1 //===-- DAGCombiner.cpp - Implement a DAG node combiner -------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by Nate Begeman and is distributed under the
6 // University of Illinois Open Source 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 // FIXME: Missing folds
14 // sdiv, udiv, srem, urem (X, const) where X is an integer can be expanded into
15 // a sequence of multiplies, shifts, and adds. This should be controlled by
16 // some kind of hint from the target that int div is expensive.
17 // various folds of mulh[s,u] by constants such as -1, powers of 2, etc.
19 // FIXME: select C, pow2, pow2 -> something smart
20 // FIXME: trunc(select X, Y, Z) -> select X, trunc(Y), trunc(Z)
21 // FIXME: Dead stores -> nuke
22 // FIXME: shr X, (and Y,31) -> shr X, Y (TRICKY!)
23 // FIXME: mul (x, const) -> shifts + adds
24 // FIXME: undef values
25 // FIXME: divide by zero is currently left unfolded. do we want to turn this
27 // FIXME: select ne (select cc, 1, 0), 0, true, false -> select cc, true, false
29 //===----------------------------------------------------------------------===//
31 #define DEBUG_TYPE "dagcombine"
32 #include "llvm/ADT/Statistic.h"
33 #include "llvm/Analysis/AliasAnalysis.h"
34 #include "llvm/CodeGen/SelectionDAG.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/MathExtras.h"
37 #include "llvm/Target/TargetLowering.h"
38 #include "llvm/Support/Compiler.h"
39 #include "llvm/Support/CommandLine.h"
46 static Statistic<> NodesCombined ("dagcombiner",
47 "Number of dag nodes combined");
49 static Statistic<> PreIndexedNodes ("pre_indexed_ops",
50 "Number of pre-indexed nodes created");
51 static Statistic<> PostIndexedNodes ("post_indexed_ops",
52 "Number of post-indexed nodes created");
55 CombinerAA("combiner-alias-analysis", cl::Hidden,
56 cl::desc("Turn on alias analysis during testing"));
59 CombinerGlobalAA("combiner-global-alias-analysis", cl::Hidden,
60 cl::desc("Include global information in alias analysis"));
62 //------------------------------ DAGCombiner ---------------------------------//
64 class VISIBILITY_HIDDEN DAGCombiner {
69 // Worklist of all of the nodes that need to be simplified.
70 std::vector<SDNode*> WorkList;
72 // AA - Used for DAG load/store alias analysis.
75 /// AddUsersToWorkList - When an instruction is simplified, add all users of
76 /// the instruction to the work lists because they might get more simplified
79 void AddUsersToWorkList(SDNode *N) {
80 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
85 /// removeFromWorkList - remove all instances of N from the worklist.
87 void removeFromWorkList(SDNode *N) {
88 WorkList.erase(std::remove(WorkList.begin(), WorkList.end(), N),
93 /// AddToWorkList - Add to the work list making sure it's instance is at the
94 /// the back (next to be processed.)
95 void AddToWorkList(SDNode *N) {
96 removeFromWorkList(N);
97 WorkList.push_back(N);
100 SDOperand CombineTo(SDNode *N, const SDOperand *To, unsigned NumTo,
102 assert(N->getNumValues() == NumTo && "Broken CombineTo call!");
104 DEBUG(std::cerr << "\nReplacing.1 "; N->dump();
105 std::cerr << "\nWith: "; To[0].Val->dump(&DAG);
106 std::cerr << " and " << NumTo-1 << " other values\n");
107 std::vector<SDNode*> NowDead;
108 DAG.ReplaceAllUsesWith(N, To, &NowDead);
111 // Push the new nodes and any users onto the worklist
112 for (unsigned i = 0, e = NumTo; i != e; ++i) {
113 AddToWorkList(To[i].Val);
114 AddUsersToWorkList(To[i].Val);
118 // Nodes can be reintroduced into the worklist. Make sure we do not
119 // process a node that has been replaced.
120 removeFromWorkList(N);
121 for (unsigned i = 0, e = NowDead.size(); i != e; ++i)
122 removeFromWorkList(NowDead[i]);
124 // Finally, since the node is now dead, remove it from the graph.
126 return SDOperand(N, 0);
129 SDOperand CombineTo(SDNode *N, SDOperand Res, bool AddTo = true) {
130 return CombineTo(N, &Res, 1, AddTo);
133 SDOperand CombineTo(SDNode *N, SDOperand Res0, SDOperand Res1,
135 SDOperand To[] = { Res0, Res1 };
136 return CombineTo(N, To, 2, AddTo);
140 /// SimplifyDemandedBits - Check the specified integer node value to see if
141 /// it can be simplified or if things it uses can be simplified by bit
142 /// propagation. If so, return true.
143 bool SimplifyDemandedBits(SDOperand Op) {
144 TargetLowering::TargetLoweringOpt TLO(DAG);
145 uint64_t KnownZero, KnownOne;
146 uint64_t Demanded = MVT::getIntVTBitMask(Op.getValueType());
147 if (!TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
151 AddToWorkList(Op.Val);
153 // Replace the old value with the new one.
155 DEBUG(std::cerr << "\nReplacing.2 "; TLO.Old.Val->dump();
156 std::cerr << "\nWith: "; TLO.New.Val->dump(&DAG);
159 std::vector<SDNode*> NowDead;
160 DAG.ReplaceAllUsesOfValueWith(TLO.Old, TLO.New, NowDead);
162 // Push the new node and any (possibly new) users onto the worklist.
163 AddToWorkList(TLO.New.Val);
164 AddUsersToWorkList(TLO.New.Val);
166 // Nodes can end up on the worklist more than once. Make sure we do
167 // not process a node that has been replaced.
168 for (unsigned i = 0, e = NowDead.size(); i != e; ++i)
169 removeFromWorkList(NowDead[i]);
171 // Finally, if the node is now dead, remove it from the graph. The node
172 // may not be dead if the replacement process recursively simplified to
173 // something else needing this node.
174 if (TLO.Old.Val->use_empty()) {
175 removeFromWorkList(TLO.Old.Val);
176 DAG.DeleteNode(TLO.Old.Val);
181 /// CombineToPreIndexedLoadStore - Try turning a load / store and a
182 /// pre-indexed load / store when the base pointer is a add or subtract
183 /// and it has other uses besides the load / store. After the
184 /// transformation, the new indexed load / store has effectively folded
185 /// the add / subtract in and all of its other uses are redirected to the
186 /// new load / store.
187 bool CombineToPreIndexedLoadStore(SDNode *N) {
193 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
194 Ptr = LD->getBasePtr();
195 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
196 Ptr = ST->getBasePtr();
201 if ((Ptr.getOpcode() == ISD::ADD || Ptr.getOpcode() == ISD::SUB) &&
202 Ptr.Val->use_size() > 1) {
205 ISD::MemOpAddrMode AM = ISD::UNINDEXED;
206 if (TLI.getPreIndexedAddressParts(N, BasePtr, Offset, AM, DAG)) {
207 // Try turning it into a pre-indexed load / store except when
208 // 1) Another use of base ptr is a predecessor of N. If ptr is folded
209 // that would create a cycle.
210 // 2) All uses are load / store ops that use it as base ptr.
212 // Now check for #1 and #2.
213 bool RealUse = false;
214 for (SDNode::use_iterator I = Ptr.Val->use_begin(),
215 E = Ptr.Val->use_end(); I != E; ++I) {
219 if (Use->isPredecessor(N))
222 if (!((Use->getOpcode() == ISD::LOAD &&
223 cast<LoadSDNode>(Use)->getBasePtr() == Ptr) ||
224 (Use->getOpcode() == ISD::STORE) &&
225 cast<StoreSDNode>(Use)->getBasePtr() == Ptr))
231 SDOperand Result = isLoad
232 ? DAG.getIndexedLoad(SDOperand(N,0), BasePtr, Offset, AM)
233 : DAG.getIndexedStore(SDOperand(N,0), BasePtr, Offset, AM);
236 DEBUG(std::cerr << "\nReplacing.4 "; N->dump();
237 std::cerr << "\nWith: "; Result.Val->dump(&DAG);
239 std::vector<SDNode*> NowDead;
241 DAG.ReplaceAllUsesOfValueWith(SDOperand(N, 0), Result.getValue(0),
243 DAG.ReplaceAllUsesOfValueWith(SDOperand(N, 1), Result.getValue(2),
246 DAG.ReplaceAllUsesOfValueWith(SDOperand(N, 0), Result.getValue(1),
250 // Nodes can end up on the worklist more than once. Make sure we do
251 // not process a node that has been replaced.
252 for (unsigned i = 0, e = NowDead.size(); i != e; ++i)
253 removeFromWorkList(NowDead[i]);
254 // Finally, since the node is now dead, remove it from the graph.
257 // Replace the uses of Ptr with uses of the updated base value.
258 DAG.ReplaceAllUsesOfValueWith(Ptr, Result.getValue(isLoad ? 1 : 0),
260 removeFromWorkList(Ptr.Val);
261 for (unsigned i = 0, e = NowDead.size(); i != e; ++i)
262 removeFromWorkList(NowDead[i]);
263 DAG.DeleteNode(Ptr.Val);
271 /// CombineToPostIndexedLoadStore - Try combine a load / store with a
272 /// add / sub of the base pointer node into a post-indexed load / store.
273 /// The transformation folded the add / subtract into the new indexed
274 /// load / store effectively and all of its uses are redirected to the
275 /// new load / store.
276 bool CombineToPostIndexedLoadStore(SDNode *N) {
283 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
284 Ptr = LD->getBasePtr();
285 VT = LD->getLoadedVT();
286 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
287 Ptr = ST->getBasePtr();
288 VT = ST->getStoredVT();
293 if (Ptr.Val->use_size() > 1) {
294 for (SDNode::use_iterator I = Ptr.Val->use_begin(),
295 E = Ptr.Val->use_end(); I != E; ++I) {
298 (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB))
303 ISD::MemOpAddrMode AM = ISD::UNINDEXED;
304 if (TLI.getPostIndexedAddressParts(Op, VT, BasePtr, Offset, AM,DAG) &&
306 // Try turning it into a post-indexed load / store except when
307 // 1) Op must be independent of N, i.e. Op is neither a predecessor
308 // nor a successor of N. Otherwise, if Op is folded that would
310 // 2) All uses are load / store ops that use it as base ptr.
313 bool TryNext = false;
314 for (SDNode::use_iterator II = BasePtr.Val->use_begin(),
315 EE = BasePtr.Val->use_end(); II != EE; ++II) {
320 // If all the uses are load / store addresses, then don't do the
322 if (Use->getOpcode() == ISD::ADD || Use->getOpcode() == ISD::SUB){
323 bool RealUse = false;
324 for (SDNode::use_iterator III = Use->use_begin(),
325 EEE = Use->use_end(); III != EEE; ++III) {
326 SDNode *UseUse = *III;
327 if (!((UseUse->getOpcode() == ISD::LOAD &&
328 cast<LoadSDNode>(UseUse)->getBasePtr().Val == Use) ||
329 (UseUse->getOpcode() == ISD::STORE) &&
330 cast<StoreSDNode>(UseUse)->getBasePtr().Val == Use))
344 if (!Op->isPredecessor(N) && !N->isPredecessor(Op)) {
345 SDOperand Result = isLoad
346 ? DAG.getIndexedLoad(SDOperand(N,0), BasePtr, Offset, AM)
347 : DAG.getIndexedStore(SDOperand(N,0), BasePtr, Offset, AM);
350 DEBUG(std::cerr << "\nReplacing.5 "; N->dump();
351 std::cerr << "\nWith: "; Result.Val->dump(&DAG);
353 std::vector<SDNode*> NowDead;
355 DAG.ReplaceAllUsesOfValueWith(SDOperand(N, 0), Result.getValue(0),
357 DAG.ReplaceAllUsesOfValueWith(SDOperand(N, 1), Result.getValue(2),
360 DAG.ReplaceAllUsesOfValueWith(SDOperand(N, 0), Result.getValue(1),
364 // Nodes can end up on the worklist more than once. Make sure we do
365 // not process a node that has been replaced.
366 for (unsigned i = 0, e = NowDead.size(); i != e; ++i)
367 removeFromWorkList(NowDead[i]);
368 // Finally, since the node is now dead, remove it from the graph.
371 // Replace the uses of Use with uses of the updated base value.
372 DAG.ReplaceAllUsesOfValueWith(SDOperand(Op, 0),
373 Result.getValue(isLoad ? 1 : 0),
375 removeFromWorkList(Op);
376 for (unsigned i = 0, e = NowDead.size(); i != e; ++i)
377 removeFromWorkList(NowDead[i]);
388 /// visit - call the node-specific routine that knows how to fold each
389 /// particular type of node.
390 SDOperand visit(SDNode *N);
392 // Visitation implementation - Implement dag node combining for different
393 // node types. The semantics are as follows:
395 // SDOperand.Val == 0 - No change was made
396 // SDOperand.Val == N - N was replaced, is dead, and is already handled.
397 // otherwise - N should be replaced by the returned Operand.
399 SDOperand visitTokenFactor(SDNode *N);
400 SDOperand visitADD(SDNode *N);
401 SDOperand visitSUB(SDNode *N);
402 SDOperand visitMUL(SDNode *N);
403 SDOperand visitSDIV(SDNode *N);
404 SDOperand visitUDIV(SDNode *N);
405 SDOperand visitSREM(SDNode *N);
406 SDOperand visitUREM(SDNode *N);
407 SDOperand visitMULHU(SDNode *N);
408 SDOperand visitMULHS(SDNode *N);
409 SDOperand visitAND(SDNode *N);
410 SDOperand visitOR(SDNode *N);
411 SDOperand visitXOR(SDNode *N);
412 SDOperand visitVBinOp(SDNode *N, ISD::NodeType IntOp, ISD::NodeType FPOp);
413 SDOperand visitSHL(SDNode *N);
414 SDOperand visitSRA(SDNode *N);
415 SDOperand visitSRL(SDNode *N);
416 SDOperand visitCTLZ(SDNode *N);
417 SDOperand visitCTTZ(SDNode *N);
418 SDOperand visitCTPOP(SDNode *N);
419 SDOperand visitSELECT(SDNode *N);
420 SDOperand visitSELECT_CC(SDNode *N);
421 SDOperand visitSETCC(SDNode *N);
422 SDOperand visitSIGN_EXTEND(SDNode *N);
423 SDOperand visitZERO_EXTEND(SDNode *N);
424 SDOperand visitANY_EXTEND(SDNode *N);
425 SDOperand visitSIGN_EXTEND_INREG(SDNode *N);
426 SDOperand visitTRUNCATE(SDNode *N);
427 SDOperand visitBIT_CONVERT(SDNode *N);
428 SDOperand visitVBIT_CONVERT(SDNode *N);
429 SDOperand visitFADD(SDNode *N);
430 SDOperand visitFSUB(SDNode *N);
431 SDOperand visitFMUL(SDNode *N);
432 SDOperand visitFDIV(SDNode *N);
433 SDOperand visitFREM(SDNode *N);
434 SDOperand visitFCOPYSIGN(SDNode *N);
435 SDOperand visitSINT_TO_FP(SDNode *N);
436 SDOperand visitUINT_TO_FP(SDNode *N);
437 SDOperand visitFP_TO_SINT(SDNode *N);
438 SDOperand visitFP_TO_UINT(SDNode *N);
439 SDOperand visitFP_ROUND(SDNode *N);
440 SDOperand visitFP_ROUND_INREG(SDNode *N);
441 SDOperand visitFP_EXTEND(SDNode *N);
442 SDOperand visitFNEG(SDNode *N);
443 SDOperand visitFABS(SDNode *N);
444 SDOperand visitBRCOND(SDNode *N);
445 SDOperand visitBR_CC(SDNode *N);
446 SDOperand visitLOAD(SDNode *N);
447 SDOperand visitSTORE(SDNode *N);
448 SDOperand visitINSERT_VECTOR_ELT(SDNode *N);
449 SDOperand visitVINSERT_VECTOR_ELT(SDNode *N);
450 SDOperand visitVBUILD_VECTOR(SDNode *N);
451 SDOperand visitVECTOR_SHUFFLE(SDNode *N);
452 SDOperand visitVVECTOR_SHUFFLE(SDNode *N);
454 SDOperand XformToShuffleWithZero(SDNode *N);
455 SDOperand ReassociateOps(unsigned Opc, SDOperand LHS, SDOperand RHS);
457 bool SimplifySelectOps(SDNode *SELECT, SDOperand LHS, SDOperand RHS);
458 SDOperand SimplifyBinOpWithSameOpcodeHands(SDNode *N);
459 SDOperand SimplifySelect(SDOperand N0, SDOperand N1, SDOperand N2);
460 SDOperand SimplifySelectCC(SDOperand N0, SDOperand N1, SDOperand N2,
461 SDOperand N3, ISD::CondCode CC);
462 SDOperand SimplifySetCC(MVT::ValueType VT, SDOperand N0, SDOperand N1,
463 ISD::CondCode Cond, bool foldBooleans = true);
464 SDOperand ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(SDNode *, MVT::ValueType);
465 SDOperand BuildSDIV(SDNode *N);
466 SDOperand BuildUDIV(SDNode *N);
467 SDNode *MatchRotate(SDOperand LHS, SDOperand RHS);
469 /// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes,
470 /// looking for aliasing nodes and adding them to the Aliases vector.
471 void GatherAllAliases(SDNode *N, SDOperand OriginalChain,
472 SmallVector<SDOperand, 8> &Aliases);
474 /// isAlias - Return true if there is any possibility that the two addresses
476 bool isAlias(SDOperand Ptr1, int64_t Size1,
477 const Value *SrcValue1, int SrcValueOffset1,
478 SDOperand Ptr2, int64_t Size2,
479 const Value *SrcValue2, int SrcValueOffset2);
481 /// FindAliasInfo - Extracts the relevant alias information from the memory
482 /// node. Returns true if the operand was a load.
483 bool FindAliasInfo(SDNode *N,
484 SDOperand &Ptr, int64_t &Size,
485 const Value *&SrcValue, int &SrcValueOffset);
487 /// FindBetterChain - Walk up chain skipping non-aliasing memory nodes,
488 /// looking for a better chain (aliasing node.)
489 SDOperand FindBetterChain(SDNode *N, SDOperand Chain);
492 DAGCombiner(SelectionDAG &D, AliasAnalysis &A)
494 TLI(D.getTargetLoweringInfo()),
495 AfterLegalize(false),
498 /// Run - runs the dag combiner on all nodes in the work list
499 void Run(bool RunningAfterLegalize);
503 //===----------------------------------------------------------------------===//
504 // TargetLowering::DAGCombinerInfo implementation
505 //===----------------------------------------------------------------------===//
507 void TargetLowering::DAGCombinerInfo::AddToWorklist(SDNode *N) {
508 ((DAGCombiner*)DC)->AddToWorkList(N);
511 SDOperand TargetLowering::DAGCombinerInfo::
512 CombineTo(SDNode *N, const std::vector<SDOperand> &To) {
513 return ((DAGCombiner*)DC)->CombineTo(N, &To[0], To.size());
516 SDOperand TargetLowering::DAGCombinerInfo::
517 CombineTo(SDNode *N, SDOperand Res) {
518 return ((DAGCombiner*)DC)->CombineTo(N, Res);
522 SDOperand TargetLowering::DAGCombinerInfo::
523 CombineTo(SDNode *N, SDOperand Res0, SDOperand Res1) {
524 return ((DAGCombiner*)DC)->CombineTo(N, Res0, Res1);
530 //===----------------------------------------------------------------------===//
533 // isSetCCEquivalent - Return true if this node is a setcc, or is a select_cc
534 // that selects between the values 1 and 0, making it equivalent to a setcc.
535 // Also, set the incoming LHS, RHS, and CC references to the appropriate
536 // nodes based on the type of node we are checking. This simplifies life a
537 // bit for the callers.
538 static bool isSetCCEquivalent(SDOperand N, SDOperand &LHS, SDOperand &RHS,
540 if (N.getOpcode() == ISD::SETCC) {
541 LHS = N.getOperand(0);
542 RHS = N.getOperand(1);
543 CC = N.getOperand(2);
546 if (N.getOpcode() == ISD::SELECT_CC &&
547 N.getOperand(2).getOpcode() == ISD::Constant &&
548 N.getOperand(3).getOpcode() == ISD::Constant &&
549 cast<ConstantSDNode>(N.getOperand(2))->getValue() == 1 &&
550 cast<ConstantSDNode>(N.getOperand(3))->isNullValue()) {
551 LHS = N.getOperand(0);
552 RHS = N.getOperand(1);
553 CC = N.getOperand(4);
559 // isOneUseSetCC - Return true if this is a SetCC-equivalent operation with only
560 // one use. If this is true, it allows the users to invert the operation for
561 // free when it is profitable to do so.
562 static bool isOneUseSetCC(SDOperand N) {
563 SDOperand N0, N1, N2;
564 if (isSetCCEquivalent(N, N0, N1, N2) && N.Val->hasOneUse())
569 SDOperand DAGCombiner::ReassociateOps(unsigned Opc, SDOperand N0, SDOperand N1){
570 MVT::ValueType VT = N0.getValueType();
571 // reassoc. (op (op x, c1), y) -> (op (op x, y), c1) iff x+c1 has one use
572 // reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2))
573 if (N0.getOpcode() == Opc && isa<ConstantSDNode>(N0.getOperand(1))) {
574 if (isa<ConstantSDNode>(N1)) {
575 SDOperand OpNode = DAG.getNode(Opc, VT, N0.getOperand(1), N1);
576 AddToWorkList(OpNode.Val);
577 return DAG.getNode(Opc, VT, OpNode, N0.getOperand(0));
578 } else if (N0.hasOneUse()) {
579 SDOperand OpNode = DAG.getNode(Opc, VT, N0.getOperand(0), N1);
580 AddToWorkList(OpNode.Val);
581 return DAG.getNode(Opc, VT, OpNode, N0.getOperand(1));
584 // reassoc. (op y, (op x, c1)) -> (op (op x, y), c1) iff x+c1 has one use
585 // reassoc. (op c2, (op x, c1)) -> (op x, (op c1, c2))
586 if (N1.getOpcode() == Opc && isa<ConstantSDNode>(N1.getOperand(1))) {
587 if (isa<ConstantSDNode>(N0)) {
588 SDOperand OpNode = DAG.getNode(Opc, VT, N1.getOperand(1), N0);
589 AddToWorkList(OpNode.Val);
590 return DAG.getNode(Opc, VT, OpNode, N1.getOperand(0));
591 } else if (N1.hasOneUse()) {
592 SDOperand OpNode = DAG.getNode(Opc, VT, N1.getOperand(0), N0);
593 AddToWorkList(OpNode.Val);
594 return DAG.getNode(Opc, VT, OpNode, N1.getOperand(1));
600 void DAGCombiner::Run(bool RunningAfterLegalize) {
601 // set the instance variable, so that the various visit routines may use it.
602 AfterLegalize = RunningAfterLegalize;
604 // Add all the dag nodes to the worklist.
605 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
606 E = DAG.allnodes_end(); I != E; ++I)
607 WorkList.push_back(I);
609 // Create a dummy node (which is not added to allnodes), that adds a reference
610 // to the root node, preventing it from being deleted, and tracking any
611 // changes of the root.
612 HandleSDNode Dummy(DAG.getRoot());
614 // The root of the dag may dangle to deleted nodes until the dag combiner is
615 // done. Set it to null to avoid confusion.
616 DAG.setRoot(SDOperand());
618 /// DagCombineInfo - Expose the DAG combiner to the target combiner impls.
619 TargetLowering::DAGCombinerInfo
620 DagCombineInfo(DAG, !RunningAfterLegalize, this);
622 // while the worklist isn't empty, inspect the node on the end of it and
623 // try and combine it.
624 while (!WorkList.empty()) {
625 SDNode *N = WorkList.back();
628 // If N has no uses, it is dead. Make sure to revisit all N's operands once
629 // N is deleted from the DAG, since they too may now be dead or may have a
630 // reduced number of uses, allowing other xforms.
631 if (N->use_empty() && N != &Dummy) {
632 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
633 AddToWorkList(N->getOperand(i).Val);
639 SDOperand RV = visit(N);
641 // If nothing happened, try a target-specific DAG combine.
643 assert(N->getOpcode() != ISD::DELETED_NODE &&
644 "Node was deleted but visit returned NULL!");
645 if (N->getOpcode() >= ISD::BUILTIN_OP_END ||
646 TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode()))
647 RV = TLI.PerformDAGCombine(N, DagCombineInfo);
652 // If we get back the same node we passed in, rather than a new node or
653 // zero, we know that the node must have defined multiple values and
654 // CombineTo was used. Since CombineTo takes care of the worklist
655 // mechanics for us, we have no work to do in this case.
657 assert(N->getOpcode() != ISD::DELETED_NODE &&
658 RV.Val->getOpcode() != ISD::DELETED_NODE &&
659 "Node was deleted but visit returned new node!");
661 DEBUG(std::cerr << "\nReplacing.3 "; N->dump();
662 std::cerr << "\nWith: "; RV.Val->dump(&DAG);
664 std::vector<SDNode*> NowDead;
665 if (N->getNumValues() == RV.Val->getNumValues())
666 DAG.ReplaceAllUsesWith(N, RV.Val, &NowDead);
668 assert(N->getValueType(0) == RV.getValueType() && "Type mismatch");
670 DAG.ReplaceAllUsesWith(N, &OpV, &NowDead);
673 // Push the new node and any users onto the worklist
674 AddToWorkList(RV.Val);
675 AddUsersToWorkList(RV.Val);
677 // Nodes can be reintroduced into the worklist. Make sure we do not
678 // process a node that has been replaced.
679 removeFromWorkList(N);
680 for (unsigned i = 0, e = NowDead.size(); i != e; ++i)
681 removeFromWorkList(NowDead[i]);
683 // Finally, since the node is now dead, remove it from the graph.
689 // If the root changed (e.g. it was a dead load, update the root).
690 DAG.setRoot(Dummy.getValue());
693 SDOperand DAGCombiner::visit(SDNode *N) {
694 switch(N->getOpcode()) {
696 case ISD::TokenFactor: return visitTokenFactor(N);
697 case ISD::ADD: return visitADD(N);
698 case ISD::SUB: return visitSUB(N);
699 case ISD::MUL: return visitMUL(N);
700 case ISD::SDIV: return visitSDIV(N);
701 case ISD::UDIV: return visitUDIV(N);
702 case ISD::SREM: return visitSREM(N);
703 case ISD::UREM: return visitUREM(N);
704 case ISD::MULHU: return visitMULHU(N);
705 case ISD::MULHS: return visitMULHS(N);
706 case ISD::AND: return visitAND(N);
707 case ISD::OR: return visitOR(N);
708 case ISD::XOR: return visitXOR(N);
709 case ISD::SHL: return visitSHL(N);
710 case ISD::SRA: return visitSRA(N);
711 case ISD::SRL: return visitSRL(N);
712 case ISD::CTLZ: return visitCTLZ(N);
713 case ISD::CTTZ: return visitCTTZ(N);
714 case ISD::CTPOP: return visitCTPOP(N);
715 case ISD::SELECT: return visitSELECT(N);
716 case ISD::SELECT_CC: return visitSELECT_CC(N);
717 case ISD::SETCC: return visitSETCC(N);
718 case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N);
719 case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N);
720 case ISD::ANY_EXTEND: return visitANY_EXTEND(N);
721 case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N);
722 case ISD::TRUNCATE: return visitTRUNCATE(N);
723 case ISD::BIT_CONVERT: return visitBIT_CONVERT(N);
724 case ISD::VBIT_CONVERT: return visitVBIT_CONVERT(N);
725 case ISD::FADD: return visitFADD(N);
726 case ISD::FSUB: return visitFSUB(N);
727 case ISD::FMUL: return visitFMUL(N);
728 case ISD::FDIV: return visitFDIV(N);
729 case ISD::FREM: return visitFREM(N);
730 case ISD::FCOPYSIGN: return visitFCOPYSIGN(N);
731 case ISD::SINT_TO_FP: return visitSINT_TO_FP(N);
732 case ISD::UINT_TO_FP: return visitUINT_TO_FP(N);
733 case ISD::FP_TO_SINT: return visitFP_TO_SINT(N);
734 case ISD::FP_TO_UINT: return visitFP_TO_UINT(N);
735 case ISD::FP_ROUND: return visitFP_ROUND(N);
736 case ISD::FP_ROUND_INREG: return visitFP_ROUND_INREG(N);
737 case ISD::FP_EXTEND: return visitFP_EXTEND(N);
738 case ISD::FNEG: return visitFNEG(N);
739 case ISD::FABS: return visitFABS(N);
740 case ISD::BRCOND: return visitBRCOND(N);
741 case ISD::BR_CC: return visitBR_CC(N);
742 case ISD::LOAD: return visitLOAD(N);
743 case ISD::STORE: return visitSTORE(N);
744 case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N);
745 case ISD::VINSERT_VECTOR_ELT: return visitVINSERT_VECTOR_ELT(N);
746 case ISD::VBUILD_VECTOR: return visitVBUILD_VECTOR(N);
747 case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N);
748 case ISD::VVECTOR_SHUFFLE: return visitVVECTOR_SHUFFLE(N);
749 case ISD::VADD: return visitVBinOp(N, ISD::ADD , ISD::FADD);
750 case ISD::VSUB: return visitVBinOp(N, ISD::SUB , ISD::FSUB);
751 case ISD::VMUL: return visitVBinOp(N, ISD::MUL , ISD::FMUL);
752 case ISD::VSDIV: return visitVBinOp(N, ISD::SDIV, ISD::FDIV);
753 case ISD::VUDIV: return visitVBinOp(N, ISD::UDIV, ISD::UDIV);
754 case ISD::VAND: return visitVBinOp(N, ISD::AND , ISD::AND);
755 case ISD::VOR: return visitVBinOp(N, ISD::OR , ISD::OR);
756 case ISD::VXOR: return visitVBinOp(N, ISD::XOR , ISD::XOR);
761 /// getInputChainForNode - Given a node, return its input chain if it has one,
762 /// otherwise return a null sd operand.
763 static SDOperand getInputChainForNode(SDNode *N) {
764 if (unsigned NumOps = N->getNumOperands()) {
765 if (N->getOperand(0).getValueType() == MVT::Other)
766 return N->getOperand(0);
767 else if (N->getOperand(NumOps-1).getValueType() == MVT::Other)
768 return N->getOperand(NumOps-1);
769 for (unsigned i = 1; i < NumOps-1; ++i)
770 if (N->getOperand(i).getValueType() == MVT::Other)
771 return N->getOperand(i);
773 return SDOperand(0, 0);
776 SDOperand DAGCombiner::visitTokenFactor(SDNode *N) {
777 // If N has two operands, where one has an input chain equal to the other,
778 // the 'other' chain is redundant.
779 if (N->getNumOperands() == 2) {
780 if (getInputChainForNode(N->getOperand(0).Val) == N->getOperand(1))
781 return N->getOperand(0);
782 if (getInputChainForNode(N->getOperand(1).Val) == N->getOperand(0))
783 return N->getOperand(1);
787 SmallVector<SDNode *, 8> TFs; // List of token factors to visit.
788 SmallVector<SDOperand, 8> Ops; // Ops for replacing token factor.
789 bool Changed = false; // If we should replace this token factor.
791 // Start out with this token factor.
794 // Iterate through token factors. The TFs grows when new token factors are
796 for (unsigned i = 0; i < TFs.size(); ++i) {
799 // Check each of the operands.
800 for (unsigned i = 0, ie = TF->getNumOperands(); i != ie; ++i) {
801 SDOperand Op = TF->getOperand(i);
803 switch (Op.getOpcode()) {
804 case ISD::EntryToken:
805 // Entry tokens don't need to be added to the list. They are
810 case ISD::TokenFactor:
811 if ((CombinerAA || Op.hasOneUse()) &&
812 std::find(TFs.begin(), TFs.end(), Op.Val) == TFs.end()) {
813 // Queue up for processing.
814 TFs.push_back(Op.Val);
815 // Clean up in case the token factor is removed.
816 AddToWorkList(Op.Val);
823 // Only add if not there prior.
824 if (std::find(Ops.begin(), Ops.end(), Op) == Ops.end())
833 // If we've change things around then replace token factor.
835 if (Ops.size() == 0) {
836 // The entry token is the only possible outcome.
837 Result = DAG.getEntryNode();
839 // New and improved token factor.
840 Result = DAG.getNode(ISD::TokenFactor, MVT::Other, &Ops[0], Ops.size());
843 // Don't add users to work list.
844 return CombineTo(N, Result, false);
850 SDOperand DAGCombiner::visitADD(SDNode *N) {
851 SDOperand N0 = N->getOperand(0);
852 SDOperand N1 = N->getOperand(1);
853 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
854 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
855 MVT::ValueType VT = N0.getValueType();
857 // fold (add c1, c2) -> c1+c2
859 return DAG.getNode(ISD::ADD, VT, N0, N1);
860 // canonicalize constant to RHS
862 return DAG.getNode(ISD::ADD, VT, N1, N0);
863 // fold (add x, 0) -> x
864 if (N1C && N1C->isNullValue())
866 // fold ((c1-A)+c2) -> (c1+c2)-A
867 if (N1C && N0.getOpcode() == ISD::SUB)
868 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getOperand(0)))
869 return DAG.getNode(ISD::SUB, VT,
870 DAG.getConstant(N1C->getValue()+N0C->getValue(), VT),
873 SDOperand RADD = ReassociateOps(ISD::ADD, N0, N1);
876 // fold ((0-A) + B) -> B-A
877 if (N0.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N0.getOperand(0)) &&
878 cast<ConstantSDNode>(N0.getOperand(0))->isNullValue())
879 return DAG.getNode(ISD::SUB, VT, N1, N0.getOperand(1));
880 // fold (A + (0-B)) -> A-B
881 if (N1.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N1.getOperand(0)) &&
882 cast<ConstantSDNode>(N1.getOperand(0))->isNullValue())
883 return DAG.getNode(ISD::SUB, VT, N0, N1.getOperand(1));
884 // fold (A+(B-A)) -> B
885 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1))
886 return N1.getOperand(0);
888 if (!MVT::isVector(VT) && SimplifyDemandedBits(SDOperand(N, 0)))
889 return SDOperand(N, 0);
891 // fold (a+b) -> (a|b) iff a and b share no bits.
892 if (MVT::isInteger(VT) && !MVT::isVector(VT)) {
893 uint64_t LHSZero, LHSOne;
894 uint64_t RHSZero, RHSOne;
895 uint64_t Mask = MVT::getIntVTBitMask(VT);
896 TLI.ComputeMaskedBits(N0, Mask, LHSZero, LHSOne);
898 TLI.ComputeMaskedBits(N1, Mask, RHSZero, RHSOne);
900 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
901 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
902 if ((RHSZero & (~LHSZero & Mask)) == (~LHSZero & Mask) ||
903 (LHSZero & (~RHSZero & Mask)) == (~RHSZero & Mask))
904 return DAG.getNode(ISD::OR, VT, N0, N1);
911 SDOperand DAGCombiner::visitSUB(SDNode *N) {
912 SDOperand N0 = N->getOperand(0);
913 SDOperand N1 = N->getOperand(1);
914 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val);
915 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val);
916 MVT::ValueType VT = N0.getValueType();
918 // fold (sub x, x) -> 0
920 return DAG.getConstant(0, N->getValueType(0));
921 // fold (sub c1, c2) -> c1-c2
923 return DAG.getNode(ISD::SUB, VT, N0, N1);
924 // fold (sub x, c) -> (add x, -c)
926 return DAG.getNode(ISD::ADD, VT, N0, DAG.getConstant(-N1C->getValue(), VT));
928 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1)
929 return N0.getOperand(1);
931 if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1)
932 return N0.getOperand(0);
936 SDOperand DAGCombiner::visitMUL(SDNode *N) {
937 SDOperand N0 = N->getOperand(0);
938 SDOperand N1 = N->getOperand(1);
939 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
940 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
941 MVT::ValueType VT = N0.getValueType();
943 // fold (mul c1, c2) -> c1*c2
945 return DAG.getNode(ISD::MUL, VT, N0, N1);
946 // canonicalize constant to RHS
948 return DAG.getNode(ISD::MUL, VT, N1, N0);
949 // fold (mul x, 0) -> 0
950 if (N1C && N1C->isNullValue())
952 // fold (mul x, -1) -> 0-x
953 if (N1C && N1C->isAllOnesValue())
954 return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT), N0);
955 // fold (mul x, (1 << c)) -> x << c
956 if (N1C && isPowerOf2_64(N1C->getValue()))
957 return DAG.getNode(ISD::SHL, VT, N0,
958 DAG.getConstant(Log2_64(N1C->getValue()),
959 TLI.getShiftAmountTy()));
960 // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c
961 if (N1C && isPowerOf2_64(-N1C->getSignExtended())) {
962 // FIXME: If the input is something that is easily negated (e.g. a
963 // single-use add), we should put the negate there.
964 return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT),
965 DAG.getNode(ISD::SHL, VT, N0,
966 DAG.getConstant(Log2_64(-N1C->getSignExtended()),
967 TLI.getShiftAmountTy())));
970 // (mul (shl X, c1), c2) -> (mul X, c2 << c1)
971 if (N1C && N0.getOpcode() == ISD::SHL &&
972 isa<ConstantSDNode>(N0.getOperand(1))) {
973 SDOperand C3 = DAG.getNode(ISD::SHL, VT, N1, N0.getOperand(1));
974 AddToWorkList(C3.Val);
975 return DAG.getNode(ISD::MUL, VT, N0.getOperand(0), C3);
978 // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one
981 SDOperand Sh(0,0), Y(0,0);
982 // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)).
983 if (N0.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N0.getOperand(1)) &&
984 N0.Val->hasOneUse()) {
986 } else if (N1.getOpcode() == ISD::SHL &&
987 isa<ConstantSDNode>(N1.getOperand(1)) && N1.Val->hasOneUse()) {
991 SDOperand Mul = DAG.getNode(ISD::MUL, VT, Sh.getOperand(0), Y);
992 return DAG.getNode(ISD::SHL, VT, Mul, Sh.getOperand(1));
995 // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2)
996 if (N1C && N0.getOpcode() == ISD::ADD && N0.Val->hasOneUse() &&
997 isa<ConstantSDNode>(N0.getOperand(1))) {
998 return DAG.getNode(ISD::ADD, VT,
999 DAG.getNode(ISD::MUL, VT, N0.getOperand(0), N1),
1000 DAG.getNode(ISD::MUL, VT, N0.getOperand(1), N1));
1004 SDOperand RMUL = ReassociateOps(ISD::MUL, N0, N1);
1010 SDOperand DAGCombiner::visitSDIV(SDNode *N) {
1011 SDOperand N0 = N->getOperand(0);
1012 SDOperand N1 = N->getOperand(1);
1013 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val);
1014 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val);
1015 MVT::ValueType VT = N->getValueType(0);
1017 // fold (sdiv c1, c2) -> c1/c2
1018 if (N0C && N1C && !N1C->isNullValue())
1019 return DAG.getNode(ISD::SDIV, VT, N0, N1);
1020 // fold (sdiv X, 1) -> X
1021 if (N1C && N1C->getSignExtended() == 1LL)
1023 // fold (sdiv X, -1) -> 0-X
1024 if (N1C && N1C->isAllOnesValue())
1025 return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT), N0);
1026 // If we know the sign bits of both operands are zero, strength reduce to a
1027 // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2
1028 uint64_t SignBit = 1ULL << (MVT::getSizeInBits(VT)-1);
1029 if (TLI.MaskedValueIsZero(N1, SignBit) &&
1030 TLI.MaskedValueIsZero(N0, SignBit))
1031 return DAG.getNode(ISD::UDIV, N1.getValueType(), N0, N1);
1032 // fold (sdiv X, pow2) -> simple ops after legalize
1033 if (N1C && N1C->getValue() && !TLI.isIntDivCheap() &&
1034 (isPowerOf2_64(N1C->getSignExtended()) ||
1035 isPowerOf2_64(-N1C->getSignExtended()))) {
1036 // If dividing by powers of two is cheap, then don't perform the following
1038 if (TLI.isPow2DivCheap())
1040 int64_t pow2 = N1C->getSignExtended();
1041 int64_t abs2 = pow2 > 0 ? pow2 : -pow2;
1042 unsigned lg2 = Log2_64(abs2);
1043 // Splat the sign bit into the register
1044 SDOperand SGN = DAG.getNode(ISD::SRA, VT, N0,
1045 DAG.getConstant(MVT::getSizeInBits(VT)-1,
1046 TLI.getShiftAmountTy()));
1047 AddToWorkList(SGN.Val);
1048 // Add (N0 < 0) ? abs2 - 1 : 0;
1049 SDOperand SRL = DAG.getNode(ISD::SRL, VT, SGN,
1050 DAG.getConstant(MVT::getSizeInBits(VT)-lg2,
1051 TLI.getShiftAmountTy()));
1052 SDOperand ADD = DAG.getNode(ISD::ADD, VT, N0, SRL);
1053 AddToWorkList(SRL.Val);
1054 AddToWorkList(ADD.Val); // Divide by pow2
1055 SDOperand SRA = DAG.getNode(ISD::SRA, VT, ADD,
1056 DAG.getConstant(lg2, TLI.getShiftAmountTy()));
1057 // If we're dividing by a positive value, we're done. Otherwise, we must
1058 // negate the result.
1061 AddToWorkList(SRA.Val);
1062 return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT), SRA);
1064 // if integer divide is expensive and we satisfy the requirements, emit an
1065 // alternate sequence.
1066 if (N1C && (N1C->getSignExtended() < -1 || N1C->getSignExtended() > 1) &&
1067 !TLI.isIntDivCheap()) {
1068 SDOperand Op = BuildSDIV(N);
1069 if (Op.Val) return Op;
1074 SDOperand DAGCombiner::visitUDIV(SDNode *N) {
1075 SDOperand N0 = N->getOperand(0);
1076 SDOperand N1 = N->getOperand(1);
1077 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val);
1078 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val);
1079 MVT::ValueType VT = N->getValueType(0);
1081 // fold (udiv c1, c2) -> c1/c2
1082 if (N0C && N1C && !N1C->isNullValue())
1083 return DAG.getNode(ISD::UDIV, VT, N0, N1);
1084 // fold (udiv x, (1 << c)) -> x >>u c
1085 if (N1C && isPowerOf2_64(N1C->getValue()))
1086 return DAG.getNode(ISD::SRL, VT, N0,
1087 DAG.getConstant(Log2_64(N1C->getValue()),
1088 TLI.getShiftAmountTy()));
1089 // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2
1090 if (N1.getOpcode() == ISD::SHL) {
1091 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
1092 if (isPowerOf2_64(SHC->getValue())) {
1093 MVT::ValueType ADDVT = N1.getOperand(1).getValueType();
1094 SDOperand Add = DAG.getNode(ISD::ADD, ADDVT, N1.getOperand(1),
1095 DAG.getConstant(Log2_64(SHC->getValue()),
1097 AddToWorkList(Add.Val);
1098 return DAG.getNode(ISD::SRL, VT, N0, Add);
1102 // fold (udiv x, c) -> alternate
1103 if (N1C && N1C->getValue() && !TLI.isIntDivCheap()) {
1104 SDOperand Op = BuildUDIV(N);
1105 if (Op.Val) return Op;
1110 SDOperand DAGCombiner::visitSREM(SDNode *N) {
1111 SDOperand N0 = N->getOperand(0);
1112 SDOperand N1 = N->getOperand(1);
1113 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1114 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1115 MVT::ValueType VT = N->getValueType(0);
1117 // fold (srem c1, c2) -> c1%c2
1118 if (N0C && N1C && !N1C->isNullValue())
1119 return DAG.getNode(ISD::SREM, VT, N0, N1);
1120 // If we know the sign bits of both operands are zero, strength reduce to a
1121 // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15
1122 uint64_t SignBit = 1ULL << (MVT::getSizeInBits(VT)-1);
1123 if (TLI.MaskedValueIsZero(N1, SignBit) &&
1124 TLI.MaskedValueIsZero(N0, SignBit))
1125 return DAG.getNode(ISD::UREM, VT, N0, N1);
1127 // Unconditionally lower X%C -> X-X/C*C. This allows the X/C logic to hack on
1128 // the remainder operation.
1129 if (N1C && !N1C->isNullValue()) {
1130 SDOperand Div = DAG.getNode(ISD::SDIV, VT, N0, N1);
1131 SDOperand Mul = DAG.getNode(ISD::MUL, VT, Div, N1);
1132 SDOperand Sub = DAG.getNode(ISD::SUB, VT, N0, Mul);
1133 AddToWorkList(Div.Val);
1134 AddToWorkList(Mul.Val);
1141 SDOperand DAGCombiner::visitUREM(SDNode *N) {
1142 SDOperand N0 = N->getOperand(0);
1143 SDOperand N1 = N->getOperand(1);
1144 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1145 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1146 MVT::ValueType VT = N->getValueType(0);
1148 // fold (urem c1, c2) -> c1%c2
1149 if (N0C && N1C && !N1C->isNullValue())
1150 return DAG.getNode(ISD::UREM, VT, N0, N1);
1151 // fold (urem x, pow2) -> (and x, pow2-1)
1152 if (N1C && !N1C->isNullValue() && isPowerOf2_64(N1C->getValue()))
1153 return DAG.getNode(ISD::AND, VT, N0, DAG.getConstant(N1C->getValue()-1,VT));
1154 // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1))
1155 if (N1.getOpcode() == ISD::SHL) {
1156 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
1157 if (isPowerOf2_64(SHC->getValue())) {
1158 SDOperand Add = DAG.getNode(ISD::ADD, VT, N1,DAG.getConstant(~0ULL,VT));
1159 AddToWorkList(Add.Val);
1160 return DAG.getNode(ISD::AND, VT, N0, Add);
1165 // Unconditionally lower X%C -> X-X/C*C. This allows the X/C logic to hack on
1166 // the remainder operation.
1167 if (N1C && !N1C->isNullValue()) {
1168 SDOperand Div = DAG.getNode(ISD::UDIV, VT, N0, N1);
1169 SDOperand Mul = DAG.getNode(ISD::MUL, VT, Div, N1);
1170 SDOperand Sub = DAG.getNode(ISD::SUB, VT, N0, Mul);
1171 AddToWorkList(Div.Val);
1172 AddToWorkList(Mul.Val);
1179 SDOperand DAGCombiner::visitMULHS(SDNode *N) {
1180 SDOperand N0 = N->getOperand(0);
1181 SDOperand N1 = N->getOperand(1);
1182 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1184 // fold (mulhs x, 0) -> 0
1185 if (N1C && N1C->isNullValue())
1187 // fold (mulhs x, 1) -> (sra x, size(x)-1)
1188 if (N1C && N1C->getValue() == 1)
1189 return DAG.getNode(ISD::SRA, N0.getValueType(), N0,
1190 DAG.getConstant(MVT::getSizeInBits(N0.getValueType())-1,
1191 TLI.getShiftAmountTy()));
1195 SDOperand DAGCombiner::visitMULHU(SDNode *N) {
1196 SDOperand N0 = N->getOperand(0);
1197 SDOperand N1 = N->getOperand(1);
1198 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1200 // fold (mulhu x, 0) -> 0
1201 if (N1C && N1C->isNullValue())
1203 // fold (mulhu x, 1) -> 0
1204 if (N1C && N1C->getValue() == 1)
1205 return DAG.getConstant(0, N0.getValueType());
1209 /// SimplifyBinOpWithSameOpcodeHands - If this is a binary operator with
1210 /// two operands of the same opcode, try to simplify it.
1211 SDOperand DAGCombiner::SimplifyBinOpWithSameOpcodeHands(SDNode *N) {
1212 SDOperand N0 = N->getOperand(0), N1 = N->getOperand(1);
1213 MVT::ValueType VT = N0.getValueType();
1214 assert(N0.getOpcode() == N1.getOpcode() && "Bad input!");
1216 // For each of OP in AND/OR/XOR:
1217 // fold (OP (zext x), (zext y)) -> (zext (OP x, y))
1218 // fold (OP (sext x), (sext y)) -> (sext (OP x, y))
1219 // fold (OP (aext x), (aext y)) -> (aext (OP x, y))
1220 // fold (OP (trunc x), (trunc y)) -> (trunc (OP x, y))
1221 if ((N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND||
1222 N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::TRUNCATE) &&
1223 N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()) {
1224 SDOperand ORNode = DAG.getNode(N->getOpcode(),
1225 N0.getOperand(0).getValueType(),
1226 N0.getOperand(0), N1.getOperand(0));
1227 AddToWorkList(ORNode.Val);
1228 return DAG.getNode(N0.getOpcode(), VT, ORNode);
1231 // For each of OP in SHL/SRL/SRA/AND...
1232 // fold (and (OP x, z), (OP y, z)) -> (OP (and x, y), z)
1233 // fold (or (OP x, z), (OP y, z)) -> (OP (or x, y), z)
1234 // fold (xor (OP x, z), (OP y, z)) -> (OP (xor x, y), z)
1235 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL ||
1236 N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::AND) &&
1237 N0.getOperand(1) == N1.getOperand(1)) {
1238 SDOperand ORNode = DAG.getNode(N->getOpcode(),
1239 N0.getOperand(0).getValueType(),
1240 N0.getOperand(0), N1.getOperand(0));
1241 AddToWorkList(ORNode.Val);
1242 return DAG.getNode(N0.getOpcode(), VT, ORNode, N0.getOperand(1));
1248 SDOperand DAGCombiner::visitAND(SDNode *N) {
1249 SDOperand N0 = N->getOperand(0);
1250 SDOperand N1 = N->getOperand(1);
1251 SDOperand LL, LR, RL, RR, CC0, CC1;
1252 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1253 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1254 MVT::ValueType VT = N1.getValueType();
1256 // fold (and c1, c2) -> c1&c2
1258 return DAG.getNode(ISD::AND, VT, N0, N1);
1259 // canonicalize constant to RHS
1261 return DAG.getNode(ISD::AND, VT, N1, N0);
1262 // fold (and x, -1) -> x
1263 if (N1C && N1C->isAllOnesValue())
1265 // if (and x, c) is known to be zero, return 0
1266 if (N1C && TLI.MaskedValueIsZero(SDOperand(N, 0), MVT::getIntVTBitMask(VT)))
1267 return DAG.getConstant(0, VT);
1269 SDOperand RAND = ReassociateOps(ISD::AND, N0, N1);
1272 // fold (and (or x, 0xFFFF), 0xFF) -> 0xFF
1273 if (N1C && N0.getOpcode() == ISD::OR)
1274 if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
1275 if ((ORI->getValue() & N1C->getValue()) == N1C->getValue())
1277 // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits.
1278 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
1279 unsigned InMask = MVT::getIntVTBitMask(N0.getOperand(0).getValueType());
1280 if (TLI.MaskedValueIsZero(N0.getOperand(0),
1281 ~N1C->getValue() & InMask)) {
1282 SDOperand Zext = DAG.getNode(ISD::ZERO_EXTEND, N0.getValueType(),
1285 // Replace uses of the AND with uses of the Zero extend node.
1288 // We actually want to replace all uses of the any_extend with the
1289 // zero_extend, to avoid duplicating things. This will later cause this
1290 // AND to be folded.
1291 CombineTo(N0.Val, Zext);
1292 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
1295 // fold (and (setcc x), (setcc y)) -> (setcc (and x, y))
1296 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
1297 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
1298 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
1300 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
1301 MVT::isInteger(LL.getValueType())) {
1302 // fold (X == 0) & (Y == 0) -> (X|Y == 0)
1303 if (cast<ConstantSDNode>(LR)->getValue() == 0 && Op1 == ISD::SETEQ) {
1304 SDOperand ORNode = DAG.getNode(ISD::OR, LR.getValueType(), LL, RL);
1305 AddToWorkList(ORNode.Val);
1306 return DAG.getSetCC(VT, ORNode, LR, Op1);
1308 // fold (X == -1) & (Y == -1) -> (X&Y == -1)
1309 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETEQ) {
1310 SDOperand ANDNode = DAG.getNode(ISD::AND, LR.getValueType(), LL, RL);
1311 AddToWorkList(ANDNode.Val);
1312 return DAG.getSetCC(VT, ANDNode, LR, Op1);
1314 // fold (X > -1) & (Y > -1) -> (X|Y > -1)
1315 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETGT) {
1316 SDOperand ORNode = DAG.getNode(ISD::OR, LR.getValueType(), LL, RL);
1317 AddToWorkList(ORNode.Val);
1318 return DAG.getSetCC(VT, ORNode, LR, Op1);
1321 // canonicalize equivalent to ll == rl
1322 if (LL == RR && LR == RL) {
1323 Op1 = ISD::getSetCCSwappedOperands(Op1);
1326 if (LL == RL && LR == RR) {
1327 bool isInteger = MVT::isInteger(LL.getValueType());
1328 ISD::CondCode Result = ISD::getSetCCAndOperation(Op0, Op1, isInteger);
1329 if (Result != ISD::SETCC_INVALID)
1330 return DAG.getSetCC(N0.getValueType(), LL, LR, Result);
1334 // Simplify: and (op x...), (op y...) -> (op (and x, y))
1335 if (N0.getOpcode() == N1.getOpcode()) {
1336 SDOperand Tmp = SimplifyBinOpWithSameOpcodeHands(N);
1337 if (Tmp.Val) return Tmp;
1340 // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1)
1341 // fold (and (sra)) -> (and (srl)) when possible.
1342 if (!MVT::isVector(VT) &&
1343 SimplifyDemandedBits(SDOperand(N, 0)))
1344 return SDOperand(N, 0);
1345 // fold (zext_inreg (extload x)) -> (zextload x)
1346 if (ISD::isEXTLoad(N0.Val)) {
1347 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
1348 MVT::ValueType EVT = LN0->getLoadedVT();
1349 // If we zero all the possible extended bits, then we can turn this into
1350 // a zextload if we are running before legalize or the operation is legal.
1351 if (TLI.MaskedValueIsZero(N1, ~0ULL << MVT::getSizeInBits(EVT)) &&
1352 (!AfterLegalize || TLI.isLoadXLegal(ISD::ZEXTLOAD, EVT))) {
1353 SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, LN0->getChain(),
1354 LN0->getBasePtr(), LN0->getSrcValue(),
1355 LN0->getSrcValueOffset(), EVT);
1357 CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1));
1358 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
1361 // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use
1362 if (ISD::isSEXTLoad(N0.Val) && N0.hasOneUse()) {
1363 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
1364 MVT::ValueType EVT = LN0->getLoadedVT();
1365 // If we zero all the possible extended bits, then we can turn this into
1366 // a zextload if we are running before legalize or the operation is legal.
1367 if (TLI.MaskedValueIsZero(N1, ~0ULL << MVT::getSizeInBits(EVT)) &&
1368 (!AfterLegalize || TLI.isLoadXLegal(ISD::ZEXTLOAD, EVT))) {
1369 SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, LN0->getChain(),
1370 LN0->getBasePtr(), LN0->getSrcValue(),
1371 LN0->getSrcValueOffset(), EVT);
1373 CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1));
1374 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
1378 // fold (and (load x), 255) -> (zextload x, i8)
1379 // fold (and (extload x, i16), 255) -> (zextload x, i8)
1380 if (N1C && N0.getOpcode() == ISD::LOAD) {
1381 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
1382 if (LN0->getExtensionType() != ISD::SEXTLOAD &&
1384 MVT::ValueType EVT, LoadedVT;
1385 if (N1C->getValue() == 255)
1387 else if (N1C->getValue() == 65535)
1389 else if (N1C->getValue() == ~0U)
1394 LoadedVT = LN0->getLoadedVT();
1395 if (EVT != MVT::Other && LoadedVT > EVT &&
1396 (!AfterLegalize || TLI.isLoadXLegal(ISD::ZEXTLOAD, EVT))) {
1397 MVT::ValueType PtrType = N0.getOperand(1).getValueType();
1398 // For big endian targets, we need to add an offset to the pointer to
1399 // load the correct bytes. For little endian systems, we merely need to
1400 // read fewer bytes from the same pointer.
1402 (MVT::getSizeInBits(LoadedVT) - MVT::getSizeInBits(EVT)) / 8;
1403 SDOperand NewPtr = LN0->getBasePtr();
1404 if (!TLI.isLittleEndian())
1405 NewPtr = DAG.getNode(ISD::ADD, PtrType, NewPtr,
1406 DAG.getConstant(PtrOff, PtrType));
1407 AddToWorkList(NewPtr.Val);
1409 DAG.getExtLoad(ISD::ZEXTLOAD, VT, LN0->getChain(), NewPtr,
1410 LN0->getSrcValue(), LN0->getSrcValueOffset(), EVT);
1412 CombineTo(N0.Val, Load, Load.getValue(1));
1413 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
1421 SDOperand DAGCombiner::visitOR(SDNode *N) {
1422 SDOperand N0 = N->getOperand(0);
1423 SDOperand N1 = N->getOperand(1);
1424 SDOperand LL, LR, RL, RR, CC0, CC1;
1425 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1426 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1427 MVT::ValueType VT = N1.getValueType();
1428 unsigned OpSizeInBits = MVT::getSizeInBits(VT);
1430 // fold (or c1, c2) -> c1|c2
1432 return DAG.getNode(ISD::OR, VT, N0, N1);
1433 // canonicalize constant to RHS
1435 return DAG.getNode(ISD::OR, VT, N1, N0);
1436 // fold (or x, 0) -> x
1437 if (N1C && N1C->isNullValue())
1439 // fold (or x, -1) -> -1
1440 if (N1C && N1C->isAllOnesValue())
1442 // fold (or x, c) -> c iff (x & ~c) == 0
1444 TLI.MaskedValueIsZero(N0,~N1C->getValue() & (~0ULL>>(64-OpSizeInBits))))
1447 SDOperand ROR = ReassociateOps(ISD::OR, N0, N1);
1450 // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2)
1451 if (N1C && N0.getOpcode() == ISD::AND && N0.Val->hasOneUse() &&
1452 isa<ConstantSDNode>(N0.getOperand(1))) {
1453 ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1));
1454 return DAG.getNode(ISD::AND, VT, DAG.getNode(ISD::OR, VT, N0.getOperand(0),
1456 DAG.getConstant(N1C->getValue() | C1->getValue(), VT));
1458 // fold (or (setcc x), (setcc y)) -> (setcc (or x, y))
1459 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
1460 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
1461 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
1463 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
1464 MVT::isInteger(LL.getValueType())) {
1465 // fold (X != 0) | (Y != 0) -> (X|Y != 0)
1466 // fold (X < 0) | (Y < 0) -> (X|Y < 0)
1467 if (cast<ConstantSDNode>(LR)->getValue() == 0 &&
1468 (Op1 == ISD::SETNE || Op1 == ISD::SETLT)) {
1469 SDOperand ORNode = DAG.getNode(ISD::OR, LR.getValueType(), LL, RL);
1470 AddToWorkList(ORNode.Val);
1471 return DAG.getSetCC(VT, ORNode, LR, Op1);
1473 // fold (X != -1) | (Y != -1) -> (X&Y != -1)
1474 // fold (X > -1) | (Y > -1) -> (X&Y > -1)
1475 if (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
1476 (Op1 == ISD::SETNE || Op1 == ISD::SETGT)) {
1477 SDOperand ANDNode = DAG.getNode(ISD::AND, LR.getValueType(), LL, RL);
1478 AddToWorkList(ANDNode.Val);
1479 return DAG.getSetCC(VT, ANDNode, LR, Op1);
1482 // canonicalize equivalent to ll == rl
1483 if (LL == RR && LR == RL) {
1484 Op1 = ISD::getSetCCSwappedOperands(Op1);
1487 if (LL == RL && LR == RR) {
1488 bool isInteger = MVT::isInteger(LL.getValueType());
1489 ISD::CondCode Result = ISD::getSetCCOrOperation(Op0, Op1, isInteger);
1490 if (Result != ISD::SETCC_INVALID)
1491 return DAG.getSetCC(N0.getValueType(), LL, LR, Result);
1495 // Simplify: or (op x...), (op y...) -> (op (or x, y))
1496 if (N0.getOpcode() == N1.getOpcode()) {
1497 SDOperand Tmp = SimplifyBinOpWithSameOpcodeHands(N);
1498 if (Tmp.Val) return Tmp;
1501 // (X & C1) | (Y & C2) -> (X|Y) & C3 if possible.
1502 if (N0.getOpcode() == ISD::AND &&
1503 N1.getOpcode() == ISD::AND &&
1504 N0.getOperand(1).getOpcode() == ISD::Constant &&
1505 N1.getOperand(1).getOpcode() == ISD::Constant &&
1506 // Don't increase # computations.
1507 (N0.Val->hasOneUse() || N1.Val->hasOneUse())) {
1508 // We can only do this xform if we know that bits from X that are set in C2
1509 // but not in C1 are already zero. Likewise for Y.
1510 uint64_t LHSMask = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
1511 uint64_t RHSMask = cast<ConstantSDNode>(N1.getOperand(1))->getValue();
1513 if (TLI.MaskedValueIsZero(N0.getOperand(0), RHSMask&~LHSMask) &&
1514 TLI.MaskedValueIsZero(N1.getOperand(0), LHSMask&~RHSMask)) {
1515 SDOperand X =DAG.getNode(ISD::OR, VT, N0.getOperand(0), N1.getOperand(0));
1516 return DAG.getNode(ISD::AND, VT, X, DAG.getConstant(LHSMask|RHSMask, VT));
1521 // See if this is some rotate idiom.
1522 if (SDNode *Rot = MatchRotate(N0, N1))
1523 return SDOperand(Rot, 0);
1529 /// MatchRotateHalf - Match "(X shl/srl V1) & V2" where V2 may not be present.
1530 static bool MatchRotateHalf(SDOperand Op, SDOperand &Shift, SDOperand &Mask) {
1531 if (Op.getOpcode() == ISD::AND) {
1532 if (isa<ConstantSDNode>(Op.getOperand(1))) {
1533 Mask = Op.getOperand(1);
1534 Op = Op.getOperand(0);
1540 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SHL) {
1548 // MatchRotate - Handle an 'or' of two operands. If this is one of the many
1549 // idioms for rotate, and if the target supports rotation instructions, generate
1551 SDNode *DAGCombiner::MatchRotate(SDOperand LHS, SDOperand RHS) {
1552 // Must be a legal type. Expanded an promoted things won't work with rotates.
1553 MVT::ValueType VT = LHS.getValueType();
1554 if (!TLI.isTypeLegal(VT)) return 0;
1556 // The target must have at least one rotate flavor.
1557 bool HasROTL = TLI.isOperationLegal(ISD::ROTL, VT);
1558 bool HasROTR = TLI.isOperationLegal(ISD::ROTR, VT);
1559 if (!HasROTL && !HasROTR) return 0;
1561 // Match "(X shl/srl V1) & V2" where V2 may not be present.
1562 SDOperand LHSShift; // The shift.
1563 SDOperand LHSMask; // AND value if any.
1564 if (!MatchRotateHalf(LHS, LHSShift, LHSMask))
1565 return 0; // Not part of a rotate.
1567 SDOperand RHSShift; // The shift.
1568 SDOperand RHSMask; // AND value if any.
1569 if (!MatchRotateHalf(RHS, RHSShift, RHSMask))
1570 return 0; // Not part of a rotate.
1572 if (LHSShift.getOperand(0) != RHSShift.getOperand(0))
1573 return 0; // Not shifting the same value.
1575 if (LHSShift.getOpcode() == RHSShift.getOpcode())
1576 return 0; // Shifts must disagree.
1578 // Canonicalize shl to left side in a shl/srl pair.
1579 if (RHSShift.getOpcode() == ISD::SHL) {
1580 std::swap(LHS, RHS);
1581 std::swap(LHSShift, RHSShift);
1582 std::swap(LHSMask , RHSMask );
1585 unsigned OpSizeInBits = MVT::getSizeInBits(VT);
1587 // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1)
1588 // fold (or (shl x, C1), (srl x, C2)) -> (rotr x, C2)
1589 if (LHSShift.getOperand(1).getOpcode() == ISD::Constant &&
1590 RHSShift.getOperand(1).getOpcode() == ISD::Constant) {
1591 uint64_t LShVal = cast<ConstantSDNode>(LHSShift.getOperand(1))->getValue();
1592 uint64_t RShVal = cast<ConstantSDNode>(RHSShift.getOperand(1))->getValue();
1593 if ((LShVal + RShVal) != OpSizeInBits)
1598 Rot = DAG.getNode(ISD::ROTL, VT, LHSShift.getOperand(0),
1599 LHSShift.getOperand(1));
1601 Rot = DAG.getNode(ISD::ROTR, VT, LHSShift.getOperand(0),
1602 RHSShift.getOperand(1));
1604 // If there is an AND of either shifted operand, apply it to the result.
1605 if (LHSMask.Val || RHSMask.Val) {
1606 uint64_t Mask = MVT::getIntVTBitMask(VT);
1609 uint64_t RHSBits = (1ULL << LShVal)-1;
1610 Mask &= cast<ConstantSDNode>(LHSMask)->getValue() | RHSBits;
1613 uint64_t LHSBits = ~((1ULL << (OpSizeInBits-RShVal))-1);
1614 Mask &= cast<ConstantSDNode>(RHSMask)->getValue() | LHSBits;
1617 Rot = DAG.getNode(ISD::AND, VT, Rot, DAG.getConstant(Mask, VT));
1623 // If there is a mask here, and we have a variable shift, we can't be sure
1624 // that we're masking out the right stuff.
1625 if (LHSMask.Val || RHSMask.Val)
1628 // fold (or (shl x, y), (srl x, (sub 32, y))) -> (rotl x, y)
1629 // fold (or (shl x, y), (srl x, (sub 32, y))) -> (rotr x, (sub 32, y))
1630 if (RHSShift.getOperand(1).getOpcode() == ISD::SUB &&
1631 LHSShift.getOperand(1) == RHSShift.getOperand(1).getOperand(1)) {
1632 if (ConstantSDNode *SUBC =
1633 dyn_cast<ConstantSDNode>(RHSShift.getOperand(1).getOperand(0))) {
1634 if (SUBC->getValue() == OpSizeInBits)
1636 return DAG.getNode(ISD::ROTL, VT, LHSShift.getOperand(0),
1637 LHSShift.getOperand(1)).Val;
1639 return DAG.getNode(ISD::ROTR, VT, LHSShift.getOperand(0),
1640 LHSShift.getOperand(1)).Val;
1644 // fold (or (shl x, (sub 32, y)), (srl x, r)) -> (rotr x, y)
1645 // fold (or (shl x, (sub 32, y)), (srl x, r)) -> (rotl x, (sub 32, y))
1646 if (LHSShift.getOperand(1).getOpcode() == ISD::SUB &&
1647 RHSShift.getOperand(1) == LHSShift.getOperand(1).getOperand(1)) {
1648 if (ConstantSDNode *SUBC =
1649 dyn_cast<ConstantSDNode>(LHSShift.getOperand(1).getOperand(0))) {
1650 if (SUBC->getValue() == OpSizeInBits)
1652 return DAG.getNode(ISD::ROTL, VT, LHSShift.getOperand(0),
1653 LHSShift.getOperand(1)).Val;
1655 return DAG.getNode(ISD::ROTR, VT, LHSShift.getOperand(0),
1656 RHSShift.getOperand(1)).Val;
1664 SDOperand DAGCombiner::visitXOR(SDNode *N) {
1665 SDOperand N0 = N->getOperand(0);
1666 SDOperand N1 = N->getOperand(1);
1667 SDOperand LHS, RHS, CC;
1668 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1669 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1670 MVT::ValueType VT = N0.getValueType();
1672 // fold (xor c1, c2) -> c1^c2
1674 return DAG.getNode(ISD::XOR, VT, N0, N1);
1675 // canonicalize constant to RHS
1677 return DAG.getNode(ISD::XOR, VT, N1, N0);
1678 // fold (xor x, 0) -> x
1679 if (N1C && N1C->isNullValue())
1682 SDOperand RXOR = ReassociateOps(ISD::XOR, N0, N1);
1685 // fold !(x cc y) -> (x !cc y)
1686 if (N1C && N1C->getValue() == 1 && isSetCCEquivalent(N0, LHS, RHS, CC)) {
1687 bool isInt = MVT::isInteger(LHS.getValueType());
1688 ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
1690 if (N0.getOpcode() == ISD::SETCC)
1691 return DAG.getSetCC(VT, LHS, RHS, NotCC);
1692 if (N0.getOpcode() == ISD::SELECT_CC)
1693 return DAG.getSelectCC(LHS, RHS, N0.getOperand(2),N0.getOperand(3),NotCC);
1694 assert(0 && "Unhandled SetCC Equivalent!");
1697 // fold !(x or y) -> (!x and !y) iff x or y are setcc
1698 if (N1C && N1C->getValue() == 1 &&
1699 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
1700 SDOperand LHS = N0.getOperand(0), RHS = N0.getOperand(1);
1701 if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) {
1702 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
1703 LHS = DAG.getNode(ISD::XOR, VT, LHS, N1); // RHS = ~LHS
1704 RHS = DAG.getNode(ISD::XOR, VT, RHS, N1); // RHS = ~RHS
1705 AddToWorkList(LHS.Val); AddToWorkList(RHS.Val);
1706 return DAG.getNode(NewOpcode, VT, LHS, RHS);
1709 // fold !(x or y) -> (!x and !y) iff x or y are constants
1710 if (N1C && N1C->isAllOnesValue() &&
1711 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
1712 SDOperand LHS = N0.getOperand(0), RHS = N0.getOperand(1);
1713 if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) {
1714 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
1715 LHS = DAG.getNode(ISD::XOR, VT, LHS, N1); // RHS = ~LHS
1716 RHS = DAG.getNode(ISD::XOR, VT, RHS, N1); // RHS = ~RHS
1717 AddToWorkList(LHS.Val); AddToWorkList(RHS.Val);
1718 return DAG.getNode(NewOpcode, VT, LHS, RHS);
1721 // fold (xor (xor x, c1), c2) -> (xor x, c1^c2)
1722 if (N1C && N0.getOpcode() == ISD::XOR) {
1723 ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0));
1724 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
1726 return DAG.getNode(ISD::XOR, VT, N0.getOperand(1),
1727 DAG.getConstant(N1C->getValue()^N00C->getValue(), VT));
1729 return DAG.getNode(ISD::XOR, VT, N0.getOperand(0),
1730 DAG.getConstant(N1C->getValue()^N01C->getValue(), VT));
1732 // fold (xor x, x) -> 0
1734 if (!MVT::isVector(VT)) {
1735 return DAG.getConstant(0, VT);
1736 } else if (!AfterLegalize || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT)) {
1737 // Produce a vector of zeros.
1738 SDOperand El = DAG.getConstant(0, MVT::getVectorBaseType(VT));
1739 std::vector<SDOperand> Ops(MVT::getVectorNumElements(VT), El);
1740 return DAG.getNode(ISD::BUILD_VECTOR, VT, &Ops[0], Ops.size());
1744 // Simplify: xor (op x...), (op y...) -> (op (xor x, y))
1745 if (N0.getOpcode() == N1.getOpcode()) {
1746 SDOperand Tmp = SimplifyBinOpWithSameOpcodeHands(N);
1747 if (Tmp.Val) return Tmp;
1750 // Simplify the expression using non-local knowledge.
1751 if (!MVT::isVector(VT) &&
1752 SimplifyDemandedBits(SDOperand(N, 0)))
1753 return SDOperand(N, 0);
1758 SDOperand DAGCombiner::visitSHL(SDNode *N) {
1759 SDOperand N0 = N->getOperand(0);
1760 SDOperand N1 = N->getOperand(1);
1761 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1762 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1763 MVT::ValueType VT = N0.getValueType();
1764 unsigned OpSizeInBits = MVT::getSizeInBits(VT);
1766 // fold (shl c1, c2) -> c1<<c2
1768 return DAG.getNode(ISD::SHL, VT, N0, N1);
1769 // fold (shl 0, x) -> 0
1770 if (N0C && N0C->isNullValue())
1772 // fold (shl x, c >= size(x)) -> undef
1773 if (N1C && N1C->getValue() >= OpSizeInBits)
1774 return DAG.getNode(ISD::UNDEF, VT);
1775 // fold (shl x, 0) -> x
1776 if (N1C && N1C->isNullValue())
1778 // if (shl x, c) is known to be zero, return 0
1779 if (TLI.MaskedValueIsZero(SDOperand(N, 0), MVT::getIntVTBitMask(VT)))
1780 return DAG.getConstant(0, VT);
1781 if (SimplifyDemandedBits(SDOperand(N, 0)))
1782 return SDOperand(N, 0);
1783 // fold (shl (shl x, c1), c2) -> 0 or (shl x, c1+c2)
1784 if (N1C && N0.getOpcode() == ISD::SHL &&
1785 N0.getOperand(1).getOpcode() == ISD::Constant) {
1786 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
1787 uint64_t c2 = N1C->getValue();
1788 if (c1 + c2 > OpSizeInBits)
1789 return DAG.getConstant(0, VT);
1790 return DAG.getNode(ISD::SHL, VT, N0.getOperand(0),
1791 DAG.getConstant(c1 + c2, N1.getValueType()));
1793 // fold (shl (srl x, c1), c2) -> (shl (and x, -1 << c1), c2-c1) or
1794 // (srl (and x, -1 << c1), c1-c2)
1795 if (N1C && N0.getOpcode() == ISD::SRL &&
1796 N0.getOperand(1).getOpcode() == ISD::Constant) {
1797 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
1798 uint64_t c2 = N1C->getValue();
1799 SDOperand Mask = DAG.getNode(ISD::AND, VT, N0.getOperand(0),
1800 DAG.getConstant(~0ULL << c1, VT));
1802 return DAG.getNode(ISD::SHL, VT, Mask,
1803 DAG.getConstant(c2-c1, N1.getValueType()));
1805 return DAG.getNode(ISD::SRL, VT, Mask,
1806 DAG.getConstant(c1-c2, N1.getValueType()));
1808 // fold (shl (sra x, c1), c1) -> (and x, -1 << c1)
1809 if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1))
1810 return DAG.getNode(ISD::AND, VT, N0.getOperand(0),
1811 DAG.getConstant(~0ULL << N1C->getValue(), VT));
1812 // fold (shl (add x, c1), c2) -> (add (shl x, c2), c1<<c2)
1813 if (N1C && N0.getOpcode() == ISD::ADD && N0.Val->hasOneUse() &&
1814 isa<ConstantSDNode>(N0.getOperand(1))) {
1815 return DAG.getNode(ISD::ADD, VT,
1816 DAG.getNode(ISD::SHL, VT, N0.getOperand(0), N1),
1817 DAG.getNode(ISD::SHL, VT, N0.getOperand(1), N1));
1822 SDOperand DAGCombiner::visitSRA(SDNode *N) {
1823 SDOperand N0 = N->getOperand(0);
1824 SDOperand N1 = N->getOperand(1);
1825 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1826 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1827 MVT::ValueType VT = N0.getValueType();
1829 // fold (sra c1, c2) -> c1>>c2
1831 return DAG.getNode(ISD::SRA, VT, N0, N1);
1832 // fold (sra 0, x) -> 0
1833 if (N0C && N0C->isNullValue())
1835 // fold (sra -1, x) -> -1
1836 if (N0C && N0C->isAllOnesValue())
1838 // fold (sra x, c >= size(x)) -> undef
1839 if (N1C && N1C->getValue() >= MVT::getSizeInBits(VT))
1840 return DAG.getNode(ISD::UNDEF, VT);
1841 // fold (sra x, 0) -> x
1842 if (N1C && N1C->isNullValue())
1844 // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports
1846 if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) {
1847 unsigned LowBits = MVT::getSizeInBits(VT) - (unsigned)N1C->getValue();
1850 default: EVT = MVT::Other; break;
1851 case 1: EVT = MVT::i1; break;
1852 case 8: EVT = MVT::i8; break;
1853 case 16: EVT = MVT::i16; break;
1854 case 32: EVT = MVT::i32; break;
1856 if (EVT > MVT::Other && TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, EVT))
1857 return DAG.getNode(ISD::SIGN_EXTEND_INREG, VT, N0.getOperand(0),
1858 DAG.getValueType(EVT));
1861 // fold (sra (sra x, c1), c2) -> (sra x, c1+c2)
1862 if (N1C && N0.getOpcode() == ISD::SRA) {
1863 if (ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1864 unsigned Sum = N1C->getValue() + C1->getValue();
1865 if (Sum >= MVT::getSizeInBits(VT)) Sum = MVT::getSizeInBits(VT)-1;
1866 return DAG.getNode(ISD::SRA, VT, N0.getOperand(0),
1867 DAG.getConstant(Sum, N1C->getValueType(0)));
1871 // Simplify, based on bits shifted out of the LHS.
1872 if (N1C && SimplifyDemandedBits(SDOperand(N, 0)))
1873 return SDOperand(N, 0);
1876 // If the sign bit is known to be zero, switch this to a SRL.
1877 if (TLI.MaskedValueIsZero(N0, MVT::getIntVTSignBit(VT)))
1878 return DAG.getNode(ISD::SRL, VT, N0, N1);
1882 SDOperand DAGCombiner::visitSRL(SDNode *N) {
1883 SDOperand N0 = N->getOperand(0);
1884 SDOperand N1 = N->getOperand(1);
1885 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1886 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1887 MVT::ValueType VT = N0.getValueType();
1888 unsigned OpSizeInBits = MVT::getSizeInBits(VT);
1890 // fold (srl c1, c2) -> c1 >>u c2
1892 return DAG.getNode(ISD::SRL, VT, N0, N1);
1893 // fold (srl 0, x) -> 0
1894 if (N0C && N0C->isNullValue())
1896 // fold (srl x, c >= size(x)) -> undef
1897 if (N1C && N1C->getValue() >= OpSizeInBits)
1898 return DAG.getNode(ISD::UNDEF, VT);
1899 // fold (srl x, 0) -> x
1900 if (N1C && N1C->isNullValue())
1902 // if (srl x, c) is known to be zero, return 0
1903 if (N1C && TLI.MaskedValueIsZero(SDOperand(N, 0), ~0ULL >> (64-OpSizeInBits)))
1904 return DAG.getConstant(0, VT);
1905 // fold (srl (srl x, c1), c2) -> 0 or (srl x, c1+c2)
1906 if (N1C && N0.getOpcode() == ISD::SRL &&
1907 N0.getOperand(1).getOpcode() == ISD::Constant) {
1908 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
1909 uint64_t c2 = N1C->getValue();
1910 if (c1 + c2 > OpSizeInBits)
1911 return DAG.getConstant(0, VT);
1912 return DAG.getNode(ISD::SRL, VT, N0.getOperand(0),
1913 DAG.getConstant(c1 + c2, N1.getValueType()));
1916 // fold (srl (anyextend x), c) -> (anyextend (srl x, c))
1917 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
1918 // Shifting in all undef bits?
1919 MVT::ValueType SmallVT = N0.getOperand(0).getValueType();
1920 if (N1C->getValue() >= MVT::getSizeInBits(SmallVT))
1921 return DAG.getNode(ISD::UNDEF, VT);
1923 SDOperand SmallShift = DAG.getNode(ISD::SRL, SmallVT, N0.getOperand(0), N1);
1924 AddToWorkList(SmallShift.Val);
1925 return DAG.getNode(ISD::ANY_EXTEND, VT, SmallShift);
1928 // fold (srl (sra X, Y), 31) -> (srl X, 31). This srl only looks at the sign
1929 // bit, which is unmodified by sra.
1930 if (N1C && N1C->getValue()+1 == MVT::getSizeInBits(VT)) {
1931 if (N0.getOpcode() == ISD::SRA)
1932 return DAG.getNode(ISD::SRL, VT, N0.getOperand(0), N1);
1935 // fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit).
1936 if (N1C && N0.getOpcode() == ISD::CTLZ &&
1937 N1C->getValue() == Log2_32(MVT::getSizeInBits(VT))) {
1938 uint64_t KnownZero, KnownOne, Mask = MVT::getIntVTBitMask(VT);
1939 TLI.ComputeMaskedBits(N0.getOperand(0), Mask, KnownZero, KnownOne);
1941 // If any of the input bits are KnownOne, then the input couldn't be all
1942 // zeros, thus the result of the srl will always be zero.
1943 if (KnownOne) return DAG.getConstant(0, VT);
1945 // If all of the bits input the to ctlz node are known to be zero, then
1946 // the result of the ctlz is "32" and the result of the shift is one.
1947 uint64_t UnknownBits = ~KnownZero & Mask;
1948 if (UnknownBits == 0) return DAG.getConstant(1, VT);
1950 // Otherwise, check to see if there is exactly one bit input to the ctlz.
1951 if ((UnknownBits & (UnknownBits-1)) == 0) {
1952 // Okay, we know that only that the single bit specified by UnknownBits
1953 // could be set on input to the CTLZ node. If this bit is set, the SRL
1954 // will return 0, if it is clear, it returns 1. Change the CTLZ/SRL pair
1955 // to an SRL,XOR pair, which is likely to simplify more.
1956 unsigned ShAmt = CountTrailingZeros_64(UnknownBits);
1957 SDOperand Op = N0.getOperand(0);
1959 Op = DAG.getNode(ISD::SRL, VT, Op,
1960 DAG.getConstant(ShAmt, TLI.getShiftAmountTy()));
1961 AddToWorkList(Op.Val);
1963 return DAG.getNode(ISD::XOR, VT, Op, DAG.getConstant(1, VT));
1970 SDOperand DAGCombiner::visitCTLZ(SDNode *N) {
1971 SDOperand N0 = N->getOperand(0);
1972 MVT::ValueType VT = N->getValueType(0);
1974 // fold (ctlz c1) -> c2
1975 if (isa<ConstantSDNode>(N0))
1976 return DAG.getNode(ISD::CTLZ, VT, N0);
1980 SDOperand DAGCombiner::visitCTTZ(SDNode *N) {
1981 SDOperand N0 = N->getOperand(0);
1982 MVT::ValueType VT = N->getValueType(0);
1984 // fold (cttz c1) -> c2
1985 if (isa<ConstantSDNode>(N0))
1986 return DAG.getNode(ISD::CTTZ, VT, N0);
1990 SDOperand DAGCombiner::visitCTPOP(SDNode *N) {
1991 SDOperand N0 = N->getOperand(0);
1992 MVT::ValueType VT = N->getValueType(0);
1994 // fold (ctpop c1) -> c2
1995 if (isa<ConstantSDNode>(N0))
1996 return DAG.getNode(ISD::CTPOP, VT, N0);
2000 SDOperand DAGCombiner::visitSELECT(SDNode *N) {
2001 SDOperand N0 = N->getOperand(0);
2002 SDOperand N1 = N->getOperand(1);
2003 SDOperand N2 = N->getOperand(2);
2004 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2005 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2006 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
2007 MVT::ValueType VT = N->getValueType(0);
2009 // fold select C, X, X -> X
2012 // fold select true, X, Y -> X
2013 if (N0C && !N0C->isNullValue())
2015 // fold select false, X, Y -> Y
2016 if (N0C && N0C->isNullValue())
2018 // fold select C, 1, X -> C | X
2019 if (MVT::i1 == VT && N1C && N1C->getValue() == 1)
2020 return DAG.getNode(ISD::OR, VT, N0, N2);
2021 // fold select C, 0, X -> ~C & X
2022 // FIXME: this should check for C type == X type, not i1?
2023 if (MVT::i1 == VT && N1C && N1C->isNullValue()) {
2024 SDOperand XORNode = DAG.getNode(ISD::XOR, VT, N0, DAG.getConstant(1, VT));
2025 AddToWorkList(XORNode.Val);
2026 return DAG.getNode(ISD::AND, VT, XORNode, N2);
2028 // fold select C, X, 1 -> ~C | X
2029 if (MVT::i1 == VT && N2C && N2C->getValue() == 1) {
2030 SDOperand XORNode = DAG.getNode(ISD::XOR, VT, N0, DAG.getConstant(1, VT));
2031 AddToWorkList(XORNode.Val);
2032 return DAG.getNode(ISD::OR, VT, XORNode, N1);
2034 // fold select C, X, 0 -> C & X
2035 // FIXME: this should check for C type == X type, not i1?
2036 if (MVT::i1 == VT && N2C && N2C->isNullValue())
2037 return DAG.getNode(ISD::AND, VT, N0, N1);
2038 // fold X ? X : Y --> X ? 1 : Y --> X | Y
2039 if (MVT::i1 == VT && N0 == N1)
2040 return DAG.getNode(ISD::OR, VT, N0, N2);
2041 // fold X ? Y : X --> X ? Y : 0 --> X & Y
2042 if (MVT::i1 == VT && N0 == N2)
2043 return DAG.getNode(ISD::AND, VT, N0, N1);
2045 // If we can fold this based on the true/false value, do so.
2046 if (SimplifySelectOps(N, N1, N2))
2047 return SDOperand(N, 0); // Don't revisit N.
2049 // fold selects based on a setcc into other things, such as min/max/abs
2050 if (N0.getOpcode() == ISD::SETCC)
2052 // Check against MVT::Other for SELECT_CC, which is a workaround for targets
2053 // having to say they don't support SELECT_CC on every type the DAG knows
2054 // about, since there is no way to mark an opcode illegal at all value types
2055 if (TLI.isOperationLegal(ISD::SELECT_CC, MVT::Other))
2056 return DAG.getNode(ISD::SELECT_CC, VT, N0.getOperand(0), N0.getOperand(1),
2057 N1, N2, N0.getOperand(2));
2059 return SimplifySelect(N0, N1, N2);
2063 SDOperand DAGCombiner::visitSELECT_CC(SDNode *N) {
2064 SDOperand N0 = N->getOperand(0);
2065 SDOperand N1 = N->getOperand(1);
2066 SDOperand N2 = N->getOperand(2);
2067 SDOperand N3 = N->getOperand(3);
2068 SDOperand N4 = N->getOperand(4);
2069 ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get();
2071 // fold select_cc lhs, rhs, x, x, cc -> x
2075 // Determine if the condition we're dealing with is constant
2076 SDOperand SCC = SimplifySetCC(TLI.getSetCCResultTy(), N0, N1, CC, false);
2077 if (SCC.Val) AddToWorkList(SCC.Val);
2079 if (ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.Val)) {
2080 if (SCCC->getValue())
2081 return N2; // cond always true -> true val
2083 return N3; // cond always false -> false val
2086 // Fold to a simpler select_cc
2087 if (SCC.Val && SCC.getOpcode() == ISD::SETCC)
2088 return DAG.getNode(ISD::SELECT_CC, N2.getValueType(),
2089 SCC.getOperand(0), SCC.getOperand(1), N2, N3,
2092 // If we can fold this based on the true/false value, do so.
2093 if (SimplifySelectOps(N, N2, N3))
2094 return SDOperand(N, 0); // Don't revisit N.
2096 // fold select_cc into other things, such as min/max/abs
2097 return SimplifySelectCC(N0, N1, N2, N3, CC);
2100 SDOperand DAGCombiner::visitSETCC(SDNode *N) {
2101 return SimplifySetCC(N->getValueType(0), N->getOperand(0), N->getOperand(1),
2102 cast<CondCodeSDNode>(N->getOperand(2))->get());
2105 SDOperand DAGCombiner::visitSIGN_EXTEND(SDNode *N) {
2106 SDOperand N0 = N->getOperand(0);
2107 MVT::ValueType VT = N->getValueType(0);
2109 // fold (sext c1) -> c1
2110 if (isa<ConstantSDNode>(N0))
2111 return DAG.getNode(ISD::SIGN_EXTEND, VT, N0);
2113 // fold (sext (sext x)) -> (sext x)
2114 // fold (sext (aext x)) -> (sext x)
2115 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
2116 return DAG.getNode(ISD::SIGN_EXTEND, VT, N0.getOperand(0));
2118 // fold (sext (truncate x)) -> (sextinreg x).
2119 if (N0.getOpcode() == ISD::TRUNCATE &&
2120 (!AfterLegalize || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG,
2121 N0.getValueType()))) {
2122 SDOperand Op = N0.getOperand(0);
2123 if (Op.getValueType() < VT) {
2124 Op = DAG.getNode(ISD::ANY_EXTEND, VT, Op);
2125 } else if (Op.getValueType() > VT) {
2126 Op = DAG.getNode(ISD::TRUNCATE, VT, Op);
2128 return DAG.getNode(ISD::SIGN_EXTEND_INREG, VT, Op,
2129 DAG.getValueType(N0.getValueType()));
2132 // fold (sext (load x)) -> (sext (truncate (sextload x)))
2133 if (ISD::isNON_EXTLoad(N0.Val) && N0.hasOneUse() &&
2134 (!AfterLegalize||TLI.isLoadXLegal(ISD::SEXTLOAD, N0.getValueType()))){
2135 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2136 SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, LN0->getChain(),
2137 LN0->getBasePtr(), LN0->getSrcValue(),
2138 LN0->getSrcValueOffset(),
2140 CombineTo(N, ExtLoad);
2141 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
2142 ExtLoad.getValue(1));
2143 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
2146 // fold (sext (sextload x)) -> (sext (truncate (sextload x)))
2147 // fold (sext ( extload x)) -> (sext (truncate (sextload x)))
2148 if ((ISD::isSEXTLoad(N0.Val) || ISD::isEXTLoad(N0.Val)) && N0.hasOneUse()) {
2149 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2150 MVT::ValueType EVT = LN0->getLoadedVT();
2151 SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, LN0->getChain(),
2152 LN0->getBasePtr(), LN0->getSrcValue(),
2153 LN0->getSrcValueOffset(), EVT);
2154 CombineTo(N, ExtLoad);
2155 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
2156 ExtLoad.getValue(1));
2157 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
2163 SDOperand DAGCombiner::visitZERO_EXTEND(SDNode *N) {
2164 SDOperand N0 = N->getOperand(0);
2165 MVT::ValueType VT = N->getValueType(0);
2167 // fold (zext c1) -> c1
2168 if (isa<ConstantSDNode>(N0))
2169 return DAG.getNode(ISD::ZERO_EXTEND, VT, N0);
2170 // fold (zext (zext x)) -> (zext x)
2171 // fold (zext (aext x)) -> (zext x)
2172 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
2173 return DAG.getNode(ISD::ZERO_EXTEND, VT, N0.getOperand(0));
2175 // fold (zext (truncate x)) -> (and x, mask)
2176 if (N0.getOpcode() == ISD::TRUNCATE &&
2177 (!AfterLegalize || TLI.isOperationLegal(ISD::AND, VT))) {
2178 SDOperand Op = N0.getOperand(0);
2179 if (Op.getValueType() < VT) {
2180 Op = DAG.getNode(ISD::ANY_EXTEND, VT, Op);
2181 } else if (Op.getValueType() > VT) {
2182 Op = DAG.getNode(ISD::TRUNCATE, VT, Op);
2184 return DAG.getZeroExtendInReg(Op, N0.getValueType());
2187 // fold (zext (and (trunc x), cst)) -> (and x, cst).
2188 if (N0.getOpcode() == ISD::AND &&
2189 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
2190 N0.getOperand(1).getOpcode() == ISD::Constant) {
2191 SDOperand X = N0.getOperand(0).getOperand(0);
2192 if (X.getValueType() < VT) {
2193 X = DAG.getNode(ISD::ANY_EXTEND, VT, X);
2194 } else if (X.getValueType() > VT) {
2195 X = DAG.getNode(ISD::TRUNCATE, VT, X);
2197 uint64_t Mask = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
2198 return DAG.getNode(ISD::AND, VT, X, DAG.getConstant(Mask, VT));
2201 // fold (zext (load x)) -> (zext (truncate (zextload x)))
2202 if (ISD::isNON_EXTLoad(N0.Val) && N0.hasOneUse() &&
2203 (!AfterLegalize||TLI.isLoadXLegal(ISD::ZEXTLOAD, N0.getValueType()))) {
2204 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2205 SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, LN0->getChain(),
2206 LN0->getBasePtr(), LN0->getSrcValue(),
2207 LN0->getSrcValueOffset(),
2209 CombineTo(N, ExtLoad);
2210 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
2211 ExtLoad.getValue(1));
2212 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
2215 // fold (zext (zextload x)) -> (zext (truncate (zextload x)))
2216 // fold (zext ( extload x)) -> (zext (truncate (zextload x)))
2217 if ((ISD::isZEXTLoad(N0.Val) || ISD::isEXTLoad(N0.Val)) && N0.hasOneUse()) {
2218 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2219 MVT::ValueType EVT = LN0->getLoadedVT();
2220 SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, LN0->getChain(),
2221 LN0->getBasePtr(), LN0->getSrcValue(),
2222 LN0->getSrcValueOffset(), EVT);
2223 CombineTo(N, ExtLoad);
2224 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
2225 ExtLoad.getValue(1));
2226 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
2231 SDOperand DAGCombiner::visitANY_EXTEND(SDNode *N) {
2232 SDOperand N0 = N->getOperand(0);
2233 MVT::ValueType VT = N->getValueType(0);
2235 // fold (aext c1) -> c1
2236 if (isa<ConstantSDNode>(N0))
2237 return DAG.getNode(ISD::ANY_EXTEND, VT, N0);
2238 // fold (aext (aext x)) -> (aext x)
2239 // fold (aext (zext x)) -> (zext x)
2240 // fold (aext (sext x)) -> (sext x)
2241 if (N0.getOpcode() == ISD::ANY_EXTEND ||
2242 N0.getOpcode() == ISD::ZERO_EXTEND ||
2243 N0.getOpcode() == ISD::SIGN_EXTEND)
2244 return DAG.getNode(N0.getOpcode(), VT, N0.getOperand(0));
2246 // fold (aext (truncate x))
2247 if (N0.getOpcode() == ISD::TRUNCATE) {
2248 SDOperand TruncOp = N0.getOperand(0);
2249 if (TruncOp.getValueType() == VT)
2250 return TruncOp; // x iff x size == zext size.
2251 if (TruncOp.getValueType() > VT)
2252 return DAG.getNode(ISD::TRUNCATE, VT, TruncOp);
2253 return DAG.getNode(ISD::ANY_EXTEND, VT, TruncOp);
2256 // fold (aext (and (trunc x), cst)) -> (and x, cst).
2257 if (N0.getOpcode() == ISD::AND &&
2258 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
2259 N0.getOperand(1).getOpcode() == ISD::Constant) {
2260 SDOperand X = N0.getOperand(0).getOperand(0);
2261 if (X.getValueType() < VT) {
2262 X = DAG.getNode(ISD::ANY_EXTEND, VT, X);
2263 } else if (X.getValueType() > VT) {
2264 X = DAG.getNode(ISD::TRUNCATE, VT, X);
2266 uint64_t Mask = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
2267 return DAG.getNode(ISD::AND, VT, X, DAG.getConstant(Mask, VT));
2270 // fold (aext (load x)) -> (aext (truncate (extload x)))
2271 if (ISD::isNON_EXTLoad(N0.Val) && N0.hasOneUse() &&
2272 (!AfterLegalize||TLI.isLoadXLegal(ISD::EXTLOAD, N0.getValueType()))) {
2273 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2274 SDOperand ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, VT, LN0->getChain(),
2275 LN0->getBasePtr(), LN0->getSrcValue(),
2276 LN0->getSrcValueOffset(),
2278 CombineTo(N, ExtLoad);
2279 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
2280 ExtLoad.getValue(1));
2281 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
2284 // fold (aext (zextload x)) -> (aext (truncate (zextload x)))
2285 // fold (aext (sextload x)) -> (aext (truncate (sextload x)))
2286 // fold (aext ( extload x)) -> (aext (truncate (extload x)))
2287 if (N0.getOpcode() == ISD::LOAD && !ISD::isNON_EXTLoad(N0.Val) &&
2289 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2290 MVT::ValueType EVT = LN0->getLoadedVT();
2291 SDOperand ExtLoad = DAG.getExtLoad(LN0->getExtensionType(), VT,
2292 LN0->getChain(), LN0->getBasePtr(),
2294 LN0->getSrcValueOffset(), EVT);
2295 CombineTo(N, ExtLoad);
2296 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
2297 ExtLoad.getValue(1));
2298 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
2304 SDOperand DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
2305 SDOperand N0 = N->getOperand(0);
2306 SDOperand N1 = N->getOperand(1);
2307 MVT::ValueType VT = N->getValueType(0);
2308 MVT::ValueType EVT = cast<VTSDNode>(N1)->getVT();
2309 unsigned EVTBits = MVT::getSizeInBits(EVT);
2311 // fold (sext_in_reg c1) -> c1
2312 if (isa<ConstantSDNode>(N0) || N0.getOpcode() == ISD::UNDEF)
2313 return DAG.getNode(ISD::SIGN_EXTEND_INREG, VT, N0, N1);
2315 // If the input is already sign extended, just drop the extension.
2316 if (TLI.ComputeNumSignBits(N0) >= MVT::getSizeInBits(VT)-EVTBits+1)
2319 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2
2320 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
2321 EVT < cast<VTSDNode>(N0.getOperand(1))->getVT()) {
2322 return DAG.getNode(ISD::SIGN_EXTEND_INREG, VT, N0.getOperand(0), N1);
2325 // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is zero
2326 if (TLI.MaskedValueIsZero(N0, 1ULL << (EVTBits-1)))
2327 return DAG.getZeroExtendInReg(N0, EVT);
2329 // fold (sext_in_reg (srl X, 24), i8) -> sra X, 24
2330 // fold (sext_in_reg (srl X, 23), i8) -> sra X, 23 iff possible.
2331 // We already fold "(sext_in_reg (srl X, 25), i8) -> srl X, 25" above.
2332 if (N0.getOpcode() == ISD::SRL) {
2333 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
2334 if (ShAmt->getValue()+EVTBits <= MVT::getSizeInBits(VT)) {
2335 // We can turn this into an SRA iff the input to the SRL is already sign
2337 unsigned InSignBits = TLI.ComputeNumSignBits(N0.getOperand(0));
2338 if (MVT::getSizeInBits(VT)-(ShAmt->getValue()+EVTBits) < InSignBits)
2339 return DAG.getNode(ISD::SRA, VT, N0.getOperand(0), N0.getOperand(1));
2343 // fold (sext_inreg (extload x)) -> (sextload x)
2344 if (ISD::isEXTLoad(N0.Val) &&
2345 EVT == cast<LoadSDNode>(N0)->getLoadedVT() &&
2346 (!AfterLegalize || TLI.isLoadXLegal(ISD::SEXTLOAD, EVT))) {
2347 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2348 SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, LN0->getChain(),
2349 LN0->getBasePtr(), LN0->getSrcValue(),
2350 LN0->getSrcValueOffset(), EVT);
2351 CombineTo(N, ExtLoad);
2352 CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1));
2353 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
2355 // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use
2356 if (ISD::isZEXTLoad(N0.Val) && N0.hasOneUse() &&
2357 EVT == cast<LoadSDNode>(N0)->getLoadedVT() &&
2358 (!AfterLegalize || TLI.isLoadXLegal(ISD::SEXTLOAD, EVT))) {
2359 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2360 SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, LN0->getChain(),
2361 LN0->getBasePtr(), LN0->getSrcValue(),
2362 LN0->getSrcValueOffset(), EVT);
2363 CombineTo(N, ExtLoad);
2364 CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1));
2365 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
2370 SDOperand DAGCombiner::visitTRUNCATE(SDNode *N) {
2371 SDOperand N0 = N->getOperand(0);
2372 MVT::ValueType VT = N->getValueType(0);
2375 if (N0.getValueType() == N->getValueType(0))
2377 // fold (truncate c1) -> c1
2378 if (isa<ConstantSDNode>(N0))
2379 return DAG.getNode(ISD::TRUNCATE, VT, N0);
2380 // fold (truncate (truncate x)) -> (truncate x)
2381 if (N0.getOpcode() == ISD::TRUNCATE)
2382 return DAG.getNode(ISD::TRUNCATE, VT, N0.getOperand(0));
2383 // fold (truncate (ext x)) -> (ext x) or (truncate x) or x
2384 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::SIGN_EXTEND||
2385 N0.getOpcode() == ISD::ANY_EXTEND) {
2386 if (N0.getValueType() < VT)
2387 // if the source is smaller than the dest, we still need an extend
2388 return DAG.getNode(N0.getOpcode(), VT, N0.getOperand(0));
2389 else if (N0.getValueType() > VT)
2390 // if the source is larger than the dest, than we just need the truncate
2391 return DAG.getNode(ISD::TRUNCATE, VT, N0.getOperand(0));
2393 // if the source and dest are the same type, we can drop both the extend
2395 return N0.getOperand(0);
2397 // fold (truncate (load x)) -> (smaller load x)
2398 if (ISD::isNON_EXTLoad(N0.Val) && N0.hasOneUse()) {
2399 assert(MVT::getSizeInBits(N0.getValueType()) > MVT::getSizeInBits(VT) &&
2400 "Cannot truncate to larger type!");
2401 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2402 MVT::ValueType PtrType = N0.getOperand(1).getValueType();
2403 // For big endian targets, we need to add an offset to the pointer to load
2404 // the correct bytes. For little endian systems, we merely need to read
2405 // fewer bytes from the same pointer.
2407 (MVT::getSizeInBits(N0.getValueType()) - MVT::getSizeInBits(VT)) / 8;
2408 SDOperand NewPtr = TLI.isLittleEndian() ? LN0->getBasePtr() :
2409 DAG.getNode(ISD::ADD, PtrType, LN0->getBasePtr(),
2410 DAG.getConstant(PtrOff, PtrType));
2411 AddToWorkList(NewPtr.Val);
2412 SDOperand Load = DAG.getLoad(VT, LN0->getChain(), NewPtr,
2413 LN0->getSrcValue(), LN0->getSrcValueOffset());
2415 CombineTo(N0.Val, Load, Load.getValue(1));
2416 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
2421 SDOperand DAGCombiner::visitBIT_CONVERT(SDNode *N) {
2422 SDOperand N0 = N->getOperand(0);
2423 MVT::ValueType VT = N->getValueType(0);
2425 // If the input is a constant, let getNode() fold it.
2426 if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) {
2427 SDOperand Res = DAG.getNode(ISD::BIT_CONVERT, VT, N0);
2428 if (Res.Val != N) return Res;
2431 if (N0.getOpcode() == ISD::BIT_CONVERT) // conv(conv(x,t1),t2) -> conv(x,t2)
2432 return DAG.getNode(ISD::BIT_CONVERT, VT, N0.getOperand(0));
2434 // fold (conv (load x)) -> (load (conv*)x)
2435 // FIXME: These xforms need to know that the resultant load doesn't need a
2436 // higher alignment than the original!
2437 if (0 && ISD::isNON_EXTLoad(N0.Val) && N0.hasOneUse()) {
2438 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2439 SDOperand Load = DAG.getLoad(VT, LN0->getChain(), LN0->getBasePtr(),
2440 LN0->getSrcValue(), LN0->getSrcValueOffset());
2442 CombineTo(N0.Val, DAG.getNode(ISD::BIT_CONVERT, N0.getValueType(), Load),
2450 SDOperand DAGCombiner::visitVBIT_CONVERT(SDNode *N) {
2451 SDOperand N0 = N->getOperand(0);
2452 MVT::ValueType VT = N->getValueType(0);
2454 // If the input is a VBUILD_VECTOR with all constant elements, fold this now.
2455 // First check to see if this is all constant.
2456 if (N0.getOpcode() == ISD::VBUILD_VECTOR && N0.Val->hasOneUse() &&
2457 VT == MVT::Vector) {
2458 bool isSimple = true;
2459 for (unsigned i = 0, e = N0.getNumOperands()-2; i != e; ++i)
2460 if (N0.getOperand(i).getOpcode() != ISD::UNDEF &&
2461 N0.getOperand(i).getOpcode() != ISD::Constant &&
2462 N0.getOperand(i).getOpcode() != ISD::ConstantFP) {
2467 MVT::ValueType DestEltVT = cast<VTSDNode>(N->getOperand(2))->getVT();
2468 if (isSimple && !MVT::isVector(DestEltVT)) {
2469 return ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(N0.Val, DestEltVT);
2476 /// ConstantFoldVBIT_CONVERTofVBUILD_VECTOR - We know that BV is a vbuild_vector
2477 /// node with Constant, ConstantFP or Undef operands. DstEltVT indicates the
2478 /// destination element value type.
2479 SDOperand DAGCombiner::
2480 ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(SDNode *BV, MVT::ValueType DstEltVT) {
2481 MVT::ValueType SrcEltVT = BV->getOperand(0).getValueType();
2483 // If this is already the right type, we're done.
2484 if (SrcEltVT == DstEltVT) return SDOperand(BV, 0);
2486 unsigned SrcBitSize = MVT::getSizeInBits(SrcEltVT);
2487 unsigned DstBitSize = MVT::getSizeInBits(DstEltVT);
2489 // If this is a conversion of N elements of one type to N elements of another
2490 // type, convert each element. This handles FP<->INT cases.
2491 if (SrcBitSize == DstBitSize) {
2492 SmallVector<SDOperand, 8> Ops;
2493 for (unsigned i = 0, e = BV->getNumOperands()-2; i != e; ++i) {
2494 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, DstEltVT, BV->getOperand(i)));
2495 AddToWorkList(Ops.back().Val);
2497 Ops.push_back(*(BV->op_end()-2)); // Add num elements.
2498 Ops.push_back(DAG.getValueType(DstEltVT));
2499 return DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, &Ops[0], Ops.size());
2502 // Otherwise, we're growing or shrinking the elements. To avoid having to
2503 // handle annoying details of growing/shrinking FP values, we convert them to
2505 if (MVT::isFloatingPoint(SrcEltVT)) {
2506 // Convert the input float vector to a int vector where the elements are the
2508 assert((SrcEltVT == MVT::f32 || SrcEltVT == MVT::f64) && "Unknown FP VT!");
2509 MVT::ValueType IntVT = SrcEltVT == MVT::f32 ? MVT::i32 : MVT::i64;
2510 BV = ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(BV, IntVT).Val;
2514 // Now we know the input is an integer vector. If the output is a FP type,
2515 // convert to integer first, then to FP of the right size.
2516 if (MVT::isFloatingPoint(DstEltVT)) {
2517 assert((DstEltVT == MVT::f32 || DstEltVT == MVT::f64) && "Unknown FP VT!");
2518 MVT::ValueType TmpVT = DstEltVT == MVT::f32 ? MVT::i32 : MVT::i64;
2519 SDNode *Tmp = ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(BV, TmpVT).Val;
2521 // Next, convert to FP elements of the same size.
2522 return ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(Tmp, DstEltVT);
2525 // Okay, we know the src/dst types are both integers of differing types.
2526 // Handling growing first.
2527 assert(MVT::isInteger(SrcEltVT) && MVT::isInteger(DstEltVT));
2528 if (SrcBitSize < DstBitSize) {
2529 unsigned NumInputsPerOutput = DstBitSize/SrcBitSize;
2531 SmallVector<SDOperand, 8> Ops;
2532 for (unsigned i = 0, e = BV->getNumOperands()-2; i != e;
2533 i += NumInputsPerOutput) {
2534 bool isLE = TLI.isLittleEndian();
2535 uint64_t NewBits = 0;
2536 bool EltIsUndef = true;
2537 for (unsigned j = 0; j != NumInputsPerOutput; ++j) {
2538 // Shift the previously computed bits over.
2539 NewBits <<= SrcBitSize;
2540 SDOperand Op = BV->getOperand(i+ (isLE ? (NumInputsPerOutput-j-1) : j));
2541 if (Op.getOpcode() == ISD::UNDEF) continue;
2544 NewBits |= cast<ConstantSDNode>(Op)->getValue();
2548 Ops.push_back(DAG.getNode(ISD::UNDEF, DstEltVT));
2550 Ops.push_back(DAG.getConstant(NewBits, DstEltVT));
2553 Ops.push_back(DAG.getConstant(Ops.size(), MVT::i32)); // Add num elements.
2554 Ops.push_back(DAG.getValueType(DstEltVT)); // Add element size.
2555 return DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, &Ops[0], Ops.size());
2558 // Finally, this must be the case where we are shrinking elements: each input
2559 // turns into multiple outputs.
2560 unsigned NumOutputsPerInput = SrcBitSize/DstBitSize;
2561 SmallVector<SDOperand, 8> Ops;
2562 for (unsigned i = 0, e = BV->getNumOperands()-2; i != e; ++i) {
2563 if (BV->getOperand(i).getOpcode() == ISD::UNDEF) {
2564 for (unsigned j = 0; j != NumOutputsPerInput; ++j)
2565 Ops.push_back(DAG.getNode(ISD::UNDEF, DstEltVT));
2568 uint64_t OpVal = cast<ConstantSDNode>(BV->getOperand(i))->getValue();
2570 for (unsigned j = 0; j != NumOutputsPerInput; ++j) {
2571 unsigned ThisVal = OpVal & ((1ULL << DstBitSize)-1);
2572 OpVal >>= DstBitSize;
2573 Ops.push_back(DAG.getConstant(ThisVal, DstEltVT));
2576 // For big endian targets, swap the order of the pieces of each element.
2577 if (!TLI.isLittleEndian())
2578 std::reverse(Ops.end()-NumOutputsPerInput, Ops.end());
2580 Ops.push_back(DAG.getConstant(Ops.size(), MVT::i32)); // Add num elements.
2581 Ops.push_back(DAG.getValueType(DstEltVT)); // Add element size.
2582 return DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, &Ops[0], Ops.size());
2587 SDOperand DAGCombiner::visitFADD(SDNode *N) {
2588 SDOperand N0 = N->getOperand(0);
2589 SDOperand N1 = N->getOperand(1);
2590 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2591 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
2592 MVT::ValueType VT = N->getValueType(0);
2594 // fold (fadd c1, c2) -> c1+c2
2596 return DAG.getNode(ISD::FADD, VT, N0, N1);
2597 // canonicalize constant to RHS
2598 if (N0CFP && !N1CFP)
2599 return DAG.getNode(ISD::FADD, VT, N1, N0);
2600 // fold (A + (-B)) -> A-B
2601 if (N1.getOpcode() == ISD::FNEG)
2602 return DAG.getNode(ISD::FSUB, VT, N0, N1.getOperand(0));
2603 // fold ((-A) + B) -> B-A
2604 if (N0.getOpcode() == ISD::FNEG)
2605 return DAG.getNode(ISD::FSUB, VT, N1, N0.getOperand(0));
2609 SDOperand DAGCombiner::visitFSUB(SDNode *N) {
2610 SDOperand N0 = N->getOperand(0);
2611 SDOperand N1 = N->getOperand(1);
2612 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2613 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
2614 MVT::ValueType VT = N->getValueType(0);
2616 // fold (fsub c1, c2) -> c1-c2
2618 return DAG.getNode(ISD::FSUB, VT, N0, N1);
2619 // fold (A-(-B)) -> A+B
2620 if (N1.getOpcode() == ISD::FNEG)
2621 return DAG.getNode(ISD::FADD, VT, N0, N1.getOperand(0));
2625 SDOperand DAGCombiner::visitFMUL(SDNode *N) {
2626 SDOperand N0 = N->getOperand(0);
2627 SDOperand N1 = N->getOperand(1);
2628 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2629 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
2630 MVT::ValueType VT = N->getValueType(0);
2632 // fold (fmul c1, c2) -> c1*c2
2634 return DAG.getNode(ISD::FMUL, VT, N0, N1);
2635 // canonicalize constant to RHS
2636 if (N0CFP && !N1CFP)
2637 return DAG.getNode(ISD::FMUL, VT, N1, N0);
2638 // fold (fmul X, 2.0) -> (fadd X, X)
2639 if (N1CFP && N1CFP->isExactlyValue(+2.0))
2640 return DAG.getNode(ISD::FADD, VT, N0, N0);
2644 SDOperand DAGCombiner::visitFDIV(SDNode *N) {
2645 SDOperand N0 = N->getOperand(0);
2646 SDOperand N1 = N->getOperand(1);
2647 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2648 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
2649 MVT::ValueType VT = N->getValueType(0);
2651 // fold (fdiv c1, c2) -> c1/c2
2653 return DAG.getNode(ISD::FDIV, VT, N0, N1);
2657 SDOperand DAGCombiner::visitFREM(SDNode *N) {
2658 SDOperand N0 = N->getOperand(0);
2659 SDOperand N1 = N->getOperand(1);
2660 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2661 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
2662 MVT::ValueType VT = N->getValueType(0);
2664 // fold (frem c1, c2) -> fmod(c1,c2)
2666 return DAG.getNode(ISD::FREM, VT, N0, N1);
2670 SDOperand DAGCombiner::visitFCOPYSIGN(SDNode *N) {
2671 SDOperand N0 = N->getOperand(0);
2672 SDOperand N1 = N->getOperand(1);
2673 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2674 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
2675 MVT::ValueType VT = N->getValueType(0);
2677 if (N0CFP && N1CFP) // Constant fold
2678 return DAG.getNode(ISD::FCOPYSIGN, VT, N0, N1);
2681 // copysign(x, c1) -> fabs(x) iff ispos(c1)
2682 // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1)
2687 u.d = N1CFP->getValue();
2689 return DAG.getNode(ISD::FABS, VT, N0);
2691 return DAG.getNode(ISD::FNEG, VT, DAG.getNode(ISD::FABS, VT, N0));
2694 // copysign(fabs(x), y) -> copysign(x, y)
2695 // copysign(fneg(x), y) -> copysign(x, y)
2696 // copysign(copysign(x,z), y) -> copysign(x, y)
2697 if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG ||
2698 N0.getOpcode() == ISD::FCOPYSIGN)
2699 return DAG.getNode(ISD::FCOPYSIGN, VT, N0.getOperand(0), N1);
2701 // copysign(x, abs(y)) -> abs(x)
2702 if (N1.getOpcode() == ISD::FABS)
2703 return DAG.getNode(ISD::FABS, VT, N0);
2705 // copysign(x, copysign(y,z)) -> copysign(x, z)
2706 if (N1.getOpcode() == ISD::FCOPYSIGN)
2707 return DAG.getNode(ISD::FCOPYSIGN, VT, N0, N1.getOperand(1));
2709 // copysign(x, fp_extend(y)) -> copysign(x, y)
2710 // copysign(x, fp_round(y)) -> copysign(x, y)
2711 if (N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND)
2712 return DAG.getNode(ISD::FCOPYSIGN, VT, N0, N1.getOperand(0));
2719 SDOperand DAGCombiner::visitSINT_TO_FP(SDNode *N) {
2720 SDOperand N0 = N->getOperand(0);
2721 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2722 MVT::ValueType VT = N->getValueType(0);
2724 // fold (sint_to_fp c1) -> c1fp
2726 return DAG.getNode(ISD::SINT_TO_FP, VT, N0);
2730 SDOperand DAGCombiner::visitUINT_TO_FP(SDNode *N) {
2731 SDOperand N0 = N->getOperand(0);
2732 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2733 MVT::ValueType VT = N->getValueType(0);
2735 // fold (uint_to_fp c1) -> c1fp
2737 return DAG.getNode(ISD::UINT_TO_FP, VT, N0);
2741 SDOperand DAGCombiner::visitFP_TO_SINT(SDNode *N) {
2742 SDOperand N0 = N->getOperand(0);
2743 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2744 MVT::ValueType VT = N->getValueType(0);
2746 // fold (fp_to_sint c1fp) -> c1
2748 return DAG.getNode(ISD::FP_TO_SINT, VT, N0);
2752 SDOperand DAGCombiner::visitFP_TO_UINT(SDNode *N) {
2753 SDOperand N0 = N->getOperand(0);
2754 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2755 MVT::ValueType VT = N->getValueType(0);
2757 // fold (fp_to_uint c1fp) -> c1
2759 return DAG.getNode(ISD::FP_TO_UINT, VT, N0);
2763 SDOperand DAGCombiner::visitFP_ROUND(SDNode *N) {
2764 SDOperand N0 = N->getOperand(0);
2765 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2766 MVT::ValueType VT = N->getValueType(0);
2768 // fold (fp_round c1fp) -> c1fp
2770 return DAG.getNode(ISD::FP_ROUND, VT, N0);
2772 // fold (fp_round (fp_extend x)) -> x
2773 if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType())
2774 return N0.getOperand(0);
2776 // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y)
2777 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.Val->hasOneUse()) {
2778 SDOperand Tmp = DAG.getNode(ISD::FP_ROUND, VT, N0.getOperand(0));
2779 AddToWorkList(Tmp.Val);
2780 return DAG.getNode(ISD::FCOPYSIGN, VT, Tmp, N0.getOperand(1));
2786 SDOperand DAGCombiner::visitFP_ROUND_INREG(SDNode *N) {
2787 SDOperand N0 = N->getOperand(0);
2788 MVT::ValueType VT = N->getValueType(0);
2789 MVT::ValueType EVT = cast<VTSDNode>(N->getOperand(1))->getVT();
2790 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2792 // fold (fp_round_inreg c1fp) -> c1fp
2794 SDOperand Round = DAG.getConstantFP(N0CFP->getValue(), EVT);
2795 return DAG.getNode(ISD::FP_EXTEND, VT, Round);
2800 SDOperand DAGCombiner::visitFP_EXTEND(SDNode *N) {
2801 SDOperand N0 = N->getOperand(0);
2802 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2803 MVT::ValueType VT = N->getValueType(0);
2805 // fold (fp_extend c1fp) -> c1fp
2807 return DAG.getNode(ISD::FP_EXTEND, VT, N0);
2809 // fold (fpext (load x)) -> (fpext (fpround (extload x)))
2810 if (ISD::isNON_EXTLoad(N0.Val) && N0.hasOneUse() &&
2811 (!AfterLegalize||TLI.isLoadXLegal(ISD::EXTLOAD, N0.getValueType()))) {
2812 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2813 SDOperand ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, VT, LN0->getChain(),
2814 LN0->getBasePtr(), LN0->getSrcValue(),
2815 LN0->getSrcValueOffset(),
2817 CombineTo(N, ExtLoad);
2818 CombineTo(N0.Val, DAG.getNode(ISD::FP_ROUND, N0.getValueType(), ExtLoad),
2819 ExtLoad.getValue(1));
2820 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
2827 SDOperand DAGCombiner::visitFNEG(SDNode *N) {
2828 SDOperand N0 = N->getOperand(0);
2829 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2830 MVT::ValueType VT = N->getValueType(0);
2832 // fold (fneg c1) -> -c1
2834 return DAG.getNode(ISD::FNEG, VT, N0);
2835 // fold (fneg (sub x, y)) -> (sub y, x)
2836 if (N0.getOpcode() == ISD::SUB)
2837 return DAG.getNode(ISD::SUB, VT, N0.getOperand(1), N0.getOperand(0));
2838 // fold (fneg (fneg x)) -> x
2839 if (N0.getOpcode() == ISD::FNEG)
2840 return N0.getOperand(0);
2844 SDOperand DAGCombiner::visitFABS(SDNode *N) {
2845 SDOperand N0 = N->getOperand(0);
2846 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2847 MVT::ValueType VT = N->getValueType(0);
2849 // fold (fabs c1) -> fabs(c1)
2851 return DAG.getNode(ISD::FABS, VT, N0);
2852 // fold (fabs (fabs x)) -> (fabs x)
2853 if (N0.getOpcode() == ISD::FABS)
2854 return N->getOperand(0);
2855 // fold (fabs (fneg x)) -> (fabs x)
2856 // fold (fabs (fcopysign x, y)) -> (fabs x)
2857 if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN)
2858 return DAG.getNode(ISD::FABS, VT, N0.getOperand(0));
2863 SDOperand DAGCombiner::visitBRCOND(SDNode *N) {
2864 SDOperand Chain = N->getOperand(0);
2865 SDOperand N1 = N->getOperand(1);
2866 SDOperand N2 = N->getOperand(2);
2867 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2869 // never taken branch, fold to chain
2870 if (N1C && N1C->isNullValue())
2872 // unconditional branch
2873 if (N1C && N1C->getValue() == 1)
2874 return DAG.getNode(ISD::BR, MVT::Other, Chain, N2);
2875 // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal
2877 if (N1.getOpcode() == ISD::SETCC &&
2878 TLI.isOperationLegal(ISD::BR_CC, MVT::Other)) {
2879 return DAG.getNode(ISD::BR_CC, MVT::Other, Chain, N1.getOperand(2),
2880 N1.getOperand(0), N1.getOperand(1), N2);
2885 // Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB.
2887 SDOperand DAGCombiner::visitBR_CC(SDNode *N) {
2888 CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1));
2889 SDOperand CondLHS = N->getOperand(2), CondRHS = N->getOperand(3);
2891 // Use SimplifySetCC to simplify SETCC's.
2892 SDOperand Simp = SimplifySetCC(MVT::i1, CondLHS, CondRHS, CC->get(), false);
2893 if (Simp.Val) AddToWorkList(Simp.Val);
2895 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(Simp.Val);
2897 // fold br_cc true, dest -> br dest (unconditional branch)
2898 if (SCCC && SCCC->getValue())
2899 return DAG.getNode(ISD::BR, MVT::Other, N->getOperand(0),
2901 // fold br_cc false, dest -> unconditional fall through
2902 if (SCCC && SCCC->isNullValue())
2903 return N->getOperand(0);
2905 // fold to a simpler setcc
2906 if (Simp.Val && Simp.getOpcode() == ISD::SETCC)
2907 return DAG.getNode(ISD::BR_CC, MVT::Other, N->getOperand(0),
2908 Simp.getOperand(2), Simp.getOperand(0),
2909 Simp.getOperand(1), N->getOperand(4));
2913 SDOperand DAGCombiner::visitLOAD(SDNode *N) {
2914 LoadSDNode *LD = cast<LoadSDNode>(N);
2915 SDOperand Chain = LD->getChain();
2916 SDOperand Ptr = LD->getBasePtr();
2918 // If there are no uses of the loaded value, change uses of the chain value
2919 // into uses of the chain input (i.e. delete the dead load).
2920 if (N->hasNUsesOfValue(0, 0))
2921 return CombineTo(N, DAG.getNode(ISD::UNDEF, N->getValueType(0)), Chain);
2923 // If this load is directly stored, replace the load value with the stored
2925 // TODO: Handle store large -> read small portion.
2926 // TODO: Handle TRUNCSTORE/LOADEXT
2927 if (LD->getExtensionType() == ISD::NON_EXTLOAD) {
2928 if (ISD::isNON_TRUNCStore(Chain.Val)) {
2929 StoreSDNode *PrevST = cast<StoreSDNode>(Chain);
2930 if (PrevST->getBasePtr() == Ptr &&
2931 PrevST->getValue().getValueType() == N->getValueType(0))
2932 return CombineTo(N, Chain.getOperand(1), Chain);
2937 // Walk up chain skipping non-aliasing memory nodes.
2938 SDOperand BetterChain = FindBetterChain(N, Chain);
2940 // If there is a better chain.
2941 if (Chain != BetterChain) {
2944 // Replace the chain to void dependency.
2945 if (LD->getExtensionType() == ISD::NON_EXTLOAD) {
2946 ReplLoad = DAG.getLoad(N->getValueType(0), BetterChain, Ptr,
2947 LD->getSrcValue(), LD->getSrcValueOffset());
2949 ReplLoad = DAG.getExtLoad(LD->getExtensionType(),
2950 LD->getValueType(0),
2951 BetterChain, Ptr, LD->getSrcValue(),
2952 LD->getSrcValueOffset(),
2956 // Create token factor to keep old chain connected.
2957 SDOperand Token = DAG.getNode(ISD::TokenFactor, MVT::Other,
2958 Chain, ReplLoad.getValue(1));
2960 // Replace uses with load result and token factor. Don't add users
2962 return CombineTo(N, ReplLoad.getValue(0), Token, false);
2966 // Try transforming N to an indexed load.
2967 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
2968 return SDOperand(N, 0);
2973 SDOperand DAGCombiner::visitSTORE(SDNode *N) {
2974 StoreSDNode *ST = cast<StoreSDNode>(N);
2975 SDOperand Chain = ST->getChain();
2976 SDOperand Value = ST->getValue();
2977 SDOperand Ptr = ST->getBasePtr();
2979 // If this is a store of a bit convert, store the input value.
2980 // FIXME: This needs to know that the resultant store does not need a
2981 // higher alignment than the original.
2982 if (0 && Value.getOpcode() == ISD::BIT_CONVERT) {
2983 return DAG.getStore(Chain, Value.getOperand(0), Ptr, ST->getSrcValue(),
2984 ST->getSrcValueOffset());
2988 // Walk up chain skipping non-aliasing memory nodes.
2989 SDOperand BetterChain = FindBetterChain(N, Chain);
2991 // If there is a better chain.
2992 if (Chain != BetterChain) {
2993 // Replace the chain to avoid dependency.
2994 SDOperand ReplStore;
2995 if (ST->isTruncatingStore()) {
2996 ReplStore = DAG.getTruncStore(BetterChain, Value, Ptr,
2997 ST->getSrcValue(),ST->getSrcValueOffset(), ST->getStoredVT());
2999 ReplStore = DAG.getStore(BetterChain, Value, Ptr,
3000 ST->getSrcValue(), ST->getSrcValueOffset());
3003 // Create token to keep both nodes around.
3005 DAG.getNode(ISD::TokenFactor, MVT::Other, Chain, ReplStore);
3007 // Don't add users to work list.
3008 return CombineTo(N, Token, false);
3012 // Try transforming N to an indexed store.
3013 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
3014 return SDOperand(N, 0);
3019 SDOperand DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) {
3020 SDOperand InVec = N->getOperand(0);
3021 SDOperand InVal = N->getOperand(1);
3022 SDOperand EltNo = N->getOperand(2);
3024 // If the invec is a BUILD_VECTOR and if EltNo is a constant, build a new
3025 // vector with the inserted element.
3026 if (InVec.getOpcode() == ISD::BUILD_VECTOR && isa<ConstantSDNode>(EltNo)) {
3027 unsigned Elt = cast<ConstantSDNode>(EltNo)->getValue();
3028 SmallVector<SDOperand, 8> Ops(InVec.Val->op_begin(), InVec.Val->op_end());
3029 if (Elt < Ops.size())
3031 return DAG.getNode(ISD::BUILD_VECTOR, InVec.getValueType(),
3032 &Ops[0], Ops.size());
3038 SDOperand DAGCombiner::visitVINSERT_VECTOR_ELT(SDNode *N) {
3039 SDOperand InVec = N->getOperand(0);
3040 SDOperand InVal = N->getOperand(1);
3041 SDOperand EltNo = N->getOperand(2);
3042 SDOperand NumElts = N->getOperand(3);
3043 SDOperand EltType = N->getOperand(4);
3045 // If the invec is a VBUILD_VECTOR and if EltNo is a constant, build a new
3046 // vector with the inserted element.
3047 if (InVec.getOpcode() == ISD::VBUILD_VECTOR && isa<ConstantSDNode>(EltNo)) {
3048 unsigned Elt = cast<ConstantSDNode>(EltNo)->getValue();
3049 SmallVector<SDOperand, 8> Ops(InVec.Val->op_begin(), InVec.Val->op_end());
3050 if (Elt < Ops.size()-2)
3052 return DAG.getNode(ISD::VBUILD_VECTOR, InVec.getValueType(),
3053 &Ops[0], Ops.size());
3059 SDOperand DAGCombiner::visitVBUILD_VECTOR(SDNode *N) {
3060 unsigned NumInScalars = N->getNumOperands()-2;
3061 SDOperand NumElts = N->getOperand(NumInScalars);
3062 SDOperand EltType = N->getOperand(NumInScalars+1);
3064 // Check to see if this is a VBUILD_VECTOR of a bunch of VEXTRACT_VECTOR_ELT
3065 // operations. If so, and if the EXTRACT_ELT vector inputs come from at most
3066 // two distinct vectors, turn this into a shuffle node.
3067 SDOperand VecIn1, VecIn2;
3068 for (unsigned i = 0; i != NumInScalars; ++i) {
3069 // Ignore undef inputs.
3070 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
3072 // If this input is something other than a VEXTRACT_VECTOR_ELT with a
3073 // constant index, bail out.
3074 if (N->getOperand(i).getOpcode() != ISD::VEXTRACT_VECTOR_ELT ||
3075 !isa<ConstantSDNode>(N->getOperand(i).getOperand(1))) {
3076 VecIn1 = VecIn2 = SDOperand(0, 0);
3080 // If the input vector type disagrees with the result of the vbuild_vector,
3081 // we can't make a shuffle.
3082 SDOperand ExtractedFromVec = N->getOperand(i).getOperand(0);
3083 if (*(ExtractedFromVec.Val->op_end()-2) != NumElts ||
3084 *(ExtractedFromVec.Val->op_end()-1) != EltType) {
3085 VecIn1 = VecIn2 = SDOperand(0, 0);
3089 // Otherwise, remember this. We allow up to two distinct input vectors.
3090 if (ExtractedFromVec == VecIn1 || ExtractedFromVec == VecIn2)
3093 if (VecIn1.Val == 0) {
3094 VecIn1 = ExtractedFromVec;
3095 } else if (VecIn2.Val == 0) {
3096 VecIn2 = ExtractedFromVec;
3099 VecIn1 = VecIn2 = SDOperand(0, 0);
3104 // If everything is good, we can make a shuffle operation.
3106 SmallVector<SDOperand, 8> BuildVecIndices;
3107 for (unsigned i = 0; i != NumInScalars; ++i) {
3108 if (N->getOperand(i).getOpcode() == ISD::UNDEF) {
3109 BuildVecIndices.push_back(DAG.getNode(ISD::UNDEF, MVT::i32));
3113 SDOperand Extract = N->getOperand(i);
3115 // If extracting from the first vector, just use the index directly.
3116 if (Extract.getOperand(0) == VecIn1) {
3117 BuildVecIndices.push_back(Extract.getOperand(1));
3121 // Otherwise, use InIdx + VecSize
3122 unsigned Idx = cast<ConstantSDNode>(Extract.getOperand(1))->getValue();
3123 BuildVecIndices.push_back(DAG.getConstant(Idx+NumInScalars, MVT::i32));
3126 // Add count and size info.
3127 BuildVecIndices.push_back(NumElts);
3128 BuildVecIndices.push_back(DAG.getValueType(MVT::i32));
3130 // Return the new VVECTOR_SHUFFLE node.
3136 // Use an undef vbuild_vector as input for the second operand.
3137 std::vector<SDOperand> UnOps(NumInScalars,
3138 DAG.getNode(ISD::UNDEF,
3139 cast<VTSDNode>(EltType)->getVT()));
3140 UnOps.push_back(NumElts);
3141 UnOps.push_back(EltType);
3142 Ops[1] = DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector,
3143 &UnOps[0], UnOps.size());
3144 AddToWorkList(Ops[1].Val);
3146 Ops[2] = DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector,
3147 &BuildVecIndices[0], BuildVecIndices.size());
3150 return DAG.getNode(ISD::VVECTOR_SHUFFLE, MVT::Vector, Ops, 5);
3156 SDOperand DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
3157 SDOperand ShufMask = N->getOperand(2);
3158 unsigned NumElts = ShufMask.getNumOperands();
3160 // If the shuffle mask is an identity operation on the LHS, return the LHS.
3161 bool isIdentity = true;
3162 for (unsigned i = 0; i != NumElts; ++i) {
3163 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF &&
3164 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() != i) {
3169 if (isIdentity) return N->getOperand(0);
3171 // If the shuffle mask is an identity operation on the RHS, return the RHS.
3173 for (unsigned i = 0; i != NumElts; ++i) {
3174 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF &&
3175 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() != i+NumElts) {
3180 if (isIdentity) return N->getOperand(1);
3182 // Check if the shuffle is a unary shuffle, i.e. one of the vectors is not
3184 bool isUnary = true;
3185 bool isSplat = true;
3187 unsigned BaseIdx = 0;
3188 for (unsigned i = 0; i != NumElts; ++i)
3189 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF) {
3190 unsigned Idx = cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue();
3191 int V = (Idx < NumElts) ? 0 : 1;
3205 SDOperand N0 = N->getOperand(0);
3206 SDOperand N1 = N->getOperand(1);
3207 // Normalize unary shuffle so the RHS is undef.
3208 if (isUnary && VecNum == 1)
3211 // If it is a splat, check if the argument vector is a build_vector with
3212 // all scalar elements the same.
3215 if (V->getOpcode() == ISD::BIT_CONVERT)
3216 V = V->getOperand(0).Val;
3217 if (V->getOpcode() == ISD::BUILD_VECTOR) {
3218 unsigned NumElems = V->getNumOperands()-2;
3219 if (NumElems > BaseIdx) {
3221 bool AllSame = true;
3222 for (unsigned i = 0; i != NumElems; ++i) {
3223 if (V->getOperand(i).getOpcode() != ISD::UNDEF) {
3224 Base = V->getOperand(i);
3228 // Splat of <u, u, u, u>, return <u, u, u, u>
3231 for (unsigned i = 0; i != NumElems; ++i) {
3232 if (V->getOperand(i).getOpcode() != ISD::UNDEF &&
3233 V->getOperand(i) != Base) {
3238 // Splat of <x, x, x, x>, return <x, x, x, x>
3245 // If it is a unary or the LHS and the RHS are the same node, turn the RHS
3247 if (isUnary || N0 == N1) {
3248 if (N0.getOpcode() == ISD::UNDEF)
3249 return DAG.getNode(ISD::UNDEF, N->getValueType(0));
3250 // Check the SHUFFLE mask, mapping any inputs from the 2nd operand into the
3252 SmallVector<SDOperand, 8> MappedOps;
3253 for (unsigned i = 0, e = ShufMask.getNumOperands(); i != e; ++i) {
3254 if (ShufMask.getOperand(i).getOpcode() == ISD::UNDEF ||
3255 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() < NumElts) {
3256 MappedOps.push_back(ShufMask.getOperand(i));
3259 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() - NumElts;
3260 MappedOps.push_back(DAG.getConstant(NewIdx, MVT::i32));
3263 ShufMask = DAG.getNode(ISD::BUILD_VECTOR, ShufMask.getValueType(),
3264 &MappedOps[0], MappedOps.size());
3265 AddToWorkList(ShufMask.Val);
3266 return DAG.getNode(ISD::VECTOR_SHUFFLE, N->getValueType(0),
3268 DAG.getNode(ISD::UNDEF, N->getValueType(0)),
3275 SDOperand DAGCombiner::visitVVECTOR_SHUFFLE(SDNode *N) {
3276 SDOperand ShufMask = N->getOperand(2);
3277 unsigned NumElts = ShufMask.getNumOperands()-2;
3279 // If the shuffle mask is an identity operation on the LHS, return the LHS.
3280 bool isIdentity = true;
3281 for (unsigned i = 0; i != NumElts; ++i) {
3282 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF &&
3283 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() != i) {
3288 if (isIdentity) return N->getOperand(0);
3290 // If the shuffle mask is an identity operation on the RHS, return the RHS.
3292 for (unsigned i = 0; i != NumElts; ++i) {
3293 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF &&
3294 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() != i+NumElts) {
3299 if (isIdentity) return N->getOperand(1);
3301 // Check if the shuffle is a unary shuffle, i.e. one of the vectors is not
3303 bool isUnary = true;
3304 bool isSplat = true;
3306 unsigned BaseIdx = 0;
3307 for (unsigned i = 0; i != NumElts; ++i)
3308 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF) {
3309 unsigned Idx = cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue();
3310 int V = (Idx < NumElts) ? 0 : 1;
3324 SDOperand N0 = N->getOperand(0);
3325 SDOperand N1 = N->getOperand(1);
3326 // Normalize unary shuffle so the RHS is undef.
3327 if (isUnary && VecNum == 1)
3330 // If it is a splat, check if the argument vector is a build_vector with
3331 // all scalar elements the same.
3335 // If this is a vbit convert that changes the element type of the vector but
3336 // not the number of vector elements, look through it. Be careful not to
3337 // look though conversions that change things like v4f32 to v2f64.
3338 if (V->getOpcode() == ISD::VBIT_CONVERT) {
3339 SDOperand ConvInput = V->getOperand(0);
3340 if (ConvInput.getValueType() == MVT::Vector &&
3342 ConvInput.getConstantOperandVal(ConvInput.getNumOperands()-2))
3346 if (V->getOpcode() == ISD::VBUILD_VECTOR) {
3347 unsigned NumElems = V->getNumOperands()-2;
3348 if (NumElems > BaseIdx) {
3350 bool AllSame = true;
3351 for (unsigned i = 0; i != NumElems; ++i) {
3352 if (V->getOperand(i).getOpcode() != ISD::UNDEF) {
3353 Base = V->getOperand(i);
3357 // Splat of <u, u, u, u>, return <u, u, u, u>
3360 for (unsigned i = 0; i != NumElems; ++i) {
3361 if (V->getOperand(i).getOpcode() != ISD::UNDEF &&
3362 V->getOperand(i) != Base) {
3367 // Splat of <x, x, x, x>, return <x, x, x, x>
3374 // If it is a unary or the LHS and the RHS are the same node, turn the RHS
3376 if (isUnary || N0 == N1) {
3377 // Check the SHUFFLE mask, mapping any inputs from the 2nd operand into the
3379 SmallVector<SDOperand, 8> MappedOps;
3380 for (unsigned i = 0; i != NumElts; ++i) {
3381 if (ShufMask.getOperand(i).getOpcode() == ISD::UNDEF ||
3382 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() < NumElts) {
3383 MappedOps.push_back(ShufMask.getOperand(i));
3386 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() - NumElts;
3387 MappedOps.push_back(DAG.getConstant(NewIdx, MVT::i32));
3390 // Add the type/#elts values.
3391 MappedOps.push_back(ShufMask.getOperand(NumElts));
3392 MappedOps.push_back(ShufMask.getOperand(NumElts+1));
3394 ShufMask = DAG.getNode(ISD::VBUILD_VECTOR, ShufMask.getValueType(),
3395 &MappedOps[0], MappedOps.size());
3396 AddToWorkList(ShufMask.Val);
3398 // Build the undef vector.
3399 SDOperand UDVal = DAG.getNode(ISD::UNDEF, MappedOps[0].getValueType());
3400 for (unsigned i = 0; i != NumElts; ++i)
3401 MappedOps[i] = UDVal;
3402 MappedOps[NumElts ] = *(N0.Val->op_end()-2);
3403 MappedOps[NumElts+1] = *(N0.Val->op_end()-1);
3404 UDVal = DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector,
3405 &MappedOps[0], MappedOps.size());
3407 return DAG.getNode(ISD::VVECTOR_SHUFFLE, MVT::Vector,
3408 N0, UDVal, ShufMask,
3409 MappedOps[NumElts], MappedOps[NumElts+1]);
3415 /// XformToShuffleWithZero - Returns a vector_shuffle if it able to transform
3416 /// a VAND to a vector_shuffle with the destination vector and a zero vector.
3417 /// e.g. VAND V, <0xffffffff, 0, 0xffffffff, 0>. ==>
3418 /// vector_shuffle V, Zero, <0, 4, 2, 4>
3419 SDOperand DAGCombiner::XformToShuffleWithZero(SDNode *N) {
3420 SDOperand LHS = N->getOperand(0);
3421 SDOperand RHS = N->getOperand(1);
3422 if (N->getOpcode() == ISD::VAND) {
3423 SDOperand DstVecSize = *(LHS.Val->op_end()-2);
3424 SDOperand DstVecEVT = *(LHS.Val->op_end()-1);
3425 if (RHS.getOpcode() == ISD::VBIT_CONVERT)
3426 RHS = RHS.getOperand(0);
3427 if (RHS.getOpcode() == ISD::VBUILD_VECTOR) {
3428 std::vector<SDOperand> IdxOps;
3429 unsigned NumOps = RHS.getNumOperands();
3430 unsigned NumElts = NumOps-2;
3431 MVT::ValueType EVT = cast<VTSDNode>(RHS.getOperand(NumOps-1))->getVT();
3432 for (unsigned i = 0; i != NumElts; ++i) {
3433 SDOperand Elt = RHS.getOperand(i);
3434 if (!isa<ConstantSDNode>(Elt))
3436 else if (cast<ConstantSDNode>(Elt)->isAllOnesValue())
3437 IdxOps.push_back(DAG.getConstant(i, EVT));
3438 else if (cast<ConstantSDNode>(Elt)->isNullValue())
3439 IdxOps.push_back(DAG.getConstant(NumElts, EVT));
3444 // Let's see if the target supports this vector_shuffle.
3445 if (!TLI.isVectorClearMaskLegal(IdxOps, EVT, DAG))
3448 // Return the new VVECTOR_SHUFFLE node.
3449 SDOperand NumEltsNode = DAG.getConstant(NumElts, MVT::i32);
3450 SDOperand EVTNode = DAG.getValueType(EVT);
3451 std::vector<SDOperand> Ops;
3452 LHS = DAG.getNode(ISD::VBIT_CONVERT, MVT::Vector, LHS, NumEltsNode,
3455 AddToWorkList(LHS.Val);
3456 std::vector<SDOperand> ZeroOps(NumElts, DAG.getConstant(0, EVT));
3457 ZeroOps.push_back(NumEltsNode);
3458 ZeroOps.push_back(EVTNode);
3459 Ops.push_back(DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector,
3460 &ZeroOps[0], ZeroOps.size()));
3461 IdxOps.push_back(NumEltsNode);
3462 IdxOps.push_back(EVTNode);
3463 Ops.push_back(DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector,
3464 &IdxOps[0], IdxOps.size()));
3465 Ops.push_back(NumEltsNode);
3466 Ops.push_back(EVTNode);
3467 SDOperand Result = DAG.getNode(ISD::VVECTOR_SHUFFLE, MVT::Vector,
3468 &Ops[0], Ops.size());
3469 if (NumEltsNode != DstVecSize || EVTNode != DstVecEVT) {
3470 Result = DAG.getNode(ISD::VBIT_CONVERT, MVT::Vector, Result,
3471 DstVecSize, DstVecEVT);
3479 /// visitVBinOp - Visit a binary vector operation, like VADD. IntOp indicates
3480 /// the scalar operation of the vop if it is operating on an integer vector
3481 /// (e.g. ADD) and FPOp indicates the FP version (e.g. FADD).
3482 SDOperand DAGCombiner::visitVBinOp(SDNode *N, ISD::NodeType IntOp,
3483 ISD::NodeType FPOp) {
3484 MVT::ValueType EltType = cast<VTSDNode>(*(N->op_end()-1))->getVT();
3485 ISD::NodeType ScalarOp = MVT::isInteger(EltType) ? IntOp : FPOp;
3486 SDOperand LHS = N->getOperand(0);
3487 SDOperand RHS = N->getOperand(1);
3488 SDOperand Shuffle = XformToShuffleWithZero(N);
3489 if (Shuffle.Val) return Shuffle;
3491 // If the LHS and RHS are VBUILD_VECTOR nodes, see if we can constant fold
3493 if (LHS.getOpcode() == ISD::VBUILD_VECTOR &&
3494 RHS.getOpcode() == ISD::VBUILD_VECTOR) {
3495 SmallVector<SDOperand, 8> Ops;
3496 for (unsigned i = 0, e = LHS.getNumOperands()-2; i != e; ++i) {
3497 SDOperand LHSOp = LHS.getOperand(i);
3498 SDOperand RHSOp = RHS.getOperand(i);
3499 // If these two elements can't be folded, bail out.
3500 if ((LHSOp.getOpcode() != ISD::UNDEF &&
3501 LHSOp.getOpcode() != ISD::Constant &&
3502 LHSOp.getOpcode() != ISD::ConstantFP) ||
3503 (RHSOp.getOpcode() != ISD::UNDEF &&
3504 RHSOp.getOpcode() != ISD::Constant &&
3505 RHSOp.getOpcode() != ISD::ConstantFP))
3507 // Can't fold divide by zero.
3508 if (N->getOpcode() == ISD::VSDIV || N->getOpcode() == ISD::VUDIV) {
3509 if ((RHSOp.getOpcode() == ISD::Constant &&
3510 cast<ConstantSDNode>(RHSOp.Val)->isNullValue()) ||
3511 (RHSOp.getOpcode() == ISD::ConstantFP &&
3512 !cast<ConstantFPSDNode>(RHSOp.Val)->getValue()))
3515 Ops.push_back(DAG.getNode(ScalarOp, EltType, LHSOp, RHSOp));
3516 AddToWorkList(Ops.back().Val);
3517 assert((Ops.back().getOpcode() == ISD::UNDEF ||
3518 Ops.back().getOpcode() == ISD::Constant ||
3519 Ops.back().getOpcode() == ISD::ConstantFP) &&
3520 "Scalar binop didn't fold!");
3523 if (Ops.size() == LHS.getNumOperands()-2) {
3524 Ops.push_back(*(LHS.Val->op_end()-2));
3525 Ops.push_back(*(LHS.Val->op_end()-1));
3526 return DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, &Ops[0], Ops.size());
3533 SDOperand DAGCombiner::SimplifySelect(SDOperand N0, SDOperand N1, SDOperand N2){
3534 assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!");
3536 SDOperand SCC = SimplifySelectCC(N0.getOperand(0), N0.getOperand(1), N1, N2,
3537 cast<CondCodeSDNode>(N0.getOperand(2))->get());
3538 // If we got a simplified select_cc node back from SimplifySelectCC, then
3539 // break it down into a new SETCC node, and a new SELECT node, and then return
3540 // the SELECT node, since we were called with a SELECT node.
3542 // Check to see if we got a select_cc back (to turn into setcc/select).
3543 // Otherwise, just return whatever node we got back, like fabs.
3544 if (SCC.getOpcode() == ISD::SELECT_CC) {
3545 SDOperand SETCC = DAG.getNode(ISD::SETCC, N0.getValueType(),
3546 SCC.getOperand(0), SCC.getOperand(1),
3548 AddToWorkList(SETCC.Val);
3549 return DAG.getNode(ISD::SELECT, SCC.getValueType(), SCC.getOperand(2),
3550 SCC.getOperand(3), SETCC);
3557 /// SimplifySelectOps - Given a SELECT or a SELECT_CC node, where LHS and RHS
3558 /// are the two values being selected between, see if we can simplify the
3559 /// select. Callers of this should assume that TheSelect is deleted if this
3560 /// returns true. As such, they should return the appropriate thing (e.g. the
3561 /// node) back to the top-level of the DAG combiner loop to avoid it being
3564 bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDOperand LHS,
3567 // If this is a select from two identical things, try to pull the operation
3568 // through the select.
3569 if (LHS.getOpcode() == RHS.getOpcode() && LHS.hasOneUse() && RHS.hasOneUse()){
3570 // If this is a load and the token chain is identical, replace the select
3571 // of two loads with a load through a select of the address to load from.
3572 // This triggers in things like "select bool X, 10.0, 123.0" after the FP
3573 // constants have been dropped into the constant pool.
3574 if (LHS.getOpcode() == ISD::LOAD &&
3575 // Token chains must be identical.
3576 LHS.getOperand(0) == RHS.getOperand(0)) {
3577 LoadSDNode *LLD = cast<LoadSDNode>(LHS);
3578 LoadSDNode *RLD = cast<LoadSDNode>(RHS);
3580 // If this is an EXTLOAD, the VT's must match.
3581 if (LLD->getLoadedVT() == RLD->getLoadedVT()) {
3582 // FIXME: this conflates two src values, discarding one. This is not
3583 // the right thing to do, but nothing uses srcvalues now. When they do,
3584 // turn SrcValue into a list of locations.
3586 if (TheSelect->getOpcode() == ISD::SELECT)
3587 Addr = DAG.getNode(ISD::SELECT, LLD->getBasePtr().getValueType(),
3588 TheSelect->getOperand(0), LLD->getBasePtr(),
3591 Addr = DAG.getNode(ISD::SELECT_CC, LLD->getBasePtr().getValueType(),
3592 TheSelect->getOperand(0),
3593 TheSelect->getOperand(1),
3594 LLD->getBasePtr(), RLD->getBasePtr(),
3595 TheSelect->getOperand(4));
3598 if (LLD->getExtensionType() == ISD::NON_EXTLOAD)
3599 Load = DAG.getLoad(TheSelect->getValueType(0), LLD->getChain(),
3600 Addr,LLD->getSrcValue(), LLD->getSrcValueOffset());
3602 Load = DAG.getExtLoad(LLD->getExtensionType(),
3603 TheSelect->getValueType(0),
3604 LLD->getChain(), Addr, LLD->getSrcValue(),
3605 LLD->getSrcValueOffset(),
3606 LLD->getLoadedVT());
3608 // Users of the select now use the result of the load.
3609 CombineTo(TheSelect, Load);
3611 // Users of the old loads now use the new load's chain. We know the
3612 // old-load value is dead now.
3613 CombineTo(LHS.Val, Load.getValue(0), Load.getValue(1));
3614 CombineTo(RHS.Val, Load.getValue(0), Load.getValue(1));
3623 SDOperand DAGCombiner::SimplifySelectCC(SDOperand N0, SDOperand N1,
3624 SDOperand N2, SDOperand N3,
3627 MVT::ValueType VT = N2.getValueType();
3628 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val);
3629 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.Val);
3630 ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N3.Val);
3632 // Determine if the condition we're dealing with is constant
3633 SDOperand SCC = SimplifySetCC(TLI.getSetCCResultTy(), N0, N1, CC, false);
3634 if (SCC.Val) AddToWorkList(SCC.Val);
3635 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.Val);
3637 // fold select_cc true, x, y -> x
3638 if (SCCC && SCCC->getValue())
3640 // fold select_cc false, x, y -> y
3641 if (SCCC && SCCC->getValue() == 0)
3644 // Check to see if we can simplify the select into an fabs node
3645 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) {
3646 // Allow either -0.0 or 0.0
3647 if (CFP->getValue() == 0.0) {
3648 // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs
3649 if ((CC == ISD::SETGE || CC == ISD::SETGT) &&
3650 N0 == N2 && N3.getOpcode() == ISD::FNEG &&
3651 N2 == N3.getOperand(0))
3652 return DAG.getNode(ISD::FABS, VT, N0);
3654 // select (setl[te] X, +/-0.0), fneg(X), X -> fabs
3655 if ((CC == ISD::SETLT || CC == ISD::SETLE) &&
3656 N0 == N3 && N2.getOpcode() == ISD::FNEG &&
3657 N2.getOperand(0) == N3)
3658 return DAG.getNode(ISD::FABS, VT, N3);
3662 // Check to see if we can perform the "gzip trick", transforming
3663 // select_cc setlt X, 0, A, 0 -> and (sra X, size(X)-1), A
3664 if (N1C && N3C && N3C->isNullValue() && CC == ISD::SETLT &&
3665 MVT::isInteger(N0.getValueType()) &&
3666 MVT::isInteger(N2.getValueType()) &&
3667 (N1C->isNullValue() || // (a < 0) ? b : 0
3668 (N1C->getValue() == 1 && N0 == N2))) { // (a < 1) ? a : 0
3669 MVT::ValueType XType = N0.getValueType();
3670 MVT::ValueType AType = N2.getValueType();
3671 if (XType >= AType) {
3672 // and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a
3673 // single-bit constant.
3674 if (N2C && ((N2C->getValue() & (N2C->getValue()-1)) == 0)) {
3675 unsigned ShCtV = Log2_64(N2C->getValue());
3676 ShCtV = MVT::getSizeInBits(XType)-ShCtV-1;
3677 SDOperand ShCt = DAG.getConstant(ShCtV, TLI.getShiftAmountTy());
3678 SDOperand Shift = DAG.getNode(ISD::SRL, XType, N0, ShCt);
3679 AddToWorkList(Shift.Val);
3680 if (XType > AType) {
3681 Shift = DAG.getNode(ISD::TRUNCATE, AType, Shift);
3682 AddToWorkList(Shift.Val);
3684 return DAG.getNode(ISD::AND, AType, Shift, N2);
3686 SDOperand Shift = DAG.getNode(ISD::SRA, XType, N0,
3687 DAG.getConstant(MVT::getSizeInBits(XType)-1,
3688 TLI.getShiftAmountTy()));
3689 AddToWorkList(Shift.Val);
3690 if (XType > AType) {
3691 Shift = DAG.getNode(ISD::TRUNCATE, AType, Shift);
3692 AddToWorkList(Shift.Val);
3694 return DAG.getNode(ISD::AND, AType, Shift, N2);
3698 // fold select C, 16, 0 -> shl C, 4
3699 if (N2C && N3C && N3C->isNullValue() && isPowerOf2_64(N2C->getValue()) &&
3700 TLI.getSetCCResultContents() == TargetLowering::ZeroOrOneSetCCResult) {
3701 // Get a SetCC of the condition
3702 // FIXME: Should probably make sure that setcc is legal if we ever have a
3703 // target where it isn't.
3704 SDOperand Temp, SCC;
3705 // cast from setcc result type to select result type
3706 if (AfterLegalize) {
3707 SCC = DAG.getSetCC(TLI.getSetCCResultTy(), N0, N1, CC);
3708 Temp = DAG.getZeroExtendInReg(SCC, N2.getValueType());
3710 SCC = DAG.getSetCC(MVT::i1, N0, N1, CC);
3711 Temp = DAG.getNode(ISD::ZERO_EXTEND, N2.getValueType(), SCC);
3713 AddToWorkList(SCC.Val);
3714 AddToWorkList(Temp.Val);
3715 // shl setcc result by log2 n2c
3716 return DAG.getNode(ISD::SHL, N2.getValueType(), Temp,
3717 DAG.getConstant(Log2_64(N2C->getValue()),
3718 TLI.getShiftAmountTy()));
3721 // Check to see if this is the equivalent of setcc
3722 // FIXME: Turn all of these into setcc if setcc if setcc is legal
3723 // otherwise, go ahead with the folds.
3724 if (0 && N3C && N3C->isNullValue() && N2C && (N2C->getValue() == 1ULL)) {
3725 MVT::ValueType XType = N0.getValueType();
3726 if (TLI.isOperationLegal(ISD::SETCC, TLI.getSetCCResultTy())) {
3727 SDOperand Res = DAG.getSetCC(TLI.getSetCCResultTy(), N0, N1, CC);
3728 if (Res.getValueType() != VT)
3729 Res = DAG.getNode(ISD::ZERO_EXTEND, VT, Res);
3733 // seteq X, 0 -> srl (ctlz X, log2(size(X)))
3734 if (N1C && N1C->isNullValue() && CC == ISD::SETEQ &&
3735 TLI.isOperationLegal(ISD::CTLZ, XType)) {
3736 SDOperand Ctlz = DAG.getNode(ISD::CTLZ, XType, N0);
3737 return DAG.getNode(ISD::SRL, XType, Ctlz,
3738 DAG.getConstant(Log2_32(MVT::getSizeInBits(XType)),
3739 TLI.getShiftAmountTy()));
3741 // setgt X, 0 -> srl (and (-X, ~X), size(X)-1)
3742 if (N1C && N1C->isNullValue() && CC == ISD::SETGT) {
3743 SDOperand NegN0 = DAG.getNode(ISD::SUB, XType, DAG.getConstant(0, XType),
3745 SDOperand NotN0 = DAG.getNode(ISD::XOR, XType, N0,
3746 DAG.getConstant(~0ULL, XType));
3747 return DAG.getNode(ISD::SRL, XType,
3748 DAG.getNode(ISD::AND, XType, NegN0, NotN0),
3749 DAG.getConstant(MVT::getSizeInBits(XType)-1,
3750 TLI.getShiftAmountTy()));
3752 // setgt X, -1 -> xor (srl (X, size(X)-1), 1)
3753 if (N1C && N1C->isAllOnesValue() && CC == ISD::SETGT) {
3754 SDOperand Sign = DAG.getNode(ISD::SRL, XType, N0,
3755 DAG.getConstant(MVT::getSizeInBits(XType)-1,
3756 TLI.getShiftAmountTy()));
3757 return DAG.getNode(ISD::XOR, XType, Sign, DAG.getConstant(1, XType));
3761 // Check to see if this is an integer abs. select_cc setl[te] X, 0, -X, X ->
3762 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
3763 if (N1C && N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE) &&
3764 N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1)) {
3765 if (ConstantSDNode *SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0))) {
3766 MVT::ValueType XType = N0.getValueType();
3767 if (SubC->isNullValue() && MVT::isInteger(XType)) {
3768 SDOperand Shift = DAG.getNode(ISD::SRA, XType, N0,
3769 DAG.getConstant(MVT::getSizeInBits(XType)-1,
3770 TLI.getShiftAmountTy()));
3771 SDOperand Add = DAG.getNode(ISD::ADD, XType, N0, Shift);
3772 AddToWorkList(Shift.Val);
3773 AddToWorkList(Add.Val);
3774 return DAG.getNode(ISD::XOR, XType, Add, Shift);
3782 SDOperand DAGCombiner::SimplifySetCC(MVT::ValueType VT, SDOperand N0,
3783 SDOperand N1, ISD::CondCode Cond,
3784 bool foldBooleans) {
3785 // These setcc operations always fold.
3789 case ISD::SETFALSE2: return DAG.getConstant(0, VT);
3791 case ISD::SETTRUE2: return DAG.getConstant(1, VT);
3794 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val)) {
3795 uint64_t C1 = N1C->getValue();
3796 if (isa<ConstantSDNode>(N0.Val)) {
3797 return DAG.FoldSetCC(VT, N0, N1, Cond);
3799 // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an
3800 // equality comparison, then we're just comparing whether X itself is
3802 if (N0.getOpcode() == ISD::SRL && (C1 == 0 || C1 == 1) &&
3803 N0.getOperand(0).getOpcode() == ISD::CTLZ &&
3804 N0.getOperand(1).getOpcode() == ISD::Constant) {
3805 unsigned ShAmt = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
3806 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
3807 ShAmt == Log2_32(MVT::getSizeInBits(N0.getValueType()))) {
3808 if ((C1 == 0) == (Cond == ISD::SETEQ)) {
3809 // (srl (ctlz x), 5) == 0 -> X != 0
3810 // (srl (ctlz x), 5) != 1 -> X != 0
3813 // (srl (ctlz x), 5) != 0 -> X == 0
3814 // (srl (ctlz x), 5) == 1 -> X == 0
3817 SDOperand Zero = DAG.getConstant(0, N0.getValueType());
3818 return DAG.getSetCC(VT, N0.getOperand(0).getOperand(0),
3823 // If the LHS is a ZERO_EXTEND, perform the comparison on the input.
3824 if (N0.getOpcode() == ISD::ZERO_EXTEND) {
3825 unsigned InSize = MVT::getSizeInBits(N0.getOperand(0).getValueType());
3827 // If the comparison constant has bits in the upper part, the
3828 // zero-extended value could never match.
3829 if (C1 & (~0ULL << InSize)) {
3830 unsigned VSize = MVT::getSizeInBits(N0.getValueType());
3834 case ISD::SETEQ: return DAG.getConstant(0, VT);
3837 case ISD::SETNE: return DAG.getConstant(1, VT);
3840 // True if the sign bit of C1 is set.
3841 return DAG.getConstant((C1 & (1ULL << VSize)) != 0, VT);
3844 // True if the sign bit of C1 isn't set.
3845 return DAG.getConstant((C1 & (1ULL << VSize)) == 0, VT);
3851 // Otherwise, we can perform the comparison with the low bits.
3859 return DAG.getSetCC(VT, N0.getOperand(0),
3860 DAG.getConstant(C1, N0.getOperand(0).getValueType()),
3863 break; // todo, be more careful with signed comparisons
3865 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
3866 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
3867 MVT::ValueType ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT();
3868 unsigned ExtSrcTyBits = MVT::getSizeInBits(ExtSrcTy);
3869 MVT::ValueType ExtDstTy = N0.getValueType();
3870 unsigned ExtDstTyBits = MVT::getSizeInBits(ExtDstTy);
3872 // If the extended part has any inconsistent bits, it cannot ever
3873 // compare equal. In other words, they have to be all ones or all
3876 (~0ULL >> (64-ExtSrcTyBits)) & (~0ULL << (ExtDstTyBits-1));
3877 if ((C1 & ExtBits) != 0 && (C1 & ExtBits) != ExtBits)
3878 return DAG.getConstant(Cond == ISD::SETNE, VT);
3881 MVT::ValueType Op0Ty = N0.getOperand(0).getValueType();
3882 if (Op0Ty == ExtSrcTy) {
3883 ZextOp = N0.getOperand(0);
3885 int64_t Imm = ~0ULL >> (64-ExtSrcTyBits);
3886 ZextOp = DAG.getNode(ISD::AND, Op0Ty, N0.getOperand(0),
3887 DAG.getConstant(Imm, Op0Ty));
3889 AddToWorkList(ZextOp.Val);
3890 // Otherwise, make this a use of a zext.
3891 return DAG.getSetCC(VT, ZextOp,
3892 DAG.getConstant(C1 & (~0ULL>>(64-ExtSrcTyBits)),
3895 } else if ((N1C->getValue() == 0 || N1C->getValue() == 1) &&
3896 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
3898 // SETCC (SETCC), [0|1], [EQ|NE] -> SETCC
3899 if (N0.getOpcode() == ISD::SETCC) {
3900 bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (N1C->getValue() != 1);
3904 // Invert the condition.
3905 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
3906 CC = ISD::getSetCCInverse(CC,
3907 MVT::isInteger(N0.getOperand(0).getValueType()));
3908 return DAG.getSetCC(VT, N0.getOperand(0), N0.getOperand(1), CC);
3911 if ((N0.getOpcode() == ISD::XOR ||
3912 (N0.getOpcode() == ISD::AND &&
3913 N0.getOperand(0).getOpcode() == ISD::XOR &&
3914 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) &&
3915 isa<ConstantSDNode>(N0.getOperand(1)) &&
3916 cast<ConstantSDNode>(N0.getOperand(1))->getValue() == 1) {
3917 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We
3918 // can only do this if the top bits are known zero.
3919 if (TLI.MaskedValueIsZero(N0,
3920 MVT::getIntVTBitMask(N0.getValueType())-1)){
3921 // Okay, get the un-inverted input value.
3923 if (N0.getOpcode() == ISD::XOR)
3924 Val = N0.getOperand(0);
3926 assert(N0.getOpcode() == ISD::AND &&
3927 N0.getOperand(0).getOpcode() == ISD::XOR);
3928 // ((X^1)&1)^1 -> X & 1
3929 Val = DAG.getNode(ISD::AND, N0.getValueType(),
3930 N0.getOperand(0).getOperand(0),
3933 return DAG.getSetCC(VT, Val, N1,
3934 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
3939 uint64_t MinVal, MaxVal;
3940 unsigned OperandBitSize = MVT::getSizeInBits(N1C->getValueType(0));
3941 if (ISD::isSignedIntSetCC(Cond)) {
3942 MinVal = 1ULL << (OperandBitSize-1);
3943 if (OperandBitSize != 1) // Avoid X >> 64, which is undefined.
3944 MaxVal = ~0ULL >> (65-OperandBitSize);
3949 MaxVal = ~0ULL >> (64-OperandBitSize);
3952 // Canonicalize GE/LE comparisons to use GT/LT comparisons.
3953 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) {
3954 if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true
3955 --C1; // X >= C0 --> X > (C0-1)
3956 return DAG.getSetCC(VT, N0, DAG.getConstant(C1, N1.getValueType()),
3957 (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT);
3960 if (Cond == ISD::SETLE || Cond == ISD::SETULE) {
3961 if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true
3962 ++C1; // X <= C0 --> X < (C0+1)
3963 return DAG.getSetCC(VT, N0, DAG.getConstant(C1, N1.getValueType()),
3964 (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT);
3967 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal)
3968 return DAG.getConstant(0, VT); // X < MIN --> false
3970 // Canonicalize setgt X, Min --> setne X, Min
3971 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal)
3972 return DAG.getSetCC(VT, N0, N1, ISD::SETNE);
3973 // Canonicalize setlt X, Max --> setne X, Max
3974 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal)
3975 return DAG.getSetCC(VT, N0, N1, ISD::SETNE);
3977 // If we have setult X, 1, turn it into seteq X, 0
3978 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1)
3979 return DAG.getSetCC(VT, N0, DAG.getConstant(MinVal, N0.getValueType()),
3981 // If we have setugt X, Max-1, turn it into seteq X, Max
3982 else if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1)
3983 return DAG.getSetCC(VT, N0, DAG.getConstant(MaxVal, N0.getValueType()),
3986 // If we have "setcc X, C0", check to see if we can shrink the immediate
3989 // SETUGT X, SINTMAX -> SETLT X, 0
3990 if (Cond == ISD::SETUGT && OperandBitSize != 1 &&
3991 C1 == (~0ULL >> (65-OperandBitSize)))
3992 return DAG.getSetCC(VT, N0, DAG.getConstant(0, N1.getValueType()),
3995 // FIXME: Implement the rest of these.
3997 // Fold bit comparisons when we can.
3998 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
3999 VT == N0.getValueType() && N0.getOpcode() == ISD::AND)
4000 if (ConstantSDNode *AndRHS =
4001 dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
4002 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3
4003 // Perform the xform if the AND RHS is a single bit.
4004 if (isPowerOf2_64(AndRHS->getValue())) {
4005 return DAG.getNode(ISD::SRL, VT, N0,
4006 DAG.getConstant(Log2_64(AndRHS->getValue()),
4007 TLI.getShiftAmountTy()));
4009 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getValue()) {
4010 // (X & 8) == 8 --> (X & 8) >> 3
4011 // Perform the xform if C1 is a single bit.
4012 if (isPowerOf2_64(C1)) {
4013 return DAG.getNode(ISD::SRL, VT, N0,
4014 DAG.getConstant(Log2_64(C1),TLI.getShiftAmountTy()));
4019 } else if (isa<ConstantSDNode>(N0.Val)) {
4020 // Ensure that the constant occurs on the RHS.
4021 return DAG.getSetCC(VT, N1, N0, ISD::getSetCCSwappedOperands(Cond));
4024 if (isa<ConstantFPSDNode>(N0.Val)) {
4025 // Constant fold or commute setcc.
4026 SDOperand O = DAG.FoldSetCC(VT, N0, N1, Cond);
4027 if (O.Val) return O;
4031 // We can always fold X == X for integer setcc's.
4032 if (MVT::isInteger(N0.getValueType()))
4033 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
4034 unsigned UOF = ISD::getUnorderedFlavor(Cond);
4035 if (UOF == 2) // FP operators that are undefined on NaNs.
4036 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
4037 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond)))
4038 return DAG.getConstant(UOF, VT);
4039 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO
4040 // if it is not already.
4041 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO;
4042 if (NewCond != Cond)
4043 return DAG.getSetCC(VT, N0, N1, NewCond);
4046 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
4047 MVT::isInteger(N0.getValueType())) {
4048 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB ||
4049 N0.getOpcode() == ISD::XOR) {
4050 // Simplify (X+Y) == (X+Z) --> Y == Z
4051 if (N0.getOpcode() == N1.getOpcode()) {
4052 if (N0.getOperand(0) == N1.getOperand(0))
4053 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(1), Cond);
4054 if (N0.getOperand(1) == N1.getOperand(1))
4055 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(0), Cond);
4056 if (DAG.isCommutativeBinOp(N0.getOpcode())) {
4057 // If X op Y == Y op X, try other combinations.
4058 if (N0.getOperand(0) == N1.getOperand(1))
4059 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(0), Cond);
4060 if (N0.getOperand(1) == N1.getOperand(0))
4061 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(1), Cond);
4065 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) {
4066 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
4067 // Turn (X+C1) == C2 --> X == C2-C1
4068 if (N0.getOpcode() == ISD::ADD && N0.Val->hasOneUse()) {
4069 return DAG.getSetCC(VT, N0.getOperand(0),
4070 DAG.getConstant(RHSC->getValue()-LHSR->getValue(),
4071 N0.getValueType()), Cond);
4074 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0.
4075 if (N0.getOpcode() == ISD::XOR)
4076 // If we know that all of the inverted bits are zero, don't bother
4077 // performing the inversion.
4078 if (TLI.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getValue()))
4079 return DAG.getSetCC(VT, N0.getOperand(0),
4080 DAG.getConstant(LHSR->getValue()^RHSC->getValue(),
4081 N0.getValueType()), Cond);
4084 // Turn (C1-X) == C2 --> X == C1-C2
4085 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) {
4086 if (N0.getOpcode() == ISD::SUB && N0.Val->hasOneUse()) {
4087 return DAG.getSetCC(VT, N0.getOperand(1),
4088 DAG.getConstant(SUBC->getValue()-RHSC->getValue(),
4089 N0.getValueType()), Cond);
4094 // Simplify (X+Z) == X --> Z == 0
4095 if (N0.getOperand(0) == N1)
4096 return DAG.getSetCC(VT, N0.getOperand(1),
4097 DAG.getConstant(0, N0.getValueType()), Cond);
4098 if (N0.getOperand(1) == N1) {
4099 if (DAG.isCommutativeBinOp(N0.getOpcode()))
4100 return DAG.getSetCC(VT, N0.getOperand(0),
4101 DAG.getConstant(0, N0.getValueType()), Cond);
4103 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!");
4104 // (Z-X) == X --> Z == X<<1
4105 SDOperand SH = DAG.getNode(ISD::SHL, N1.getValueType(),
4107 DAG.getConstant(1,TLI.getShiftAmountTy()));
4108 AddToWorkList(SH.Val);
4109 return DAG.getSetCC(VT, N0.getOperand(0), SH, Cond);
4114 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB ||
4115 N1.getOpcode() == ISD::XOR) {
4116 // Simplify X == (X+Z) --> Z == 0
4117 if (N1.getOperand(0) == N0) {
4118 return DAG.getSetCC(VT, N1.getOperand(1),
4119 DAG.getConstant(0, N1.getValueType()), Cond);
4120 } else if (N1.getOperand(1) == N0) {
4121 if (DAG.isCommutativeBinOp(N1.getOpcode())) {
4122 return DAG.getSetCC(VT, N1.getOperand(0),
4123 DAG.getConstant(0, N1.getValueType()), Cond);
4125 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!");
4126 // X == (Z-X) --> X<<1 == Z
4127 SDOperand SH = DAG.getNode(ISD::SHL, N1.getValueType(), N0,
4128 DAG.getConstant(1,TLI.getShiftAmountTy()));
4129 AddToWorkList(SH.Val);
4130 return DAG.getSetCC(VT, SH, N1.getOperand(0), Cond);
4136 // Fold away ALL boolean setcc's.
4138 if (N0.getValueType() == MVT::i1 && foldBooleans) {
4140 default: assert(0 && "Unknown integer setcc!");
4141 case ISD::SETEQ: // X == Y -> (X^Y)^1
4142 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, N1);
4143 N0 = DAG.getNode(ISD::XOR, MVT::i1, Temp, DAG.getConstant(1, MVT::i1));
4144 AddToWorkList(Temp.Val);
4146 case ISD::SETNE: // X != Y --> (X^Y)
4147 N0 = DAG.getNode(ISD::XOR, MVT::i1, N0, N1);
4149 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> X^1 & Y
4150 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> X^1 & Y
4151 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1));
4152 N0 = DAG.getNode(ISD::AND, MVT::i1, N1, Temp);
4153 AddToWorkList(Temp.Val);
4155 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> Y^1 & X
4156 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> Y^1 & X
4157 Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1));
4158 N0 = DAG.getNode(ISD::AND, MVT::i1, N0, Temp);
4159 AddToWorkList(Temp.Val);
4161 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> X^1 | Y
4162 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> X^1 | Y
4163 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1));
4164 N0 = DAG.getNode(ISD::OR, MVT::i1, N1, Temp);
4165 AddToWorkList(Temp.Val);
4167 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> Y^1 | X
4168 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> Y^1 | X
4169 Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1));
4170 N0 = DAG.getNode(ISD::OR, MVT::i1, N0, Temp);
4173 if (VT != MVT::i1) {
4174 AddToWorkList(N0.Val);
4175 // FIXME: If running after legalize, we probably can't do this.
4176 N0 = DAG.getNode(ISD::ZERO_EXTEND, VT, N0);
4181 // Could not fold it.
4185 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
4186 /// return a DAG expression to select that will generate the same value by
4187 /// multiplying by a magic number. See:
4188 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
4189 SDOperand DAGCombiner::BuildSDIV(SDNode *N) {
4190 std::vector<SDNode*> Built;
4191 SDOperand S = TLI.BuildSDIV(N, DAG, &Built);
4193 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end();
4199 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
4200 /// return a DAG expression to select that will generate the same value by
4201 /// multiplying by a magic number. See:
4202 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
4203 SDOperand DAGCombiner::BuildUDIV(SDNode *N) {
4204 std::vector<SDNode*> Built;
4205 SDOperand S = TLI.BuildUDIV(N, DAG, &Built);
4207 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end();
4213 /// FindBaseOffset - Return true if base is known not to alias with anything
4214 /// but itself. Provides base object and offset as results.
4215 static bool FindBaseOffset(SDOperand Ptr, SDOperand &Base, int64_t &Offset) {
4216 // Assume it is a primitive operation.
4217 Base = Ptr; Offset = 0;
4219 // If it's an adding a simple constant then integrate the offset.
4220 if (Base.getOpcode() == ISD::ADD) {
4221 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Base.getOperand(1))) {
4222 Base = Base.getOperand(0);
4223 Offset += C->getValue();
4227 // If it's any of the following then it can't alias with anything but itself.
4228 return isa<FrameIndexSDNode>(Base) ||
4229 isa<ConstantPoolSDNode>(Base) ||
4230 isa<GlobalAddressSDNode>(Base);
4233 /// isAlias - Return true if there is any possibility that the two addresses
4235 bool DAGCombiner::isAlias(SDOperand Ptr1, int64_t Size1,
4236 const Value *SrcValue1, int SrcValueOffset1,
4237 SDOperand Ptr2, int64_t Size2,
4238 const Value *SrcValue2, int SrcValueOffset2)
4240 // If they are the same then they must be aliases.
4241 if (Ptr1 == Ptr2) return true;
4243 // Gather base node and offset information.
4244 SDOperand Base1, Base2;
4245 int64_t Offset1, Offset2;
4246 bool KnownBase1 = FindBaseOffset(Ptr1, Base1, Offset1);
4247 bool KnownBase2 = FindBaseOffset(Ptr2, Base2, Offset2);
4249 // If they have a same base address then...
4250 if (Base1 == Base2) {
4251 // Check to see if the addresses overlap.
4252 return!((Offset1 + Size1) <= Offset2 || (Offset2 + Size2) <= Offset1);
4255 // If we know both bases then they can't alias.
4256 if (KnownBase1 && KnownBase2) return false;
4258 if (CombinerGlobalAA) {
4259 // Use alias analysis information.
4260 int Overlap1 = Size1 + SrcValueOffset1 + Offset1;
4261 int Overlap2 = Size2 + SrcValueOffset2 + Offset2;
4262 AliasAnalysis::AliasResult AAResult =
4263 AA.alias(SrcValue1, Overlap1, SrcValue2, Overlap2);
4264 if (AAResult == AliasAnalysis::NoAlias)
4268 // Otherwise we have to assume they alias.
4272 /// FindAliasInfo - Extracts the relevant alias information from the memory
4273 /// node. Returns true if the operand was a load.
4274 bool DAGCombiner::FindAliasInfo(SDNode *N,
4275 SDOperand &Ptr, int64_t &Size,
4276 const Value *&SrcValue, int &SrcValueOffset) {
4277 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4278 Ptr = LD->getBasePtr();
4279 Size = MVT::getSizeInBits(LD->getLoadedVT()) >> 3;
4280 SrcValue = LD->getSrcValue();
4281 SrcValueOffset = LD->getSrcValueOffset();
4283 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4284 Ptr = ST->getBasePtr();
4285 Size = MVT::getSizeInBits(ST->getStoredVT()) >> 3;
4286 SrcValue = ST->getSrcValue();
4287 SrcValueOffset = ST->getSrcValueOffset();
4289 assert(0 && "FindAliasInfo expected a memory operand");
4295 /// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes,
4296 /// looking for aliasing nodes and adding them to the Aliases vector.
4297 void DAGCombiner::GatherAllAliases(SDNode *N, SDOperand OriginalChain,
4298 SmallVector<SDOperand, 8> &Aliases) {
4299 SmallVector<SDOperand, 8> Chains; // List of chains to visit.
4300 std::set<SDNode *> Visited; // Visited node set.
4302 // Get alias information for node.
4305 const Value *SrcValue;
4307 bool IsLoad = FindAliasInfo(N, Ptr, Size, SrcValue, SrcValueOffset);
4310 Chains.push_back(OriginalChain);
4312 // Look at each chain and determine if it is an alias. If so, add it to the
4313 // aliases list. If not, then continue up the chain looking for the next
4315 while (!Chains.empty()) {
4316 SDOperand Chain = Chains.back();
4319 // Don't bother if we've been before.
4320 if (Visited.find(Chain.Val) != Visited.end()) continue;
4321 Visited.insert(Chain.Val);
4323 switch (Chain.getOpcode()) {
4324 case ISD::EntryToken:
4325 // Entry token is ideal chain operand, but handled in FindBetterChain.
4330 // Get alias information for Chain.
4333 const Value *OpSrcValue;
4334 int OpSrcValueOffset;
4335 bool IsOpLoad = FindAliasInfo(Chain.Val, OpPtr, OpSize,
4336 OpSrcValue, OpSrcValueOffset);
4338 // If chain is alias then stop here.
4339 if (!(IsLoad && IsOpLoad) &&
4340 isAlias(Ptr, Size, SrcValue, SrcValueOffset,
4341 OpPtr, OpSize, OpSrcValue, OpSrcValueOffset)) {
4342 Aliases.push_back(Chain);
4344 // Look further up the chain.
4345 Chains.push_back(Chain.getOperand(0));
4346 // Clean up old chain.
4347 AddToWorkList(Chain.Val);
4352 case ISD::TokenFactor:
4353 // We have to check each of the operands of the token factor, so we queue
4354 // then up. Adding the operands to the queue (stack) in reverse order
4355 // maintains the original order and increases the likelihood that getNode
4356 // will find a matching token factor (CSE.)
4357 for (unsigned n = Chain.getNumOperands(); n;)
4358 Chains.push_back(Chain.getOperand(--n));
4359 // Eliminate the token factor if we can.
4360 AddToWorkList(Chain.Val);
4364 // For all other instructions we will just have to take what we can get.
4365 Aliases.push_back(Chain);
4371 /// FindBetterChain - Walk up chain skipping non-aliasing memory nodes, looking
4372 /// for a better chain (aliasing node.)
4373 SDOperand DAGCombiner::FindBetterChain(SDNode *N, SDOperand OldChain) {
4374 SmallVector<SDOperand, 8> Aliases; // Ops for replacing token factor.
4376 // Accumulate all the aliases to this node.
4377 GatherAllAliases(N, OldChain, Aliases);
4379 if (Aliases.size() == 0) {
4380 // If no operands then chain to entry token.
4381 return DAG.getEntryNode();
4382 } else if (Aliases.size() == 1) {
4383 // If a single operand then chain to it. We don't need to revisit it.
4387 // Construct a custom tailored token factor.
4388 SDOperand NewChain = DAG.getNode(ISD::TokenFactor, MVT::Other,
4389 &Aliases[0], Aliases.size());
4391 // Make sure the old chain gets cleaned up.
4392 if (NewChain != OldChain) AddToWorkList(OldChain.Val);
4397 // SelectionDAG::Combine - This is the entry point for the file.
4399 void SelectionDAG::Combine(bool RunningAfterLegalize, AliasAnalysis &AA) {
4400 /// run - This is the main entry point to this class.
4402 DAGCombiner(*this, AA).Run(RunningAfterLegalize);