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 /// \brief Set of nodes which have been combined (at least once).
109 /// This is used to allow us to reliably add any operands of a DAG node
110 /// which have not yet been combined to the worklist.
111 SmallPtrSet<SDNode *, 64> CombinedNodes;
113 // AA - Used for DAG load/store alias analysis.
116 /// AddUsersToWorklist - When an instruction is simplified, add all users of
117 /// the instruction to the work lists because they might get more simplified
120 void AddUsersToWorklist(SDNode *N) {
121 for (SDNode *Node : N->uses())
125 /// visit - call the node-specific routine that knows how to fold each
126 /// particular type of node.
127 SDValue visit(SDNode *N);
130 /// AddToWorklist - Add to the work list making sure its instance is at the
131 /// back (next to be processed.)
132 void AddToWorklist(SDNode *N) {
133 // Skip handle nodes as they can't usefully be combined and confuse the
134 // zero-use deletion strategy.
135 if (N->getOpcode() == ISD::HANDLENODE)
138 if (WorklistMap.insert(std::make_pair(N, Worklist.size())).second)
139 Worklist.push_back(N);
142 /// removeFromWorklist - remove all instances of N from the worklist.
144 void removeFromWorklist(SDNode *N) {
145 CombinedNodes.erase(N);
147 auto It = WorklistMap.find(N);
148 if (It == WorklistMap.end())
149 return; // Not in the worklist.
151 // Null out the entry rather than erasing it to avoid a linear operation.
152 Worklist[It->second] = nullptr;
153 WorklistMap.erase(It);
156 bool recursivelyDeleteUnusedNodes(SDNode *N);
158 SDValue CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
161 SDValue CombineTo(SDNode *N, SDValue Res, bool AddTo = true) {
162 return CombineTo(N, &Res, 1, AddTo);
165 SDValue CombineTo(SDNode *N, SDValue Res0, SDValue Res1,
167 SDValue To[] = { Res0, Res1 };
168 return CombineTo(N, To, 2, AddTo);
171 void CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO);
175 /// SimplifyDemandedBits - Check the specified integer node value to see if
176 /// it can be simplified or if things it uses can be simplified by bit
177 /// propagation. If so, return true.
178 bool SimplifyDemandedBits(SDValue Op) {
179 unsigned BitWidth = Op.getValueType().getScalarType().getSizeInBits();
180 APInt Demanded = APInt::getAllOnesValue(BitWidth);
181 return SimplifyDemandedBits(Op, Demanded);
184 bool SimplifyDemandedBits(SDValue Op, const APInt &Demanded);
186 bool CombineToPreIndexedLoadStore(SDNode *N);
187 bool CombineToPostIndexedLoadStore(SDNode *N);
188 bool SliceUpLoad(SDNode *N);
190 void ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad);
191 SDValue PromoteOperand(SDValue Op, EVT PVT, bool &Replace);
192 SDValue SExtPromoteOperand(SDValue Op, EVT PVT);
193 SDValue ZExtPromoteOperand(SDValue Op, EVT PVT);
194 SDValue PromoteIntBinOp(SDValue Op);
195 SDValue PromoteIntShiftOp(SDValue Op);
196 SDValue PromoteExtend(SDValue Op);
197 bool PromoteLoad(SDValue Op);
199 void ExtendSetCCUses(const SmallVectorImpl<SDNode *> &SetCCs,
200 SDValue Trunc, SDValue ExtLoad, SDLoc DL,
201 ISD::NodeType ExtType);
203 /// combine - call the node-specific routine that knows how to fold each
204 /// particular type of node. If that doesn't do anything, try the
205 /// target-specific DAG combines.
206 SDValue combine(SDNode *N);
208 // Visitation implementation - Implement dag node combining for different
209 // node types. The semantics are as follows:
211 // SDValue.getNode() == 0 - No change was made
212 // SDValue.getNode() == N - N was replaced, is dead and has been handled.
213 // otherwise - N should be replaced by the returned Operand.
215 SDValue visitTokenFactor(SDNode *N);
216 SDValue visitMERGE_VALUES(SDNode *N);
217 SDValue visitADD(SDNode *N);
218 SDValue visitSUB(SDNode *N);
219 SDValue visitADDC(SDNode *N);
220 SDValue visitSUBC(SDNode *N);
221 SDValue visitADDE(SDNode *N);
222 SDValue visitSUBE(SDNode *N);
223 SDValue visitMUL(SDNode *N);
224 SDValue visitSDIV(SDNode *N);
225 SDValue visitUDIV(SDNode *N);
226 SDValue visitSREM(SDNode *N);
227 SDValue visitUREM(SDNode *N);
228 SDValue visitMULHU(SDNode *N);
229 SDValue visitMULHS(SDNode *N);
230 SDValue visitSMUL_LOHI(SDNode *N);
231 SDValue visitUMUL_LOHI(SDNode *N);
232 SDValue visitSMULO(SDNode *N);
233 SDValue visitUMULO(SDNode *N);
234 SDValue visitSDIVREM(SDNode *N);
235 SDValue visitUDIVREM(SDNode *N);
236 SDValue visitAND(SDNode *N);
237 SDValue visitOR(SDNode *N);
238 SDValue visitXOR(SDNode *N);
239 SDValue SimplifyVBinOp(SDNode *N);
240 SDValue SimplifyVUnaryOp(SDNode *N);
241 SDValue visitSHL(SDNode *N);
242 SDValue visitSRA(SDNode *N);
243 SDValue visitSRL(SDNode *N);
244 SDValue visitRotate(SDNode *N);
245 SDValue visitCTLZ(SDNode *N);
246 SDValue visitCTLZ_ZERO_UNDEF(SDNode *N);
247 SDValue visitCTTZ(SDNode *N);
248 SDValue visitCTTZ_ZERO_UNDEF(SDNode *N);
249 SDValue visitCTPOP(SDNode *N);
250 SDValue visitSELECT(SDNode *N);
251 SDValue visitVSELECT(SDNode *N);
252 SDValue visitSELECT_CC(SDNode *N);
253 SDValue visitSETCC(SDNode *N);
254 SDValue visitSIGN_EXTEND(SDNode *N);
255 SDValue visitZERO_EXTEND(SDNode *N);
256 SDValue visitANY_EXTEND(SDNode *N);
257 SDValue visitSIGN_EXTEND_INREG(SDNode *N);
258 SDValue visitTRUNCATE(SDNode *N);
259 SDValue visitBITCAST(SDNode *N);
260 SDValue visitBUILD_PAIR(SDNode *N);
261 SDValue visitFADD(SDNode *N);
262 SDValue visitFSUB(SDNode *N);
263 SDValue visitFMUL(SDNode *N);
264 SDValue visitFMA(SDNode *N);
265 SDValue visitFDIV(SDNode *N);
266 SDValue visitFREM(SDNode *N);
267 SDValue visitFCOPYSIGN(SDNode *N);
268 SDValue visitSINT_TO_FP(SDNode *N);
269 SDValue visitUINT_TO_FP(SDNode *N);
270 SDValue visitFP_TO_SINT(SDNode *N);
271 SDValue visitFP_TO_UINT(SDNode *N);
272 SDValue visitFP_ROUND(SDNode *N);
273 SDValue visitFP_ROUND_INREG(SDNode *N);
274 SDValue visitFP_EXTEND(SDNode *N);
275 SDValue visitFNEG(SDNode *N);
276 SDValue visitFABS(SDNode *N);
277 SDValue visitFCEIL(SDNode *N);
278 SDValue visitFTRUNC(SDNode *N);
279 SDValue visitFFLOOR(SDNode *N);
280 SDValue visitBRCOND(SDNode *N);
281 SDValue visitBR_CC(SDNode *N);
282 SDValue visitLOAD(SDNode *N);
283 SDValue visitSTORE(SDNode *N);
284 SDValue visitINSERT_VECTOR_ELT(SDNode *N);
285 SDValue visitEXTRACT_VECTOR_ELT(SDNode *N);
286 SDValue visitBUILD_VECTOR(SDNode *N);
287 SDValue visitCONCAT_VECTORS(SDNode *N);
288 SDValue visitEXTRACT_SUBVECTOR(SDNode *N);
289 SDValue visitVECTOR_SHUFFLE(SDNode *N);
290 SDValue visitINSERT_SUBVECTOR(SDNode *N);
292 SDValue XformToShuffleWithZero(SDNode *N);
293 SDValue ReassociateOps(unsigned Opc, SDLoc DL, SDValue LHS, SDValue RHS);
295 SDValue visitShiftByConstant(SDNode *N, ConstantSDNode *Amt);
297 bool SimplifySelectOps(SDNode *SELECT, SDValue LHS, SDValue RHS);
298 SDValue SimplifyBinOpWithSameOpcodeHands(SDNode *N);
299 SDValue SimplifySelect(SDLoc DL, SDValue N0, SDValue N1, SDValue N2);
300 SDValue SimplifySelectCC(SDLoc DL, SDValue N0, SDValue N1, SDValue N2,
301 SDValue N3, ISD::CondCode CC,
302 bool NotExtCompare = false);
303 SDValue SimplifySetCC(EVT VT, SDValue N0, SDValue N1, ISD::CondCode Cond,
304 SDLoc DL, bool foldBooleans = true);
306 bool isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS,
308 bool isOneUseSetCC(SDValue N) const;
310 SDValue SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
312 SDValue CombineConsecutiveLoads(SDNode *N, EVT VT);
313 SDValue ConstantFoldBITCASTofBUILD_VECTOR(SDNode *, EVT);
314 SDValue BuildSDIV(SDNode *N);
315 SDValue BuildSDIVPow2(SDNode *N);
316 SDValue BuildUDIV(SDNode *N);
317 SDValue MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
318 bool DemandHighBits = true);
319 SDValue MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1);
320 SDNode *MatchRotatePosNeg(SDValue Shifted, SDValue Pos, SDValue Neg,
321 SDValue InnerPos, SDValue InnerNeg,
322 unsigned PosOpcode, unsigned NegOpcode,
324 SDNode *MatchRotate(SDValue LHS, SDValue RHS, SDLoc DL);
325 SDValue ReduceLoadWidth(SDNode *N);
326 SDValue ReduceLoadOpStoreWidth(SDNode *N);
327 SDValue TransformFPLoadStorePair(SDNode *N);
328 SDValue reduceBuildVecExtToExtBuildVec(SDNode *N);
329 SDValue reduceBuildVecConvertToConvertBuildVec(SDNode *N);
331 SDValue GetDemandedBits(SDValue V, const APInt &Mask);
333 /// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes,
334 /// looking for aliasing nodes and adding them to the Aliases vector.
335 void GatherAllAliases(SDNode *N, SDValue OriginalChain,
336 SmallVectorImpl<SDValue> &Aliases);
338 /// isAlias - Return true if there is any possibility that the two addresses
340 bool isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) const;
342 /// FindBetterChain - Walk up chain skipping non-aliasing memory nodes,
343 /// looking for a better chain (aliasing node.)
344 SDValue FindBetterChain(SDNode *N, SDValue Chain);
346 /// Merge consecutive store operations into a wide store.
347 /// This optimization uses wide integers or vectors when possible.
348 /// \return True if some memory operations were changed.
349 bool MergeConsecutiveStores(StoreSDNode *N);
351 /// \brief Try to transform a truncation where C is a constant:
352 /// (trunc (and X, C)) -> (and (trunc X), (trunc C))
354 /// \p N needs to be a truncation and its first operand an AND. Other
355 /// requirements are checked by the function (e.g. that trunc is
356 /// single-use) and if missed an empty SDValue is returned.
357 SDValue distributeTruncateThroughAnd(SDNode *N);
360 DAGCombiner(SelectionDAG &D, AliasAnalysis &A, CodeGenOpt::Level OL)
361 : DAG(D), TLI(D.getTargetLoweringInfo()), Level(BeforeLegalizeTypes),
362 OptLevel(OL), LegalOperations(false), LegalTypes(false), AA(A) {
363 AttributeSet FnAttrs =
364 DAG.getMachineFunction().getFunction()->getAttributes();
366 FnAttrs.hasAttribute(AttributeSet::FunctionIndex,
367 Attribute::OptimizeForSize) ||
368 FnAttrs.hasAttribute(AttributeSet::FunctionIndex, Attribute::MinSize);
371 /// Run - runs the dag combiner on all nodes in the work list
372 void Run(CombineLevel AtLevel);
374 SelectionDAG &getDAG() const { return DAG; }
376 /// getShiftAmountTy - Returns a type large enough to hold any valid
377 /// shift amount - before type legalization these can be huge.
378 EVT getShiftAmountTy(EVT LHSTy) {
379 assert(LHSTy.isInteger() && "Shift amount is not an integer type!");
380 if (LHSTy.isVector())
382 return LegalTypes ? TLI.getScalarShiftAmountTy(LHSTy)
383 : TLI.getPointerTy();
386 /// isTypeLegal - This method returns true if we are running before type
387 /// legalization or if the specified VT is legal.
388 bool isTypeLegal(const EVT &VT) {
389 if (!LegalTypes) return true;
390 return TLI.isTypeLegal(VT);
393 /// getSetCCResultType - Convenience wrapper around
394 /// TargetLowering::getSetCCResultType
395 EVT getSetCCResultType(EVT VT) const {
396 return TLI.getSetCCResultType(*DAG.getContext(), VT);
403 /// WorklistRemover - This class is a DAGUpdateListener that removes any deleted
404 /// nodes from the worklist.
405 class WorklistRemover : public SelectionDAG::DAGUpdateListener {
408 explicit WorklistRemover(DAGCombiner &dc)
409 : SelectionDAG::DAGUpdateListener(dc.getDAG()), DC(dc) {}
411 void NodeDeleted(SDNode *N, SDNode *E) override {
412 DC.removeFromWorklist(N);
417 //===----------------------------------------------------------------------===//
418 // TargetLowering::DAGCombinerInfo implementation
419 //===----------------------------------------------------------------------===//
421 void TargetLowering::DAGCombinerInfo::AddToWorklist(SDNode *N) {
422 ((DAGCombiner*)DC)->AddToWorklist(N);
425 void TargetLowering::DAGCombinerInfo::RemoveFromWorklist(SDNode *N) {
426 ((DAGCombiner*)DC)->removeFromWorklist(N);
429 SDValue TargetLowering::DAGCombinerInfo::
430 CombineTo(SDNode *N, const std::vector<SDValue> &To, bool AddTo) {
431 return ((DAGCombiner*)DC)->CombineTo(N, &To[0], To.size(), AddTo);
434 SDValue TargetLowering::DAGCombinerInfo::
435 CombineTo(SDNode *N, SDValue Res, bool AddTo) {
436 return ((DAGCombiner*)DC)->CombineTo(N, Res, AddTo);
440 SDValue TargetLowering::DAGCombinerInfo::
441 CombineTo(SDNode *N, SDValue Res0, SDValue Res1, bool AddTo) {
442 return ((DAGCombiner*)DC)->CombineTo(N, Res0, Res1, AddTo);
445 void TargetLowering::DAGCombinerInfo::
446 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
447 return ((DAGCombiner*)DC)->CommitTargetLoweringOpt(TLO);
450 //===----------------------------------------------------------------------===//
452 //===----------------------------------------------------------------------===//
454 /// isNegatibleForFree - Return 1 if we can compute the negated form of the
455 /// specified expression for the same cost as the expression itself, or 2 if we
456 /// can compute the negated form more cheaply than the expression itself.
457 static char isNegatibleForFree(SDValue Op, bool LegalOperations,
458 const TargetLowering &TLI,
459 const TargetOptions *Options,
460 unsigned Depth = 0) {
461 // fneg is removable even if it has multiple uses.
462 if (Op.getOpcode() == ISD::FNEG) return 2;
464 // Don't allow anything with multiple uses.
465 if (!Op.hasOneUse()) return 0;
467 // Don't recurse exponentially.
468 if (Depth > 6) return 0;
470 switch (Op.getOpcode()) {
471 default: return false;
472 case ISD::ConstantFP:
473 // Don't invert constant FP values after legalize. The negated constant
474 // isn't necessarily legal.
475 return LegalOperations ? 0 : 1;
477 // FIXME: determine better conditions for this xform.
478 if (!Options->UnsafeFPMath) return 0;
480 // After operation legalization, it might not be legal to create new FSUBs.
481 if (LegalOperations &&
482 !TLI.isOperationLegalOrCustom(ISD::FSUB, Op.getValueType()))
485 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B)
486 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI,
489 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
490 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options,
493 // We can't turn -(A-B) into B-A when we honor signed zeros.
494 if (!Options->UnsafeFPMath) return 0;
496 // fold (fneg (fsub A, B)) -> (fsub B, A)
501 if (Options->HonorSignDependentRoundingFPMath()) return 0;
503 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y) or (fmul X, (fneg Y))
504 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI,
508 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options,
514 return isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI, Options,
519 /// GetNegatedExpression - If isNegatibleForFree returns true, this function
520 /// returns the newly negated expression.
521 static SDValue GetNegatedExpression(SDValue Op, SelectionDAG &DAG,
522 bool LegalOperations, unsigned Depth = 0) {
523 // fneg is removable even if it has multiple uses.
524 if (Op.getOpcode() == ISD::FNEG) return Op.getOperand(0);
526 // Don't allow anything with multiple uses.
527 assert(Op.hasOneUse() && "Unknown reuse!");
529 assert(Depth <= 6 && "GetNegatedExpression doesn't match isNegatibleForFree");
530 switch (Op.getOpcode()) {
531 default: llvm_unreachable("Unknown code");
532 case ISD::ConstantFP: {
533 APFloat V = cast<ConstantFPSDNode>(Op)->getValueAPF();
535 return DAG.getConstantFP(V, Op.getValueType());
538 // FIXME: determine better conditions for this xform.
539 assert(DAG.getTarget().Options.UnsafeFPMath);
541 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B)
542 if (isNegatibleForFree(Op.getOperand(0), LegalOperations,
543 DAG.getTargetLoweringInfo(),
544 &DAG.getTarget().Options, Depth+1))
545 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
546 GetNegatedExpression(Op.getOperand(0), DAG,
547 LegalOperations, Depth+1),
549 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
550 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
551 GetNegatedExpression(Op.getOperand(1), DAG,
552 LegalOperations, Depth+1),
555 // We can't turn -(A-B) into B-A when we honor signed zeros.
556 assert(DAG.getTarget().Options.UnsafeFPMath);
558 // fold (fneg (fsub 0, B)) -> B
559 if (ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(Op.getOperand(0)))
560 if (N0CFP->getValueAPF().isZero())
561 return Op.getOperand(1);
563 // fold (fneg (fsub A, B)) -> (fsub B, A)
564 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
565 Op.getOperand(1), Op.getOperand(0));
569 assert(!DAG.getTarget().Options.HonorSignDependentRoundingFPMath());
571 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y)
572 if (isNegatibleForFree(Op.getOperand(0), LegalOperations,
573 DAG.getTargetLoweringInfo(),
574 &DAG.getTarget().Options, Depth+1))
575 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
576 GetNegatedExpression(Op.getOperand(0), DAG,
577 LegalOperations, Depth+1),
580 // fold (fneg (fmul X, Y)) -> (fmul X, (fneg Y))
581 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
583 GetNegatedExpression(Op.getOperand(1), DAG,
584 LegalOperations, Depth+1));
588 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
589 GetNegatedExpression(Op.getOperand(0), DAG,
590 LegalOperations, Depth+1));
592 return DAG.getNode(ISD::FP_ROUND, SDLoc(Op), Op.getValueType(),
593 GetNegatedExpression(Op.getOperand(0), DAG,
594 LegalOperations, Depth+1),
599 // isSetCCEquivalent - Return true if this node is a setcc, or is a select_cc
600 // that selects between the target values used for true and false, making it
601 // equivalent to a setcc. Also, set the incoming LHS, RHS, and CC references to
602 // the appropriate nodes based on the type of node we are checking. This
603 // simplifies life a bit for the callers.
604 bool DAGCombiner::isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS,
606 if (N.getOpcode() == ISD::SETCC) {
607 LHS = N.getOperand(0);
608 RHS = N.getOperand(1);
609 CC = N.getOperand(2);
613 if (N.getOpcode() != ISD::SELECT_CC ||
614 !TLI.isConstTrueVal(N.getOperand(2).getNode()) ||
615 !TLI.isConstFalseVal(N.getOperand(3).getNode()))
618 LHS = N.getOperand(0);
619 RHS = N.getOperand(1);
620 CC = N.getOperand(4);
624 // isOneUseSetCC - Return true if this is a SetCC-equivalent operation with only
625 // one use. If this is true, it allows the users to invert the operation for
626 // free when it is profitable to do so.
627 bool DAGCombiner::isOneUseSetCC(SDValue N) const {
629 if (isSetCCEquivalent(N, N0, N1, N2) && N.getNode()->hasOneUse())
634 /// isConstantSplatVector - Returns true if N is a BUILD_VECTOR node whose
635 /// elements are all the same constant or undefined.
636 static bool isConstantSplatVector(SDNode *N, APInt& SplatValue) {
637 BuildVectorSDNode *C = dyn_cast<BuildVectorSDNode>(N);
642 unsigned SplatBitSize;
644 EVT EltVT = N->getValueType(0).getVectorElementType();
645 return (C->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
647 EltVT.getSizeInBits() >= SplatBitSize);
650 // \brief Returns the SDNode if it is a constant BuildVector or constant.
651 static SDNode *isConstantBuildVectorOrConstantInt(SDValue N) {
652 if (isa<ConstantSDNode>(N))
654 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N);
655 if(BV && BV->isConstant())
660 // \brief Returns the SDNode if it is a constant splat BuildVector or constant
662 static ConstantSDNode *isConstOrConstSplat(SDValue N) {
663 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N))
666 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N)) {
667 BitVector UndefElements;
668 ConstantSDNode *CN = BV->getConstantSplatNode(&UndefElements);
670 // BuildVectors can truncate their operands. Ignore that case here.
671 // FIXME: We blindly ignore splats which include undef which is overly
673 if (CN && UndefElements.none() &&
674 CN->getValueType(0) == N.getValueType().getScalarType())
681 SDValue DAGCombiner::ReassociateOps(unsigned Opc, SDLoc DL,
682 SDValue N0, SDValue N1) {
683 EVT VT = N0.getValueType();
684 if (N0.getOpcode() == Opc) {
685 if (SDNode *L = isConstantBuildVectorOrConstantInt(N0.getOperand(1))) {
686 if (SDNode *R = isConstantBuildVectorOrConstantInt(N1)) {
687 // reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2))
688 SDValue OpNode = DAG.FoldConstantArithmetic(Opc, VT, L, R);
689 if (!OpNode.getNode())
691 return DAG.getNode(Opc, DL, VT, N0.getOperand(0), OpNode);
693 if (N0.hasOneUse()) {
694 // reassoc. (op (op x, c1), y) -> (op (op x, y), c1) iff x+c1 has one
696 SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N0.getOperand(0), N1);
697 if (!OpNode.getNode())
699 AddToWorklist(OpNode.getNode());
700 return DAG.getNode(Opc, DL, VT, OpNode, N0.getOperand(1));
705 if (N1.getOpcode() == Opc) {
706 if (SDNode *R = isConstantBuildVectorOrConstantInt(N1.getOperand(1))) {
707 if (SDNode *L = isConstantBuildVectorOrConstantInt(N0)) {
708 // reassoc. (op c2, (op x, c1)) -> (op x, (op c1, c2))
709 SDValue OpNode = DAG.FoldConstantArithmetic(Opc, VT, R, L);
710 if (!OpNode.getNode())
712 return DAG.getNode(Opc, DL, VT, N1.getOperand(0), OpNode);
714 if (N1.hasOneUse()) {
715 // reassoc. (op y, (op x, c1)) -> (op (op x, y), c1) iff x+c1 has one
717 SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N1.getOperand(0), N0);
718 if (!OpNode.getNode())
720 AddToWorklist(OpNode.getNode());
721 return DAG.getNode(Opc, DL, VT, OpNode, N1.getOperand(1));
729 SDValue DAGCombiner::CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
731 assert(N->getNumValues() == NumTo && "Broken CombineTo call!");
733 DEBUG(dbgs() << "\nReplacing.1 ";
735 dbgs() << "\nWith: ";
736 To[0].getNode()->dump(&DAG);
737 dbgs() << " and " << NumTo-1 << " other values\n";
738 for (unsigned i = 0, e = NumTo; i != e; ++i)
739 assert((!To[i].getNode() ||
740 N->getValueType(i) == To[i].getValueType()) &&
741 "Cannot combine value to value of different type!"));
742 WorklistRemover DeadNodes(*this);
743 DAG.ReplaceAllUsesWith(N, To);
745 // Push the new nodes and any users onto the worklist
746 for (unsigned i = 0, e = NumTo; i != e; ++i) {
747 if (To[i].getNode()) {
748 AddToWorklist(To[i].getNode());
749 AddUsersToWorklist(To[i].getNode());
754 // Finally, if the node is now dead, remove it from the graph. The node
755 // may not be dead if the replacement process recursively simplified to
756 // something else needing this node.
757 if (N->use_empty()) {
758 // Nodes can be reintroduced into the worklist. Make sure we do not
759 // process a node that has been replaced.
760 removeFromWorklist(N);
762 // Finally, since the node is now dead, remove it from the graph.
765 return SDValue(N, 0);
769 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
770 // Replace all uses. If any nodes become isomorphic to other nodes and
771 // are deleted, make sure to remove them from our worklist.
772 WorklistRemover DeadNodes(*this);
773 DAG.ReplaceAllUsesOfValueWith(TLO.Old, TLO.New);
775 // Push the new node and any (possibly new) users onto the worklist.
776 AddToWorklist(TLO.New.getNode());
777 AddUsersToWorklist(TLO.New.getNode());
779 // Finally, if the node is now dead, remove it from the graph. The node
780 // may not be dead if the replacement process recursively simplified to
781 // something else needing this node.
782 if (TLO.Old.getNode()->use_empty()) {
783 removeFromWorklist(TLO.Old.getNode());
785 // If the operands of this node are only used by the node, they will now
786 // be dead. Make sure to visit them first to delete dead nodes early.
787 for (unsigned i = 0, e = TLO.Old.getNode()->getNumOperands(); i != e; ++i)
788 if (TLO.Old.getNode()->getOperand(i).getNode()->hasOneUse())
789 AddToWorklist(TLO.Old.getNode()->getOperand(i).getNode());
791 DAG.DeleteNode(TLO.Old.getNode());
795 /// SimplifyDemandedBits - Check the specified integer node value to see if
796 /// it can be simplified or if things it uses can be simplified by bit
797 /// propagation. If so, return true.
798 bool DAGCombiner::SimplifyDemandedBits(SDValue Op, const APInt &Demanded) {
799 TargetLowering::TargetLoweringOpt TLO(DAG, LegalTypes, LegalOperations);
800 APInt KnownZero, KnownOne;
801 if (!TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
805 AddToWorklist(Op.getNode());
807 // Replace the old value with the new one.
809 DEBUG(dbgs() << "\nReplacing.2 ";
810 TLO.Old.getNode()->dump(&DAG);
811 dbgs() << "\nWith: ";
812 TLO.New.getNode()->dump(&DAG);
815 CommitTargetLoweringOpt(TLO);
819 void DAGCombiner::ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad) {
821 EVT VT = Load->getValueType(0);
822 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, VT, SDValue(ExtLoad, 0));
824 DEBUG(dbgs() << "\nReplacing.9 ";
826 dbgs() << "\nWith: ";
827 Trunc.getNode()->dump(&DAG);
829 WorklistRemover DeadNodes(*this);
830 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 0), Trunc);
831 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), SDValue(ExtLoad, 1));
832 removeFromWorklist(Load);
833 DAG.DeleteNode(Load);
834 AddToWorklist(Trunc.getNode());
837 SDValue DAGCombiner::PromoteOperand(SDValue Op, EVT PVT, bool &Replace) {
840 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Op)) {
841 EVT MemVT = LD->getMemoryVT();
842 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
843 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT) ? ISD::ZEXTLOAD
845 : LD->getExtensionType();
847 return DAG.getExtLoad(ExtType, dl, PVT,
848 LD->getChain(), LD->getBasePtr(),
849 MemVT, LD->getMemOperand());
852 unsigned Opc = Op.getOpcode();
855 case ISD::AssertSext:
856 return DAG.getNode(ISD::AssertSext, dl, PVT,
857 SExtPromoteOperand(Op.getOperand(0), PVT),
859 case ISD::AssertZext:
860 return DAG.getNode(ISD::AssertZext, dl, PVT,
861 ZExtPromoteOperand(Op.getOperand(0), PVT),
863 case ISD::Constant: {
865 Op.getValueType().isByteSized() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
866 return DAG.getNode(ExtOpc, dl, PVT, Op);
870 if (!TLI.isOperationLegal(ISD::ANY_EXTEND, PVT))
872 return DAG.getNode(ISD::ANY_EXTEND, dl, PVT, Op);
875 SDValue DAGCombiner::SExtPromoteOperand(SDValue Op, EVT PVT) {
876 if (!TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, PVT))
878 EVT OldVT = Op.getValueType();
880 bool Replace = false;
881 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
882 if (!NewOp.getNode())
884 AddToWorklist(NewOp.getNode());
887 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
888 return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, NewOp.getValueType(), NewOp,
889 DAG.getValueType(OldVT));
892 SDValue DAGCombiner::ZExtPromoteOperand(SDValue Op, EVT PVT) {
893 EVT OldVT = Op.getValueType();
895 bool Replace = false;
896 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
897 if (!NewOp.getNode())
899 AddToWorklist(NewOp.getNode());
902 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
903 return DAG.getZeroExtendInReg(NewOp, dl, OldVT);
906 /// PromoteIntBinOp - Promote the specified integer binary operation if the
907 /// target indicates it is beneficial. e.g. On x86, it's usually better to
908 /// promote i16 operations to i32 since i16 instructions are longer.
909 SDValue DAGCombiner::PromoteIntBinOp(SDValue Op) {
910 if (!LegalOperations)
913 EVT VT = Op.getValueType();
914 if (VT.isVector() || !VT.isInteger())
917 // If operation type is 'undesirable', e.g. i16 on x86, consider
919 unsigned Opc = Op.getOpcode();
920 if (TLI.isTypeDesirableForOp(Opc, VT))
924 // Consult target whether it is a good idea to promote this operation and
925 // what's the right type to promote it to.
926 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
927 assert(PVT != VT && "Don't know what type to promote to!");
929 bool Replace0 = false;
930 SDValue N0 = Op.getOperand(0);
931 SDValue NN0 = PromoteOperand(N0, PVT, Replace0);
935 bool Replace1 = false;
936 SDValue N1 = Op.getOperand(1);
941 NN1 = PromoteOperand(N1, PVT, Replace1);
946 AddToWorklist(NN0.getNode());
948 AddToWorklist(NN1.getNode());
951 ReplaceLoadWithPromotedLoad(N0.getNode(), NN0.getNode());
953 ReplaceLoadWithPromotedLoad(N1.getNode(), NN1.getNode());
955 DEBUG(dbgs() << "\nPromoting ";
956 Op.getNode()->dump(&DAG));
958 return DAG.getNode(ISD::TRUNCATE, dl, VT,
959 DAG.getNode(Opc, dl, PVT, NN0, NN1));
964 /// PromoteIntShiftOp - Promote the specified integer shift operation if the
965 /// target indicates it is beneficial. e.g. On x86, it's usually better to
966 /// promote i16 operations to i32 since i16 instructions are longer.
967 SDValue DAGCombiner::PromoteIntShiftOp(SDValue Op) {
968 if (!LegalOperations)
971 EVT VT = Op.getValueType();
972 if (VT.isVector() || !VT.isInteger())
975 // If operation type is 'undesirable', e.g. i16 on x86, consider
977 unsigned Opc = Op.getOpcode();
978 if (TLI.isTypeDesirableForOp(Opc, VT))
982 // Consult target whether it is a good idea to promote this operation and
983 // what's the right type to promote it to.
984 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
985 assert(PVT != VT && "Don't know what type to promote to!");
987 bool Replace = false;
988 SDValue N0 = Op.getOperand(0);
990 N0 = SExtPromoteOperand(Op.getOperand(0), PVT);
991 else if (Opc == ISD::SRL)
992 N0 = ZExtPromoteOperand(Op.getOperand(0), PVT);
994 N0 = PromoteOperand(N0, PVT, Replace);
998 AddToWorklist(N0.getNode());
1000 ReplaceLoadWithPromotedLoad(Op.getOperand(0).getNode(), N0.getNode());
1002 DEBUG(dbgs() << "\nPromoting ";
1003 Op.getNode()->dump(&DAG));
1005 return DAG.getNode(ISD::TRUNCATE, dl, VT,
1006 DAG.getNode(Opc, dl, PVT, N0, Op.getOperand(1)));
1011 SDValue DAGCombiner::PromoteExtend(SDValue Op) {
1012 if (!LegalOperations)
1015 EVT VT = Op.getValueType();
1016 if (VT.isVector() || !VT.isInteger())
1019 // If operation type is 'undesirable', e.g. i16 on x86, consider
1021 unsigned Opc = Op.getOpcode();
1022 if (TLI.isTypeDesirableForOp(Opc, VT))
1026 // Consult target whether it is a good idea to promote this operation and
1027 // what's the right type to promote it to.
1028 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1029 assert(PVT != VT && "Don't know what type to promote to!");
1030 // fold (aext (aext x)) -> (aext x)
1031 // fold (aext (zext x)) -> (zext x)
1032 // fold (aext (sext x)) -> (sext x)
1033 DEBUG(dbgs() << "\nPromoting ";
1034 Op.getNode()->dump(&DAG));
1035 return DAG.getNode(Op.getOpcode(), SDLoc(Op), VT, Op.getOperand(0));
1040 bool DAGCombiner::PromoteLoad(SDValue Op) {
1041 if (!LegalOperations)
1044 EVT VT = Op.getValueType();
1045 if (VT.isVector() || !VT.isInteger())
1048 // If operation type is 'undesirable', e.g. i16 on x86, consider
1050 unsigned Opc = Op.getOpcode();
1051 if (TLI.isTypeDesirableForOp(Opc, VT))
1055 // Consult target whether it is a good idea to promote this operation and
1056 // what's the right type to promote it to.
1057 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1058 assert(PVT != VT && "Don't know what type to promote to!");
1061 SDNode *N = Op.getNode();
1062 LoadSDNode *LD = cast<LoadSDNode>(N);
1063 EVT MemVT = LD->getMemoryVT();
1064 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
1065 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT) ? ISD::ZEXTLOAD
1067 : LD->getExtensionType();
1068 SDValue NewLD = DAG.getExtLoad(ExtType, dl, PVT,
1069 LD->getChain(), LD->getBasePtr(),
1070 MemVT, LD->getMemOperand());
1071 SDValue Result = DAG.getNode(ISD::TRUNCATE, dl, VT, NewLD);
1073 DEBUG(dbgs() << "\nPromoting ";
1076 Result.getNode()->dump(&DAG);
1078 WorklistRemover DeadNodes(*this);
1079 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result);
1080 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), NewLD.getValue(1));
1081 removeFromWorklist(N);
1083 AddToWorklist(Result.getNode());
1089 /// \brief Recursively delete a node which has no uses and any operands for
1090 /// which it is the only use.
1092 /// Note that this both deletes the nodes and removes them from the worklist.
1093 /// It also adds any nodes who have had a user deleted to the worklist as they
1094 /// may now have only one use and subject to other combines.
1095 bool DAGCombiner::recursivelyDeleteUnusedNodes(SDNode *N) {
1096 if (!N->use_empty())
1099 SmallSetVector<SDNode *, 16> Nodes;
1102 N = Nodes.pop_back_val();
1106 if (N->use_empty()) {
1107 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1108 Nodes.insert(N->getOperand(i).getNode());
1110 removeFromWorklist(N);
1115 } while (!Nodes.empty());
1119 //===----------------------------------------------------------------------===//
1120 // Main DAG Combiner implementation
1121 //===----------------------------------------------------------------------===//
1123 void DAGCombiner::Run(CombineLevel AtLevel) {
1124 // set the instance variables, so that the various visit routines may use it.
1126 LegalOperations = Level >= AfterLegalizeVectorOps;
1127 LegalTypes = Level >= AfterLegalizeTypes;
1129 // Add all the dag nodes to the worklist.
1130 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
1131 E = DAG.allnodes_end(); I != E; ++I)
1134 // Create a dummy node (which is not added to allnodes), that adds a reference
1135 // to the root node, preventing it from being deleted, and tracking any
1136 // changes of the root.
1137 HandleSDNode Dummy(DAG.getRoot());
1139 // while the worklist isn't empty, find a node and
1140 // try and combine it.
1141 while (!WorklistMap.empty()) {
1143 // The Worklist holds the SDNodes in order, but it may contain null entries.
1145 N = Worklist.pop_back_val();
1148 bool GoodWorklistEntry = WorklistMap.erase(N);
1149 (void)GoodWorklistEntry;
1150 assert(GoodWorklistEntry &&
1151 "Found a worklist entry without a corresponding map entry!");
1153 // If N has no uses, it is dead. Make sure to revisit all N's operands once
1154 // N is deleted from the DAG, since they too may now be dead or may have a
1155 // reduced number of uses, allowing other xforms.
1156 if (recursivelyDeleteUnusedNodes(N))
1159 DEBUG(dbgs() << "\nCombining: ";
1162 // Add any operands of the new node which have not yet been combined to the
1163 // worklist as well. Because the worklist uniques things already, this
1164 // won't repeatedly process the same operand.
1165 CombinedNodes.insert(N);
1166 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1167 if (!CombinedNodes.count(N->getOperand(i).getNode()))
1168 AddToWorklist(N->getOperand(i).getNode());
1170 WorklistRemover DeadNodes(*this);
1172 // If this combine is running after legalizing the DAG, re-legalize any
1173 // nodes pulled off the worklist.
1174 if (Level == AfterLegalizeDAG) {
1175 SmallSetVector<SDNode *, 16> UpdatedNodes;
1176 bool NIsValid = DAG.LegalizeOp(N, UpdatedNodes);
1178 for (SDNode *LN : UpdatedNodes) {
1180 AddUsersToWorklist(LN);
1186 SDValue RV = combine(N);
1193 // If we get back the same node we passed in, rather than a new node or
1194 // zero, we know that the node must have defined multiple values and
1195 // CombineTo was used. Since CombineTo takes care of the worklist
1196 // mechanics for us, we have no work to do in this case.
1197 if (RV.getNode() == N)
1200 assert(N->getOpcode() != ISD::DELETED_NODE &&
1201 RV.getNode()->getOpcode() != ISD::DELETED_NODE &&
1202 "Node was deleted but visit returned new node!");
1204 DEBUG(dbgs() << " ... into: ";
1205 RV.getNode()->dump(&DAG));
1207 // Transfer debug value.
1208 DAG.TransferDbgValues(SDValue(N, 0), RV);
1209 if (N->getNumValues() == RV.getNode()->getNumValues())
1210 DAG.ReplaceAllUsesWith(N, RV.getNode());
1212 assert(N->getValueType(0) == RV.getValueType() &&
1213 N->getNumValues() == 1 && "Type mismatch");
1215 DAG.ReplaceAllUsesWith(N, &OpV);
1218 // Push the new node and any users onto the worklist
1219 AddToWorklist(RV.getNode());
1220 AddUsersToWorklist(RV.getNode());
1222 // Finally, if the node is now dead, remove it from the graph. The node
1223 // may not be dead if the replacement process recursively simplified to
1224 // something else needing this node. This will also take care of adding any
1225 // operands which have lost a user to the worklist.
1226 recursivelyDeleteUnusedNodes(N);
1229 // If the root changed (e.g. it was a dead load, update the root).
1230 DAG.setRoot(Dummy.getValue());
1231 DAG.RemoveDeadNodes();
1234 SDValue DAGCombiner::visit(SDNode *N) {
1235 switch (N->getOpcode()) {
1237 case ISD::TokenFactor: return visitTokenFactor(N);
1238 case ISD::MERGE_VALUES: return visitMERGE_VALUES(N);
1239 case ISD::ADD: return visitADD(N);
1240 case ISD::SUB: return visitSUB(N);
1241 case ISD::ADDC: return visitADDC(N);
1242 case ISD::SUBC: return visitSUBC(N);
1243 case ISD::ADDE: return visitADDE(N);
1244 case ISD::SUBE: return visitSUBE(N);
1245 case ISD::MUL: return visitMUL(N);
1246 case ISD::SDIV: return visitSDIV(N);
1247 case ISD::UDIV: return visitUDIV(N);
1248 case ISD::SREM: return visitSREM(N);
1249 case ISD::UREM: return visitUREM(N);
1250 case ISD::MULHU: return visitMULHU(N);
1251 case ISD::MULHS: return visitMULHS(N);
1252 case ISD::SMUL_LOHI: return visitSMUL_LOHI(N);
1253 case ISD::UMUL_LOHI: return visitUMUL_LOHI(N);
1254 case ISD::SMULO: return visitSMULO(N);
1255 case ISD::UMULO: return visitUMULO(N);
1256 case ISD::SDIVREM: return visitSDIVREM(N);
1257 case ISD::UDIVREM: return visitUDIVREM(N);
1258 case ISD::AND: return visitAND(N);
1259 case ISD::OR: return visitOR(N);
1260 case ISD::XOR: return visitXOR(N);
1261 case ISD::SHL: return visitSHL(N);
1262 case ISD::SRA: return visitSRA(N);
1263 case ISD::SRL: return visitSRL(N);
1265 case ISD::ROTL: return visitRotate(N);
1266 case ISD::CTLZ: return visitCTLZ(N);
1267 case ISD::CTLZ_ZERO_UNDEF: return visitCTLZ_ZERO_UNDEF(N);
1268 case ISD::CTTZ: return visitCTTZ(N);
1269 case ISD::CTTZ_ZERO_UNDEF: return visitCTTZ_ZERO_UNDEF(N);
1270 case ISD::CTPOP: return visitCTPOP(N);
1271 case ISD::SELECT: return visitSELECT(N);
1272 case ISD::VSELECT: return visitVSELECT(N);
1273 case ISD::SELECT_CC: return visitSELECT_CC(N);
1274 case ISD::SETCC: return visitSETCC(N);
1275 case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N);
1276 case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N);
1277 case ISD::ANY_EXTEND: return visitANY_EXTEND(N);
1278 case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N);
1279 case ISD::TRUNCATE: return visitTRUNCATE(N);
1280 case ISD::BITCAST: return visitBITCAST(N);
1281 case ISD::BUILD_PAIR: return visitBUILD_PAIR(N);
1282 case ISD::FADD: return visitFADD(N);
1283 case ISD::FSUB: return visitFSUB(N);
1284 case ISD::FMUL: return visitFMUL(N);
1285 case ISD::FMA: return visitFMA(N);
1286 case ISD::FDIV: return visitFDIV(N);
1287 case ISD::FREM: return visitFREM(N);
1288 case ISD::FCOPYSIGN: return visitFCOPYSIGN(N);
1289 case ISD::SINT_TO_FP: return visitSINT_TO_FP(N);
1290 case ISD::UINT_TO_FP: return visitUINT_TO_FP(N);
1291 case ISD::FP_TO_SINT: return visitFP_TO_SINT(N);
1292 case ISD::FP_TO_UINT: return visitFP_TO_UINT(N);
1293 case ISD::FP_ROUND: return visitFP_ROUND(N);
1294 case ISD::FP_ROUND_INREG: return visitFP_ROUND_INREG(N);
1295 case ISD::FP_EXTEND: return visitFP_EXTEND(N);
1296 case ISD::FNEG: return visitFNEG(N);
1297 case ISD::FABS: return visitFABS(N);
1298 case ISD::FFLOOR: return visitFFLOOR(N);
1299 case ISD::FCEIL: return visitFCEIL(N);
1300 case ISD::FTRUNC: return visitFTRUNC(N);
1301 case ISD::BRCOND: return visitBRCOND(N);
1302 case ISD::BR_CC: return visitBR_CC(N);
1303 case ISD::LOAD: return visitLOAD(N);
1304 case ISD::STORE: return visitSTORE(N);
1305 case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N);
1306 case ISD::EXTRACT_VECTOR_ELT: return visitEXTRACT_VECTOR_ELT(N);
1307 case ISD::BUILD_VECTOR: return visitBUILD_VECTOR(N);
1308 case ISD::CONCAT_VECTORS: return visitCONCAT_VECTORS(N);
1309 case ISD::EXTRACT_SUBVECTOR: return visitEXTRACT_SUBVECTOR(N);
1310 case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N);
1311 case ISD::INSERT_SUBVECTOR: return visitINSERT_SUBVECTOR(N);
1316 SDValue DAGCombiner::combine(SDNode *N) {
1317 SDValue RV = visit(N);
1319 // If nothing happened, try a target-specific DAG combine.
1320 if (!RV.getNode()) {
1321 assert(N->getOpcode() != ISD::DELETED_NODE &&
1322 "Node was deleted but visit returned NULL!");
1324 if (N->getOpcode() >= ISD::BUILTIN_OP_END ||
1325 TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode())) {
1327 // Expose the DAG combiner to the target combiner impls.
1328 TargetLowering::DAGCombinerInfo
1329 DagCombineInfo(DAG, Level, false, this);
1331 RV = TLI.PerformDAGCombine(N, DagCombineInfo);
1335 // If nothing happened still, try promoting the operation.
1336 if (!RV.getNode()) {
1337 switch (N->getOpcode()) {
1345 RV = PromoteIntBinOp(SDValue(N, 0));
1350 RV = PromoteIntShiftOp(SDValue(N, 0));
1352 case ISD::SIGN_EXTEND:
1353 case ISD::ZERO_EXTEND:
1354 case ISD::ANY_EXTEND:
1355 RV = PromoteExtend(SDValue(N, 0));
1358 if (PromoteLoad(SDValue(N, 0)))
1364 // If N is a commutative binary node, try commuting it to enable more
1366 if (!RV.getNode() && SelectionDAG::isCommutativeBinOp(N->getOpcode()) &&
1367 N->getNumValues() == 1) {
1368 SDValue N0 = N->getOperand(0);
1369 SDValue N1 = N->getOperand(1);
1371 // Constant operands are canonicalized to RHS.
1372 if (isa<ConstantSDNode>(N0) || !isa<ConstantSDNode>(N1)) {
1373 SDValue Ops[] = {N1, N0};
1375 if (const BinaryWithFlagsSDNode *BinNode =
1376 dyn_cast<BinaryWithFlagsSDNode>(N)) {
1377 CSENode = DAG.getNodeIfExists(
1378 N->getOpcode(), N->getVTList(), Ops, BinNode->hasNoUnsignedWrap(),
1379 BinNode->hasNoSignedWrap(), BinNode->isExact());
1381 CSENode = DAG.getNodeIfExists(N->getOpcode(), N->getVTList(), Ops);
1384 return SDValue(CSENode, 0);
1391 /// getInputChainForNode - Given a node, return its input chain if it has one,
1392 /// otherwise return a null sd operand.
1393 static SDValue getInputChainForNode(SDNode *N) {
1394 if (unsigned NumOps = N->getNumOperands()) {
1395 if (N->getOperand(0).getValueType() == MVT::Other)
1396 return N->getOperand(0);
1397 if (N->getOperand(NumOps-1).getValueType() == MVT::Other)
1398 return N->getOperand(NumOps-1);
1399 for (unsigned i = 1; i < NumOps-1; ++i)
1400 if (N->getOperand(i).getValueType() == MVT::Other)
1401 return N->getOperand(i);
1406 SDValue DAGCombiner::visitTokenFactor(SDNode *N) {
1407 // If N has two operands, where one has an input chain equal to the other,
1408 // the 'other' chain is redundant.
1409 if (N->getNumOperands() == 2) {
1410 if (getInputChainForNode(N->getOperand(0).getNode()) == N->getOperand(1))
1411 return N->getOperand(0);
1412 if (getInputChainForNode(N->getOperand(1).getNode()) == N->getOperand(0))
1413 return N->getOperand(1);
1416 SmallVector<SDNode *, 8> TFs; // List of token factors to visit.
1417 SmallVector<SDValue, 8> Ops; // Ops for replacing token factor.
1418 SmallPtrSet<SDNode*, 16> SeenOps;
1419 bool Changed = false; // If we should replace this token factor.
1421 // Start out with this token factor.
1424 // Iterate through token factors. The TFs grows when new token factors are
1426 for (unsigned i = 0; i < TFs.size(); ++i) {
1427 SDNode *TF = TFs[i];
1429 // Check each of the operands.
1430 for (unsigned i = 0, ie = TF->getNumOperands(); i != ie; ++i) {
1431 SDValue Op = TF->getOperand(i);
1433 switch (Op.getOpcode()) {
1434 case ISD::EntryToken:
1435 // Entry tokens don't need to be added to the list. They are
1440 case ISD::TokenFactor:
1441 if (Op.hasOneUse() &&
1442 std::find(TFs.begin(), TFs.end(), Op.getNode()) == TFs.end()) {
1443 // Queue up for processing.
1444 TFs.push_back(Op.getNode());
1445 // Clean up in case the token factor is removed.
1446 AddToWorklist(Op.getNode());
1453 // Only add if it isn't already in the list.
1454 if (SeenOps.insert(Op.getNode()))
1465 // If we've change things around then replace token factor.
1468 // The entry token is the only possible outcome.
1469 Result = DAG.getEntryNode();
1471 // New and improved token factor.
1472 Result = DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, Ops);
1475 // Don't add users to work list.
1476 return CombineTo(N, Result, false);
1482 /// MERGE_VALUES can always be eliminated.
1483 SDValue DAGCombiner::visitMERGE_VALUES(SDNode *N) {
1484 WorklistRemover DeadNodes(*this);
1485 // Replacing results may cause a different MERGE_VALUES to suddenly
1486 // be CSE'd with N, and carry its uses with it. Iterate until no
1487 // uses remain, to ensure that the node can be safely deleted.
1488 // First add the users of this node to the work list so that they
1489 // can be tried again once they have new operands.
1490 AddUsersToWorklist(N);
1492 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1493 DAG.ReplaceAllUsesOfValueWith(SDValue(N, i), N->getOperand(i));
1494 } while (!N->use_empty());
1495 removeFromWorklist(N);
1497 return SDValue(N, 0); // Return N so it doesn't get rechecked!
1501 SDValue combineShlAddConstant(SDLoc DL, SDValue N0, SDValue N1,
1502 SelectionDAG &DAG) {
1503 EVT VT = N0.getValueType();
1504 SDValue N00 = N0.getOperand(0);
1505 SDValue N01 = N0.getOperand(1);
1506 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N01);
1508 if (N01C && N00.getOpcode() == ISD::ADD && N00.getNode()->hasOneUse() &&
1509 isa<ConstantSDNode>(N00.getOperand(1))) {
1510 // fold (add (shl (add x, c1), c2), ) -> (add (add (shl x, c2), c1<<c2), )
1511 N0 = DAG.getNode(ISD::ADD, SDLoc(N0), VT,
1512 DAG.getNode(ISD::SHL, SDLoc(N00), VT,
1513 N00.getOperand(0), N01),
1514 DAG.getNode(ISD::SHL, SDLoc(N01), VT,
1515 N00.getOperand(1), N01));
1516 return DAG.getNode(ISD::ADD, DL, VT, N0, N1);
1522 SDValue DAGCombiner::visitADD(SDNode *N) {
1523 SDValue N0 = N->getOperand(0);
1524 SDValue N1 = N->getOperand(1);
1525 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1526 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1527 EVT VT = N0.getValueType();
1530 if (VT.isVector()) {
1531 SDValue FoldedVOp = SimplifyVBinOp(N);
1532 if (FoldedVOp.getNode()) return FoldedVOp;
1534 // fold (add x, 0) -> x, vector edition
1535 if (ISD::isBuildVectorAllZeros(N1.getNode()))
1537 if (ISD::isBuildVectorAllZeros(N0.getNode()))
1541 // fold (add x, undef) -> undef
1542 if (N0.getOpcode() == ISD::UNDEF)
1544 if (N1.getOpcode() == ISD::UNDEF)
1546 // fold (add c1, c2) -> c1+c2
1548 return DAG.FoldConstantArithmetic(ISD::ADD, VT, N0C, N1C);
1549 // canonicalize constant to RHS
1551 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N1, N0);
1552 // fold (add x, 0) -> x
1553 if (N1C && N1C->isNullValue())
1555 // fold (add Sym, c) -> Sym+c
1556 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1557 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA) && N1C &&
1558 GA->getOpcode() == ISD::GlobalAddress)
1559 return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT,
1561 (uint64_t)N1C->getSExtValue());
1562 // fold ((c1-A)+c2) -> (c1+c2)-A
1563 if (N1C && N0.getOpcode() == ISD::SUB)
1564 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getOperand(0)))
1565 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1566 DAG.getConstant(N1C->getAPIntValue()+
1567 N0C->getAPIntValue(), VT),
1570 SDValue RADD = ReassociateOps(ISD::ADD, SDLoc(N), N0, N1);
1573 // fold ((0-A) + B) -> B-A
1574 if (N0.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N0.getOperand(0)) &&
1575 cast<ConstantSDNode>(N0.getOperand(0))->isNullValue())
1576 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1, N0.getOperand(1));
1577 // fold (A + (0-B)) -> A-B
1578 if (N1.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N1.getOperand(0)) &&
1579 cast<ConstantSDNode>(N1.getOperand(0))->isNullValue())
1580 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, N1.getOperand(1));
1581 // fold (A+(B-A)) -> B
1582 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1))
1583 return N1.getOperand(0);
1584 // fold ((B-A)+A) -> B
1585 if (N0.getOpcode() == ISD::SUB && N1 == N0.getOperand(1))
1586 return N0.getOperand(0);
1587 // fold (A+(B-(A+C))) to (B-C)
1588 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1589 N0 == N1.getOperand(1).getOperand(0))
1590 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1.getOperand(0),
1591 N1.getOperand(1).getOperand(1));
1592 // fold (A+(B-(C+A))) to (B-C)
1593 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1594 N0 == N1.getOperand(1).getOperand(1))
1595 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1.getOperand(0),
1596 N1.getOperand(1).getOperand(0));
1597 // fold (A+((B-A)+or-C)) to (B+or-C)
1598 if ((N1.getOpcode() == ISD::SUB || N1.getOpcode() == ISD::ADD) &&
1599 N1.getOperand(0).getOpcode() == ISD::SUB &&
1600 N0 == N1.getOperand(0).getOperand(1))
1601 return DAG.getNode(N1.getOpcode(), SDLoc(N), VT,
1602 N1.getOperand(0).getOperand(0), N1.getOperand(1));
1604 // fold (A-B)+(C-D) to (A+C)-(B+D) when A or C is constant
1605 if (N0.getOpcode() == ISD::SUB && N1.getOpcode() == ISD::SUB) {
1606 SDValue N00 = N0.getOperand(0);
1607 SDValue N01 = N0.getOperand(1);
1608 SDValue N10 = N1.getOperand(0);
1609 SDValue N11 = N1.getOperand(1);
1611 if (isa<ConstantSDNode>(N00) || isa<ConstantSDNode>(N10))
1612 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1613 DAG.getNode(ISD::ADD, SDLoc(N0), VT, N00, N10),
1614 DAG.getNode(ISD::ADD, SDLoc(N1), VT, N01, N11));
1617 if (!VT.isVector() && SimplifyDemandedBits(SDValue(N, 0)))
1618 return SDValue(N, 0);
1620 // fold (a+b) -> (a|b) iff a and b share no bits.
1621 if (VT.isInteger() && !VT.isVector()) {
1622 APInt LHSZero, LHSOne;
1623 APInt RHSZero, RHSOne;
1624 DAG.computeKnownBits(N0, LHSZero, LHSOne);
1626 if (LHSZero.getBoolValue()) {
1627 DAG.computeKnownBits(N1, RHSZero, RHSOne);
1629 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1630 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1631 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero){
1632 if (!LegalOperations || TLI.isOperationLegal(ISD::OR, VT))
1633 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1);
1638 // fold (add (shl (add x, c1), c2), ) -> (add (add (shl x, c2), c1<<c2), )
1639 if (N0.getOpcode() == ISD::SHL && N0.getNode()->hasOneUse()) {
1640 SDValue Result = combineShlAddConstant(SDLoc(N), N0, N1, DAG);
1641 if (Result.getNode()) return Result;
1643 if (N1.getOpcode() == ISD::SHL && N1.getNode()->hasOneUse()) {
1644 SDValue Result = combineShlAddConstant(SDLoc(N), N1, N0, DAG);
1645 if (Result.getNode()) return Result;
1648 // fold (add x, shl(0 - y, n)) -> sub(x, shl(y, n))
1649 if (N1.getOpcode() == ISD::SHL &&
1650 N1.getOperand(0).getOpcode() == ISD::SUB)
1651 if (ConstantSDNode *C =
1652 dyn_cast<ConstantSDNode>(N1.getOperand(0).getOperand(0)))
1653 if (C->getAPIntValue() == 0)
1654 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N0,
1655 DAG.getNode(ISD::SHL, SDLoc(N), VT,
1656 N1.getOperand(0).getOperand(1),
1658 if (N0.getOpcode() == ISD::SHL &&
1659 N0.getOperand(0).getOpcode() == ISD::SUB)
1660 if (ConstantSDNode *C =
1661 dyn_cast<ConstantSDNode>(N0.getOperand(0).getOperand(0)))
1662 if (C->getAPIntValue() == 0)
1663 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1,
1664 DAG.getNode(ISD::SHL, SDLoc(N), VT,
1665 N0.getOperand(0).getOperand(1),
1668 if (N1.getOpcode() == ISD::AND) {
1669 SDValue AndOp0 = N1.getOperand(0);
1670 ConstantSDNode *AndOp1 = dyn_cast<ConstantSDNode>(N1->getOperand(1));
1671 unsigned NumSignBits = DAG.ComputeNumSignBits(AndOp0);
1672 unsigned DestBits = VT.getScalarType().getSizeInBits();
1674 // (add z, (and (sbbl x, x), 1)) -> (sub z, (sbbl x, x))
1675 // and similar xforms where the inner op is either ~0 or 0.
1676 if (NumSignBits == DestBits && AndOp1 && AndOp1->isOne()) {
1678 return DAG.getNode(ISD::SUB, DL, VT, N->getOperand(0), AndOp0);
1682 // add (sext i1), X -> sub X, (zext i1)
1683 if (N0.getOpcode() == ISD::SIGN_EXTEND &&
1684 N0.getOperand(0).getValueType() == MVT::i1 &&
1685 !TLI.isOperationLegal(ISD::SIGN_EXTEND, MVT::i1)) {
1687 SDValue ZExt = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0));
1688 return DAG.getNode(ISD::SUB, DL, VT, N1, ZExt);
1694 SDValue DAGCombiner::visitADDC(SDNode *N) {
1695 SDValue N0 = N->getOperand(0);
1696 SDValue N1 = N->getOperand(1);
1697 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1698 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1699 EVT VT = N0.getValueType();
1701 // If the flag result is dead, turn this into an ADD.
1702 if (!N->hasAnyUseOfValue(1))
1703 return CombineTo(N, DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N1),
1704 DAG.getNode(ISD::CARRY_FALSE,
1705 SDLoc(N), MVT::Glue));
1707 // canonicalize constant to RHS.
1709 return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N1, N0);
1711 // fold (addc x, 0) -> x + no carry out
1712 if (N1C && N1C->isNullValue())
1713 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE,
1714 SDLoc(N), MVT::Glue));
1716 // fold (addc a, b) -> (or a, b), CARRY_FALSE iff a and b share no bits.
1717 APInt LHSZero, LHSOne;
1718 APInt RHSZero, RHSOne;
1719 DAG.computeKnownBits(N0, LHSZero, LHSOne);
1721 if (LHSZero.getBoolValue()) {
1722 DAG.computeKnownBits(N1, RHSZero, RHSOne);
1724 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1725 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1726 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero)
1727 return CombineTo(N, DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1),
1728 DAG.getNode(ISD::CARRY_FALSE,
1729 SDLoc(N), MVT::Glue));
1735 SDValue DAGCombiner::visitADDE(SDNode *N) {
1736 SDValue N0 = N->getOperand(0);
1737 SDValue N1 = N->getOperand(1);
1738 SDValue CarryIn = N->getOperand(2);
1739 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1740 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1742 // canonicalize constant to RHS
1744 return DAG.getNode(ISD::ADDE, SDLoc(N), N->getVTList(),
1747 // fold (adde x, y, false) -> (addc x, y)
1748 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1749 return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N0, N1);
1754 // Since it may not be valid to emit a fold to zero for vector initializers
1755 // check if we can before folding.
1756 static SDValue tryFoldToZero(SDLoc DL, const TargetLowering &TLI, EVT VT,
1758 bool LegalOperations, bool LegalTypes) {
1760 return DAG.getConstant(0, VT);
1761 if (!LegalOperations || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
1762 return DAG.getConstant(0, VT);
1766 SDValue DAGCombiner::visitSUB(SDNode *N) {
1767 SDValue N0 = N->getOperand(0);
1768 SDValue N1 = N->getOperand(1);
1769 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1770 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1771 ConstantSDNode *N1C1 = N1.getOpcode() != ISD::ADD ? nullptr :
1772 dyn_cast<ConstantSDNode>(N1.getOperand(1).getNode());
1773 EVT VT = N0.getValueType();
1776 if (VT.isVector()) {
1777 SDValue FoldedVOp = SimplifyVBinOp(N);
1778 if (FoldedVOp.getNode()) return FoldedVOp;
1780 // fold (sub x, 0) -> x, vector edition
1781 if (ISD::isBuildVectorAllZeros(N1.getNode()))
1785 // fold (sub x, x) -> 0
1786 // FIXME: Refactor this and xor and other similar operations together.
1788 return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations, LegalTypes);
1789 // fold (sub c1, c2) -> c1-c2
1791 return DAG.FoldConstantArithmetic(ISD::SUB, VT, N0C, N1C);
1792 // fold (sub x, c) -> (add x, -c)
1794 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0,
1795 DAG.getConstant(-N1C->getAPIntValue(), VT));
1796 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1)
1797 if (N0C && N0C->isAllOnesValue())
1798 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0);
1799 // fold A-(A-B) -> B
1800 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(0))
1801 return N1.getOperand(1);
1802 // fold (A+B)-A -> B
1803 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1)
1804 return N0.getOperand(1);
1805 // fold (A+B)-B -> A
1806 if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1)
1807 return N0.getOperand(0);
1808 // fold C2-(A+C1) -> (C2-C1)-A
1809 if (N1.getOpcode() == ISD::ADD && N0C && N1C1) {
1810 SDValue NewC = DAG.getConstant(N0C->getAPIntValue() - N1C1->getAPIntValue(),
1812 return DAG.getNode(ISD::SUB, SDLoc(N), VT, NewC,
1815 // fold ((A+(B+or-C))-B) -> A+or-C
1816 if (N0.getOpcode() == ISD::ADD &&
1817 (N0.getOperand(1).getOpcode() == ISD::SUB ||
1818 N0.getOperand(1).getOpcode() == ISD::ADD) &&
1819 N0.getOperand(1).getOperand(0) == N1)
1820 return DAG.getNode(N0.getOperand(1).getOpcode(), SDLoc(N), VT,
1821 N0.getOperand(0), N0.getOperand(1).getOperand(1));
1822 // fold ((A+(C+B))-B) -> A+C
1823 if (N0.getOpcode() == ISD::ADD &&
1824 N0.getOperand(1).getOpcode() == ISD::ADD &&
1825 N0.getOperand(1).getOperand(1) == N1)
1826 return DAG.getNode(ISD::ADD, SDLoc(N), VT,
1827 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1828 // fold ((A-(B-C))-C) -> A-B
1829 if (N0.getOpcode() == ISD::SUB &&
1830 N0.getOperand(1).getOpcode() == ISD::SUB &&
1831 N0.getOperand(1).getOperand(1) == N1)
1832 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1833 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1835 // If either operand of a sub is undef, the result is undef
1836 if (N0.getOpcode() == ISD::UNDEF)
1838 if (N1.getOpcode() == ISD::UNDEF)
1841 // If the relocation model supports it, consider symbol offsets.
1842 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1843 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA)) {
1844 // fold (sub Sym, c) -> Sym-c
1845 if (N1C && GA->getOpcode() == ISD::GlobalAddress)
1846 return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT,
1848 (uint64_t)N1C->getSExtValue());
1849 // fold (sub Sym+c1, Sym+c2) -> c1-c2
1850 if (GlobalAddressSDNode *GB = dyn_cast<GlobalAddressSDNode>(N1))
1851 if (GA->getGlobal() == GB->getGlobal())
1852 return DAG.getConstant((uint64_t)GA->getOffset() - GB->getOffset(),
1859 SDValue DAGCombiner::visitSUBC(SDNode *N) {
1860 SDValue N0 = N->getOperand(0);
1861 SDValue N1 = N->getOperand(1);
1862 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1863 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1864 EVT VT = N0.getValueType();
1866 // If the flag result is dead, turn this into an SUB.
1867 if (!N->hasAnyUseOfValue(1))
1868 return CombineTo(N, DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, N1),
1869 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1872 // fold (subc x, x) -> 0 + no borrow
1874 return CombineTo(N, DAG.getConstant(0, VT),
1875 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1878 // fold (subc x, 0) -> x + no borrow
1879 if (N1C && N1C->isNullValue())
1880 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1883 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1) + no borrow
1884 if (N0C && N0C->isAllOnesValue())
1885 return CombineTo(N, DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0),
1886 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1892 SDValue DAGCombiner::visitSUBE(SDNode *N) {
1893 SDValue N0 = N->getOperand(0);
1894 SDValue N1 = N->getOperand(1);
1895 SDValue CarryIn = N->getOperand(2);
1897 // fold (sube x, y, false) -> (subc x, y)
1898 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1899 return DAG.getNode(ISD::SUBC, SDLoc(N), N->getVTList(), N0, N1);
1904 SDValue DAGCombiner::visitMUL(SDNode *N) {
1905 SDValue N0 = N->getOperand(0);
1906 SDValue N1 = N->getOperand(1);
1907 EVT VT = N0.getValueType();
1909 // fold (mul x, undef) -> 0
1910 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
1911 return DAG.getConstant(0, VT);
1913 bool N0IsConst = false;
1914 bool N1IsConst = false;
1915 APInt ConstValue0, ConstValue1;
1917 if (VT.isVector()) {
1918 SDValue FoldedVOp = SimplifyVBinOp(N);
1919 if (FoldedVOp.getNode()) return FoldedVOp;
1921 N0IsConst = isConstantSplatVector(N0.getNode(), ConstValue0);
1922 N1IsConst = isConstantSplatVector(N1.getNode(), ConstValue1);
1924 N0IsConst = dyn_cast<ConstantSDNode>(N0) != nullptr;
1925 ConstValue0 = N0IsConst ? (dyn_cast<ConstantSDNode>(N0))->getAPIntValue()
1927 N1IsConst = dyn_cast<ConstantSDNode>(N1) != nullptr;
1928 ConstValue1 = N1IsConst ? (dyn_cast<ConstantSDNode>(N1))->getAPIntValue()
1932 // fold (mul c1, c2) -> c1*c2
1933 if (N0IsConst && N1IsConst)
1934 return DAG.FoldConstantArithmetic(ISD::MUL, VT, N0.getNode(), N1.getNode());
1936 // canonicalize constant to RHS
1937 if (N0IsConst && !N1IsConst)
1938 return DAG.getNode(ISD::MUL, SDLoc(N), VT, N1, N0);
1939 // fold (mul x, 0) -> 0
1940 if (N1IsConst && ConstValue1 == 0)
1942 // We require a splat of the entire scalar bit width for non-contiguous
1945 ConstValue1.getBitWidth() == VT.getScalarType().getSizeInBits();
1946 // fold (mul x, 1) -> x
1947 if (N1IsConst && ConstValue1 == 1 && IsFullSplat)
1949 // fold (mul x, -1) -> 0-x
1950 if (N1IsConst && ConstValue1.isAllOnesValue())
1951 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1952 DAG.getConstant(0, VT), N0);
1953 // fold (mul x, (1 << c)) -> x << c
1954 if (N1IsConst && ConstValue1.isPowerOf2() && IsFullSplat)
1955 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0,
1956 DAG.getConstant(ConstValue1.logBase2(),
1957 getShiftAmountTy(N0.getValueType())));
1958 // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c
1959 if (N1IsConst && (-ConstValue1).isPowerOf2() && IsFullSplat) {
1960 unsigned Log2Val = (-ConstValue1).logBase2();
1961 // FIXME: If the input is something that is easily negated (e.g. a
1962 // single-use add), we should put the negate there.
1963 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1964 DAG.getConstant(0, VT),
1965 DAG.getNode(ISD::SHL, SDLoc(N), VT, N0,
1966 DAG.getConstant(Log2Val,
1967 getShiftAmountTy(N0.getValueType()))));
1971 // (mul (shl X, c1), c2) -> (mul X, c2 << c1)
1972 if (N1IsConst && N0.getOpcode() == ISD::SHL &&
1973 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
1974 isa<ConstantSDNode>(N0.getOperand(1)))) {
1975 SDValue C3 = DAG.getNode(ISD::SHL, SDLoc(N), VT,
1976 N1, N0.getOperand(1));
1977 AddToWorklist(C3.getNode());
1978 return DAG.getNode(ISD::MUL, SDLoc(N), VT,
1979 N0.getOperand(0), C3);
1982 // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one
1985 SDValue Sh(nullptr,0), Y(nullptr,0);
1986 // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)).
1987 if (N0.getOpcode() == ISD::SHL &&
1988 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
1989 isa<ConstantSDNode>(N0.getOperand(1))) &&
1990 N0.getNode()->hasOneUse()) {
1992 } else if (N1.getOpcode() == ISD::SHL &&
1993 isa<ConstantSDNode>(N1.getOperand(1)) &&
1994 N1.getNode()->hasOneUse()) {
1999 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2000 Sh.getOperand(0), Y);
2001 return DAG.getNode(ISD::SHL, SDLoc(N), VT,
2002 Mul, Sh.getOperand(1));
2006 // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2)
2007 if (N1IsConst && N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() &&
2008 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
2009 isa<ConstantSDNode>(N0.getOperand(1))))
2010 return DAG.getNode(ISD::ADD, SDLoc(N), VT,
2011 DAG.getNode(ISD::MUL, SDLoc(N0), VT,
2012 N0.getOperand(0), N1),
2013 DAG.getNode(ISD::MUL, SDLoc(N1), VT,
2014 N0.getOperand(1), N1));
2017 SDValue RMUL = ReassociateOps(ISD::MUL, SDLoc(N), N0, N1);
2024 SDValue DAGCombiner::visitSDIV(SDNode *N) {
2025 SDValue N0 = N->getOperand(0);
2026 SDValue N1 = N->getOperand(1);
2027 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2028 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2029 EVT VT = N->getValueType(0);
2032 if (VT.isVector()) {
2033 SDValue FoldedVOp = SimplifyVBinOp(N);
2034 if (FoldedVOp.getNode()) return FoldedVOp;
2037 // fold (sdiv c1, c2) -> c1/c2
2038 if (N0C && N1C && !N1C->isNullValue())
2039 return DAG.FoldConstantArithmetic(ISD::SDIV, VT, N0C, N1C);
2040 // fold (sdiv X, 1) -> X
2041 if (N1C && N1C->getAPIntValue() == 1LL)
2043 // fold (sdiv X, -1) -> 0-X
2044 if (N1C && N1C->isAllOnesValue())
2045 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
2046 DAG.getConstant(0, VT), N0);
2047 // If we know the sign bits of both operands are zero, strength reduce to a
2048 // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2
2049 if (!VT.isVector()) {
2050 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
2051 return DAG.getNode(ISD::UDIV, SDLoc(N), N1.getValueType(),
2055 // fold (sdiv X, pow2) -> simple ops after legalize
2056 if (N1C && !N1C->isNullValue() && (N1C->getAPIntValue().isPowerOf2() ||
2057 (-N1C->getAPIntValue()).isPowerOf2())) {
2058 // If dividing by powers of two is cheap, then don't perform the following
2060 if (TLI.isPow2DivCheap())
2063 // Target-specific implementation of sdiv x, pow2.
2064 SDValue Res = BuildSDIVPow2(N);
2068 unsigned lg2 = N1C->getAPIntValue().countTrailingZeros();
2070 // Splat the sign bit into the register
2072 DAG.getNode(ISD::SRA, SDLoc(N), VT, N0,
2073 DAG.getConstant(VT.getScalarSizeInBits() - 1,
2074 getShiftAmountTy(N0.getValueType())));
2075 AddToWorklist(SGN.getNode());
2077 // Add (N0 < 0) ? abs2 - 1 : 0;
2079 DAG.getNode(ISD::SRL, SDLoc(N), VT, SGN,
2080 DAG.getConstant(VT.getScalarSizeInBits() - lg2,
2081 getShiftAmountTy(SGN.getValueType())));
2082 SDValue ADD = DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, SRL);
2083 AddToWorklist(SRL.getNode());
2084 AddToWorklist(ADD.getNode()); // Divide by pow2
2085 SDValue SRA = DAG.getNode(ISD::SRA, SDLoc(N), VT, ADD,
2086 DAG.getConstant(lg2, getShiftAmountTy(ADD.getValueType())));
2088 // If we're dividing by a positive value, we're done. Otherwise, we must
2089 // negate the result.
2090 if (N1C->getAPIntValue().isNonNegative())
2093 AddToWorklist(SRA.getNode());
2094 return DAG.getNode(ISD::SUB, SDLoc(N), VT, DAG.getConstant(0, VT), SRA);
2097 // if integer divide is expensive and we satisfy the requirements, emit an
2098 // alternate sequence.
2099 if (N1C && !TLI.isIntDivCheap()) {
2100 SDValue Op = BuildSDIV(N);
2101 if (Op.getNode()) return Op;
2105 if (N0.getOpcode() == ISD::UNDEF)
2106 return DAG.getConstant(0, VT);
2107 // X / undef -> undef
2108 if (N1.getOpcode() == ISD::UNDEF)
2114 SDValue DAGCombiner::visitUDIV(SDNode *N) {
2115 SDValue N0 = N->getOperand(0);
2116 SDValue N1 = N->getOperand(1);
2117 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2118 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2119 EVT VT = N->getValueType(0);
2122 if (VT.isVector()) {
2123 SDValue FoldedVOp = SimplifyVBinOp(N);
2124 if (FoldedVOp.getNode()) return FoldedVOp;
2127 // fold (udiv c1, c2) -> c1/c2
2128 if (N0C && N1C && !N1C->isNullValue())
2129 return DAG.FoldConstantArithmetic(ISD::UDIV, VT, N0C, N1C);
2130 // fold (udiv x, (1 << c)) -> x >>u c
2131 if (N1C && N1C->getAPIntValue().isPowerOf2())
2132 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0,
2133 DAG.getConstant(N1C->getAPIntValue().logBase2(),
2134 getShiftAmountTy(N0.getValueType())));
2135 // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2
2136 if (N1.getOpcode() == ISD::SHL) {
2137 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
2138 if (SHC->getAPIntValue().isPowerOf2()) {
2139 EVT ADDVT = N1.getOperand(1).getValueType();
2140 SDValue Add = DAG.getNode(ISD::ADD, SDLoc(N), ADDVT,
2142 DAG.getConstant(SHC->getAPIntValue()
2145 AddToWorklist(Add.getNode());
2146 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, Add);
2150 // fold (udiv x, c) -> alternate
2151 if (N1C && !TLI.isIntDivCheap()) {
2152 SDValue Op = BuildUDIV(N);
2153 if (Op.getNode()) return Op;
2157 if (N0.getOpcode() == ISD::UNDEF)
2158 return DAG.getConstant(0, VT);
2159 // X / undef -> undef
2160 if (N1.getOpcode() == ISD::UNDEF)
2166 SDValue DAGCombiner::visitSREM(SDNode *N) {
2167 SDValue N0 = N->getOperand(0);
2168 SDValue N1 = N->getOperand(1);
2169 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2170 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2171 EVT VT = N->getValueType(0);
2173 // fold (srem c1, c2) -> c1%c2
2174 if (N0C && N1C && !N1C->isNullValue())
2175 return DAG.FoldConstantArithmetic(ISD::SREM, VT, N0C, N1C);
2176 // If we know the sign bits of both operands are zero, strength reduce to a
2177 // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15
2178 if (!VT.isVector()) {
2179 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
2180 return DAG.getNode(ISD::UREM, SDLoc(N), VT, N0, N1);
2183 // If X/C can be simplified by the division-by-constant logic, lower
2184 // X%C to the equivalent of X-X/C*C.
2185 if (N1C && !N1C->isNullValue()) {
2186 SDValue Div = DAG.getNode(ISD::SDIV, SDLoc(N), VT, N0, N1);
2187 AddToWorklist(Div.getNode());
2188 SDValue OptimizedDiv = combine(Div.getNode());
2189 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2190 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2192 SDValue Sub = DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, Mul);
2193 AddToWorklist(Mul.getNode());
2199 if (N0.getOpcode() == ISD::UNDEF)
2200 return DAG.getConstant(0, VT);
2201 // X % undef -> undef
2202 if (N1.getOpcode() == ISD::UNDEF)
2208 SDValue DAGCombiner::visitUREM(SDNode *N) {
2209 SDValue N0 = N->getOperand(0);
2210 SDValue N1 = N->getOperand(1);
2211 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2212 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2213 EVT VT = N->getValueType(0);
2215 // fold (urem c1, c2) -> c1%c2
2216 if (N0C && N1C && !N1C->isNullValue())
2217 return DAG.FoldConstantArithmetic(ISD::UREM, VT, N0C, N1C);
2218 // fold (urem x, pow2) -> (and x, pow2-1)
2219 if (N1C && !N1C->isNullValue() && N1C->getAPIntValue().isPowerOf2())
2220 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0,
2221 DAG.getConstant(N1C->getAPIntValue()-1,VT));
2222 // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1))
2223 if (N1.getOpcode() == ISD::SHL) {
2224 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
2225 if (SHC->getAPIntValue().isPowerOf2()) {
2227 DAG.getNode(ISD::ADD, SDLoc(N), VT, N1,
2228 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()),
2230 AddToWorklist(Add.getNode());
2231 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, Add);
2236 // If X/C can be simplified by the division-by-constant logic, lower
2237 // X%C to the equivalent of X-X/C*C.
2238 if (N1C && !N1C->isNullValue()) {
2239 SDValue Div = DAG.getNode(ISD::UDIV, SDLoc(N), VT, N0, N1);
2240 AddToWorklist(Div.getNode());
2241 SDValue OptimizedDiv = combine(Div.getNode());
2242 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2243 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2245 SDValue Sub = DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, Mul);
2246 AddToWorklist(Mul.getNode());
2252 if (N0.getOpcode() == ISD::UNDEF)
2253 return DAG.getConstant(0, VT);
2254 // X % undef -> undef
2255 if (N1.getOpcode() == ISD::UNDEF)
2261 SDValue DAGCombiner::visitMULHS(SDNode *N) {
2262 SDValue N0 = N->getOperand(0);
2263 SDValue N1 = N->getOperand(1);
2264 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2265 EVT VT = N->getValueType(0);
2268 // fold (mulhs x, 0) -> 0
2269 if (N1C && N1C->isNullValue())
2271 // fold (mulhs x, 1) -> (sra x, size(x)-1)
2272 if (N1C && N1C->getAPIntValue() == 1)
2273 return DAG.getNode(ISD::SRA, SDLoc(N), N0.getValueType(), N0,
2274 DAG.getConstant(N0.getValueType().getSizeInBits() - 1,
2275 getShiftAmountTy(N0.getValueType())));
2276 // fold (mulhs x, undef) -> 0
2277 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2278 return DAG.getConstant(0, VT);
2280 // If the type twice as wide is legal, transform the mulhs to a wider multiply
2282 if (VT.isSimple() && !VT.isVector()) {
2283 MVT Simple = VT.getSimpleVT();
2284 unsigned SimpleSize = Simple.getSizeInBits();
2285 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2286 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2287 N0 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N0);
2288 N1 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N1);
2289 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2290 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2291 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2292 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2299 SDValue DAGCombiner::visitMULHU(SDNode *N) {
2300 SDValue N0 = N->getOperand(0);
2301 SDValue N1 = N->getOperand(1);
2302 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2303 EVT VT = N->getValueType(0);
2306 // fold (mulhu x, 0) -> 0
2307 if (N1C && N1C->isNullValue())
2309 // fold (mulhu x, 1) -> 0
2310 if (N1C && N1C->getAPIntValue() == 1)
2311 return DAG.getConstant(0, N0.getValueType());
2312 // fold (mulhu x, undef) -> 0
2313 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2314 return DAG.getConstant(0, VT);
2316 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2318 if (VT.isSimple() && !VT.isVector()) {
2319 MVT Simple = VT.getSimpleVT();
2320 unsigned SimpleSize = Simple.getSizeInBits();
2321 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2322 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2323 N0 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N0);
2324 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N1);
2325 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2326 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2327 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2328 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2335 /// SimplifyNodeWithTwoResults - Perform optimizations common to nodes that
2336 /// compute two values. LoOp and HiOp give the opcodes for the two computations
2337 /// that are being performed. Return true if a simplification was made.
2339 SDValue DAGCombiner::SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
2341 // If the high half is not needed, just compute the low half.
2342 bool HiExists = N->hasAnyUseOfValue(1);
2344 (!LegalOperations ||
2345 TLI.isOperationLegalOrCustom(LoOp, N->getValueType(0)))) {
2346 SDValue Res = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0),
2347 ArrayRef<SDUse>(N->op_begin(), N->op_end()));
2348 return CombineTo(N, Res, Res);
2351 // If the low half is not needed, just compute the high half.
2352 bool LoExists = N->hasAnyUseOfValue(0);
2354 (!LegalOperations ||
2355 TLI.isOperationLegal(HiOp, N->getValueType(1)))) {
2356 SDValue Res = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1),
2357 ArrayRef<SDUse>(N->op_begin(), N->op_end()));
2358 return CombineTo(N, Res, Res);
2361 // If both halves are used, return as it is.
2362 if (LoExists && HiExists)
2365 // If the two computed results can be simplified separately, separate them.
2367 SDValue Lo = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0),
2368 ArrayRef<SDUse>(N->op_begin(), N->op_end()));
2369 AddToWorklist(Lo.getNode());
2370 SDValue LoOpt = combine(Lo.getNode());
2371 if (LoOpt.getNode() && LoOpt.getNode() != Lo.getNode() &&
2372 (!LegalOperations ||
2373 TLI.isOperationLegal(LoOpt.getOpcode(), LoOpt.getValueType())))
2374 return CombineTo(N, LoOpt, LoOpt);
2378 SDValue Hi = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1),
2379 ArrayRef<SDUse>(N->op_begin(), N->op_end()));
2380 AddToWorklist(Hi.getNode());
2381 SDValue HiOpt = combine(Hi.getNode());
2382 if (HiOpt.getNode() && HiOpt != Hi &&
2383 (!LegalOperations ||
2384 TLI.isOperationLegal(HiOpt.getOpcode(), HiOpt.getValueType())))
2385 return CombineTo(N, HiOpt, HiOpt);
2391 SDValue DAGCombiner::visitSMUL_LOHI(SDNode *N) {
2392 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHS);
2393 if (Res.getNode()) return Res;
2395 EVT VT = N->getValueType(0);
2398 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2400 if (VT.isSimple() && !VT.isVector()) {
2401 MVT Simple = VT.getSimpleVT();
2402 unsigned SimpleSize = Simple.getSizeInBits();
2403 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2404 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2405 SDValue Lo = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(0));
2406 SDValue Hi = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(1));
2407 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2408 // Compute the high part as N1.
2409 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2410 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2411 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2412 // Compute the low part as N0.
2413 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2414 return CombineTo(N, Lo, Hi);
2421 SDValue DAGCombiner::visitUMUL_LOHI(SDNode *N) {
2422 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHU);
2423 if (Res.getNode()) return Res;
2425 EVT VT = N->getValueType(0);
2428 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2430 if (VT.isSimple() && !VT.isVector()) {
2431 MVT Simple = VT.getSimpleVT();
2432 unsigned SimpleSize = Simple.getSizeInBits();
2433 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2434 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2435 SDValue Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(0));
2436 SDValue Hi = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(1));
2437 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2438 // Compute the high part as N1.
2439 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2440 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2441 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2442 // Compute the low part as N0.
2443 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2444 return CombineTo(N, Lo, Hi);
2451 SDValue DAGCombiner::visitSMULO(SDNode *N) {
2452 // (smulo x, 2) -> (saddo x, x)
2453 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2454 if (C2->getAPIntValue() == 2)
2455 return DAG.getNode(ISD::SADDO, SDLoc(N), N->getVTList(),
2456 N->getOperand(0), N->getOperand(0));
2461 SDValue DAGCombiner::visitUMULO(SDNode *N) {
2462 // (umulo x, 2) -> (uaddo x, x)
2463 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2464 if (C2->getAPIntValue() == 2)
2465 return DAG.getNode(ISD::UADDO, SDLoc(N), N->getVTList(),
2466 N->getOperand(0), N->getOperand(0));
2471 SDValue DAGCombiner::visitSDIVREM(SDNode *N) {
2472 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::SDIV, ISD::SREM);
2473 if (Res.getNode()) return Res;
2478 SDValue DAGCombiner::visitUDIVREM(SDNode *N) {
2479 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::UDIV, ISD::UREM);
2480 if (Res.getNode()) return Res;
2485 /// SimplifyBinOpWithSameOpcodeHands - If this is a binary operator with
2486 /// two operands of the same opcode, try to simplify it.
2487 SDValue DAGCombiner::SimplifyBinOpWithSameOpcodeHands(SDNode *N) {
2488 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
2489 EVT VT = N0.getValueType();
2490 assert(N0.getOpcode() == N1.getOpcode() && "Bad input!");
2492 // Bail early if none of these transforms apply.
2493 if (N0.getNode()->getNumOperands() == 0) return SDValue();
2495 // For each of OP in AND/OR/XOR:
2496 // fold (OP (zext x), (zext y)) -> (zext (OP x, y))
2497 // fold (OP (sext x), (sext y)) -> (sext (OP x, y))
2498 // fold (OP (aext x), (aext y)) -> (aext (OP x, y))
2499 // fold (OP (trunc x), (trunc y)) -> (trunc (OP x, y)) (if trunc isn't free)
2501 // do not sink logical op inside of a vector extend, since it may combine
2503 EVT Op0VT = N0.getOperand(0).getValueType();
2504 if ((N0.getOpcode() == ISD::ZERO_EXTEND ||
2505 N0.getOpcode() == ISD::SIGN_EXTEND ||
2506 // Avoid infinite looping with PromoteIntBinOp.
2507 (N0.getOpcode() == ISD::ANY_EXTEND &&
2508 (!LegalTypes || TLI.isTypeDesirableForOp(N->getOpcode(), Op0VT))) ||
2509 (N0.getOpcode() == ISD::TRUNCATE &&
2510 (!TLI.isZExtFree(VT, Op0VT) ||
2511 !TLI.isTruncateFree(Op0VT, VT)) &&
2512 TLI.isTypeLegal(Op0VT))) &&
2514 Op0VT == N1.getOperand(0).getValueType() &&
2515 (!LegalOperations || TLI.isOperationLegal(N->getOpcode(), Op0VT))) {
2516 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0),
2517 N0.getOperand(0).getValueType(),
2518 N0.getOperand(0), N1.getOperand(0));
2519 AddToWorklist(ORNode.getNode());
2520 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, ORNode);
2523 // For each of OP in SHL/SRL/SRA/AND...
2524 // fold (and (OP x, z), (OP y, z)) -> (OP (and x, y), z)
2525 // fold (or (OP x, z), (OP y, z)) -> (OP (or x, y), z)
2526 // fold (xor (OP x, z), (OP y, z)) -> (OP (xor x, y), z)
2527 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL ||
2528 N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::AND) &&
2529 N0.getOperand(1) == N1.getOperand(1)) {
2530 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0),
2531 N0.getOperand(0).getValueType(),
2532 N0.getOperand(0), N1.getOperand(0));
2533 AddToWorklist(ORNode.getNode());
2534 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
2535 ORNode, N0.getOperand(1));
2538 // Simplify xor/and/or (bitcast(A), bitcast(B)) -> bitcast(op (A,B))
2539 // Only perform this optimization after type legalization and before
2540 // LegalizeVectorOprs. LegalizeVectorOprs promotes vector operations by
2541 // adding bitcasts. For example (xor v4i32) is promoted to (v2i64), and
2542 // we don't want to undo this promotion.
2543 // We also handle SCALAR_TO_VECTOR because xor/or/and operations are cheaper
2545 if ((N0.getOpcode() == ISD::BITCAST ||
2546 N0.getOpcode() == ISD::SCALAR_TO_VECTOR) &&
2547 Level == AfterLegalizeTypes) {
2548 SDValue In0 = N0.getOperand(0);
2549 SDValue In1 = N1.getOperand(0);
2550 EVT In0Ty = In0.getValueType();
2551 EVT In1Ty = In1.getValueType();
2553 // If both incoming values are integers, and the original types are the
2555 if (In0Ty.isInteger() && In1Ty.isInteger() && In0Ty == In1Ty) {
2556 SDValue Op = DAG.getNode(N->getOpcode(), DL, In0Ty, In0, In1);
2557 SDValue BC = DAG.getNode(N0.getOpcode(), DL, VT, Op);
2558 AddToWorklist(Op.getNode());
2563 // Xor/and/or are indifferent to the swizzle operation (shuffle of one value).
2564 // Simplify xor/and/or (shuff(A), shuff(B)) -> shuff(op (A,B))
2565 // If both shuffles use the same mask, and both shuffle within a single
2566 // vector, then it is worthwhile to move the swizzle after the operation.
2567 // The type-legalizer generates this pattern when loading illegal
2568 // vector types from memory. In many cases this allows additional shuffle
2570 // There are other cases where moving the shuffle after the xor/and/or
2571 // is profitable even if shuffles don't perform a swizzle.
2572 // If both shuffles use the same mask, and both shuffles have the same first
2573 // or second operand, then it might still be profitable to move the shuffle
2574 // after the xor/and/or operation.
2575 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG) {
2576 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(N0);
2577 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(N1);
2579 assert(N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType() &&
2580 "Inputs to shuffles are not the same type");
2582 // Check that both shuffles use the same mask. The masks are known to be of
2583 // the same length because the result vector type is the same.
2584 // Check also that shuffles have only one use to avoid introducing extra
2586 if (SVN0->hasOneUse() && SVN1->hasOneUse() &&
2587 SVN0->getMask().equals(SVN1->getMask())) {
2588 SDValue ShOp = N0->getOperand(1);
2590 // Don't try to fold this node if it requires introducing a
2591 // build vector of all zeros that might be illegal at this stage.
2592 if (N->getOpcode() == ISD::XOR && ShOp.getOpcode() != ISD::UNDEF) {
2594 ShOp = DAG.getConstant(0, VT);
2599 // (AND (shuf (A, C), shuf (B, C)) -> shuf (AND (A, B), C)
2600 // (OR (shuf (A, C), shuf (B, C)) -> shuf (OR (A, B), C)
2601 // (XOR (shuf (A, C), shuf (B, C)) -> shuf (XOR (A, B), V_0)
2602 if (N0.getOperand(1) == N1.getOperand(1) && ShOp.getNode()) {
2603 SDValue NewNode = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
2604 N0->getOperand(0), N1->getOperand(0));
2605 AddToWorklist(NewNode.getNode());
2606 return DAG.getVectorShuffle(VT, SDLoc(N), NewNode, ShOp,
2607 &SVN0->getMask()[0]);
2610 // Don't try to fold this node if it requires introducing a
2611 // build vector of all zeros that might be illegal at this stage.
2612 ShOp = N0->getOperand(0);
2613 if (N->getOpcode() == ISD::XOR && ShOp.getOpcode() != ISD::UNDEF) {
2615 ShOp = DAG.getConstant(0, VT);
2620 // (AND (shuf (C, A), shuf (C, B)) -> shuf (C, AND (A, B))
2621 // (OR (shuf (C, A), shuf (C, B)) -> shuf (C, OR (A, B))
2622 // (XOR (shuf (C, A), shuf (C, B)) -> shuf (V_0, XOR (A, B))
2623 if (N0->getOperand(0) == N1->getOperand(0) && ShOp.getNode()) {
2624 SDValue NewNode = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
2625 N0->getOperand(1), N1->getOperand(1));
2626 AddToWorklist(NewNode.getNode());
2627 return DAG.getVectorShuffle(VT, SDLoc(N), ShOp, NewNode,
2628 &SVN0->getMask()[0]);
2636 SDValue DAGCombiner::visitAND(SDNode *N) {
2637 SDValue N0 = N->getOperand(0);
2638 SDValue N1 = N->getOperand(1);
2639 SDValue LL, LR, RL, RR, CC0, CC1;
2640 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2641 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2642 EVT VT = N1.getValueType();
2643 unsigned BitWidth = VT.getScalarType().getSizeInBits();
2646 if (VT.isVector()) {
2647 SDValue FoldedVOp = SimplifyVBinOp(N);
2648 if (FoldedVOp.getNode()) return FoldedVOp;
2650 // fold (and x, 0) -> 0, vector edition
2651 if (ISD::isBuildVectorAllZeros(N0.getNode()))
2653 if (ISD::isBuildVectorAllZeros(N1.getNode()))
2656 // fold (and x, -1) -> x, vector edition
2657 if (ISD::isBuildVectorAllOnes(N0.getNode()))
2659 if (ISD::isBuildVectorAllOnes(N1.getNode()))
2663 // fold (and x, undef) -> 0
2664 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2665 return DAG.getConstant(0, VT);
2666 // fold (and c1, c2) -> c1&c2
2668 return DAG.FoldConstantArithmetic(ISD::AND, VT, N0C, N1C);
2669 // canonicalize constant to RHS
2671 return DAG.getNode(ISD::AND, SDLoc(N), VT, N1, N0);
2672 // fold (and x, -1) -> x
2673 if (N1C && N1C->isAllOnesValue())
2675 // if (and x, c) is known to be zero, return 0
2676 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
2677 APInt::getAllOnesValue(BitWidth)))
2678 return DAG.getConstant(0, VT);
2680 SDValue RAND = ReassociateOps(ISD::AND, SDLoc(N), N0, N1);
2683 // fold (and (or x, C), D) -> D if (C & D) == D
2684 if (N1C && N0.getOpcode() == ISD::OR)
2685 if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
2686 if ((ORI->getAPIntValue() & N1C->getAPIntValue()) == N1C->getAPIntValue())
2688 // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits.
2689 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
2690 SDValue N0Op0 = N0.getOperand(0);
2691 APInt Mask = ~N1C->getAPIntValue();
2692 Mask = Mask.trunc(N0Op0.getValueSizeInBits());
2693 if (DAG.MaskedValueIsZero(N0Op0, Mask)) {
2694 SDValue Zext = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N),
2695 N0.getValueType(), N0Op0);
2697 // Replace uses of the AND with uses of the Zero extend node.
2700 // We actually want to replace all uses of the any_extend with the
2701 // zero_extend, to avoid duplicating things. This will later cause this
2702 // AND to be folded.
2703 CombineTo(N0.getNode(), Zext);
2704 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2707 // similarly fold (and (X (load ([non_ext|any_ext|zero_ext] V))), c) ->
2708 // (X (load ([non_ext|zero_ext] V))) if 'and' only clears top bits which must
2709 // already be zero by virtue of the width of the base type of the load.
2711 // the 'X' node here can either be nothing or an extract_vector_elt to catch
2713 if ((N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
2714 N0.getOperand(0).getOpcode() == ISD::LOAD) ||
2715 N0.getOpcode() == ISD::LOAD) {
2716 LoadSDNode *Load = cast<LoadSDNode>( (N0.getOpcode() == ISD::LOAD) ?
2717 N0 : N0.getOperand(0) );
2719 // Get the constant (if applicable) the zero'th operand is being ANDed with.
2720 // This can be a pure constant or a vector splat, in which case we treat the
2721 // vector as a scalar and use the splat value.
2722 APInt Constant = APInt::getNullValue(1);
2723 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
2724 Constant = C->getAPIntValue();
2725 } else if (BuildVectorSDNode *Vector = dyn_cast<BuildVectorSDNode>(N1)) {
2726 APInt SplatValue, SplatUndef;
2727 unsigned SplatBitSize;
2729 bool IsSplat = Vector->isConstantSplat(SplatValue, SplatUndef,
2730 SplatBitSize, HasAnyUndefs);
2732 // Undef bits can contribute to a possible optimisation if set, so
2734 SplatValue |= SplatUndef;
2736 // The splat value may be something like "0x00FFFFFF", which means 0 for
2737 // the first vector value and FF for the rest, repeating. We need a mask
2738 // that will apply equally to all members of the vector, so AND all the
2739 // lanes of the constant together.
2740 EVT VT = Vector->getValueType(0);
2741 unsigned BitWidth = VT.getVectorElementType().getSizeInBits();
2743 // If the splat value has been compressed to a bitlength lower
2744 // than the size of the vector lane, we need to re-expand it to
2746 if (BitWidth > SplatBitSize)
2747 for (SplatValue = SplatValue.zextOrTrunc(BitWidth);
2748 SplatBitSize < BitWidth;
2749 SplatBitSize = SplatBitSize * 2)
2750 SplatValue |= SplatValue.shl(SplatBitSize);
2752 Constant = APInt::getAllOnesValue(BitWidth);
2753 for (unsigned i = 0, n = SplatBitSize/BitWidth; i < n; ++i)
2754 Constant &= SplatValue.lshr(i*BitWidth).zextOrTrunc(BitWidth);
2758 // If we want to change an EXTLOAD to a ZEXTLOAD, ensure a ZEXTLOAD is
2759 // actually legal and isn't going to get expanded, else this is a false
2761 bool CanZextLoadProfitably = TLI.isLoadExtLegal(ISD::ZEXTLOAD,
2762 Load->getMemoryVT());
2764 // Resize the constant to the same size as the original memory access before
2765 // extension. If it is still the AllOnesValue then this AND is completely
2768 Constant.zextOrTrunc(Load->getMemoryVT().getScalarType().getSizeInBits());
2771 switch (Load->getExtensionType()) {
2772 default: B = false; break;
2773 case ISD::EXTLOAD: B = CanZextLoadProfitably; break;
2775 case ISD::NON_EXTLOAD: B = true; break;
2778 if (B && Constant.isAllOnesValue()) {
2779 // If the load type was an EXTLOAD, convert to ZEXTLOAD in order to
2780 // preserve semantics once we get rid of the AND.
2781 SDValue NewLoad(Load, 0);
2782 if (Load->getExtensionType() == ISD::EXTLOAD) {
2783 NewLoad = DAG.getLoad(Load->getAddressingMode(), ISD::ZEXTLOAD,
2784 Load->getValueType(0), SDLoc(Load),
2785 Load->getChain(), Load->getBasePtr(),
2786 Load->getOffset(), Load->getMemoryVT(),
2787 Load->getMemOperand());
2788 // Replace uses of the EXTLOAD with the new ZEXTLOAD.
2789 if (Load->getNumValues() == 3) {
2790 // PRE/POST_INC loads have 3 values.
2791 SDValue To[] = { NewLoad.getValue(0), NewLoad.getValue(1),
2792 NewLoad.getValue(2) };
2793 CombineTo(Load, To, 3, true);
2795 CombineTo(Load, NewLoad.getValue(0), NewLoad.getValue(1));
2799 // Fold the AND away, taking care not to fold to the old load node if we
2801 CombineTo(N, (N0.getNode() == Load) ? NewLoad : N0);
2803 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2806 // fold (and (setcc x), (setcc y)) -> (setcc (and x, y))
2807 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
2808 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
2809 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
2811 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
2812 LL.getValueType().isInteger()) {
2813 // fold (and (seteq X, 0), (seteq Y, 0)) -> (seteq (or X, Y), 0)
2814 if (cast<ConstantSDNode>(LR)->isNullValue() && Op1 == ISD::SETEQ) {
2815 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(N0),
2816 LR.getValueType(), LL, RL);
2817 AddToWorklist(ORNode.getNode());
2818 return DAG.getSetCC(SDLoc(N), VT, ORNode, LR, Op1);
2820 // fold (and (seteq X, -1), (seteq Y, -1)) -> (seteq (and X, Y), -1)
2821 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETEQ) {
2822 SDValue ANDNode = DAG.getNode(ISD::AND, SDLoc(N0),
2823 LR.getValueType(), LL, RL);
2824 AddToWorklist(ANDNode.getNode());
2825 return DAG.getSetCC(SDLoc(N), VT, ANDNode, LR, Op1);
2827 // fold (and (setgt X, -1), (setgt Y, -1)) -> (setgt (or X, Y), -1)
2828 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETGT) {
2829 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(N0),
2830 LR.getValueType(), LL, RL);
2831 AddToWorklist(ORNode.getNode());
2832 return DAG.getSetCC(SDLoc(N), VT, ORNode, LR, Op1);
2835 // Simplify (and (setne X, 0), (setne X, -1)) -> (setuge (add X, 1), 2)
2836 if (LL == RL && isa<ConstantSDNode>(LR) && isa<ConstantSDNode>(RR) &&
2837 Op0 == Op1 && LL.getValueType().isInteger() &&
2838 Op0 == ISD::SETNE && ((cast<ConstantSDNode>(LR)->isNullValue() &&
2839 cast<ConstantSDNode>(RR)->isAllOnesValue()) ||
2840 (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
2841 cast<ConstantSDNode>(RR)->isNullValue()))) {
2842 SDValue ADDNode = DAG.getNode(ISD::ADD, SDLoc(N0), LL.getValueType(),
2843 LL, DAG.getConstant(1, LL.getValueType()));
2844 AddToWorklist(ADDNode.getNode());
2845 return DAG.getSetCC(SDLoc(N), VT, ADDNode,
2846 DAG.getConstant(2, LL.getValueType()), ISD::SETUGE);
2848 // canonicalize equivalent to ll == rl
2849 if (LL == RR && LR == RL) {
2850 Op1 = ISD::getSetCCSwappedOperands(Op1);
2853 if (LL == RL && LR == RR) {
2854 bool isInteger = LL.getValueType().isInteger();
2855 ISD::CondCode Result = ISD::getSetCCAndOperation(Op0, Op1, isInteger);
2856 if (Result != ISD::SETCC_INVALID &&
2857 (!LegalOperations ||
2858 (TLI.isCondCodeLegal(Result, LL.getSimpleValueType()) &&
2859 TLI.isOperationLegal(ISD::SETCC,
2860 getSetCCResultType(N0.getSimpleValueType())))))
2861 return DAG.getSetCC(SDLoc(N), N0.getValueType(),
2866 // Simplify: (and (op x...), (op y...)) -> (op (and x, y))
2867 if (N0.getOpcode() == N1.getOpcode()) {
2868 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
2869 if (Tmp.getNode()) return Tmp;
2872 // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1)
2873 // fold (and (sra)) -> (and (srl)) when possible.
2874 if (!VT.isVector() &&
2875 SimplifyDemandedBits(SDValue(N, 0)))
2876 return SDValue(N, 0);
2878 // fold (zext_inreg (extload x)) -> (zextload x)
2879 if (ISD::isEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode())) {
2880 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2881 EVT MemVT = LN0->getMemoryVT();
2882 // If we zero all the possible extended bits, then we can turn this into
2883 // a zextload if we are running before legalize or the operation is legal.
2884 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
2885 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
2886 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
2887 ((!LegalOperations && !LN0->isVolatile()) ||
2888 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
2889 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT,
2890 LN0->getChain(), LN0->getBasePtr(),
2891 MemVT, LN0->getMemOperand());
2893 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
2894 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2897 // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use
2898 if (ISD::isSEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
2900 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2901 EVT MemVT = LN0->getMemoryVT();
2902 // If we zero all the possible extended bits, then we can turn this into
2903 // a zextload if we are running before legalize or the operation is legal.
2904 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
2905 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
2906 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
2907 ((!LegalOperations && !LN0->isVolatile()) ||
2908 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
2909 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT,
2910 LN0->getChain(), LN0->getBasePtr(),
2911 MemVT, LN0->getMemOperand());
2913 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
2914 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2918 // fold (and (load x), 255) -> (zextload x, i8)
2919 // fold (and (extload x, i16), 255) -> (zextload x, i8)
2920 // fold (and (any_ext (extload x, i16)), 255) -> (zextload x, i8)
2921 if (N1C && (N0.getOpcode() == ISD::LOAD ||
2922 (N0.getOpcode() == ISD::ANY_EXTEND &&
2923 N0.getOperand(0).getOpcode() == ISD::LOAD))) {
2924 bool HasAnyExt = N0.getOpcode() == ISD::ANY_EXTEND;
2925 LoadSDNode *LN0 = HasAnyExt
2926 ? cast<LoadSDNode>(N0.getOperand(0))
2927 : cast<LoadSDNode>(N0);
2928 if (LN0->getExtensionType() != ISD::SEXTLOAD &&
2929 LN0->isUnindexed() && N0.hasOneUse() && SDValue(LN0, 0).hasOneUse()) {
2930 uint32_t ActiveBits = N1C->getAPIntValue().getActiveBits();
2931 if (ActiveBits > 0 && APIntOps::isMask(ActiveBits, N1C->getAPIntValue())){
2932 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits);
2933 EVT LoadedVT = LN0->getMemoryVT();
2935 if (ExtVT == LoadedVT &&
2936 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) {
2937 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
2940 DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy,
2941 LN0->getChain(), LN0->getBasePtr(), ExtVT,
2942 LN0->getMemOperand());
2944 CombineTo(LN0, NewLoad, NewLoad.getValue(1));
2945 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2948 // Do not change the width of a volatile load.
2949 // Do not generate loads of non-round integer types since these can
2950 // be expensive (and would be wrong if the type is not byte sized).
2951 if (!LN0->isVolatile() && LoadedVT.bitsGT(ExtVT) && ExtVT.isRound() &&
2952 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) {
2953 EVT PtrType = LN0->getOperand(1).getValueType();
2955 unsigned Alignment = LN0->getAlignment();
2956 SDValue NewPtr = LN0->getBasePtr();
2958 // For big endian targets, we need to add an offset to the pointer
2959 // to load the correct bytes. For little endian systems, we merely
2960 // need to read fewer bytes from the same pointer.
2961 if (TLI.isBigEndian()) {
2962 unsigned LVTStoreBytes = LoadedVT.getStoreSize();
2963 unsigned EVTStoreBytes = ExtVT.getStoreSize();
2964 unsigned PtrOff = LVTStoreBytes - EVTStoreBytes;
2965 NewPtr = DAG.getNode(ISD::ADD, SDLoc(LN0), PtrType,
2966 NewPtr, DAG.getConstant(PtrOff, PtrType));
2967 Alignment = MinAlign(Alignment, PtrOff);
2970 AddToWorklist(NewPtr.getNode());
2972 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
2974 DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy,
2975 LN0->getChain(), NewPtr,
2976 LN0->getPointerInfo(),
2977 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
2978 Alignment, LN0->getAAInfo());
2980 CombineTo(LN0, Load, Load.getValue(1));
2981 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2987 if (N0.getOpcode() == ISD::ADD && N1.getOpcode() == ISD::SRL &&
2988 VT.getSizeInBits() <= 64) {
2989 if (ConstantSDNode *ADDI = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
2990 APInt ADDC = ADDI->getAPIntValue();
2991 if (!TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
2992 // Look for (and (add x, c1), (lshr y, c2)). If C1 wasn't a legal
2993 // immediate for an add, but it is legal if its top c2 bits are set,
2994 // transform the ADD so the immediate doesn't need to be materialized
2996 if (ConstantSDNode *SRLI = dyn_cast<ConstantSDNode>(N1.getOperand(1))) {
2997 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(),
2998 SRLI->getZExtValue());
2999 if (DAG.MaskedValueIsZero(N0.getOperand(1), Mask)) {
3001 if (TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
3003 DAG.getNode(ISD::ADD, SDLoc(N0), VT,
3004 N0.getOperand(0), DAG.getConstant(ADDC, VT));
3005 CombineTo(N0.getNode(), NewAdd);
3006 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3014 // fold (and (or (srl N, 8), (shl N, 8)), 0xffff) -> (srl (bswap N), const)
3015 if (N1C && N1C->getAPIntValue() == 0xffff && N0.getOpcode() == ISD::OR) {
3016 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
3017 N0.getOperand(1), false);
3018 if (BSwap.getNode())
3025 /// MatchBSwapHWord - Match (a >> 8) | (a << 8) as (bswap a) >> 16
3027 SDValue DAGCombiner::MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
3028 bool DemandHighBits) {
3029 if (!LegalOperations)
3032 EVT VT = N->getValueType(0);
3033 if (VT != MVT::i64 && VT != MVT::i32 && VT != MVT::i16)
3035 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
3038 // Recognize (and (shl a, 8), 0xff), (and (srl a, 8), 0xff00)
3039 bool LookPassAnd0 = false;
3040 bool LookPassAnd1 = false;
3041 if (N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode() == ISD::SRL)
3043 if (N1.getOpcode() == ISD::AND && N1.getOperand(0).getOpcode() == ISD::SHL)
3045 if (N0.getOpcode() == ISD::AND) {
3046 if (!N0.getNode()->hasOneUse())
3048 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3049 if (!N01C || N01C->getZExtValue() != 0xFF00)
3051 N0 = N0.getOperand(0);
3052 LookPassAnd0 = true;
3055 if (N1.getOpcode() == ISD::AND) {
3056 if (!N1.getNode()->hasOneUse())
3058 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
3059 if (!N11C || N11C->getZExtValue() != 0xFF)
3061 N1 = N1.getOperand(0);
3062 LookPassAnd1 = true;
3065 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
3067 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
3069 if (!N0.getNode()->hasOneUse() ||
3070 !N1.getNode()->hasOneUse())
3073 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3074 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
3077 if (N01C->getZExtValue() != 8 || N11C->getZExtValue() != 8)
3080 // Look for (shl (and a, 0xff), 8), (srl (and a, 0xff00), 8)
3081 SDValue N00 = N0->getOperand(0);
3082 if (!LookPassAnd0 && N00.getOpcode() == ISD::AND) {
3083 if (!N00.getNode()->hasOneUse())
3085 ConstantSDNode *N001C = dyn_cast<ConstantSDNode>(N00.getOperand(1));
3086 if (!N001C || N001C->getZExtValue() != 0xFF)
3088 N00 = N00.getOperand(0);
3089 LookPassAnd0 = true;
3092 SDValue N10 = N1->getOperand(0);
3093 if (!LookPassAnd1 && N10.getOpcode() == ISD::AND) {
3094 if (!N10.getNode()->hasOneUse())
3096 ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N10.getOperand(1));
3097 if (!N101C || N101C->getZExtValue() != 0xFF00)
3099 N10 = N10.getOperand(0);
3100 LookPassAnd1 = true;
3106 // Make sure everything beyond the low halfword gets set to zero since the SRL
3107 // 16 will clear the top bits.
3108 unsigned OpSizeInBits = VT.getSizeInBits();
3109 if (DemandHighBits && OpSizeInBits > 16) {
3110 // If the left-shift isn't masked out then the only way this is a bswap is
3111 // if all bits beyond the low 8 are 0. In that case the entire pattern
3112 // reduces to a left shift anyway: leave it for other parts of the combiner.
3116 // However, if the right shift isn't masked out then it might be because
3117 // it's not needed. See if we can spot that too.
3118 if (!LookPassAnd1 &&
3119 !DAG.MaskedValueIsZero(
3120 N10, APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - 16)))
3124 SDValue Res = DAG.getNode(ISD::BSWAP, SDLoc(N), VT, N00);
3125 if (OpSizeInBits > 16)
3126 Res = DAG.getNode(ISD::SRL, SDLoc(N), VT, Res,
3127 DAG.getConstant(OpSizeInBits-16, getShiftAmountTy(VT)));
3131 /// isBSwapHWordElement - Return true if the specified node is an element
3132 /// that makes up a 32-bit packed halfword byteswap. i.e.
3133 /// ((x&0xff)<<8)|((x&0xff00)>>8)|((x&0x00ff0000)<<8)|((x&0xff000000)>>8)
3134 static bool isBSwapHWordElement(SDValue N, SmallVectorImpl<SDNode *> &Parts) {
3135 if (!N.getNode()->hasOneUse())
3138 unsigned Opc = N.getOpcode();
3139 if (Opc != ISD::AND && Opc != ISD::SHL && Opc != ISD::SRL)
3142 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3147 switch (N1C->getZExtValue()) {
3150 case 0xFF: Num = 0; break;
3151 case 0xFF00: Num = 1; break;
3152 case 0xFF0000: Num = 2; break;
3153 case 0xFF000000: Num = 3; break;
3156 // Look for (x & 0xff) << 8 as well as ((x << 8) & 0xff00).
3157 SDValue N0 = N.getOperand(0);
3158 if (Opc == ISD::AND) {
3159 if (Num == 0 || Num == 2) {
3161 // (x >> 8) & 0xff0000
3162 if (N0.getOpcode() != ISD::SRL)
3164 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3165 if (!C || C->getZExtValue() != 8)
3168 // (x << 8) & 0xff00
3169 // (x << 8) & 0xff000000
3170 if (N0.getOpcode() != ISD::SHL)
3172 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3173 if (!C || C->getZExtValue() != 8)
3176 } else if (Opc == ISD::SHL) {
3178 // (x & 0xff0000) << 8
3179 if (Num != 0 && Num != 2)
3181 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3182 if (!C || C->getZExtValue() != 8)
3184 } else { // Opc == ISD::SRL
3185 // (x & 0xff00) >> 8
3186 // (x & 0xff000000) >> 8
3187 if (Num != 1 && Num != 3)
3189 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3190 if (!C || C->getZExtValue() != 8)
3197 Parts[Num] = N0.getOperand(0).getNode();
3201 /// MatchBSwapHWord - Match a 32-bit packed halfword bswap. That is
3202 /// ((x&0xff)<<8)|((x&0xff00)>>8)|((x&0x00ff0000)<<8)|((x&0xff000000)>>8)
3203 /// => (rotl (bswap x), 16)
3204 SDValue DAGCombiner::MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1) {
3205 if (!LegalOperations)
3208 EVT VT = N->getValueType(0);
3211 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
3214 SmallVector<SDNode*,4> Parts(4, (SDNode*)nullptr);
3216 // (or (or (and), (and)), (or (and), (and)))
3217 // (or (or (or (and), (and)), (and)), (and))
3218 if (N0.getOpcode() != ISD::OR)
3220 SDValue N00 = N0.getOperand(0);
3221 SDValue N01 = N0.getOperand(1);
3223 if (N1.getOpcode() == ISD::OR &&
3224 N00.getNumOperands() == 2 && N01.getNumOperands() == 2) {
3225 // (or (or (and), (and)), (or (and), (and)))
3226 SDValue N000 = N00.getOperand(0);
3227 if (!isBSwapHWordElement(N000, Parts))
3230 SDValue N001 = N00.getOperand(1);
3231 if (!isBSwapHWordElement(N001, Parts))
3233 SDValue N010 = N01.getOperand(0);
3234 if (!isBSwapHWordElement(N010, Parts))
3236 SDValue N011 = N01.getOperand(1);
3237 if (!isBSwapHWordElement(N011, Parts))
3240 // (or (or (or (and), (and)), (and)), (and))
3241 if (!isBSwapHWordElement(N1, Parts))
3243 if (!isBSwapHWordElement(N01, Parts))
3245 if (N00.getOpcode() != ISD::OR)
3247 SDValue N000 = N00.getOperand(0);
3248 if (!isBSwapHWordElement(N000, Parts))
3250 SDValue N001 = N00.getOperand(1);
3251 if (!isBSwapHWordElement(N001, Parts))
3255 // Make sure the parts are all coming from the same node.
3256 if (Parts[0] != Parts[1] || Parts[0] != Parts[2] || Parts[0] != Parts[3])
3259 SDValue BSwap = DAG.getNode(ISD::BSWAP, SDLoc(N), VT,
3260 SDValue(Parts[0],0));
3262 // Result of the bswap should be rotated by 16. If it's not legal, then
3263 // do (x << 16) | (x >> 16).
3264 SDValue ShAmt = DAG.getConstant(16, getShiftAmountTy(VT));
3265 if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT))
3266 return DAG.getNode(ISD::ROTL, SDLoc(N), VT, BSwap, ShAmt);
3267 if (TLI.isOperationLegalOrCustom(ISD::ROTR, VT))
3268 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, BSwap, ShAmt);
3269 return DAG.getNode(ISD::OR, SDLoc(N), VT,
3270 DAG.getNode(ISD::SHL, SDLoc(N), VT, BSwap, ShAmt),
3271 DAG.getNode(ISD::SRL, SDLoc(N), VT, BSwap, ShAmt));
3274 SDValue DAGCombiner::visitOR(SDNode *N) {
3275 SDValue N0 = N->getOperand(0);
3276 SDValue N1 = N->getOperand(1);
3277 SDValue LL, LR, RL, RR, CC0, CC1;
3278 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3279 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3280 EVT VT = N1.getValueType();
3283 if (VT.isVector()) {
3284 SDValue FoldedVOp = SimplifyVBinOp(N);
3285 if (FoldedVOp.getNode()) return FoldedVOp;
3287 // fold (or x, 0) -> x, vector edition
3288 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3290 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3293 // fold (or x, -1) -> -1, vector edition
3294 if (ISD::isBuildVectorAllOnes(N0.getNode()))
3296 if (ISD::isBuildVectorAllOnes(N1.getNode()))
3299 // fold (or (shuf A, V_0, MA), (shuf B, V_0, MB)) -> (shuf A, B, Mask1)
3300 // fold (or (shuf A, V_0, MA), (shuf B, V_0, MB)) -> (shuf B, A, Mask2)
3301 // Do this only if the resulting shuffle is legal.
3302 if (isa<ShuffleVectorSDNode>(N0) &&
3303 isa<ShuffleVectorSDNode>(N1) &&
3304 // Avoid folding a node with illegal type.
3305 TLI.isTypeLegal(VT) &&
3306 N0->getOperand(1) == N1->getOperand(1) &&
3307 ISD::isBuildVectorAllZeros(N0.getOperand(1).getNode())) {
3308 bool CanFold = true;
3309 unsigned NumElts = VT.getVectorNumElements();
3310 const ShuffleVectorSDNode *SV0 = cast<ShuffleVectorSDNode>(N0);
3311 const ShuffleVectorSDNode *SV1 = cast<ShuffleVectorSDNode>(N1);
3312 // We construct two shuffle masks:
3313 // - Mask1 is a shuffle mask for a shuffle with N0 as the first operand
3314 // and N1 as the second operand.
3315 // - Mask2 is a shuffle mask for a shuffle with N1 as the first operand
3316 // and N0 as the second operand.
3317 // We do this because OR is commutable and therefore there might be
3318 // two ways to fold this node into a shuffle.
3319 SmallVector<int,4> Mask1;
3320 SmallVector<int,4> Mask2;
3322 for (unsigned i = 0; i != NumElts && CanFold; ++i) {
3323 int M0 = SV0->getMaskElt(i);
3324 int M1 = SV1->getMaskElt(i);
3326 // Both shuffle indexes are undef. Propagate Undef.
3327 if (M0 < 0 && M1 < 0) {
3328 Mask1.push_back(M0);
3329 Mask2.push_back(M0);
3333 if (M0 < 0 || M1 < 0 ||
3334 (M0 < (int)NumElts && M1 < (int)NumElts) ||
3335 (M0 >= (int)NumElts && M1 >= (int)NumElts)) {
3340 Mask1.push_back(M0 < (int)NumElts ? M0 : M1 + NumElts);
3341 Mask2.push_back(M1 < (int)NumElts ? M1 : M0 + NumElts);
3345 // Fold this sequence only if the resulting shuffle is 'legal'.
3346 if (TLI.isShuffleMaskLegal(Mask1, VT))
3347 return DAG.getVectorShuffle(VT, SDLoc(N), N0->getOperand(0),
3348 N1->getOperand(0), &Mask1[0]);
3349 if (TLI.isShuffleMaskLegal(Mask2, VT))
3350 return DAG.getVectorShuffle(VT, SDLoc(N), N1->getOperand(0),
3351 N0->getOperand(0), &Mask2[0]);
3356 // fold (or x, undef) -> -1
3357 if (!LegalOperations &&
3358 (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)) {
3359 EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT;
3360 return DAG.getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()), VT);
3362 // fold (or c1, c2) -> c1|c2
3364 return DAG.FoldConstantArithmetic(ISD::OR, VT, N0C, N1C);
3365 // canonicalize constant to RHS
3367 return DAG.getNode(ISD::OR, SDLoc(N), VT, N1, N0);
3368 // fold (or x, 0) -> x
3369 if (N1C && N1C->isNullValue())
3371 // fold (or x, -1) -> -1
3372 if (N1C && N1C->isAllOnesValue())
3374 // fold (or x, c) -> c iff (x & ~c) == 0
3375 if (N1C && DAG.MaskedValueIsZero(N0, ~N1C->getAPIntValue()))
3378 // Recognize halfword bswaps as (bswap + rotl 16) or (bswap + shl 16)
3379 SDValue BSwap = MatchBSwapHWord(N, N0, N1);
3380 if (BSwap.getNode())
3382 BSwap = MatchBSwapHWordLow(N, N0, N1);
3383 if (BSwap.getNode())
3387 SDValue ROR = ReassociateOps(ISD::OR, SDLoc(N), N0, N1);
3390 // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2)
3391 // iff (c1 & c2) == 0.
3392 if (N1C && N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
3393 isa<ConstantSDNode>(N0.getOperand(1))) {
3394 ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1));
3395 if ((C1->getAPIntValue() & N1C->getAPIntValue()) != 0) {
3396 SDValue COR = DAG.FoldConstantArithmetic(ISD::OR, VT, N1C, C1);
3399 return DAG.getNode(ISD::AND, SDLoc(N), VT,
3400 DAG.getNode(ISD::OR, SDLoc(N0), VT,
3401 N0.getOperand(0), N1), COR);
3404 // fold (or (setcc x), (setcc y)) -> (setcc (or x, y))
3405 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
3406 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
3407 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
3409 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
3410 LL.getValueType().isInteger()) {
3411 // fold (or (setne X, 0), (setne Y, 0)) -> (setne (or X, Y), 0)
3412 // fold (or (setlt X, 0), (setlt Y, 0)) -> (setne (or X, Y), 0)
3413 if (cast<ConstantSDNode>(LR)->isNullValue() &&
3414 (Op1 == ISD::SETNE || Op1 == ISD::SETLT)) {
3415 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(LR),
3416 LR.getValueType(), LL, RL);
3417 AddToWorklist(ORNode.getNode());
3418 return DAG.getSetCC(SDLoc(N), VT, ORNode, LR, Op1);
3420 // fold (or (setne X, -1), (setne Y, -1)) -> (setne (and X, Y), -1)
3421 // fold (or (setgt X, -1), (setgt Y -1)) -> (setgt (and X, Y), -1)
3422 if (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
3423 (Op1 == ISD::SETNE || Op1 == ISD::SETGT)) {
3424 SDValue ANDNode = DAG.getNode(ISD::AND, SDLoc(LR),
3425 LR.getValueType(), LL, RL);
3426 AddToWorklist(ANDNode.getNode());
3427 return DAG.getSetCC(SDLoc(N), VT, ANDNode, LR, Op1);
3430 // canonicalize equivalent to ll == rl
3431 if (LL == RR && LR == RL) {
3432 Op1 = ISD::getSetCCSwappedOperands(Op1);
3435 if (LL == RL && LR == RR) {
3436 bool isInteger = LL.getValueType().isInteger();
3437 ISD::CondCode Result = ISD::getSetCCOrOperation(Op0, Op1, isInteger);
3438 if (Result != ISD::SETCC_INVALID &&
3439 (!LegalOperations ||
3440 (TLI.isCondCodeLegal(Result, LL.getSimpleValueType()) &&
3441 TLI.isOperationLegal(ISD::SETCC,
3442 getSetCCResultType(N0.getValueType())))))
3443 return DAG.getSetCC(SDLoc(N), N0.getValueType(),
3448 // Simplify: (or (op x...), (op y...)) -> (op (or x, y))
3449 if (N0.getOpcode() == N1.getOpcode()) {
3450 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3451 if (Tmp.getNode()) return Tmp;
3454 // (or (and X, C1), (and Y, C2)) -> (and (or X, Y), C3) if possible.
3455 if (N0.getOpcode() == ISD::AND &&
3456 N1.getOpcode() == ISD::AND &&
3457 N0.getOperand(1).getOpcode() == ISD::Constant &&
3458 N1.getOperand(1).getOpcode() == ISD::Constant &&
3459 // Don't increase # computations.
3460 (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) {
3461 // We can only do this xform if we know that bits from X that are set in C2
3462 // but not in C1 are already zero. Likewise for Y.
3463 const APInt &LHSMask =
3464 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
3465 const APInt &RHSMask =
3466 cast<ConstantSDNode>(N1.getOperand(1))->getAPIntValue();
3468 if (DAG.MaskedValueIsZero(N0.getOperand(0), RHSMask&~LHSMask) &&
3469 DAG.MaskedValueIsZero(N1.getOperand(0), LHSMask&~RHSMask)) {
3470 SDValue X = DAG.getNode(ISD::OR, SDLoc(N0), VT,
3471 N0.getOperand(0), N1.getOperand(0));
3472 return DAG.getNode(ISD::AND, SDLoc(N), VT, X,
3473 DAG.getConstant(LHSMask | RHSMask, VT));
3477 // See if this is some rotate idiom.
3478 if (SDNode *Rot = MatchRotate(N0, N1, SDLoc(N)))
3479 return SDValue(Rot, 0);
3481 // Simplify the operands using demanded-bits information.
3482 if (!VT.isVector() &&
3483 SimplifyDemandedBits(SDValue(N, 0)))
3484 return SDValue(N, 0);
3489 /// MatchRotateHalf - Match "(X shl/srl V1) & V2" where V2 may not be present.
3490 static bool MatchRotateHalf(SDValue Op, SDValue &Shift, SDValue &Mask) {
3491 if (Op.getOpcode() == ISD::AND) {
3492 if (isa<ConstantSDNode>(Op.getOperand(1))) {
3493 Mask = Op.getOperand(1);
3494 Op = Op.getOperand(0);
3500 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SHL) {
3508 // Return true if we can prove that, whenever Neg and Pos are both in the
3509 // range [0, OpSize), Neg == (Pos == 0 ? 0 : OpSize - Pos). This means that
3510 // for two opposing shifts shift1 and shift2 and a value X with OpBits bits:
3512 // (or (shift1 X, Neg), (shift2 X, Pos))
3514 // reduces to a rotate in direction shift2 by Pos or (equivalently) a rotate
3515 // in direction shift1 by Neg. The range [0, OpSize) means that we only need
3516 // to consider shift amounts with defined behavior.
3517 static bool matchRotateSub(SDValue Pos, SDValue Neg, unsigned OpSize) {
3518 // If OpSize is a power of 2 then:
3520 // (a) (Pos == 0 ? 0 : OpSize - Pos) == (OpSize - Pos) & (OpSize - 1)
3521 // (b) Neg == Neg & (OpSize - 1) whenever Neg is in [0, OpSize).
3523 // So if OpSize is a power of 2 and Neg is (and Neg', OpSize-1), we check
3524 // for the stronger condition:
3526 // Neg & (OpSize - 1) == (OpSize - Pos) & (OpSize - 1) [A]
3528 // for all Neg and Pos. Since Neg & (OpSize - 1) == Neg' & (OpSize - 1)
3529 // we can just replace Neg with Neg' for the rest of the function.
3531 // In other cases we check for the even stronger condition:
3533 // Neg == OpSize - Pos [B]
3535 // for all Neg and Pos. Note that the (or ...) then invokes undefined
3536 // behavior if Pos == 0 (and consequently Neg == OpSize).
3538 // We could actually use [A] whenever OpSize is a power of 2, but the
3539 // only extra cases that it would match are those uninteresting ones
3540 // where Neg and Pos are never in range at the same time. E.g. for
3541 // OpSize == 32, using [A] would allow a Neg of the form (sub 64, Pos)
3542 // as well as (sub 32, Pos), but:
3544 // (or (shift1 X, (sub 64, Pos)), (shift2 X, Pos))
3546 // always invokes undefined behavior for 32-bit X.
3548 // Below, Mask == OpSize - 1 when using [A] and is all-ones otherwise.
3549 unsigned MaskLoBits = 0;
3550 if (Neg.getOpcode() == ISD::AND &&
3551 isPowerOf2_64(OpSize) &&
3552 Neg.getOperand(1).getOpcode() == ISD::Constant &&
3553 cast<ConstantSDNode>(Neg.getOperand(1))->getAPIntValue() == OpSize - 1) {
3554 Neg = Neg.getOperand(0);
3555 MaskLoBits = Log2_64(OpSize);
3558 // Check whether Neg has the form (sub NegC, NegOp1) for some NegC and NegOp1.
3559 if (Neg.getOpcode() != ISD::SUB)
3561 ConstantSDNode *NegC = dyn_cast<ConstantSDNode>(Neg.getOperand(0));
3564 SDValue NegOp1 = Neg.getOperand(1);
3566 // On the RHS of [A], if Pos is Pos' & (OpSize - 1), just replace Pos with
3567 // Pos'. The truncation is redundant for the purpose of the equality.
3569 Pos.getOpcode() == ISD::AND &&
3570 Pos.getOperand(1).getOpcode() == ISD::Constant &&
3571 cast<ConstantSDNode>(Pos.getOperand(1))->getAPIntValue() == OpSize - 1)
3572 Pos = Pos.getOperand(0);
3574 // The condition we need is now:
3576 // (NegC - NegOp1) & Mask == (OpSize - Pos) & Mask
3578 // If NegOp1 == Pos then we need:
3580 // OpSize & Mask == NegC & Mask
3582 // (because "x & Mask" is a truncation and distributes through subtraction).
3585 Width = NegC->getAPIntValue();
3586 // Check for cases where Pos has the form (add NegOp1, PosC) for some PosC.
3587 // Then the condition we want to prove becomes:
3589 // (NegC - NegOp1) & Mask == (OpSize - (NegOp1 + PosC)) & Mask
3591 // which, again because "x & Mask" is a truncation, becomes:
3593 // NegC & Mask == (OpSize - PosC) & Mask
3594 // OpSize & Mask == (NegC + PosC) & Mask
3595 else if (Pos.getOpcode() == ISD::ADD &&
3596 Pos.getOperand(0) == NegOp1 &&
3597 Pos.getOperand(1).getOpcode() == ISD::Constant)
3598 Width = (cast<ConstantSDNode>(Pos.getOperand(1))->getAPIntValue() +
3599 NegC->getAPIntValue());
3603 // Now we just need to check that OpSize & Mask == Width & Mask.
3605 // Opsize & Mask is 0 since Mask is Opsize - 1.
3606 return Width.getLoBits(MaskLoBits) == 0;
3607 return Width == OpSize;
3610 // A subroutine of MatchRotate used once we have found an OR of two opposite
3611 // shifts of Shifted. If Neg == <operand size> - Pos then the OR reduces
3612 // to both (PosOpcode Shifted, Pos) and (NegOpcode Shifted, Neg), with the
3613 // former being preferred if supported. InnerPos and InnerNeg are Pos and
3614 // Neg with outer conversions stripped away.
3615 SDNode *DAGCombiner::MatchRotatePosNeg(SDValue Shifted, SDValue Pos,
3616 SDValue Neg, SDValue InnerPos,
3617 SDValue InnerNeg, unsigned PosOpcode,
3618 unsigned NegOpcode, SDLoc DL) {
3619 // fold (or (shl x, (*ext y)),
3620 // (srl x, (*ext (sub 32, y)))) ->
3621 // (rotl x, y) or (rotr x, (sub 32, y))
3623 // fold (or (shl x, (*ext (sub 32, y))),
3624 // (srl x, (*ext y))) ->
3625 // (rotr x, y) or (rotl x, (sub 32, y))
3626 EVT VT = Shifted.getValueType();
3627 if (matchRotateSub(InnerPos, InnerNeg, VT.getSizeInBits())) {
3628 bool HasPos = TLI.isOperationLegalOrCustom(PosOpcode, VT);
3629 return DAG.getNode(HasPos ? PosOpcode : NegOpcode, DL, VT, Shifted,
3630 HasPos ? Pos : Neg).getNode();
3636 // MatchRotate - Handle an 'or' of two operands. If this is one of the many
3637 // idioms for rotate, and if the target supports rotation instructions, generate
3639 SDNode *DAGCombiner::MatchRotate(SDValue LHS, SDValue RHS, SDLoc DL) {
3640 // Must be a legal type. Expanded 'n promoted things won't work with rotates.
3641 EVT VT = LHS.getValueType();
3642 if (!TLI.isTypeLegal(VT)) return nullptr;
3644 // The target must have at least one rotate flavor.
3645 bool HasROTL = TLI.isOperationLegalOrCustom(ISD::ROTL, VT);
3646 bool HasROTR = TLI.isOperationLegalOrCustom(ISD::ROTR, VT);
3647 if (!HasROTL && !HasROTR) return nullptr;
3649 // Match "(X shl/srl V1) & V2" where V2 may not be present.
3650 SDValue LHSShift; // The shift.
3651 SDValue LHSMask; // AND value if any.
3652 if (!MatchRotateHalf(LHS, LHSShift, LHSMask))
3653 return nullptr; // Not part of a rotate.
3655 SDValue RHSShift; // The shift.
3656 SDValue RHSMask; // AND value if any.
3657 if (!MatchRotateHalf(RHS, RHSShift, RHSMask))
3658 return nullptr; // Not part of a rotate.
3660 if (LHSShift.getOperand(0) != RHSShift.getOperand(0))
3661 return nullptr; // Not shifting the same value.
3663 if (LHSShift.getOpcode() == RHSShift.getOpcode())
3664 return nullptr; // Shifts must disagree.
3666 // Canonicalize shl to left side in a shl/srl pair.
3667 if (RHSShift.getOpcode() == ISD::SHL) {
3668 std::swap(LHS, RHS);
3669 std::swap(LHSShift, RHSShift);
3670 std::swap(LHSMask , RHSMask );
3673 unsigned OpSizeInBits = VT.getSizeInBits();
3674 SDValue LHSShiftArg = LHSShift.getOperand(0);
3675 SDValue LHSShiftAmt = LHSShift.getOperand(1);
3676 SDValue RHSShiftArg = RHSShift.getOperand(0);
3677 SDValue RHSShiftAmt = RHSShift.getOperand(1);
3679 // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1)
3680 // fold (or (shl x, C1), (srl x, C2)) -> (rotr x, C2)
3681 if (LHSShiftAmt.getOpcode() == ISD::Constant &&
3682 RHSShiftAmt.getOpcode() == ISD::Constant) {
3683 uint64_t LShVal = cast<ConstantSDNode>(LHSShiftAmt)->getZExtValue();
3684 uint64_t RShVal = cast<ConstantSDNode>(RHSShiftAmt)->getZExtValue();
3685 if ((LShVal + RShVal) != OpSizeInBits)
3688 SDValue Rot = DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT,
3689 LHSShiftArg, HasROTL ? LHSShiftAmt : RHSShiftAmt);
3691 // If there is an AND of either shifted operand, apply it to the result.
3692 if (LHSMask.getNode() || RHSMask.getNode()) {
3693 APInt Mask = APInt::getAllOnesValue(OpSizeInBits);
3695 if (LHSMask.getNode()) {
3696 APInt RHSBits = APInt::getLowBitsSet(OpSizeInBits, LShVal);
3697 Mask &= cast<ConstantSDNode>(LHSMask)->getAPIntValue() | RHSBits;
3699 if (RHSMask.getNode()) {
3700 APInt LHSBits = APInt::getHighBitsSet(OpSizeInBits, RShVal);
3701 Mask &= cast<ConstantSDNode>(RHSMask)->getAPIntValue() | LHSBits;
3704 Rot = DAG.getNode(ISD::AND, DL, VT, Rot, DAG.getConstant(Mask, VT));
3707 return Rot.getNode();
3710 // If there is a mask here, and we have a variable shift, we can't be sure
3711 // that we're masking out the right stuff.
3712 if (LHSMask.getNode() || RHSMask.getNode())
3715 // If the shift amount is sign/zext/any-extended just peel it off.
3716 SDValue LExtOp0 = LHSShiftAmt;
3717 SDValue RExtOp0 = RHSShiftAmt;
3718 if ((LHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3719 LHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3720 LHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3721 LHSShiftAmt.getOpcode() == ISD::TRUNCATE) &&
3722 (RHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3723 RHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3724 RHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3725 RHSShiftAmt.getOpcode() == ISD::TRUNCATE)) {
3726 LExtOp0 = LHSShiftAmt.getOperand(0);
3727 RExtOp0 = RHSShiftAmt.getOperand(0);
3730 SDNode *TryL = MatchRotatePosNeg(LHSShiftArg, LHSShiftAmt, RHSShiftAmt,
3731 LExtOp0, RExtOp0, ISD::ROTL, ISD::ROTR, DL);
3735 SDNode *TryR = MatchRotatePosNeg(RHSShiftArg, RHSShiftAmt, LHSShiftAmt,
3736 RExtOp0, LExtOp0, ISD::ROTR, ISD::ROTL, DL);
3743 SDValue DAGCombiner::visitXOR(SDNode *N) {
3744 SDValue N0 = N->getOperand(0);
3745 SDValue N1 = N->getOperand(1);
3746 SDValue LHS, RHS, CC;
3747 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3748 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3749 EVT VT = N0.getValueType();
3752 if (VT.isVector()) {
3753 SDValue FoldedVOp = SimplifyVBinOp(N);
3754 if (FoldedVOp.getNode()) return FoldedVOp;
3756 // fold (xor x, 0) -> x, vector edition
3757 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3759 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3763 // fold (xor undef, undef) -> 0. This is a common idiom (misuse).
3764 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
3765 return DAG.getConstant(0, VT);
3766 // fold (xor x, undef) -> undef
3767 if (N0.getOpcode() == ISD::UNDEF)
3769 if (N1.getOpcode() == ISD::UNDEF)
3771 // fold (xor c1, c2) -> c1^c2
3773 return DAG.FoldConstantArithmetic(ISD::XOR, VT, N0C, N1C);
3774 // canonicalize constant to RHS
3776 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0);
3777 // fold (xor x, 0) -> x
3778 if (N1C && N1C->isNullValue())
3781 SDValue RXOR = ReassociateOps(ISD::XOR, SDLoc(N), N0, N1);
3785 // fold !(x cc y) -> (x !cc y)
3786 if (N1C && N1C->getAPIntValue() == 1 && isSetCCEquivalent(N0, LHS, RHS, CC)) {
3787 bool isInt = LHS.getValueType().isInteger();
3788 ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
3791 if (!LegalOperations ||
3792 TLI.isCondCodeLegal(NotCC, LHS.getSimpleValueType())) {
3793 switch (N0.getOpcode()) {
3795 llvm_unreachable("Unhandled SetCC Equivalent!");
3797 return DAG.getSetCC(SDLoc(N), VT, LHS, RHS, NotCC);
3798 case ISD::SELECT_CC:
3799 return DAG.getSelectCC(SDLoc(N), LHS, RHS, N0.getOperand(2),
3800 N0.getOperand(3), NotCC);
3805 // fold (not (zext (setcc x, y))) -> (zext (not (setcc x, y)))
3806 if (N1C && N1C->getAPIntValue() == 1 && N0.getOpcode() == ISD::ZERO_EXTEND &&
3807 N0.getNode()->hasOneUse() &&
3808 isSetCCEquivalent(N0.getOperand(0), LHS, RHS, CC)){
3809 SDValue V = N0.getOperand(0);
3810 V = DAG.getNode(ISD::XOR, SDLoc(N0), V.getValueType(), V,
3811 DAG.getConstant(1, V.getValueType()));
3812 AddToWorklist(V.getNode());
3813 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, V);
3816 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are setcc
3817 if (N1C && N1C->getAPIntValue() == 1 && VT == MVT::i1 &&
3818 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3819 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3820 if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) {
3821 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3822 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS
3823 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS
3824 AddToWorklist(LHS.getNode()); AddToWorklist(RHS.getNode());
3825 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS);
3828 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are constants
3829 if (N1C && N1C->isAllOnesValue() &&
3830 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3831 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3832 if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) {
3833 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3834 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS
3835 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS
3836 AddToWorklist(LHS.getNode()); AddToWorklist(RHS.getNode());
3837 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS);
3840 // fold (xor (and x, y), y) -> (and (not x), y)
3841 if (N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
3842 N0->getOperand(1) == N1) {
3843 SDValue X = N0->getOperand(0);
3844 SDValue NotX = DAG.getNOT(SDLoc(X), X, VT);
3845 AddToWorklist(NotX.getNode());
3846 return DAG.getNode(ISD::AND, SDLoc(N), VT, NotX, N1);
3848 // fold (xor (xor x, c1), c2) -> (xor x, (xor c1, c2))
3849 if (N1C && N0.getOpcode() == ISD::XOR) {
3850 ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0));
3851 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3853 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N0.getOperand(1),
3854 DAG.getConstant(N1C->getAPIntValue() ^
3855 N00C->getAPIntValue(), VT));
3857 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N0.getOperand(0),
3858 DAG.getConstant(N1C->getAPIntValue() ^
3859 N01C->getAPIntValue(), VT));
3861 // fold (xor x, x) -> 0
3863 return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations, LegalTypes);
3865 // Simplify: xor (op x...), (op y...) -> (op (xor x, y))
3866 if (N0.getOpcode() == N1.getOpcode()) {
3867 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3868 if (Tmp.getNode()) return Tmp;
3871 // Simplify the expression using non-local knowledge.
3872 if (!VT.isVector() &&
3873 SimplifyDemandedBits(SDValue(N, 0)))
3874 return SDValue(N, 0);
3879 /// visitShiftByConstant - Handle transforms common to the three shifts, when
3880 /// the shift amount is a constant.
3881 SDValue DAGCombiner::visitShiftByConstant(SDNode *N, ConstantSDNode *Amt) {
3882 // We can't and shouldn't fold opaque constants.
3883 if (Amt->isOpaque())
3886 SDNode *LHS = N->getOperand(0).getNode();
3887 if (!LHS->hasOneUse()) return SDValue();
3889 // We want to pull some binops through shifts, so that we have (and (shift))
3890 // instead of (shift (and)), likewise for add, or, xor, etc. This sort of
3891 // thing happens with address calculations, so it's important to canonicalize
3893 bool HighBitSet = false; // Can we transform this if the high bit is set?
3895 switch (LHS->getOpcode()) {
3896 default: return SDValue();
3899 HighBitSet = false; // We can only transform sra if the high bit is clear.
3902 HighBitSet = true; // We can only transform sra if the high bit is set.
3905 if (N->getOpcode() != ISD::SHL)
3906 return SDValue(); // only shl(add) not sr[al](add).
3907 HighBitSet = false; // We can only transform sra if the high bit is clear.
3911 // We require the RHS of the binop to be a constant and not opaque as well.
3912 ConstantSDNode *BinOpCst = dyn_cast<ConstantSDNode>(LHS->getOperand(1));
3913 if (!BinOpCst || BinOpCst->isOpaque()) return SDValue();
3915 // FIXME: disable this unless the input to the binop is a shift by a constant.
3916 // If it is not a shift, it pessimizes some common cases like:
3918 // void foo(int *X, int i) { X[i & 1235] = 1; }
3919 // int bar(int *X, int i) { return X[i & 255]; }
3920 SDNode *BinOpLHSVal = LHS->getOperand(0).getNode();
3921 if ((BinOpLHSVal->getOpcode() != ISD::SHL &&
3922 BinOpLHSVal->getOpcode() != ISD::SRA &&
3923 BinOpLHSVal->getOpcode() != ISD::SRL) ||
3924 !isa<ConstantSDNode>(BinOpLHSVal->getOperand(1)))
3927 EVT VT = N->getValueType(0);
3929 // If this is a signed shift right, and the high bit is modified by the
3930 // logical operation, do not perform the transformation. The highBitSet
3931 // boolean indicates the value of the high bit of the constant which would
3932 // cause it to be modified for this operation.
3933 if (N->getOpcode() == ISD::SRA) {
3934 bool BinOpRHSSignSet = BinOpCst->getAPIntValue().isNegative();
3935 if (BinOpRHSSignSet != HighBitSet)
3939 if (!TLI.isDesirableToCommuteWithShift(LHS))
3942 // Fold the constants, shifting the binop RHS by the shift amount.
3943 SDValue NewRHS = DAG.getNode(N->getOpcode(), SDLoc(LHS->getOperand(1)),
3945 LHS->getOperand(1), N->getOperand(1));
3946 assert(isa<ConstantSDNode>(NewRHS) && "Folding was not successful!");
3948 // Create the new shift.
3949 SDValue NewShift = DAG.getNode(N->getOpcode(),
3950 SDLoc(LHS->getOperand(0)),
3951 VT, LHS->getOperand(0), N->getOperand(1));
3953 // Create the new binop.
3954 return DAG.getNode(LHS->getOpcode(), SDLoc(N), VT, NewShift, NewRHS);
3957 SDValue DAGCombiner::distributeTruncateThroughAnd(SDNode *N) {
3958 assert(N->getOpcode() == ISD::TRUNCATE);
3959 assert(N->getOperand(0).getOpcode() == ISD::AND);
3961 // (truncate:TruncVT (and N00, N01C)) -> (and (truncate:TruncVT N00), TruncC)
3962 if (N->hasOneUse() && N->getOperand(0).hasOneUse()) {
3963 SDValue N01 = N->getOperand(0).getOperand(1);
3965 if (ConstantSDNode *N01C = isConstOrConstSplat(N01)) {
3966 EVT TruncVT = N->getValueType(0);
3967 SDValue N00 = N->getOperand(0).getOperand(0);
3968 APInt TruncC = N01C->getAPIntValue();
3969 TruncC = TruncC.trunc(TruncVT.getScalarSizeInBits());
3971 return DAG.getNode(ISD::AND, SDLoc(N), TruncVT,
3972 DAG.getNode(ISD::TRUNCATE, SDLoc(N), TruncVT, N00),
3973 DAG.getConstant(TruncC, TruncVT));
3980 SDValue DAGCombiner::visitRotate(SDNode *N) {
3981 // fold (rot* x, (trunc (and y, c))) -> (rot* x, (and (trunc y), (trunc c))).
3982 if (N->getOperand(1).getOpcode() == ISD::TRUNCATE &&
3983 N->getOperand(1).getOperand(0).getOpcode() == ISD::AND) {
3984 SDValue NewOp1 = distributeTruncateThroughAnd(N->getOperand(1).getNode());
3985 if (NewOp1.getNode())
3986 return DAG.getNode(N->getOpcode(), SDLoc(N), N->getValueType(0),
3987 N->getOperand(0), NewOp1);
3992 SDValue DAGCombiner::visitSHL(SDNode *N) {
3993 SDValue N0 = N->getOperand(0);
3994 SDValue N1 = N->getOperand(1);
3995 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3996 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3997 EVT VT = N0.getValueType();
3998 unsigned OpSizeInBits = VT.getScalarSizeInBits();
4001 if (VT.isVector()) {
4002 SDValue FoldedVOp = SimplifyVBinOp(N);
4003 if (FoldedVOp.getNode()) return FoldedVOp;
4005 BuildVectorSDNode *N1CV = dyn_cast<BuildVectorSDNode>(N1);
4006 // If setcc produces all-one true value then:
4007 // (shl (and (setcc) N01CV) N1CV) -> (and (setcc) N01CV<<N1CV)
4008 if (N1CV && N1CV->isConstant()) {
4009 if (N0.getOpcode() == ISD::AND) {
4010 SDValue N00 = N0->getOperand(0);
4011 SDValue N01 = N0->getOperand(1);
4012 BuildVectorSDNode *N01CV = dyn_cast<BuildVectorSDNode>(N01);
4014 if (N01CV && N01CV->isConstant() && N00.getOpcode() == ISD::SETCC &&
4015 TLI.getBooleanContents(N00.getOperand(0).getValueType()) ==
4016 TargetLowering::ZeroOrNegativeOneBooleanContent) {
4017 SDValue C = DAG.FoldConstantArithmetic(ISD::SHL, VT, N01CV, N1CV);
4019 return DAG.getNode(ISD::AND, SDLoc(N), VT, N00, C);
4022 N1C = isConstOrConstSplat(N1);
4027 // fold (shl c1, c2) -> c1<<c2
4029 return DAG.FoldConstantArithmetic(ISD::SHL, VT, N0C, N1C);
4030 // fold (shl 0, x) -> 0
4031 if (N0C && N0C->isNullValue())
4033 // fold (shl x, c >= size(x)) -> undef
4034 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4035 return DAG.getUNDEF(VT);
4036 // fold (shl x, 0) -> x
4037 if (N1C && N1C->isNullValue())
4039 // fold (shl undef, x) -> 0
4040 if (N0.getOpcode() == ISD::UNDEF)
4041 return DAG.getConstant(0, VT);
4042 // if (shl x, c) is known to be zero, return 0
4043 if (DAG.MaskedValueIsZero(SDValue(N, 0),
4044 APInt::getAllOnesValue(OpSizeInBits)))
4045 return DAG.getConstant(0, VT);
4046 // fold (shl x, (trunc (and y, c))) -> (shl x, (and (trunc y), (trunc c))).
4047 if (N1.getOpcode() == ISD::TRUNCATE &&
4048 N1.getOperand(0).getOpcode() == ISD::AND) {
4049 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4050 if (NewOp1.getNode())
4051 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0, NewOp1);
4054 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4055 return SDValue(N, 0);
4057 // fold (shl (shl x, c1), c2) -> 0 or (shl x, (add c1, c2))
4058 if (N1C && N0.getOpcode() == ISD::SHL) {
4059 if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
4060 uint64_t c1 = N0C1->getZExtValue();
4061 uint64_t c2 = N1C->getZExtValue();
4062 if (c1 + c2 >= OpSizeInBits)
4063 return DAG.getConstant(0, VT);
4064 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0),
4065 DAG.getConstant(c1 + c2, N1.getValueType()));
4069 // fold (shl (ext (shl x, c1)), c2) -> (ext (shl x, (add c1, c2)))
4070 // For this to be valid, the second form must not preserve any of the bits
4071 // that are shifted out by the inner shift in the first form. This means
4072 // the outer shift size must be >= the number of bits added by the ext.
4073 // As a corollary, we don't care what kind of ext it is.
4074 if (N1C && (N0.getOpcode() == ISD::ZERO_EXTEND ||
4075 N0.getOpcode() == ISD::ANY_EXTEND ||
4076 N0.getOpcode() == ISD::SIGN_EXTEND) &&
4077 N0.getOperand(0).getOpcode() == ISD::SHL) {
4078 SDValue N0Op0 = N0.getOperand(0);
4079 if (ConstantSDNode *N0Op0C1 = isConstOrConstSplat(N0Op0.getOperand(1))) {
4080 uint64_t c1 = N0Op0C1->getZExtValue();
4081 uint64_t c2 = N1C->getZExtValue();
4082 EVT InnerShiftVT = N0Op0.getValueType();
4083 uint64_t InnerShiftSize = InnerShiftVT.getScalarSizeInBits();
4084 if (c2 >= OpSizeInBits - InnerShiftSize) {
4085 if (c1 + c2 >= OpSizeInBits)
4086 return DAG.getConstant(0, VT);
4087 return DAG.getNode(ISD::SHL, SDLoc(N0), VT,
4088 DAG.getNode(N0.getOpcode(), SDLoc(N0), VT,
4089 N0Op0->getOperand(0)),
4090 DAG.getConstant(c1 + c2, N1.getValueType()));
4095 // fold (shl (zext (srl x, C)), C) -> (zext (shl (srl x, C), C))
4096 // Only fold this if the inner zext has no other uses to avoid increasing
4097 // the total number of instructions.
4098 if (N1C && N0.getOpcode() == ISD::ZERO_EXTEND && N0.hasOneUse() &&
4099 N0.getOperand(0).getOpcode() == ISD::SRL) {
4100 SDValue N0Op0 = N0.getOperand(0);
4101 if (ConstantSDNode *N0Op0C1 = isConstOrConstSplat(N0Op0.getOperand(1))) {
4102 uint64_t c1 = N0Op0C1->getZExtValue();
4103 if (c1 < VT.getScalarSizeInBits()) {
4104 uint64_t c2 = N1C->getZExtValue();
4106 SDValue NewOp0 = N0.getOperand(0);
4107 EVT CountVT = NewOp0.getOperand(1).getValueType();
4108 SDValue NewSHL = DAG.getNode(ISD::SHL, SDLoc(N), NewOp0.getValueType(),
4109 NewOp0, DAG.getConstant(c2, CountVT));
4110 AddToWorklist(NewSHL.getNode());
4111 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N0), VT, NewSHL);
4117 // fold (shl (srl x, c1), c2) -> (and (shl x, (sub c2, c1), MASK) or
4118 // (and (srl x, (sub c1, c2), MASK)
4119 // Only fold this if the inner shift has no other uses -- if it does, folding
4120 // this will increase the total number of instructions.
4121 if (N1C && N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
4122 if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
4123 uint64_t c1 = N0C1->getZExtValue();
4124 if (c1 < OpSizeInBits) {
4125 uint64_t c2 = N1C->getZExtValue();
4126 APInt Mask = APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - c1);
4129 Mask = Mask.shl(c2 - c1);
4130 Shift = DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0),
4131 DAG.getConstant(c2 - c1, N1.getValueType()));
4133 Mask = Mask.lshr(c1 - c2);
4134 Shift = DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0),
4135 DAG.getConstant(c1 - c2, N1.getValueType()));
4137 return DAG.getNode(ISD::AND, SDLoc(N0), VT, Shift,
4138 DAG.getConstant(Mask, VT));
4142 // fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1))
4143 if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1)) {
4144 unsigned BitSize = VT.getScalarSizeInBits();
4145 SDValue HiBitsMask =
4146 DAG.getConstant(APInt::getHighBitsSet(BitSize,
4147 BitSize - N1C->getZExtValue()), VT);
4148 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0.getOperand(0),
4153 SDValue NewSHL = visitShiftByConstant(N, N1C);
4154 if (NewSHL.getNode())
4161 SDValue DAGCombiner::visitSRA(SDNode *N) {
4162 SDValue N0 = N->getOperand(0);
4163 SDValue N1 = N->getOperand(1);
4164 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4165 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4166 EVT VT = N0.getValueType();
4167 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
4170 if (VT.isVector()) {
4171 SDValue FoldedVOp = SimplifyVBinOp(N);
4172 if (FoldedVOp.getNode()) return FoldedVOp;
4174 N1C = isConstOrConstSplat(N1);
4177 // fold (sra c1, c2) -> (sra c1, c2)
4179 return DAG.FoldConstantArithmetic(ISD::SRA, VT, N0C, N1C);
4180 // fold (sra 0, x) -> 0
4181 if (N0C && N0C->isNullValue())
4183 // fold (sra -1, x) -> -1
4184 if (N0C && N0C->isAllOnesValue())
4186 // fold (sra x, (setge c, size(x))) -> undef
4187 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4188 return DAG.getUNDEF(VT);
4189 // fold (sra x, 0) -> x
4190 if (N1C && N1C->isNullValue())
4192 // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports
4194 if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) {
4195 unsigned LowBits = OpSizeInBits - (unsigned)N1C->getZExtValue();
4196 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), LowBits);
4198 ExtVT = EVT::getVectorVT(*DAG.getContext(),
4199 ExtVT, VT.getVectorNumElements());
4200 if ((!LegalOperations ||
4201 TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, ExtVT)))
4202 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
4203 N0.getOperand(0), DAG.getValueType(ExtVT));
4206 // fold (sra (sra x, c1), c2) -> (sra x, (add c1, c2))
4207 if (N1C && N0.getOpcode() == ISD::SRA) {
4208 if (ConstantSDNode *C1 = isConstOrConstSplat(N0.getOperand(1))) {
4209 unsigned Sum = N1C->getZExtValue() + C1->getZExtValue();
4210 if (Sum >= OpSizeInBits)
4211 Sum = OpSizeInBits - 1;
4212 return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0.getOperand(0),
4213 DAG.getConstant(Sum, N1.getValueType()));
4217 // fold (sra (shl X, m), (sub result_size, n))
4218 // -> (sign_extend (trunc (shl X, (sub (sub result_size, n), m)))) for
4219 // result_size - n != m.
4220 // If truncate is free for the target sext(shl) is likely to result in better
4222 if (N0.getOpcode() == ISD::SHL && N1C) {
4223 // Get the two constanst of the shifts, CN0 = m, CN = n.
4224 const ConstantSDNode *N01C = isConstOrConstSplat(N0.getOperand(1));
4226 LLVMContext &Ctx = *DAG.getContext();
4227 // Determine what the truncate's result bitsize and type would be.
4228 EVT TruncVT = EVT::getIntegerVT(Ctx, OpSizeInBits - N1C->getZExtValue());
4231 TruncVT = EVT::getVectorVT(Ctx, TruncVT, VT.getVectorNumElements());
4233 // Determine the residual right-shift amount.
4234 signed ShiftAmt = N1C->getZExtValue() - N01C->getZExtValue();
4236 // If the shift is not a no-op (in which case this should be just a sign
4237 // extend already), the truncated to type is legal, sign_extend is legal
4238 // on that type, and the truncate to that type is both legal and free,
4239 // perform the transform.
4240 if ((ShiftAmt > 0) &&
4241 TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND, TruncVT) &&
4242 TLI.isOperationLegalOrCustom(ISD::TRUNCATE, VT) &&
4243 TLI.isTruncateFree(VT, TruncVT)) {
4245 SDValue Amt = DAG.getConstant(ShiftAmt,
4246 getShiftAmountTy(N0.getOperand(0).getValueType()));
4247 SDValue Shift = DAG.getNode(ISD::SRL, SDLoc(N0), VT,
4248 N0.getOperand(0), Amt);
4249 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), TruncVT,
4251 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N),
4252 N->getValueType(0), Trunc);
4257 // fold (sra x, (trunc (and y, c))) -> (sra x, (and (trunc y), (trunc c))).
4258 if (N1.getOpcode() == ISD::TRUNCATE &&
4259 N1.getOperand(0).getOpcode() == ISD::AND) {
4260 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4261 if (NewOp1.getNode())
4262 return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0, NewOp1);
4265 // fold (sra (trunc (srl x, c1)), c2) -> (trunc (sra x, c1 + c2))
4266 // if c1 is equal to the number of bits the trunc removes
4267 if (N0.getOpcode() == ISD::TRUNCATE &&
4268 (N0.getOperand(0).getOpcode() == ISD::SRL ||
4269 N0.getOperand(0).getOpcode() == ISD::SRA) &&
4270 N0.getOperand(0).hasOneUse() &&
4271 N0.getOperand(0).getOperand(1).hasOneUse() &&
4273 SDValue N0Op0 = N0.getOperand(0);
4274 if (ConstantSDNode *LargeShift = isConstOrConstSplat(N0Op0.getOperand(1))) {
4275 unsigned LargeShiftVal = LargeShift->getZExtValue();
4276 EVT LargeVT = N0Op0.getValueType();
4278 if (LargeVT.getScalarSizeInBits() - OpSizeInBits == LargeShiftVal) {
4280 DAG.getConstant(LargeShiftVal + N1C->getZExtValue(),
4281 getShiftAmountTy(N0Op0.getOperand(0).getValueType()));
4282 SDValue SRA = DAG.getNode(ISD::SRA, SDLoc(N), LargeVT,
4283 N0Op0.getOperand(0), Amt);
4284 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, SRA);
4289 // Simplify, based on bits shifted out of the LHS.
4290 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4291 return SDValue(N, 0);
4294 // If the sign bit is known to be zero, switch this to a SRL.
4295 if (DAG.SignBitIsZero(N0))
4296 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, N1);
4299 SDValue NewSRA = visitShiftByConstant(N, N1C);
4300 if (NewSRA.getNode())
4307 SDValue DAGCombiner::visitSRL(SDNode *N) {
4308 SDValue N0 = N->getOperand(0);
4309 SDValue N1 = N->getOperand(1);
4310 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4311 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4312 EVT VT = N0.getValueType();
4313 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
4316 if (VT.isVector()) {
4317 SDValue FoldedVOp = SimplifyVBinOp(N);
4318 if (FoldedVOp.getNode()) return FoldedVOp;
4320 N1C = isConstOrConstSplat(N1);
4323 // fold (srl c1, c2) -> c1 >>u c2
4325 return DAG.FoldConstantArithmetic(ISD::SRL, VT, N0C, N1C);
4326 // fold (srl 0, x) -> 0
4327 if (N0C && N0C->isNullValue())
4329 // fold (srl x, c >= size(x)) -> undef
4330 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4331 return DAG.getUNDEF(VT);
4332 // fold (srl x, 0) -> x
4333 if (N1C && N1C->isNullValue())
4335 // if (srl x, c) is known to be zero, return 0
4336 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
4337 APInt::getAllOnesValue(OpSizeInBits)))
4338 return DAG.getConstant(0, VT);
4340 // fold (srl (srl x, c1), c2) -> 0 or (srl x, (add c1, c2))
4341 if (N1C && N0.getOpcode() == ISD::SRL) {
4342 if (ConstantSDNode *N01C = isConstOrConstSplat(N0.getOperand(1))) {
4343 uint64_t c1 = N01C->getZExtValue();
4344 uint64_t c2 = N1C->getZExtValue();
4345 if (c1 + c2 >= OpSizeInBits)
4346 return DAG.getConstant(0, VT);
4347 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0),
4348 DAG.getConstant(c1 + c2, N1.getValueType()));
4352 // fold (srl (trunc (srl x, c1)), c2) -> 0 or (trunc (srl x, (add c1, c2)))
4353 if (N1C && N0.getOpcode() == ISD::TRUNCATE &&
4354 N0.getOperand(0).getOpcode() == ISD::SRL &&
4355 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
4357 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
4358 uint64_t c2 = N1C->getZExtValue();
4359 EVT InnerShiftVT = N0.getOperand(0).getValueType();
4360 EVT ShiftCountVT = N0.getOperand(0)->getOperand(1).getValueType();
4361 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
4362 // This is only valid if the OpSizeInBits + c1 = size of inner shift.
4363 if (c1 + OpSizeInBits == InnerShiftSize) {
4364 if (c1 + c2 >= InnerShiftSize)
4365 return DAG.getConstant(0, VT);
4366 return DAG.getNode(ISD::TRUNCATE, SDLoc(N0), VT,
4367 DAG.getNode(ISD::SRL, SDLoc(N0), InnerShiftVT,
4368 N0.getOperand(0)->getOperand(0),
4369 DAG.getConstant(c1 + c2, ShiftCountVT)));
4373 // fold (srl (shl x, c), c) -> (and x, cst2)
4374 if (N1C && N0.getOpcode() == ISD::SHL && N0.getOperand(1) == N1) {
4375 unsigned BitSize = N0.getScalarValueSizeInBits();
4376 if (BitSize <= 64) {
4377 uint64_t ShAmt = N1C->getZExtValue() + 64 - BitSize;
4378 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0.getOperand(0),
4379 DAG.getConstant(~0ULL >> ShAmt, VT));
4383 // fold (srl (anyextend x), c) -> (and (anyextend (srl x, c)), mask)
4384 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
4385 // Shifting in all undef bits?
4386 EVT SmallVT = N0.getOperand(0).getValueType();
4387 unsigned BitSize = SmallVT.getScalarSizeInBits();
4388 if (N1C->getZExtValue() >= BitSize)
4389 return DAG.getUNDEF(VT);
4391 if (!LegalTypes || TLI.isTypeDesirableForOp(ISD::SRL, SmallVT)) {
4392 uint64_t ShiftAmt = N1C->getZExtValue();
4393 SDValue SmallShift = DAG.getNode(ISD::SRL, SDLoc(N0), SmallVT,
4395 DAG.getConstant(ShiftAmt, getShiftAmountTy(SmallVT)));
4396 AddToWorklist(SmallShift.getNode());
4397 APInt Mask = APInt::getAllOnesValue(OpSizeInBits).lshr(ShiftAmt);
4398 return DAG.getNode(ISD::AND, SDLoc(N), VT,
4399 DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, SmallShift),
4400 DAG.getConstant(Mask, VT));
4404 // fold (srl (sra X, Y), 31) -> (srl X, 31). This srl only looks at the sign
4405 // bit, which is unmodified by sra.
4406 if (N1C && N1C->getZExtValue() + 1 == OpSizeInBits) {
4407 if (N0.getOpcode() == ISD::SRA)
4408 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0), N1);
4411 // fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit).
4412 if (N1C && N0.getOpcode() == ISD::CTLZ &&
4413 N1C->getAPIntValue() == Log2_32(OpSizeInBits)) {
4414 APInt KnownZero, KnownOne;
4415 DAG.computeKnownBits(N0.getOperand(0), KnownZero, KnownOne);
4417 // If any of the input bits are KnownOne, then the input couldn't be all
4418 // zeros, thus the result of the srl will always be zero.
4419 if (KnownOne.getBoolValue()) return DAG.getConstant(0, VT);
4421 // If all of the bits input the to ctlz node are known to be zero, then
4422 // the result of the ctlz is "32" and the result of the shift is one.
4423 APInt UnknownBits = ~KnownZero;
4424 if (UnknownBits == 0) return DAG.getConstant(1, VT);
4426 // Otherwise, check to see if there is exactly one bit input to the ctlz.
4427 if ((UnknownBits & (UnknownBits - 1)) == 0) {
4428 // Okay, we know that only that the single bit specified by UnknownBits
4429 // could be set on input to the CTLZ node. If this bit is set, the SRL
4430 // will return 0, if it is clear, it returns 1. Change the CTLZ/SRL pair
4431 // to an SRL/XOR pair, which is likely to simplify more.
4432 unsigned ShAmt = UnknownBits.countTrailingZeros();
4433 SDValue Op = N0.getOperand(0);
4436 Op = DAG.getNode(ISD::SRL, SDLoc(N0), VT, Op,
4437 DAG.getConstant(ShAmt, getShiftAmountTy(Op.getValueType())));
4438 AddToWorklist(Op.getNode());
4441 return DAG.getNode(ISD::XOR, SDLoc(N), VT,
4442 Op, DAG.getConstant(1, VT));
4446 // fold (srl x, (trunc (and y, c))) -> (srl x, (and (trunc y), (trunc c))).
4447 if (N1.getOpcode() == ISD::TRUNCATE &&
4448 N1.getOperand(0).getOpcode() == ISD::AND) {
4449 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4450 if (NewOp1.getNode())
4451 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, NewOp1);
4454 // fold operands of srl based on knowledge that the low bits are not
4456 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4457 return SDValue(N, 0);
4460 SDValue NewSRL = visitShiftByConstant(N, N1C);
4461 if (NewSRL.getNode())
4465 // Attempt to convert a srl of a load into a narrower zero-extending load.
4466 SDValue NarrowLoad = ReduceLoadWidth(N);
4467 if (NarrowLoad.getNode())
4470 // Here is a common situation. We want to optimize:
4473 // %b = and i32 %a, 2
4474 // %c = srl i32 %b, 1
4475 // brcond i32 %c ...
4481 // %c = setcc eq %b, 0
4484 // However when after the source operand of SRL is optimized into AND, the SRL
4485 // itself may not be optimized further. Look for it and add the BRCOND into
4487 if (N->hasOneUse()) {
4488 SDNode *Use = *N->use_begin();
4489 if (Use->getOpcode() == ISD::BRCOND)
4491 else if (Use->getOpcode() == ISD::TRUNCATE && Use->hasOneUse()) {
4492 // Also look pass the truncate.
4493 Use = *Use->use_begin();
4494 if (Use->getOpcode() == ISD::BRCOND)
4502 SDValue DAGCombiner::visitCTLZ(SDNode *N) {
4503 SDValue N0 = N->getOperand(0);
4504 EVT VT = N->getValueType(0);
4506 // fold (ctlz c1) -> c2
4507 if (isa<ConstantSDNode>(N0))
4508 return DAG.getNode(ISD::CTLZ, SDLoc(N), VT, N0);
4512 SDValue DAGCombiner::visitCTLZ_ZERO_UNDEF(SDNode *N) {
4513 SDValue N0 = N->getOperand(0);
4514 EVT VT = N->getValueType(0);
4516 // fold (ctlz_zero_undef c1) -> c2
4517 if (isa<ConstantSDNode>(N0))
4518 return DAG.getNode(ISD::CTLZ_ZERO_UNDEF, SDLoc(N), VT, N0);
4522 SDValue DAGCombiner::visitCTTZ(SDNode *N) {
4523 SDValue N0 = N->getOperand(0);
4524 EVT VT = N->getValueType(0);
4526 // fold (cttz c1) -> c2
4527 if (isa<ConstantSDNode>(N0))
4528 return DAG.getNode(ISD::CTTZ, SDLoc(N), VT, N0);
4532 SDValue DAGCombiner::visitCTTZ_ZERO_UNDEF(SDNode *N) {
4533 SDValue N0 = N->getOperand(0);
4534 EVT VT = N->getValueType(0);
4536 // fold (cttz_zero_undef c1) -> c2
4537 if (isa<ConstantSDNode>(N0))
4538 return DAG.getNode(ISD::CTTZ_ZERO_UNDEF, SDLoc(N), VT, N0);
4542 SDValue DAGCombiner::visitCTPOP(SDNode *N) {
4543 SDValue N0 = N->getOperand(0);
4544 EVT VT = N->getValueType(0);
4546 // fold (ctpop c1) -> c2
4547 if (isa<ConstantSDNode>(N0))
4548 return DAG.getNode(ISD::CTPOP, SDLoc(N), VT, N0);
4552 SDValue DAGCombiner::visitSELECT(SDNode *N) {
4553 SDValue N0 = N->getOperand(0);
4554 SDValue N1 = N->getOperand(1);
4555 SDValue N2 = N->getOperand(2);
4556 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4557 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4558 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
4559 EVT VT = N->getValueType(0);
4560 EVT VT0 = N0.getValueType();
4562 // fold (select C, X, X) -> X
4565 // fold (select true, X, Y) -> X
4566 if (N0C && !N0C->isNullValue())
4568 // fold (select false, X, Y) -> Y
4569 if (N0C && N0C->isNullValue())
4571 // fold (select C, 1, X) -> (or C, X)
4572 if (VT == MVT::i1 && N1C && N1C->getAPIntValue() == 1)
4573 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N2);
4574 // fold (select C, 0, 1) -> (xor C, 1)
4575 // We can't do this reliably if integer based booleans have different contents
4576 // to floating point based booleans. This is because we can't tell whether we
4577 // have an integer-based boolean or a floating-point-based boolean unless we
4578 // can find the SETCC that produced it and inspect its operands. This is
4579 // fairly easy if C is the SETCC node, but it can potentially be
4580 // undiscoverable (or not reasonably discoverable). For example, it could be
4581 // in another basic block or it could require searching a complicated
4583 if (VT.isInteger() &&
4584 (VT0 == MVT::i1 || (VT0.isInteger() &&
4585 TLI.getBooleanContents(false, false) ==
4586 TLI.getBooleanContents(false, true) &&
4587 TLI.getBooleanContents(false, false) ==
4588 TargetLowering::ZeroOrOneBooleanContent)) &&
4589 N1C && N2C && N1C->isNullValue() && N2C->getAPIntValue() == 1) {
4592 return DAG.getNode(ISD::XOR, SDLoc(N), VT0,
4593 N0, DAG.getConstant(1, VT0));
4594 XORNode = DAG.getNode(ISD::XOR, SDLoc(N0), VT0,
4595 N0, DAG.getConstant(1, VT0));
4596 AddToWorklist(XORNode.getNode());
4598 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, XORNode);
4599 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, XORNode);
4601 // fold (select C, 0, X) -> (and (not C), X)
4602 if (VT == VT0 && VT == MVT::i1 && N1C && N1C->isNullValue()) {
4603 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT);
4604 AddToWorklist(NOTNode.getNode());
4605 return DAG.getNode(ISD::AND, SDLoc(N), VT, NOTNode, N2);
4607 // fold (select C, X, 1) -> (or (not C), X)
4608 if (VT == VT0 && VT == MVT::i1 && N2C && N2C->getAPIntValue() == 1) {
4609 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT);
4610 AddToWorklist(NOTNode.getNode());
4611 return DAG.getNode(ISD::OR, SDLoc(N), VT, NOTNode, N1);
4613 // fold (select C, X, 0) -> (and C, X)
4614 if (VT == MVT::i1 && N2C && N2C->isNullValue())
4615 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, N1);
4616 // fold (select X, X, Y) -> (or X, Y)
4617 // fold (select X, 1, Y) -> (or X, Y)
4618 if (VT == MVT::i1 && (N0 == N1 || (N1C && N1C->getAPIntValue() == 1)))
4619 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N2);
4620 // fold (select X, Y, X) -> (and X, Y)
4621 // fold (select X, Y, 0) -> (and X, Y)
4622 if (VT == MVT::i1 && (N0 == N2 || (N2C && N2C->getAPIntValue() == 0)))
4623 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, N1);
4625 // If we can fold this based on the true/false value, do so.
4626 if (SimplifySelectOps(N, N1, N2))
4627 return SDValue(N, 0); // Don't revisit N.
4629 // fold selects based on a setcc into other things, such as min/max/abs
4630 if (N0.getOpcode() == ISD::SETCC) {
4631 if ((!LegalOperations &&
4632 TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT)) ||
4633 TLI.isOperationLegal(ISD::SELECT_CC, VT))
4634 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT,
4635 N0.getOperand(0), N0.getOperand(1),
4636 N1, N2, N0.getOperand(2));
4637 return SimplifySelect(SDLoc(N), N0, N1, N2);
4644 std::pair<SDValue, SDValue> SplitVSETCC(const SDNode *N, SelectionDAG &DAG) {
4647 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0));
4649 // Split the inputs.
4650 SDValue Lo, Hi, LL, LH, RL, RH;
4651 std::tie(LL, LH) = DAG.SplitVectorOperand(N, 0);
4652 std::tie(RL, RH) = DAG.SplitVectorOperand(N, 1);
4654 Lo = DAG.getNode(N->getOpcode(), DL, LoVT, LL, RL, N->getOperand(2));
4655 Hi = DAG.getNode(N->getOpcode(), DL, HiVT, LH, RH, N->getOperand(2));
4657 return std::make_pair(Lo, Hi);
4660 // This function assumes all the vselect's arguments are CONCAT_VECTOR
4661 // nodes and that the condition is a BV of ConstantSDNodes (or undefs).
4662 static SDValue ConvertSelectToConcatVector(SDNode *N, SelectionDAG &DAG) {
4664 SDValue Cond = N->getOperand(0);
4665 SDValue LHS = N->getOperand(1);
4666 SDValue RHS = N->getOperand(2);
4667 MVT VT = N->getSimpleValueType(0);
4668 int NumElems = VT.getVectorNumElements();
4669 assert(LHS.getOpcode() == ISD::CONCAT_VECTORS &&
4670 RHS.getOpcode() == ISD::CONCAT_VECTORS &&
4671 Cond.getOpcode() == ISD::BUILD_VECTOR);
4673 // We're sure we have an even number of elements due to the
4674 // concat_vectors we have as arguments to vselect.
4675 // Skip BV elements until we find one that's not an UNDEF
4676 // After we find an UNDEF element, keep looping until we get to half the
4677 // length of the BV and see if all the non-undef nodes are the same.
4678 ConstantSDNode *BottomHalf = nullptr;
4679 for (int i = 0; i < NumElems / 2; ++i) {
4680 if (Cond->getOperand(i)->getOpcode() == ISD::UNDEF)
4683 if (BottomHalf == nullptr)
4684 BottomHalf = cast<ConstantSDNode>(Cond.getOperand(i));
4685 else if (Cond->getOperand(i).getNode() != BottomHalf)
4689 // Do the same for the second half of the BuildVector
4690 ConstantSDNode *TopHalf = nullptr;
4691 for (int i = NumElems / 2; i < NumElems; ++i) {
4692 if (Cond->getOperand(i)->getOpcode() == ISD::UNDEF)
4695 if (TopHalf == nullptr)
4696 TopHalf = cast<ConstantSDNode>(Cond.getOperand(i));
4697 else if (Cond->getOperand(i).getNode() != TopHalf)
4701 assert(TopHalf && BottomHalf &&
4702 "One half of the selector was all UNDEFs and the other was all the "
4703 "same value. This should have been addressed before this function.");
4705 ISD::CONCAT_VECTORS, dl, VT,
4706 BottomHalf->isNullValue() ? RHS->getOperand(0) : LHS->getOperand(0),
4707 TopHalf->isNullValue() ? RHS->getOperand(1) : LHS->getOperand(1));
4710 SDValue DAGCombiner::visitVSELECT(SDNode *N) {
4711 SDValue N0 = N->getOperand(0);
4712 SDValue N1 = N->getOperand(1);
4713 SDValue N2 = N->getOperand(2);
4716 // Canonicalize integer abs.
4717 // vselect (setg[te] X, 0), X, -X ->
4718 // vselect (setgt X, -1), X, -X ->
4719 // vselect (setl[te] X, 0), -X, X ->
4720 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
4721 if (N0.getOpcode() == ISD::SETCC) {
4722 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
4723 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
4725 bool RHSIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
4727 if (((RHSIsAllZeros && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
4728 (ISD::isBuildVectorAllOnes(RHS.getNode()) && CC == ISD::SETGT)) &&
4729 N1 == LHS && N2.getOpcode() == ISD::SUB && N1 == N2.getOperand(1))
4730 isAbs = ISD::isBuildVectorAllZeros(N2.getOperand(0).getNode());
4731 else if ((RHSIsAllZeros && (CC == ISD::SETLT || CC == ISD::SETLE)) &&
4732 N2 == LHS && N1.getOpcode() == ISD::SUB && N2 == N1.getOperand(1))
4733 isAbs = ISD::isBuildVectorAllZeros(N1.getOperand(0).getNode());
4736 EVT VT = LHS.getValueType();
4737 SDValue Shift = DAG.getNode(
4738 ISD::SRA, DL, VT, LHS,
4739 DAG.getConstant(VT.getScalarType().getSizeInBits() - 1, VT));
4740 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, LHS, Shift);
4741 AddToWorklist(Shift.getNode());
4742 AddToWorklist(Add.getNode());
4743 return DAG.getNode(ISD::XOR, DL, VT, Add, Shift);
4747 // If the VSELECT result requires splitting and the mask is provided by a
4748 // SETCC, then split both nodes and its operands before legalization. This
4749 // prevents the type legalizer from unrolling SETCC into scalar comparisons
4750 // and enables future optimizations (e.g. min/max pattern matching on X86).
4751 if (N0.getOpcode() == ISD::SETCC) {
4752 EVT VT = N->getValueType(0);
4754 // Check if any splitting is required.
4755 if (TLI.getTypeAction(*DAG.getContext(), VT) !=
4756 TargetLowering::TypeSplitVector)
4759 SDValue Lo, Hi, CCLo, CCHi, LL, LH, RL, RH;
4760 std::tie(CCLo, CCHi) = SplitVSETCC(N0.getNode(), DAG);
4761 std::tie(LL, LH) = DAG.SplitVectorOperand(N, 1);
4762 std::tie(RL, RH) = DAG.SplitVectorOperand(N, 2);
4764 Lo = DAG.getNode(N->getOpcode(), DL, LL.getValueType(), CCLo, LL, RL);
4765 Hi = DAG.getNode(N->getOpcode(), DL, LH.getValueType(), CCHi, LH, RH);
4767 // Add the new VSELECT nodes to the work list in case they need to be split
4769 AddToWorklist(Lo.getNode());
4770 AddToWorklist(Hi.getNode());
4772 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
4775 // Fold (vselect (build_vector all_ones), N1, N2) -> N1
4776 if (ISD::isBuildVectorAllOnes(N0.getNode()))
4778 // Fold (vselect (build_vector all_zeros), N1, N2) -> N2
4779 if (ISD::isBuildVectorAllZeros(N0.getNode()))
4782 // The ConvertSelectToConcatVector function is assuming both the above
4783 // checks for (vselect (build_vector all{ones,zeros) ...) have been made
4785 if (N1.getOpcode() == ISD::CONCAT_VECTORS &&
4786 N2.getOpcode() == ISD::CONCAT_VECTORS &&
4787 ISD::isBuildVectorOfConstantSDNodes(N0.getNode())) {
4788 SDValue CV = ConvertSelectToConcatVector(N, DAG);
4796 SDValue DAGCombiner::visitSELECT_CC(SDNode *N) {
4797 SDValue N0 = N->getOperand(0);
4798 SDValue N1 = N->getOperand(1);
4799 SDValue N2 = N->getOperand(2);
4800 SDValue N3 = N->getOperand(3);
4801 SDValue N4 = N->getOperand(4);
4802 ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get();
4804 // fold select_cc lhs, rhs, x, x, cc -> x
4808 // Determine if the condition we're dealing with is constant
4809 SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()),
4810 N0, N1, CC, SDLoc(N), false);
4811 if (SCC.getNode()) {
4812 AddToWorklist(SCC.getNode());
4814 if (ConstantSDNode *SCCC = dyn_cast<ConstantSDNode>(SCC.getNode())) {
4815 if (!SCCC->isNullValue())
4816 return N2; // cond always true -> true val
4818 return N3; // cond always false -> false val
4821 // Fold to a simpler select_cc
4822 if (SCC.getOpcode() == ISD::SETCC)
4823 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), N2.getValueType(),
4824 SCC.getOperand(0), SCC.getOperand(1), N2, N3,
4828 // If we can fold this based on the true/false value, do so.
4829 if (SimplifySelectOps(N, N2, N3))
4830 return SDValue(N, 0); // Don't revisit N.
4832 // fold select_cc into other things, such as min/max/abs
4833 return SimplifySelectCC(SDLoc(N), N0, N1, N2, N3, CC);
4836 SDValue DAGCombiner::visitSETCC(SDNode *N) {
4837 return SimplifySetCC(N->getValueType(0), N->getOperand(0), N->getOperand(1),
4838 cast<CondCodeSDNode>(N->getOperand(2))->get(),
4842 // tryToFoldExtendOfConstant - Try to fold a sext/zext/aext
4843 // dag node into a ConstantSDNode or a build_vector of constants.
4844 // This function is called by the DAGCombiner when visiting sext/zext/aext
4845 // dag nodes (see for example method DAGCombiner::visitSIGN_EXTEND).
4846 // Vector extends are not folded if operations are legal; this is to
4847 // avoid introducing illegal build_vector dag nodes.
4848 static SDNode *tryToFoldExtendOfConstant(SDNode *N, const TargetLowering &TLI,
4849 SelectionDAG &DAG, bool LegalTypes,
4850 bool LegalOperations) {
4851 unsigned Opcode = N->getOpcode();
4852 SDValue N0 = N->getOperand(0);
4853 EVT VT = N->getValueType(0);
4855 assert((Opcode == ISD::SIGN_EXTEND || Opcode == ISD::ZERO_EXTEND ||
4856 Opcode == ISD::ANY_EXTEND) && "Expected EXTEND dag node in input!");
4858 // fold (sext c1) -> c1
4859 // fold (zext c1) -> c1
4860 // fold (aext c1) -> c1
4861 if (isa<ConstantSDNode>(N0))
4862 return DAG.getNode(Opcode, SDLoc(N), VT, N0).getNode();
4864 // fold (sext (build_vector AllConstants) -> (build_vector AllConstants)
4865 // fold (zext (build_vector AllConstants) -> (build_vector AllConstants)
4866 // fold (aext (build_vector AllConstants) -> (build_vector AllConstants)
4867 EVT SVT = VT.getScalarType();
4868 if (!(VT.isVector() &&
4869 (!LegalTypes || (!LegalOperations && TLI.isTypeLegal(SVT))) &&
4870 ISD::isBuildVectorOfConstantSDNodes(N0.getNode())))
4873 // We can fold this node into a build_vector.
4874 unsigned VTBits = SVT.getSizeInBits();
4875 unsigned EVTBits = N0->getValueType(0).getScalarType().getSizeInBits();
4876 unsigned ShAmt = VTBits - EVTBits;
4877 SmallVector<SDValue, 8> Elts;
4878 unsigned NumElts = N0->getNumOperands();
4881 for (unsigned i=0; i != NumElts; ++i) {
4882 SDValue Op = N0->getOperand(i);
4883 if (Op->getOpcode() == ISD::UNDEF) {
4884 Elts.push_back(DAG.getUNDEF(SVT));
4888 ConstantSDNode *CurrentND = cast<ConstantSDNode>(Op);
4889 const APInt &C = APInt(VTBits, CurrentND->getAPIntValue().getZExtValue());
4890 if (Opcode == ISD::SIGN_EXTEND)
4891 Elts.push_back(DAG.getConstant(C.shl(ShAmt).ashr(ShAmt).getZExtValue(),
4894 Elts.push_back(DAG.getConstant(C.shl(ShAmt).lshr(ShAmt).getZExtValue(),
4898 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Elts).getNode();
4901 // ExtendUsesToFormExtLoad - Trying to extend uses of a load to enable this:
4902 // "fold ({s|z|a}ext (load x)) -> ({s|z|a}ext (truncate ({s|z|a}extload x)))"
4903 // transformation. Returns true if extension are possible and the above
4904 // mentioned transformation is profitable.
4905 static bool ExtendUsesToFormExtLoad(SDNode *N, SDValue N0,
4907 SmallVectorImpl<SDNode *> &ExtendNodes,
4908 const TargetLowering &TLI) {
4909 bool HasCopyToRegUses = false;
4910 bool isTruncFree = TLI.isTruncateFree(N->getValueType(0), N0.getValueType());
4911 for (SDNode::use_iterator UI = N0.getNode()->use_begin(),
4912 UE = N0.getNode()->use_end();
4917 if (UI.getUse().getResNo() != N0.getResNo())
4919 // FIXME: Only extend SETCC N, N and SETCC N, c for now.
4920 if (ExtOpc != ISD::ANY_EXTEND && User->getOpcode() == ISD::SETCC) {
4921 ISD::CondCode CC = cast<CondCodeSDNode>(User->getOperand(2))->get();
4922 if (ExtOpc == ISD::ZERO_EXTEND && ISD::isSignedIntSetCC(CC))
4923 // Sign bits will be lost after a zext.
4926 for (unsigned i = 0; i != 2; ++i) {
4927 SDValue UseOp = User->getOperand(i);
4930 if (!isa<ConstantSDNode>(UseOp))
4935 ExtendNodes.push_back(User);
4938 // If truncates aren't free and there are users we can't
4939 // extend, it isn't worthwhile.
4942 // Remember if this value is live-out.
4943 if (User->getOpcode() == ISD::CopyToReg)
4944 HasCopyToRegUses = true;
4947 if (HasCopyToRegUses) {
4948 bool BothLiveOut = false;
4949 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
4951 SDUse &Use = UI.getUse();
4952 if (Use.getResNo() == 0 && Use.getUser()->getOpcode() == ISD::CopyToReg) {
4958 // Both unextended and extended values are live out. There had better be
4959 // a good reason for the transformation.
4960 return ExtendNodes.size();
4965 void DAGCombiner::ExtendSetCCUses(const SmallVectorImpl<SDNode *> &SetCCs,
4966 SDValue Trunc, SDValue ExtLoad, SDLoc DL,
4967 ISD::NodeType ExtType) {
4968 // Extend SetCC uses if necessary.
4969 for (unsigned i = 0, e = SetCCs.size(); i != e; ++i) {
4970 SDNode *SetCC = SetCCs[i];
4971 SmallVector<SDValue, 4> Ops;
4973 for (unsigned j = 0; j != 2; ++j) {
4974 SDValue SOp = SetCC->getOperand(j);
4976 Ops.push_back(ExtLoad);
4978 Ops.push_back(DAG.getNode(ExtType, DL, ExtLoad->getValueType(0), SOp));
4981 Ops.push_back(SetCC->getOperand(2));
4982 CombineTo(SetCC, DAG.getNode(ISD::SETCC, DL, SetCC->getValueType(0), Ops));
4986 SDValue DAGCombiner::visitSIGN_EXTEND(SDNode *N) {
4987 SDValue N0 = N->getOperand(0);
4988 EVT VT = N->getValueType(0);
4990 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
4992 return SDValue(Res, 0);
4994 // fold (sext (sext x)) -> (sext x)
4995 // fold (sext (aext x)) -> (sext x)
4996 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
4997 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT,
5000 if (N0.getOpcode() == ISD::TRUNCATE) {
5001 // fold (sext (truncate (load x))) -> (sext (smaller load x))
5002 // fold (sext (truncate (srl (load x), c))) -> (sext (smaller load (x+c/n)))
5003 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5004 if (NarrowLoad.getNode()) {
5005 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5006 if (NarrowLoad.getNode() != N0.getNode()) {
5007 CombineTo(N0.getNode(), NarrowLoad);
5008 // CombineTo deleted the truncate, if needed, but not what's under it.
5011 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5014 // See if the value being truncated is already sign extended. If so, just
5015 // eliminate the trunc/sext pair.
5016 SDValue Op = N0.getOperand(0);
5017 unsigned OpBits = Op.getValueType().getScalarType().getSizeInBits();
5018 unsigned MidBits = N0.getValueType().getScalarType().getSizeInBits();
5019 unsigned DestBits = VT.getScalarType().getSizeInBits();
5020 unsigned NumSignBits = DAG.ComputeNumSignBits(Op);
5022 if (OpBits == DestBits) {
5023 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
5024 // bits, it is already ready.
5025 if (NumSignBits > DestBits-MidBits)
5027 } else if (OpBits < DestBits) {
5028 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
5029 // bits, just sext from i32.
5030 if (NumSignBits > OpBits-MidBits)
5031 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, Op);
5033 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
5034 // bits, just truncate to i32.
5035 if (NumSignBits > OpBits-MidBits)
5036 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
5039 // fold (sext (truncate x)) -> (sextinreg x).
5040 if (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG,
5041 N0.getValueType())) {
5042 if (OpBits < DestBits)
5043 Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N0), VT, Op);
5044 else if (OpBits > DestBits)
5045 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), VT, Op);
5046 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, Op,
5047 DAG.getValueType(N0.getValueType()));
5051 // fold (sext (load x)) -> (sext (truncate (sextload x)))
5052 // None of the supported targets knows how to perform load and sign extend
5053 // on vectors in one instruction. We only perform this transformation on
5055 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
5056 ISD::isUNINDEXEDLoad(N0.getNode()) &&
5057 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5058 TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()))) {
5059 bool DoXform = true;
5060 SmallVector<SDNode*, 4> SetCCs;
5061 if (!N0.hasOneUse())
5062 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::SIGN_EXTEND, SetCCs, TLI);
5064 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5065 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5067 LN0->getBasePtr(), N0.getValueType(),
5068 LN0->getMemOperand());
5069 CombineTo(N, ExtLoad);
5070 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5071 N0.getValueType(), ExtLoad);
5072 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
5073 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5075 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5079 // fold (sext (sextload x)) -> (sext (truncate (sextload x)))
5080 // fold (sext ( extload x)) -> (sext (truncate (sextload x)))
5081 if ((ISD::isSEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
5082 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
5083 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5084 EVT MemVT = LN0->getMemoryVT();
5085 if ((!LegalOperations && !LN0->isVolatile()) ||
5086 TLI.isLoadExtLegal(ISD::SEXTLOAD, MemVT)) {
5087 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5089 LN0->getBasePtr(), MemVT,
5090 LN0->getMemOperand());
5091 CombineTo(N, ExtLoad);
5092 CombineTo(N0.getNode(),
5093 DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5094 N0.getValueType(), ExtLoad),
5095 ExtLoad.getValue(1));
5096 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5100 // fold (sext (and/or/xor (load x), cst)) ->
5101 // (and/or/xor (sextload x), (sext cst))
5102 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
5103 N0.getOpcode() == ISD::XOR) &&
5104 isa<LoadSDNode>(N0.getOperand(0)) &&
5105 N0.getOperand(1).getOpcode() == ISD::Constant &&
5106 TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()) &&
5107 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
5108 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
5109 if (LN0->getExtensionType() != ISD::ZEXTLOAD && LN0->isUnindexed()) {
5110 bool DoXform = true;
5111 SmallVector<SDNode*, 4> SetCCs;
5112 if (!N0.hasOneUse())
5113 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::SIGN_EXTEND,
5116 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(LN0), VT,
5117 LN0->getChain(), LN0->getBasePtr(),
5119 LN0->getMemOperand());
5120 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5121 Mask = Mask.sext(VT.getSizeInBits());
5122 SDValue And = DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
5123 ExtLoad, DAG.getConstant(Mask, VT));
5124 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
5125 SDLoc(N0.getOperand(0)),
5126 N0.getOperand(0).getValueType(), ExtLoad);
5128 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
5129 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5131 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5136 if (N0.getOpcode() == ISD::SETCC) {
5137 EVT N0VT = N0.getOperand(0).getValueType();
5138 // sext(setcc) -> sext_in_reg(vsetcc) for vectors.
5139 // Only do this before legalize for now.
5140 if (VT.isVector() && !LegalOperations &&
5141 TLI.getBooleanContents(N0VT) ==
5142 TargetLowering::ZeroOrNegativeOneBooleanContent) {
5143 // On some architectures (such as SSE/NEON/etc) the SETCC result type is
5144 // of the same size as the compared operands. Only optimize sext(setcc())
5145 // if this is the case.
5146 EVT SVT = getSetCCResultType(N0VT);
5148 // We know that the # elements of the results is the same as the
5149 // # elements of the compare (and the # elements of the compare result
5150 // for that matter). Check to see that they are the same size. If so,
5151 // we know that the element size of the sext'd result matches the
5152 // element size of the compare operands.
5153 if (VT.getSizeInBits() == SVT.getSizeInBits())
5154 return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
5156 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5158 // If the desired elements are smaller or larger than the source
5159 // elements we can use a matching integer vector type and then
5160 // truncate/sign extend
5161 EVT MatchingVectorType = N0VT.changeVectorElementTypeToInteger();
5162 if (SVT == MatchingVectorType) {
5163 SDValue VsetCC = DAG.getSetCC(SDLoc(N), MatchingVectorType,
5164 N0.getOperand(0), N0.getOperand(1),
5165 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5166 return DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT);
5170 // sext(setcc x, y, cc) -> (select (setcc x, y, cc), -1, 0)
5171 unsigned ElementWidth = VT.getScalarType().getSizeInBits();
5173 DAG.getConstant(APInt::getAllOnesValue(ElementWidth), VT);
5175 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
5176 NegOne, DAG.getConstant(0, VT),
5177 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
5178 if (SCC.getNode()) return SCC;
5180 if (!VT.isVector()) {
5181 EVT SetCCVT = getSetCCResultType(N0.getOperand(0).getValueType());
5182 if (!LegalOperations || TLI.isOperationLegal(ISD::SETCC, SetCCVT)) {
5184 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
5185 SDValue SetCC = DAG.getSetCC(DL,
5187 N0.getOperand(0), N0.getOperand(1), CC);
5188 EVT SelectVT = getSetCCResultType(VT);
5189 return DAG.getSelect(DL, VT,
5190 DAG.getSExtOrTrunc(SetCC, DL, SelectVT),
5191 NegOne, DAG.getConstant(0, VT));
5197 // fold (sext x) -> (zext x) if the sign bit is known zero.
5198 if ((!LegalOperations || TLI.isOperationLegal(ISD::ZERO_EXTEND, VT)) &&
5199 DAG.SignBitIsZero(N0))
5200 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, N0);
5205 // isTruncateOf - If N is a truncate of some other value, return true, record
5206 // the value being truncated in Op and which of Op's bits are zero in KnownZero.
5207 // This function computes KnownZero to avoid a duplicated call to
5208 // computeKnownBits in the caller.
5209 static bool isTruncateOf(SelectionDAG &DAG, SDValue N, SDValue &Op,
5212 if (N->getOpcode() == ISD::TRUNCATE) {
5213 Op = N->getOperand(0);
5214 DAG.computeKnownBits(Op, KnownZero, KnownOne);
5218 if (N->getOpcode() != ISD::SETCC || N->getValueType(0) != MVT::i1 ||
5219 cast<CondCodeSDNode>(N->getOperand(2))->get() != ISD::SETNE)
5222 SDValue Op0 = N->getOperand(0);
5223 SDValue Op1 = N->getOperand(1);
5224 assert(Op0.getValueType() == Op1.getValueType());
5226 ConstantSDNode *COp0 = dyn_cast<ConstantSDNode>(Op0);
5227 ConstantSDNode *COp1 = dyn_cast<ConstantSDNode>(Op1);
5228 if (COp0 && COp0->isNullValue())
5230 else if (COp1 && COp1->isNullValue())
5235 DAG.computeKnownBits(Op, KnownZero, KnownOne);
5237 if (!(KnownZero | APInt(Op.getValueSizeInBits(), 1)).isAllOnesValue())
5243 SDValue DAGCombiner::visitZERO_EXTEND(SDNode *N) {
5244 SDValue N0 = N->getOperand(0);
5245 EVT VT = N->getValueType(0);
5247 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
5249 return SDValue(Res, 0);
5251 // fold (zext (zext x)) -> (zext x)
5252 // fold (zext (aext x)) -> (zext x)
5253 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
5254 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT,
5257 // fold (zext (truncate x)) -> (zext x) or
5258 // (zext (truncate x)) -> (truncate x)
5259 // This is valid when the truncated bits of x are already zero.
5260 // FIXME: We should extend this to work for vectors too.
5263 if (!VT.isVector() && isTruncateOf(DAG, N0, Op, KnownZero)) {
5264 APInt TruncatedBits =
5265 (Op.getValueSizeInBits() == N0.getValueSizeInBits()) ?
5266 APInt(Op.getValueSizeInBits(), 0) :
5267 APInt::getBitsSet(Op.getValueSizeInBits(),
5268 N0.getValueSizeInBits(),
5269 std::min(Op.getValueSizeInBits(),
5270 VT.getSizeInBits()));
5271 if (TruncatedBits == (KnownZero & TruncatedBits)) {
5272 if (VT.bitsGT(Op.getValueType()))
5273 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, Op);
5274 if (VT.bitsLT(Op.getValueType()))
5275 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
5281 // fold (zext (truncate (load x))) -> (zext (smaller load x))
5282 // fold (zext (truncate (srl (load x), c))) -> (zext (small load (x+c/n)))
5283 if (N0.getOpcode() == ISD::TRUNCATE) {
5284 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5285 if (NarrowLoad.getNode()) {
5286 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5287 if (NarrowLoad.getNode() != N0.getNode()) {
5288 CombineTo(N0.getNode(), NarrowLoad);
5289 // CombineTo deleted the truncate, if needed, but not what's under it.
5292 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5296 // fold (zext (truncate x)) -> (and x, mask)
5297 if (N0.getOpcode() == ISD::TRUNCATE &&
5298 (!LegalOperations || TLI.isOperationLegal(ISD::AND, VT))) {
5300 // fold (zext (truncate (load x))) -> (zext (smaller load x))
5301 // fold (zext (truncate (srl (load x), c))) -> (zext (smaller load (x+c/n)))
5302 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5303 if (NarrowLoad.getNode()) {
5304 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5305 if (NarrowLoad.getNode() != N0.getNode()) {
5306 CombineTo(N0.getNode(), NarrowLoad);
5307 // CombineTo deleted the truncate, if needed, but not what's under it.
5310 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5313 SDValue Op = N0.getOperand(0);
5314 if (Op.getValueType().bitsLT(VT)) {
5315 Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, Op);
5316 AddToWorklist(Op.getNode());
5317 } else if (Op.getValueType().bitsGT(VT)) {
5318 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
5319 AddToWorklist(Op.getNode());
5321 return DAG.getZeroExtendInReg(Op, SDLoc(N),
5322 N0.getValueType().getScalarType());
5325 // Fold (zext (and (trunc x), cst)) -> (and x, cst),
5326 // if either of the casts is not free.
5327 if (N0.getOpcode() == ISD::AND &&
5328 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
5329 N0.getOperand(1).getOpcode() == ISD::Constant &&
5330 (!TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
5331 N0.getValueType()) ||
5332 !TLI.isZExtFree(N0.getValueType(), VT))) {
5333 SDValue X = N0.getOperand(0).getOperand(0);
5334 if (X.getValueType().bitsLT(VT)) {
5335 X = DAG.getNode(ISD::ANY_EXTEND, SDLoc(X), VT, X);
5336 } else if (X.getValueType().bitsGT(VT)) {
5337 X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X);
5339 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5340 Mask = Mask.zext(VT.getSizeInBits());
5341 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5342 X, DAG.getConstant(Mask, VT));
5345 // fold (zext (load x)) -> (zext (truncate (zextload x)))
5346 // None of the supported targets knows how to perform load and vector_zext
5347 // on vectors in one instruction. We only perform this transformation on
5349 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
5350 ISD::isUNINDEXEDLoad(N0.getNode()) &&
5351 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5352 TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()))) {
5353 bool DoXform = true;
5354 SmallVector<SDNode*, 4> SetCCs;
5355 if (!N0.hasOneUse())
5356 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ZERO_EXTEND, SetCCs, TLI);
5358 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5359 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT,
5361 LN0->getBasePtr(), N0.getValueType(),
5362 LN0->getMemOperand());
5363 CombineTo(N, ExtLoad);
5364 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5365 N0.getValueType(), ExtLoad);
5366 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
5368 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5370 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5374 // fold (zext (and/or/xor (load x), cst)) ->
5375 // (and/or/xor (zextload x), (zext cst))
5376 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
5377 N0.getOpcode() == ISD::XOR) &&
5378 isa<LoadSDNode>(N0.getOperand(0)) &&
5379 N0.getOperand(1).getOpcode() == ISD::Constant &&
5380 TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()) &&
5381 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
5382 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
5383 if (LN0->getExtensionType() != ISD::SEXTLOAD && LN0->isUnindexed()) {
5384 bool DoXform = true;
5385 SmallVector<SDNode*, 4> SetCCs;
5386 if (!N0.hasOneUse())
5387 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::ZERO_EXTEND,
5390 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), VT,
5391 LN0->getChain(), LN0->getBasePtr(),
5393 LN0->getMemOperand());
5394 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5395 Mask = Mask.zext(VT.getSizeInBits());
5396 SDValue And = DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
5397 ExtLoad, DAG.getConstant(Mask, VT));
5398 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
5399 SDLoc(N0.getOperand(0)),
5400 N0.getOperand(0).getValueType(), ExtLoad);
5402 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
5403 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5405 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5410 // fold (zext (zextload x)) -> (zext (truncate (zextload x)))
5411 // fold (zext ( extload x)) -> (zext (truncate (zextload x)))
5412 if ((ISD::isZEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
5413 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
5414 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5415 EVT MemVT = LN0->getMemoryVT();
5416 if ((!LegalOperations && !LN0->isVolatile()) ||
5417 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT)) {
5418 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT,
5420 LN0->getBasePtr(), MemVT,
5421 LN0->getMemOperand());
5422 CombineTo(N, ExtLoad);
5423 CombineTo(N0.getNode(),
5424 DAG.getNode(ISD::TRUNCATE, SDLoc(N0), N0.getValueType(),
5426 ExtLoad.getValue(1));
5427 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5431 if (N0.getOpcode() == ISD::SETCC) {
5432 if (!LegalOperations && VT.isVector() &&
5433 N0.getValueType().getVectorElementType() == MVT::i1) {
5434 EVT N0VT = N0.getOperand(0).getValueType();
5435 if (getSetCCResultType(N0VT) == N0.getValueType())
5438 // zext(setcc) -> (and (vsetcc), (1, 1, ...) for vectors.
5439 // Only do this before legalize for now.
5440 EVT EltVT = VT.getVectorElementType();
5441 SmallVector<SDValue,8> OneOps(VT.getVectorNumElements(),
5442 DAG.getConstant(1, EltVT));
5443 if (VT.getSizeInBits() == N0VT.getSizeInBits())
5444 // We know that the # elements of the results is the same as the
5445 // # elements of the compare (and the # elements of the compare result
5446 // for that matter). Check to see that they are the same size. If so,
5447 // we know that the element size of the sext'd result matches the
5448 // element size of the compare operands.
5449 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5450 DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
5452 cast<CondCodeSDNode>(N0.getOperand(2))->get()),
5453 DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT,
5456 // If the desired elements are smaller or larger than the source
5457 // elements we can use a matching integer vector type and then
5458 // truncate/sign extend
5459 EVT MatchingElementType =
5460 EVT::getIntegerVT(*DAG.getContext(),
5461 N0VT.getScalarType().getSizeInBits());
5462 EVT MatchingVectorType =
5463 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
5464 N0VT.getVectorNumElements());
5466 DAG.getSetCC(SDLoc(N), MatchingVectorType, N0.getOperand(0),
5468 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5469 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5470 DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT),
5471 DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, OneOps));
5474 // zext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
5476 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
5477 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
5478 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
5479 if (SCC.getNode()) return SCC;
5482 // (zext (shl (zext x), cst)) -> (shl (zext x), cst)
5483 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL) &&
5484 isa<ConstantSDNode>(N0.getOperand(1)) &&
5485 N0.getOperand(0).getOpcode() == ISD::ZERO_EXTEND &&
5487 SDValue ShAmt = N0.getOperand(1);
5488 unsigned ShAmtVal = cast<ConstantSDNode>(ShAmt)->getZExtValue();
5489 if (N0.getOpcode() == ISD::SHL) {
5490 SDValue InnerZExt = N0.getOperand(0);
5491 // If the original shl may be shifting out bits, do not perform this
5493 unsigned KnownZeroBits = InnerZExt.getValueType().getSizeInBits() -
5494 InnerZExt.getOperand(0).getValueType().getSizeInBits();
5495 if (ShAmtVal > KnownZeroBits)
5501 // Ensure that the shift amount is wide enough for the shifted value.
5502 if (VT.getSizeInBits() >= 256)
5503 ShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, ShAmt);
5505 return DAG.getNode(N0.getOpcode(), DL, VT,
5506 DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0)),
5513 SDValue DAGCombiner::visitANY_EXTEND(SDNode *N) {
5514 SDValue N0 = N->getOperand(0);
5515 EVT VT = N->getValueType(0);
5517 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
5519 return SDValue(Res, 0);
5521 // fold (aext (aext x)) -> (aext x)
5522 // fold (aext (zext x)) -> (zext x)
5523 // fold (aext (sext x)) -> (sext x)
5524 if (N0.getOpcode() == ISD::ANY_EXTEND ||
5525 N0.getOpcode() == ISD::ZERO_EXTEND ||
5526 N0.getOpcode() == ISD::SIGN_EXTEND)
5527 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, N0.getOperand(0));
5529 // fold (aext (truncate (load x))) -> (aext (smaller load x))
5530 // fold (aext (truncate (srl (load x), c))) -> (aext (small load (x+c/n)))
5531 if (N0.getOpcode() == ISD::TRUNCATE) {
5532 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5533 if (NarrowLoad.getNode()) {
5534 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5535 if (NarrowLoad.getNode() != N0.getNode()) {
5536 CombineTo(N0.getNode(), NarrowLoad);
5537 // CombineTo deleted the truncate, if needed, but not what's under it.
5540 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5544 // fold (aext (truncate x))
5545 if (N0.getOpcode() == ISD::TRUNCATE) {
5546 SDValue TruncOp = N0.getOperand(0);
5547 if (TruncOp.getValueType() == VT)
5548 return TruncOp; // x iff x size == zext size.
5549 if (TruncOp.getValueType().bitsGT(VT))
5550 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, TruncOp);
5551 return DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, TruncOp);
5554 // Fold (aext (and (trunc x), cst)) -> (and x, cst)
5555 // if the trunc is not free.
5556 if (N0.getOpcode() == ISD::AND &&
5557 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
5558 N0.getOperand(1).getOpcode() == ISD::Constant &&
5559 !TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
5560 N0.getValueType())) {
5561 SDValue X = N0.getOperand(0).getOperand(0);
5562 if (X.getValueType().bitsLT(VT)) {
5563 X = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, X);
5564 } else if (X.getValueType().bitsGT(VT)) {
5565 X = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, X);
5567 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5568 Mask = Mask.zext(VT.getSizeInBits());
5569 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5570 X, DAG.getConstant(Mask, VT));
5573 // fold (aext (load x)) -> (aext (truncate (extload x)))
5574 // None of the supported targets knows how to perform load and any_ext
5575 // on vectors in one instruction. We only perform this transformation on
5577 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
5578 ISD::isUNINDEXEDLoad(N0.getNode()) &&
5579 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType())) {
5580 bool DoXform = true;
5581 SmallVector<SDNode*, 4> SetCCs;
5582 if (!N0.hasOneUse())
5583 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ANY_EXTEND, SetCCs, TLI);
5585 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5586 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT,
5588 LN0->getBasePtr(), N0.getValueType(),
5589 LN0->getMemOperand());
5590 CombineTo(N, ExtLoad);
5591 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5592 N0.getValueType(), ExtLoad);
5593 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
5594 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5596 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5600 // fold (aext (zextload x)) -> (aext (truncate (zextload x)))
5601 // fold (aext (sextload x)) -> (aext (truncate (sextload x)))
5602 // fold (aext ( extload x)) -> (aext (truncate (extload x)))
5603 if (N0.getOpcode() == ISD::LOAD &&
5604 !ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
5606 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5607 ISD::LoadExtType ExtType = LN0->getExtensionType();
5608 EVT MemVT = LN0->getMemoryVT();
5609 if (!LegalOperations || TLI.isLoadExtLegal(ExtType, MemVT)) {
5610 SDValue ExtLoad = DAG.getExtLoad(ExtType, SDLoc(N),
5611 VT, LN0->getChain(), LN0->getBasePtr(),
5612 MemVT, LN0->getMemOperand());
5613 CombineTo(N, ExtLoad);
5614 CombineTo(N0.getNode(),
5615 DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5616 N0.getValueType(), ExtLoad),
5617 ExtLoad.getValue(1));
5618 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5622 if (N0.getOpcode() == ISD::SETCC) {
5624 // aext(setcc) -> vsetcc
5625 // aext(setcc) -> truncate(vsetcc)
5626 // aext(setcc) -> aext(vsetcc)
5627 // Only do this before legalize for now.
5628 if (VT.isVector() && !LegalOperations) {
5629 EVT N0VT = N0.getOperand(0).getValueType();
5630 // We know that the # elements of the results is the same as the
5631 // # elements of the compare (and the # elements of the compare result
5632 // for that matter). Check to see that they are the same size. If so,
5633 // we know that the element size of the sext'd result matches the
5634 // element size of the compare operands.
5635 if (VT.getSizeInBits() == N0VT.getSizeInBits())
5636 return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
5638 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5639 // If the desired elements are smaller or larger than the source
5640 // elements we can use a matching integer vector type and then
5641 // truncate/any extend
5643 EVT MatchingVectorType = N0VT.changeVectorElementTypeToInteger();
5645 DAG.getSetCC(SDLoc(N), MatchingVectorType, N0.getOperand(0),
5647 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5648 return DAG.getAnyExtOrTrunc(VsetCC, SDLoc(N), VT);
5652 // aext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
5654 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
5655 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
5656 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
5664 /// GetDemandedBits - See if the specified operand can be simplified with the
5665 /// knowledge that only the bits specified by Mask are used. If so, return the
5666 /// simpler operand, otherwise return a null SDValue.
5667 SDValue DAGCombiner::GetDemandedBits(SDValue V, const APInt &Mask) {
5668 switch (V.getOpcode()) {
5670 case ISD::Constant: {
5671 const ConstantSDNode *CV = cast<ConstantSDNode>(V.getNode());
5672 assert(CV && "Const value should be ConstSDNode.");
5673 const APInt &CVal = CV->getAPIntValue();
5674 APInt NewVal = CVal & Mask;
5676 return DAG.getConstant(NewVal, V.getValueType());
5681 // If the LHS or RHS don't contribute bits to the or, drop them.
5682 if (DAG.MaskedValueIsZero(V.getOperand(0), Mask))
5683 return V.getOperand(1);
5684 if (DAG.MaskedValueIsZero(V.getOperand(1), Mask))
5685 return V.getOperand(0);
5688 // Only look at single-use SRLs.
5689 if (!V.getNode()->hasOneUse())
5691 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(V.getOperand(1))) {
5692 // See if we can recursively simplify the LHS.
5693 unsigned Amt = RHSC->getZExtValue();
5695 // Watch out for shift count overflow though.
5696 if (Amt >= Mask.getBitWidth()) break;
5697 APInt NewMask = Mask << Amt;
5698 SDValue SimplifyLHS = GetDemandedBits(V.getOperand(0), NewMask);
5699 if (SimplifyLHS.getNode())
5700 return DAG.getNode(ISD::SRL, SDLoc(V), V.getValueType(),
5701 SimplifyLHS, V.getOperand(1));
5707 /// ReduceLoadWidth - If the result of a wider load is shifted to right of N
5708 /// bits and then truncated to a narrower type and where N is a multiple
5709 /// of number of bits of the narrower type, transform it to a narrower load
5710 /// from address + N / num of bits of new type. If the result is to be
5711 /// extended, also fold the extension to form a extending load.
5712 SDValue DAGCombiner::ReduceLoadWidth(SDNode *N) {
5713 unsigned Opc = N->getOpcode();
5715 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
5716 SDValue N0 = N->getOperand(0);
5717 EVT VT = N->getValueType(0);
5720 // This transformation isn't valid for vector loads.
5724 // Special case: SIGN_EXTEND_INREG is basically truncating to ExtVT then
5726 if (Opc == ISD::SIGN_EXTEND_INREG) {
5727 ExtType = ISD::SEXTLOAD;
5728 ExtVT = cast<VTSDNode>(N->getOperand(1))->getVT();
5729 } else if (Opc == ISD::SRL) {
5730 // Another special-case: SRL is basically zero-extending a narrower value.
5731 ExtType = ISD::ZEXTLOAD;
5733 ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1));
5734 if (!N01) return SDValue();
5735 ExtVT = EVT::getIntegerVT(*DAG.getContext(),
5736 VT.getSizeInBits() - N01->getZExtValue());
5738 if (LegalOperations && !TLI.isLoadExtLegal(ExtType, ExtVT))
5741 unsigned EVTBits = ExtVT.getSizeInBits();
5743 // Do not generate loads of non-round integer types since these can
5744 // be expensive (and would be wrong if the type is not byte sized).
5745 if (!ExtVT.isRound())
5749 if (N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
5750 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
5751 ShAmt = N01->getZExtValue();
5752 // Is the shift amount a multiple of size of VT?
5753 if ((ShAmt & (EVTBits-1)) == 0) {
5754 N0 = N0.getOperand(0);
5755 // Is the load width a multiple of size of VT?
5756 if ((N0.getValueType().getSizeInBits() & (EVTBits-1)) != 0)
5760 // At this point, we must have a load or else we can't do the transform.
5761 if (!isa<LoadSDNode>(N0)) return SDValue();
5763 // Because a SRL must be assumed to *need* to zero-extend the high bits
5764 // (as opposed to anyext the high bits), we can't combine the zextload
5765 // lowering of SRL and an sextload.
5766 if (cast<LoadSDNode>(N0)->getExtensionType() == ISD::SEXTLOAD)
5769 // If the shift amount is larger than the input type then we're not
5770 // accessing any of the loaded bytes. If the load was a zextload/extload
5771 // then the result of the shift+trunc is zero/undef (handled elsewhere).
5772 if (ShAmt >= cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits())
5777 // If the load is shifted left (and the result isn't shifted back right),
5778 // we can fold the truncate through the shift.
5779 unsigned ShLeftAmt = 0;
5780 if (ShAmt == 0 && N0.getOpcode() == ISD::SHL && N0.hasOneUse() &&
5781 ExtVT == VT && TLI.isNarrowingProfitable(N0.getValueType(), VT)) {
5782 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
5783 ShLeftAmt = N01->getZExtValue();
5784 N0 = N0.getOperand(0);
5788 // If we haven't found a load, we can't narrow it. Don't transform one with
5789 // multiple uses, this would require adding a new load.
5790 if (!isa<LoadSDNode>(N0) || !N0.hasOneUse())
5793 // Don't change the width of a volatile load.
5794 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5795 if (LN0->isVolatile())
5798 // Verify that we are actually reducing a load width here.
5799 if (LN0->getMemoryVT().getSizeInBits() < EVTBits)
5802 // For the transform to be legal, the load must produce only two values
5803 // (the value loaded and the chain). Don't transform a pre-increment
5804 // load, for example, which produces an extra value. Otherwise the
5805 // transformation is not equivalent, and the downstream logic to replace
5806 // uses gets things wrong.
5807 if (LN0->getNumValues() > 2)
5810 // If the load that we're shrinking is an extload and we're not just
5811 // discarding the extension we can't simply shrink the load. Bail.
5812 // TODO: It would be possible to merge the extensions in some cases.
5813 if (LN0->getExtensionType() != ISD::NON_EXTLOAD &&
5814 LN0->getMemoryVT().getSizeInBits() < ExtVT.getSizeInBits() + ShAmt)
5817 EVT PtrType = N0.getOperand(1).getValueType();
5819 if (PtrType == MVT::Untyped || PtrType.isExtended())
5820 // It's not possible to generate a constant of extended or untyped type.
5823 // For big endian targets, we need to adjust the offset to the pointer to
5824 // load the correct bytes.
5825 if (TLI.isBigEndian()) {
5826 unsigned LVTStoreBits = LN0->getMemoryVT().getStoreSizeInBits();
5827 unsigned EVTStoreBits = ExtVT.getStoreSizeInBits();
5828 ShAmt = LVTStoreBits - EVTStoreBits - ShAmt;
5831 uint64_t PtrOff = ShAmt / 8;
5832 unsigned NewAlign = MinAlign(LN0->getAlignment(), PtrOff);
5833 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(LN0),
5834 PtrType, LN0->getBasePtr(),
5835 DAG.getConstant(PtrOff, PtrType));
5836 AddToWorklist(NewPtr.getNode());
5839 if (ExtType == ISD::NON_EXTLOAD)
5840 Load = DAG.getLoad(VT, SDLoc(N0), LN0->getChain(), NewPtr,
5841 LN0->getPointerInfo().getWithOffset(PtrOff),
5842 LN0->isVolatile(), LN0->isNonTemporal(),
5843 LN0->isInvariant(), NewAlign, LN0->getAAInfo());
5845 Load = DAG.getExtLoad(ExtType, SDLoc(N0), VT, LN0->getChain(),NewPtr,
5846 LN0->getPointerInfo().getWithOffset(PtrOff),
5847 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
5848 NewAlign, LN0->getAAInfo());
5850 // Replace the old load's chain with the new load's chain.
5851 WorklistRemover DeadNodes(*this);
5852 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1));
5854 // Shift the result left, if we've swallowed a left shift.
5855 SDValue Result = Load;
5856 if (ShLeftAmt != 0) {
5857 EVT ShImmTy = getShiftAmountTy(Result.getValueType());
5858 if (!isUIntN(ShImmTy.getSizeInBits(), ShLeftAmt))
5860 // If the shift amount is as large as the result size (but, presumably,
5861 // no larger than the source) then the useful bits of the result are
5862 // zero; we can't simply return the shortened shift, because the result
5863 // of that operation is undefined.
5864 if (ShLeftAmt >= VT.getSizeInBits())
5865 Result = DAG.getConstant(0, VT);
5867 Result = DAG.getNode(ISD::SHL, SDLoc(N0), VT,
5868 Result, DAG.getConstant(ShLeftAmt, ShImmTy));
5871 // Return the new loaded value.
5875 SDValue DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
5876 SDValue N0 = N->getOperand(0);
5877 SDValue N1 = N->getOperand(1);
5878 EVT VT = N->getValueType(0);
5879 EVT EVT = cast<VTSDNode>(N1)->getVT();
5880 unsigned VTBits = VT.getScalarType().getSizeInBits();
5881 unsigned EVTBits = EVT.getScalarType().getSizeInBits();
5883 // fold (sext_in_reg c1) -> c1
5884 if (isa<ConstantSDNode>(N0) || N0.getOpcode() == ISD::UNDEF)
5885 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, N0, N1);
5887 // If the input is already sign extended, just drop the extension.
5888 if (DAG.ComputeNumSignBits(N0) >= VTBits-EVTBits+1)
5891 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2
5892 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
5893 EVT.bitsLT(cast<VTSDNode>(N0.getOperand(1))->getVT()))
5894 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
5895 N0.getOperand(0), N1);
5897 // fold (sext_in_reg (sext x)) -> (sext x)
5898 // fold (sext_in_reg (aext x)) -> (sext x)
5899 // if x is small enough.
5900 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) {
5901 SDValue N00 = N0.getOperand(0);
5902 if (N00.getValueType().getScalarType().getSizeInBits() <= EVTBits &&
5903 (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND, VT)))
5904 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, N00, N1);
5907 // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is known zero.
5908 if (DAG.MaskedValueIsZero(N0, APInt::getBitsSet(VTBits, EVTBits-1, EVTBits)))
5909 return DAG.getZeroExtendInReg(N0, SDLoc(N), EVT);
5911 // fold operands of sext_in_reg based on knowledge that the top bits are not
5913 if (SimplifyDemandedBits(SDValue(N, 0)))
5914 return SDValue(N, 0);
5916 // fold (sext_in_reg (load x)) -> (smaller sextload x)
5917 // fold (sext_in_reg (srl (load x), c)) -> (smaller sextload (x+c/evtbits))
5918 SDValue NarrowLoad = ReduceLoadWidth(N);
5919 if (NarrowLoad.getNode())
5922 // fold (sext_in_reg (srl X, 24), i8) -> (sra X, 24)
5923 // fold (sext_in_reg (srl X, 23), i8) -> (sra X, 23) iff possible.
5924 // We already fold "(sext_in_reg (srl X, 25), i8) -> srl X, 25" above.
5925 if (N0.getOpcode() == ISD::SRL) {
5926 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
5927 if (ShAmt->getZExtValue()+EVTBits <= VTBits) {
5928 // We can turn this into an SRA iff the input to the SRL is already sign
5930 unsigned InSignBits = DAG.ComputeNumSignBits(N0.getOperand(0));
5931 if (VTBits-(ShAmt->getZExtValue()+EVTBits) < InSignBits)
5932 return DAG.getNode(ISD::SRA, SDLoc(N), VT,
5933 N0.getOperand(0), N0.getOperand(1));
5937 // fold (sext_inreg (extload x)) -> (sextload x)
5938 if (ISD::isEXTLoad(N0.getNode()) &&
5939 ISD::isUNINDEXEDLoad(N0.getNode()) &&
5940 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
5941 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5942 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) {
5943 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5944 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5946 LN0->getBasePtr(), EVT,
5947 LN0->getMemOperand());
5948 CombineTo(N, ExtLoad);
5949 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
5950 AddToWorklist(ExtLoad.getNode());
5951 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5953 // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use
5954 if (ISD::isZEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
5956 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
5957 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5958 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) {
5959 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5960 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5962 LN0->getBasePtr(), EVT,
5963 LN0->getMemOperand());
5964 CombineTo(N, ExtLoad);
5965 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
5966 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5969 // Form (sext_inreg (bswap >> 16)) or (sext_inreg (rotl (bswap) 16))
5970 if (EVTBits <= 16 && N0.getOpcode() == ISD::OR) {
5971 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
5972 N0.getOperand(1), false);
5973 if (BSwap.getNode())
5974 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
5978 // Fold a sext_inreg of a build_vector of ConstantSDNodes or undefs
5979 // into a build_vector.
5980 if (ISD::isBuildVectorOfConstantSDNodes(N0.getNode())) {
5981 SmallVector<SDValue, 8> Elts;
5982 unsigned NumElts = N0->getNumOperands();
5983 unsigned ShAmt = VTBits - EVTBits;
5985 for (unsigned i = 0; i != NumElts; ++i) {
5986 SDValue Op = N0->getOperand(i);
5987 if (Op->getOpcode() == ISD::UNDEF) {
5992 ConstantSDNode *CurrentND = cast<ConstantSDNode>(Op);
5993 const APInt &C = APInt(VTBits, CurrentND->getAPIntValue().getZExtValue());
5994 Elts.push_back(DAG.getConstant(C.shl(ShAmt).ashr(ShAmt).getZExtValue(),
5995 Op.getValueType()));
5998 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Elts);
6004 SDValue DAGCombiner::visitTRUNCATE(SDNode *N) {
6005 SDValue N0 = N->getOperand(0);
6006 EVT VT = N->getValueType(0);
6007 bool isLE = TLI.isLittleEndian();
6010 if (N0.getValueType() == N->getValueType(0))
6012 // fold (truncate c1) -> c1
6013 if (isa<ConstantSDNode>(N0))
6014 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0);
6015 // fold (truncate (truncate x)) -> (truncate x)
6016 if (N0.getOpcode() == ISD::TRUNCATE)
6017 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0));
6018 // fold (truncate (ext x)) -> (ext x) or (truncate x) or x
6019 if (N0.getOpcode() == ISD::ZERO_EXTEND ||
6020 N0.getOpcode() == ISD::SIGN_EXTEND ||
6021 N0.getOpcode() == ISD::ANY_EXTEND) {
6022 if (N0.getOperand(0).getValueType().bitsLT(VT))
6023 // if the source is smaller than the dest, we still need an extend
6024 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
6026 if (N0.getOperand(0).getValueType().bitsGT(VT))
6027 // if the source is larger than the dest, than we just need the truncate
6028 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0));
6029 // if the source and dest are the same type, we can drop both the extend
6030 // and the truncate.
6031 return N0.getOperand(0);
6034 // Fold extract-and-trunc into a narrow extract. For example:
6035 // i64 x = EXTRACT_VECTOR_ELT(v2i64 val, i32 1)
6036 // i32 y = TRUNCATE(i64 x)
6038 // v16i8 b = BITCAST (v2i64 val)
6039 // i8 x = EXTRACT_VECTOR_ELT(v16i8 b, i32 8)
6041 // Note: We only run this optimization after type legalization (which often
6042 // creates this pattern) and before operation legalization after which
6043 // we need to be more careful about the vector instructions that we generate.
6044 if (N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
6045 LegalTypes && !LegalOperations && N0->hasOneUse() && VT != MVT::i1) {
6047 EVT VecTy = N0.getOperand(0).getValueType();
6048 EVT ExTy = N0.getValueType();
6049 EVT TrTy = N->getValueType(0);
6051 unsigned NumElem = VecTy.getVectorNumElements();
6052 unsigned SizeRatio = ExTy.getSizeInBits()/TrTy.getSizeInBits();
6054 EVT NVT = EVT::getVectorVT(*DAG.getContext(), TrTy, SizeRatio * NumElem);
6055 assert(NVT.getSizeInBits() == VecTy.getSizeInBits() && "Invalid Size");
6057 SDValue EltNo = N0->getOperand(1);
6058 if (isa<ConstantSDNode>(EltNo) && isTypeLegal(NVT)) {
6059 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
6060 EVT IndexTy = TLI.getVectorIdxTy();
6061 int Index = isLE ? (Elt*SizeRatio) : (Elt*SizeRatio + (SizeRatio-1));
6063 SDValue V = DAG.getNode(ISD::BITCAST, SDLoc(N),
6064 NVT, N0.getOperand(0));
6066 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
6068 DAG.getConstant(Index, IndexTy));
6072 // trunc (select c, a, b) -> select c, (trunc a), (trunc b)
6073 if (N0.getOpcode() == ISD::SELECT) {
6074 EVT SrcVT = N0.getValueType();
6075 if ((!LegalOperations || TLI.isOperationLegal(ISD::SELECT, SrcVT)) &&
6076 TLI.isTruncateFree(SrcVT, VT)) {
6078 SDValue Cond = N0.getOperand(0);
6079 SDValue TruncOp0 = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(1));
6080 SDValue TruncOp1 = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(2));
6081 return DAG.getNode(ISD::SELECT, SDLoc(N), VT, Cond, TruncOp0, TruncOp1);
6085 // Fold a series of buildvector, bitcast, and truncate if possible.
6087 // (2xi32 trunc (bitcast ((4xi32)buildvector x, x, y, y) 2xi64)) to
6088 // (2xi32 (buildvector x, y)).
6089 if (Level == AfterLegalizeVectorOps && VT.isVector() &&
6090 N0.getOpcode() == ISD::BITCAST && N0.hasOneUse() &&
6091 N0.getOperand(0).getOpcode() == ISD::BUILD_VECTOR &&
6092 N0.getOperand(0).hasOneUse()) {
6094 SDValue BuildVect = N0.getOperand(0);
6095 EVT BuildVectEltTy = BuildVect.getValueType().getVectorElementType();
6096 EVT TruncVecEltTy = VT.getVectorElementType();
6098 // Check that the element types match.
6099 if (BuildVectEltTy == TruncVecEltTy) {
6100 // Now we only need to compute the offset of the truncated elements.
6101 unsigned BuildVecNumElts = BuildVect.getNumOperands();
6102 unsigned TruncVecNumElts = VT.getVectorNumElements();
6103 unsigned TruncEltOffset = BuildVecNumElts / TruncVecNumElts;
6105 assert((BuildVecNumElts % TruncVecNumElts) == 0 &&
6106 "Invalid number of elements");
6108 SmallVector<SDValue, 8> Opnds;
6109 for (unsigned i = 0, e = BuildVecNumElts; i != e; i += TruncEltOffset)
6110 Opnds.push_back(BuildVect.getOperand(i));
6112 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Opnds);
6116 // See if we can simplify the input to this truncate through knowledge that
6117 // only the low bits are being used.
6118 // For example "trunc (or (shl x, 8), y)" // -> trunc y
6119 // Currently we only perform this optimization on scalars because vectors
6120 // may have different active low bits.
6121 if (!VT.isVector()) {
6123 GetDemandedBits(N0, APInt::getLowBitsSet(N0.getValueSizeInBits(),
6124 VT.getSizeInBits()));
6125 if (Shorter.getNode())
6126 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Shorter);
6128 // fold (truncate (load x)) -> (smaller load x)
6129 // fold (truncate (srl (load x), c)) -> (smaller load (x+c/evtbits))
6130 if (!LegalTypes || TLI.isTypeDesirableForOp(N0.getOpcode(), VT)) {
6131 SDValue Reduced = ReduceLoadWidth(N);
6132 if (Reduced.getNode())
6134 // Handle the case where the load remains an extending load even
6135 // after truncation.
6136 if (N0.hasOneUse() && ISD::isUNINDEXEDLoad(N0.getNode())) {
6137 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6138 if (!LN0->isVolatile() &&
6139 LN0->getMemoryVT().getStoreSizeInBits() < VT.getSizeInBits()) {
6140 SDValue NewLoad = DAG.getExtLoad(LN0->getExtensionType(), SDLoc(LN0),
6141 VT, LN0->getChain(), LN0->getBasePtr(),
6143 LN0->getMemOperand());
6144 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLoad.getValue(1));
6149 // fold (trunc (concat ... x ...)) -> (concat ..., (trunc x), ...)),
6150 // where ... are all 'undef'.
6151 if (N0.getOpcode() == ISD::CONCAT_VECTORS && !LegalTypes) {
6152 SmallVector<EVT, 8> VTs;
6155 unsigned NumDefs = 0;
6157 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) {
6158 SDValue X = N0.getOperand(i);
6159 if (X.getOpcode() != ISD::UNDEF) {
6164 // Stop if more than one members are non-undef.
6167 VTs.push_back(EVT::getVectorVT(*DAG.getContext(),
6168 VT.getVectorElementType(),
6169 X.getValueType().getVectorNumElements()));
6173 return DAG.getUNDEF(VT);
6176 assert(V.getNode() && "The single defined operand is empty!");
6177 SmallVector<SDValue, 8> Opnds;
6178 for (unsigned i = 0, e = VTs.size(); i != e; ++i) {
6180 Opnds.push_back(DAG.getUNDEF(VTs[i]));
6183 SDValue NV = DAG.getNode(ISD::TRUNCATE, SDLoc(V), VTs[i], V);
6184 AddToWorklist(NV.getNode());
6185 Opnds.push_back(NV);
6187 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Opnds);
6191 // Simplify the operands using demanded-bits information.
6192 if (!VT.isVector() &&
6193 SimplifyDemandedBits(SDValue(N, 0)))
6194 return SDValue(N, 0);
6199 static SDNode *getBuildPairElt(SDNode *N, unsigned i) {
6200 SDValue Elt = N->getOperand(i);
6201 if (Elt.getOpcode() != ISD::MERGE_VALUES)
6202 return Elt.getNode();
6203 return Elt.getOperand(Elt.getResNo()).getNode();
6206 /// CombineConsecutiveLoads - build_pair (load, load) -> load
6207 /// if load locations are consecutive.
6208 SDValue DAGCombiner::CombineConsecutiveLoads(SDNode *N, EVT VT) {
6209 assert(N->getOpcode() == ISD::BUILD_PAIR);
6211 LoadSDNode *LD1 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 0));
6212 LoadSDNode *LD2 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 1));
6213 if (!LD1 || !LD2 || !ISD::isNON_EXTLoad(LD1) || !LD1->hasOneUse() ||
6214 LD1->getAddressSpace() != LD2->getAddressSpace())
6216 EVT LD1VT = LD1->getValueType(0);
6218 if (ISD::isNON_EXTLoad(LD2) &&
6220 // If both are volatile this would reduce the number of volatile loads.
6221 // If one is volatile it might be ok, but play conservative and bail out.
6222 !LD1->isVolatile() &&
6223 !LD2->isVolatile() &&
6224 DAG.isConsecutiveLoad(LD2, LD1, LD1VT.getSizeInBits()/8, 1)) {
6225 unsigned Align = LD1->getAlignment();
6226 unsigned NewAlign = TLI.getDataLayout()->
6227 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
6229 if (NewAlign <= Align &&
6230 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)))
6231 return DAG.getLoad(VT, SDLoc(N), LD1->getChain(),
6232 LD1->getBasePtr(), LD1->getPointerInfo(),
6233 false, false, false, Align);
6239 SDValue DAGCombiner::visitBITCAST(SDNode *N) {
6240 SDValue N0 = N->getOperand(0);
6241 EVT VT = N->getValueType(0);
6243 // If the input is a BUILD_VECTOR with all constant elements, fold this now.
6244 // Only do this before legalize, since afterward the target may be depending
6245 // on the bitconvert.
6246 // First check to see if this is all constant.
6248 N0.getOpcode() == ISD::BUILD_VECTOR && N0.getNode()->hasOneUse() &&
6250 bool isSimple = cast<BuildVectorSDNode>(N0)->isConstant();
6252 EVT DestEltVT = N->getValueType(0).getVectorElementType();
6253 assert(!DestEltVT.isVector() &&
6254 "Element type of vector ValueType must not be vector!");
6256 return ConstantFoldBITCASTofBUILD_VECTOR(N0.getNode(), DestEltVT);
6259 // If the input is a constant, let getNode fold it.
6260 if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) {
6261 SDValue Res = DAG.getNode(ISD::BITCAST, SDLoc(N), VT, N0);
6262 if (Res.getNode() != N) {
6263 if (!LegalOperations ||
6264 TLI.isOperationLegal(Res.getNode()->getOpcode(), VT))
6267 // Folding it resulted in an illegal node, and it's too late to
6268 // do that. Clean up the old node and forego the transformation.
6269 // Ideally this won't happen very often, because instcombine
6270 // and the earlier dagcombine runs (where illegal nodes are
6271 // permitted) should have folded most of them already.
6272 DAG.DeleteNode(Res.getNode());
6276 // (conv (conv x, t1), t2) -> (conv x, t2)
6277 if (N0.getOpcode() == ISD::BITCAST)
6278 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT,
6281 // fold (conv (load x)) -> (load (conv*)x)
6282 // If the resultant load doesn't need a higher alignment than the original!
6283 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
6284 // Do not change the width of a volatile load.
6285 !cast<LoadSDNode>(N0)->isVolatile() &&
6286 // Do not remove the cast if the types differ in endian layout.
6287 TLI.hasBigEndianPartOrdering(N0.getValueType()) ==
6288 TLI.hasBigEndianPartOrdering(VT) &&
6289 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)) &&
6290 TLI.isLoadBitCastBeneficial(N0.getValueType(), VT)) {
6291 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6292 unsigned Align = TLI.getDataLayout()->
6293 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
6294 unsigned OrigAlign = LN0->getAlignment();
6296 if (Align <= OrigAlign) {
6297 SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(),
6298 LN0->getBasePtr(), LN0->getPointerInfo(),
6299 LN0->isVolatile(), LN0->isNonTemporal(),
6300 LN0->isInvariant(), OrigAlign,
6303 CombineTo(N0.getNode(),
6304 DAG.getNode(ISD::BITCAST, SDLoc(N0),
6305 N0.getValueType(), Load),
6311 // fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit)
6312 // fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit))
6313 // This often reduces constant pool loads.
6314 if (((N0.getOpcode() == ISD::FNEG && !TLI.isFNegFree(N0.getValueType())) ||
6315 (N0.getOpcode() == ISD::FABS && !TLI.isFAbsFree(N0.getValueType()))) &&
6316 N0.getNode()->hasOneUse() && VT.isInteger() &&
6317 !VT.isVector() && !N0.getValueType().isVector()) {
6318 SDValue NewConv = DAG.getNode(ISD::BITCAST, SDLoc(N0), VT,
6320 AddToWorklist(NewConv.getNode());
6322 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
6323 if (N0.getOpcode() == ISD::FNEG)
6324 return DAG.getNode(ISD::XOR, SDLoc(N), VT,
6325 NewConv, DAG.getConstant(SignBit, VT));
6326 assert(N0.getOpcode() == ISD::FABS);
6327 return DAG.getNode(ISD::AND, SDLoc(N), VT,
6328 NewConv, DAG.getConstant(~SignBit, VT));
6331 // fold (bitconvert (fcopysign cst, x)) ->
6332 // (or (and (bitconvert x), sign), (and cst, (not sign)))
6333 // Note that we don't handle (copysign x, cst) because this can always be
6334 // folded to an fneg or fabs.
6335 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse() &&
6336 isa<ConstantFPSDNode>(N0.getOperand(0)) &&
6337 VT.isInteger() && !VT.isVector()) {
6338 unsigned OrigXWidth = N0.getOperand(1).getValueType().getSizeInBits();
6339 EVT IntXVT = EVT::getIntegerVT(*DAG.getContext(), OrigXWidth);
6340 if (isTypeLegal(IntXVT)) {
6341 SDValue X = DAG.getNode(ISD::BITCAST, SDLoc(N0),
6342 IntXVT, N0.getOperand(1));
6343 AddToWorklist(X.getNode());
6345 // If X has a different width than the result/lhs, sext it or truncate it.
6346 unsigned VTWidth = VT.getSizeInBits();
6347 if (OrigXWidth < VTWidth) {
6348 X = DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, X);
6349 AddToWorklist(X.getNode());
6350 } else if (OrigXWidth > VTWidth) {
6351 // To get the sign bit in the right place, we have to shift it right
6352 // before truncating.
6353 X = DAG.getNode(ISD::SRL, SDLoc(X),
6354 X.getValueType(), X,
6355 DAG.getConstant(OrigXWidth-VTWidth, X.getValueType()));
6356 AddToWorklist(X.getNode());
6357 X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X);
6358 AddToWorklist(X.getNode());
6361 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
6362 X = DAG.getNode(ISD::AND, SDLoc(X), VT,
6363 X, DAG.getConstant(SignBit, VT));
6364 AddToWorklist(X.getNode());
6366 SDValue Cst = DAG.getNode(ISD::BITCAST, SDLoc(N0),
6367 VT, N0.getOperand(0));
6368 Cst = DAG.getNode(ISD::AND, SDLoc(Cst), VT,
6369 Cst, DAG.getConstant(~SignBit, VT));
6370 AddToWorklist(Cst.getNode());
6372 return DAG.getNode(ISD::OR, SDLoc(N), VT, X, Cst);
6376 // bitconvert(build_pair(ld, ld)) -> ld iff load locations are consecutive.
6377 if (N0.getOpcode() == ISD::BUILD_PAIR) {
6378 SDValue CombineLD = CombineConsecutiveLoads(N0.getNode(), VT);
6379 if (CombineLD.getNode())
6386 SDValue DAGCombiner::visitBUILD_PAIR(SDNode *N) {
6387 EVT VT = N->getValueType(0);
6388 return CombineConsecutiveLoads(N, VT);
6391 /// ConstantFoldBITCASTofBUILD_VECTOR - We know that BV is a build_vector
6392 /// node with Constant, ConstantFP or Undef operands. DstEltVT indicates the
6393 /// destination element value type.
6394 SDValue DAGCombiner::
6395 ConstantFoldBITCASTofBUILD_VECTOR(SDNode *BV, EVT DstEltVT) {
6396 EVT SrcEltVT = BV->getValueType(0).getVectorElementType();
6398 // If this is already the right type, we're done.
6399 if (SrcEltVT == DstEltVT) return SDValue(BV, 0);
6401 unsigned SrcBitSize = SrcEltVT.getSizeInBits();
6402 unsigned DstBitSize = DstEltVT.getSizeInBits();
6404 // If this is a conversion of N elements of one type to N elements of another
6405 // type, convert each element. This handles FP<->INT cases.
6406 if (SrcBitSize == DstBitSize) {
6407 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
6408 BV->getValueType(0).getVectorNumElements());
6410 // Due to the FP element handling below calling this routine recursively,
6411 // we can end up with a scalar-to-vector node here.
6412 if (BV->getOpcode() == ISD::SCALAR_TO_VECTOR)
6413 return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(BV), VT,
6414 DAG.getNode(ISD::BITCAST, SDLoc(BV),
6415 DstEltVT, BV->getOperand(0)));
6417 SmallVector<SDValue, 8> Ops;
6418 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
6419 SDValue Op = BV->getOperand(i);
6420 // If the vector element type is not legal, the BUILD_VECTOR operands
6421 // are promoted and implicitly truncated. Make that explicit here.
6422 if (Op.getValueType() != SrcEltVT)
6423 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(BV), SrcEltVT, Op);
6424 Ops.push_back(DAG.getNode(ISD::BITCAST, SDLoc(BV),
6426 AddToWorklist(Ops.back().getNode());
6428 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
6431 // Otherwise, we're growing or shrinking the elements. To avoid having to
6432 // handle annoying details of growing/shrinking FP values, we convert them to
6434 if (SrcEltVT.isFloatingPoint()) {
6435 // Convert the input float vector to a int vector where the elements are the
6437 assert((SrcEltVT == MVT::f32 || SrcEltVT == MVT::f64) && "Unknown FP VT!");
6438 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), SrcEltVT.getSizeInBits());
6439 BV = ConstantFoldBITCASTofBUILD_VECTOR(BV, IntVT).getNode();
6443 // Now we know the input is an integer vector. If the output is a FP type,
6444 // convert to integer first, then to FP of the right size.
6445 if (DstEltVT.isFloatingPoint()) {
6446 assert((DstEltVT == MVT::f32 || DstEltVT == MVT::f64) && "Unknown FP VT!");
6447 EVT TmpVT = EVT::getIntegerVT(*DAG.getContext(), DstEltVT.getSizeInBits());
6448 SDNode *Tmp = ConstantFoldBITCASTofBUILD_VECTOR(BV, TmpVT).getNode();
6450 // Next, convert to FP elements of the same size.
6451 return ConstantFoldBITCASTofBUILD_VECTOR(Tmp, DstEltVT);
6454 // Okay, we know the src/dst types are both integers of differing types.
6455 // Handling growing first.
6456 assert(SrcEltVT.isInteger() && DstEltVT.isInteger());
6457 if (SrcBitSize < DstBitSize) {
6458 unsigned NumInputsPerOutput = DstBitSize/SrcBitSize;
6460 SmallVector<SDValue, 8> Ops;
6461 for (unsigned i = 0, e = BV->getNumOperands(); i != e;
6462 i += NumInputsPerOutput) {
6463 bool isLE = TLI.isLittleEndian();
6464 APInt NewBits = APInt(DstBitSize, 0);
6465 bool EltIsUndef = true;
6466 for (unsigned j = 0; j != NumInputsPerOutput; ++j) {
6467 // Shift the previously computed bits over.
6468 NewBits <<= SrcBitSize;
6469 SDValue Op = BV->getOperand(i+ (isLE ? (NumInputsPerOutput-j-1) : j));
6470 if (Op.getOpcode() == ISD::UNDEF) continue;
6473 NewBits |= cast<ConstantSDNode>(Op)->getAPIntValue().
6474 zextOrTrunc(SrcBitSize).zext(DstBitSize);
6478 Ops.push_back(DAG.getUNDEF(DstEltVT));
6480 Ops.push_back(DAG.getConstant(NewBits, DstEltVT));
6483 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, Ops.size());
6484 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
6487 // Finally, this must be the case where we are shrinking elements: each input
6488 // turns into multiple outputs.
6489 bool isS2V = ISD::isScalarToVector(BV);
6490 unsigned NumOutputsPerInput = SrcBitSize/DstBitSize;
6491 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
6492 NumOutputsPerInput*BV->getNumOperands());
6493 SmallVector<SDValue, 8> Ops;
6495 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
6496 if (BV->getOperand(i).getOpcode() == ISD::UNDEF) {
6497 for (unsigned j = 0; j != NumOutputsPerInput; ++j)
6498 Ops.push_back(DAG.getUNDEF(DstEltVT));
6502 APInt OpVal = cast<ConstantSDNode>(BV->getOperand(i))->
6503 getAPIntValue().zextOrTrunc(SrcBitSize);
6505 for (unsigned j = 0; j != NumOutputsPerInput; ++j) {
6506 APInt ThisVal = OpVal.trunc(DstBitSize);
6507 Ops.push_back(DAG.getConstant(ThisVal, DstEltVT));
6508 if (isS2V && i == 0 && j == 0 && ThisVal.zext(SrcBitSize) == OpVal)
6509 // Simply turn this into a SCALAR_TO_VECTOR of the new type.
6510 return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(BV), VT,
6512 OpVal = OpVal.lshr(DstBitSize);
6515 // For big endian targets, swap the order of the pieces of each element.
6516 if (TLI.isBigEndian())
6517 std::reverse(Ops.end()-NumOutputsPerInput, Ops.end());
6520 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
6523 SDValue DAGCombiner::visitFADD(SDNode *N) {
6524 SDValue N0 = N->getOperand(0);
6525 SDValue N1 = N->getOperand(1);
6526 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6527 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6528 EVT VT = N->getValueType(0);
6531 if (VT.isVector()) {
6532 SDValue FoldedVOp = SimplifyVBinOp(N);
6533 if (FoldedVOp.getNode()) return FoldedVOp;
6536 // fold (fadd c1, c2) -> c1 + c2
6538 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N1);
6539 // canonicalize constant to RHS
6540 if (N0CFP && !N1CFP)
6541 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N1, N0);
6542 // fold (fadd A, 0) -> A
6543 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
6544 N1CFP->getValueAPF().isZero())
6546 // fold (fadd A, (fneg B)) -> (fsub A, B)
6547 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
6548 isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options) == 2)
6549 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N0,
6550 GetNegatedExpression(N1, DAG, LegalOperations));
6551 // fold (fadd (fneg A), B) -> (fsub B, A)
6552 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
6553 isNegatibleForFree(N0, LegalOperations, TLI, &DAG.getTarget().Options) == 2)
6554 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N1,
6555 GetNegatedExpression(N0, DAG, LegalOperations));
6557 // If allowed, fold (fadd (fadd x, c1), c2) -> (fadd x, (fadd c1, c2))
6558 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
6559 N0.getOpcode() == ISD::FADD && N0.getNode()->hasOneUse() &&
6560 isa<ConstantFPSDNode>(N0.getOperand(1)))
6561 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0.getOperand(0),
6562 DAG.getNode(ISD::FADD, SDLoc(N), VT,
6563 N0.getOperand(1), N1));
6565 // No FP constant should be created after legalization as Instruction
6566 // Selection pass has hard time in dealing with FP constant.
6568 // We don't need test this condition for transformation like following, as
6569 // the DAG being transformed implies it is legal to take FP constant as
6572 // (fadd (fmul c, x), x) -> (fmul c+1, x)
6574 bool AllowNewFpConst = (Level < AfterLegalizeDAG);
6576 // If allow, fold (fadd (fneg x), x) -> 0.0
6577 if (AllowNewFpConst && DAG.getTarget().Options.UnsafeFPMath &&
6578 N0.getOpcode() == ISD::FNEG && N0.getOperand(0) == N1)
6579 return DAG.getConstantFP(0.0, VT);
6581 // If allow, fold (fadd x, (fneg x)) -> 0.0
6582 if (AllowNewFpConst && DAG.getTarget().Options.UnsafeFPMath &&
6583 N1.getOpcode() == ISD::FNEG && N1.getOperand(0) == N0)
6584 return DAG.getConstantFP(0.0, VT);
6586 // In unsafe math mode, we can fold chains of FADD's of the same value
6587 // into multiplications. This transform is not safe in general because
6588 // we are reducing the number of rounding steps.
6589 if (DAG.getTarget().Options.UnsafeFPMath &&
6590 TLI.isOperationLegalOrCustom(ISD::FMUL, VT) &&
6592 if (N0.getOpcode() == ISD::FMUL) {
6593 ConstantFPSDNode *CFP00 = dyn_cast<ConstantFPSDNode>(N0.getOperand(0));
6594 ConstantFPSDNode *CFP01 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
6596 // (fadd (fmul c, x), x) -> (fmul x, c+1)
6597 if (CFP00 && !CFP01 && N0.getOperand(1) == N1) {
6598 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6600 DAG.getConstantFP(1.0, VT));
6601 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6605 // (fadd (fmul x, c), x) -> (fmul x, c+1)
6606 if (CFP01 && !CFP00 && N0.getOperand(0) == N1) {
6607 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6609 DAG.getConstantFP(1.0, VT));
6610 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6614 // (fadd (fmul c, x), (fadd x, x)) -> (fmul x, c+2)
6615 if (CFP00 && !CFP01 && N1.getOpcode() == ISD::FADD &&
6616 N1.getOperand(0) == N1.getOperand(1) &&
6617 N0.getOperand(1) == N1.getOperand(0)) {
6618 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6620 DAG.getConstantFP(2.0, VT));
6621 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6622 N0.getOperand(1), NewCFP);
6625 // (fadd (fmul x, c), (fadd x, x)) -> (fmul x, c+2)
6626 if (CFP01 && !CFP00 && N1.getOpcode() == ISD::FADD &&
6627 N1.getOperand(0) == N1.getOperand(1) &&
6628 N0.getOperand(0) == N1.getOperand(0)) {
6629 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6631 DAG.getConstantFP(2.0, VT));
6632 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6633 N0.getOperand(0), NewCFP);
6637 if (N1.getOpcode() == ISD::FMUL) {
6638 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0));
6639 ConstantFPSDNode *CFP11 = dyn_cast<ConstantFPSDNode>(N1.getOperand(1));
6641 // (fadd x, (fmul c, x)) -> (fmul x, c+1)
6642 if (CFP10 && !CFP11 && N1.getOperand(1) == N0) {
6643 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6645 DAG.getConstantFP(1.0, VT));
6646 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6650 // (fadd x, (fmul x, c)) -> (fmul x, c+1)
6651 if (CFP11 && !CFP10 && N1.getOperand(0) == N0) {
6652 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6654 DAG.getConstantFP(1.0, VT));
6655 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6660 // (fadd (fadd x, x), (fmul c, x)) -> (fmul x, c+2)
6661 if (CFP10 && !CFP11 && N0.getOpcode() == ISD::FADD &&
6662 N0.getOperand(0) == N0.getOperand(1) &&
6663 N1.getOperand(1) == N0.getOperand(0)) {
6664 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6666 DAG.getConstantFP(2.0, VT));
6667 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6668 N1.getOperand(1), NewCFP);
6671 // (fadd (fadd x, x), (fmul x, c)) -> (fmul x, c+2)
6672 if (CFP11 && !CFP10 && N0.getOpcode() == ISD::FADD &&
6673 N0.getOperand(0) == N0.getOperand(1) &&
6674 N1.getOperand(0) == N0.getOperand(0)) {
6675 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6677 DAG.getConstantFP(2.0, VT));
6678 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6679 N1.getOperand(0), NewCFP);
6683 if (N0.getOpcode() == ISD::FADD && AllowNewFpConst) {
6684 ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N0.getOperand(0));
6685 // (fadd (fadd x, x), x) -> (fmul x, 3.0)
6686 if (!CFP && N0.getOperand(0) == N0.getOperand(1) &&
6687 (N0.getOperand(0) == N1))
6688 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6689 N1, DAG.getConstantFP(3.0, VT));
6692 if (N1.getOpcode() == ISD::FADD && AllowNewFpConst) {
6693 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0));
6694 // (fadd x, (fadd x, x)) -> (fmul x, 3.0)
6695 if (!CFP10 && N1.getOperand(0) == N1.getOperand(1) &&
6696 N1.getOperand(0) == N0)
6697 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6698 N0, DAG.getConstantFP(3.0, VT));
6701 // (fadd (fadd x, x), (fadd x, x)) -> (fmul x, 4.0)
6702 if (AllowNewFpConst &&
6703 N0.getOpcode() == ISD::FADD && N1.getOpcode() == ISD::FADD &&
6704 N0.getOperand(0) == N0.getOperand(1) &&
6705 N1.getOperand(0) == N1.getOperand(1) &&
6706 N0.getOperand(0) == N1.getOperand(0))
6707 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6709 DAG.getConstantFP(4.0, VT));
6712 // FADD -> FMA combines:
6713 if ((DAG.getTarget().Options.AllowFPOpFusion == FPOpFusion::Fast ||
6714 DAG.getTarget().Options.UnsafeFPMath) &&
6715 DAG.getTarget().getTargetLowering()->isFMAFasterThanFMulAndFAdd(VT) &&
6716 (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FMA, VT))) {
6718 // fold (fadd (fmul x, y), z) -> (fma x, y, z)
6719 if (N0.getOpcode() == ISD::FMUL && N0->hasOneUse())
6720 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
6721 N0.getOperand(0), N0.getOperand(1), N1);
6723 // fold (fadd x, (fmul y, z)) -> (fma y, z, x)
6724 // Note: Commutes FADD operands.
6725 if (N1.getOpcode() == ISD::FMUL && N1->hasOneUse())
6726 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
6727 N1.getOperand(0), N1.getOperand(1), N0);
6733 SDValue DAGCombiner::visitFSUB(SDNode *N) {
6734 SDValue N0 = N->getOperand(0);
6735 SDValue N1 = N->getOperand(1);
6736 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6737 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6738 EVT VT = N->getValueType(0);
6742 if (VT.isVector()) {
6743 SDValue FoldedVOp = SimplifyVBinOp(N);
6744 if (FoldedVOp.getNode()) return FoldedVOp;
6747 // fold (fsub c1, c2) -> c1-c2
6749 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N0, N1);
6750 // fold (fsub A, 0) -> A
6751 if (DAG.getTarget().Options.UnsafeFPMath &&
6752 N1CFP && N1CFP->getValueAPF().isZero())
6754 // fold (fsub 0, B) -> -B
6755 if (DAG.getTarget().Options.UnsafeFPMath &&
6756 N0CFP && N0CFP->getValueAPF().isZero()) {
6757 if (isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options))
6758 return GetNegatedExpression(N1, DAG, LegalOperations);
6759 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
6760 return DAG.getNode(ISD::FNEG, dl, VT, N1);
6762 // fold (fsub A, (fneg B)) -> (fadd A, B)
6763 if (isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options))
6764 return DAG.getNode(ISD::FADD, dl, VT, N0,
6765 GetNegatedExpression(N1, DAG, LegalOperations));
6767 // If 'unsafe math' is enabled, fold
6768 // (fsub x, x) -> 0.0 &
6769 // (fsub x, (fadd x, y)) -> (fneg y) &
6770 // (fsub x, (fadd y, x)) -> (fneg y)
6771 if (DAG.getTarget().Options.UnsafeFPMath) {
6773 return DAG.getConstantFP(0.0f, VT);
6775 if (N1.getOpcode() == ISD::FADD) {
6776 SDValue N10 = N1->getOperand(0);
6777 SDValue N11 = N1->getOperand(1);
6779 if (N10 == N0 && isNegatibleForFree(N11, LegalOperations, TLI,
6780 &DAG.getTarget().Options))
6781 return GetNegatedExpression(N11, DAG, LegalOperations);
6783 if (N11 == N0 && isNegatibleForFree(N10, LegalOperations, TLI,
6784 &DAG.getTarget().Options))
6785 return GetNegatedExpression(N10, DAG, LegalOperations);
6789 // FSUB -> FMA combines:
6790 if ((DAG.getTarget().Options.AllowFPOpFusion == FPOpFusion::Fast ||
6791 DAG.getTarget().Options.UnsafeFPMath) &&
6792 DAG.getTarget().getTargetLowering()->isFMAFasterThanFMulAndFAdd(VT) &&
6793 (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FMA, VT))) {
6795 // fold (fsub (fmul x, y), z) -> (fma x, y, (fneg z))
6796 if (N0.getOpcode() == ISD::FMUL && N0->hasOneUse())
6797 return DAG.getNode(ISD::FMA, dl, VT,
6798 N0.getOperand(0), N0.getOperand(1),
6799 DAG.getNode(ISD::FNEG, dl, VT, N1));
6801 // fold (fsub x, (fmul y, z)) -> (fma (fneg y), z, x)
6802 // Note: Commutes FSUB operands.
6803 if (N1.getOpcode() == ISD::FMUL && N1->hasOneUse())
6804 return DAG.getNode(ISD::FMA, dl, VT,
6805 DAG.getNode(ISD::FNEG, dl, VT,
6807 N1.getOperand(1), N0);
6809 // fold (fsub (fneg (fmul, x, y)), z) -> (fma (fneg x), y, (fneg z))
6810 if (N0.getOpcode() == ISD::FNEG &&
6811 N0.getOperand(0).getOpcode() == ISD::FMUL &&
6812 N0->hasOneUse() && N0.getOperand(0).hasOneUse()) {
6813 SDValue N00 = N0.getOperand(0).getOperand(0);
6814 SDValue N01 = N0.getOperand(0).getOperand(1);
6815 return DAG.getNode(ISD::FMA, dl, VT,
6816 DAG.getNode(ISD::FNEG, dl, VT, N00), N01,
6817 DAG.getNode(ISD::FNEG, dl, VT, N1));
6824 SDValue DAGCombiner::visitFMUL(SDNode *N) {
6825 SDValue N0 = N->getOperand(0);
6826 SDValue N1 = N->getOperand(1);
6827 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6828 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6829 EVT VT = N->getValueType(0);
6830 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6833 if (VT.isVector()) {
6834 SDValue FoldedVOp = SimplifyVBinOp(N);
6835 if (FoldedVOp.getNode()) return FoldedVOp;
6838 // fold (fmul c1, c2) -> c1*c2
6840 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0, N1);
6841 // canonicalize constant to RHS
6842 if (N0CFP && !N1CFP)
6843 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N1, N0);
6844 // fold (fmul A, 0) -> 0
6845 if (DAG.getTarget().Options.UnsafeFPMath &&
6846 N1CFP && N1CFP->getValueAPF().isZero())
6848 // fold (fmul A, 0) -> 0, vector edition.
6849 if (DAG.getTarget().Options.UnsafeFPMath &&
6850 ISD::isBuildVectorAllZeros(N1.getNode()))
6852 // fold (fmul A, 1.0) -> A
6853 if (N1CFP && N1CFP->isExactlyValue(1.0))
6855 // fold (fmul X, 2.0) -> (fadd X, X)
6856 if (N1CFP && N1CFP->isExactlyValue(+2.0))
6857 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N0);
6858 // fold (fmul X, -1.0) -> (fneg X)
6859 if (N1CFP && N1CFP->isExactlyValue(-1.0))
6860 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
6861 return DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0);
6863 // fold (fmul (fneg X), (fneg Y)) -> (fmul X, Y)
6864 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI,
6865 &DAG.getTarget().Options)) {
6866 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI,
6867 &DAG.getTarget().Options)) {
6868 // Both can be negated for free, check to see if at least one is cheaper
6870 if (LHSNeg == 2 || RHSNeg == 2)
6871 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6872 GetNegatedExpression(N0, DAG, LegalOperations),
6873 GetNegatedExpression(N1, DAG, LegalOperations));
6877 // If allowed, fold (fmul (fmul x, c1), c2) -> (fmul x, (fmul c1, c2))
6878 if (DAG.getTarget().Options.UnsafeFPMath &&
6879 N1CFP && N0.getOpcode() == ISD::FMUL &&
6880 N0.getNode()->hasOneUse() && isa<ConstantFPSDNode>(N0.getOperand(1)))
6881 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0.getOperand(0),
6882 DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6883 N0.getOperand(1), N1));
6888 SDValue DAGCombiner::visitFMA(SDNode *N) {
6889 SDValue N0 = N->getOperand(0);
6890 SDValue N1 = N->getOperand(1);
6891 SDValue N2 = N->getOperand(2);
6892 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6893 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6894 EVT VT = N->getValueType(0);
6897 if (DAG.getTarget().Options.UnsafeFPMath) {
6898 if (N0CFP && N0CFP->isZero())
6900 if (N1CFP && N1CFP->isZero())
6903 if (N0CFP && N0CFP->isExactlyValue(1.0))
6904 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N1, N2);
6905 if (N1CFP && N1CFP->isExactlyValue(1.0))
6906 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N2);
6908 // Canonicalize (fma c, x, y) -> (fma x, c, y)
6909 if (N0CFP && !N1CFP)
6910 return DAG.getNode(ISD::FMA, SDLoc(N), VT, N1, N0, N2);
6912 // (fma x, c1, (fmul x, c2)) -> (fmul x, c1+c2)
6913 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
6914 N2.getOpcode() == ISD::FMUL &&
6915 N0 == N2.getOperand(0) &&
6916 N2.getOperand(1).getOpcode() == ISD::ConstantFP) {
6917 return DAG.getNode(ISD::FMUL, dl, VT, N0,
6918 DAG.getNode(ISD::FADD, dl, VT, N1, N2.getOperand(1)));
6922 // (fma (fmul x, c1), c2, y) -> (fma x, c1*c2, y)
6923 if (DAG.getTarget().Options.UnsafeFPMath &&
6924 N0.getOpcode() == ISD::FMUL && N1CFP &&
6925 N0.getOperand(1).getOpcode() == ISD::ConstantFP) {
6926 return DAG.getNode(ISD::FMA, dl, VT,
6928 DAG.getNode(ISD::FMUL, dl, VT, N1, N0.getOperand(1)),
6932 // (fma x, 1, y) -> (fadd x, y)
6933 // (fma x, -1, y) -> (fadd (fneg x), y)
6935 if (N1CFP->isExactlyValue(1.0))
6936 return DAG.getNode(ISD::FADD, dl, VT, N0, N2);
6938 if (N1CFP->isExactlyValue(-1.0) &&
6939 (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))) {
6940 SDValue RHSNeg = DAG.getNode(ISD::FNEG, dl, VT, N0);
6941 AddToWorklist(RHSNeg.getNode());
6942 return DAG.getNode(ISD::FADD, dl, VT, N2, RHSNeg);
6946 // (fma x, c, x) -> (fmul x, (c+1))
6947 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP && N0 == N2)
6948 return DAG.getNode(ISD::FMUL, dl, VT, N0,
6949 DAG.getNode(ISD::FADD, dl, VT,
6950 N1, DAG.getConstantFP(1.0, VT)));
6952 // (fma x, c, (fneg x)) -> (fmul x, (c-1))
6953 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
6954 N2.getOpcode() == ISD::FNEG && N2.getOperand(0) == N0)
6955 return DAG.getNode(ISD::FMUL, dl, VT, N0,
6956 DAG.getNode(ISD::FADD, dl, VT,
6957 N1, DAG.getConstantFP(-1.0, VT)));
6963 SDValue DAGCombiner::visitFDIV(SDNode *N) {
6964 SDValue N0 = N->getOperand(0);
6965 SDValue N1 = N->getOperand(1);
6966 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6967 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6968 EVT VT = N->getValueType(0);
6969 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6972 if (VT.isVector()) {
6973 SDValue FoldedVOp = SimplifyVBinOp(N);
6974 if (FoldedVOp.getNode()) return FoldedVOp;
6977 // fold (fdiv c1, c2) -> c1/c2
6979 return DAG.getNode(ISD::FDIV, SDLoc(N), VT, N0, N1);
6981 // fold (fdiv X, c2) -> fmul X, 1/c2 if losing precision is acceptable.
6982 if (N1CFP && DAG.getTarget().Options.UnsafeFPMath) {
6983 // Compute the reciprocal 1.0 / c2.
6984 APFloat N1APF = N1CFP->getValueAPF();
6985 APFloat Recip(N1APF.getSemantics(), 1); // 1.0
6986 APFloat::opStatus st = Recip.divide(N1APF, APFloat::rmNearestTiesToEven);
6987 // Only do the transform if the reciprocal is a legal fp immediate that
6988 // isn't too nasty (eg NaN, denormal, ...).
6989 if ((st == APFloat::opOK || st == APFloat::opInexact) && // Not too nasty
6990 (!LegalOperations ||
6991 // FIXME: custom lowering of ConstantFP might fail (see e.g. ARM
6992 // backend)... we should handle this gracefully after Legalize.
6993 // TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT) ||
6994 TLI.isOperationLegal(llvm::ISD::ConstantFP, VT) ||
6995 TLI.isFPImmLegal(Recip, VT)))
6996 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0,
6997 DAG.getConstantFP(Recip, VT));
7000 // (fdiv (fneg X), (fneg Y)) -> (fdiv X, Y)
7001 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI,
7002 &DAG.getTarget().Options)) {
7003 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI,
7004 &DAG.getTarget().Options)) {
7005 // Both can be negated for free, check to see if at least one is cheaper
7007 if (LHSNeg == 2 || RHSNeg == 2)
7008 return DAG.getNode(ISD::FDIV, SDLoc(N), VT,
7009 GetNegatedExpression(N0, DAG, LegalOperations),
7010 GetNegatedExpression(N1, DAG, LegalOperations));
7017 SDValue DAGCombiner::visitFREM(SDNode *N) {
7018 SDValue N0 = N->getOperand(0);
7019 SDValue N1 = N->getOperand(1);
7020 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7021 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
7022 EVT VT = N->getValueType(0);
7024 // fold (frem c1, c2) -> fmod(c1,c2)
7026 return DAG.getNode(ISD::FREM, SDLoc(N), VT, N0, N1);
7031 SDValue DAGCombiner::visitFCOPYSIGN(SDNode *N) {
7032 SDValue N0 = N->getOperand(0);
7033 SDValue N1 = N->getOperand(1);
7034 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7035 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
7036 EVT VT = N->getValueType(0);
7038 if (N0CFP && N1CFP) // Constant fold
7039 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT, N0, N1);
7042 const APFloat& V = N1CFP->getValueAPF();
7043 // copysign(x, c1) -> fabs(x) iff ispos(c1)
7044 // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1)
7045 if (!V.isNegative()) {
7046 if (!LegalOperations || TLI.isOperationLegal(ISD::FABS, VT))
7047 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
7049 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
7050 return DAG.getNode(ISD::FNEG, SDLoc(N), VT,
7051 DAG.getNode(ISD::FABS, SDLoc(N0), VT, N0));
7055 // copysign(fabs(x), y) -> copysign(x, y)
7056 // copysign(fneg(x), y) -> copysign(x, y)
7057 // copysign(copysign(x,z), y) -> copysign(x, y)
7058 if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG ||
7059 N0.getOpcode() == ISD::FCOPYSIGN)
7060 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
7061 N0.getOperand(0), N1);
7063 // copysign(x, abs(y)) -> abs(x)
7064 if (N1.getOpcode() == ISD::FABS)
7065 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
7067 // copysign(x, copysign(y,z)) -> copysign(x, z)
7068 if (N1.getOpcode() == ISD::FCOPYSIGN)
7069 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
7070 N0, N1.getOperand(1));
7072 // copysign(x, fp_extend(y)) -> copysign(x, y)
7073 // copysign(x, fp_round(y)) -> copysign(x, y)
7074 if (N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND)
7075 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
7076 N0, N1.getOperand(0));
7081 SDValue DAGCombiner::visitSINT_TO_FP(SDNode *N) {
7082 SDValue N0 = N->getOperand(0);
7083 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
7084 EVT VT = N->getValueType(0);
7085 EVT OpVT = N0.getValueType();
7087 // fold (sint_to_fp c1) -> c1fp
7089 // ...but only if the target supports immediate floating-point values
7090 (!LegalOperations ||
7091 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
7092 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0);
7094 // If the input is a legal type, and SINT_TO_FP is not legal on this target,
7095 // but UINT_TO_FP is legal on this target, try to convert.
7096 if (!TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT) &&
7097 TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT)) {
7098 // If the sign bit is known to be zero, we can change this to UINT_TO_FP.
7099 if (DAG.SignBitIsZero(N0))
7100 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0);
7103 // The next optimizations are desirable only if SELECT_CC can be lowered.
7104 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT) || !LegalOperations) {
7105 // fold (sint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
7106 if (N0.getOpcode() == ISD::SETCC && N0.getValueType() == MVT::i1 &&
7108 (!LegalOperations ||
7109 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
7111 { N0.getOperand(0), N0.getOperand(1),
7112 DAG.getConstantFP(-1.0, VT) , DAG.getConstantFP(0.0, VT),
7114 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops);
7117 // fold (sint_to_fp (zext (setcc x, y, cc))) ->
7118 // (select_cc x, y, 1.0, 0.0,, cc)
7119 if (N0.getOpcode() == ISD::ZERO_EXTEND &&
7120 N0.getOperand(0).getOpcode() == ISD::SETCC &&!VT.isVector() &&
7121 (!LegalOperations ||
7122 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
7124 { N0.getOperand(0).getOperand(0), N0.getOperand(0).getOperand(1),
7125 DAG.getConstantFP(1.0, VT) , DAG.getConstantFP(0.0, VT),
7126 N0.getOperand(0).getOperand(2) };
7127 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops);
7134 SDValue DAGCombiner::visitUINT_TO_FP(SDNode *N) {
7135 SDValue N0 = N->getOperand(0);
7136 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
7137 EVT VT = N->getValueType(0);
7138 EVT OpVT = N0.getValueType();
7140 // fold (uint_to_fp c1) -> c1fp
7142 // ...but only if the target supports immediate floating-point values
7143 (!LegalOperations ||
7144 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
7145 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0);
7147 // If the input is a legal type, and UINT_TO_FP is not legal on this target,
7148 // but SINT_TO_FP is legal on this target, try to convert.
7149 if (!TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT) &&
7150 TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT)) {
7151 // If the sign bit is known to be zero, we can change this to SINT_TO_FP.
7152 if (DAG.SignBitIsZero(N0))
7153 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0);
7156 // The next optimizations are desirable only if SELECT_CC can be lowered.
7157 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT) || !LegalOperations) {
7158 // fold (uint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
7160 if (N0.getOpcode() == ISD::SETCC && !VT.isVector() &&
7161 (!LegalOperations ||
7162 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
7164 { N0.getOperand(0), N0.getOperand(1),
7165 DAG.getConstantFP(1.0, VT), DAG.getConstantFP(0.0, VT),
7167 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops);
7174 SDValue DAGCombiner::visitFP_TO_SINT(SDNode *N) {
7175 SDValue N0 = N->getOperand(0);
7176 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7177 EVT VT = N->getValueType(0);
7179 // fold (fp_to_sint c1fp) -> c1
7181 return DAG.getNode(ISD::FP_TO_SINT, SDLoc(N), VT, N0);
7186 SDValue DAGCombiner::visitFP_TO_UINT(SDNode *N) {
7187 SDValue N0 = N->getOperand(0);
7188 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7189 EVT VT = N->getValueType(0);
7191 // fold (fp_to_uint c1fp) -> c1
7193 return DAG.getNode(ISD::FP_TO_UINT, SDLoc(N), VT, N0);
7198 SDValue DAGCombiner::visitFP_ROUND(SDNode *N) {
7199 SDValue N0 = N->getOperand(0);
7200 SDValue N1 = N->getOperand(1);
7201 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7202 EVT VT = N->getValueType(0);
7204 // fold (fp_round c1fp) -> c1fp
7206 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, N0, N1);
7208 // fold (fp_round (fp_extend x)) -> x
7209 if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType())
7210 return N0.getOperand(0);
7212 // fold (fp_round (fp_round x)) -> (fp_round x)
7213 if (N0.getOpcode() == ISD::FP_ROUND) {
7214 // This is a value preserving truncation if both round's are.
7215 bool IsTrunc = N->getConstantOperandVal(1) == 1 &&
7216 N0.getNode()->getConstantOperandVal(1) == 1;
7217 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, N0.getOperand(0),
7218 DAG.getIntPtrConstant(IsTrunc));
7221 // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y)
7222 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse()) {
7223 SDValue Tmp = DAG.getNode(ISD::FP_ROUND, SDLoc(N0), VT,
7224 N0.getOperand(0), N1);
7225 AddToWorklist(Tmp.getNode());
7226 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
7227 Tmp, N0.getOperand(1));
7233 SDValue DAGCombiner::visitFP_ROUND_INREG(SDNode *N) {
7234 SDValue N0 = N->getOperand(0);
7235 EVT VT = N->getValueType(0);
7236 EVT EVT = cast<VTSDNode>(N->getOperand(1))->getVT();
7237 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7239 // fold (fp_round_inreg c1fp) -> c1fp
7240 if (N0CFP && isTypeLegal(EVT)) {
7241 SDValue Round = DAG.getConstantFP(*N0CFP->getConstantFPValue(), EVT);
7242 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, Round);
7248 SDValue DAGCombiner::visitFP_EXTEND(SDNode *N) {
7249 SDValue N0 = N->getOperand(0);
7250 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7251 EVT VT = N->getValueType(0);
7253 // If this is fp_round(fpextend), don't fold it, allow ourselves to be folded.
7254 if (N->hasOneUse() &&
7255 N->use_begin()->getOpcode() == ISD::FP_ROUND)
7258 // fold (fp_extend c1fp) -> c1fp
7260 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, N0);
7262 // Turn fp_extend(fp_round(X, 1)) -> x since the fp_round doesn't affect the
7264 if (N0.getOpcode() == ISD::FP_ROUND
7265 && N0.getNode()->getConstantOperandVal(1) == 1) {
7266 SDValue In = N0.getOperand(0);
7267 if (In.getValueType() == VT) return In;
7268 if (VT.bitsLT(In.getValueType()))
7269 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT,
7270 In, N0.getOperand(1));
7271 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, In);
7274 // fold (fpext (load x)) -> (fpext (fptrunc (extload x)))
7275 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
7276 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType())) {
7277 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
7278 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT,
7280 LN0->getBasePtr(), N0.getValueType(),
7281 LN0->getMemOperand());
7282 CombineTo(N, ExtLoad);
7283 CombineTo(N0.getNode(),
7284 DAG.getNode(ISD::FP_ROUND, SDLoc(N0),
7285 N0.getValueType(), ExtLoad, DAG.getIntPtrConstant(1)),
7286 ExtLoad.getValue(1));
7287 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7293 SDValue DAGCombiner::visitFNEG(SDNode *N) {
7294 SDValue N0 = N->getOperand(0);
7295 EVT VT = N->getValueType(0);
7297 if (VT.isVector()) {
7298 SDValue FoldedVOp = SimplifyVUnaryOp(N);
7299 if (FoldedVOp.getNode()) return FoldedVOp;
7302 if (isNegatibleForFree(N0, LegalOperations, DAG.getTargetLoweringInfo(),
7303 &DAG.getTarget().Options))
7304 return GetNegatedExpression(N0, DAG, LegalOperations);
7306 // Transform fneg(bitconvert(x)) -> bitconvert(x^sign) to avoid loading
7307 // constant pool values.
7308 if (!TLI.isFNegFree(VT) && N0.getOpcode() == ISD::BITCAST &&
7310 N0.getNode()->hasOneUse() &&
7311 N0.getOperand(0).getValueType().isInteger()) {
7312 SDValue Int = N0.getOperand(0);
7313 EVT IntVT = Int.getValueType();
7314 if (IntVT.isInteger() && !IntVT.isVector()) {
7315 Int = DAG.getNode(ISD::XOR, SDLoc(N0), IntVT, Int,
7316 DAG.getConstant(APInt::getSignBit(IntVT.getSizeInBits()), IntVT));
7317 AddToWorklist(Int.getNode());
7318 return DAG.getNode(ISD::BITCAST, SDLoc(N),
7323 // (fneg (fmul c, x)) -> (fmul -c, x)
7324 if (N0.getOpcode() == ISD::FMUL) {
7325 ConstantFPSDNode *CFP1 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
7327 APFloat CVal = CFP1->getValueAPF();
7329 if (Level >= AfterLegalizeDAG &&
7330 (TLI.isFPImmLegal(CVal, N->getValueType(0)) ||
7331 TLI.isOperationLegal(ISD::ConstantFP, N->getValueType(0))))
7333 ISD::FMUL, SDLoc(N), VT, N0.getOperand(0),
7334 DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0.getOperand(1)));
7341 SDValue DAGCombiner::visitFCEIL(SDNode *N) {
7342 SDValue N0 = N->getOperand(0);
7343 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7344 EVT VT = N->getValueType(0);
7346 // fold (fceil c1) -> fceil(c1)
7348 return DAG.getNode(ISD::FCEIL, SDLoc(N), VT, N0);
7353 SDValue DAGCombiner::visitFTRUNC(SDNode *N) {
7354 SDValue N0 = N->getOperand(0);
7355 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7356 EVT VT = N->getValueType(0);
7358 // fold (ftrunc c1) -> ftrunc(c1)
7360 return DAG.getNode(ISD::FTRUNC, SDLoc(N), VT, N0);
7365 SDValue DAGCombiner::visitFFLOOR(SDNode *N) {
7366 SDValue N0 = N->getOperand(0);
7367 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7368 EVT VT = N->getValueType(0);
7370 // fold (ffloor c1) -> ffloor(c1)
7372 return DAG.getNode(ISD::FFLOOR, SDLoc(N), VT, N0);
7377 SDValue DAGCombiner::visitFABS(SDNode *N) {
7378 SDValue N0 = N->getOperand(0);
7379 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7380 EVT VT = N->getValueType(0);
7382 if (VT.isVector()) {
7383 SDValue FoldedVOp = SimplifyVUnaryOp(N);
7384 if (FoldedVOp.getNode()) return FoldedVOp;
7387 // fold (fabs c1) -> fabs(c1)
7389 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
7390 // fold (fabs (fabs x)) -> (fabs x)
7391 if (N0.getOpcode() == ISD::FABS)
7392 return N->getOperand(0);
7393 // fold (fabs (fneg x)) -> (fabs x)
7394 // fold (fabs (fcopysign x, y)) -> (fabs x)
7395 if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN)
7396 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0.getOperand(0));
7398 // Transform fabs(bitconvert(x)) -> bitconvert(x&~sign) to avoid loading
7399 // constant pool values.
7400 if (!TLI.isFAbsFree(VT) &&
7401 N0.getOpcode() == ISD::BITCAST && N0.getNode()->hasOneUse() &&
7402 N0.getOperand(0).getValueType().isInteger() &&
7403 !N0.getOperand(0).getValueType().isVector()) {
7404 SDValue Int = N0.getOperand(0);
7405 EVT IntVT = Int.getValueType();
7406 if (IntVT.isInteger() && !IntVT.isVector()) {
7407 Int = DAG.getNode(ISD::AND, SDLoc(N0), IntVT, Int,
7408 DAG.getConstant(~APInt::getSignBit(IntVT.getSizeInBits()), IntVT));
7409 AddToWorklist(Int.getNode());
7410 return DAG.getNode(ISD::BITCAST, SDLoc(N),
7411 N->getValueType(0), Int);
7418 SDValue DAGCombiner::visitBRCOND(SDNode *N) {
7419 SDValue Chain = N->getOperand(0);
7420 SDValue N1 = N->getOperand(1);
7421 SDValue N2 = N->getOperand(2);
7423 // If N is a constant we could fold this into a fallthrough or unconditional
7424 // branch. However that doesn't happen very often in normal code, because
7425 // Instcombine/SimplifyCFG should have handled the available opportunities.
7426 // If we did this folding here, it would be necessary to update the
7427 // MachineBasicBlock CFG, which is awkward.
7429 // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal
7431 if (N1.getOpcode() == ISD::SETCC &&
7432 TLI.isOperationLegalOrCustom(ISD::BR_CC,
7433 N1.getOperand(0).getValueType())) {
7434 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other,
7435 Chain, N1.getOperand(2),
7436 N1.getOperand(0), N1.getOperand(1), N2);
7439 if ((N1.hasOneUse() && N1.getOpcode() == ISD::SRL) ||
7440 ((N1.getOpcode() == ISD::TRUNCATE && N1.hasOneUse()) &&
7441 (N1.getOperand(0).hasOneUse() &&
7442 N1.getOperand(0).getOpcode() == ISD::SRL))) {
7443 SDNode *Trunc = nullptr;
7444 if (N1.getOpcode() == ISD::TRUNCATE) {
7445 // Look pass the truncate.
7446 Trunc = N1.getNode();
7447 N1 = N1.getOperand(0);
7450 // Match this pattern so that we can generate simpler code:
7453 // %b = and i32 %a, 2
7454 // %c = srl i32 %b, 1
7455 // brcond i32 %c ...
7460 // %b = and i32 %a, 2
7461 // %c = setcc eq %b, 0
7464 // This applies only when the AND constant value has one bit set and the
7465 // SRL constant is equal to the log2 of the AND constant. The back-end is
7466 // smart enough to convert the result into a TEST/JMP sequence.
7467 SDValue Op0 = N1.getOperand(0);
7468 SDValue Op1 = N1.getOperand(1);
7470 if (Op0.getOpcode() == ISD::AND &&
7471 Op1.getOpcode() == ISD::Constant) {
7472 SDValue AndOp1 = Op0.getOperand(1);
7474 if (AndOp1.getOpcode() == ISD::Constant) {
7475 const APInt &AndConst = cast<ConstantSDNode>(AndOp1)->getAPIntValue();
7477 if (AndConst.isPowerOf2() &&
7478 cast<ConstantSDNode>(Op1)->getAPIntValue()==AndConst.logBase2()) {
7480 DAG.getSetCC(SDLoc(N),
7481 getSetCCResultType(Op0.getValueType()),
7482 Op0, DAG.getConstant(0, Op0.getValueType()),
7485 SDValue NewBRCond = DAG.getNode(ISD::BRCOND, SDLoc(N),
7486 MVT::Other, Chain, SetCC, N2);
7487 // Don't add the new BRCond into the worklist or else SimplifySelectCC
7488 // will convert it back to (X & C1) >> C2.
7489 CombineTo(N, NewBRCond, false);
7490 // Truncate is dead.
7492 removeFromWorklist(Trunc);
7493 DAG.DeleteNode(Trunc);
7495 // Replace the uses of SRL with SETCC
7496 WorklistRemover DeadNodes(*this);
7497 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
7498 removeFromWorklist(N1.getNode());
7499 DAG.DeleteNode(N1.getNode());
7500 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7506 // Restore N1 if the above transformation doesn't match.
7507 N1 = N->getOperand(1);
7510 // Transform br(xor(x, y)) -> br(x != y)
7511 // Transform br(xor(xor(x,y), 1)) -> br (x == y)
7512 if (N1.hasOneUse() && N1.getOpcode() == ISD::XOR) {
7513 SDNode *TheXor = N1.getNode();
7514 SDValue Op0 = TheXor->getOperand(0);
7515 SDValue Op1 = TheXor->getOperand(1);
7516 if (Op0.getOpcode() == Op1.getOpcode()) {
7517 // Avoid missing important xor optimizations.
7518 SDValue Tmp = visitXOR(TheXor);
7519 if (Tmp.getNode()) {
7520 if (Tmp.getNode() != TheXor) {
7521 DEBUG(dbgs() << "\nReplacing.8 ";
7523 dbgs() << "\nWith: ";
7524 Tmp.getNode()->dump(&DAG);
7526 WorklistRemover DeadNodes(*this);
7527 DAG.ReplaceAllUsesOfValueWith(N1, Tmp);
7528 removeFromWorklist(TheXor);
7529 DAG.DeleteNode(TheXor);
7530 return DAG.getNode(ISD::BRCOND, SDLoc(N),
7531 MVT::Other, Chain, Tmp, N2);
7534 // visitXOR has changed XOR's operands or replaced the XOR completely,
7536 return SDValue(N, 0);
7540 if (Op0.getOpcode() != ISD::SETCC && Op1.getOpcode() != ISD::SETCC) {
7542 if (ConstantSDNode *RHSCI = dyn_cast<ConstantSDNode>(Op0))
7543 if (RHSCI->getAPIntValue() == 1 && Op0.hasOneUse() &&
7544 Op0.getOpcode() == ISD::XOR) {
7545 TheXor = Op0.getNode();
7549 EVT SetCCVT = N1.getValueType();
7551 SetCCVT = getSetCCResultType(SetCCVT);
7552 SDValue SetCC = DAG.getSetCC(SDLoc(TheXor),
7555 Equal ? ISD::SETEQ : ISD::SETNE);
7556 // Replace the uses of XOR with SETCC
7557 WorklistRemover DeadNodes(*this);
7558 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
7559 removeFromWorklist(N1.getNode());
7560 DAG.DeleteNode(N1.getNode());
7561 return DAG.getNode(ISD::BRCOND, SDLoc(N),
7562 MVT::Other, Chain, SetCC, N2);
7569 // Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB.
7571 SDValue DAGCombiner::visitBR_CC(SDNode *N) {
7572 CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1));
7573 SDValue CondLHS = N->getOperand(2), CondRHS = N->getOperand(3);
7575 // If N is a constant we could fold this into a fallthrough or unconditional
7576 // branch. However that doesn't happen very often in normal code, because
7577 // Instcombine/SimplifyCFG should have handled the available opportunities.
7578 // If we did this folding here, it would be necessary to update the
7579 // MachineBasicBlock CFG, which is awkward.
7581 // Use SimplifySetCC to simplify SETCC's.
7582 SDValue Simp = SimplifySetCC(getSetCCResultType(CondLHS.getValueType()),
7583 CondLHS, CondRHS, CC->get(), SDLoc(N),
7585 if (Simp.getNode()) AddToWorklist(Simp.getNode());
7587 // fold to a simpler setcc
7588 if (Simp.getNode() && Simp.getOpcode() == ISD::SETCC)
7589 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other,
7590 N->getOperand(0), Simp.getOperand(2),
7591 Simp.getOperand(0), Simp.getOperand(1),
7597 /// canFoldInAddressingMode - Return true if 'Use' is a load or a store that
7598 /// uses N as its base pointer and that N may be folded in the load / store
7599 /// addressing mode.
7600 static bool canFoldInAddressingMode(SDNode *N, SDNode *Use,
7602 const TargetLowering &TLI) {
7604 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Use)) {
7605 if (LD->isIndexed() || LD->getBasePtr().getNode() != N)
7607 VT = Use->getValueType(0);
7608 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(Use)) {
7609 if (ST->isIndexed() || ST->getBasePtr().getNode() != N)
7611 VT = ST->getValue().getValueType();
7615 TargetLowering::AddrMode AM;
7616 if (N->getOpcode() == ISD::ADD) {
7617 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
7620 AM.BaseOffs = Offset->getSExtValue();
7624 } else if (N->getOpcode() == ISD::SUB) {
7625 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
7628 AM.BaseOffs = -Offset->getSExtValue();
7635 return TLI.isLegalAddressingMode(AM, VT.getTypeForEVT(*DAG.getContext()));
7638 /// CombineToPreIndexedLoadStore - Try turning a load / store into a
7639 /// pre-indexed load / store when the base pointer is an add or subtract
7640 /// and it has other uses besides the load / store. After the
7641 /// transformation, the new indexed load / store has effectively folded
7642 /// the add / subtract in and all of its other uses are redirected to the
7643 /// new load / store.
7644 bool DAGCombiner::CombineToPreIndexedLoadStore(SDNode *N) {
7645 if (Level < AfterLegalizeDAG)
7651 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7652 if (LD->isIndexed())
7654 VT = LD->getMemoryVT();
7655 if (!TLI.isIndexedLoadLegal(ISD::PRE_INC, VT) &&
7656 !TLI.isIndexedLoadLegal(ISD::PRE_DEC, VT))
7658 Ptr = LD->getBasePtr();
7659 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7660 if (ST->isIndexed())
7662 VT = ST->getMemoryVT();
7663 if (!TLI.isIndexedStoreLegal(ISD::PRE_INC, VT) &&
7664 !TLI.isIndexedStoreLegal(ISD::PRE_DEC, VT))
7666 Ptr = ST->getBasePtr();
7672 // If the pointer is not an add/sub, or if it doesn't have multiple uses, bail
7673 // out. There is no reason to make this a preinc/predec.
7674 if ((Ptr.getOpcode() != ISD::ADD && Ptr.getOpcode() != ISD::SUB) ||
7675 Ptr.getNode()->hasOneUse())
7678 // Ask the target to do addressing mode selection.
7681 ISD::MemIndexedMode AM = ISD::UNINDEXED;
7682 if (!TLI.getPreIndexedAddressParts(N, BasePtr, Offset, AM, DAG))
7685 // Backends without true r+i pre-indexed forms may need to pass a
7686 // constant base with a variable offset so that constant coercion
7687 // will work with the patterns in canonical form.
7688 bool Swapped = false;
7689 if (isa<ConstantSDNode>(BasePtr)) {
7690 std::swap(BasePtr, Offset);
7694 // Don't create a indexed load / store with zero offset.
7695 if (isa<ConstantSDNode>(Offset) &&
7696 cast<ConstantSDNode>(Offset)->isNullValue())
7699 // Try turning it into a pre-indexed load / store except when:
7700 // 1) The new base ptr is a frame index.
7701 // 2) If N is a store and the new base ptr is either the same as or is a
7702 // predecessor of the value being stored.
7703 // 3) Another use of old base ptr is a predecessor of N. If ptr is folded
7704 // that would create a cycle.
7705 // 4) All uses are load / store ops that use it as old base ptr.
7707 // Check #1. Preinc'ing a frame index would require copying the stack pointer
7708 // (plus the implicit offset) to a register to preinc anyway.
7709 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
7714 SDValue Val = cast<StoreSDNode>(N)->getValue();
7715 if (Val == BasePtr || BasePtr.getNode()->isPredecessorOf(Val.getNode()))
7719 // If the offset is a constant, there may be other adds of constants that
7720 // can be folded with this one. We should do this to avoid having to keep
7721 // a copy of the original base pointer.
7722 SmallVector<SDNode *, 16> OtherUses;
7723 if (isa<ConstantSDNode>(Offset))
7724 for (SDNode *Use : BasePtr.getNode()->uses()) {
7725 if (Use == Ptr.getNode())
7728 if (Use->isPredecessorOf(N))
7731 if (Use->getOpcode() != ISD::ADD && Use->getOpcode() != ISD::SUB) {
7736 SDValue Op0 = Use->getOperand(0), Op1 = Use->getOperand(1);
7737 if (Op1.getNode() == BasePtr.getNode())
7738 std::swap(Op0, Op1);
7739 assert(Op0.getNode() == BasePtr.getNode() &&
7740 "Use of ADD/SUB but not an operand");
7742 if (!isa<ConstantSDNode>(Op1)) {
7747 // FIXME: In some cases, we can be smarter about this.
7748 if (Op1.getValueType() != Offset.getValueType()) {
7753 OtherUses.push_back(Use);
7757 std::swap(BasePtr, Offset);
7759 // Now check for #3 and #4.
7760 bool RealUse = false;
7762 // Caches for hasPredecessorHelper
7763 SmallPtrSet<const SDNode *, 32> Visited;
7764 SmallVector<const SDNode *, 16> Worklist;
7766 for (SDNode *Use : Ptr.getNode()->uses()) {
7769 if (N->hasPredecessorHelper(Use, Visited, Worklist))
7772 // If Ptr may be folded in addressing mode of other use, then it's
7773 // not profitable to do this transformation.
7774 if (!canFoldInAddressingMode(Ptr.getNode(), Use, DAG, TLI))
7783 Result = DAG.getIndexedLoad(SDValue(N,0), SDLoc(N),
7784 BasePtr, Offset, AM);
7786 Result = DAG.getIndexedStore(SDValue(N,0), SDLoc(N),
7787 BasePtr, Offset, AM);
7790 DEBUG(dbgs() << "\nReplacing.4 ";
7792 dbgs() << "\nWith: ";
7793 Result.getNode()->dump(&DAG);
7795 WorklistRemover DeadNodes(*this);
7797 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
7798 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
7800 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
7803 // Finally, since the node is now dead, remove it from the graph.
7807 std::swap(BasePtr, Offset);
7809 // Replace other uses of BasePtr that can be updated to use Ptr
7810 for (unsigned i = 0, e = OtherUses.size(); i != e; ++i) {
7811 unsigned OffsetIdx = 1;
7812 if (OtherUses[i]->getOperand(OffsetIdx).getNode() == BasePtr.getNode())
7814 assert(OtherUses[i]->getOperand(!OffsetIdx).getNode() ==
7815 BasePtr.getNode() && "Expected BasePtr operand");
7817 // We need to replace ptr0 in the following expression:
7818 // x0 * offset0 + y0 * ptr0 = t0
7820 // x1 * offset1 + y1 * ptr0 = t1 (the indexed load/store)
7822 // where x0, x1, y0 and y1 in {-1, 1} are given by the types of the
7823 // indexed load/store and the expresion that needs to be re-written.
7825 // Therefore, we have:
7826 // t0 = (x0 * offset0 - x1 * y0 * y1 *offset1) + (y0 * y1) * t1
7828 ConstantSDNode *CN =
7829 cast<ConstantSDNode>(OtherUses[i]->getOperand(OffsetIdx));
7831 APInt Offset0 = CN->getAPIntValue();
7832 APInt Offset1 = cast<ConstantSDNode>(Offset)->getAPIntValue();
7834 X0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 1) ? -1 : 1;
7835 Y0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 0) ? -1 : 1;
7836 X1 = (AM == ISD::PRE_DEC && !Swapped) ? -1 : 1;
7837 Y1 = (AM == ISD::PRE_DEC && Swapped) ? -1 : 1;
7839 unsigned Opcode = (Y0 * Y1 < 0) ? ISD::SUB : ISD::ADD;
7841 APInt CNV = Offset0;
7842 if (X0 < 0) CNV = -CNV;
7843 if (X1 * Y0 * Y1 < 0) CNV = CNV + Offset1;
7844 else CNV = CNV - Offset1;
7846 // We can now generate the new expression.
7847 SDValue NewOp1 = DAG.getConstant(CNV, CN->getValueType(0));
7848 SDValue NewOp2 = Result.getValue(isLoad ? 1 : 0);
7850 SDValue NewUse = DAG.getNode(Opcode,
7851 SDLoc(OtherUses[i]),
7852 OtherUses[i]->getValueType(0), NewOp1, NewOp2);
7853 DAG.ReplaceAllUsesOfValueWith(SDValue(OtherUses[i], 0), NewUse);
7854 removeFromWorklist(OtherUses[i]);
7855 DAG.DeleteNode(OtherUses[i]);
7858 // Replace the uses of Ptr with uses of the updated base value.
7859 DAG.ReplaceAllUsesOfValueWith(Ptr, Result.getValue(isLoad ? 1 : 0));
7860 removeFromWorklist(Ptr.getNode());
7861 DAG.DeleteNode(Ptr.getNode());
7866 /// CombineToPostIndexedLoadStore - Try to combine a load / store with a
7867 /// add / sub of the base pointer node into a post-indexed load / store.
7868 /// The transformation folded the add / subtract into the new indexed
7869 /// load / store effectively and all of its uses are redirected to the
7870 /// new load / store.
7871 bool DAGCombiner::CombineToPostIndexedLoadStore(SDNode *N) {
7872 if (Level < AfterLegalizeDAG)
7878 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7879 if (LD->isIndexed())
7881 VT = LD->getMemoryVT();
7882 if (!TLI.isIndexedLoadLegal(ISD::POST_INC, VT) &&
7883 !TLI.isIndexedLoadLegal(ISD::POST_DEC, VT))
7885 Ptr = LD->getBasePtr();
7886 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7887 if (ST->isIndexed())
7889 VT = ST->getMemoryVT();
7890 if (!TLI.isIndexedStoreLegal(ISD::POST_INC, VT) &&
7891 !TLI.isIndexedStoreLegal(ISD::POST_DEC, VT))
7893 Ptr = ST->getBasePtr();
7899 if (Ptr.getNode()->hasOneUse())
7902 for (SDNode *Op : Ptr.getNode()->uses()) {
7904 (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB))
7909 ISD::MemIndexedMode AM = ISD::UNINDEXED;
7910 if (TLI.getPostIndexedAddressParts(N, Op, BasePtr, Offset, AM, DAG)) {
7911 // Don't create a indexed load / store with zero offset.
7912 if (isa<ConstantSDNode>(Offset) &&
7913 cast<ConstantSDNode>(Offset)->isNullValue())
7916 // Try turning it into a post-indexed load / store except when
7917 // 1) All uses are load / store ops that use it as base ptr (and
7918 // it may be folded as addressing mmode).
7919 // 2) Op must be independent of N, i.e. Op is neither a predecessor
7920 // nor a successor of N. Otherwise, if Op is folded that would
7923 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
7927 bool TryNext = false;
7928 for (SDNode *Use : BasePtr.getNode()->uses()) {
7929 if (Use == Ptr.getNode())
7932 // If all the uses are load / store addresses, then don't do the
7934 if (Use->getOpcode() == ISD::ADD || Use->getOpcode() == ISD::SUB){
7935 bool RealUse = false;
7936 for (SDNode *UseUse : Use->uses()) {
7937 if (!canFoldInAddressingMode(Use, UseUse, DAG, TLI))
7952 if (!Op->isPredecessorOf(N) && !N->isPredecessorOf(Op)) {
7953 SDValue Result = isLoad
7954 ? DAG.getIndexedLoad(SDValue(N,0), SDLoc(N),
7955 BasePtr, Offset, AM)
7956 : DAG.getIndexedStore(SDValue(N,0), SDLoc(N),
7957 BasePtr, Offset, AM);
7960 DEBUG(dbgs() << "\nReplacing.5 ";
7962 dbgs() << "\nWith: ";
7963 Result.getNode()->dump(&DAG);
7965 WorklistRemover DeadNodes(*this);
7967 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
7968 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
7970 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
7973 // Finally, since the node is now dead, remove it from the graph.
7976 // Replace the uses of Use with uses of the updated base value.
7977 DAG.ReplaceAllUsesOfValueWith(SDValue(Op, 0),
7978 Result.getValue(isLoad ? 1 : 0));
7979 removeFromWorklist(Op);
7989 SDValue DAGCombiner::visitLOAD(SDNode *N) {
7990 LoadSDNode *LD = cast<LoadSDNode>(N);
7991 SDValue Chain = LD->getChain();
7992 SDValue Ptr = LD->getBasePtr();
7994 // If load is not volatile and there are no uses of the loaded value (and
7995 // the updated indexed value in case of indexed loads), change uses of the
7996 // chain value into uses of the chain input (i.e. delete the dead load).
7997 if (!LD->isVolatile()) {
7998 if (N->getValueType(1) == MVT::Other) {
8000 if (!N->hasAnyUseOfValue(0)) {
8001 // It's not safe to use the two value CombineTo variant here. e.g.
8002 // v1, chain2 = load chain1, loc
8003 // v2, chain3 = load chain2, loc
8005 // Now we replace use of chain2 with chain1. This makes the second load
8006 // isomorphic to the one we are deleting, and thus makes this load live.
8007 DEBUG(dbgs() << "\nReplacing.6 ";
8009 dbgs() << "\nWith chain: ";
8010 Chain.getNode()->dump(&DAG);
8012 WorklistRemover DeadNodes(*this);
8013 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
8015 if (N->use_empty()) {
8016 removeFromWorklist(N);
8020 return SDValue(N, 0); // Return N so it doesn't get rechecked!
8024 assert(N->getValueType(2) == MVT::Other && "Malformed indexed loads?");
8025 if (!N->hasAnyUseOfValue(0) && !N->hasAnyUseOfValue(1)) {
8026 SDValue Undef = DAG.getUNDEF(N->getValueType(0));
8027 DEBUG(dbgs() << "\nReplacing.7 ";
8029 dbgs() << "\nWith: ";
8030 Undef.getNode()->dump(&DAG);
8031 dbgs() << " and 2 other values\n");
8032 WorklistRemover DeadNodes(*this);
8033 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Undef);
8034 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1),
8035 DAG.getUNDEF(N->getValueType(1)));
8036 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 2), Chain);
8037 removeFromWorklist(N);
8039 return SDValue(N, 0); // Return N so it doesn't get rechecked!
8044 // If this load is directly stored, replace the load value with the stored
8046 // TODO: Handle store large -> read small portion.
8047 // TODO: Handle TRUNCSTORE/LOADEXT
8048 if (ISD::isNormalLoad(N) && !LD->isVolatile()) {
8049 if (ISD::isNON_TRUNCStore(Chain.getNode())) {
8050 StoreSDNode *PrevST = cast<StoreSDNode>(Chain);
8051 if (PrevST->getBasePtr() == Ptr &&
8052 PrevST->getValue().getValueType() == N->getValueType(0))
8053 return CombineTo(N, Chain.getOperand(1), Chain);
8057 // Try to infer better alignment information than the load already has.
8058 if (OptLevel != CodeGenOpt::None && LD->isUnindexed()) {
8059 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
8060 if (Align > LD->getMemOperand()->getBaseAlignment()) {
8062 DAG.getExtLoad(LD->getExtensionType(), SDLoc(N),
8063 LD->getValueType(0),
8064 Chain, Ptr, LD->getPointerInfo(),
8066 LD->isVolatile(), LD->isNonTemporal(), Align,
8068 return CombineTo(N, NewLoad, SDValue(NewLoad.getNode(), 1), true);
8073 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA :
8074 TLI.getTargetMachine().getSubtarget<TargetSubtargetInfo>().useAA();
8076 if (CombinerAAOnlyFunc.getNumOccurrences() &&
8077 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
8080 if (UseAA && LD->isUnindexed()) {
8081 // Walk up chain skipping non-aliasing memory nodes.
8082 SDValue BetterChain = FindBetterChain(N, Chain);
8084 // If there is a better chain.
8085 if (Chain != BetterChain) {
8088 // Replace the chain to void dependency.
8089 if (LD->getExtensionType() == ISD::NON_EXTLOAD) {
8090 ReplLoad = DAG.getLoad(N->getValueType(0), SDLoc(LD),
8091 BetterChain, Ptr, LD->getMemOperand());
8093 ReplLoad = DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD),
8094 LD->getValueType(0),
8095 BetterChain, Ptr, LD->getMemoryVT(),
8096 LD->getMemOperand());
8099 // Create token factor to keep old chain connected.
8100 SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N),
8101 MVT::Other, Chain, ReplLoad.getValue(1));
8103 // Make sure the new and old chains are cleaned up.
8104 AddToWorklist(Token.getNode());
8106 // Replace uses with load result and token factor. Don't add users
8108 return CombineTo(N, ReplLoad.getValue(0), Token, false);
8112 // Try transforming N to an indexed load.
8113 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
8114 return SDValue(N, 0);
8116 // Try to slice up N to more direct loads if the slices are mapped to
8117 // different register banks or pairing can take place.
8119 return SDValue(N, 0);
8125 /// \brief Helper structure used to slice a load in smaller loads.
8126 /// Basically a slice is obtained from the following sequence:
8127 /// Origin = load Ty1, Base
8128 /// Shift = srl Ty1 Origin, CstTy Amount
8129 /// Inst = trunc Shift to Ty2
8131 /// Then, it will be rewriten into:
8132 /// Slice = load SliceTy, Base + SliceOffset
8133 /// [Inst = zext Slice to Ty2], only if SliceTy <> Ty2
8135 /// SliceTy is deduced from the number of bits that are actually used to
8137 struct LoadedSlice {
8138 /// \brief Helper structure used to compute the cost of a slice.
8140 /// Are we optimizing for code size.
8145 unsigned CrossRegisterBanksCopies;
8149 Cost(bool ForCodeSize = false)
8150 : ForCodeSize(ForCodeSize), Loads(0), Truncates(0),
8151 CrossRegisterBanksCopies(0), ZExts(0), Shift(0) {}
8153 /// \brief Get the cost of one isolated slice.
8154 Cost(const LoadedSlice &LS, bool ForCodeSize = false)
8155 : ForCodeSize(ForCodeSize), Loads(1), Truncates(0),
8156 CrossRegisterBanksCopies(0), ZExts(0), Shift(0) {
8157 EVT TruncType = LS.Inst->getValueType(0);
8158 EVT LoadedType = LS.getLoadedType();
8159 if (TruncType != LoadedType &&
8160 !LS.DAG->getTargetLoweringInfo().isZExtFree(LoadedType, TruncType))
8164 /// \brief Account for slicing gain in the current cost.
8165 /// Slicing provide a few gains like removing a shift or a
8166 /// truncate. This method allows to grow the cost of the original
8167 /// load with the gain from this slice.
8168 void addSliceGain(const LoadedSlice &LS) {
8169 // Each slice saves a truncate.
8170 const TargetLowering &TLI = LS.DAG->getTargetLoweringInfo();
8171 if (!TLI.isTruncateFree(LS.Inst->getValueType(0),
8172 LS.Inst->getOperand(0).getValueType()))
8174 // If there is a shift amount, this slice gets rid of it.
8177 // If this slice can merge a cross register bank copy, account for it.
8178 if (LS.canMergeExpensiveCrossRegisterBankCopy())
8179 ++CrossRegisterBanksCopies;
8182 Cost &operator+=(const Cost &RHS) {
8184 Truncates += RHS.Truncates;
8185 CrossRegisterBanksCopies += RHS.CrossRegisterBanksCopies;
8191 bool operator==(const Cost &RHS) const {
8192 return Loads == RHS.Loads && Truncates == RHS.Truncates &&
8193 CrossRegisterBanksCopies == RHS.CrossRegisterBanksCopies &&
8194 ZExts == RHS.ZExts && Shift == RHS.Shift;
8197 bool operator!=(const Cost &RHS) const { return !(*this == RHS); }
8199 bool operator<(const Cost &RHS) const {
8200 // Assume cross register banks copies are as expensive as loads.
8201 // FIXME: Do we want some more target hooks?
8202 unsigned ExpensiveOpsLHS = Loads + CrossRegisterBanksCopies;
8203 unsigned ExpensiveOpsRHS = RHS.Loads + RHS.CrossRegisterBanksCopies;
8204 // Unless we are optimizing for code size, consider the
8205 // expensive operation first.
8206 if (!ForCodeSize && ExpensiveOpsLHS != ExpensiveOpsRHS)
8207 return ExpensiveOpsLHS < ExpensiveOpsRHS;
8208 return (Truncates + ZExts + Shift + ExpensiveOpsLHS) <
8209 (RHS.Truncates + RHS.ZExts + RHS.Shift + ExpensiveOpsRHS);
8212 bool operator>(const Cost &RHS) const { return RHS < *this; }
8214 bool operator<=(const Cost &RHS) const { return !(RHS < *this); }
8216 bool operator>=(const Cost &RHS) const { return !(*this < RHS); }
8218 // The last instruction that represent the slice. This should be a
8219 // truncate instruction.
8221 // The original load instruction.
8223 // The right shift amount in bits from the original load.
8225 // The DAG from which Origin came from.
8226 // This is used to get some contextual information about legal types, etc.
8229 LoadedSlice(SDNode *Inst = nullptr, LoadSDNode *Origin = nullptr,
8230 unsigned Shift = 0, SelectionDAG *DAG = nullptr)
8231 : Inst(Inst), Origin(Origin), Shift(Shift), DAG(DAG) {}
8233 LoadedSlice(const LoadedSlice &LS)
8234 : Inst(LS.Inst), Origin(LS.Origin), Shift(LS.Shift), DAG(LS.DAG) {}
8236 /// \brief Get the bits used in a chunk of bits \p BitWidth large.
8237 /// \return Result is \p BitWidth and has used bits set to 1 and
8238 /// not used bits set to 0.
8239 APInt getUsedBits() const {
8240 // Reproduce the trunc(lshr) sequence:
8241 // - Start from the truncated value.
8242 // - Zero extend to the desired bit width.
8244 assert(Origin && "No original load to compare against.");
8245 unsigned BitWidth = Origin->getValueSizeInBits(0);
8246 assert(Inst && "This slice is not bound to an instruction");
8247 assert(Inst->getValueSizeInBits(0) <= BitWidth &&
8248 "Extracted slice is bigger than the whole type!");
8249 APInt UsedBits(Inst->getValueSizeInBits(0), 0);
8250 UsedBits.setAllBits();
8251 UsedBits = UsedBits.zext(BitWidth);
8256 /// \brief Get the size of the slice to be loaded in bytes.
8257 unsigned getLoadedSize() const {
8258 unsigned SliceSize = getUsedBits().countPopulation();
8259 assert(!(SliceSize & 0x7) && "Size is not a multiple of a byte.");
8260 return SliceSize / 8;
8263 /// \brief Get the type that will be loaded for this slice.
8264 /// Note: This may not be the final type for the slice.
8265 EVT getLoadedType() const {
8266 assert(DAG && "Missing context");
8267 LLVMContext &Ctxt = *DAG->getContext();
8268 return EVT::getIntegerVT(Ctxt, getLoadedSize() * 8);
8271 /// \brief Get the alignment of the load used for this slice.
8272 unsigned getAlignment() const {
8273 unsigned Alignment = Origin->getAlignment();
8274 unsigned Offset = getOffsetFromBase();
8276 Alignment = MinAlign(Alignment, Alignment + Offset);
8280 /// \brief Check if this slice can be rewritten with legal operations.
8281 bool isLegal() const {
8282 // An invalid slice is not legal.
8283 if (!Origin || !Inst || !DAG)
8286 // Offsets are for indexed load only, we do not handle that.
8287 if (Origin->getOffset().getOpcode() != ISD::UNDEF)
8290 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
8292 // Check that the type is legal.
8293 EVT SliceType = getLoadedType();
8294 if (!TLI.isTypeLegal(SliceType))
8297 // Check that the load is legal for this type.
8298 if (!TLI.isOperationLegal(ISD::LOAD, SliceType))
8301 // Check that the offset can be computed.
8302 // 1. Check its type.
8303 EVT PtrType = Origin->getBasePtr().getValueType();
8304 if (PtrType == MVT::Untyped || PtrType.isExtended())
8307 // 2. Check that it fits in the immediate.
8308 if (!TLI.isLegalAddImmediate(getOffsetFromBase()))
8311 // 3. Check that the computation is legal.
8312 if (!TLI.isOperationLegal(ISD::ADD, PtrType))
8315 // Check that the zext is legal if it needs one.
8316 EVT TruncateType = Inst->getValueType(0);
8317 if (TruncateType != SliceType &&
8318 !TLI.isOperationLegal(ISD::ZERO_EXTEND, TruncateType))
8324 /// \brief Get the offset in bytes of this slice in the original chunk of
8326 /// \pre DAG != nullptr.
8327 uint64_t getOffsetFromBase() const {
8328 assert(DAG && "Missing context.");
8330 DAG->getTargetLoweringInfo().getDataLayout()->isBigEndian();
8331 assert(!(Shift & 0x7) && "Shifts not aligned on Bytes are not supported.");
8332 uint64_t Offset = Shift / 8;
8333 unsigned TySizeInBytes = Origin->getValueSizeInBits(0) / 8;
8334 assert(!(Origin->getValueSizeInBits(0) & 0x7) &&
8335 "The size of the original loaded type is not a multiple of a"
8337 // If Offset is bigger than TySizeInBytes, it means we are loading all
8338 // zeros. This should have been optimized before in the process.
8339 assert(TySizeInBytes > Offset &&
8340 "Invalid shift amount for given loaded size");
8342 Offset = TySizeInBytes - Offset - getLoadedSize();
8346 /// \brief Generate the sequence of instructions to load the slice
8347 /// represented by this object and redirect the uses of this slice to
8348 /// this new sequence of instructions.
8349 /// \pre this->Inst && this->Origin are valid Instructions and this
8350 /// object passed the legal check: LoadedSlice::isLegal returned true.
8351 /// \return The last instruction of the sequence used to load the slice.
8352 SDValue loadSlice() const {
8353 assert(Inst && Origin && "Unable to replace a non-existing slice.");
8354 const SDValue &OldBaseAddr = Origin->getBasePtr();
8355 SDValue BaseAddr = OldBaseAddr;
8356 // Get the offset in that chunk of bytes w.r.t. the endianess.
8357 int64_t Offset = static_cast<int64_t>(getOffsetFromBase());
8358 assert(Offset >= 0 && "Offset too big to fit in int64_t!");
8360 // BaseAddr = BaseAddr + Offset.
8361 EVT ArithType = BaseAddr.getValueType();
8362 BaseAddr = DAG->getNode(ISD::ADD, SDLoc(Origin), ArithType, BaseAddr,
8363 DAG->getConstant(Offset, ArithType));
8366 // Create the type of the loaded slice according to its size.
8367 EVT SliceType = getLoadedType();
8369 // Create the load for the slice.
8370 SDValue LastInst = DAG->getLoad(
8371 SliceType, SDLoc(Origin), Origin->getChain(), BaseAddr,
8372 Origin->getPointerInfo().getWithOffset(Offset), Origin->isVolatile(),
8373 Origin->isNonTemporal(), Origin->isInvariant(), getAlignment());
8374 // If the final type is not the same as the loaded type, this means that
8375 // we have to pad with zero. Create a zero extend for that.
8376 EVT FinalType = Inst->getValueType(0);
8377 if (SliceType != FinalType)
8379 DAG->getNode(ISD::ZERO_EXTEND, SDLoc(LastInst), FinalType, LastInst);
8383 /// \brief Check if this slice can be merged with an expensive cross register
8384 /// bank copy. E.g.,
8386 /// f = bitcast i32 i to float
8387 bool canMergeExpensiveCrossRegisterBankCopy() const {
8388 if (!Inst || !Inst->hasOneUse())
8390 SDNode *Use = *Inst->use_begin();
8391 if (Use->getOpcode() != ISD::BITCAST)
8393 assert(DAG && "Missing context");
8394 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
8395 EVT ResVT = Use->getValueType(0);
8396 const TargetRegisterClass *ResRC = TLI.getRegClassFor(ResVT.getSimpleVT());
8397 const TargetRegisterClass *ArgRC =
8398 TLI.getRegClassFor(Use->getOperand(0).getValueType().getSimpleVT());
8399 if (ArgRC == ResRC || !TLI.isOperationLegal(ISD::LOAD, ResVT))
8402 // At this point, we know that we perform a cross-register-bank copy.
8403 // Check if it is expensive.
8404 const TargetRegisterInfo *TRI = TLI.getTargetMachine().getRegisterInfo();
8405 // Assume bitcasts are cheap, unless both register classes do not
8406 // explicitly share a common sub class.
8407 if (!TRI || TRI->getCommonSubClass(ArgRC, ResRC))
8410 // Check if it will be merged with the load.
8411 // 1. Check the alignment constraint.
8412 unsigned RequiredAlignment = TLI.getDataLayout()->getABITypeAlignment(
8413 ResVT.getTypeForEVT(*DAG->getContext()));
8415 if (RequiredAlignment > getAlignment())
8418 // 2. Check that the load is a legal operation for that type.
8419 if (!TLI.isOperationLegal(ISD::LOAD, ResVT))
8422 // 3. Check that we do not have a zext in the way.
8423 if (Inst->getValueType(0) != getLoadedType())
8431 /// \brief Check that all bits set in \p UsedBits form a dense region, i.e.,
8432 /// \p UsedBits looks like 0..0 1..1 0..0.
8433 static bool areUsedBitsDense(const APInt &UsedBits) {
8434 // If all the bits are one, this is dense!
8435 if (UsedBits.isAllOnesValue())
8438 // Get rid of the unused bits on the right.
8439 APInt NarrowedUsedBits = UsedBits.lshr(UsedBits.countTrailingZeros());
8440 // Get rid of the unused bits on the left.
8441 if (NarrowedUsedBits.countLeadingZeros())
8442 NarrowedUsedBits = NarrowedUsedBits.trunc(NarrowedUsedBits.getActiveBits());
8443 // Check that the chunk of bits is completely used.
8444 return NarrowedUsedBits.isAllOnesValue();
8447 /// \brief Check whether or not \p First and \p Second are next to each other
8448 /// in memory. This means that there is no hole between the bits loaded
8449 /// by \p First and the bits loaded by \p Second.
8450 static bool areSlicesNextToEachOther(const LoadedSlice &First,
8451 const LoadedSlice &Second) {
8452 assert(First.Origin == Second.Origin && First.Origin &&
8453 "Unable to match different memory origins.");
8454 APInt UsedBits = First.getUsedBits();
8455 assert((UsedBits & Second.getUsedBits()) == 0 &&
8456 "Slices are not supposed to overlap.");
8457 UsedBits |= Second.getUsedBits();
8458 return areUsedBitsDense(UsedBits);
8461 /// \brief Adjust the \p GlobalLSCost according to the target
8462 /// paring capabilities and the layout of the slices.
8463 /// \pre \p GlobalLSCost should account for at least as many loads as
8464 /// there is in the slices in \p LoadedSlices.
8465 static void adjustCostForPairing(SmallVectorImpl<LoadedSlice> &LoadedSlices,
8466 LoadedSlice::Cost &GlobalLSCost) {
8467 unsigned NumberOfSlices = LoadedSlices.size();
8468 // If there is less than 2 elements, no pairing is possible.
8469 if (NumberOfSlices < 2)
8472 // Sort the slices so that elements that are likely to be next to each
8473 // other in memory are next to each other in the list.
8474 std::sort(LoadedSlices.begin(), LoadedSlices.end(),
8475 [](const LoadedSlice &LHS, const LoadedSlice &RHS) {
8476 assert(LHS.Origin == RHS.Origin && "Different bases not implemented.");
8477 return LHS.getOffsetFromBase() < RHS.getOffsetFromBase();
8479 const TargetLowering &TLI = LoadedSlices[0].DAG->getTargetLoweringInfo();
8480 // First (resp. Second) is the first (resp. Second) potentially candidate
8481 // to be placed in a paired load.
8482 const LoadedSlice *First = nullptr;
8483 const LoadedSlice *Second = nullptr;
8484 for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice,
8485 // Set the beginning of the pair.
8488 Second = &LoadedSlices[CurrSlice];
8490 // If First is NULL, it means we start a new pair.
8491 // Get to the next slice.
8495 EVT LoadedType = First->getLoadedType();
8497 // If the types of the slices are different, we cannot pair them.
8498 if (LoadedType != Second->getLoadedType())
8501 // Check if the target supplies paired loads for this type.
8502 unsigned RequiredAlignment = 0;
8503 if (!TLI.hasPairedLoad(LoadedType, RequiredAlignment)) {
8504 // move to the next pair, this type is hopeless.
8508 // Check if we meet the alignment requirement.
8509 if (RequiredAlignment > First->getAlignment())
8512 // Check that both loads are next to each other in memory.
8513 if (!areSlicesNextToEachOther(*First, *Second))
8516 assert(GlobalLSCost.Loads > 0 && "We save more loads than we created!");
8517 --GlobalLSCost.Loads;
8518 // Move to the next pair.
8523 /// \brief Check the profitability of all involved LoadedSlice.
8524 /// Currently, it is considered profitable if there is exactly two
8525 /// involved slices (1) which are (2) next to each other in memory, and
8526 /// whose cost (\see LoadedSlice::Cost) is smaller than the original load (3).
8528 /// Note: The order of the elements in \p LoadedSlices may be modified, but not
8529 /// the elements themselves.
8531 /// FIXME: When the cost model will be mature enough, we can relax
8532 /// constraints (1) and (2).
8533 static bool isSlicingProfitable(SmallVectorImpl<LoadedSlice> &LoadedSlices,
8534 const APInt &UsedBits, bool ForCodeSize) {
8535 unsigned NumberOfSlices = LoadedSlices.size();
8536 if (StressLoadSlicing)
8537 return NumberOfSlices > 1;
8540 if (NumberOfSlices != 2)
8544 if (!areUsedBitsDense(UsedBits))
8548 LoadedSlice::Cost OrigCost(ForCodeSize), GlobalSlicingCost(ForCodeSize);
8549 // The original code has one big load.
8551 for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice) {
8552 const LoadedSlice &LS = LoadedSlices[CurrSlice];
8553 // Accumulate the cost of all the slices.
8554 LoadedSlice::Cost SliceCost(LS, ForCodeSize);
8555 GlobalSlicingCost += SliceCost;
8557 // Account as cost in the original configuration the gain obtained
8558 // with the current slices.
8559 OrigCost.addSliceGain(LS);
8562 // If the target supports paired load, adjust the cost accordingly.
8563 adjustCostForPairing(LoadedSlices, GlobalSlicingCost);
8564 return OrigCost > GlobalSlicingCost;
8567 /// \brief If the given load, \p LI, is used only by trunc or trunc(lshr)
8568 /// operations, split it in the various pieces being extracted.
8570 /// This sort of thing is introduced by SROA.
8571 /// This slicing takes care not to insert overlapping loads.
8572 /// \pre LI is a simple load (i.e., not an atomic or volatile load).
8573 bool DAGCombiner::SliceUpLoad(SDNode *N) {
8574 if (Level < AfterLegalizeDAG)
8577 LoadSDNode *LD = cast<LoadSDNode>(N);
8578 if (LD->isVolatile() || !ISD::isNormalLoad(LD) ||
8579 !LD->getValueType(0).isInteger())
8582 // Keep track of already used bits to detect overlapping values.
8583 // In that case, we will just abort the transformation.
8584 APInt UsedBits(LD->getValueSizeInBits(0), 0);
8586 SmallVector<LoadedSlice, 4> LoadedSlices;
8588 // Check if this load is used as several smaller chunks of bits.
8589 // Basically, look for uses in trunc or trunc(lshr) and record a new chain
8590 // of computation for each trunc.
8591 for (SDNode::use_iterator UI = LD->use_begin(), UIEnd = LD->use_end();
8592 UI != UIEnd; ++UI) {
8593 // Skip the uses of the chain.
8594 if (UI.getUse().getResNo() != 0)
8600 // Check if this is a trunc(lshr).
8601 if (User->getOpcode() == ISD::SRL && User->hasOneUse() &&
8602 isa<ConstantSDNode>(User->getOperand(1))) {
8603 Shift = cast<ConstantSDNode>(User->getOperand(1))->getZExtValue();
8604 User = *User->use_begin();
8607 // At this point, User is a Truncate, iff we encountered, trunc or
8609 if (User->getOpcode() != ISD::TRUNCATE)
8612 // The width of the type must be a power of 2 and greater than 8-bits.
8613 // Otherwise the load cannot be represented in LLVM IR.
8614 // Moreover, if we shifted with a non-8-bits multiple, the slice
8615 // will be across several bytes. We do not support that.
8616 unsigned Width = User->getValueSizeInBits(0);
8617 if (Width < 8 || !isPowerOf2_32(Width) || (Shift & 0x7))
8620 // Build the slice for this chain of computations.
8621 LoadedSlice LS(User, LD, Shift, &DAG);
8622 APInt CurrentUsedBits = LS.getUsedBits();
8624 // Check if this slice overlaps with another.
8625 if ((CurrentUsedBits & UsedBits) != 0)
8627 // Update the bits used globally.
8628 UsedBits |= CurrentUsedBits;
8630 // Check if the new slice would be legal.
8634 // Record the slice.
8635 LoadedSlices.push_back(LS);
8638 // Abort slicing if it does not seem to be profitable.
8639 if (!isSlicingProfitable(LoadedSlices, UsedBits, ForCodeSize))
8644 // Rewrite each chain to use an independent load.
8645 // By construction, each chain can be represented by a unique load.
8647 // Prepare the argument for the new token factor for all the slices.
8648 SmallVector<SDValue, 8> ArgChains;
8649 for (SmallVectorImpl<LoadedSlice>::const_iterator
8650 LSIt = LoadedSlices.begin(),
8651 LSItEnd = LoadedSlices.end();
8652 LSIt != LSItEnd; ++LSIt) {
8653 SDValue SliceInst = LSIt->loadSlice();
8654 CombineTo(LSIt->Inst, SliceInst, true);
8655 if (SliceInst.getNode()->getOpcode() != ISD::LOAD)
8656 SliceInst = SliceInst.getOperand(0);
8657 assert(SliceInst->getOpcode() == ISD::LOAD &&
8658 "It takes more than a zext to get to the loaded slice!!");
8659 ArgChains.push_back(SliceInst.getValue(1));
8662 SDValue Chain = DAG.getNode(ISD::TokenFactor, SDLoc(LD), MVT::Other,
8664 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
8668 /// CheckForMaskedLoad - Check to see if V is (and load (ptr), imm), where the
8669 /// load is having specific bytes cleared out. If so, return the byte size
8670 /// being masked out and the shift amount.
8671 static std::pair<unsigned, unsigned>
8672 CheckForMaskedLoad(SDValue V, SDValue Ptr, SDValue Chain) {
8673 std::pair<unsigned, unsigned> Result(0, 0);
8675 // Check for the structure we're looking for.
8676 if (V->getOpcode() != ISD::AND ||
8677 !isa<ConstantSDNode>(V->getOperand(1)) ||
8678 !ISD::isNormalLoad(V->getOperand(0).getNode()))
8681 // Check the chain and pointer.
8682 LoadSDNode *LD = cast<LoadSDNode>(V->getOperand(0));
8683 if (LD->getBasePtr() != Ptr) return Result; // Not from same pointer.
8685 // The store should be chained directly to the load or be an operand of a
8687 if (LD == Chain.getNode())
8689 else if (Chain->getOpcode() != ISD::TokenFactor)
8690 return Result; // Fail.
8693 for (unsigned i = 0, e = Chain->getNumOperands(); i != e; ++i)
8694 if (Chain->getOperand(i).getNode() == LD) {
8698 if (!isOk) return Result;
8701 // This only handles simple types.
8702 if (V.getValueType() != MVT::i16 &&
8703 V.getValueType() != MVT::i32 &&
8704 V.getValueType() != MVT::i64)
8707 // Check the constant mask. Invert it so that the bits being masked out are
8708 // 0 and the bits being kept are 1. Use getSExtValue so that leading bits
8709 // follow the sign bit for uniformity.
8710 uint64_t NotMask = ~cast<ConstantSDNode>(V->getOperand(1))->getSExtValue();
8711 unsigned NotMaskLZ = countLeadingZeros(NotMask);
8712 if (NotMaskLZ & 7) return Result; // Must be multiple of a byte.
8713 unsigned NotMaskTZ = countTrailingZeros(NotMask);
8714 if (NotMaskTZ & 7) return Result; // Must be multiple of a byte.
8715 if (NotMaskLZ == 64) return Result; // All zero mask.
8717 // See if we have a continuous run of bits. If so, we have 0*1+0*
8718 if (CountTrailingOnes_64(NotMask >> NotMaskTZ)+NotMaskTZ+NotMaskLZ != 64)
8721 // Adjust NotMaskLZ down to be from the actual size of the int instead of i64.
8722 if (V.getValueType() != MVT::i64 && NotMaskLZ)
8723 NotMaskLZ -= 64-V.getValueSizeInBits();
8725 unsigned MaskedBytes = (V.getValueSizeInBits()-NotMaskLZ-NotMaskTZ)/8;
8726 switch (MaskedBytes) {
8730 default: return Result; // All one mask, or 5-byte mask.
8733 // Verify that the first bit starts at a multiple of mask so that the access
8734 // is aligned the same as the access width.
8735 if (NotMaskTZ && NotMaskTZ/8 % MaskedBytes) return Result;
8737 Result.first = MaskedBytes;
8738 Result.second = NotMaskTZ/8;
8743 /// ShrinkLoadReplaceStoreWithStore - Check to see if IVal is something that
8744 /// provides a value as specified by MaskInfo. If so, replace the specified
8745 /// store with a narrower store of truncated IVal.
8747 ShrinkLoadReplaceStoreWithStore(const std::pair<unsigned, unsigned> &MaskInfo,
8748 SDValue IVal, StoreSDNode *St,
8750 unsigned NumBytes = MaskInfo.first;
8751 unsigned ByteShift = MaskInfo.second;
8752 SelectionDAG &DAG = DC->getDAG();
8754 // Check to see if IVal is all zeros in the part being masked in by the 'or'
8755 // that uses this. If not, this is not a replacement.
8756 APInt Mask = ~APInt::getBitsSet(IVal.getValueSizeInBits(),
8757 ByteShift*8, (ByteShift+NumBytes)*8);
8758 if (!DAG.MaskedValueIsZero(IVal, Mask)) return nullptr;
8760 // Check that it is legal on the target to do this. It is legal if the new
8761 // VT we're shrinking to (i8/i16/i32) is legal or we're still before type
8763 MVT VT = MVT::getIntegerVT(NumBytes*8);
8764 if (!DC->isTypeLegal(VT))
8767 // Okay, we can do this! Replace the 'St' store with a store of IVal that is
8768 // shifted by ByteShift and truncated down to NumBytes.
8770 IVal = DAG.getNode(ISD::SRL, SDLoc(IVal), IVal.getValueType(), IVal,
8771 DAG.getConstant(ByteShift*8,
8772 DC->getShiftAmountTy(IVal.getValueType())));
8774 // Figure out the offset for the store and the alignment of the access.
8776 unsigned NewAlign = St->getAlignment();
8778 if (DAG.getTargetLoweringInfo().isLittleEndian())
8779 StOffset = ByteShift;
8781 StOffset = IVal.getValueType().getStoreSize() - ByteShift - NumBytes;
8783 SDValue Ptr = St->getBasePtr();
8785 Ptr = DAG.getNode(ISD::ADD, SDLoc(IVal), Ptr.getValueType(),
8786 Ptr, DAG.getConstant(StOffset, Ptr.getValueType()));
8787 NewAlign = MinAlign(NewAlign, StOffset);
8790 // Truncate down to the new size.
8791 IVal = DAG.getNode(ISD::TRUNCATE, SDLoc(IVal), VT, IVal);
8794 return DAG.getStore(St->getChain(), SDLoc(St), IVal, Ptr,
8795 St->getPointerInfo().getWithOffset(StOffset),
8796 false, false, NewAlign).getNode();
8800 /// ReduceLoadOpStoreWidth - Look for sequence of load / op / store where op is
8801 /// one of 'or', 'xor', and 'and' of immediates. If 'op' is only touching some
8802 /// of the loaded bits, try narrowing the load and store if it would end up
8803 /// being a win for performance or code size.
8804 SDValue DAGCombiner::ReduceLoadOpStoreWidth(SDNode *N) {
8805 StoreSDNode *ST = cast<StoreSDNode>(N);
8806 if (ST->isVolatile())
8809 SDValue Chain = ST->getChain();
8810 SDValue Value = ST->getValue();
8811 SDValue Ptr = ST->getBasePtr();
8812 EVT VT = Value.getValueType();
8814 if (ST->isTruncatingStore() || VT.isVector() || !Value.hasOneUse())
8817 unsigned Opc = Value.getOpcode();
8819 // If this is "store (or X, Y), P" and X is "(and (load P), cst)", where cst
8820 // is a byte mask indicating a consecutive number of bytes, check to see if
8821 // Y is known to provide just those bytes. If so, we try to replace the
8822 // load + replace + store sequence with a single (narrower) store, which makes
8824 if (Opc == ISD::OR) {
8825 std::pair<unsigned, unsigned> MaskedLoad;
8826 MaskedLoad = CheckForMaskedLoad(Value.getOperand(0), Ptr, Chain);
8827 if (MaskedLoad.first)
8828 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
8829 Value.getOperand(1), ST,this))
8830 return SDValue(NewST, 0);
8832 // Or is commutative, so try swapping X and Y.
8833 MaskedLoad = CheckForMaskedLoad(Value.getOperand(1), Ptr, Chain);
8834 if (MaskedLoad.first)
8835 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
8836 Value.getOperand(0), ST,this))
8837 return SDValue(NewST, 0);
8840 if ((Opc != ISD::OR && Opc != ISD::XOR && Opc != ISD::AND) ||
8841 Value.getOperand(1).getOpcode() != ISD::Constant)
8844 SDValue N0 = Value.getOperand(0);
8845 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
8846 Chain == SDValue(N0.getNode(), 1)) {
8847 LoadSDNode *LD = cast<LoadSDNode>(N0);
8848 if (LD->getBasePtr() != Ptr ||
8849 LD->getPointerInfo().getAddrSpace() !=
8850 ST->getPointerInfo().getAddrSpace())
8853 // Find the type to narrow it the load / op / store to.
8854 SDValue N1 = Value.getOperand(1);
8855 unsigned BitWidth = N1.getValueSizeInBits();
8856 APInt Imm = cast<ConstantSDNode>(N1)->getAPIntValue();
8857 if (Opc == ISD::AND)
8858 Imm ^= APInt::getAllOnesValue(BitWidth);
8859 if (Imm == 0 || Imm.isAllOnesValue())
8861 unsigned ShAmt = Imm.countTrailingZeros();
8862 unsigned MSB = BitWidth - Imm.countLeadingZeros() - 1;
8863 unsigned NewBW = NextPowerOf2(MSB - ShAmt);
8864 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
8865 while (NewBW < BitWidth &&
8866 !(TLI.isOperationLegalOrCustom(Opc, NewVT) &&
8867 TLI.isNarrowingProfitable(VT, NewVT))) {
8868 NewBW = NextPowerOf2(NewBW);
8869 NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
8871 if (NewBW >= BitWidth)
8874 // If the lsb changed does not start at the type bitwidth boundary,
8875 // start at the previous one.
8877 ShAmt = (((ShAmt + NewBW - 1) / NewBW) * NewBW) - NewBW;
8878 APInt Mask = APInt::getBitsSet(BitWidth, ShAmt,
8879 std::min(BitWidth, ShAmt + NewBW));
8880 if ((Imm & Mask) == Imm) {
8881 APInt NewImm = (Imm & Mask).lshr(ShAmt).trunc(NewBW);
8882 if (Opc == ISD::AND)
8883 NewImm ^= APInt::getAllOnesValue(NewBW);
8884 uint64_t PtrOff = ShAmt / 8;
8885 // For big endian targets, we need to adjust the offset to the pointer to
8886 // load the correct bytes.
8887 if (TLI.isBigEndian())
8888 PtrOff = (BitWidth + 7 - NewBW) / 8 - PtrOff;
8890 unsigned NewAlign = MinAlign(LD->getAlignment(), PtrOff);
8891 Type *NewVTTy = NewVT.getTypeForEVT(*DAG.getContext());
8892 if (NewAlign < TLI.getDataLayout()->getABITypeAlignment(NewVTTy))
8895 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(LD),
8896 Ptr.getValueType(), Ptr,
8897 DAG.getConstant(PtrOff, Ptr.getValueType()));
8898 SDValue NewLD = DAG.getLoad(NewVT, SDLoc(N0),
8899 LD->getChain(), NewPtr,
8900 LD->getPointerInfo().getWithOffset(PtrOff),
8901 LD->isVolatile(), LD->isNonTemporal(),
8902 LD->isInvariant(), NewAlign,
8904 SDValue NewVal = DAG.getNode(Opc, SDLoc(Value), NewVT, NewLD,
8905 DAG.getConstant(NewImm, NewVT));
8906 SDValue NewST = DAG.getStore(Chain, SDLoc(N),
8908 ST->getPointerInfo().getWithOffset(PtrOff),
8909 false, false, NewAlign);
8911 AddToWorklist(NewPtr.getNode());
8912 AddToWorklist(NewLD.getNode());
8913 AddToWorklist(NewVal.getNode());
8914 WorklistRemover DeadNodes(*this);
8915 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLD.getValue(1));
8924 /// TransformFPLoadStorePair - For a given floating point load / store pair,
8925 /// if the load value isn't used by any other operations, then consider
8926 /// transforming the pair to integer load / store operations if the target
8927 /// deems the transformation profitable.
8928 SDValue DAGCombiner::TransformFPLoadStorePair(SDNode *N) {
8929 StoreSDNode *ST = cast<StoreSDNode>(N);
8930 SDValue Chain = ST->getChain();
8931 SDValue Value = ST->getValue();
8932 if (ISD::isNormalStore(ST) && ISD::isNormalLoad(Value.getNode()) &&
8933 Value.hasOneUse() &&
8934 Chain == SDValue(Value.getNode(), 1)) {
8935 LoadSDNode *LD = cast<LoadSDNode>(Value);
8936 EVT VT = LD->getMemoryVT();
8937 if (!VT.isFloatingPoint() ||
8938 VT != ST->getMemoryVT() ||
8939 LD->isNonTemporal() ||
8940 ST->isNonTemporal() ||
8941 LD->getPointerInfo().getAddrSpace() != 0 ||
8942 ST->getPointerInfo().getAddrSpace() != 0)
8945 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
8946 if (!TLI.isOperationLegal(ISD::LOAD, IntVT) ||
8947 !TLI.isOperationLegal(ISD::STORE, IntVT) ||
8948 !TLI.isDesirableToTransformToIntegerOp(ISD::LOAD, VT) ||
8949 !TLI.isDesirableToTransformToIntegerOp(ISD::STORE, VT))
8952 unsigned LDAlign = LD->getAlignment();
8953 unsigned STAlign = ST->getAlignment();
8954 Type *IntVTTy = IntVT.getTypeForEVT(*DAG.getContext());
8955 unsigned ABIAlign = TLI.getDataLayout()->getABITypeAlignment(IntVTTy);
8956 if (LDAlign < ABIAlign || STAlign < ABIAlign)
8959 SDValue NewLD = DAG.getLoad(IntVT, SDLoc(Value),
8960 LD->getChain(), LD->getBasePtr(),
8961 LD->getPointerInfo(),
8962 false, false, false, LDAlign);
8964 SDValue NewST = DAG.getStore(NewLD.getValue(1), SDLoc(N),
8965 NewLD, ST->getBasePtr(),
8966 ST->getPointerInfo(),
8967 false, false, STAlign);
8969 AddToWorklist(NewLD.getNode());
8970 AddToWorklist(NewST.getNode());
8971 WorklistRemover DeadNodes(*this);
8972 DAG.ReplaceAllUsesOfValueWith(Value.getValue(1), NewLD.getValue(1));
8980 /// Helper struct to parse and store a memory address as base + index + offset.
8981 /// We ignore sign extensions when it is safe to do so.
8982 /// The following two expressions are not equivalent. To differentiate we need
8983 /// to store whether there was a sign extension involved in the index
8985 /// (load (i64 add (i64 copyfromreg %c)
8986 /// (i64 signextend (add (i8 load %index)
8990 /// (load (i64 add (i64 copyfromreg %c)
8991 /// (i64 signextend (i32 add (i32 signextend (i8 load %index))
8993 struct BaseIndexOffset {
8997 bool IsIndexSignExt;
8999 BaseIndexOffset() : Offset(0), IsIndexSignExt(false) {}
9001 BaseIndexOffset(SDValue Base, SDValue Index, int64_t Offset,
9002 bool IsIndexSignExt) :
9003 Base(Base), Index(Index), Offset(Offset), IsIndexSignExt(IsIndexSignExt) {}
9005 bool equalBaseIndex(const BaseIndexOffset &Other) {
9006 return Other.Base == Base && Other.Index == Index &&
9007 Other.IsIndexSignExt == IsIndexSignExt;
9010 /// Parses tree in Ptr for base, index, offset addresses.
9011 static BaseIndexOffset match(SDValue Ptr) {
9012 bool IsIndexSignExt = false;
9014 // We only can pattern match BASE + INDEX + OFFSET. If Ptr is not an ADD
9015 // instruction, then it could be just the BASE or everything else we don't
9016 // know how to handle. Just use Ptr as BASE and give up.
9017 if (Ptr->getOpcode() != ISD::ADD)
9018 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
9020 // We know that we have at least an ADD instruction. Try to pattern match
9021 // the simple case of BASE + OFFSET.
9022 if (isa<ConstantSDNode>(Ptr->getOperand(1))) {
9023 int64_t Offset = cast<ConstantSDNode>(Ptr->getOperand(1))->getSExtValue();
9024 return BaseIndexOffset(Ptr->getOperand(0), SDValue(), Offset,
9028 // Inside a loop the current BASE pointer is calculated using an ADD and a
9029 // MUL instruction. In this case Ptr is the actual BASE pointer.
9030 // (i64 add (i64 %array_ptr)
9031 // (i64 mul (i64 %induction_var)
9032 // (i64 %element_size)))
9033 if (Ptr->getOperand(1)->getOpcode() == ISD::MUL)
9034 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
9036 // Look at Base + Index + Offset cases.
9037 SDValue Base = Ptr->getOperand(0);
9038 SDValue IndexOffset = Ptr->getOperand(1);
9040 // Skip signextends.
9041 if (IndexOffset->getOpcode() == ISD::SIGN_EXTEND) {
9042 IndexOffset = IndexOffset->getOperand(0);
9043 IsIndexSignExt = true;
9046 // Either the case of Base + Index (no offset) or something else.
9047 if (IndexOffset->getOpcode() != ISD::ADD)
9048 return BaseIndexOffset(Base, IndexOffset, 0, IsIndexSignExt);
9050 // Now we have the case of Base + Index + offset.
9051 SDValue Index = IndexOffset->getOperand(0);
9052 SDValue Offset = IndexOffset->getOperand(1);
9054 if (!isa<ConstantSDNode>(Offset))
9055 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
9057 // Ignore signextends.
9058 if (Index->getOpcode() == ISD::SIGN_EXTEND) {
9059 Index = Index->getOperand(0);
9060 IsIndexSignExt = true;
9061 } else IsIndexSignExt = false;
9063 int64_t Off = cast<ConstantSDNode>(Offset)->getSExtValue();
9064 return BaseIndexOffset(Base, Index, Off, IsIndexSignExt);
9068 /// Holds a pointer to an LSBaseSDNode as well as information on where it
9069 /// is located in a sequence of memory operations connected by a chain.
9071 MemOpLink (LSBaseSDNode *N, int64_t Offset, unsigned Seq):
9072 MemNode(N), OffsetFromBase(Offset), SequenceNum(Seq) { }
9073 // Ptr to the mem node.
9074 LSBaseSDNode *MemNode;
9075 // Offset from the base ptr.
9076 int64_t OffsetFromBase;
9077 // What is the sequence number of this mem node.
9078 // Lowest mem operand in the DAG starts at zero.
9079 unsigned SequenceNum;
9082 bool DAGCombiner::MergeConsecutiveStores(StoreSDNode* St) {
9083 EVT MemVT = St->getMemoryVT();
9084 int64_t ElementSizeBytes = MemVT.getSizeInBits()/8;
9085 bool NoVectors = DAG.getMachineFunction().getFunction()->getAttributes().
9086 hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat);
9088 // Don't merge vectors into wider inputs.
9089 if (MemVT.isVector() || !MemVT.isSimple())
9092 // Perform an early exit check. Do not bother looking at stored values that
9093 // are not constants or loads.
9094 SDValue StoredVal = St->getValue();
9095 bool IsLoadSrc = isa<LoadSDNode>(StoredVal);
9096 if (!isa<ConstantSDNode>(StoredVal) && !isa<ConstantFPSDNode>(StoredVal) &&
9100 // Only look at ends of store sequences.
9101 SDValue Chain = SDValue(St, 0);
9102 if (Chain->hasOneUse() && Chain->use_begin()->getOpcode() == ISD::STORE)
9105 // This holds the base pointer, index, and the offset in bytes from the base
9107 BaseIndexOffset BasePtr = BaseIndexOffset::match(St->getBasePtr());
9109 // We must have a base and an offset.
9110 if (!BasePtr.Base.getNode())
9113 // Do not handle stores to undef base pointers.
9114 if (BasePtr.Base.getOpcode() == ISD::UNDEF)
9117 // Save the LoadSDNodes that we find in the chain.
9118 // We need to make sure that these nodes do not interfere with
9119 // any of the store nodes.
9120 SmallVector<LSBaseSDNode*, 8> AliasLoadNodes;
9122 // Save the StoreSDNodes that we find in the chain.
9123 SmallVector<MemOpLink, 8> StoreNodes;
9125 // Walk up the chain and look for nodes with offsets from the same
9126 // base pointer. Stop when reaching an instruction with a different kind
9127 // or instruction which has a different base pointer.
9129 StoreSDNode *Index = St;
9131 // If the chain has more than one use, then we can't reorder the mem ops.
9132 if (Index != St && !SDValue(Index, 0)->hasOneUse())
9135 // Find the base pointer and offset for this memory node.
9136 BaseIndexOffset Ptr = BaseIndexOffset::match(Index->getBasePtr());
9138 // Check that the base pointer is the same as the original one.
9139 if (!Ptr.equalBaseIndex(BasePtr))
9142 // Check that the alignment is the same.
9143 if (Index->getAlignment() != St->getAlignment())
9146 // The memory operands must not be volatile.
9147 if (Index->isVolatile() || Index->isIndexed())
9151 if (StoreSDNode *St = dyn_cast<StoreSDNode>(Index))
9152 if (St->isTruncatingStore())
9155 // The stored memory type must be the same.
9156 if (Index->getMemoryVT() != MemVT)
9159 // We do not allow unaligned stores because we want to prevent overriding
9161 if (Index->getAlignment()*8 != MemVT.getSizeInBits())
9164 // We found a potential memory operand to merge.
9165 StoreNodes.push_back(MemOpLink(Index, Ptr.Offset, Seq++));
9167 // Find the next memory operand in the chain. If the next operand in the
9168 // chain is a store then move up and continue the scan with the next
9169 // memory operand. If the next operand is a load save it and use alias
9170 // information to check if it interferes with anything.
9171 SDNode *NextInChain = Index->getChain().getNode();
9173 if (StoreSDNode *STn = dyn_cast<StoreSDNode>(NextInChain)) {
9174 // We found a store node. Use it for the next iteration.
9177 } else if (LoadSDNode *Ldn = dyn_cast<LoadSDNode>(NextInChain)) {
9178 if (Ldn->isVolatile()) {
9183 // Save the load node for later. Continue the scan.
9184 AliasLoadNodes.push_back(Ldn);
9185 NextInChain = Ldn->getChain().getNode();
9194 // Check if there is anything to merge.
9195 if (StoreNodes.size() < 2)
9198 // Sort the memory operands according to their distance from the base pointer.
9199 std::sort(StoreNodes.begin(), StoreNodes.end(),
9200 [](MemOpLink LHS, MemOpLink RHS) {
9201 return LHS.OffsetFromBase < RHS.OffsetFromBase ||
9202 (LHS.OffsetFromBase == RHS.OffsetFromBase &&
9203 LHS.SequenceNum > RHS.SequenceNum);
9206 // Scan the memory operations on the chain and find the first non-consecutive
9207 // store memory address.
9208 unsigned LastConsecutiveStore = 0;
9209 int64_t StartAddress = StoreNodes[0].OffsetFromBase;
9210 for (unsigned i = 0, e = StoreNodes.size(); i < e; ++i) {
9212 // Check that the addresses are consecutive starting from the second
9213 // element in the list of stores.
9215 int64_t CurrAddress = StoreNodes[i].OffsetFromBase;
9216 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
9221 // Check if this store interferes with any of the loads that we found.
9222 for (unsigned ld = 0, lde = AliasLoadNodes.size(); ld < lde; ++ld)
9223 if (isAlias(AliasLoadNodes[ld], StoreNodes[i].MemNode)) {
9227 // We found a load that alias with this store. Stop the sequence.
9231 // Mark this node as useful.
9232 LastConsecutiveStore = i;
9235 // The node with the lowest store address.
9236 LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode;
9238 // Store the constants into memory as one consecutive store.
9240 unsigned LastLegalType = 0;
9241 unsigned LastLegalVectorType = 0;
9242 bool NonZero = false;
9243 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
9244 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
9245 SDValue StoredVal = St->getValue();
9247 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(StoredVal)) {
9248 NonZero |= !C->isNullValue();
9249 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(StoredVal)) {
9250 NonZero |= !C->getConstantFPValue()->isNullValue();
9256 // Find a legal type for the constant store.
9257 unsigned StoreBW = (i+1) * ElementSizeBytes * 8;
9258 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
9259 if (TLI.isTypeLegal(StoreTy))
9260 LastLegalType = i+1;
9261 // Or check whether a truncstore is legal.
9262 else if (TLI.getTypeAction(*DAG.getContext(), StoreTy) ==
9263 TargetLowering::TypePromoteInteger) {
9264 EVT LegalizedStoredValueTy =
9265 TLI.getTypeToTransformTo(*DAG.getContext(), StoredVal.getValueType());
9266 if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy))
9267 LastLegalType = i+1;
9270 // Find a legal type for the vector store.
9271 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
9272 if (TLI.isTypeLegal(Ty))
9273 LastLegalVectorType = i + 1;
9276 // We only use vectors if the constant is known to be zero and the
9277 // function is not marked with the noimplicitfloat attribute.
9278 if (NonZero || NoVectors)
9279 LastLegalVectorType = 0;
9281 // Check if we found a legal integer type to store.
9282 if (LastLegalType == 0 && LastLegalVectorType == 0)
9285 bool UseVector = (LastLegalVectorType > LastLegalType) && !NoVectors;
9286 unsigned NumElem = UseVector ? LastLegalVectorType : LastLegalType;
9288 // Make sure we have something to merge.
9292 unsigned EarliestNodeUsed = 0;
9293 for (unsigned i=0; i < NumElem; ++i) {
9294 // Find a chain for the new wide-store operand. Notice that some
9295 // of the store nodes that we found may not be selected for inclusion
9296 // in the wide store. The chain we use needs to be the chain of the
9297 // earliest store node which is *used* and replaced by the wide store.
9298 if (StoreNodes[i].SequenceNum > StoreNodes[EarliestNodeUsed].SequenceNum)
9299 EarliestNodeUsed = i;
9302 // The earliest Node in the DAG.
9303 LSBaseSDNode *EarliestOp = StoreNodes[EarliestNodeUsed].MemNode;
9304 SDLoc DL(StoreNodes[0].MemNode);
9308 // Find a legal type for the vector store.
9309 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
9310 assert(TLI.isTypeLegal(Ty) && "Illegal vector store");
9311 StoredVal = DAG.getConstant(0, Ty);
9313 unsigned StoreBW = NumElem * ElementSizeBytes * 8;
9314 APInt StoreInt(StoreBW, 0);
9316 // Construct a single integer constant which is made of the smaller
9318 bool IsLE = TLI.isLittleEndian();
9319 for (unsigned i = 0; i < NumElem ; ++i) {
9320 unsigned Idx = IsLE ?(NumElem - 1 - i) : i;
9321 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[Idx].MemNode);
9322 SDValue Val = St->getValue();
9323 StoreInt<<=ElementSizeBytes*8;
9324 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val)) {
9325 StoreInt|=C->getAPIntValue().zext(StoreBW);
9326 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Val)) {
9327 StoreInt|= C->getValueAPF().bitcastToAPInt().zext(StoreBW);
9329 assert(false && "Invalid constant element type");
9333 // Create the new Load and Store operations.
9334 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
9335 StoredVal = DAG.getConstant(StoreInt, StoreTy);
9338 SDValue NewStore = DAG.getStore(EarliestOp->getChain(), DL, StoredVal,
9339 FirstInChain->getBasePtr(),
9340 FirstInChain->getPointerInfo(),
9342 FirstInChain->getAlignment());
9344 // Replace the first store with the new store
9345 CombineTo(EarliestOp, NewStore);
9346 // Erase all other stores.
9347 for (unsigned i = 0; i < NumElem ; ++i) {
9348 if (StoreNodes[i].MemNode == EarliestOp)
9350 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
9351 // ReplaceAllUsesWith will replace all uses that existed when it was
9352 // called, but graph optimizations may cause new ones to appear. For
9353 // example, the case in pr14333 looks like
9355 // St's chain -> St -> another store -> X
9357 // And the only difference from St to the other store is the chain.
9358 // When we change it's chain to be St's chain they become identical,
9359 // get CSEed and the net result is that X is now a use of St.
9360 // Since we know that St is redundant, just iterate.
9361 while (!St->use_empty())
9362 DAG.ReplaceAllUsesWith(SDValue(St, 0), St->getChain());
9363 removeFromWorklist(St);
9370 // Below we handle the case of multiple consecutive stores that
9371 // come from multiple consecutive loads. We merge them into a single
9372 // wide load and a single wide store.
9374 // Look for load nodes which are used by the stored values.
9375 SmallVector<MemOpLink, 8> LoadNodes;
9377 // Find acceptable loads. Loads need to have the same chain (token factor),
9378 // must not be zext, volatile, indexed, and they must be consecutive.
9379 BaseIndexOffset LdBasePtr;
9380 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
9381 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
9382 LoadSDNode *Ld = dyn_cast<LoadSDNode>(St->getValue());
9385 // Loads must only have one use.
9386 if (!Ld->hasNUsesOfValue(1, 0))
9389 // Check that the alignment is the same as the stores.
9390 if (Ld->getAlignment() != St->getAlignment())
9393 // The memory operands must not be volatile.
9394 if (Ld->isVolatile() || Ld->isIndexed())
9397 // We do not accept ext loads.
9398 if (Ld->getExtensionType() != ISD::NON_EXTLOAD)
9401 // The stored memory type must be the same.
9402 if (Ld->getMemoryVT() != MemVT)
9405 BaseIndexOffset LdPtr = BaseIndexOffset::match(Ld->getBasePtr());
9406 // If this is not the first ptr that we check.
9407 if (LdBasePtr.Base.getNode()) {
9408 // The base ptr must be the same.
9409 if (!LdPtr.equalBaseIndex(LdBasePtr))
9412 // Check that all other base pointers are the same as this one.
9416 // We found a potential memory operand to merge.
9417 LoadNodes.push_back(MemOpLink(Ld, LdPtr.Offset, 0));
9420 if (LoadNodes.size() < 2)
9423 // Scan the memory operations on the chain and find the first non-consecutive
9424 // load memory address. These variables hold the index in the store node
9426 unsigned LastConsecutiveLoad = 0;
9427 // This variable refers to the size and not index in the array.
9428 unsigned LastLegalVectorType = 0;
9429 unsigned LastLegalIntegerType = 0;
9430 StartAddress = LoadNodes[0].OffsetFromBase;
9431 SDValue FirstChain = LoadNodes[0].MemNode->getChain();
9432 for (unsigned i = 1; i < LoadNodes.size(); ++i) {
9433 // All loads much share the same chain.
9434 if (LoadNodes[i].MemNode->getChain() != FirstChain)
9437 int64_t CurrAddress = LoadNodes[i].OffsetFromBase;
9438 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
9440 LastConsecutiveLoad = i;
9442 // Find a legal type for the vector store.
9443 EVT StoreTy = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
9444 if (TLI.isTypeLegal(StoreTy))
9445 LastLegalVectorType = i + 1;
9447 // Find a legal type for the integer store.
9448 unsigned StoreBW = (i+1) * ElementSizeBytes * 8;
9449 StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
9450 if (TLI.isTypeLegal(StoreTy))
9451 LastLegalIntegerType = i + 1;
9452 // Or check whether a truncstore and extload is legal.
9453 else if (TLI.getTypeAction(*DAG.getContext(), StoreTy) ==
9454 TargetLowering::TypePromoteInteger) {
9455 EVT LegalizedStoredValueTy =
9456 TLI.getTypeToTransformTo(*DAG.getContext(), StoreTy);
9457 if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy) &&
9458 TLI.isLoadExtLegal(ISD::ZEXTLOAD, StoreTy) &&
9459 TLI.isLoadExtLegal(ISD::SEXTLOAD, StoreTy) &&
9460 TLI.isLoadExtLegal(ISD::EXTLOAD, StoreTy))
9461 LastLegalIntegerType = i+1;
9465 // Only use vector types if the vector type is larger than the integer type.
9466 // If they are the same, use integers.
9467 bool UseVectorTy = LastLegalVectorType > LastLegalIntegerType && !NoVectors;
9468 unsigned LastLegalType = std::max(LastLegalVectorType, LastLegalIntegerType);
9470 // We add +1 here because the LastXXX variables refer to location while
9471 // the NumElem refers to array/index size.
9472 unsigned NumElem = std::min(LastConsecutiveStore, LastConsecutiveLoad) + 1;
9473 NumElem = std::min(LastLegalType, NumElem);
9478 // The earliest Node in the DAG.
9479 unsigned EarliestNodeUsed = 0;
9480 LSBaseSDNode *EarliestOp = StoreNodes[EarliestNodeUsed].MemNode;
9481 for (unsigned i=1; i<NumElem; ++i) {
9482 // Find a chain for the new wide-store operand. Notice that some
9483 // of the store nodes that we found may not be selected for inclusion
9484 // in the wide store. The chain we use needs to be the chain of the
9485 // earliest store node which is *used* and replaced by the wide store.
9486 if (StoreNodes[i].SequenceNum > StoreNodes[EarliestNodeUsed].SequenceNum)
9487 EarliestNodeUsed = i;
9490 // Find if it is better to use vectors or integers to load and store
9494 JointMemOpVT = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
9496 unsigned StoreBW = NumElem * ElementSizeBytes * 8;
9497 JointMemOpVT = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
9500 SDLoc LoadDL(LoadNodes[0].MemNode);
9501 SDLoc StoreDL(StoreNodes[0].MemNode);
9503 LoadSDNode *FirstLoad = cast<LoadSDNode>(LoadNodes[0].MemNode);
9504 SDValue NewLoad = DAG.getLoad(JointMemOpVT, LoadDL,
9505 FirstLoad->getChain(),
9506 FirstLoad->getBasePtr(),
9507 FirstLoad->getPointerInfo(),
9508 false, false, false,
9509 FirstLoad->getAlignment());
9511 SDValue NewStore = DAG.getStore(EarliestOp->getChain(), StoreDL, NewLoad,
9512 FirstInChain->getBasePtr(),
9513 FirstInChain->getPointerInfo(), false, false,
9514 FirstInChain->getAlignment());
9516 // Replace one of the loads with the new load.
9517 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[0].MemNode);
9518 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1),
9519 SDValue(NewLoad.getNode(), 1));
9521 // Remove the rest of the load chains.
9522 for (unsigned i = 1; i < NumElem ; ++i) {
9523 // Replace all chain users of the old load nodes with the chain of the new
9525 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[i].MemNode);
9526 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Ld->getChain());
9529 // Replace the first store with the new store.
9530 CombineTo(EarliestOp, NewStore);
9531 // Erase all other stores.
9532 for (unsigned i = 0; i < NumElem ; ++i) {
9533 // Remove all Store nodes.
9534 if (StoreNodes[i].MemNode == EarliestOp)
9536 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
9537 DAG.ReplaceAllUsesOfValueWith(SDValue(St, 0), St->getChain());
9538 removeFromWorklist(St);
9545 SDValue DAGCombiner::visitSTORE(SDNode *N) {
9546 StoreSDNode *ST = cast<StoreSDNode>(N);
9547 SDValue Chain = ST->getChain();
9548 SDValue Value = ST->getValue();
9549 SDValue Ptr = ST->getBasePtr();
9551 // If this is a store of a bit convert, store the input value if the
9552 // resultant store does not need a higher alignment than the original.
9553 if (Value.getOpcode() == ISD::BITCAST && !ST->isTruncatingStore() &&
9554 ST->isUnindexed()) {
9555 unsigned OrigAlign = ST->getAlignment();
9556 EVT SVT = Value.getOperand(0).getValueType();
9557 unsigned Align = TLI.getDataLayout()->
9558 getABITypeAlignment(SVT.getTypeForEVT(*DAG.getContext()));
9559 if (Align <= OrigAlign &&
9560 ((!LegalOperations && !ST->isVolatile()) ||
9561 TLI.isOperationLegalOrCustom(ISD::STORE, SVT)))
9562 return DAG.getStore(Chain, SDLoc(N), Value.getOperand(0),
9563 Ptr, ST->getPointerInfo(), ST->isVolatile(),
9564 ST->isNonTemporal(), OrigAlign,
9568 // Turn 'store undef, Ptr' -> nothing.
9569 if (Value.getOpcode() == ISD::UNDEF && ST->isUnindexed())
9572 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
9573 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Value)) {
9574 // NOTE: If the original store is volatile, this transform must not increase
9575 // the number of stores. For example, on x86-32 an f64 can be stored in one
9576 // processor operation but an i64 (which is not legal) requires two. So the
9577 // transform should not be done in this case.
9578 if (Value.getOpcode() != ISD::TargetConstantFP) {
9580 switch (CFP->getSimpleValueType(0).SimpleTy) {
9581 default: llvm_unreachable("Unknown FP type");
9582 case MVT::f16: // We don't do this for these yet.
9588 if ((isTypeLegal(MVT::i32) && !LegalOperations && !ST->isVolatile()) ||
9589 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
9590 Tmp = DAG.getConstant((uint32_t)CFP->getValueAPF().
9591 bitcastToAPInt().getZExtValue(), MVT::i32);
9592 return DAG.getStore(Chain, SDLoc(N), Tmp,
9593 Ptr, ST->getMemOperand());
9597 if ((TLI.isTypeLegal(MVT::i64) && !LegalOperations &&
9598 !ST->isVolatile()) ||
9599 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i64)) {
9600 Tmp = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
9601 getZExtValue(), MVT::i64);
9602 return DAG.getStore(Chain, SDLoc(N), Tmp,
9603 Ptr, ST->getMemOperand());
9606 if (!ST->isVolatile() &&
9607 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
9608 // Many FP stores are not made apparent until after legalize, e.g. for
9609 // argument passing. Since this is so common, custom legalize the
9610 // 64-bit integer store into two 32-bit stores.
9611 uint64_t Val = CFP->getValueAPF().bitcastToAPInt().getZExtValue();
9612 SDValue Lo = DAG.getConstant(Val & 0xFFFFFFFF, MVT::i32);
9613 SDValue Hi = DAG.getConstant(Val >> 32, MVT::i32);
9614 if (TLI.isBigEndian()) std::swap(Lo, Hi);
9616 unsigned Alignment = ST->getAlignment();
9617 bool isVolatile = ST->isVolatile();
9618 bool isNonTemporal = ST->isNonTemporal();
9619 AAMDNodes AAInfo = ST->getAAInfo();
9621 SDValue St0 = DAG.getStore(Chain, SDLoc(ST), Lo,
9622 Ptr, ST->getPointerInfo(),
9623 isVolatile, isNonTemporal,
9624 ST->getAlignment(), AAInfo);
9625 Ptr = DAG.getNode(ISD::ADD, SDLoc(N), Ptr.getValueType(), Ptr,
9626 DAG.getConstant(4, Ptr.getValueType()));
9627 Alignment = MinAlign(Alignment, 4U);
9628 SDValue St1 = DAG.getStore(Chain, SDLoc(ST), Hi,
9629 Ptr, ST->getPointerInfo().getWithOffset(4),
9630 isVolatile, isNonTemporal,
9632 return DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other,
9641 // Try to infer better alignment information than the store already has.
9642 if (OptLevel != CodeGenOpt::None && ST->isUnindexed()) {
9643 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
9644 if (Align > ST->getAlignment())
9645 return DAG.getTruncStore(Chain, SDLoc(N), Value,
9646 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
9647 ST->isVolatile(), ST->isNonTemporal(), Align,
9652 // Try transforming a pair floating point load / store ops to integer
9653 // load / store ops.
9654 SDValue NewST = TransformFPLoadStorePair(N);
9655 if (NewST.getNode())
9658 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA :
9659 TLI.getTargetMachine().getSubtarget<TargetSubtargetInfo>().useAA();
9661 if (CombinerAAOnlyFunc.getNumOccurrences() &&
9662 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
9665 if (UseAA && ST->isUnindexed()) {
9666 // Walk up chain skipping non-aliasing memory nodes.
9667 SDValue BetterChain = FindBetterChain(N, Chain);
9669 // If there is a better chain.
9670 if (Chain != BetterChain) {
9673 // Replace the chain to avoid dependency.
9674 if (ST->isTruncatingStore()) {
9675 ReplStore = DAG.getTruncStore(BetterChain, SDLoc(N), Value, Ptr,
9676 ST->getMemoryVT(), ST->getMemOperand());
9678 ReplStore = DAG.getStore(BetterChain, SDLoc(N), Value, Ptr,
9679 ST->getMemOperand());
9682 // Create token to keep both nodes around.
9683 SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N),
9684 MVT::Other, Chain, ReplStore);
9686 // Make sure the new and old chains are cleaned up.
9687 AddToWorklist(Token.getNode());
9689 // Don't add users to work list.
9690 return CombineTo(N, Token, false);
9694 // Try transforming N to an indexed store.
9695 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
9696 return SDValue(N, 0);
9698 // FIXME: is there such a thing as a truncating indexed store?
9699 if (ST->isTruncatingStore() && ST->isUnindexed() &&
9700 Value.getValueType().isInteger()) {
9701 // See if we can simplify the input to this truncstore with knowledge that
9702 // only the low bits are being used. For example:
9703 // "truncstore (or (shl x, 8), y), i8" -> "truncstore y, i8"
9705 GetDemandedBits(Value,
9706 APInt::getLowBitsSet(
9707 Value.getValueType().getScalarType().getSizeInBits(),
9708 ST->getMemoryVT().getScalarType().getSizeInBits()));
9709 AddToWorklist(Value.getNode());
9710 if (Shorter.getNode())
9711 return DAG.getTruncStore(Chain, SDLoc(N), Shorter,
9712 Ptr, ST->getMemoryVT(), ST->getMemOperand());
9714 // Otherwise, see if we can simplify the operation with
9715 // SimplifyDemandedBits, which only works if the value has a single use.
9716 if (SimplifyDemandedBits(Value,
9717 APInt::getLowBitsSet(
9718 Value.getValueType().getScalarType().getSizeInBits(),
9719 ST->getMemoryVT().getScalarType().getSizeInBits())))
9720 return SDValue(N, 0);
9723 // If this is a load followed by a store to the same location, then the store
9725 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Value)) {
9726 if (Ld->getBasePtr() == Ptr && ST->getMemoryVT() == Ld->getMemoryVT() &&
9727 ST->isUnindexed() && !ST->isVolatile() &&
9728 // There can't be any side effects between the load and store, such as
9730 Chain.reachesChainWithoutSideEffects(SDValue(Ld, 1))) {
9731 // The store is dead, remove it.
9736 // If this is an FP_ROUND or TRUNC followed by a store, fold this into a
9737 // truncating store. We can do this even if this is already a truncstore.
9738 if ((Value.getOpcode() == ISD::FP_ROUND || Value.getOpcode() == ISD::TRUNCATE)
9739 && Value.getNode()->hasOneUse() && ST->isUnindexed() &&
9740 TLI.isTruncStoreLegal(Value.getOperand(0).getValueType(),
9741 ST->getMemoryVT())) {
9742 return DAG.getTruncStore(Chain, SDLoc(N), Value.getOperand(0),
9743 Ptr, ST->getMemoryVT(), ST->getMemOperand());
9746 // Only perform this optimization before the types are legal, because we
9747 // don't want to perform this optimization on every DAGCombine invocation.
9749 bool EverChanged = false;
9752 // There can be multiple store sequences on the same chain.
9753 // Keep trying to merge store sequences until we are unable to do so
9754 // or until we merge the last store on the chain.
9755 bool Changed = MergeConsecutiveStores(ST);
9756 EverChanged |= Changed;
9757 if (!Changed) break;
9758 } while (ST->getOpcode() != ISD::DELETED_NODE);
9761 return SDValue(N, 0);
9764 return ReduceLoadOpStoreWidth(N);
9767 SDValue DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) {
9768 SDValue InVec = N->getOperand(0);
9769 SDValue InVal = N->getOperand(1);
9770 SDValue EltNo = N->getOperand(2);
9773 // If the inserted element is an UNDEF, just use the input vector.
9774 if (InVal.getOpcode() == ISD::UNDEF)
9777 EVT VT = InVec.getValueType();
9779 // If we can't generate a legal BUILD_VECTOR, exit
9780 if (LegalOperations && !TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
9783 // Check that we know which element is being inserted
9784 if (!isa<ConstantSDNode>(EltNo))
9786 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
9788 // Canonicalize insert_vector_elt dag nodes.
9790 // (insert_vector_elt (insert_vector_elt A, Idx0), Idx1)
9791 // -> (insert_vector_elt (insert_vector_elt A, Idx1), Idx0)
9793 // Do this only if the child insert_vector node has one use; also
9794 // do this only if indices are both constants and Idx1 < Idx0.
9795 if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT && InVec.hasOneUse()
9796 && isa<ConstantSDNode>(InVec.getOperand(2))) {
9798 cast<ConstantSDNode>(InVec.getOperand(2))->getZExtValue();
9799 if (Elt < OtherElt) {
9801 SDValue NewOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(N), VT,
9802 InVec.getOperand(0), InVal, EltNo);
9803 AddToWorklist(NewOp.getNode());
9804 return DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(InVec.getNode()),
9805 VT, NewOp, InVec.getOperand(1), InVec.getOperand(2));
9809 // Check that the operand is a BUILD_VECTOR (or UNDEF, which can essentially
9810 // be converted to a BUILD_VECTOR). Fill in the Ops vector with the
9812 SmallVector<SDValue, 8> Ops;
9813 // Do not combine these two vectors if the output vector will not replace
9814 // the input vector.
9815 if (InVec.getOpcode() == ISD::BUILD_VECTOR && InVec.hasOneUse()) {
9816 Ops.append(InVec.getNode()->op_begin(),
9817 InVec.getNode()->op_end());
9818 } else if (InVec.getOpcode() == ISD::UNDEF) {
9819 unsigned NElts = VT.getVectorNumElements();
9820 Ops.append(NElts, DAG.getUNDEF(InVal.getValueType()));
9825 // Insert the element
9826 if (Elt < Ops.size()) {
9827 // All the operands of BUILD_VECTOR must have the same type;
9828 // we enforce that here.
9829 EVT OpVT = Ops[0].getValueType();
9830 if (InVal.getValueType() != OpVT)
9831 InVal = OpVT.bitsGT(InVal.getValueType()) ?
9832 DAG.getNode(ISD::ANY_EXTEND, dl, OpVT, InVal) :
9833 DAG.getNode(ISD::TRUNCATE, dl, OpVT, InVal);
9837 // Return the new vector
9838 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
9841 SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) {
9842 // (vextract (scalar_to_vector val, 0) -> val
9843 SDValue InVec = N->getOperand(0);
9844 EVT VT = InVec.getValueType();
9845 EVT NVT = N->getValueType(0);
9847 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
9848 // Check if the result type doesn't match the inserted element type. A
9849 // SCALAR_TO_VECTOR may truncate the inserted element and the
9850 // EXTRACT_VECTOR_ELT may widen the extracted vector.
9851 SDValue InOp = InVec.getOperand(0);
9852 if (InOp.getValueType() != NVT) {
9853 assert(InOp.getValueType().isInteger() && NVT.isInteger());
9854 return DAG.getSExtOrTrunc(InOp, SDLoc(InVec), NVT);
9859 SDValue EltNo = N->getOperand(1);
9860 bool ConstEltNo = isa<ConstantSDNode>(EltNo);
9862 // Transform: (EXTRACT_VECTOR_ELT( VECTOR_SHUFFLE )) -> EXTRACT_VECTOR_ELT.
9863 // We only perform this optimization before the op legalization phase because
9864 // we may introduce new vector instructions which are not backed by TD
9865 // patterns. For example on AVX, extracting elements from a wide vector
9866 // without using extract_subvector. However, if we can find an underlying
9867 // scalar value, then we can always use that.
9868 if (InVec.getOpcode() == ISD::VECTOR_SHUFFLE
9870 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
9871 int NumElem = VT.getVectorNumElements();
9872 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(InVec);
9873 // Find the new index to extract from.
9874 int OrigElt = SVOp->getMaskElt(Elt);
9876 // Extracting an undef index is undef.
9878 return DAG.getUNDEF(NVT);
9880 // Select the right vector half to extract from.
9882 if (OrigElt < NumElem) {
9883 SVInVec = InVec->getOperand(0);
9885 SVInVec = InVec->getOperand(1);
9889 if (SVInVec.getOpcode() == ISD::BUILD_VECTOR) {
9890 SDValue InOp = SVInVec.getOperand(OrigElt);
9891 if (InOp.getValueType() != NVT) {
9892 assert(InOp.getValueType().isInteger() && NVT.isInteger());
9893 InOp = DAG.getSExtOrTrunc(InOp, SDLoc(SVInVec), NVT);
9899 // FIXME: We should handle recursing on other vector shuffles and
9900 // scalar_to_vector here as well.
9902 if (!LegalOperations) {
9903 EVT IndexTy = TLI.getVectorIdxTy();
9904 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(N), NVT,
9905 SVInVec, DAG.getConstant(OrigElt, IndexTy));
9909 // Perform only after legalization to ensure build_vector / vector_shuffle
9910 // optimizations have already been done.
9911 if (!LegalOperations) return SDValue();
9913 // (vextract (v4f32 load $addr), c) -> (f32 load $addr+c*size)
9914 // (vextract (v4f32 s2v (f32 load $addr)), c) -> (f32 load $addr+c*size)
9915 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), 0) -> (f32 load $addr)
9918 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
9919 bool NewLoad = false;
9920 bool BCNumEltsChanged = false;
9921 EVT ExtVT = VT.getVectorElementType();
9924 // If the result of load has to be truncated, then it's not necessarily
9926 if (NVT.bitsLT(LVT) && !TLI.isTruncateFree(LVT, NVT))
9929 if (InVec.getOpcode() == ISD::BITCAST) {
9930 // Don't duplicate a load with other uses.
9931 if (!InVec.hasOneUse())
9934 EVT BCVT = InVec.getOperand(0).getValueType();
9935 if (!BCVT.isVector() || ExtVT.bitsGT(BCVT.getVectorElementType()))
9937 if (VT.getVectorNumElements() != BCVT.getVectorNumElements())
9938 BCNumEltsChanged = true;
9939 InVec = InVec.getOperand(0);
9940 ExtVT = BCVT.getVectorElementType();
9944 LoadSDNode *LN0 = nullptr;
9945 const ShuffleVectorSDNode *SVN = nullptr;
9946 if (ISD::isNormalLoad(InVec.getNode())) {
9947 LN0 = cast<LoadSDNode>(InVec);
9948 } else if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR &&
9949 InVec.getOperand(0).getValueType() == ExtVT &&
9950 ISD::isNormalLoad(InVec.getOperand(0).getNode())) {
9951 // Don't duplicate a load with other uses.
9952 if (!InVec.hasOneUse())
9955 LN0 = cast<LoadSDNode>(InVec.getOperand(0));
9956 } else if ((SVN = dyn_cast<ShuffleVectorSDNode>(InVec))) {
9957 // (vextract (vector_shuffle (load $addr), v2, <1, u, u, u>), 1)
9959 // (load $addr+1*size)
9961 // Don't duplicate a load with other uses.
9962 if (!InVec.hasOneUse())
9965 // If the bit convert changed the number of elements, it is unsafe
9966 // to examine the mask.
9967 if (BCNumEltsChanged)
9970 // Select the input vector, guarding against out of range extract vector.
9971 unsigned NumElems = VT.getVectorNumElements();
9972 int Idx = (Elt > (int)NumElems) ? -1 : SVN->getMaskElt(Elt);
9973 InVec = (Idx < (int)NumElems) ? InVec.getOperand(0) : InVec.getOperand(1);
9975 if (InVec.getOpcode() == ISD::BITCAST) {
9976 // Don't duplicate a load with other uses.
9977 if (!InVec.hasOneUse())
9980 InVec = InVec.getOperand(0);
9982 if (ISD::isNormalLoad(InVec.getNode())) {
9983 LN0 = cast<LoadSDNode>(InVec);
9984 Elt = (Idx < (int)NumElems) ? Idx : Idx - (int)NumElems;
9988 // Make sure we found a non-volatile load and the extractelement is
9990 if (!LN0 || !LN0->hasNUsesOfValue(1,0) || LN0->isVolatile())
9993 // If Idx was -1 above, Elt is going to be -1, so just return undef.
9995 return DAG.getUNDEF(LVT);
9997 unsigned Align = LN0->getAlignment();
9999 // Check the resultant load doesn't need a higher alignment than the
10001 unsigned NewAlign =
10002 TLI.getDataLayout()
10003 ->getABITypeAlignment(LVT.getTypeForEVT(*DAG.getContext()));
10005 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, LVT))
10011 SDValue NewPtr = LN0->getBasePtr();
10012 unsigned PtrOff = 0;
10015 PtrOff = LVT.getSizeInBits() * Elt / 8;
10016 EVT PtrType = NewPtr.getValueType();
10017 if (TLI.isBigEndian())
10018 PtrOff = VT.getSizeInBits() / 8 - PtrOff;
10019 NewPtr = DAG.getNode(ISD::ADD, SDLoc(N), PtrType, NewPtr,
10020 DAG.getConstant(PtrOff, PtrType));
10023 // The replacement we need to do here is a little tricky: we need to
10024 // replace an extractelement of a load with a load.
10025 // Use ReplaceAllUsesOfValuesWith to do the replacement.
10026 // Note that this replacement assumes that the extractvalue is the only
10027 // use of the load; that's okay because we don't want to perform this
10028 // transformation in other cases anyway.
10031 if (NVT.bitsGT(LVT)) {
10032 // If the result type of vextract is wider than the load, then issue an
10033 // extending load instead.
10034 ISD::LoadExtType ExtType = TLI.isLoadExtLegal(ISD::ZEXTLOAD, LVT)
10035 ? ISD::ZEXTLOAD : ISD::EXTLOAD;
10036 Load = DAG.getExtLoad(ExtType, SDLoc(N), NVT, LN0->getChain(),
10037 NewPtr, LN0->getPointerInfo().getWithOffset(PtrOff),
10038 LVT, LN0->isVolatile(), LN0->isNonTemporal(),
10039 Align, LN0->getAAInfo());
10040 Chain = Load.getValue(1);
10042 Load = DAG.getLoad(LVT, SDLoc(N), LN0->getChain(), NewPtr,
10043 LN0->getPointerInfo().getWithOffset(PtrOff),
10044 LN0->isVolatile(), LN0->isNonTemporal(),
10045 LN0->isInvariant(), Align, LN0->getAAInfo());
10046 Chain = Load.getValue(1);
10047 if (NVT.bitsLT(LVT))
10048 Load = DAG.getNode(ISD::TRUNCATE, SDLoc(N), NVT, Load);
10050 Load = DAG.getNode(ISD::BITCAST, SDLoc(N), NVT, Load);
10052 WorklistRemover DeadNodes(*this);
10053 SDValue From[] = { SDValue(N, 0), SDValue(LN0,1) };
10054 SDValue To[] = { Load, Chain };
10055 DAG.ReplaceAllUsesOfValuesWith(From, To, 2);
10056 // Since we're explcitly calling ReplaceAllUses, add the new node to the
10057 // worklist explicitly as well.
10058 AddToWorklist(Load.getNode());
10059 AddUsersToWorklist(Load.getNode()); // Add users too
10060 // Make sure to revisit this node to clean it up; it will usually be dead.
10062 return SDValue(N, 0);
10068 // Simplify (build_vec (ext )) to (bitcast (build_vec ))
10069 SDValue DAGCombiner::reduceBuildVecExtToExtBuildVec(SDNode *N) {
10070 // We perform this optimization post type-legalization because
10071 // the type-legalizer often scalarizes integer-promoted vectors.
10072 // Performing this optimization before may create bit-casts which
10073 // will be type-legalized to complex code sequences.
10074 // We perform this optimization only before the operation legalizer because we
10075 // may introduce illegal operations.
10076 if (Level != AfterLegalizeVectorOps && Level != AfterLegalizeTypes)
10079 unsigned NumInScalars = N->getNumOperands();
10081 EVT VT = N->getValueType(0);
10083 // Check to see if this is a BUILD_VECTOR of a bunch of values
10084 // which come from any_extend or zero_extend nodes. If so, we can create
10085 // a new BUILD_VECTOR using bit-casts which may enable other BUILD_VECTOR
10086 // optimizations. We do not handle sign-extend because we can't fill the sign
10088 EVT SourceType = MVT::Other;
10089 bool AllAnyExt = true;
10091 for (unsigned i = 0; i != NumInScalars; ++i) {
10092 SDValue In = N->getOperand(i);
10093 // Ignore undef inputs.
10094 if (In.getOpcode() == ISD::UNDEF) continue;
10096 bool AnyExt = In.getOpcode() == ISD::ANY_EXTEND;
10097 bool ZeroExt = In.getOpcode() == ISD::ZERO_EXTEND;
10099 // Abort if the element is not an extension.
10100 if (!ZeroExt && !AnyExt) {
10101 SourceType = MVT::Other;
10105 // The input is a ZeroExt or AnyExt. Check the original type.
10106 EVT InTy = In.getOperand(0).getValueType();
10108 // Check that all of the widened source types are the same.
10109 if (SourceType == MVT::Other)
10112 else if (InTy != SourceType) {
10113 // Multiple income types. Abort.
10114 SourceType = MVT::Other;
10118 // Check if all of the extends are ANY_EXTENDs.
10119 AllAnyExt &= AnyExt;
10122 // In order to have valid types, all of the inputs must be extended from the
10123 // same source type and all of the inputs must be any or zero extend.
10124 // Scalar sizes must be a power of two.
10125 EVT OutScalarTy = VT.getScalarType();
10126 bool ValidTypes = SourceType != MVT::Other &&
10127 isPowerOf2_32(OutScalarTy.getSizeInBits()) &&
10128 isPowerOf2_32(SourceType.getSizeInBits());
10130 // Create a new simpler BUILD_VECTOR sequence which other optimizations can
10131 // turn into a single shuffle instruction.
10135 bool isLE = TLI.isLittleEndian();
10136 unsigned ElemRatio = OutScalarTy.getSizeInBits()/SourceType.getSizeInBits();
10137 assert(ElemRatio > 1 && "Invalid element size ratio");
10138 SDValue Filler = AllAnyExt ? DAG.getUNDEF(SourceType):
10139 DAG.getConstant(0, SourceType);
10141 unsigned NewBVElems = ElemRatio * VT.getVectorNumElements();
10142 SmallVector<SDValue, 8> Ops(NewBVElems, Filler);
10144 // Populate the new build_vector
10145 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
10146 SDValue Cast = N->getOperand(i);
10147 assert((Cast.getOpcode() == ISD::ANY_EXTEND ||
10148 Cast.getOpcode() == ISD::ZERO_EXTEND ||
10149 Cast.getOpcode() == ISD::UNDEF) && "Invalid cast opcode");
10151 if (Cast.getOpcode() == ISD::UNDEF)
10152 In = DAG.getUNDEF(SourceType);
10154 In = Cast->getOperand(0);
10155 unsigned Index = isLE ? (i * ElemRatio) :
10156 (i * ElemRatio + (ElemRatio - 1));
10158 assert(Index < Ops.size() && "Invalid index");
10162 // The type of the new BUILD_VECTOR node.
10163 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), SourceType, NewBVElems);
10164 assert(VecVT.getSizeInBits() == VT.getSizeInBits() &&
10165 "Invalid vector size");
10166 // Check if the new vector type is legal.
10167 if (!isTypeLegal(VecVT)) return SDValue();
10169 // Make the new BUILD_VECTOR.
10170 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
10172 // The new BUILD_VECTOR node has the potential to be further optimized.
10173 AddToWorklist(BV.getNode());
10174 // Bitcast to the desired type.
10175 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
10178 SDValue DAGCombiner::reduceBuildVecConvertToConvertBuildVec(SDNode *N) {
10179 EVT VT = N->getValueType(0);
10181 unsigned NumInScalars = N->getNumOperands();
10184 EVT SrcVT = MVT::Other;
10185 unsigned Opcode = ISD::DELETED_NODE;
10186 unsigned NumDefs = 0;
10188 for (unsigned i = 0; i != NumInScalars; ++i) {
10189 SDValue In = N->getOperand(i);
10190 unsigned Opc = In.getOpcode();
10192 if (Opc == ISD::UNDEF)
10195 // If all scalar values are floats and converted from integers.
10196 if (Opcode == ISD::DELETED_NODE &&
10197 (Opc == ISD::UINT_TO_FP || Opc == ISD::SINT_TO_FP)) {
10204 EVT InVT = In.getOperand(0).getValueType();
10206 // If all scalar values are typed differently, bail out. It's chosen to
10207 // simplify BUILD_VECTOR of integer types.
10208 if (SrcVT == MVT::Other)
10215 // If the vector has just one element defined, it's not worth to fold it into
10216 // a vectorized one.
10220 assert((Opcode == ISD::UINT_TO_FP || Opcode == ISD::SINT_TO_FP)
10221 && "Should only handle conversion from integer to float.");
10222 assert(SrcVT != MVT::Other && "Cannot determine source type!");
10224 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SrcVT, NumInScalars);
10226 if (!TLI.isOperationLegalOrCustom(Opcode, NVT))
10229 SmallVector<SDValue, 8> Opnds;
10230 for (unsigned i = 0; i != NumInScalars; ++i) {
10231 SDValue In = N->getOperand(i);
10233 if (In.getOpcode() == ISD::UNDEF)
10234 Opnds.push_back(DAG.getUNDEF(SrcVT));
10236 Opnds.push_back(In.getOperand(0));
10238 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NVT, Opnds);
10239 AddToWorklist(BV.getNode());
10241 return DAG.getNode(Opcode, dl, VT, BV);
10244 SDValue DAGCombiner::visitBUILD_VECTOR(SDNode *N) {
10245 unsigned NumInScalars = N->getNumOperands();
10247 EVT VT = N->getValueType(0);
10249 // A vector built entirely of undefs is undef.
10250 if (ISD::allOperandsUndef(N))
10251 return DAG.getUNDEF(VT);
10253 SDValue V = reduceBuildVecExtToExtBuildVec(N);
10257 V = reduceBuildVecConvertToConvertBuildVec(N);
10261 // Check to see if this is a BUILD_VECTOR of a bunch of EXTRACT_VECTOR_ELT
10262 // operations. If so, and if the EXTRACT_VECTOR_ELT vector inputs come from
10263 // at most two distinct vectors, turn this into a shuffle node.
10265 // May only combine to shuffle after legalize if shuffle is legal.
10266 if (LegalOperations &&
10267 !TLI.isOperationLegalOrCustom(ISD::VECTOR_SHUFFLE, VT))
10270 SDValue VecIn1, VecIn2;
10271 for (unsigned i = 0; i != NumInScalars; ++i) {
10272 // Ignore undef inputs.
10273 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
10275 // If this input is something other than a EXTRACT_VECTOR_ELT with a
10276 // constant index, bail out.
10277 if (N->getOperand(i).getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
10278 !isa<ConstantSDNode>(N->getOperand(i).getOperand(1))) {
10279 VecIn1 = VecIn2 = SDValue(nullptr, 0);
10283 // We allow up to two distinct input vectors.
10284 SDValue ExtractedFromVec = N->getOperand(i).getOperand(0);
10285 if (ExtractedFromVec == VecIn1 || ExtractedFromVec == VecIn2)
10288 if (!VecIn1.getNode()) {
10289 VecIn1 = ExtractedFromVec;
10290 } else if (!VecIn2.getNode()) {
10291 VecIn2 = ExtractedFromVec;
10293 // Too many inputs.
10294 VecIn1 = VecIn2 = SDValue(nullptr, 0);
10299 // If everything is good, we can make a shuffle operation.
10300 if (VecIn1.getNode()) {
10301 SmallVector<int, 8> Mask;
10302 for (unsigned i = 0; i != NumInScalars; ++i) {
10303 if (N->getOperand(i).getOpcode() == ISD::UNDEF) {
10304 Mask.push_back(-1);
10308 // If extracting from the first vector, just use the index directly.
10309 SDValue Extract = N->getOperand(i);
10310 SDValue ExtVal = Extract.getOperand(1);
10311 if (Extract.getOperand(0) == VecIn1) {
10312 unsigned ExtIndex = cast<ConstantSDNode>(ExtVal)->getZExtValue();
10313 if (ExtIndex > VT.getVectorNumElements())
10316 Mask.push_back(ExtIndex);
10320 // Otherwise, use InIdx + VecSize
10321 unsigned Idx = cast<ConstantSDNode>(ExtVal)->getZExtValue();
10322 Mask.push_back(Idx+NumInScalars);
10325 // We can't generate a shuffle node with mismatched input and output types.
10326 // Attempt to transform a single input vector to the correct type.
10327 if ((VT != VecIn1.getValueType())) {
10328 // We don't support shuffeling between TWO values of different types.
10329 if (VecIn2.getNode())
10332 // We only support widening of vectors which are half the size of the
10333 // output registers. For example XMM->YMM widening on X86 with AVX.
10334 if (VecIn1.getValueType().getSizeInBits()*2 != VT.getSizeInBits())
10337 // If the input vector type has a different base type to the output
10338 // vector type, bail out.
10339 if (VecIn1.getValueType().getVectorElementType() !=
10340 VT.getVectorElementType())
10343 // Widen the input vector by adding undef values.
10344 VecIn1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
10345 VecIn1, DAG.getUNDEF(VecIn1.getValueType()));
10348 // If VecIn2 is unused then change it to undef.
10349 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
10351 // Check that we were able to transform all incoming values to the same
10353 if (VecIn2.getValueType() != VecIn1.getValueType() ||
10354 VecIn1.getValueType() != VT)
10357 // Only type-legal BUILD_VECTOR nodes are converted to shuffle nodes.
10358 if (!isTypeLegal(VT))
10361 // Return the new VECTOR_SHUFFLE node.
10365 return DAG.getVectorShuffle(VT, dl, Ops[0], Ops[1], &Mask[0]);
10371 SDValue DAGCombiner::visitCONCAT_VECTORS(SDNode *N) {
10372 // TODO: Check to see if this is a CONCAT_VECTORS of a bunch of
10373 // EXTRACT_SUBVECTOR operations. If so, and if the EXTRACT_SUBVECTOR vector
10374 // inputs come from at most two distinct vectors, turn this into a shuffle
10377 // If we only have one input vector, we don't need to do any concatenation.
10378 if (N->getNumOperands() == 1)
10379 return N->getOperand(0);
10381 // Check if all of the operands are undefs.
10382 EVT VT = N->getValueType(0);
10383 if (ISD::allOperandsUndef(N))
10384 return DAG.getUNDEF(VT);
10386 // Optimize concat_vectors where one of the vectors is undef.
10387 if (N->getNumOperands() == 2 &&
10388 N->getOperand(1)->getOpcode() == ISD::UNDEF) {
10389 SDValue In = N->getOperand(0);
10390 assert(In.getValueType().isVector() && "Must concat vectors");
10392 // Transform: concat_vectors(scalar, undef) -> scalar_to_vector(sclr).
10393 if (In->getOpcode() == ISD::BITCAST &&
10394 !In->getOperand(0)->getValueType(0).isVector()) {
10395 SDValue Scalar = In->getOperand(0);
10396 EVT SclTy = Scalar->getValueType(0);
10398 if (!SclTy.isFloatingPoint() && !SclTy.isInteger())
10401 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SclTy,
10402 VT.getSizeInBits() / SclTy.getSizeInBits());
10403 if (!TLI.isTypeLegal(NVT) || !TLI.isTypeLegal(Scalar.getValueType()))
10406 SDLoc dl = SDLoc(N);
10407 SDValue Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, NVT, Scalar);
10408 return DAG.getNode(ISD::BITCAST, dl, VT, Res);
10412 // fold (concat_vectors (BUILD_VECTOR A, B, ...), (BUILD_VECTOR C, D, ...))
10413 // -> (BUILD_VECTOR A, B, ..., C, D, ...)
10414 if (N->getNumOperands() == 2 &&
10415 N->getOperand(0).getOpcode() == ISD::BUILD_VECTOR &&
10416 N->getOperand(1).getOpcode() == ISD::BUILD_VECTOR) {
10417 EVT VT = N->getValueType(0);
10418 SDValue N0 = N->getOperand(0);
10419 SDValue N1 = N->getOperand(1);
10420 SmallVector<SDValue, 8> Opnds;
10421 unsigned BuildVecNumElts = N0.getNumOperands();
10423 EVT SclTy0 = N0.getOperand(0)->getValueType(0);
10424 EVT SclTy1 = N1.getOperand(0)->getValueType(0);
10425 if (SclTy0.isFloatingPoint()) {
10426 for (unsigned i = 0; i != BuildVecNumElts; ++i)
10427 Opnds.push_back(N0.getOperand(i));
10428 for (unsigned i = 0; i != BuildVecNumElts; ++i)
10429 Opnds.push_back(N1.getOperand(i));
10431 // If BUILD_VECTOR are from built from integer, they may have different
10432 // operand types. Get the smaller type and truncate all operands to it.
10433 EVT MinTy = SclTy0.bitsLE(SclTy1) ? SclTy0 : SclTy1;
10434 for (unsigned i = 0; i != BuildVecNumElts; ++i)
10435 Opnds.push_back(DAG.getNode(ISD::TRUNCATE, SDLoc(N), MinTy,
10436 N0.getOperand(i)));
10437 for (unsigned i = 0; i != BuildVecNumElts; ++i)
10438 Opnds.push_back(DAG.getNode(ISD::TRUNCATE, SDLoc(N), MinTy,
10439 N1.getOperand(i)));
10442 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Opnds);
10445 // Type legalization of vectors and DAG canonicalization of SHUFFLE_VECTOR
10446 // nodes often generate nop CONCAT_VECTOR nodes.
10447 // Scan the CONCAT_VECTOR operands and look for a CONCAT operations that
10448 // place the incoming vectors at the exact same location.
10449 SDValue SingleSource = SDValue();
10450 unsigned PartNumElem = N->getOperand(0).getValueType().getVectorNumElements();
10452 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
10453 SDValue Op = N->getOperand(i);
10455 if (Op.getOpcode() == ISD::UNDEF)
10458 // Check if this is the identity extract:
10459 if (Op.getOpcode() != ISD::EXTRACT_SUBVECTOR)
10462 // Find the single incoming vector for the extract_subvector.
10463 if (SingleSource.getNode()) {
10464 if (Op.getOperand(0) != SingleSource)
10467 SingleSource = Op.getOperand(0);
10469 // Check the source type is the same as the type of the result.
10470 // If not, this concat may extend the vector, so we can not
10471 // optimize it away.
10472 if (SingleSource.getValueType() != N->getValueType(0))
10476 unsigned IdentityIndex = i * PartNumElem;
10477 ConstantSDNode *CS = dyn_cast<ConstantSDNode>(Op.getOperand(1));
10478 // The extract index must be constant.
10482 // Check that we are reading from the identity index.
10483 if (CS->getZExtValue() != IdentityIndex)
10487 if (SingleSource.getNode())
10488 return SingleSource;
10493 SDValue DAGCombiner::visitEXTRACT_SUBVECTOR(SDNode* N) {
10494 EVT NVT = N->getValueType(0);
10495 SDValue V = N->getOperand(0);
10497 if (V->getOpcode() == ISD::CONCAT_VECTORS) {
10499 // (extract_subvec (concat V1, V2, ...), i)
10502 // Only operand 0 is checked as 'concat' assumes all inputs of the same
10504 if (V->getOperand(0).getValueType() != NVT)
10506 unsigned Idx = dyn_cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
10507 unsigned NumElems = NVT.getVectorNumElements();
10508 assert((Idx % NumElems) == 0 &&
10509 "IDX in concat is not a multiple of the result vector length.");
10510 return V->getOperand(Idx / NumElems);
10514 if (V->getOpcode() == ISD::BITCAST)
10515 V = V.getOperand(0);
10517 if (V->getOpcode() == ISD::INSERT_SUBVECTOR) {
10519 // Handle only simple case where vector being inserted and vector
10520 // being extracted are of same type, and are half size of larger vectors.
10521 EVT BigVT = V->getOperand(0).getValueType();
10522 EVT SmallVT = V->getOperand(1).getValueType();
10523 if (!NVT.bitsEq(SmallVT) || NVT.getSizeInBits()*2 != BigVT.getSizeInBits())
10526 // Only handle cases where both indexes are constants with the same type.
10527 ConstantSDNode *ExtIdx = dyn_cast<ConstantSDNode>(N->getOperand(1));
10528 ConstantSDNode *InsIdx = dyn_cast<ConstantSDNode>(V->getOperand(2));
10530 if (InsIdx && ExtIdx &&
10531 InsIdx->getValueType(0).getSizeInBits() <= 64 &&
10532 ExtIdx->getValueType(0).getSizeInBits() <= 64) {
10534 // (extract_subvec (insert_subvec V1, V2, InsIdx), ExtIdx)
10536 // indices are equal or bit offsets are equal => V1
10537 // otherwise => (extract_subvec V1, ExtIdx)
10538 if (InsIdx->getZExtValue() * SmallVT.getScalarType().getSizeInBits() ==
10539 ExtIdx->getZExtValue() * NVT.getScalarType().getSizeInBits())
10540 return DAG.getNode(ISD::BITCAST, dl, NVT, V->getOperand(1));
10541 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, NVT,
10542 DAG.getNode(ISD::BITCAST, dl,
10543 N->getOperand(0).getValueType(),
10544 V->getOperand(0)), N->getOperand(1));
10551 // Tries to turn a shuffle of two CONCAT_VECTORS into a single concat.
10552 static SDValue partitionShuffleOfConcats(SDNode *N, SelectionDAG &DAG) {
10553 EVT VT = N->getValueType(0);
10554 unsigned NumElts = VT.getVectorNumElements();
10556 SDValue N0 = N->getOperand(0);
10557 SDValue N1 = N->getOperand(1);
10558 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
10560 SmallVector<SDValue, 4> Ops;
10561 EVT ConcatVT = N0.getOperand(0).getValueType();
10562 unsigned NumElemsPerConcat = ConcatVT.getVectorNumElements();
10563 unsigned NumConcats = NumElts / NumElemsPerConcat;
10565 // Look at every vector that's inserted. We're looking for exact
10566 // subvector-sized copies from a concatenated vector
10567 for (unsigned I = 0; I != NumConcats; ++I) {
10568 // Make sure we're dealing with a copy.
10569 unsigned Begin = I * NumElemsPerConcat;
10570 bool AllUndef = true, NoUndef = true;
10571 for (unsigned J = Begin; J != Begin + NumElemsPerConcat; ++J) {
10572 if (SVN->getMaskElt(J) >= 0)
10579 if (SVN->getMaskElt(Begin) % NumElemsPerConcat != 0)
10582 for (unsigned J = 1; J != NumElemsPerConcat; ++J)
10583 if (SVN->getMaskElt(Begin + J - 1) + 1 != SVN->getMaskElt(Begin + J))
10586 unsigned FirstElt = SVN->getMaskElt(Begin) / NumElemsPerConcat;
10587 if (FirstElt < N0.getNumOperands())
10588 Ops.push_back(N0.getOperand(FirstElt));
10590 Ops.push_back(N1.getOperand(FirstElt - N0.getNumOperands()));
10592 } else if (AllUndef) {
10593 Ops.push_back(DAG.getUNDEF(N0.getOperand(0).getValueType()));
10594 } else { // Mixed with general masks and undefs, can't do optimization.
10599 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Ops);
10602 SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
10603 EVT VT = N->getValueType(0);
10604 unsigned NumElts = VT.getVectorNumElements();
10606 SDValue N0 = N->getOperand(0);
10607 SDValue N1 = N->getOperand(1);
10609 assert(N0.getValueType() == VT && "Vector shuffle must be normalized in DAG");
10611 // Canonicalize shuffle undef, undef -> undef
10612 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
10613 return DAG.getUNDEF(VT);
10615 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
10617 // Canonicalize shuffle v, v -> v, undef
10619 SmallVector<int, 8> NewMask;
10620 for (unsigned i = 0; i != NumElts; ++i) {
10621 int Idx = SVN->getMaskElt(i);
10622 if (Idx >= (int)NumElts) Idx -= NumElts;
10623 NewMask.push_back(Idx);
10625 return DAG.getVectorShuffle(VT, SDLoc(N), N0, DAG.getUNDEF(VT),
10629 // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask.
10630 if (N0.getOpcode() == ISD::UNDEF) {
10631 SmallVector<int, 8> NewMask;
10632 for (unsigned i = 0; i != NumElts; ++i) {
10633 int Idx = SVN->getMaskElt(i);
10635 if (Idx >= (int)NumElts)
10638 Idx = -1; // remove reference to lhs
10640 NewMask.push_back(Idx);
10642 return DAG.getVectorShuffle(VT, SDLoc(N), N1, DAG.getUNDEF(VT),
10646 // Remove references to rhs if it is undef
10647 if (N1.getOpcode() == ISD::UNDEF) {
10648 bool Changed = false;
10649 SmallVector<int, 8> NewMask;
10650 for (unsigned i = 0; i != NumElts; ++i) {
10651 int Idx = SVN->getMaskElt(i);
10652 if (Idx >= (int)NumElts) {
10656 NewMask.push_back(Idx);
10659 return DAG.getVectorShuffle(VT, SDLoc(N), N0, N1, &NewMask[0]);
10662 // If it is a splat, check if the argument vector is another splat or a
10663 // build_vector with all scalar elements the same.
10664 if (SVN->isSplat() && SVN->getSplatIndex() < (int)NumElts) {
10665 SDNode *V = N0.getNode();
10667 // If this is a bit convert that changes the element type of the vector but
10668 // not the number of vector elements, look through it. Be careful not to
10669 // look though conversions that change things like v4f32 to v2f64.
10670 if (V->getOpcode() == ISD::BITCAST) {
10671 SDValue ConvInput = V->getOperand(0);
10672 if (ConvInput.getValueType().isVector() &&
10673 ConvInput.getValueType().getVectorNumElements() == NumElts)
10674 V = ConvInput.getNode();
10677 if (V->getOpcode() == ISD::BUILD_VECTOR) {
10678 assert(V->getNumOperands() == NumElts &&
10679 "BUILD_VECTOR has wrong number of operands");
10681 bool AllSame = true;
10682 for (unsigned i = 0; i != NumElts; ++i) {
10683 if (V->getOperand(i).getOpcode() != ISD::UNDEF) {
10684 Base = V->getOperand(i);
10688 // Splat of <u, u, u, u>, return <u, u, u, u>
10689 if (!Base.getNode())
10691 for (unsigned i = 0; i != NumElts; ++i) {
10692 if (V->getOperand(i) != Base) {
10697 // Splat of <x, x, x, x>, return <x, x, x, x>
10703 if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
10704 Level < AfterLegalizeVectorOps &&
10705 (N1.getOpcode() == ISD::UNDEF ||
10706 (N1.getOpcode() == ISD::CONCAT_VECTORS &&
10707 N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()))) {
10708 SDValue V = partitionShuffleOfConcats(N, DAG);
10714 // If this shuffle node is simply a swizzle of another shuffle node,
10715 // then try to simplify it.
10716 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
10717 N1.getOpcode() == ISD::UNDEF) {
10719 ShuffleVectorSDNode *OtherSV = cast<ShuffleVectorSDNode>(N0);
10721 // The incoming shuffle must be of the same type as the result of the
10722 // current shuffle.
10723 assert(OtherSV->getOperand(0).getValueType() == VT &&
10724 "Shuffle types don't match");
10726 SmallVector<int, 4> Mask;
10727 // Compute the combined shuffle mask.
10728 for (unsigned i = 0; i != NumElts; ++i) {
10729 int Idx = SVN->getMaskElt(i);
10730 assert(Idx < (int)NumElts && "Index references undef operand");
10731 // Next, this index comes from the first value, which is the incoming
10732 // shuffle. Adopt the incoming index.
10734 Idx = OtherSV->getMaskElt(Idx);
10735 Mask.push_back(Idx);
10738 bool CommuteOperands = false;
10739 if (N0.getOperand(1).getOpcode() != ISD::UNDEF) {
10740 // To be valid, the combine shuffle mask should only reference elements
10741 // from one of the two vectors in input to the inner shufflevector.
10742 bool IsValidMask = true;
10743 for (unsigned i = 0; i != NumElts && IsValidMask; ++i)
10744 // See if the combined mask only reference undefs or elements coming
10745 // from the first shufflevector operand.
10746 IsValidMask = Mask[i] < 0 || (unsigned)Mask[i] < NumElts;
10748 if (!IsValidMask) {
10749 IsValidMask = true;
10750 for (unsigned i = 0; i != NumElts && IsValidMask; ++i)
10751 // Check that all the elements come from the second shuffle operand.
10752 IsValidMask = Mask[i] < 0 || (unsigned)Mask[i] >= NumElts;
10753 CommuteOperands = IsValidMask;
10756 // Early exit if the combined shuffle mask is not valid.
10761 // See if this pair of shuffles can be safely folded according to either
10762 // of the following rules:
10763 // shuffle(shuffle(x, y), undef) -> x
10764 // shuffle(shuffle(x, undef), undef) -> x
10765 // shuffle(shuffle(x, y), undef) -> y
10766 bool IsIdentityMask = true;
10767 unsigned BaseMaskIndex = CommuteOperands ? NumElts : 0;
10768 for (unsigned i = 0; i != NumElts && IsIdentityMask; ++i) {
10773 // The combined shuffle must map each index to itself.
10774 IsIdentityMask = (unsigned)Mask[i] == i + BaseMaskIndex;
10777 if (IsIdentityMask) {
10778 if (CommuteOperands)
10779 // optimize shuffle(shuffle(x, y), undef) -> y.
10780 return OtherSV->getOperand(1);
10782 // optimize shuffle(shuffle(x, undef), undef) -> x
10783 // optimize shuffle(shuffle(x, y), undef) -> x
10784 return OtherSV->getOperand(0);
10787 // It may still be beneficial to combine the two shuffles if the
10788 // resulting shuffle is legal.
10789 if (TLI.isTypeLegal(VT) && TLI.isShuffleMaskLegal(Mask, VT)) {
10790 if (!CommuteOperands)
10791 // shuffle(shuffle(x, undef, M1), undef, M2) -> shuffle(x, undef, M3).
10792 // shuffle(shuffle(x, y, M1), undef, M2) -> shuffle(x, undef, M3)
10793 return DAG.getVectorShuffle(VT, SDLoc(N), N0->getOperand(0), N1,
10796 // shuffle(shuffle(x, y, M1), undef, M2) -> shuffle(undef, y, M3)
10797 return DAG.getVectorShuffle(VT, SDLoc(N), N1, N0->getOperand(1),
10802 // Canonicalize shuffles according to rules:
10803 // shuffle(A, shuffle(A, B)) -> shuffle(shuffle(A,B), A)
10804 // shuffle(B, shuffle(A, B)) -> shuffle(shuffle(A,B), B)
10805 // shuffle(B, shuffle(A, Undef)) -> shuffle(shuffle(A, Undef), B)
10806 if (N1.getOpcode() == ISD::VECTOR_SHUFFLE && N0.getOpcode() != ISD::UNDEF &&
10807 N0.getOpcode() != ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
10808 TLI.isTypeLegal(VT)) {
10809 // The incoming shuffle must be of the same type as the result of the
10810 // current shuffle.
10811 assert(N1->getOperand(0).getValueType() == VT &&
10812 "Shuffle types don't match");
10814 SDValue SV0 = N1->getOperand(0);
10815 SDValue SV1 = N1->getOperand(1);
10816 bool HasSameOp0 = N0 == SV0;
10817 bool IsSV1Undef = SV1.getOpcode() == ISD::UNDEF;
10818 if (HasSameOp0 || IsSV1Undef || N0 == SV1)
10819 // Commute the operands of this shuffle so that next rule
10821 return DAG.getCommutedVectorShuffle(*SVN);
10824 // Try to fold according to rules:
10825 // shuffle(shuffle(A, B, M0), B, M1) -> shuffle(A, B, M2)
10826 // shuffle(shuffle(A, B, M0), A, M1) -> shuffle(A, B, M2)
10827 // shuffle(shuffle(A, Undef, M0), B, M1) -> shuffle(A, B, M2)
10828 // shuffle(shuffle(A, Undef, M0), A, M1) -> shuffle(A, Undef, M2)
10829 // Don't try to fold shuffles with illegal type.
10830 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
10831 N1.getOpcode() != ISD::UNDEF && TLI.isTypeLegal(VT)) {
10832 ShuffleVectorSDNode *OtherSV = cast<ShuffleVectorSDNode>(N0);
10834 // The incoming shuffle must be of the same type as the result of the
10835 // current shuffle.
10836 assert(OtherSV->getOperand(0).getValueType() == VT &&
10837 "Shuffle types don't match");
10839 SDValue SV0 = OtherSV->getOperand(0);
10840 SDValue SV1 = OtherSV->getOperand(1);
10841 bool HasSameOp0 = N1 == SV0;
10842 bool IsSV1Undef = SV1.getOpcode() == ISD::UNDEF;
10843 if (!HasSameOp0 && !IsSV1Undef && N1 != SV1)
10847 SmallVector<int, 4> Mask;
10848 // Compute the combined shuffle mask for a shuffle with SV0 as the first
10849 // operand, and SV1 as the second operand.
10850 for (unsigned i = 0; i != NumElts; ++i) {
10851 int Idx = SVN->getMaskElt(i);
10853 // Propagate Undef.
10854 Mask.push_back(Idx);
10858 if (Idx < (int)NumElts) {
10859 Idx = OtherSV->getMaskElt(Idx);
10860 if (IsSV1Undef && Idx >= (int) NumElts)
10861 Idx = -1; // Propagate Undef.
10863 Idx = HasSameOp0 ? Idx - NumElts : Idx;
10865 Mask.push_back(Idx);
10868 // Avoid introducing shuffles with illegal mask.
10869 if (TLI.isShuffleMaskLegal(Mask, VT)) {
10871 // shuffle(shuffle(A, Undef, M0), B, M1) -> shuffle(A, B, M2)
10872 // shuffle(shuffle(A, Undef, M0), A, M1) -> shuffle(A, Undef, M2)
10873 return DAG.getVectorShuffle(VT, SDLoc(N), SV0, N1, &Mask[0]);
10874 return DAG.getVectorShuffle(VT, SDLoc(N), SV0, SV1, &Mask[0]);
10881 SDValue DAGCombiner::visitINSERT_SUBVECTOR(SDNode *N) {
10882 SDValue N0 = N->getOperand(0);
10883 SDValue N2 = N->getOperand(2);
10885 // If the input vector is a concatenation, and the insert replaces
10886 // one of the halves, we can optimize into a single concat_vectors.
10887 if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
10888 N0->getNumOperands() == 2 && N2.getOpcode() == ISD::Constant) {
10889 APInt InsIdx = cast<ConstantSDNode>(N2)->getAPIntValue();
10890 EVT VT = N->getValueType(0);
10892 // Lower half: fold (insert_subvector (concat_vectors X, Y), Z) ->
10893 // (concat_vectors Z, Y)
10895 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
10896 N->getOperand(1), N0.getOperand(1));
10898 // Upper half: fold (insert_subvector (concat_vectors X, Y), Z) ->
10899 // (concat_vectors X, Z)
10900 if (InsIdx == VT.getVectorNumElements()/2)
10901 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
10902 N0.getOperand(0), N->getOperand(1));
10908 /// XformToShuffleWithZero - Returns a vector_shuffle if it able to transform
10909 /// an AND to a vector_shuffle with the destination vector and a zero vector.
10910 /// e.g. AND V, <0xffffffff, 0, 0xffffffff, 0>. ==>
10911 /// vector_shuffle V, Zero, <0, 4, 2, 4>
10912 SDValue DAGCombiner::XformToShuffleWithZero(SDNode *N) {
10913 EVT VT = N->getValueType(0);
10915 SDValue LHS = N->getOperand(0);
10916 SDValue RHS = N->getOperand(1);
10917 if (N->getOpcode() == ISD::AND) {
10918 if (RHS.getOpcode() == ISD::BITCAST)
10919 RHS = RHS.getOperand(0);
10920 if (RHS.getOpcode() == ISD::BUILD_VECTOR) {
10921 SmallVector<int, 8> Indices;
10922 unsigned NumElts = RHS.getNumOperands();
10923 for (unsigned i = 0; i != NumElts; ++i) {
10924 SDValue Elt = RHS.getOperand(i);
10925 if (!isa<ConstantSDNode>(Elt))
10928 if (cast<ConstantSDNode>(Elt)->isAllOnesValue())
10929 Indices.push_back(i);
10930 else if (cast<ConstantSDNode>(Elt)->isNullValue())
10931 Indices.push_back(NumElts);
10936 // Let's see if the target supports this vector_shuffle.
10937 EVT RVT = RHS.getValueType();
10938 if (!TLI.isVectorClearMaskLegal(Indices, RVT))
10941 // Return the new VECTOR_SHUFFLE node.
10942 EVT EltVT = RVT.getVectorElementType();
10943 SmallVector<SDValue,8> ZeroOps(RVT.getVectorNumElements(),
10944 DAG.getConstant(0, EltVT));
10945 SDValue Zero = DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), RVT, ZeroOps);
10946 LHS = DAG.getNode(ISD::BITCAST, dl, RVT, LHS);
10947 SDValue Shuf = DAG.getVectorShuffle(RVT, dl, LHS, Zero, &Indices[0]);
10948 return DAG.getNode(ISD::BITCAST, dl, VT, Shuf);
10955 /// SimplifyVBinOp - Visit a binary vector operation, like ADD.
10956 SDValue DAGCombiner::SimplifyVBinOp(SDNode *N) {
10957 assert(N->getValueType(0).isVector() &&
10958 "SimplifyVBinOp only works on vectors!");
10960 SDValue LHS = N->getOperand(0);
10961 SDValue RHS = N->getOperand(1);
10962 SDValue Shuffle = XformToShuffleWithZero(N);
10963 if (Shuffle.getNode()) return Shuffle;
10965 // If the LHS and RHS are BUILD_VECTOR nodes, see if we can constant fold
10967 if (LHS.getOpcode() == ISD::BUILD_VECTOR &&
10968 RHS.getOpcode() == ISD::BUILD_VECTOR) {
10969 // Check if both vectors are constants. If not bail out.
10970 if (!(cast<BuildVectorSDNode>(LHS)->isConstant() &&
10971 cast<BuildVectorSDNode>(RHS)->isConstant()))
10974 SmallVector<SDValue, 8> Ops;
10975 for (unsigned i = 0, e = LHS.getNumOperands(); i != e; ++i) {
10976 SDValue LHSOp = LHS.getOperand(i);
10977 SDValue RHSOp = RHS.getOperand(i);
10979 // Can't fold divide by zero.
10980 if (N->getOpcode() == ISD::SDIV || N->getOpcode() == ISD::UDIV ||
10981 N->getOpcode() == ISD::FDIV) {
10982 if ((RHSOp.getOpcode() == ISD::Constant &&
10983 cast<ConstantSDNode>(RHSOp.getNode())->isNullValue()) ||
10984 (RHSOp.getOpcode() == ISD::ConstantFP &&
10985 cast<ConstantFPSDNode>(RHSOp.getNode())->getValueAPF().isZero()))
10989 EVT VT = LHSOp.getValueType();
10990 EVT RVT = RHSOp.getValueType();
10992 // Integer BUILD_VECTOR operands may have types larger than the element
10993 // size (e.g., when the element type is not legal). Prior to type
10994 // legalization, the types may not match between the two BUILD_VECTORS.
10995 // Truncate one of the operands to make them match.
10996 if (RVT.getSizeInBits() > VT.getSizeInBits()) {
10997 RHSOp = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, RHSOp);
10999 LHSOp = DAG.getNode(ISD::TRUNCATE, SDLoc(N), RVT, LHSOp);
11003 SDValue FoldOp = DAG.getNode(N->getOpcode(), SDLoc(LHS), VT,
11005 if (FoldOp.getOpcode() != ISD::UNDEF &&
11006 FoldOp.getOpcode() != ISD::Constant &&
11007 FoldOp.getOpcode() != ISD::ConstantFP)
11009 Ops.push_back(FoldOp);
11010 AddToWorklist(FoldOp.getNode());
11013 if (Ops.size() == LHS.getNumOperands())
11014 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), LHS.getValueType(), Ops);
11017 // Type legalization might introduce new shuffles in the DAG.
11018 // Fold (VBinOp (shuffle (A, Undef, Mask)), (shuffle (B, Undef, Mask)))
11019 // -> (shuffle (VBinOp (A, B)), Undef, Mask).
11020 if (LegalTypes && isa<ShuffleVectorSDNode>(LHS) &&
11021 isa<ShuffleVectorSDNode>(RHS) && LHS.hasOneUse() && RHS.hasOneUse() &&
11022 LHS.getOperand(1).getOpcode() == ISD::UNDEF &&
11023 RHS.getOperand(1).getOpcode() == ISD::UNDEF) {
11024 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(LHS);
11025 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(RHS);
11027 if (SVN0->getMask().equals(SVN1->getMask())) {
11028 EVT VT = N->getValueType(0);
11029 SDValue UndefVector = LHS.getOperand(1);
11030 SDValue NewBinOp = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
11031 LHS.getOperand(0), RHS.getOperand(0));
11032 AddUsersToWorklist(N);
11033 return DAG.getVectorShuffle(VT, SDLoc(N), NewBinOp, UndefVector,
11034 &SVN0->getMask()[0]);
11041 /// SimplifyVUnaryOp - Visit a binary vector operation, like FABS/FNEG.
11042 SDValue DAGCombiner::SimplifyVUnaryOp(SDNode *N) {
11043 assert(N->getValueType(0).isVector() &&
11044 "SimplifyVUnaryOp only works on vectors!");
11046 SDValue N0 = N->getOperand(0);
11048 if (N0.getOpcode() != ISD::BUILD_VECTOR)
11051 // Operand is a BUILD_VECTOR node, see if we can constant fold it.
11052 SmallVector<SDValue, 8> Ops;
11053 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) {
11054 SDValue Op = N0.getOperand(i);
11055 if (Op.getOpcode() != ISD::UNDEF &&
11056 Op.getOpcode() != ISD::ConstantFP)
11058 EVT EltVT = Op.getValueType();
11059 SDValue FoldOp = DAG.getNode(N->getOpcode(), SDLoc(N0), EltVT, Op);
11060 if (FoldOp.getOpcode() != ISD::UNDEF &&
11061 FoldOp.getOpcode() != ISD::ConstantFP)
11063 Ops.push_back(FoldOp);
11064 AddToWorklist(FoldOp.getNode());
11067 if (Ops.size() != N0.getNumOperands())
11070 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), N0.getValueType(), Ops);
11073 SDValue DAGCombiner::SimplifySelect(SDLoc DL, SDValue N0,
11074 SDValue N1, SDValue N2){
11075 assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!");
11077 SDValue SCC = SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1), N1, N2,
11078 cast<CondCodeSDNode>(N0.getOperand(2))->get());
11080 // If we got a simplified select_cc node back from SimplifySelectCC, then
11081 // break it down into a new SETCC node, and a new SELECT node, and then return
11082 // the SELECT node, since we were called with a SELECT node.
11083 if (SCC.getNode()) {
11084 // Check to see if we got a select_cc back (to turn into setcc/select).
11085 // Otherwise, just return whatever node we got back, like fabs.
11086 if (SCC.getOpcode() == ISD::SELECT_CC) {
11087 SDValue SETCC = DAG.getNode(ISD::SETCC, SDLoc(N0),
11089 SCC.getOperand(0), SCC.getOperand(1),
11090 SCC.getOperand(4));
11091 AddToWorklist(SETCC.getNode());
11092 return DAG.getSelect(SDLoc(SCC), SCC.getValueType(),
11093 SCC.getOperand(2), SCC.getOperand(3), SETCC);
11101 /// SimplifySelectOps - Given a SELECT or a SELECT_CC node, where LHS and RHS
11102 /// are the two values being selected between, see if we can simplify the
11103 /// select. Callers of this should assume that TheSelect is deleted if this
11104 /// returns true. As such, they should return the appropriate thing (e.g. the
11105 /// node) back to the top-level of the DAG combiner loop to avoid it being
11107 bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDValue LHS,
11110 // Cannot simplify select with vector condition
11111 if (TheSelect->getOperand(0).getValueType().isVector()) return false;
11113 // If this is a select from two identical things, try to pull the operation
11114 // through the select.
11115 if (LHS.getOpcode() != RHS.getOpcode() ||
11116 !LHS.hasOneUse() || !RHS.hasOneUse())
11119 // If this is a load and the token chain is identical, replace the select
11120 // of two loads with a load through a select of the address to load from.
11121 // This triggers in things like "select bool X, 10.0, 123.0" after the FP
11122 // constants have been dropped into the constant pool.
11123 if (LHS.getOpcode() == ISD::LOAD) {
11124 LoadSDNode *LLD = cast<LoadSDNode>(LHS);
11125 LoadSDNode *RLD = cast<LoadSDNode>(RHS);
11127 // Token chains must be identical.
11128 if (LHS.getOperand(0) != RHS.getOperand(0) ||
11129 // Do not let this transformation reduce the number of volatile loads.
11130 LLD->isVolatile() || RLD->isVolatile() ||
11131 // If this is an EXTLOAD, the VT's must match.
11132 LLD->getMemoryVT() != RLD->getMemoryVT() ||
11133 // If this is an EXTLOAD, the kind of extension must match.
11134 (LLD->getExtensionType() != RLD->getExtensionType() &&
11135 // The only exception is if one of the extensions is anyext.
11136 LLD->getExtensionType() != ISD::EXTLOAD &&
11137 RLD->getExtensionType() != ISD::EXTLOAD) ||
11138 // FIXME: this discards src value information. This is
11139 // over-conservative. It would be beneficial to be able to remember
11140 // both potential memory locations. Since we are discarding
11141 // src value info, don't do the transformation if the memory
11142 // locations are not in the default address space.
11143 LLD->getPointerInfo().getAddrSpace() != 0 ||
11144 RLD->getPointerInfo().getAddrSpace() != 0 ||
11145 !TLI.isOperationLegalOrCustom(TheSelect->getOpcode(),
11146 LLD->getBasePtr().getValueType()))
11149 // Check that the select condition doesn't reach either load. If so,
11150 // folding this will induce a cycle into the DAG. If not, this is safe to
11151 // xform, so create a select of the addresses.
11153 if (TheSelect->getOpcode() == ISD::SELECT) {
11154 SDNode *CondNode = TheSelect->getOperand(0).getNode();
11155 if ((LLD->hasAnyUseOfValue(1) && LLD->isPredecessorOf(CondNode)) ||
11156 (RLD->hasAnyUseOfValue(1) && RLD->isPredecessorOf(CondNode)))
11158 // The loads must not depend on one another.
11159 if (LLD->isPredecessorOf(RLD) ||
11160 RLD->isPredecessorOf(LLD))
11162 Addr = DAG.getSelect(SDLoc(TheSelect),
11163 LLD->getBasePtr().getValueType(),
11164 TheSelect->getOperand(0), LLD->getBasePtr(),
11165 RLD->getBasePtr());
11166 } else { // Otherwise SELECT_CC
11167 SDNode *CondLHS = TheSelect->getOperand(0).getNode();
11168 SDNode *CondRHS = TheSelect->getOperand(1).getNode();
11170 if ((LLD->hasAnyUseOfValue(1) &&
11171 (LLD->isPredecessorOf(CondLHS) || LLD->isPredecessorOf(CondRHS))) ||
11172 (RLD->hasAnyUseOfValue(1) &&
11173 (RLD->isPredecessorOf(CondLHS) || RLD->isPredecessorOf(CondRHS))))
11176 Addr = DAG.getNode(ISD::SELECT_CC, SDLoc(TheSelect),
11177 LLD->getBasePtr().getValueType(),
11178 TheSelect->getOperand(0),
11179 TheSelect->getOperand(1),
11180 LLD->getBasePtr(), RLD->getBasePtr(),
11181 TheSelect->getOperand(4));
11185 if (LLD->getExtensionType() == ISD::NON_EXTLOAD) {
11186 Load = DAG.getLoad(TheSelect->getValueType(0),
11188 // FIXME: Discards pointer and AA info.
11189 LLD->getChain(), Addr, MachinePointerInfo(),
11190 LLD->isVolatile(), LLD->isNonTemporal(),
11191 LLD->isInvariant(), LLD->getAlignment());
11193 Load = DAG.getExtLoad(LLD->getExtensionType() == ISD::EXTLOAD ?
11194 RLD->getExtensionType() : LLD->getExtensionType(),
11196 TheSelect->getValueType(0),
11197 // FIXME: Discards pointer and AA info.
11198 LLD->getChain(), Addr, MachinePointerInfo(),
11199 LLD->getMemoryVT(), LLD->isVolatile(),
11200 LLD->isNonTemporal(), LLD->getAlignment());
11203 // Users of the select now use the result of the load.
11204 CombineTo(TheSelect, Load);
11206 // Users of the old loads now use the new load's chain. We know the
11207 // old-load value is dead now.
11208 CombineTo(LHS.getNode(), Load.getValue(0), Load.getValue(1));
11209 CombineTo(RHS.getNode(), Load.getValue(0), Load.getValue(1));
11216 /// SimplifySelectCC - Simplify an expression of the form (N0 cond N1) ? N2 : N3
11217 /// where 'cond' is the comparison specified by CC.
11218 SDValue DAGCombiner::SimplifySelectCC(SDLoc DL, SDValue N0, SDValue N1,
11219 SDValue N2, SDValue N3,
11220 ISD::CondCode CC, bool NotExtCompare) {
11221 // (x ? y : y) -> y.
11222 if (N2 == N3) return N2;
11224 EVT VT = N2.getValueType();
11225 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
11226 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
11227 ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N3.getNode());
11229 // Determine if the condition we're dealing with is constant
11230 SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()),
11231 N0, N1, CC, DL, false);
11232 if (SCC.getNode()) AddToWorklist(SCC.getNode());
11233 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode());
11235 // fold select_cc true, x, y -> x
11236 if (SCCC && !SCCC->isNullValue())
11238 // fold select_cc false, x, y -> y
11239 if (SCCC && SCCC->isNullValue())
11242 // Check to see if we can simplify the select into an fabs node
11243 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) {
11244 // Allow either -0.0 or 0.0
11245 if (CFP->getValueAPF().isZero()) {
11246 // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs
11247 if ((CC == ISD::SETGE || CC == ISD::SETGT) &&
11248 N0 == N2 && N3.getOpcode() == ISD::FNEG &&
11249 N2 == N3.getOperand(0))
11250 return DAG.getNode(ISD::FABS, DL, VT, N0);
11252 // select (setl[te] X, +/-0.0), fneg(X), X -> fabs
11253 if ((CC == ISD::SETLT || CC == ISD::SETLE) &&
11254 N0 == N3 && N2.getOpcode() == ISD::FNEG &&
11255 N2.getOperand(0) == N3)
11256 return DAG.getNode(ISD::FABS, DL, VT, N3);
11260 // Turn "(a cond b) ? 1.0f : 2.0f" into "load (tmp + ((a cond b) ? 0 : 4)"
11261 // where "tmp" is a constant pool entry containing an array with 1.0 and 2.0
11262 // in it. This is a win when the constant is not otherwise available because
11263 // it replaces two constant pool loads with one. We only do this if the FP
11264 // type is known to be legal, because if it isn't, then we are before legalize
11265 // types an we want the other legalization to happen first (e.g. to avoid
11266 // messing with soft float) and if the ConstantFP is not legal, because if
11267 // it is legal, we may not need to store the FP constant in a constant pool.
11268 if (ConstantFPSDNode *TV = dyn_cast<ConstantFPSDNode>(N2))
11269 if (ConstantFPSDNode *FV = dyn_cast<ConstantFPSDNode>(N3)) {
11270 if (TLI.isTypeLegal(N2.getValueType()) &&
11271 (TLI.getOperationAction(ISD::ConstantFP, N2.getValueType()) !=
11272 TargetLowering::Legal &&
11273 !TLI.isFPImmLegal(TV->getValueAPF(), TV->getValueType(0)) &&
11274 !TLI.isFPImmLegal(FV->getValueAPF(), FV->getValueType(0))) &&
11275 // If both constants have multiple uses, then we won't need to do an
11276 // extra load, they are likely around in registers for other users.
11277 (TV->hasOneUse() || FV->hasOneUse())) {
11278 Constant *Elts[] = {
11279 const_cast<ConstantFP*>(FV->getConstantFPValue()),
11280 const_cast<ConstantFP*>(TV->getConstantFPValue())
11282 Type *FPTy = Elts[0]->getType();
11283 const DataLayout &TD = *TLI.getDataLayout();
11285 // Create a ConstantArray of the two constants.
11286 Constant *CA = ConstantArray::get(ArrayType::get(FPTy, 2), Elts);
11287 SDValue CPIdx = DAG.getConstantPool(CA, TLI.getPointerTy(),
11288 TD.getPrefTypeAlignment(FPTy));
11289 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
11291 // Get the offsets to the 0 and 1 element of the array so that we can
11292 // select between them.
11293 SDValue Zero = DAG.getIntPtrConstant(0);
11294 unsigned EltSize = (unsigned)TD.getTypeAllocSize(Elts[0]->getType());
11295 SDValue One = DAG.getIntPtrConstant(EltSize);
11297 SDValue Cond = DAG.getSetCC(DL,
11298 getSetCCResultType(N0.getValueType()),
11300 AddToWorklist(Cond.getNode());
11301 SDValue CstOffset = DAG.getSelect(DL, Zero.getValueType(),
11303 AddToWorklist(CstOffset.getNode());
11304 CPIdx = DAG.getNode(ISD::ADD, DL, CPIdx.getValueType(), CPIdx,
11306 AddToWorklist(CPIdx.getNode());
11307 return DAG.getLoad(TV->getValueType(0), DL, DAG.getEntryNode(), CPIdx,
11308 MachinePointerInfo::getConstantPool(), false,
11309 false, false, Alignment);
11314 // Check to see if we can perform the "gzip trick", transforming
11315 // (select_cc setlt X, 0, A, 0) -> (and (sra X, (sub size(X), 1), A)
11316 if (N1C && N3C && N3C->isNullValue() && CC == ISD::SETLT &&
11317 (N1C->isNullValue() || // (a < 0) ? b : 0
11318 (N1C->getAPIntValue() == 1 && N0 == N2))) { // (a < 1) ? a : 0
11319 EVT XType = N0.getValueType();
11320 EVT AType = N2.getValueType();
11321 if (XType.bitsGE(AType)) {
11322 // and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a
11323 // single-bit constant.
11324 if (N2C && ((N2C->getAPIntValue() & (N2C->getAPIntValue()-1)) == 0)) {
11325 unsigned ShCtV = N2C->getAPIntValue().logBase2();
11326 ShCtV = XType.getSizeInBits()-ShCtV-1;
11327 SDValue ShCt = DAG.getConstant(ShCtV,
11328 getShiftAmountTy(N0.getValueType()));
11329 SDValue Shift = DAG.getNode(ISD::SRL, SDLoc(N0),
11331 AddToWorklist(Shift.getNode());
11333 if (XType.bitsGT(AType)) {
11334 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
11335 AddToWorklist(Shift.getNode());
11338 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
11341 SDValue Shift = DAG.getNode(ISD::SRA, SDLoc(N0),
11343 DAG.getConstant(XType.getSizeInBits()-1,
11344 getShiftAmountTy(N0.getValueType())));
11345 AddToWorklist(Shift.getNode());
11347 if (XType.bitsGT(AType)) {
11348 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
11349 AddToWorklist(Shift.getNode());
11352 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
11356 // fold (select_cc seteq (and x, y), 0, 0, A) -> (and (shr (shl x)) A)
11357 // where y is has a single bit set.
11358 // A plaintext description would be, we can turn the SELECT_CC into an AND
11359 // when the condition can be materialized as an all-ones register. Any
11360 // single bit-test can be materialized as an all-ones register with
11361 // shift-left and shift-right-arith.
11362 if (CC == ISD::SETEQ && N0->getOpcode() == ISD::AND &&
11363 N0->getValueType(0) == VT &&
11364 N1C && N1C->isNullValue() &&
11365 N2C && N2C->isNullValue()) {
11366 SDValue AndLHS = N0->getOperand(0);
11367 ConstantSDNode *ConstAndRHS = dyn_cast<ConstantSDNode>(N0->getOperand(1));
11368 if (ConstAndRHS && ConstAndRHS->getAPIntValue().countPopulation() == 1) {
11369 // Shift the tested bit over the sign bit.
11370 APInt AndMask = ConstAndRHS->getAPIntValue();
11372 DAG.getConstant(AndMask.countLeadingZeros(),
11373 getShiftAmountTy(AndLHS.getValueType()));
11374 SDValue Shl = DAG.getNode(ISD::SHL, SDLoc(N0), VT, AndLHS, ShlAmt);
11376 // Now arithmetic right shift it all the way over, so the result is either
11377 // all-ones, or zero.
11379 DAG.getConstant(AndMask.getBitWidth()-1,
11380 getShiftAmountTy(Shl.getValueType()));
11381 SDValue Shr = DAG.getNode(ISD::SRA, SDLoc(N0), VT, Shl, ShrAmt);
11383 return DAG.getNode(ISD::AND, DL, VT, Shr, N3);
11387 // fold select C, 16, 0 -> shl C, 4
11388 if (N2C && N3C && N3C->isNullValue() && N2C->getAPIntValue().isPowerOf2() &&
11389 TLI.getBooleanContents(N0.getValueType()) ==
11390 TargetLowering::ZeroOrOneBooleanContent) {
11392 // If the caller doesn't want us to simplify this into a zext of a compare,
11394 if (NotExtCompare && N2C->getAPIntValue() == 1)
11397 // Get a SetCC of the condition
11398 // NOTE: Don't create a SETCC if it's not legal on this target.
11399 if (!LegalOperations ||
11400 TLI.isOperationLegal(ISD::SETCC,
11401 LegalTypes ? getSetCCResultType(N0.getValueType()) : MVT::i1)) {
11403 // cast from setcc result type to select result type
11405 SCC = DAG.getSetCC(DL, getSetCCResultType(N0.getValueType()),
11407 if (N2.getValueType().bitsLT(SCC.getValueType()))
11408 Temp = DAG.getZeroExtendInReg(SCC, SDLoc(N2),
11409 N2.getValueType());
11411 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2),
11412 N2.getValueType(), SCC);
11414 SCC = DAG.getSetCC(SDLoc(N0), MVT::i1, N0, N1, CC);
11415 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2),
11416 N2.getValueType(), SCC);
11419 AddToWorklist(SCC.getNode());
11420 AddToWorklist(Temp.getNode());
11422 if (N2C->getAPIntValue() == 1)
11425 // shl setcc result by log2 n2c
11426 return DAG.getNode(
11427 ISD::SHL, DL, N2.getValueType(), Temp,
11428 DAG.getConstant(N2C->getAPIntValue().logBase2(),
11429 getShiftAmountTy(Temp.getValueType())));
11433 // Check to see if this is the equivalent of setcc
11434 // FIXME: Turn all of these into setcc if setcc if setcc is legal
11435 // otherwise, go ahead with the folds.
11436 if (0 && N3C && N3C->isNullValue() && N2C && (N2C->getAPIntValue() == 1ULL)) {
11437 EVT XType = N0.getValueType();
11438 if (!LegalOperations ||
11439 TLI.isOperationLegal(ISD::SETCC, getSetCCResultType(XType))) {
11440 SDValue Res = DAG.getSetCC(DL, getSetCCResultType(XType), N0, N1, CC);
11441 if (Res.getValueType() != VT)
11442 Res = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Res);
11446 // fold (seteq X, 0) -> (srl (ctlz X, log2(size(X))))
11447 if (N1C && N1C->isNullValue() && CC == ISD::SETEQ &&
11448 (!LegalOperations ||
11449 TLI.isOperationLegal(ISD::CTLZ, XType))) {
11450 SDValue Ctlz = DAG.getNode(ISD::CTLZ, SDLoc(N0), XType, N0);
11451 return DAG.getNode(ISD::SRL, DL, XType, Ctlz,
11452 DAG.getConstant(Log2_32(XType.getSizeInBits()),
11453 getShiftAmountTy(Ctlz.getValueType())));
11455 // fold (setgt X, 0) -> (srl (and (-X, ~X), size(X)-1))
11456 if (N1C && N1C->isNullValue() && CC == ISD::SETGT) {
11457 SDValue NegN0 = DAG.getNode(ISD::SUB, SDLoc(N0),
11458 XType, DAG.getConstant(0, XType), N0);
11459 SDValue NotN0 = DAG.getNOT(SDLoc(N0), N0, XType);
11460 return DAG.getNode(ISD::SRL, DL, XType,
11461 DAG.getNode(ISD::AND, DL, XType, NegN0, NotN0),
11462 DAG.getConstant(XType.getSizeInBits()-1,
11463 getShiftAmountTy(XType)));
11465 // fold (setgt X, -1) -> (xor (srl (X, size(X)-1), 1))
11466 if (N1C && N1C->isAllOnesValue() && CC == ISD::SETGT) {
11467 SDValue Sign = DAG.getNode(ISD::SRL, SDLoc(N0), XType, N0,
11468 DAG.getConstant(XType.getSizeInBits()-1,
11469 getShiftAmountTy(N0.getValueType())));
11470 return DAG.getNode(ISD::XOR, DL, XType, Sign, DAG.getConstant(1, XType));
11474 // Check to see if this is an integer abs.
11475 // select_cc setg[te] X, 0, X, -X ->
11476 // select_cc setgt X, -1, X, -X ->
11477 // select_cc setl[te] X, 0, -X, X ->
11478 // select_cc setlt X, 1, -X, X ->
11479 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
11481 ConstantSDNode *SubC = nullptr;
11482 if (((N1C->isNullValue() && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
11483 (N1C->isAllOnesValue() && CC == ISD::SETGT)) &&
11484 N0 == N2 && N3.getOpcode() == ISD::SUB && N0 == N3.getOperand(1))
11485 SubC = dyn_cast<ConstantSDNode>(N3.getOperand(0));
11486 else if (((N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE)) ||
11487 (N1C->isOne() && CC == ISD::SETLT)) &&
11488 N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1))
11489 SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0));
11491 EVT XType = N0.getValueType();
11492 if (SubC && SubC->isNullValue() && XType.isInteger()) {
11493 SDValue Shift = DAG.getNode(ISD::SRA, SDLoc(N0), XType,
11495 DAG.getConstant(XType.getSizeInBits()-1,
11496 getShiftAmountTy(N0.getValueType())));
11497 SDValue Add = DAG.getNode(ISD::ADD, SDLoc(N0),
11499 AddToWorklist(Shift.getNode());
11500 AddToWorklist(Add.getNode());
11501 return DAG.getNode(ISD::XOR, DL, XType, Add, Shift);
11508 /// SimplifySetCC - This is a stub for TargetLowering::SimplifySetCC.
11509 SDValue DAGCombiner::SimplifySetCC(EVT VT, SDValue N0,
11510 SDValue N1, ISD::CondCode Cond,
11511 SDLoc DL, bool foldBooleans) {
11512 TargetLowering::DAGCombinerInfo
11513 DagCombineInfo(DAG, Level, false, this);
11514 return TLI.SimplifySetCC(VT, N0, N1, Cond, foldBooleans, DagCombineInfo, DL);
11517 /// BuildSDIV - Given an ISD::SDIV node expressing a divide by constant, return
11518 /// a DAG expression to select that will generate the same value by multiplying
11519 /// by a magic number. See:
11520 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
11521 SDValue DAGCombiner::BuildSDIV(SDNode *N) {
11522 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
11526 // Avoid division by zero.
11527 if (!C->getAPIntValue())
11530 std::vector<SDNode*> Built;
11532 TLI.BuildSDIV(N, C->getAPIntValue(), DAG, LegalOperations, &Built);
11534 for (SDNode *N : Built)
11539 /// BuildSDIVPow2 - Given an ISD::SDIV node expressing a divide by constant
11540 /// power of 2, return a DAG expression to select that will generate the same
11541 /// value by right shifting.
11542 SDValue DAGCombiner::BuildSDIVPow2(SDNode *N) {
11543 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
11547 // Avoid division by zero.
11548 if (!C->getAPIntValue())
11551 std::vector<SDNode *> Built;
11552 SDValue S = TLI.BuildSDIVPow2(N, C->getAPIntValue(), DAG, &Built);
11554 for (SDNode *N : Built)
11559 /// BuildUDIV - Given an ISD::UDIV node expressing a divide by constant,
11560 /// return a DAG expression to select that will generate the same value by
11561 /// multiplying by a magic number. See:
11562 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
11563 SDValue DAGCombiner::BuildUDIV(SDNode *N) {
11564 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
11568 // Avoid division by zero.
11569 if (!C->getAPIntValue())
11572 std::vector<SDNode*> Built;
11574 TLI.BuildUDIV(N, C->getAPIntValue(), DAG, LegalOperations, &Built);
11576 for (SDNode *N : Built)
11581 /// FindBaseOffset - Return true if base is a frame index, which is known not
11582 // to alias with anything but itself. Provides base object and offset as
11584 static bool FindBaseOffset(SDValue Ptr, SDValue &Base, int64_t &Offset,
11585 const GlobalValue *&GV, const void *&CV) {
11586 // Assume it is a primitive operation.
11587 Base = Ptr; Offset = 0; GV = nullptr; CV = nullptr;
11589 // If it's an adding a simple constant then integrate the offset.
11590 if (Base.getOpcode() == ISD::ADD) {
11591 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Base.getOperand(1))) {
11592 Base = Base.getOperand(0);
11593 Offset += C->getZExtValue();
11597 // Return the underlying GlobalValue, and update the Offset. Return false
11598 // for GlobalAddressSDNode since the same GlobalAddress may be represented
11599 // by multiple nodes with different offsets.
11600 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Base)) {
11601 GV = G->getGlobal();
11602 Offset += G->getOffset();
11606 // Return the underlying Constant value, and update the Offset. Return false
11607 // for ConstantSDNodes since the same constant pool entry may be represented
11608 // by multiple nodes with different offsets.
11609 if (ConstantPoolSDNode *C = dyn_cast<ConstantPoolSDNode>(Base)) {
11610 CV = C->isMachineConstantPoolEntry() ? (const void *)C->getMachineCPVal()
11611 : (const void *)C->getConstVal();
11612 Offset += C->getOffset();
11615 // If it's any of the following then it can't alias with anything but itself.
11616 return isa<FrameIndexSDNode>(Base);
11619 /// isAlias - Return true if there is any possibility that the two addresses
11621 bool DAGCombiner::isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) const {
11622 // If they are the same then they must be aliases.
11623 if (Op0->getBasePtr() == Op1->getBasePtr()) return true;
11625 // If they are both volatile then they cannot be reordered.
11626 if (Op0->isVolatile() && Op1->isVolatile()) return true;
11628 // Gather base node and offset information.
11629 SDValue Base1, Base2;
11630 int64_t Offset1, Offset2;
11631 const GlobalValue *GV1, *GV2;
11632 const void *CV1, *CV2;
11633 bool isFrameIndex1 = FindBaseOffset(Op0->getBasePtr(),
11634 Base1, Offset1, GV1, CV1);
11635 bool isFrameIndex2 = FindBaseOffset(Op1->getBasePtr(),
11636 Base2, Offset2, GV2, CV2);
11638 // If they have a same base address then check to see if they overlap.
11639 if (Base1 == Base2 || (GV1 && (GV1 == GV2)) || (CV1 && (CV1 == CV2)))
11640 return !((Offset1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= Offset2 ||
11641 (Offset2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= Offset1);
11643 // It is possible for different frame indices to alias each other, mostly
11644 // when tail call optimization reuses return address slots for arguments.
11645 // To catch this case, look up the actual index of frame indices to compute
11646 // the real alias relationship.
11647 if (isFrameIndex1 && isFrameIndex2) {
11648 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
11649 Offset1 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base1)->getIndex());
11650 Offset2 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base2)->getIndex());
11651 return !((Offset1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= Offset2 ||
11652 (Offset2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= Offset1);
11655 // Otherwise, if we know what the bases are, and they aren't identical, then
11656 // we know they cannot alias.
11657 if ((isFrameIndex1 || CV1 || GV1) && (isFrameIndex2 || CV2 || GV2))
11660 // If we know required SrcValue1 and SrcValue2 have relatively large alignment
11661 // compared to the size and offset of the access, we may be able to prove they
11662 // do not alias. This check is conservative for now to catch cases created by
11663 // splitting vector types.
11664 if ((Op0->getOriginalAlignment() == Op1->getOriginalAlignment()) &&
11665 (Op0->getSrcValueOffset() != Op1->getSrcValueOffset()) &&
11666 (Op0->getMemoryVT().getSizeInBits() >> 3 ==
11667 Op1->getMemoryVT().getSizeInBits() >> 3) &&
11668 (Op0->getOriginalAlignment() > Op0->getMemoryVT().getSizeInBits()) >> 3) {
11669 int64_t OffAlign1 = Op0->getSrcValueOffset() % Op0->getOriginalAlignment();
11670 int64_t OffAlign2 = Op1->getSrcValueOffset() % Op1->getOriginalAlignment();
11672 // There is no overlap between these relatively aligned accesses of similar
11673 // size, return no alias.
11674 if ((OffAlign1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= OffAlign2 ||
11675 (OffAlign2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= OffAlign1)
11679 bool UseAA = CombinerGlobalAA.getNumOccurrences() > 0 ? CombinerGlobalAA :
11680 TLI.getTargetMachine().getSubtarget<TargetSubtargetInfo>().useAA();
11682 if (CombinerAAOnlyFunc.getNumOccurrences() &&
11683 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
11687 Op0->getMemOperand()->getValue() && Op1->getMemOperand()->getValue()) {
11688 // Use alias analysis information.
11689 int64_t MinOffset = std::min(Op0->getSrcValueOffset(),
11690 Op1->getSrcValueOffset());
11691 int64_t Overlap1 = (Op0->getMemoryVT().getSizeInBits() >> 3) +
11692 Op0->getSrcValueOffset() - MinOffset;
11693 int64_t Overlap2 = (Op1->getMemoryVT().getSizeInBits() >> 3) +
11694 Op1->getSrcValueOffset() - MinOffset;
11695 AliasAnalysis::AliasResult AAResult =
11696 AA.alias(AliasAnalysis::Location(Op0->getMemOperand()->getValue(),
11698 UseTBAA ? Op0->getAAInfo() : AAMDNodes()),
11699 AliasAnalysis::Location(Op1->getMemOperand()->getValue(),
11701 UseTBAA ? Op1->getAAInfo() : AAMDNodes()));
11702 if (AAResult == AliasAnalysis::NoAlias)
11706 // Otherwise we have to assume they alias.
11710 /// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes,
11711 /// looking for aliasing nodes and adding them to the Aliases vector.
11712 void DAGCombiner::GatherAllAliases(SDNode *N, SDValue OriginalChain,
11713 SmallVectorImpl<SDValue> &Aliases) {
11714 SmallVector<SDValue, 8> Chains; // List of chains to visit.
11715 SmallPtrSet<SDNode *, 16> Visited; // Visited node set.
11717 // Get alias information for node.
11718 bool IsLoad = isa<LoadSDNode>(N) && !cast<LSBaseSDNode>(N)->isVolatile();
11721 Chains.push_back(OriginalChain);
11722 unsigned Depth = 0;
11724 // Look at each chain and determine if it is an alias. If so, add it to the
11725 // aliases list. If not, then continue up the chain looking for the next
11727 while (!Chains.empty()) {
11728 SDValue Chain = Chains.back();
11731 // For TokenFactor nodes, look at each operand and only continue up the
11732 // chain until we find two aliases. If we've seen two aliases, assume we'll
11733 // find more and revert to original chain since the xform is unlikely to be
11736 // FIXME: The depth check could be made to return the last non-aliasing
11737 // chain we found before we hit a tokenfactor rather than the original
11739 if (Depth > 6 || Aliases.size() == 2) {
11741 Aliases.push_back(OriginalChain);
11745 // Don't bother if we've been before.
11746 if (!Visited.insert(Chain.getNode()))
11749 switch (Chain.getOpcode()) {
11750 case ISD::EntryToken:
11751 // Entry token is ideal chain operand, but handled in FindBetterChain.
11756 // Get alias information for Chain.
11757 bool IsOpLoad = isa<LoadSDNode>(Chain.getNode()) &&
11758 !cast<LSBaseSDNode>(Chain.getNode())->isVolatile();
11760 // If chain is alias then stop here.
11761 if (!(IsLoad && IsOpLoad) &&
11762 isAlias(cast<LSBaseSDNode>(N), cast<LSBaseSDNode>(Chain.getNode()))) {
11763 Aliases.push_back(Chain);
11765 // Look further up the chain.
11766 Chains.push_back(Chain.getOperand(0));
11772 case ISD::TokenFactor:
11773 // We have to check each of the operands of the token factor for "small"
11774 // token factors, so we queue them up. Adding the operands to the queue
11775 // (stack) in reverse order maintains the original order and increases the
11776 // likelihood that getNode will find a matching token factor (CSE.)
11777 if (Chain.getNumOperands() > 16) {
11778 Aliases.push_back(Chain);
11781 for (unsigned n = Chain.getNumOperands(); n;)
11782 Chains.push_back(Chain.getOperand(--n));
11787 // For all other instructions we will just have to take what we can get.
11788 Aliases.push_back(Chain);
11793 // We need to be careful here to also search for aliases through the
11794 // value operand of a store, etc. Consider the following situation:
11796 // L1 = load Token1, %52
11797 // S1 = store Token1, L1, %51
11798 // L2 = load Token1, %52+8
11799 // S2 = store Token1, L2, %51+8
11800 // Token2 = Token(S1, S2)
11801 // L3 = load Token2, %53
11802 // S3 = store Token2, L3, %52
11803 // L4 = load Token2, %53+8
11804 // S4 = store Token2, L4, %52+8
11805 // If we search for aliases of S3 (which loads address %52), and we look
11806 // only through the chain, then we'll miss the trivial dependence on L1
11807 // (which also loads from %52). We then might change all loads and
11808 // stores to use Token1 as their chain operand, which could result in
11809 // copying %53 into %52 before copying %52 into %51 (which should
11812 // The problem is, however, that searching for such data dependencies
11813 // can become expensive, and the cost is not directly related to the
11814 // chain depth. Instead, we'll rule out such configurations here by
11815 // insisting that we've visited all chain users (except for users
11816 // of the original chain, which is not necessary). When doing this,
11817 // we need to look through nodes we don't care about (otherwise, things
11818 // like register copies will interfere with trivial cases).
11820 SmallVector<const SDNode *, 16> Worklist;
11821 for (SmallPtrSet<SDNode *, 16>::iterator I = Visited.begin(),
11822 IE = Visited.end(); I != IE; ++I)
11823 if (*I != OriginalChain.getNode())
11824 Worklist.push_back(*I);
11826 while (!Worklist.empty()) {
11827 const SDNode *M = Worklist.pop_back_val();
11829 // We have already visited M, and want to make sure we've visited any uses
11830 // of M that we care about. For uses that we've not visisted, and don't
11831 // care about, queue them to the worklist.
11833 for (SDNode::use_iterator UI = M->use_begin(),
11834 UIE = M->use_end(); UI != UIE; ++UI)
11835 if (UI.getUse().getValueType() == MVT::Other && Visited.insert(*UI)) {
11836 if (isa<MemIntrinsicSDNode>(*UI) || isa<MemSDNode>(*UI)) {
11837 // We've not visited this use, and we care about it (it could have an
11838 // ordering dependency with the original node).
11840 Aliases.push_back(OriginalChain);
11844 // We've not visited this use, but we don't care about it. Mark it as
11845 // visited and enqueue it to the worklist.
11846 Worklist.push_back(*UI);
11851 /// FindBetterChain - Walk up chain skipping non-aliasing memory nodes, looking
11852 /// for a better chain (aliasing node.)
11853 SDValue DAGCombiner::FindBetterChain(SDNode *N, SDValue OldChain) {
11854 SmallVector<SDValue, 8> Aliases; // Ops for replacing token factor.
11856 // Accumulate all the aliases to this node.
11857 GatherAllAliases(N, OldChain, Aliases);
11859 // If no operands then chain to entry token.
11860 if (Aliases.size() == 0)
11861 return DAG.getEntryNode();
11863 // If a single operand then chain to it. We don't need to revisit it.
11864 if (Aliases.size() == 1)
11867 // Construct a custom tailored token factor.
11868 return DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, Aliases);
11871 // SelectionDAG::Combine - This is the entry point for the file.
11873 void SelectionDAG::Combine(CombineLevel Level, AliasAnalysis &AA,
11874 CodeGenOpt::Level OptLevel) {
11875 /// run - This is the main entry point to this class.
11877 DAGCombiner(*this, AA, OptLevel).Run(Level);