectorNumElements();
10537 unsigned NumConcats = NumElts / NumElemsPerConcat;
10539 // Look at every vector that's inserted. We're looking for exact
10540 // subvector-sized copies from a concatenated vector
10541 for (unsigned I = 0; I != NumConcats; ++I) {
10542 // Make sure we're dealing with a copy.
10543 unsigned Begin = I * NumElemsPerConcat;
10544 bool AllUndef = true, NoUndef = true;
10545 for (unsigned J = Begin; J != Begin + NumElemsPerConcat; ++J) {
10546 if (SVN->getMaskElt(J) >= 0)
10553 if (SVN->getMaskElt(Begin) % NumElemsPerConcat != 0)
10556 for (unsigned J = 1; J != NumElemsPerConcat; ++J)
10557 if (SVN->getMaskElt(Begin + J - 1) + 1 != SVN->getMaskElt(Begin + J))
10560 unsigned FirstElt = SVN->getMaskElt(Begin) / NumElemsPerConcat;
10561 if (FirstElt < N0.getNumOperands())
10562 Ops.push_back(N0.getOperand(FirstElt));
10564 Ops.push_back(N1.getOperand(FirstElt - N0.getNumOperands()));
10566 } else if (AllUndef) {
10567 Ops.push_back(DAG.getUNDEF(N0.getOperand(0).getValueType()));
10568 } else { // Mixed with general masks and undefs, can't do optimization.
10573 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Ops);
10576 SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
10577 EVT VT = N->getValueType(0);
10578 unsigned NumElts = VT.getVectorNumElements();
10580 SDValue N0 = N->getOperand(0);
10581 SDValue N1 = N->getOperand(1);
10583 assert(N0.getValueType() == VT && "Vector shuffle must be normalized in DAG");
10585 // Canonicalize shuffle undef, undef -> undef
10586 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
10587 return DAG.getUNDEF(VT);
10589 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
10591 // Canonicalize shuffle v, v -> v, undef
10593 SmallVector<int, 8> NewMask;
10594 for (unsigned i = 0; i != NumElts; ++i) {
10595 int Idx = SVN->getMaskElt(i);
10596 if (Idx >= (int)NumElts) Idx -= NumElts;
10597 NewMask.push_back(Idx);
10599 return DAG.getVectorShuffle(VT, SDLoc(N), N0, DAG.getUNDEF(VT),
10603 // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask.
10604 if (N0.getOpcode() == ISD::UNDEF) {
10605 SmallVector<int, 8> NewMask;
10606 for (unsigned i = 0; i != NumElts; ++i) {
10607 int Idx = SVN->getMaskElt(i);
10609 if (Idx >= (int)NumElts)
10612 Idx = -1; // remove reference to lhs
10614 NewMask.push_back(Idx);
10616 return DAG.getVectorShuffle(VT, SDLoc(N), N1, DAG.getUNDEF(VT),
10620 // Remove references to rhs if it is undef
10621 if (N1.getOpcode() == ISD::UNDEF) {
10622 bool Changed = false;
10623 SmallVector<int, 8> NewMask;
10624 for (unsigned i = 0; i != NumElts; ++i) {
10625 int Idx = SVN->getMaskElt(i);
10626 if (Idx >= (int)NumElts) {
10630 NewMask.push_back(Idx);
10633 return DAG.getVectorShuffle(VT, SDLoc(N), N0, N1, &NewMask[0]);
10636 // If it is a splat, check if the argument vector is another splat or a
10637 // build_vector with all scalar elements the same.
10638 if (SVN->isSplat() && SVN->getSplatIndex() < (int)NumElts) {
10639 SDNode *V = N0.getNode();
10641 // If this is a bit convert that changes the element type of the vector but
10642 // not the number of vector elements, look through it. Be careful not to
10643 // look though conversions that change things like v4f32 to v2f64.
10644 if (V->getOpcode() == ISD::BITCAST) {
10645 SDValue ConvInput = V->getOperand(0);
10646 if (ConvInput.getValueType().isVector() &&
10647 ConvInput.getValueType().getVectorNumElements() == NumElts)
10648 V = ConvInput.getNode();
10651 if (V->getOpcode() == ISD::BUILD_VECTOR) {
10652 assert(V->getNumOperands() == NumElts &&
10653 "BUILD_VECTOR has wrong number of operands");
10655 bool AllSame = true;
10656 for (unsigned i = 0; i != NumElts; ++i) {
10657 if (V->getOperand(i).getOpcode() != ISD::UNDEF) {
10658 Base = V->getOperand(i);
10662 // Splat of <u, u, u, u>, return <u, u, u, u>
10663 if (!Base.getNode())
10665 for (unsigned i = 0; i != NumElts; ++i) {
10666 if (V->getOperand(i) != Base) {
10671 // Splat of <x, x, x, x>, return <x, x, x, x>
10677 if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
10678 Level < AfterLegalizeVectorOps &&
10679 (N1.getOpcode() == ISD::UNDEF ||
10680 (N1.getOpcode() == ISD::CONCAT_VECTORS &&
10681 N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()))) {
10682 SDValue V = partitionShuffleOfConcats(N, DAG);
10688 // If this shuffle node is simply a swizzle of another shuffle node,
10689 // then try to simplify it.
10690 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
10691 N1.getOpcode() == ISD::UNDEF) {
10693 ShuffleVectorSDNode *OtherSV = cast<ShuffleVectorSDNode>(N0);
10695 // The incoming shuffle must be of the same type as the result of the
10696 // current shuffle.
10697 assert(OtherSV->getOperand(0).getValueType() == VT &&
10698 "Shuffle types don't match");
10700 SmallVector<int, 4> Mask;
10701 // Compute the combined shuffle mask.
10702 for (unsigned i = 0; i != NumElts; ++i) {
10703 int Idx = SVN->getMaskElt(i);
10704 assert(Idx < (int)NumElts && "Index references undef operand");
10705 // Next, this index comes from the first value, which is the incoming
10706 // shuffle. Adopt the incoming index.
10708 Idx = OtherSV->getMaskElt(Idx);
10709 Mask.push_back(Idx);
10712 bool CommuteOperands = false;
10713 if (N0.getOperand(1).getOpcode() != ISD::UNDEF) {
10714 // To be valid, the combine shuffle mask should only reference elements
10715 // from one of the two vectors in input to the inner shufflevector.
10716 bool IsValidMask = true;
10717 for (unsigned i = 0; i != NumElts && IsValidMask; ++i)
10718 // See if the combined mask only reference undefs or elements coming
10719 // from the first shufflevector operand.
10720 IsValidMask = Mask[i] < 0 || (unsigned)Mask[i] < NumElts;
10722 if (!IsValidMask) {
10723 IsValidMask = true;
10724 for (unsigned i = 0; i != NumElts && IsValidMask; ++i)
10725 // Check that all the elements come from the second shuffle operand.
10726 IsValidMask = Mask[i] < 0 || (unsigned)Mask[i] >= NumElts;
10727 CommuteOperands = IsValidMask;
10730 // Early exit if the combined shuffle mask is not valid.
10735 // See if this pair of shuffles can be safely folded according to either
10736 // of the following rules:
10737 // shuffle(shuffle(x, y), undef) -> x
10738 // shuffle(shuffle(x, undef), undef) -> x
10739 // shuffle(shuffle(x, y), undef) -> y
10740 bool IsIdentityMask = true;
10741 unsigned BaseMaskIndex = CommuteOperands ? NumElts : 0;
10742 for (unsigned i = 0; i != NumElts && IsIdentityMask; ++i) {
10747 // The combined shuffle must map each index to itself.
10748 IsIdentityMask = (unsigned)Mask[i] == i + BaseMaskIndex;
10751 if (IsIdentityMask) {
10752 if (CommuteOperands)
10753 // optimize shuffle(shuffle(x, y), undef) -> y.
10754 return OtherSV->getOperand(1);
10756 // optimize shuffle(shuffle(x, undef), undef) -> x
10757 // optimize shuffle(shuffle(x, y), undef) -> x
10758 return OtherSV->getOperand(0);
10761 // It may still be beneficial to combine the two shuffles if the
10762 // resulting shuffle is legal.
10763 if (TLI.isTypeLegal(VT) && TLI.isShuffleMaskLegal(Mask, VT)) {
10764 if (!CommuteOperands)
10765 // shuffle(shuffle(x, undef, M1), undef, M2) -> shuffle(x, undef, M3).
10766 // shuffle(shuffle(x, y, M1), undef, M2) -> shuffle(x, undef, M3)
10767 return DAG.getVectorShuffle(VT, SDLoc(N), N0->getOperand(0), N1,
10770 // shuffle(shuffle(x, y, M1), undef, M2) -> shuffle(undef, y, M3)
10771 return DAG.getVectorShuffle(VT, SDLoc(N), N1, N0->getOperand(1),
10776 // Canonicalize shuffles according to rules:
10777 // shuffle(A, shuffle(A, B)) -> shuffle(shuffle(A,B), A)
10778 // shuffle(B, shuffle(A, B)) -> shuffle(shuffle(A,B), B)
10779 // shuffle(B, shuffle(A, Undef)) -> shuffle(shuffle(A, Undef), B)
10780 if (N1.getOpcode() == ISD::VECTOR_SHUFFLE && N0.getOpcode() != ISD::UNDEF &&
10781 N0.getOpcode() != ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
10782 TLI.isTypeLegal(VT)) {
10783 // The incoming shuffle must be of the same type as the result of the
10784 // current shuffle.
10785 assert(N1->getOperand(0).getValueType() == VT &&
10786 "Shuffle types don't match");
10788 SDValue SV0 = N1->getOperand(0);
10789 SDValue SV1 = N1->getOperand(1);
10790 bool HasSameOp0 = N0 == SV0;
10791 bool IsSV1Undef = SV1.getOpcode() == ISD::UNDEF;
10792 if (HasSameOp0 || IsSV1Undef || N0 == SV1)
10793 // Commute the operands of this shuffle so that next rule
10795 return DAG.getCommutedVectorShuffle(*SVN);
10798 // Try to fold according to rules:
10799 // shuffle(shuffle(A, B, M0), B, M1) -> shuffle(A, B, M2)
10800 // shuffle(shuffle(A, B, M0), A, M1) -> shuffle(A, B, M2)
10801 // shuffle(shuffle(A, Undef, M0), B, M1) -> shuffle(A, B, M2)
10802 // shuffle(shuffle(A, Undef, M0), A, M1) -> shuffle(A, Undef, M2)
10803 // Don't try to fold shuffles with illegal type.
10804 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
10805 N1.getOpcode() != ISD::UNDEF && TLI.isTypeLegal(VT)) {
10806 ShuffleVectorSDNode *OtherSV = cast<ShuffleVectorSDNode>(N0);
10808 // The incoming shuffle must be of the same type as the result of the
10809 // current shuffle.
10810 assert(OtherSV->getOperand(0).getValueType() == VT &&
10811 "Shuffle types don't match");
10813 SDValue SV0 = OtherSV->getOperand(0);
10814 SDValue SV1 = OtherSV->getOperand(1);
10815 bool HasSameOp0 = N1 == SV0;
10816 bool IsSV1Undef = SV1.getOpcode() == ISD::UNDEF;
10817 if (!HasSameOp0 && !IsSV1Undef && N1 != SV1)
10821 SmallVector<int, 4> Mask;
10822 // Compute the combined shuffle mask for a shuffle with SV0 as the first
10823 // operand, and SV1 as the second operand.
10824 for (unsigned i = 0; i != NumElts; ++i) {
10825 int Idx = SVN->getMaskElt(i);
10827 // Propagate Undef.
10828 Mask.push_back(Idx);
10832 if (Idx < (int)NumElts) {
10833 Idx = OtherSV->getMaskElt(Idx);
10834 if (IsSV1Undef && Idx >= (int) NumElts)
10835 Idx = -1; // Propagate Undef.
10837 Idx = HasSameOp0 ? Idx - NumElts : Idx;
10839 Mask.push_back(Idx);
10842 // Avoid introducing shuffles with illegal mask.
10843 if (TLI.isShuffleMaskLegal(Mask, VT)) {
10845 // shuffle(shuffle(A, Undef, M0), B, M1) -> shuffle(A, B, M2)
10846 // shuffle(shuffle(A, Undef, M0), A, M1) -> shuffle(A, Undef, M2)
10847 return DAG.getVectorShuffle(VT, SDLoc(N), SV0, N1, &Mask[0]);
10848 return DAG.getVectorShuffle(VT, SDLoc(N), SV0, SV1, &Mask[0]);
10855 SDValue DAGCombiner::visitINSERT_SUBVECTOR(SDNode *N) {
10856 SDValue N0 = N->getOperand(0);
10857 SDValue N2 = N->getOperand(2);
10859 // If the input vector is a concatenation, and the insert replaces
10860 // one of the halves, we can optimize into a single concat_vectors.
10861 if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
10862 N0->getNumOperands() == 2 && N2.getOpcode() == ISD::Constant) {
10863 APInt InsIdx = cast<ConstantSDNode>(N2)->getAPIntValue();
10864 EVT VT = N->getValueType(0);
10866 // Lower half: fold (insert_subvector (concat_vectors X, Y), Z) ->
10867 // (concat_vectors Z, Y)
10869 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
10870 N->getOperand(1), N0.getOperand(1));
10872 // Upper half: fold (insert_subvector (concat_vectors X, Y), Z) ->
10873 // (concat_vectors X, Z)
10874 if (InsIdx == VT.getVectorNumElements()/2)
10875 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
10876 N0.getOperand(0), N->getOperand(1));
10882 /// XformToShuffleWithZero - Returns a vector_shuffle if it able to transform
10883 /// an AND to a vector_shuffle with the destination vector and a zero vector.
10884 /// e.g. AND V, <0xffffffff, 0, 0xffffffff, 0>. ==>
10885 /// vector_shuffle V, Zero, <0, 4, 2, 4>
10886 SDValue DAGCombiner::XformToShuffleWithZero(SDNode *N) {
10887 EVT VT = N->getValueType(0);
10889 SDValue LHS = N->getOperand(0);
10890 SDValue RHS = N->getOperand(1);
10891 if (N->getOpcode() == ISD::AND) {
10892 if (RHS.getOpcode() == ISD::BITCAST)
10893 RHS = RHS.getOperand(0);
10894 if (RHS.getOpcode() == ISD::BUILD_VECTOR) {
10895 SmallVector<int, 8> Indices;
10896 unsigned NumElts = RHS.getNumOperands();
10897 for (unsigned i = 0; i != NumElts; ++i) {
10898 SDValue Elt = RHS.getOperand(i);
10899 if (!isa<ConstantSDNode>(Elt))
10902 if (cast<ConstantSDNode>(Elt)->isAllOnesValue())
10903 Indices.push_back(i);
10904 else if (cast<ConstantSDNode>(Elt)->isNullValue())
10905 Indices.push_back(NumElts);
10910 // Let's see if the target supports this vector_shuffle.
10911 EVT RVT = RHS.getValueType();
10912 if (!TLI.isVectorClearMaskLegal(Indices, RVT))
10915 // Return the new VECTOR_SHUFFLE node.
10916 EVT EltVT = RVT.getVectorElementType();
10917 SmallVector<SDValue,8> ZeroOps(RVT.getVectorNumElements(),
10918 DAG.getConstant(0, EltVT));
10919 SDValue Zero = DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), RVT, ZeroOps);
10920 LHS = DAG.getNode(ISD::BITCAST, dl, RVT, LHS);
10921 SDValue Shuf = DAG.getVectorShuffle(RVT, dl, LHS, Zero, &Indices[0]);
10922 return DAG.getNode(ISD::BITCAST, dl, VT, Shuf);
10929 /// SimplifyVBinOp - Visit a binary vector operation, like ADD.
10930 SDValue DAGCombiner::SimplifyVBinOp(SDNode *N) {
10931 assert(N->getValueType(0).isVector() &&
10932 "SimplifyVBinOp only works on vectors!");
10934 SDValue LHS = N->getOperand(0);
10935 SDValue RHS = N->getOperand(1);
10936 SDValue Shuffle = XformToShuffleWithZero(N);
10937 if (Shuffle.getNode()) return Shuffle;
10939 // If the LHS and RHS are BUILD_VECTOR nodes, see if we can constant fold
10941 if (LHS.getOpcode() == ISD::BUILD_VECTOR &&
10942 RHS.getOpcode() == ISD::BUILD_VECTOR) {
10943 // Check if both vectors are constants. If not bail out.
10944 if (!(cast<BuildVectorSDNode>(LHS)->isConstant() &&
10945 cast<BuildVectorSDNode>(RHS)->isConstant()))
10948 SmallVector<SDValue, 8> Ops;
10949 for (unsigned i = 0, e = LHS.getNumOperands(); i != e; ++i) {
10950 SDValue LHSOp = LHS.getOperand(i);
10951 SDValue RHSOp = RHS.getOperand(i);
10953 // Can't fold divide by zero.
10954 if (N->getOpcode() == ISD::SDIV || N->getOpcode() == ISD::UDIV ||
10955 N->getOpcode() == ISD::FDIV) {
10956 if ((RHSOp.getOpcode() == ISD::Constant &&
10957 cast<ConstantSDNode>(RHSOp.getNode())->isNullValue()) ||
10958 (RHSOp.getOpcode() == ISD::ConstantFP &&
10959 cast<ConstantFPSDNode>(RHSOp.getNode())->getValueAPF().isZero()))
10963 EVT VT = LHSOp.getValueType();
10964 EVT RVT = RHSOp.getValueType();
10966 // Integer BUILD_VECTOR operands may have types larger than the element
10967 // size (e.g., when the element type is not legal). Prior to type
10968 // legalization, the types may not match between the two BUILD_VECTORS.
10969 // Truncate one of the operands to make them match.
10970 if (RVT.getSizeInBits() > VT.getSizeInBits()) {
10971 RHSOp = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, RHSOp);
10973 LHSOp = DAG.getNode(ISD::TRUNCATE, SDLoc(N), RVT, LHSOp);
10977 SDValue FoldOp = DAG.getNode(N->getOpcode(), SDLoc(LHS), VT,
10979 if (FoldOp.getOpcode() != ISD::UNDEF &&
10980 FoldOp.getOpcode() != ISD::Constant &&
10981 FoldOp.getOpcode() != ISD::ConstantFP)
10983 Ops.push_back(FoldOp);
10984 AddToWorklist(FoldOp.getNode());
10987 if (Ops.size() == LHS.getNumOperands())
10988 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), LHS.getValueType(), Ops);
10991 // Type legalization might introduce new shuffles in the DAG.
10992 // Fold (VBinOp (shuffle (A, Undef, Mask)), (shuffle (B, Undef, Mask)))
10993 // -> (shuffle (VBinOp (A, B)), Undef, Mask).
10994 if (LegalTypes && isa<ShuffleVectorSDNode>(LHS) &&
10995 isa<ShuffleVectorSDNode>(RHS) && LHS.hasOneUse() && RHS.hasOneUse() &&
10996 LHS.getOperand(1).getOpcode() == ISD::UNDEF &&
10997 RHS.getOperand(1).getOpcode() == ISD::UNDEF) {
10998 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(LHS);
10999 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(RHS);
11001 if (SVN0->getMask().equals(SVN1->getMask())) {
11002 EVT VT = N->getValueType(0);
11003 SDValue UndefVector = LHS.getOperand(1);
11004 SDValue NewBinOp = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
11005 LHS.getOperand(0), RHS.getOperand(0));
11006 AddUsersToWorklist(N);
11007 return DAG.getVectorShuffle(VT, SDLoc(N), NewBinOp, UndefVector,
11008 &SVN0->getMask()[0]);
11015 /// SimplifyVUnaryOp - Visit a binary vector operation, like FABS/FNEG.
11016 SDValue DAGCombiner::SimplifyVUnaryOp(SDNode *N) {
11017 assert(N->getValueType(0).isVector() &&
11018 "SimplifyVUnaryOp only works on vectors!");
11020 SDValue N0 = N->getOperand(0);
11022 if (N0.getOpcode() != ISD::BUILD_VECTOR)
11025 // Operand is a BUILD_VECTOR node, see if we can constant fold it.
11026 SmallVector<SDValue, 8> Ops;
11027 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) {
11028 SDValue Op = N0.getOperand(i);
11029 if (Op.getOpcode() != ISD::UNDEF &&
11030 Op.getOpcode() != ISD::ConstantFP)
11032 EVT EltVT = Op.getValueType();
11033 SDValue FoldOp = DAG.getNode(N->getOpcode(), SDLoc(N0), EltVT, Op);
11034 if (FoldOp.getOpcode() != ISD::UNDEF &&
11035 FoldOp.getOpcode() != ISD::ConstantFP)
11037 Ops.push_back(FoldOp);
11038 AddToWorklist(FoldOp.getNode());
11041 if (Ops.size() != N0.getNumOperands())
11044 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), N0.getValueType(), Ops);
11047 SDValue DAGCombiner::SimplifySelect(SDLoc DL, SDValue N0,
11048 SDValue N1, SDValue N2){
11049 assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!");
11051 SDValue SCC = SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1), N1, N2,
11052 cast<CondCodeSDNode>(N0.getOperand(2))->get());
11054 // If we got a simplified select_cc node back from SimplifySelectCC, then
11055 // break it down into a new SETCC node, and a new SELECT node, and then return
11056 // the SELECT node, since we were called with a SELECT node.
11057 if (SCC.getNode()) {
11058 // Check to see if we got a select_cc back (to turn into setcc/select).
11059 // Otherwise, just return whatever node we got back, like fabs.
11060 if (SCC.getOpcode() == ISD::SELECT_CC) {
11061 SDValue SETCC = DAG.getNode(ISD::SETCC, SDLoc(N0),
11063 SCC.getOperand(0), SCC.getOperand(1),
11064 SCC.getOperand(4));
11065 AddToWorklist(SETCC.getNode());
11066 return DAG.getSelect(SDLoc(SCC), SCC.getValueType(),
11067 SCC.getOperand(2), SCC.getOperand(3), SETCC);
11075 /// SimplifySelectOps - Given a SELECT or a SELECT_CC node, where LHS and RHS
11076 /// are the two values being selected between, see if we can simplify the
11077 /// select. Callers of this should assume that TheSelect is deleted if this
11078 /// returns true. As such, they should return the appropriate thing (e.g. the
11079 /// node) back to the top-level of the DAG combiner loop to avoid it being
11081 bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDValue LHS,
11084 // Cannot simplify select with vector condition
11085 if (TheSelect->getOperand(0).getValueType().isVector()) return false;
11087 // If this is a select from two identical things, try to pull the operation
11088 // through the select.
11089 if (LHS.getOpcode() != RHS.getOpcode() ||
11090 !LHS.hasOneUse() || !RHS.hasOneUse())
11093 // If this is a load and the token chain is identical, replace the select
11094 // of two loads with a load through a select of the address to load from.
11095 // This triggers in things like "select bool X, 10.0, 123.0" after the FP
11096 // constants have been dropped into the constant pool.
11097 if (LHS.getOpcode() == ISD::LOAD) {
11098 LoadSDNode *LLD = cast<LoadSDNode>(LHS);
11099 LoadSDNode *RLD = cast<LoadSDNode>(RHS);
11101 // Token chains must be identical.
11102 if (LHS.getOperand(0) != RHS.getOperand(0) ||
11103 // Do not let this transformation reduce the number of volatile loads.
11104 LLD->isVolatile() || RLD->isVolatile() ||
11105 // If this is an EXTLOAD, the VT's must match.
11106 LLD->getMemoryVT() != RLD->getMemoryVT() ||
11107 // If this is an EXTLOAD, the kind of extension must match.
11108 (LLD->getExtensionType() != RLD->getExtensionType() &&
11109 // The only exception is if one of the extensions is anyext.
11110 LLD->getExtensionType() != ISD::EXTLOAD &&
11111 RLD->getExtensionType() != ISD::EXTLOAD) ||
11112 // FIXME: this discards src value information. This is
11113 // over-conservative. It would be beneficial to be able to remember
11114 // both potential memory locations. Since we are discarding
11115 // src value info, don't do the transformation if the memory
11116 // locations are not in the default address space.
11117 LLD->getPointerInfo().getAddrSpace() != 0 ||
11118 RLD->getPointerInfo().getAddrSpace() != 0 ||
11119 !TLI.isOperationLegalOrCustom(TheSelect->getOpcode(),
11120 LLD->getBasePtr().getValueType()))
11123 // Check that the select condition doesn't reach either load. If so,
11124 // folding this will induce a cycle into the DAG. If not, this is safe to
11125 // xform, so create a select of the addresses.
11127 if (TheSelect->getOpcode() == ISD::SELECT) {
11128 SDNode *CondNode = TheSelect->getOperand(0).getNode();
11129 if ((LLD->hasAnyUseOfValue(1) && LLD->isPredecessorOf(CondNode)) ||
11130 (RLD->hasAnyUseOfValue(1) && RLD->isPredecessorOf(CondNode)))
11132 // The loads must not depend on one another.
11133 if (LLD->isPredecessorOf(RLD) ||
11134 RLD->isPredecessorOf(LLD))
11136 Addr = DAG.getSelect(SDLoc(TheSelect),
11137 LLD->getBasePtr().getValueType(),
11138 TheSelect->getOperand(0), LLD->getBasePtr(),
11139 RLD->getBasePtr());
11140 } else { // Otherwise SELECT_CC
11141 SDNode *CondLHS = TheSelect->getOperand(0).getNode();
11142 SDNode *CondRHS = TheSelect->getOperand(1).getNode();
11144 if ((LLD->hasAnyUseOfValue(1) &&
11145 (LLD->isPredecessorOf(CondLHS) || LLD->isPredecessorOf(CondRHS))) ||
11146 (RLD->hasAnyUseOfValue(1) &&
11147 (RLD->isPredecessorOf(CondLHS) || RLD->isPredecessorOf(CondRHS))))
11150 Addr = DAG.getNode(ISD::SELECT_CC, SDLoc(TheSelect),
11151 LLD->getBasePtr().getValueType(),
11152 TheSelect->getOperand(0),
11153 TheSelect->getOperand(1),
11154 LLD->getBasePtr(), RLD->getBasePtr(),
11155 TheSelect->getOperand(4));
11159 if (LLD->getExtensionType() == ISD::NON_EXTLOAD) {
11160 Load = DAG.getLoad(TheSelect->getValueType(0),
11162 // FIXME: Discards pointer and AA info.
11163 LLD->getChain(), Addr, MachinePointerInfo(),
11164 LLD->isVolatile(), LLD->isNonTemporal(),
11165 LLD->isInvariant(), LLD->getAlignment());
11167 Load = DAG.getExtLoad(LLD->getExtensionType() == ISD::EXTLOAD ?
11168 RLD->getExtensionType() : LLD->getExtensionType(),
11170 TheSelect->getValueType(0),
11171 // FIXME: Discards pointer and AA info.
11172 LLD->getChain(), Addr, MachinePointerInfo(),
11173 LLD->getMemoryVT(), LLD->isVolatile(),
11174 LLD->isNonTemporal(), LLD->getAlignment());
11177 // Users of the select now use the result of the load.
11178 CombineTo(TheSelect, Load);
11180 // Users of the old loads now use the new load's chain. We know the
11181 // old-load value is dead now.
11182 CombineTo(LHS.getNode(), Load.getValue(0), Load.getValue(1));
11183 CombineTo(RHS.getNode(), Load.getValue(0), Load.getValue(1));
11190 /// SimplifySelectCC - Simplify an expression of the form (N0 cond N1) ? N2 : N3
11191 /// where 'cond' is the comparison specified by CC.
11192 SDValue DAGCombiner::SimplifySelectCC(SDLoc DL, SDValue N0, SDValue N1,
11193 SDValue N2, SDValue N3,
11194 ISD::CondCode CC, bool NotExtCompare) {
11195 // (x ? y : y) -> y.
11196 if (N2 == N3) return N2;
11198 EVT VT = N2.getValueType();
11199 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
11200 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
11201 ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N3.getNode());
11203 // Determine if the condition we're dealing with is constant
11204 SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()),
11205 N0, N1, CC, DL, false);
11206 if (SCC.getNode()) AddToWorklist(SCC.getNode());
11207 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode());
11209 // fold select_cc true, x, y -> x
11210 if (SCCC && !SCCC->isNullValue())
11212 // fold select_cc false, x, y -> y
11213 if (SCCC && SCCC->isNullValue())
11216 // Check to see if we can simplify the select into an fabs node
11217 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) {
11218 // Allow either -0.0 or 0.0
11219 if (CFP->getValueAPF().isZero()) {
11220 // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs
11221 if ((CC == ISD::SETGE || CC == ISD::SETGT) &&
11222 N0 == N2 && N3.getOpcode() == ISD::FNEG &&
11223 N2 == N3.getOperand(0))
11224 return DAG.getNode(ISD::FABS, DL, VT, N0);
11226 // select (setl[te] X, +/-0.0), fneg(X), X -> fabs
11227 if ((CC == ISD::SETLT || CC == ISD::SETLE) &&
11228 N0 == N3 && N2.getOpcode() == ISD::FNEG &&
11229 N2.getOperand(0) == N3)
11230 return DAG.getNode(ISD::FABS, DL, VT, N3);
11234 // Turn "(a cond b) ? 1.0f : 2.0f" into "load (tmp + ((a cond b) ? 0 : 4)"
11235 // where "tmp" is a constant pool entry containing an array with 1.0 and 2.0
11236 // in it. This is a win when the constant is not otherwise available because
11237 // it replaces two constant pool loads with one. We only do this if the FP
11238 // type is known to be legal, because if it isn't, then we are before legalize
11239 // types an we want the other legalization to happen first (e.g. to avoid
11240 // messing with soft float) and if the ConstantFP is not legal, because if
11241 // it is legal, we may not need to store the FP constant in a constant pool.
11242 if (ConstantFPSDNode *TV = dyn_cast<ConstantFPSDNode>(N2))
11243 if (ConstantFPSDNode *FV = dyn_cast<ConstantFPSDNode>(N3)) {
11244 if (TLI.isTypeLegal(N2.getValueType()) &&
11245 (TLI.getOperationAction(ISD::ConstantFP, N2.getValueType()) !=
11246 TargetLowering::Legal &&
11247 !TLI.isFPImmLegal(TV->getValueAPF(), TV->getValueType(0)) &&
11248 !TLI.isFPImmLegal(FV->getValueAPF(), FV->getValueType(0))) &&
11249 // If both constants have multiple uses, then we won't need to do an
11250 // extra load, they are likely around in registers for other users.
11251 (TV->hasOneUse() || FV->hasOneUse())) {
11252 Constant *Elts[] = {
11253 const_cast<ConstantFP*>(FV->getConstantFPValue()),
11254 const_cast<ConstantFP*>(TV->getConstantFPValue())
11256 Type *FPTy = Elts[0]->getType();
11257 const DataLayout &TD = *TLI.getDataLayout();
11259 // Create a ConstantArray of the two constants.
11260 Constant *CA = ConstantArray::get(ArrayType::get(FPTy, 2), Elts);
11261 SDValue CPIdx = DAG.getConstantPool(CA, TLI.getPointerTy(),
11262 TD.getPrefTypeAlignment(FPTy));
11263 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
11265 // Get the offsets to the 0 and 1 element of the array so that we can
11266 // select between them.
11267 SDValue Zero = DAG.getIntPtrConstant(0);
11268 unsigned EltSize = (unsigned)TD.getTypeAllocSize(Elts[0]->getType());
11269 SDValue One = DAG.getIntPtrConstant(EltSize);
11271 SDValue Cond = DAG.getSetCC(DL,
11272 getSetCCResultType(N0.getValueType()),
11274 AddToWorklist(Cond.getNode());
11275 SDValue CstOffset = DAG.getSelect(DL, Zero.getValueType(),
11277 AddToWorklist(CstOffset.getNode());
11278 CPIdx = DAG.getNode(ISD::ADD, DL, CPIdx.getValueType(), CPIdx,
11280 AddToWorklist(CPIdx.getNode());
11281 return DAG.getLoad(TV->getValueType(0), DL, DAG.getEntryNode(), CPIdx,
11282 MachinePointerInfo::getConstantPool(), false,
11283 false, false, Alignment);
11288 // Check to see if we can perform the "gzip trick", transforming
11289 // (select_cc setlt X, 0, A, 0) -> (and (sra X, (sub size(X), 1), A)
11290 if (N1C && N3C && N3C->isNullValue() && CC == ISD::SETLT &&
11291 (N1C->isNullValue() || // (a < 0) ? b : 0
11292 (N1C->getAPIntValue() == 1 && N0 == N2))) { // (a < 1) ? a : 0
11293 EVT XType = N0.getValueType();
11294 EVT AType = N2.getValueType();
11295 if (XType.bitsGE(AType)) {
11296 // and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a
11297 // single-bit constant.
11298 if (N2C && ((N2C->getAPIntValue() & (N2C->getAPIntValue()-1)) == 0)) {
11299 unsigned ShCtV = N2C->getAPIntValue().logBase2();
11300 ShCtV = XType.getSizeInBits()-ShCtV-1;
11301 SDValue ShCt = DAG.getConstant(ShCtV,
11302 getShiftAmountTy(N0.getValueType()));
11303 SDValue Shift = DAG.getNode(ISD::SRL, SDLoc(N0),
11305 AddToWorklist(Shift.getNode());
11307 if (XType.bitsGT(AType)) {
11308 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
11309 AddToWorklist(Shift.getNode());
11312 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
11315 SDValue Shift = DAG.getNode(ISD::SRA, SDLoc(N0),
11317 DAG.getConstant(XType.getSizeInBits()-1,
11318 getShiftAmountTy(N0.getValueType())));
11319 AddToWorklist(Shift.getNode());
11321 if (XType.bitsGT(AType)) {
11322 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
11323 AddToWorklist(Shift.getNode());
11326 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
11330 // fold (select_cc seteq (and x, y), 0, 0, A) -> (and (shr (shl x)) A)
11331 // where y is has a single bit set.
11332 // A plaintext description would be, we can turn the SELECT_CC into an AND
11333 // when the condition can be materialized as an all-ones register. Any
11334 // single bit-test can be materialized as an all-ones register with
11335 // shift-left and shift-right-arith.
11336 if (CC == ISD::SETEQ && N0->getOpcode() == ISD::AND &&
11337 N0->getValueType(0) == VT &&
11338 N1C && N1C->isNullValue() &&
11339 N2C && N2C->isNullValue()) {
11340 SDValue AndLHS = N0->getOperand(0);
11341 ConstantSDNode *ConstAndRHS = dyn_cast<ConstantSDNode>(N0->getOperand(1));
11342 if (ConstAndRHS && ConstAndRHS->getAPIntValue().countPopulation() == 1) {
11343 // Shift the tested bit over the sign bit.
11344 APInt AndMask = ConstAndRHS->getAPIntValue();
11346 DAG.getConstant(AndMask.countLeadingZeros(),
11347 getShiftAmountTy(AndLHS.getValueType()));
11348 SDValue Shl = DAG.getNode(ISD::SHL, SDLoc(N0), VT, AndLHS, ShlAmt);
11350 // Now arithmetic right shift it all the way over, so the result is either
11351 // all-ones, or zero.
11353 DAG.getConstant(AndMask.getBitWidth()-1,
11354 getShiftAmountTy(Shl.getValueType()));
11355 SDValue Shr = DAG.getNode(ISD::SRA, SDLoc(N0), VT, Shl, ShrAmt);
11357 return DAG.getNode(ISD::AND, DL, VT, Shr, N3);
11361 // fold select C, 16, 0 -> shl C, 4
11362 if (N2C && N3C && N3C->isNullValue() && N2C->getAPIntValue().isPowerOf2() &&
11363 TLI.getBooleanContents(N0.getValueType()) ==
11364 TargetLowering::ZeroOrOneBooleanContent) {
11366 // If the caller doesn't want us to simplify this into a zext of a compare,
11368 if (NotExtCompare && N2C->getAPIntValue() == 1)
11371 // Get a SetCC of the condition
11372 // NOTE: Don't create a SETCC if it's not legal on this target.
11373 if (!LegalOperations ||
11374 TLI.isOperationLegal(ISD::SETCC,
11375 LegalTypes ? getSetCCResultType(N0.getValueType()) : MVT::i1)) {
11377 // cast from setcc result type to select result type
11379 SCC = DAG.getSetCC(DL, getSetCCResultType(N0.getValueType()),
11381 if (N2.getValueType().bitsLT(SCC.getValueType()))
11382 Temp = DAG.getZeroExtendInReg(SCC, SDLoc(N2),
11383 N2.getValueType());
11385 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2),
11386 N2.getValueType(), SCC);
11388 SCC = DAG.getSetCC(SDLoc(N0), MVT::i1, N0, N1, CC);
11389 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2),
11390 N2.getValueType(), SCC);
11393 AddToWorklist(SCC.getNode());
11394 AddToWorklist(Temp.getNode());
11396 if (N2C->getAPIntValue() == 1)
11399 // shl setcc result by log2 n2c
11400 return DAG.getNode(
11401 ISD::SHL, DL, N2.getValueType(), Temp,
11402 DAG.getConstant(N2C->getAPIntValue().logBase2(),
11403 getShiftAmountTy(Temp.getValueType())));
11407 // Check to see if this is the equivalent of setcc
11408 // FIXME: Turn all of these into setcc if setcc if setcc is legal
11409 // otherwise, go ahead with the folds.
11410 if (0 && N3C && N3C->isNullValue() && N2C && (N2C->getAPIntValue() == 1ULL)) {
11411 EVT XType = N0.getValueType();
11412 if (!LegalOperations ||
11413 TLI.isOperationLegal(ISD::SETCC, getSetCCResultType(XType))) {
11414 SDValue Res = DAG.getSetCC(DL, getSetCCResultType(XType), N0, N1, CC);
11415 if (Res.getValueType() != VT)
11416 Res = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Res);
11420 // fold (seteq X, 0) -> (srl (ctlz X, log2(size(X))))
11421 if (N1C && N1C->isNullValue() && CC == ISD::SETEQ &&
11422 (!LegalOperations ||
11423 TLI.isOperationLegal(ISD::CTLZ, XType))) {
11424 SDValue Ctlz = DAG.getNode(ISD::CTLZ, SDLoc(N0), XType, N0);
11425 return DAG.getNode(ISD::SRL, DL, XType, Ctlz,
11426 DAG.getConstant(Log2_32(XType.getSizeInBits()),
11427 getShiftAmountTy(Ctlz.getValueType())));
11429 // fold (setgt X, 0) -> (srl (and (-X, ~X), size(X)-1))
11430 if (N1C && N1C->isNullValue() && CC == ISD::SETGT) {
11431 SDValue NegN0 = DAG.getNode(ISD::SUB, SDLoc(N0),
11432 XType, DAG.getConstant(0, XType), N0);
11433 SDValue NotN0 = DAG.getNOT(SDLoc(N0), N0, XType);
11434 return DAG.getNode(ISD::SRL, DL, XType,
11435 DAG.getNode(ISD::AND, DL, XType, NegN0, NotN0),
11436 DAG.getConstant(XType.getSizeInBits()-1,
11437 getShiftAmountTy(XType)));
11439 // fold (setgt X, -1) -> (xor (srl (X, size(X)-1), 1))
11440 if (N1C && N1C->isAllOnesValue() && CC == ISD::SETGT) {
11441 SDValue Sign = DAG.getNode(ISD::SRL, SDLoc(N0), XType, N0,
11442 DAG.getConstant(XType.getSizeInBits()-1,
11443 getShiftAmountTy(N0.getValueType())));
11444 return DAG.getNode(ISD::XOR, DL, XType, Sign, DAG.getConstant(1, XType));
11448 // Check to see if this is an integer abs.
11449 // select_cc setg[te] X, 0, X, -X ->
11450 // select_cc setgt X, -1, X, -X ->
11451 // select_cc setl[te] X, 0, -X, X ->
11452 // select_cc setlt X, 1, -X, X ->
11453 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
11455 ConstantSDNode *SubC = nullptr;
11456 if (((N1C->isNullValue() && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
11457 (N1C->isAllOnesValue() && CC == ISD::SETGT)) &&
11458 N0 == N2 && N3.getOpcode() == ISD::SUB && N0 == N3.getOperand(1))
11459 SubC = dyn_cast<ConstantSDNode>(N3.getOperand(0));
11460 else if (((N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE)) ||
11461 (N1C->isOne() && CC == ISD::SETLT)) &&
11462 N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1))
11463 SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0));
11465 EVT XType = N0.getValueType();
11466 if (SubC && SubC->isNullValue() && XType.isInteger()) {
11467 SDValue Shift = DAG.getNode(ISD::SRA, SDLoc(N0), XType,
11469 DAG.getConstant(XType.getSizeInBits()-1,
11470 getShiftAmountTy(N0.getValueType())));
11471 SDValue Add = DAG.getNode(ISD::ADD, SDLoc(N0),
11473 AddToWorklist(Shift.getNode());
11474 AddToWorklist(Add.getNode());
11475 return DAG.getNode(ISD::XOR, DL, XType, Add, Shift);
11482 /// SimplifySetCC - This is a stub for TargetLowering::SimplifySetCC.
11483 SDValue DAGCombiner::SimplifySetCC(EVT VT, SDValue N0,
11484 SDValue N1, ISD::CondCode Cond,
11485 SDLoc DL, bool foldBooleans) {
11486 TargetLowering::DAGCombinerInfo
11487 DagCombineInfo(DAG, Level, false, this);
11488 return TLI.SimplifySetCC(VT, N0, N1, Cond, foldBooleans, DagCombineInfo, DL);
11491 /// BuildSDIV - Given an ISD::SDIV node expressing a divide by constant, return
11492 /// a DAG expression to select that will generate the same value by multiplying
11493 /// by a magic number. See:
11494 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
11495 SDValue DAGCombiner::BuildSDIV(SDNode *N) {
11496 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
11500 // Avoid division by zero.
11501 if (!C->getAPIntValue())
11504 std::vector<SDNode*> Built;
11506 TLI.BuildSDIV(N, C->getAPIntValue(), DAG, LegalOperations, &Built);
11508 for (SDNode *N : Built)
11513 /// BuildSDIVPow2 - Given an ISD::SDIV node expressing a divide by constant
11514 /// power of 2, return a DAG expression to select that will generate the same
11515 /// value by right shifting.
11516 SDValue DAGCombiner::BuildSDIVPow2(SDNode *N) {
11517 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
11521 // Avoid division by zero.
11522 if (!C->getAPIntValue())
11525 std::vector<SDNode *> Built;
11526 SDValue S = TLI.BuildSDIVPow2(N, C->getAPIntValue(), DAG, &Built);
11528 for (SDNode *N : Built)
11533 /// BuildUDIV - Given an ISD::UDIV node expressing a divide by constant,
11534 /// return a DAG expression to select that will generate the same value by
11535 /// multiplying by a magic number. See:
11536 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
11537 SDValue DAGCombiner::BuildUDIV(SDNode *N) {
11538 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
11542 // Avoid division by zero.
11543 if (!C->getAPIntValue())
11546 std::vector<SDNode*> Built;
11548 TLI.BuildUDIV(N, C->getAPIntValue(), DAG, LegalOperations, &Built);
11550 for (SDNode *N : Built)
11555 /// FindBaseOffset - Return true if base is a frame index, which is known not
11556 // to alias with anything but itself. Provides base object and offset as
11558 static bool FindBaseOffset(SDValue Ptr, SDValue &Base, int64_t &Offset,
11559 const GlobalValue *&GV, const void *&CV) {
11560 // Assume it is a primitive operation.
11561 Base = Ptr; Offset = 0; GV = nullptr; CV = nullptr;
11563 // If it's an adding a simple constant then integrate the offset.
11564 if (Base.getOpcode() == ISD::ADD) {
11565 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Base.getOperand(1))) {
11566 Base = Base.getOperand(0);
11567 Offset += C->getZExtValue();
11571 // Return the underlying GlobalValue, and update the Offset. Return false
11572 // for GlobalAddressSDNode since the same GlobalAddress may be represented
11573 // by multiple nodes with different offsets.
11574 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Base)) {
11575 GV = G->getGlobal();
11576 Offset += G->getOffset();
11580 // Return the underlying Constant value, and update the Offset. Return false
11581 // for ConstantSDNodes since the same constant pool entry may be represented
11582 // by multiple nodes with different offsets.
11583 if (ConstantPoolSDNode *C = dyn_cast<ConstantPoolSDNode>(Base)) {
11584 CV = C->isMachineConstantPoolEntry() ? (const void *)C->getMachineCPVal()
11585 : (const void *)C->getConstVal();
11586 Offset += C->getOffset();
11589 // If it's any of the following then it can't alias with anything but itself.
11590 return isa<FrameIndexSDNode>(Base);
11593 /// isAlias - Return true if there is any possibility that the two addresses
11595 bool DAGCombiner::isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) const {
11596 // If they are the same then they must be aliases.
11597 if (Op0->getBasePtr() == Op1->getBasePtr()) return true;
11599 // If they are both volatile then they cannot be reordered.
11600 if (Op0->isVolatile() && Op1->isVolatile()) return true;
11602 // Gather base node and offset information.
11603 SDValue Base1, Base2;
11604 int64_t Offset1, Offset2;
11605 const GlobalValue *GV1, *GV2;
11606 const void *CV1, *CV2;
11607 bool isFrameIndex1 = FindBaseOffset(Op0->getBasePtr(),
11608 Base1, Offset1, GV1, CV1);
11609 bool isFrameIndex2 = FindBaseOffset(Op1->getBasePtr(),
11610 Base2, Offset2, GV2, CV2);
11612 // If they have a same base address then check to see if they overlap.
11613 if (Base1 == Base2 || (GV1 && (GV1 == GV2)) || (CV1 && (CV1 == CV2)))
11614 return !((Offset1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= Offset2 ||
11615 (Offset2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= Offset1);
11617 // It is possible for different frame indices to alias each other, mostly
11618 // when tail call optimization reuses return address slots for arguments.
11619 // To catch this case, look up the actual index of frame indices to compute
11620 // the real alias relationship.
11621 if (isFrameIndex1 && isFrameIndex2) {
11622 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
11623 Offset1 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base1)->getIndex());
11624 Offset2 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base2)->getIndex());
11625 return !((Offset1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= Offset2 ||
11626 (Offset2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= Offset1);
11629 // Otherwise, if we know what the bases are, and they aren't identical, then
11630 // we know they cannot alias.
11631 if ((isFrameIndex1 || CV1 || GV1) && (isFrameIndex2 || CV2 || GV2))
11634 // If we know required SrcValue1 and SrcValue2 have relatively large alignment
11635 // compared to the size and offset of the access, we may be able to prove they
11636 // do not alias. This check is conservative for now to catch cases created by
11637 // splitting vector types.
11638 if ((Op0->getOriginalAlignment() == Op1->getOriginalAlignment()) &&
11639 (Op0->getSrcValueOffset() != Op1->getSrcValueOffset()) &&
11640 (Op0->getMemoryVT().getSizeInBits() >> 3 ==
11641 Op1->getMemoryVT().getSizeInBits() >> 3) &&
11642 (Op0->getOriginalAlignment() > Op0->getMemoryVT().getSizeInBits()) >> 3) {
11643 int64_t OffAlign1 = Op0->getSrcValueOffset() % Op0->getOriginalAlignment();
11644 int64_t OffAlign2 = Op1->getSrcValueOffset() % Op1->getOriginalAlignment();
11646 // There is no overlap between these relatively aligned accesses of similar
11647 // size, return no alias.
11648 if ((OffAlign1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= OffAlign2 ||
11649 (OffAlign2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= OffAlign1)
11653 bool UseAA = CombinerGlobalAA.getNumOccurrences() > 0 ? CombinerGlobalAA :
11654 TLI.getTargetMachine().getSubtarget<TargetSubtargetInfo>().useAA();
11656 if (CombinerAAOnlyFunc.getNumOccurrences() &&
11657 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
11661 Op0->getMemOperand()->getValue() && Op1->getMemOperand()->getValue()) {
11662 // Use alias analysis information.
11663 int64_t MinOffset = std::min(Op0->getSrcValueOffset(),
11664 Op1->getSrcValueOffset());
11665 int64_t Overlap1 = (Op0->getMemoryVT().getSizeInBits() >> 3) +
11666 Op0->getSrcValueOffset() - MinOffset;
11667 int64_t Overlap2 = (Op1->getMemoryVT().getSizeInBits() >> 3) +
11668 Op1->getSrcValueOffset() - MinOffset;
11669 AliasAnalysis::AliasResult AAResult =
11670 AA.alias(AliasAnalysis::Location(Op0->getMemOperand()->getValue(),
11672 UseTBAA ? Op0->getAAInfo() : AAMDNodes()),
11673 AliasAnalysis::Location(Op1->getMemOperand()->getValue(),
11675 UseTBAA ? Op1->getAAInfo() : AAMDNodes()));
11676 if (AAResult == AliasAnalysis::NoAlias)
11680 // Otherwise we have to assume they alias.
11684 /// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes,
11685 /// looking for aliasing nodes and adding them to the Aliases vector.
11686 void DAGCombiner::GatherAllAliases(SDNode *N, SDValue OriginalChain,
11687 SmallVectorImpl<SDValue> &Aliases) {
11688 SmallVector<SDValue, 8> Chains; // List of chains to visit.
11689 SmallPtrSet<SDNode *, 16> Visited; // Visited node set.
11691 // Get alias information for node.
11692 bool IsLoad = isa<LoadSDNode>(N) && !cast<LSBaseSDNode>(N)->isVolatile();
11695 Chains.push_back(OriginalChain);
11696 unsigned Depth = 0;
11698 // Look at each chain and determine if it is an alias. If so, add it to the
11699 // aliases list. If not, then continue up the chain looking for the next
11701 while (!Chains.empty()) {
11702 SDValue Chain = Chains.back();
11705 // For TokenFactor nodes, look at each operand and only continue up the
11706 // chain until we find two aliases. If we've seen two aliases, assume we'll
11707 // find more and revert to original chain since the xform is unlikely to be
11710 // FIXME: The depth check could be made to return the last non-aliasing
11711 // chain we found before we hit a tokenfactor rather than the original
11713 if (Depth > 6 || Aliases.size() == 2) {
11715 Aliases.push_back(OriginalChain);
11719 // Don't bother if we've been before.
11720 if (!Visited.insert(Chain.getNode()))
11723 switch (Chain.getOpcode()) {
11724 case ISD::EntryToken:
11725 // Entry token is ideal chain operand, but handled in FindBetterChain.
11730 // Get alias information for Chain.
11731 bool IsOpLoad = isa<LoadSDNode>(Chain.getNode()) &&
11732 !cast<LSBaseSDNode>(Chain.getNode())->isVolatile();
11734 // If chain is alias then stop here.
11735 if (!(IsLoad && IsOpLoad) &&
11736 isAlias(cast<LSBaseSDNode>(N), cast<LSBaseSDNode>(Chain.getNode()))) {
11737 Aliases.push_back(Chain);
11739 // Look further up the chain.
11740 Chains.push_back(Chain.getOperand(0));
11746 case ISD::TokenFactor:
11747 // We have to check each of the operands of the token factor for "small"
11748 // token factors, so we queue them up. Adding the operands to the queue
11749 // (stack) in reverse order maintains the original order and increases the
11750 // likelihood that getNode will find a matching token factor (CSE.)
11751 if (Chain.getNumOperands() > 16) {
11752 Aliases.push_back(Chain);
11755 for (unsigned n = Chain.getNumOperands(); n;)
11756 Chains.push_back(Chain.getOperand(--n));
11761 // For all other instructions we will just have to take what we can get.
11762 Aliases.push_back(Chain);
11767 // We need to be careful here to also search for aliases through the
11768 // value operand of a store, etc. Consider the following situation:
11770 // L1 = load Token1, %52
11771 // S1 = store Token1, L1, %51
11772 // L2 = load Token1, %52+8
11773 // S2 = store Token1, L2, %51+8
11774 // Token2 = Token(S1, S2)
11775 // L3 = load Token2, %53
11776 // S3 = store Token2, L3, %52
11777 // L4 = load Token2, %53+8
11778 // S4 = store Token2, L4, %52+8
11779 // If we search for aliases of S3 (which loads address %52), and we look
11780 // only through the chain, then we'll miss the trivial dependence on L1
11781 // (which also loads from %52). We then might change all loads and
11782 // stores to use Token1 as their chain operand, which could result in
11783 // copying %53 into %52 before copying %52 into %51 (which should
11786 // The problem is, however, that searching for such data dependencies
11787 // can become expensive, and the cost is not directly related to the
11788 // chain depth. Instead, we'll rule out such configurations here by
11789 // insisting that we've visited all chain users (except for users
11790 // of the original chain, which is not necessary). When doing this,
11791 // we need to look through nodes we don't care about (otherwise, things
11792 // like register copies will interfere with trivial cases).
11794 SmallVector<const SDNode *, 16> Worklist;
11795 for (SmallPtrSet<SDNode *, 16>::iterator I = Visited.begin(),
11796 IE = Visited.end(); I != IE; ++I)
11797 if (*I != OriginalChain.getNode())
11798 Worklist.push_back(*I);
11800 while (!Worklist.empty()) {
11801 const SDNode *M = Worklist.pop_back_val();
11803 // We have already visited M, and want to make sure we've visited any uses
11804 // of M that we care about. For uses that we've not visisted, and don't
11805 // care about, queue them to the worklist.
11807 for (SDNode::use_iterator UI = M->use_begin(),
11808 UIE = M->use_end(); UI != UIE; ++UI)
11809 if (UI.getUse().getValueType() == MVT::Other && Visited.insert(*UI)) {
11810 if (isa<MemIntrinsicSDNode>(*UI) || isa<MemSDNode>(*UI)) {
11811 // We've not visited this use, and we care about it (it could have an
11812 // ordering dependency with the original node).
11814 Aliases.push_back(OriginalChain);
11818 // We've not visited this use, but we don't care about it. Mark it as
11819 // visited and enqueue it to the worklist.
11820 Worklist.push_back(*UI);
11825 /// FindBetterChain - Walk up chain skipping non-aliasing memory nodes, looking
11826 /// for a better chain (aliasing node.)
11827 SDValue DAGCombiner::FindBetterChain(SDNode *N, SDValue OldChain) {
11828 SmallVector<SDValue, 8> Aliases; // Ops for replacing token factor.
11830 // Accumulate all the aliases to this node.
11831 GatherAllAliases(N, OldChain, Aliases);
11833 // If no operands then chain to entry token.
11834 if (Aliases.size() == 0)
11835 return DAG.getEntryNode();
11837 // If a single operand then chain to it. We don't need to revisit it.
11838 if (Aliases.size() == 1)
11841 // Construct a custom tailored token factor.
11842 return DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, Aliases);
11845 // SelectionDAG::Combine - This is the entry point for the file.
11847 void SelectionDAG::Combine(CombineLevel Level, AliasAnalysis &AA,
11848 CodeGenOpt::Level OptLevel) {
11849 /// run - This is the main entry point to this class.
11851 DAGCombiner(*this, AA, OptLevel).Run(Level);