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"
47 static Statistic<> NodesCombined ("dagcombiner",
48 "Number of dag nodes combined");
51 CombinerAA("combiner-alias-analysis", cl::Hidden,
52 cl::desc("Turn on alias analysis during testing"));
55 CombinerGlobalAA("combiner-global-alias-analysis", cl::Hidden,
56 cl::desc("Include global information in alias analysis"));
58 //------------------------------ DAGCombiner ---------------------------------//
60 class VISIBILITY_HIDDEN DAGCombiner {
65 // Worklist of all of the nodes that need to be simplified.
66 std::vector<SDNode*> WorkList;
68 // AA - Used for DAG load/store alias analysis.
71 /// AddUsersToWorkList - When an instruction is simplified, add all users of
72 /// the instruction to the work lists because they might get more simplified
75 void AddUsersToWorkList(SDNode *N) {
76 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
81 /// removeFromWorkList - remove all instances of N from the worklist.
83 void removeFromWorkList(SDNode *N) {
84 WorkList.erase(std::remove(WorkList.begin(), WorkList.end(), N),
89 /// AddToWorkList - Add to the work list making sure it's instance is at the
90 /// the back (next to be processed.)
91 void AddToWorkList(SDNode *N) {
92 removeFromWorkList(N);
93 WorkList.push_back(N);
96 SDOperand CombineTo(SDNode *N, const SDOperand *To, unsigned NumTo,
98 assert(N->getNumValues() == NumTo && "Broken CombineTo call!");
100 DEBUG(std::cerr << "\nReplacing.1 "; N->dump();
101 std::cerr << "\nWith: "; To[0].Val->dump(&DAG);
102 std::cerr << " and " << NumTo-1 << " other values\n");
103 std::vector<SDNode*> NowDead;
104 DAG.ReplaceAllUsesWith(N, To, &NowDead);
107 // Push the new nodes and any users onto the worklist
108 for (unsigned i = 0, e = NumTo; i != e; ++i) {
109 AddToWorkList(To[i].Val);
110 AddUsersToWorkList(To[i].Val);
114 // Nodes can be reintroduced into the worklist. Make sure we do not
115 // process a node that has been replaced.
116 removeFromWorkList(N);
117 for (unsigned i = 0, e = NowDead.size(); i != e; ++i)
118 removeFromWorkList(NowDead[i]);
120 // Finally, since the node is now dead, remove it from the graph.
122 return SDOperand(N, 0);
125 SDOperand CombineTo(SDNode *N, SDOperand Res, bool AddTo = true) {
126 return CombineTo(N, &Res, 1, AddTo);
129 SDOperand CombineTo(SDNode *N, SDOperand Res0, SDOperand Res1,
131 SDOperand To[] = { Res0, Res1 };
132 return CombineTo(N, To, 2, AddTo);
136 /// SimplifyDemandedBits - Check the specified integer node value to see if
137 /// it can be simplified or if things it uses can be simplified by bit
138 /// propagation. If so, return true.
139 bool SimplifyDemandedBits(SDOperand Op) {
140 TargetLowering::TargetLoweringOpt TLO(DAG);
141 uint64_t KnownZero, KnownOne;
142 uint64_t Demanded = MVT::getIntVTBitMask(Op.getValueType());
143 if (!TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
147 AddToWorkList(Op.Val);
149 // Replace the old value with the new one.
151 DEBUG(std::cerr << "\nReplacing.2 "; TLO.Old.Val->dump();
152 std::cerr << "\nWith: "; TLO.New.Val->dump(&DAG);
155 std::vector<SDNode*> NowDead;
156 DAG.ReplaceAllUsesOfValueWith(TLO.Old, TLO.New, NowDead);
158 // Push the new node and any (possibly new) users onto the worklist.
159 AddToWorkList(TLO.New.Val);
160 AddUsersToWorkList(TLO.New.Val);
162 // Nodes can end up on the worklist more than once. Make sure we do
163 // not process a node that has been replaced.
164 for (unsigned i = 0, e = NowDead.size(); i != e; ++i)
165 removeFromWorkList(NowDead[i]);
167 // Finally, if the node is now dead, remove it from the graph. The node
168 // may not be dead if the replacement process recursively simplified to
169 // something else needing this node.
170 if (TLO.Old.Val->use_empty()) {
171 removeFromWorkList(TLO.Old.Val);
172 DAG.DeleteNode(TLO.Old.Val);
177 /// CombineToPreIndexedLoadStore - Try turning a load / store into a
178 /// pre-indexed load store when the base pointer is a add or subtract
179 /// and it has other uses besides the load / store. When the
180 /// transformation is done, the new indexed load / store effectively
181 /// folded the add / subtract in and all of its other uses are redirected
182 /// to the new load / store.
183 bool CombineToPreIndexedLoadStore(SDNode *N) {
186 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
187 Ptr = LD->getBasePtr();
188 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
189 Ptr = ST->getBasePtr();
195 (Ptr.getOpcode() == ISD::ADD || Ptr.getOpcode() == ISD::SUB) &&
196 Ptr.Val->use_size() > 1) {
199 ISD::MemOpAddrMode AM = ISD::UNINDEXED;
200 if (TLI.getPreIndexedAddressParts(N, BasePtr, Offset, AM, DAG)) {
201 // Try turning it into a pre-indexed load / store except when
202 // 1) Another use of base ptr is a predecessor of N. If ptr is folded
203 // that would create a cycle.
204 // 2) All uses are load / store ops that use it as base ptr and offset
205 // is just an addressing mode immediate.
206 // 3) If the would-be new base may not to be dead at N. FIXME: The
207 // proper check is too expensive (in turns of compile time) to
208 // check. Just make sure other uses of the new base are not also
209 // themselves use of loads / stores.
211 bool OffIsAMImm = Offset.getOpcode() == ISD::Constant &&
212 TLI.isLegalAddressImmediate(
213 cast<ConstantSDNode>(Offset)->getValue());
216 if (OffIsAMImm && BasePtr.Val->use_size() > 1) {
217 for (SDNode::use_iterator I = BasePtr.Val->use_begin(),
218 E = BasePtr.Val->use_end(); I != E; ++I) {
222 if (Use->getOpcode() == ISD::ADD || Use->getOpcode() == ISD::SUB){
223 for (SDNode::use_iterator II = Use->use_begin(),
224 EE = Use->use_end(); II != EE; ++II) {
225 SDNode *UseUse = *II;
226 if (UseUse->getOpcode() == ISD::LOAD &&
227 cast<LoadSDNode>(UseUse)->getBasePtr().Val == Use)
229 else if (UseUse->getOpcode() == ISD::STORE &&
230 cast<StoreSDNode>(UseUse)->getBasePtr().Val == Use)
237 // Now check for #1 and #2.
238 unsigned NumRealUses = 0;
239 for (SDNode::use_iterator I = Ptr.Val->use_begin(),
240 E = Ptr.Val->use_end(); I != E; ++I) {
244 if (Use->isPredecessor(N))
249 } else if (Use->getOpcode() == ISD::LOAD) {
250 if (cast<LoadSDNode>(Use)->getBasePtr().Val != Ptr.Val)
252 } else if (Use->getOpcode() == ISD::STORE) {
253 if (cast<StoreSDNode>(Use)->getBasePtr().Val != Ptr.Val)
258 if (NumRealUses == 0)
261 SDOperand Result = isLoad
262 ? DAG.getIndexedLoad(SDOperand(N,0), BasePtr, Offset, AM)
263 : DAG.getIndexedStore(SDOperand(N,0), BasePtr, Offset, AM);
265 DEBUG(std::cerr << "\nReplacing.4 "; N->dump();
266 std::cerr << "\nWith: "; Result.Val->dump(&DAG);
268 std::vector<SDNode*> NowDead;
270 DAG.ReplaceAllUsesOfValueWith(SDOperand(N, 0), Result.getValue(0),
272 DAG.ReplaceAllUsesOfValueWith(SDOperand(N, 1), Result.getValue(2),
275 DAG.ReplaceAllUsesOfValueWith(SDOperand(N, 0), Result.getValue(1),
279 // Nodes can end up on the worklist more than once. Make sure we do
280 // not process a node that has been replaced.
281 for (unsigned i = 0, e = NowDead.size(); i != e; ++i)
282 removeFromWorkList(NowDead[i]);
283 // Finally, since the node is now dead, remove it from the graph.
286 // Replace the uses of Ptr with uses of the updated base value.
287 DAG.ReplaceAllUsesOfValueWith(Ptr, Result.getValue(isLoad ? 1 : 0),
289 removeFromWorkList(Ptr.Val);
290 for (unsigned i = 0, e = NowDead.size(); i != e; ++i)
291 removeFromWorkList(NowDead[i]);
292 DAG.DeleteNode(Ptr.Val);
301 /// visit - call the node-specific routine that knows how to fold each
302 /// particular type of node.
303 SDOperand visit(SDNode *N);
305 // Visitation implementation - Implement dag node combining for different
306 // node types. The semantics are as follows:
308 // SDOperand.Val == 0 - No change was made
309 // SDOperand.Val == N - N was replaced, is dead, and is already handled.
310 // otherwise - N should be replaced by the returned Operand.
312 SDOperand visitTokenFactor(SDNode *N);
313 SDOperand visitADD(SDNode *N);
314 SDOperand visitSUB(SDNode *N);
315 SDOperand visitMUL(SDNode *N);
316 SDOperand visitSDIV(SDNode *N);
317 SDOperand visitUDIV(SDNode *N);
318 SDOperand visitSREM(SDNode *N);
319 SDOperand visitUREM(SDNode *N);
320 SDOperand visitMULHU(SDNode *N);
321 SDOperand visitMULHS(SDNode *N);
322 SDOperand visitAND(SDNode *N);
323 SDOperand visitOR(SDNode *N);
324 SDOperand visitXOR(SDNode *N);
325 SDOperand visitVBinOp(SDNode *N, ISD::NodeType IntOp, ISD::NodeType FPOp);
326 SDOperand visitSHL(SDNode *N);
327 SDOperand visitSRA(SDNode *N);
328 SDOperand visitSRL(SDNode *N);
329 SDOperand visitCTLZ(SDNode *N);
330 SDOperand visitCTTZ(SDNode *N);
331 SDOperand visitCTPOP(SDNode *N);
332 SDOperand visitSELECT(SDNode *N);
333 SDOperand visitSELECT_CC(SDNode *N);
334 SDOperand visitSETCC(SDNode *N);
335 SDOperand visitSIGN_EXTEND(SDNode *N);
336 SDOperand visitZERO_EXTEND(SDNode *N);
337 SDOperand visitANY_EXTEND(SDNode *N);
338 SDOperand visitSIGN_EXTEND_INREG(SDNode *N);
339 SDOperand visitTRUNCATE(SDNode *N);
340 SDOperand visitBIT_CONVERT(SDNode *N);
341 SDOperand visitVBIT_CONVERT(SDNode *N);
342 SDOperand visitFADD(SDNode *N);
343 SDOperand visitFSUB(SDNode *N);
344 SDOperand visitFMUL(SDNode *N);
345 SDOperand visitFDIV(SDNode *N);
346 SDOperand visitFREM(SDNode *N);
347 SDOperand visitFCOPYSIGN(SDNode *N);
348 SDOperand visitSINT_TO_FP(SDNode *N);
349 SDOperand visitUINT_TO_FP(SDNode *N);
350 SDOperand visitFP_TO_SINT(SDNode *N);
351 SDOperand visitFP_TO_UINT(SDNode *N);
352 SDOperand visitFP_ROUND(SDNode *N);
353 SDOperand visitFP_ROUND_INREG(SDNode *N);
354 SDOperand visitFP_EXTEND(SDNode *N);
355 SDOperand visitFNEG(SDNode *N);
356 SDOperand visitFABS(SDNode *N);
357 SDOperand visitBRCOND(SDNode *N);
358 SDOperand visitBR_CC(SDNode *N);
359 SDOperand visitLOAD(SDNode *N);
360 SDOperand visitSTORE(SDNode *N);
361 SDOperand visitINSERT_VECTOR_ELT(SDNode *N);
362 SDOperand visitVINSERT_VECTOR_ELT(SDNode *N);
363 SDOperand visitVBUILD_VECTOR(SDNode *N);
364 SDOperand visitVECTOR_SHUFFLE(SDNode *N);
365 SDOperand visitVVECTOR_SHUFFLE(SDNode *N);
367 SDOperand XformToShuffleWithZero(SDNode *N);
368 SDOperand ReassociateOps(unsigned Opc, SDOperand LHS, SDOperand RHS);
370 bool SimplifySelectOps(SDNode *SELECT, SDOperand LHS, SDOperand RHS);
371 SDOperand SimplifyBinOpWithSameOpcodeHands(SDNode *N);
372 SDOperand SimplifySelect(SDOperand N0, SDOperand N1, SDOperand N2);
373 SDOperand SimplifySelectCC(SDOperand N0, SDOperand N1, SDOperand N2,
374 SDOperand N3, ISD::CondCode CC);
375 SDOperand SimplifySetCC(MVT::ValueType VT, SDOperand N0, SDOperand N1,
376 ISD::CondCode Cond, bool foldBooleans = true);
377 SDOperand ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(SDNode *, MVT::ValueType);
378 SDOperand BuildSDIV(SDNode *N);
379 SDOperand BuildUDIV(SDNode *N);
380 SDNode *MatchRotate(SDOperand LHS, SDOperand RHS);
382 /// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes,
383 /// looking for aliasing nodes and adding them to the Aliases vector.
384 void GatherAllAliases(SDNode *N, SDOperand OriginalChain,
385 SmallVector<SDOperand, 8> &Aliases);
387 /// isAlias - Return true if there is any possibility that the two addresses
389 bool isAlias(SDOperand Ptr1, int64_t Size1,
390 const Value *SrcValue1, int SrcValueOffset1,
391 SDOperand Ptr2, int64_t Size2,
392 const Value *SrcValue2, int SrcValueOffset2);
394 /// FindAliasInfo - Extracts the relevant alias information from the memory
395 /// node. Returns true if the operand was a load.
396 bool FindAliasInfo(SDNode *N,
397 SDOperand &Ptr, int64_t &Size,
398 const Value *&SrcValue, int &SrcValueOffset);
400 /// FindBetterChain - Walk up chain skipping non-aliasing memory nodes,
401 /// looking for a better chain (aliasing node.)
402 SDOperand FindBetterChain(SDNode *N, SDOperand Chain);
405 DAGCombiner(SelectionDAG &D, AliasAnalysis &A)
407 TLI(D.getTargetLoweringInfo()),
408 AfterLegalize(false),
411 /// Run - runs the dag combiner on all nodes in the work list
412 void Run(bool RunningAfterLegalize);
416 //===----------------------------------------------------------------------===//
417 // TargetLowering::DAGCombinerInfo implementation
418 //===----------------------------------------------------------------------===//
420 void TargetLowering::DAGCombinerInfo::AddToWorklist(SDNode *N) {
421 ((DAGCombiner*)DC)->AddToWorkList(N);
424 SDOperand TargetLowering::DAGCombinerInfo::
425 CombineTo(SDNode *N, const std::vector<SDOperand> &To) {
426 return ((DAGCombiner*)DC)->CombineTo(N, &To[0], To.size());
429 SDOperand TargetLowering::DAGCombinerInfo::
430 CombineTo(SDNode *N, SDOperand Res) {
431 return ((DAGCombiner*)DC)->CombineTo(N, Res);
435 SDOperand TargetLowering::DAGCombinerInfo::
436 CombineTo(SDNode *N, SDOperand Res0, SDOperand Res1) {
437 return ((DAGCombiner*)DC)->CombineTo(N, Res0, Res1);
443 //===----------------------------------------------------------------------===//
446 // isSetCCEquivalent - Return true if this node is a setcc, or is a select_cc
447 // that selects between the values 1 and 0, making it equivalent to a setcc.
448 // Also, set the incoming LHS, RHS, and CC references to the appropriate
449 // nodes based on the type of node we are checking. This simplifies life a
450 // bit for the callers.
451 static bool isSetCCEquivalent(SDOperand N, SDOperand &LHS, SDOperand &RHS,
453 if (N.getOpcode() == ISD::SETCC) {
454 LHS = N.getOperand(0);
455 RHS = N.getOperand(1);
456 CC = N.getOperand(2);
459 if (N.getOpcode() == ISD::SELECT_CC &&
460 N.getOperand(2).getOpcode() == ISD::Constant &&
461 N.getOperand(3).getOpcode() == ISD::Constant &&
462 cast<ConstantSDNode>(N.getOperand(2))->getValue() == 1 &&
463 cast<ConstantSDNode>(N.getOperand(3))->isNullValue()) {
464 LHS = N.getOperand(0);
465 RHS = N.getOperand(1);
466 CC = N.getOperand(4);
472 // isOneUseSetCC - Return true if this is a SetCC-equivalent operation with only
473 // one use. If this is true, it allows the users to invert the operation for
474 // free when it is profitable to do so.
475 static bool isOneUseSetCC(SDOperand N) {
476 SDOperand N0, N1, N2;
477 if (isSetCCEquivalent(N, N0, N1, N2) && N.Val->hasOneUse())
482 SDOperand DAGCombiner::ReassociateOps(unsigned Opc, SDOperand N0, SDOperand N1){
483 MVT::ValueType VT = N0.getValueType();
484 // reassoc. (op (op x, c1), y) -> (op (op x, y), c1) iff x+c1 has one use
485 // reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2))
486 if (N0.getOpcode() == Opc && isa<ConstantSDNode>(N0.getOperand(1))) {
487 if (isa<ConstantSDNode>(N1)) {
488 SDOperand OpNode = DAG.getNode(Opc, VT, N0.getOperand(1), N1);
489 AddToWorkList(OpNode.Val);
490 return DAG.getNode(Opc, VT, OpNode, N0.getOperand(0));
491 } else if (N0.hasOneUse()) {
492 SDOperand OpNode = DAG.getNode(Opc, VT, N0.getOperand(0), N1);
493 AddToWorkList(OpNode.Val);
494 return DAG.getNode(Opc, VT, OpNode, N0.getOperand(1));
497 // reassoc. (op y, (op x, c1)) -> (op (op x, y), c1) iff x+c1 has one use
498 // reassoc. (op c2, (op x, c1)) -> (op x, (op c1, c2))
499 if (N1.getOpcode() == Opc && isa<ConstantSDNode>(N1.getOperand(1))) {
500 if (isa<ConstantSDNode>(N0)) {
501 SDOperand OpNode = DAG.getNode(Opc, VT, N1.getOperand(1), N0);
502 AddToWorkList(OpNode.Val);
503 return DAG.getNode(Opc, VT, OpNode, N1.getOperand(0));
504 } else if (N1.hasOneUse()) {
505 SDOperand OpNode = DAG.getNode(Opc, VT, N1.getOperand(0), N0);
506 AddToWorkList(OpNode.Val);
507 return DAG.getNode(Opc, VT, OpNode, N1.getOperand(1));
513 void DAGCombiner::Run(bool RunningAfterLegalize) {
514 // set the instance variable, so that the various visit routines may use it.
515 AfterLegalize = RunningAfterLegalize;
517 // Add all the dag nodes to the worklist.
518 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
519 E = DAG.allnodes_end(); I != E; ++I)
520 WorkList.push_back(I);
522 // Create a dummy node (which is not added to allnodes), that adds a reference
523 // to the root node, preventing it from being deleted, and tracking any
524 // changes of the root.
525 HandleSDNode Dummy(DAG.getRoot());
527 // The root of the dag may dangle to deleted nodes until the dag combiner is
528 // done. Set it to null to avoid confusion.
529 DAG.setRoot(SDOperand());
531 /// DagCombineInfo - Expose the DAG combiner to the target combiner impls.
532 TargetLowering::DAGCombinerInfo
533 DagCombineInfo(DAG, !RunningAfterLegalize, this);
535 // while the worklist isn't empty, inspect the node on the end of it and
536 // try and combine it.
537 while (!WorkList.empty()) {
538 SDNode *N = WorkList.back();
541 // If N has no uses, it is dead. Make sure to revisit all N's operands once
542 // N is deleted from the DAG, since they too may now be dead or may have a
543 // reduced number of uses, allowing other xforms.
544 if (N->use_empty() && N != &Dummy) {
545 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
546 AddToWorkList(N->getOperand(i).Val);
552 SDOperand RV = visit(N);
554 // If nothing happened, try a target-specific DAG combine.
556 assert(N->getOpcode() != ISD::DELETED_NODE &&
557 "Node was deleted but visit returned NULL!");
558 if (N->getOpcode() >= ISD::BUILTIN_OP_END ||
559 TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode()))
560 RV = TLI.PerformDAGCombine(N, DagCombineInfo);
565 // If we get back the same node we passed in, rather than a new node or
566 // zero, we know that the node must have defined multiple values and
567 // CombineTo was used. Since CombineTo takes care of the worklist
568 // mechanics for us, we have no work to do in this case.
570 assert(N->getOpcode() != ISD::DELETED_NODE &&
571 RV.Val->getOpcode() != ISD::DELETED_NODE &&
572 "Node was deleted but visit returned new node!");
574 DEBUG(std::cerr << "\nReplacing.3 "; N->dump();
575 std::cerr << "\nWith: "; RV.Val->dump(&DAG);
577 std::vector<SDNode*> NowDead;
578 if (N->getNumValues() == RV.Val->getNumValues())
579 DAG.ReplaceAllUsesWith(N, RV.Val, &NowDead);
581 assert(N->getValueType(0) == RV.getValueType() && "Type mismatch");
583 DAG.ReplaceAllUsesWith(N, &OpV, &NowDead);
586 // Push the new node and any users onto the worklist
587 AddToWorkList(RV.Val);
588 AddUsersToWorkList(RV.Val);
590 // Nodes can be reintroduced into the worklist. Make sure we do not
591 // process a node that has been replaced.
592 removeFromWorkList(N);
593 for (unsigned i = 0, e = NowDead.size(); i != e; ++i)
594 removeFromWorkList(NowDead[i]);
596 // Finally, since the node is now dead, remove it from the graph.
602 // If the root changed (e.g. it was a dead load, update the root).
603 DAG.setRoot(Dummy.getValue());
606 SDOperand DAGCombiner::visit(SDNode *N) {
607 switch(N->getOpcode()) {
609 case ISD::TokenFactor: return visitTokenFactor(N);
610 case ISD::ADD: return visitADD(N);
611 case ISD::SUB: return visitSUB(N);
612 case ISD::MUL: return visitMUL(N);
613 case ISD::SDIV: return visitSDIV(N);
614 case ISD::UDIV: return visitUDIV(N);
615 case ISD::SREM: return visitSREM(N);
616 case ISD::UREM: return visitUREM(N);
617 case ISD::MULHU: return visitMULHU(N);
618 case ISD::MULHS: return visitMULHS(N);
619 case ISD::AND: return visitAND(N);
620 case ISD::OR: return visitOR(N);
621 case ISD::XOR: return visitXOR(N);
622 case ISD::SHL: return visitSHL(N);
623 case ISD::SRA: return visitSRA(N);
624 case ISD::SRL: return visitSRL(N);
625 case ISD::CTLZ: return visitCTLZ(N);
626 case ISD::CTTZ: return visitCTTZ(N);
627 case ISD::CTPOP: return visitCTPOP(N);
628 case ISD::SELECT: return visitSELECT(N);
629 case ISD::SELECT_CC: return visitSELECT_CC(N);
630 case ISD::SETCC: return visitSETCC(N);
631 case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N);
632 case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N);
633 case ISD::ANY_EXTEND: return visitANY_EXTEND(N);
634 case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N);
635 case ISD::TRUNCATE: return visitTRUNCATE(N);
636 case ISD::BIT_CONVERT: return visitBIT_CONVERT(N);
637 case ISD::VBIT_CONVERT: return visitVBIT_CONVERT(N);
638 case ISD::FADD: return visitFADD(N);
639 case ISD::FSUB: return visitFSUB(N);
640 case ISD::FMUL: return visitFMUL(N);
641 case ISD::FDIV: return visitFDIV(N);
642 case ISD::FREM: return visitFREM(N);
643 case ISD::FCOPYSIGN: return visitFCOPYSIGN(N);
644 case ISD::SINT_TO_FP: return visitSINT_TO_FP(N);
645 case ISD::UINT_TO_FP: return visitUINT_TO_FP(N);
646 case ISD::FP_TO_SINT: return visitFP_TO_SINT(N);
647 case ISD::FP_TO_UINT: return visitFP_TO_UINT(N);
648 case ISD::FP_ROUND: return visitFP_ROUND(N);
649 case ISD::FP_ROUND_INREG: return visitFP_ROUND_INREG(N);
650 case ISD::FP_EXTEND: return visitFP_EXTEND(N);
651 case ISD::FNEG: return visitFNEG(N);
652 case ISD::FABS: return visitFABS(N);
653 case ISD::BRCOND: return visitBRCOND(N);
654 case ISD::BR_CC: return visitBR_CC(N);
655 case ISD::LOAD: return visitLOAD(N);
656 case ISD::STORE: return visitSTORE(N);
657 case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N);
658 case ISD::VINSERT_VECTOR_ELT: return visitVINSERT_VECTOR_ELT(N);
659 case ISD::VBUILD_VECTOR: return visitVBUILD_VECTOR(N);
660 case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N);
661 case ISD::VVECTOR_SHUFFLE: return visitVVECTOR_SHUFFLE(N);
662 case ISD::VADD: return visitVBinOp(N, ISD::ADD , ISD::FADD);
663 case ISD::VSUB: return visitVBinOp(N, ISD::SUB , ISD::FSUB);
664 case ISD::VMUL: return visitVBinOp(N, ISD::MUL , ISD::FMUL);
665 case ISD::VSDIV: return visitVBinOp(N, ISD::SDIV, ISD::FDIV);
666 case ISD::VUDIV: return visitVBinOp(N, ISD::UDIV, ISD::UDIV);
667 case ISD::VAND: return visitVBinOp(N, ISD::AND , ISD::AND);
668 case ISD::VOR: return visitVBinOp(N, ISD::OR , ISD::OR);
669 case ISD::VXOR: return visitVBinOp(N, ISD::XOR , ISD::XOR);
674 /// getInputChainForNode - Given a node, return its input chain if it has one,
675 /// otherwise return a null sd operand.
676 static SDOperand getInputChainForNode(SDNode *N) {
677 if (unsigned NumOps = N->getNumOperands()) {
678 if (N->getOperand(0).getValueType() == MVT::Other)
679 return N->getOperand(0);
680 else if (N->getOperand(NumOps-1).getValueType() == MVT::Other)
681 return N->getOperand(NumOps-1);
682 for (unsigned i = 1; i < NumOps-1; ++i)
683 if (N->getOperand(i).getValueType() == MVT::Other)
684 return N->getOperand(i);
686 return SDOperand(0, 0);
689 SDOperand DAGCombiner::visitTokenFactor(SDNode *N) {
690 // If N has two operands, where one has an input chain equal to the other,
691 // the 'other' chain is redundant.
692 if (N->getNumOperands() == 2) {
693 if (getInputChainForNode(N->getOperand(0).Val) == N->getOperand(1))
694 return N->getOperand(0);
695 if (getInputChainForNode(N->getOperand(1).Val) == N->getOperand(0))
696 return N->getOperand(1);
700 SmallVector<SDNode *, 8> TFs; // List of token factors to visit.
701 SmallVector<SDOperand, 8> Ops; // Ops for replacing token factor.
702 bool Changed = false; // If we should replace this token factor.
704 // Start out with this token factor.
707 // Iterate through token factors. The TFs grows when new token factors are
709 for (unsigned i = 0; i < TFs.size(); ++i) {
712 // Check each of the operands.
713 for (unsigned i = 0, ie = TF->getNumOperands(); i != ie; ++i) {
714 SDOperand Op = TF->getOperand(i);
716 switch (Op.getOpcode()) {
717 case ISD::EntryToken:
718 // Entry tokens don't need to be added to the list. They are
723 case ISD::TokenFactor:
724 if ((CombinerAA || Op.hasOneUse()) &&
725 std::find(TFs.begin(), TFs.end(), Op.Val) == TFs.end()) {
726 // Queue up for processing.
727 TFs.push_back(Op.Val);
728 // Clean up in case the token factor is removed.
729 AddToWorkList(Op.Val);
736 // Only add if not there prior.
737 if (std::find(Ops.begin(), Ops.end(), Op) == Ops.end())
746 // If we've change things around then replace token factor.
748 if (Ops.size() == 0) {
749 // The entry token is the only possible outcome.
750 Result = DAG.getEntryNode();
752 // New and improved token factor.
753 Result = DAG.getNode(ISD::TokenFactor, MVT::Other, &Ops[0], Ops.size());
756 // Don't add users to work list.
757 return CombineTo(N, Result, false);
763 SDOperand DAGCombiner::visitADD(SDNode *N) {
764 SDOperand N0 = N->getOperand(0);
765 SDOperand N1 = N->getOperand(1);
766 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
767 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
768 MVT::ValueType VT = N0.getValueType();
770 // fold (add c1, c2) -> c1+c2
772 return DAG.getNode(ISD::ADD, VT, N0, N1);
773 // canonicalize constant to RHS
775 return DAG.getNode(ISD::ADD, VT, N1, N0);
776 // fold (add x, 0) -> x
777 if (N1C && N1C->isNullValue())
779 // fold ((c1-A)+c2) -> (c1+c2)-A
780 if (N1C && N0.getOpcode() == ISD::SUB)
781 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getOperand(0)))
782 return DAG.getNode(ISD::SUB, VT,
783 DAG.getConstant(N1C->getValue()+N0C->getValue(), VT),
786 SDOperand RADD = ReassociateOps(ISD::ADD, N0, N1);
789 // fold ((0-A) + B) -> B-A
790 if (N0.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N0.getOperand(0)) &&
791 cast<ConstantSDNode>(N0.getOperand(0))->isNullValue())
792 return DAG.getNode(ISD::SUB, VT, N1, N0.getOperand(1));
793 // fold (A + (0-B)) -> A-B
794 if (N1.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N1.getOperand(0)) &&
795 cast<ConstantSDNode>(N1.getOperand(0))->isNullValue())
796 return DAG.getNode(ISD::SUB, VT, N0, N1.getOperand(1));
797 // fold (A+(B-A)) -> B
798 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1))
799 return N1.getOperand(0);
801 if (!MVT::isVector(VT) && SimplifyDemandedBits(SDOperand(N, 0)))
802 return SDOperand(N, 0);
804 // fold (a+b) -> (a|b) iff a and b share no bits.
805 if (MVT::isInteger(VT) && !MVT::isVector(VT)) {
806 uint64_t LHSZero, LHSOne;
807 uint64_t RHSZero, RHSOne;
808 uint64_t Mask = MVT::getIntVTBitMask(VT);
809 TLI.ComputeMaskedBits(N0, Mask, LHSZero, LHSOne);
811 TLI.ComputeMaskedBits(N1, Mask, RHSZero, RHSOne);
813 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
814 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
815 if ((RHSZero & (~LHSZero & Mask)) == (~LHSZero & Mask) ||
816 (LHSZero & (~RHSZero & Mask)) == (~RHSZero & Mask))
817 return DAG.getNode(ISD::OR, VT, N0, N1);
824 SDOperand DAGCombiner::visitSUB(SDNode *N) {
825 SDOperand N0 = N->getOperand(0);
826 SDOperand N1 = N->getOperand(1);
827 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val);
828 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val);
829 MVT::ValueType VT = N0.getValueType();
831 // fold (sub x, x) -> 0
833 return DAG.getConstant(0, N->getValueType(0));
834 // fold (sub c1, c2) -> c1-c2
836 return DAG.getNode(ISD::SUB, VT, N0, N1);
837 // fold (sub x, c) -> (add x, -c)
839 return DAG.getNode(ISD::ADD, VT, N0, DAG.getConstant(-N1C->getValue(), VT));
841 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1)
842 return N0.getOperand(1);
844 if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1)
845 return N0.getOperand(0);
849 SDOperand DAGCombiner::visitMUL(SDNode *N) {
850 SDOperand N0 = N->getOperand(0);
851 SDOperand N1 = N->getOperand(1);
852 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
853 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
854 MVT::ValueType VT = N0.getValueType();
856 // fold (mul c1, c2) -> c1*c2
858 return DAG.getNode(ISD::MUL, VT, N0, N1);
859 // canonicalize constant to RHS
861 return DAG.getNode(ISD::MUL, VT, N1, N0);
862 // fold (mul x, 0) -> 0
863 if (N1C && N1C->isNullValue())
865 // fold (mul x, -1) -> 0-x
866 if (N1C && N1C->isAllOnesValue())
867 return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT), N0);
868 // fold (mul x, (1 << c)) -> x << c
869 if (N1C && isPowerOf2_64(N1C->getValue()))
870 return DAG.getNode(ISD::SHL, VT, N0,
871 DAG.getConstant(Log2_64(N1C->getValue()),
872 TLI.getShiftAmountTy()));
873 // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c
874 if (N1C && isPowerOf2_64(-N1C->getSignExtended())) {
875 // FIXME: If the input is something that is easily negated (e.g. a
876 // single-use add), we should put the negate there.
877 return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT),
878 DAG.getNode(ISD::SHL, VT, N0,
879 DAG.getConstant(Log2_64(-N1C->getSignExtended()),
880 TLI.getShiftAmountTy())));
883 // (mul (shl X, c1), c2) -> (mul X, c2 << c1)
884 if (N1C && N0.getOpcode() == ISD::SHL &&
885 isa<ConstantSDNode>(N0.getOperand(1))) {
886 SDOperand C3 = DAG.getNode(ISD::SHL, VT, N1, N0.getOperand(1));
887 AddToWorkList(C3.Val);
888 return DAG.getNode(ISD::MUL, VT, N0.getOperand(0), C3);
891 // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one
894 SDOperand Sh(0,0), Y(0,0);
895 // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)).
896 if (N0.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N0.getOperand(1)) &&
897 N0.Val->hasOneUse()) {
899 } else if (N1.getOpcode() == ISD::SHL &&
900 isa<ConstantSDNode>(N1.getOperand(1)) && N1.Val->hasOneUse()) {
904 SDOperand Mul = DAG.getNode(ISD::MUL, VT, Sh.getOperand(0), Y);
905 return DAG.getNode(ISD::SHL, VT, Mul, Sh.getOperand(1));
908 // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2)
909 if (N1C && N0.getOpcode() == ISD::ADD && N0.Val->hasOneUse() &&
910 isa<ConstantSDNode>(N0.getOperand(1))) {
911 return DAG.getNode(ISD::ADD, VT,
912 DAG.getNode(ISD::MUL, VT, N0.getOperand(0), N1),
913 DAG.getNode(ISD::MUL, VT, N0.getOperand(1), N1));
917 SDOperand RMUL = ReassociateOps(ISD::MUL, N0, N1);
923 SDOperand DAGCombiner::visitSDIV(SDNode *N) {
924 SDOperand N0 = N->getOperand(0);
925 SDOperand N1 = N->getOperand(1);
926 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val);
927 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val);
928 MVT::ValueType VT = N->getValueType(0);
930 // fold (sdiv c1, c2) -> c1/c2
931 if (N0C && N1C && !N1C->isNullValue())
932 return DAG.getNode(ISD::SDIV, VT, N0, N1);
933 // fold (sdiv X, 1) -> X
934 if (N1C && N1C->getSignExtended() == 1LL)
936 // fold (sdiv X, -1) -> 0-X
937 if (N1C && N1C->isAllOnesValue())
938 return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT), N0);
939 // If we know the sign bits of both operands are zero, strength reduce to a
940 // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2
941 uint64_t SignBit = 1ULL << (MVT::getSizeInBits(VT)-1);
942 if (TLI.MaskedValueIsZero(N1, SignBit) &&
943 TLI.MaskedValueIsZero(N0, SignBit))
944 return DAG.getNode(ISD::UDIV, N1.getValueType(), N0, N1);
945 // fold (sdiv X, pow2) -> simple ops after legalize
946 if (N1C && N1C->getValue() && !TLI.isIntDivCheap() &&
947 (isPowerOf2_64(N1C->getSignExtended()) ||
948 isPowerOf2_64(-N1C->getSignExtended()))) {
949 // If dividing by powers of two is cheap, then don't perform the following
951 if (TLI.isPow2DivCheap())
953 int64_t pow2 = N1C->getSignExtended();
954 int64_t abs2 = pow2 > 0 ? pow2 : -pow2;
955 unsigned lg2 = Log2_64(abs2);
956 // Splat the sign bit into the register
957 SDOperand SGN = DAG.getNode(ISD::SRA, VT, N0,
958 DAG.getConstant(MVT::getSizeInBits(VT)-1,
959 TLI.getShiftAmountTy()));
960 AddToWorkList(SGN.Val);
961 // Add (N0 < 0) ? abs2 - 1 : 0;
962 SDOperand SRL = DAG.getNode(ISD::SRL, VT, SGN,
963 DAG.getConstant(MVT::getSizeInBits(VT)-lg2,
964 TLI.getShiftAmountTy()));
965 SDOperand ADD = DAG.getNode(ISD::ADD, VT, N0, SRL);
966 AddToWorkList(SRL.Val);
967 AddToWorkList(ADD.Val); // Divide by pow2
968 SDOperand SRA = DAG.getNode(ISD::SRA, VT, ADD,
969 DAG.getConstant(lg2, TLI.getShiftAmountTy()));
970 // If we're dividing by a positive value, we're done. Otherwise, we must
971 // negate the result.
974 AddToWorkList(SRA.Val);
975 return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT), SRA);
977 // if integer divide is expensive and we satisfy the requirements, emit an
978 // alternate sequence.
979 if (N1C && (N1C->getSignExtended() < -1 || N1C->getSignExtended() > 1) &&
980 !TLI.isIntDivCheap()) {
981 SDOperand Op = BuildSDIV(N);
982 if (Op.Val) return Op;
987 SDOperand DAGCombiner::visitUDIV(SDNode *N) {
988 SDOperand N0 = N->getOperand(0);
989 SDOperand N1 = N->getOperand(1);
990 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val);
991 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val);
992 MVT::ValueType VT = N->getValueType(0);
994 // fold (udiv c1, c2) -> c1/c2
995 if (N0C && N1C && !N1C->isNullValue())
996 return DAG.getNode(ISD::UDIV, VT, N0, N1);
997 // fold (udiv x, (1 << c)) -> x >>u c
998 if (N1C && isPowerOf2_64(N1C->getValue()))
999 return DAG.getNode(ISD::SRL, VT, N0,
1000 DAG.getConstant(Log2_64(N1C->getValue()),
1001 TLI.getShiftAmountTy()));
1002 // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2
1003 if (N1.getOpcode() == ISD::SHL) {
1004 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
1005 if (isPowerOf2_64(SHC->getValue())) {
1006 MVT::ValueType ADDVT = N1.getOperand(1).getValueType();
1007 SDOperand Add = DAG.getNode(ISD::ADD, ADDVT, N1.getOperand(1),
1008 DAG.getConstant(Log2_64(SHC->getValue()),
1010 AddToWorkList(Add.Val);
1011 return DAG.getNode(ISD::SRL, VT, N0, Add);
1015 // fold (udiv x, c) -> alternate
1016 if (N1C && N1C->getValue() && !TLI.isIntDivCheap()) {
1017 SDOperand Op = BuildUDIV(N);
1018 if (Op.Val) return Op;
1023 SDOperand DAGCombiner::visitSREM(SDNode *N) {
1024 SDOperand N0 = N->getOperand(0);
1025 SDOperand N1 = N->getOperand(1);
1026 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1027 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1028 MVT::ValueType VT = N->getValueType(0);
1030 // fold (srem c1, c2) -> c1%c2
1031 if (N0C && N1C && !N1C->isNullValue())
1032 return DAG.getNode(ISD::SREM, VT, N0, N1);
1033 // If we know the sign bits of both operands are zero, strength reduce to a
1034 // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15
1035 uint64_t SignBit = 1ULL << (MVT::getSizeInBits(VT)-1);
1036 if (TLI.MaskedValueIsZero(N1, SignBit) &&
1037 TLI.MaskedValueIsZero(N0, SignBit))
1038 return DAG.getNode(ISD::UREM, VT, N0, N1);
1040 // Unconditionally lower X%C -> X-X/C*C. This allows the X/C logic to hack on
1041 // the remainder operation.
1042 if (N1C && !N1C->isNullValue()) {
1043 SDOperand Div = DAG.getNode(ISD::SDIV, VT, N0, N1);
1044 SDOperand Mul = DAG.getNode(ISD::MUL, VT, Div, N1);
1045 SDOperand Sub = DAG.getNode(ISD::SUB, VT, N0, Mul);
1046 AddToWorkList(Div.Val);
1047 AddToWorkList(Mul.Val);
1054 SDOperand DAGCombiner::visitUREM(SDNode *N) {
1055 SDOperand N0 = N->getOperand(0);
1056 SDOperand N1 = N->getOperand(1);
1057 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1058 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1059 MVT::ValueType VT = N->getValueType(0);
1061 // fold (urem c1, c2) -> c1%c2
1062 if (N0C && N1C && !N1C->isNullValue())
1063 return DAG.getNode(ISD::UREM, VT, N0, N1);
1064 // fold (urem x, pow2) -> (and x, pow2-1)
1065 if (N1C && !N1C->isNullValue() && isPowerOf2_64(N1C->getValue()))
1066 return DAG.getNode(ISD::AND, VT, N0, DAG.getConstant(N1C->getValue()-1,VT));
1067 // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1))
1068 if (N1.getOpcode() == ISD::SHL) {
1069 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
1070 if (isPowerOf2_64(SHC->getValue())) {
1071 SDOperand Add = DAG.getNode(ISD::ADD, VT, N1,DAG.getConstant(~0ULL,VT));
1072 AddToWorkList(Add.Val);
1073 return DAG.getNode(ISD::AND, VT, N0, Add);
1078 // Unconditionally lower X%C -> X-X/C*C. This allows the X/C logic to hack on
1079 // the remainder operation.
1080 if (N1C && !N1C->isNullValue()) {
1081 SDOperand Div = DAG.getNode(ISD::UDIV, VT, N0, N1);
1082 SDOperand Mul = DAG.getNode(ISD::MUL, VT, Div, N1);
1083 SDOperand Sub = DAG.getNode(ISD::SUB, VT, N0, Mul);
1084 AddToWorkList(Div.Val);
1085 AddToWorkList(Mul.Val);
1092 SDOperand DAGCombiner::visitMULHS(SDNode *N) {
1093 SDOperand N0 = N->getOperand(0);
1094 SDOperand N1 = N->getOperand(1);
1095 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1097 // fold (mulhs x, 0) -> 0
1098 if (N1C && N1C->isNullValue())
1100 // fold (mulhs x, 1) -> (sra x, size(x)-1)
1101 if (N1C && N1C->getValue() == 1)
1102 return DAG.getNode(ISD::SRA, N0.getValueType(), N0,
1103 DAG.getConstant(MVT::getSizeInBits(N0.getValueType())-1,
1104 TLI.getShiftAmountTy()));
1108 SDOperand DAGCombiner::visitMULHU(SDNode *N) {
1109 SDOperand N0 = N->getOperand(0);
1110 SDOperand N1 = N->getOperand(1);
1111 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1113 // fold (mulhu x, 0) -> 0
1114 if (N1C && N1C->isNullValue())
1116 // fold (mulhu x, 1) -> 0
1117 if (N1C && N1C->getValue() == 1)
1118 return DAG.getConstant(0, N0.getValueType());
1122 /// SimplifyBinOpWithSameOpcodeHands - If this is a binary operator with
1123 /// two operands of the same opcode, try to simplify it.
1124 SDOperand DAGCombiner::SimplifyBinOpWithSameOpcodeHands(SDNode *N) {
1125 SDOperand N0 = N->getOperand(0), N1 = N->getOperand(1);
1126 MVT::ValueType VT = N0.getValueType();
1127 assert(N0.getOpcode() == N1.getOpcode() && "Bad input!");
1129 // For each of OP in AND/OR/XOR:
1130 // fold (OP (zext x), (zext y)) -> (zext (OP x, y))
1131 // fold (OP (sext x), (sext y)) -> (sext (OP x, y))
1132 // fold (OP (aext x), (aext y)) -> (aext (OP x, y))
1133 // fold (OP (trunc x), (trunc y)) -> (trunc (OP x, y))
1134 if ((N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND||
1135 N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::TRUNCATE) &&
1136 N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()) {
1137 SDOperand ORNode = DAG.getNode(N->getOpcode(),
1138 N0.getOperand(0).getValueType(),
1139 N0.getOperand(0), N1.getOperand(0));
1140 AddToWorkList(ORNode.Val);
1141 return DAG.getNode(N0.getOpcode(), VT, ORNode);
1144 // For each of OP in SHL/SRL/SRA/AND...
1145 // fold (and (OP x, z), (OP y, z)) -> (OP (and x, y), z)
1146 // fold (or (OP x, z), (OP y, z)) -> (OP (or x, y), z)
1147 // fold (xor (OP x, z), (OP y, z)) -> (OP (xor x, y), z)
1148 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL ||
1149 N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::AND) &&
1150 N0.getOperand(1) == N1.getOperand(1)) {
1151 SDOperand ORNode = DAG.getNode(N->getOpcode(),
1152 N0.getOperand(0).getValueType(),
1153 N0.getOperand(0), N1.getOperand(0));
1154 AddToWorkList(ORNode.Val);
1155 return DAG.getNode(N0.getOpcode(), VT, ORNode, N0.getOperand(1));
1161 SDOperand DAGCombiner::visitAND(SDNode *N) {
1162 SDOperand N0 = N->getOperand(0);
1163 SDOperand N1 = N->getOperand(1);
1164 SDOperand LL, LR, RL, RR, CC0, CC1;
1165 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1166 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1167 MVT::ValueType VT = N1.getValueType();
1169 // fold (and c1, c2) -> c1&c2
1171 return DAG.getNode(ISD::AND, VT, N0, N1);
1172 // canonicalize constant to RHS
1174 return DAG.getNode(ISD::AND, VT, N1, N0);
1175 // fold (and x, -1) -> x
1176 if (N1C && N1C->isAllOnesValue())
1178 // if (and x, c) is known to be zero, return 0
1179 if (N1C && TLI.MaskedValueIsZero(SDOperand(N, 0), MVT::getIntVTBitMask(VT)))
1180 return DAG.getConstant(0, VT);
1182 SDOperand RAND = ReassociateOps(ISD::AND, N0, N1);
1185 // fold (and (or x, 0xFFFF), 0xFF) -> 0xFF
1186 if (N1C && N0.getOpcode() == ISD::OR)
1187 if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
1188 if ((ORI->getValue() & N1C->getValue()) == N1C->getValue())
1190 // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits.
1191 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
1192 unsigned InMask = MVT::getIntVTBitMask(N0.getOperand(0).getValueType());
1193 if (TLI.MaskedValueIsZero(N0.getOperand(0),
1194 ~N1C->getValue() & InMask)) {
1195 SDOperand Zext = DAG.getNode(ISD::ZERO_EXTEND, N0.getValueType(),
1198 // Replace uses of the AND with uses of the Zero extend node.
1201 // We actually want to replace all uses of the any_extend with the
1202 // zero_extend, to avoid duplicating things. This will later cause this
1203 // AND to be folded.
1204 CombineTo(N0.Val, Zext);
1205 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
1208 // fold (and (setcc x), (setcc y)) -> (setcc (and x, y))
1209 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
1210 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
1211 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
1213 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
1214 MVT::isInteger(LL.getValueType())) {
1215 // fold (X == 0) & (Y == 0) -> (X|Y == 0)
1216 if (cast<ConstantSDNode>(LR)->getValue() == 0 && Op1 == ISD::SETEQ) {
1217 SDOperand ORNode = DAG.getNode(ISD::OR, LR.getValueType(), LL, RL);
1218 AddToWorkList(ORNode.Val);
1219 return DAG.getSetCC(VT, ORNode, LR, Op1);
1221 // fold (X == -1) & (Y == -1) -> (X&Y == -1)
1222 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETEQ) {
1223 SDOperand ANDNode = DAG.getNode(ISD::AND, LR.getValueType(), LL, RL);
1224 AddToWorkList(ANDNode.Val);
1225 return DAG.getSetCC(VT, ANDNode, LR, Op1);
1227 // fold (X > -1) & (Y > -1) -> (X|Y > -1)
1228 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETGT) {
1229 SDOperand ORNode = DAG.getNode(ISD::OR, LR.getValueType(), LL, RL);
1230 AddToWorkList(ORNode.Val);
1231 return DAG.getSetCC(VT, ORNode, LR, Op1);
1234 // canonicalize equivalent to ll == rl
1235 if (LL == RR && LR == RL) {
1236 Op1 = ISD::getSetCCSwappedOperands(Op1);
1239 if (LL == RL && LR == RR) {
1240 bool isInteger = MVT::isInteger(LL.getValueType());
1241 ISD::CondCode Result = ISD::getSetCCAndOperation(Op0, Op1, isInteger);
1242 if (Result != ISD::SETCC_INVALID)
1243 return DAG.getSetCC(N0.getValueType(), LL, LR, Result);
1247 // Simplify: and (op x...), (op y...) -> (op (and x, y))
1248 if (N0.getOpcode() == N1.getOpcode()) {
1249 SDOperand Tmp = SimplifyBinOpWithSameOpcodeHands(N);
1250 if (Tmp.Val) return Tmp;
1253 // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1)
1254 // fold (and (sra)) -> (and (srl)) when possible.
1255 if (!MVT::isVector(VT) &&
1256 SimplifyDemandedBits(SDOperand(N, 0)))
1257 return SDOperand(N, 0);
1258 // fold (zext_inreg (extload x)) -> (zextload x)
1259 if (ISD::isEXTLoad(N0.Val)) {
1260 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
1261 MVT::ValueType EVT = LN0->getLoadedVT();
1262 // If we zero all the possible extended bits, then we can turn this into
1263 // a zextload if we are running before legalize or the operation is legal.
1264 if (TLI.MaskedValueIsZero(N1, ~0ULL << MVT::getSizeInBits(EVT)) &&
1265 (!AfterLegalize || TLI.isLoadXLegal(ISD::ZEXTLOAD, EVT))) {
1266 SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, LN0->getChain(),
1267 LN0->getBasePtr(), LN0->getSrcValue(),
1268 LN0->getSrcValueOffset(), EVT);
1270 CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1));
1271 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
1274 // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use
1275 if (ISD::isSEXTLoad(N0.Val) && N0.hasOneUse()) {
1276 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
1277 MVT::ValueType EVT = LN0->getLoadedVT();
1278 // If we zero all the possible extended bits, then we can turn this into
1279 // a zextload if we are running before legalize or the operation is legal.
1280 if (TLI.MaskedValueIsZero(N1, ~0ULL << MVT::getSizeInBits(EVT)) &&
1281 (!AfterLegalize || TLI.isLoadXLegal(ISD::ZEXTLOAD, EVT))) {
1282 SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, LN0->getChain(),
1283 LN0->getBasePtr(), LN0->getSrcValue(),
1284 LN0->getSrcValueOffset(), EVT);
1286 CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1));
1287 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
1291 // fold (and (load x), 255) -> (zextload x, i8)
1292 // fold (and (extload x, i16), 255) -> (zextload x, i8)
1293 if (N1C && N0.getOpcode() == ISD::LOAD) {
1294 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
1295 if (LN0->getExtensionType() != ISD::SEXTLOAD &&
1297 MVT::ValueType EVT, LoadedVT;
1298 if (N1C->getValue() == 255)
1300 else if (N1C->getValue() == 65535)
1302 else if (N1C->getValue() == ~0U)
1307 LoadedVT = LN0->getLoadedVT();
1308 if (EVT != MVT::Other && LoadedVT > EVT &&
1309 (!AfterLegalize || TLI.isLoadXLegal(ISD::ZEXTLOAD, EVT))) {
1310 MVT::ValueType PtrType = N0.getOperand(1).getValueType();
1311 // For big endian targets, we need to add an offset to the pointer to
1312 // load the correct bytes. For little endian systems, we merely need to
1313 // read fewer bytes from the same pointer.
1315 (MVT::getSizeInBits(LoadedVT) - MVT::getSizeInBits(EVT)) / 8;
1316 SDOperand NewPtr = LN0->getBasePtr();
1317 if (!TLI.isLittleEndian())
1318 NewPtr = DAG.getNode(ISD::ADD, PtrType, NewPtr,
1319 DAG.getConstant(PtrOff, PtrType));
1320 AddToWorkList(NewPtr.Val);
1322 DAG.getExtLoad(ISD::ZEXTLOAD, VT, LN0->getChain(), NewPtr,
1323 LN0->getSrcValue(), LN0->getSrcValueOffset(), EVT);
1325 CombineTo(N0.Val, Load, Load.getValue(1));
1326 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
1334 SDOperand DAGCombiner::visitOR(SDNode *N) {
1335 SDOperand N0 = N->getOperand(0);
1336 SDOperand N1 = N->getOperand(1);
1337 SDOperand LL, LR, RL, RR, CC0, CC1;
1338 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1339 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1340 MVT::ValueType VT = N1.getValueType();
1341 unsigned OpSizeInBits = MVT::getSizeInBits(VT);
1343 // fold (or c1, c2) -> c1|c2
1345 return DAG.getNode(ISD::OR, VT, N0, N1);
1346 // canonicalize constant to RHS
1348 return DAG.getNode(ISD::OR, VT, N1, N0);
1349 // fold (or x, 0) -> x
1350 if (N1C && N1C->isNullValue())
1352 // fold (or x, -1) -> -1
1353 if (N1C && N1C->isAllOnesValue())
1355 // fold (or x, c) -> c iff (x & ~c) == 0
1357 TLI.MaskedValueIsZero(N0,~N1C->getValue() & (~0ULL>>(64-OpSizeInBits))))
1360 SDOperand ROR = ReassociateOps(ISD::OR, N0, N1);
1363 // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2)
1364 if (N1C && N0.getOpcode() == ISD::AND && N0.Val->hasOneUse() &&
1365 isa<ConstantSDNode>(N0.getOperand(1))) {
1366 ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1));
1367 return DAG.getNode(ISD::AND, VT, DAG.getNode(ISD::OR, VT, N0.getOperand(0),
1369 DAG.getConstant(N1C->getValue() | C1->getValue(), VT));
1371 // fold (or (setcc x), (setcc y)) -> (setcc (or x, y))
1372 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
1373 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
1374 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
1376 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
1377 MVT::isInteger(LL.getValueType())) {
1378 // fold (X != 0) | (Y != 0) -> (X|Y != 0)
1379 // fold (X < 0) | (Y < 0) -> (X|Y < 0)
1380 if (cast<ConstantSDNode>(LR)->getValue() == 0 &&
1381 (Op1 == ISD::SETNE || Op1 == ISD::SETLT)) {
1382 SDOperand ORNode = DAG.getNode(ISD::OR, LR.getValueType(), LL, RL);
1383 AddToWorkList(ORNode.Val);
1384 return DAG.getSetCC(VT, ORNode, LR, Op1);
1386 // fold (X != -1) | (Y != -1) -> (X&Y != -1)
1387 // fold (X > -1) | (Y > -1) -> (X&Y > -1)
1388 if (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
1389 (Op1 == ISD::SETNE || Op1 == ISD::SETGT)) {
1390 SDOperand ANDNode = DAG.getNode(ISD::AND, LR.getValueType(), LL, RL);
1391 AddToWorkList(ANDNode.Val);
1392 return DAG.getSetCC(VT, ANDNode, LR, Op1);
1395 // canonicalize equivalent to ll == rl
1396 if (LL == RR && LR == RL) {
1397 Op1 = ISD::getSetCCSwappedOperands(Op1);
1400 if (LL == RL && LR == RR) {
1401 bool isInteger = MVT::isInteger(LL.getValueType());
1402 ISD::CondCode Result = ISD::getSetCCOrOperation(Op0, Op1, isInteger);
1403 if (Result != ISD::SETCC_INVALID)
1404 return DAG.getSetCC(N0.getValueType(), LL, LR, Result);
1408 // Simplify: or (op x...), (op y...) -> (op (or x, y))
1409 if (N0.getOpcode() == N1.getOpcode()) {
1410 SDOperand Tmp = SimplifyBinOpWithSameOpcodeHands(N);
1411 if (Tmp.Val) return Tmp;
1414 // (X & C1) | (Y & C2) -> (X|Y) & C3 if possible.
1415 if (N0.getOpcode() == ISD::AND &&
1416 N1.getOpcode() == ISD::AND &&
1417 N0.getOperand(1).getOpcode() == ISD::Constant &&
1418 N1.getOperand(1).getOpcode() == ISD::Constant &&
1419 // Don't increase # computations.
1420 (N0.Val->hasOneUse() || N1.Val->hasOneUse())) {
1421 // We can only do this xform if we know that bits from X that are set in C2
1422 // but not in C1 are already zero. Likewise for Y.
1423 uint64_t LHSMask = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
1424 uint64_t RHSMask = cast<ConstantSDNode>(N1.getOperand(1))->getValue();
1426 if (TLI.MaskedValueIsZero(N0.getOperand(0), RHSMask&~LHSMask) &&
1427 TLI.MaskedValueIsZero(N1.getOperand(0), LHSMask&~RHSMask)) {
1428 SDOperand X =DAG.getNode(ISD::OR, VT, N0.getOperand(0), N1.getOperand(0));
1429 return DAG.getNode(ISD::AND, VT, X, DAG.getConstant(LHSMask|RHSMask, VT));
1434 // See if this is some rotate idiom.
1435 if (SDNode *Rot = MatchRotate(N0, N1))
1436 return SDOperand(Rot, 0);
1442 /// MatchRotateHalf - Match "(X shl/srl V1) & V2" where V2 may not be present.
1443 static bool MatchRotateHalf(SDOperand Op, SDOperand &Shift, SDOperand &Mask) {
1444 if (Op.getOpcode() == ISD::AND) {
1445 if (isa<ConstantSDNode>(Op.getOperand(1))) {
1446 Mask = Op.getOperand(1);
1447 Op = Op.getOperand(0);
1453 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SHL) {
1461 // MatchRotate - Handle an 'or' of two operands. If this is one of the many
1462 // idioms for rotate, and if the target supports rotation instructions, generate
1464 SDNode *DAGCombiner::MatchRotate(SDOperand LHS, SDOperand RHS) {
1465 // Must be a legal type. Expanded an promoted things won't work with rotates.
1466 MVT::ValueType VT = LHS.getValueType();
1467 if (!TLI.isTypeLegal(VT)) return 0;
1469 // The target must have at least one rotate flavor.
1470 bool HasROTL = TLI.isOperationLegal(ISD::ROTL, VT);
1471 bool HasROTR = TLI.isOperationLegal(ISD::ROTR, VT);
1472 if (!HasROTL && !HasROTR) return 0;
1474 // Match "(X shl/srl V1) & V2" where V2 may not be present.
1475 SDOperand LHSShift; // The shift.
1476 SDOperand LHSMask; // AND value if any.
1477 if (!MatchRotateHalf(LHS, LHSShift, LHSMask))
1478 return 0; // Not part of a rotate.
1480 SDOperand RHSShift; // The shift.
1481 SDOperand RHSMask; // AND value if any.
1482 if (!MatchRotateHalf(RHS, RHSShift, RHSMask))
1483 return 0; // Not part of a rotate.
1485 if (LHSShift.getOperand(0) != RHSShift.getOperand(0))
1486 return 0; // Not shifting the same value.
1488 if (LHSShift.getOpcode() == RHSShift.getOpcode())
1489 return 0; // Shifts must disagree.
1491 // Canonicalize shl to left side in a shl/srl pair.
1492 if (RHSShift.getOpcode() == ISD::SHL) {
1493 std::swap(LHS, RHS);
1494 std::swap(LHSShift, RHSShift);
1495 std::swap(LHSMask , RHSMask );
1498 unsigned OpSizeInBits = MVT::getSizeInBits(VT);
1500 // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1)
1501 // fold (or (shl x, C1), (srl x, C2)) -> (rotr x, C2)
1502 if (LHSShift.getOperand(1).getOpcode() == ISD::Constant &&
1503 RHSShift.getOperand(1).getOpcode() == ISD::Constant) {
1504 uint64_t LShVal = cast<ConstantSDNode>(LHSShift.getOperand(1))->getValue();
1505 uint64_t RShVal = cast<ConstantSDNode>(RHSShift.getOperand(1))->getValue();
1506 if ((LShVal + RShVal) != OpSizeInBits)
1511 Rot = DAG.getNode(ISD::ROTL, VT, LHSShift.getOperand(0),
1512 LHSShift.getOperand(1));
1514 Rot = DAG.getNode(ISD::ROTR, VT, LHSShift.getOperand(0),
1515 RHSShift.getOperand(1));
1517 // If there is an AND of either shifted operand, apply it to the result.
1518 if (LHSMask.Val || RHSMask.Val) {
1519 uint64_t Mask = MVT::getIntVTBitMask(VT);
1522 uint64_t RHSBits = (1ULL << LShVal)-1;
1523 Mask &= cast<ConstantSDNode>(LHSMask)->getValue() | RHSBits;
1526 uint64_t LHSBits = ~((1ULL << (OpSizeInBits-RShVal))-1);
1527 Mask &= cast<ConstantSDNode>(RHSMask)->getValue() | LHSBits;
1530 Rot = DAG.getNode(ISD::AND, VT, Rot, DAG.getConstant(Mask, VT));
1536 // If there is a mask here, and we have a variable shift, we can't be sure
1537 // that we're masking out the right stuff.
1538 if (LHSMask.Val || RHSMask.Val)
1541 // fold (or (shl x, y), (srl x, (sub 32, y))) -> (rotl x, y)
1542 // fold (or (shl x, y), (srl x, (sub 32, y))) -> (rotr x, (sub 32, y))
1543 if (RHSShift.getOperand(1).getOpcode() == ISD::SUB &&
1544 LHSShift.getOperand(1) == RHSShift.getOperand(1).getOperand(1)) {
1545 if (ConstantSDNode *SUBC =
1546 dyn_cast<ConstantSDNode>(RHSShift.getOperand(1).getOperand(0))) {
1547 if (SUBC->getValue() == OpSizeInBits)
1549 return DAG.getNode(ISD::ROTL, VT, LHSShift.getOperand(0),
1550 LHSShift.getOperand(1)).Val;
1552 return DAG.getNode(ISD::ROTR, VT, LHSShift.getOperand(0),
1553 LHSShift.getOperand(1)).Val;
1557 // fold (or (shl x, (sub 32, y)), (srl x, r)) -> (rotr x, y)
1558 // fold (or (shl x, (sub 32, y)), (srl x, r)) -> (rotl x, (sub 32, y))
1559 if (LHSShift.getOperand(1).getOpcode() == ISD::SUB &&
1560 RHSShift.getOperand(1) == LHSShift.getOperand(1).getOperand(1)) {
1561 if (ConstantSDNode *SUBC =
1562 dyn_cast<ConstantSDNode>(LHSShift.getOperand(1).getOperand(0))) {
1563 if (SUBC->getValue() == OpSizeInBits)
1565 return DAG.getNode(ISD::ROTL, VT, LHSShift.getOperand(0),
1566 LHSShift.getOperand(1)).Val;
1568 return DAG.getNode(ISD::ROTR, VT, LHSShift.getOperand(0),
1569 RHSShift.getOperand(1)).Val;
1577 SDOperand DAGCombiner::visitXOR(SDNode *N) {
1578 SDOperand N0 = N->getOperand(0);
1579 SDOperand N1 = N->getOperand(1);
1580 SDOperand LHS, RHS, CC;
1581 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1582 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1583 MVT::ValueType VT = N0.getValueType();
1585 // fold (xor c1, c2) -> c1^c2
1587 return DAG.getNode(ISD::XOR, VT, N0, N1);
1588 // canonicalize constant to RHS
1590 return DAG.getNode(ISD::XOR, VT, N1, N0);
1591 // fold (xor x, 0) -> x
1592 if (N1C && N1C->isNullValue())
1595 SDOperand RXOR = ReassociateOps(ISD::XOR, N0, N1);
1598 // fold !(x cc y) -> (x !cc y)
1599 if (N1C && N1C->getValue() == 1 && isSetCCEquivalent(N0, LHS, RHS, CC)) {
1600 bool isInt = MVT::isInteger(LHS.getValueType());
1601 ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
1603 if (N0.getOpcode() == ISD::SETCC)
1604 return DAG.getSetCC(VT, LHS, RHS, NotCC);
1605 if (N0.getOpcode() == ISD::SELECT_CC)
1606 return DAG.getSelectCC(LHS, RHS, N0.getOperand(2),N0.getOperand(3),NotCC);
1607 assert(0 && "Unhandled SetCC Equivalent!");
1610 // fold !(x or y) -> (!x and !y) iff x or y are setcc
1611 if (N1C && N1C->getValue() == 1 &&
1612 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
1613 SDOperand LHS = N0.getOperand(0), RHS = N0.getOperand(1);
1614 if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) {
1615 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
1616 LHS = DAG.getNode(ISD::XOR, VT, LHS, N1); // RHS = ~LHS
1617 RHS = DAG.getNode(ISD::XOR, VT, RHS, N1); // RHS = ~RHS
1618 AddToWorkList(LHS.Val); AddToWorkList(RHS.Val);
1619 return DAG.getNode(NewOpcode, VT, LHS, RHS);
1622 // fold !(x or y) -> (!x and !y) iff x or y are constants
1623 if (N1C && N1C->isAllOnesValue() &&
1624 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
1625 SDOperand LHS = N0.getOperand(0), RHS = N0.getOperand(1);
1626 if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) {
1627 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
1628 LHS = DAG.getNode(ISD::XOR, VT, LHS, N1); // RHS = ~LHS
1629 RHS = DAG.getNode(ISD::XOR, VT, RHS, N1); // RHS = ~RHS
1630 AddToWorkList(LHS.Val); AddToWorkList(RHS.Val);
1631 return DAG.getNode(NewOpcode, VT, LHS, RHS);
1634 // fold (xor (xor x, c1), c2) -> (xor x, c1^c2)
1635 if (N1C && N0.getOpcode() == ISD::XOR) {
1636 ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0));
1637 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
1639 return DAG.getNode(ISD::XOR, VT, N0.getOperand(1),
1640 DAG.getConstant(N1C->getValue()^N00C->getValue(), VT));
1642 return DAG.getNode(ISD::XOR, VT, N0.getOperand(0),
1643 DAG.getConstant(N1C->getValue()^N01C->getValue(), VT));
1645 // fold (xor x, x) -> 0
1647 if (!MVT::isVector(VT)) {
1648 return DAG.getConstant(0, VT);
1649 } else if (!AfterLegalize || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT)) {
1650 // Produce a vector of zeros.
1651 SDOperand El = DAG.getConstant(0, MVT::getVectorBaseType(VT));
1652 std::vector<SDOperand> Ops(MVT::getVectorNumElements(VT), El);
1653 return DAG.getNode(ISD::BUILD_VECTOR, VT, &Ops[0], Ops.size());
1657 // Simplify: xor (op x...), (op y...) -> (op (xor x, y))
1658 if (N0.getOpcode() == N1.getOpcode()) {
1659 SDOperand Tmp = SimplifyBinOpWithSameOpcodeHands(N);
1660 if (Tmp.Val) return Tmp;
1663 // Simplify the expression using non-local knowledge.
1664 if (!MVT::isVector(VT) &&
1665 SimplifyDemandedBits(SDOperand(N, 0)))
1666 return SDOperand(N, 0);
1671 SDOperand DAGCombiner::visitSHL(SDNode *N) {
1672 SDOperand N0 = N->getOperand(0);
1673 SDOperand N1 = N->getOperand(1);
1674 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1675 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1676 MVT::ValueType VT = N0.getValueType();
1677 unsigned OpSizeInBits = MVT::getSizeInBits(VT);
1679 // fold (shl c1, c2) -> c1<<c2
1681 return DAG.getNode(ISD::SHL, VT, N0, N1);
1682 // fold (shl 0, x) -> 0
1683 if (N0C && N0C->isNullValue())
1685 // fold (shl x, c >= size(x)) -> undef
1686 if (N1C && N1C->getValue() >= OpSizeInBits)
1687 return DAG.getNode(ISD::UNDEF, VT);
1688 // fold (shl x, 0) -> x
1689 if (N1C && N1C->isNullValue())
1691 // if (shl x, c) is known to be zero, return 0
1692 if (TLI.MaskedValueIsZero(SDOperand(N, 0), MVT::getIntVTBitMask(VT)))
1693 return DAG.getConstant(0, VT);
1694 if (SimplifyDemandedBits(SDOperand(N, 0)))
1695 return SDOperand(N, 0);
1696 // fold (shl (shl x, c1), c2) -> 0 or (shl x, c1+c2)
1697 if (N1C && N0.getOpcode() == ISD::SHL &&
1698 N0.getOperand(1).getOpcode() == ISD::Constant) {
1699 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
1700 uint64_t c2 = N1C->getValue();
1701 if (c1 + c2 > OpSizeInBits)
1702 return DAG.getConstant(0, VT);
1703 return DAG.getNode(ISD::SHL, VT, N0.getOperand(0),
1704 DAG.getConstant(c1 + c2, N1.getValueType()));
1706 // fold (shl (srl x, c1), c2) -> (shl (and x, -1 << c1), c2-c1) or
1707 // (srl (and x, -1 << c1), c1-c2)
1708 if (N1C && N0.getOpcode() == ISD::SRL &&
1709 N0.getOperand(1).getOpcode() == ISD::Constant) {
1710 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
1711 uint64_t c2 = N1C->getValue();
1712 SDOperand Mask = DAG.getNode(ISD::AND, VT, N0.getOperand(0),
1713 DAG.getConstant(~0ULL << c1, VT));
1715 return DAG.getNode(ISD::SHL, VT, Mask,
1716 DAG.getConstant(c2-c1, N1.getValueType()));
1718 return DAG.getNode(ISD::SRL, VT, Mask,
1719 DAG.getConstant(c1-c2, N1.getValueType()));
1721 // fold (shl (sra x, c1), c1) -> (and x, -1 << c1)
1722 if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1))
1723 return DAG.getNode(ISD::AND, VT, N0.getOperand(0),
1724 DAG.getConstant(~0ULL << N1C->getValue(), VT));
1725 // fold (shl (add x, c1), c2) -> (add (shl x, c2), c1<<c2)
1726 if (N1C && N0.getOpcode() == ISD::ADD && N0.Val->hasOneUse() &&
1727 isa<ConstantSDNode>(N0.getOperand(1))) {
1728 return DAG.getNode(ISD::ADD, VT,
1729 DAG.getNode(ISD::SHL, VT, N0.getOperand(0), N1),
1730 DAG.getNode(ISD::SHL, VT, N0.getOperand(1), N1));
1735 SDOperand DAGCombiner::visitSRA(SDNode *N) {
1736 SDOperand N0 = N->getOperand(0);
1737 SDOperand N1 = N->getOperand(1);
1738 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1739 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1740 MVT::ValueType VT = N0.getValueType();
1742 // fold (sra c1, c2) -> c1>>c2
1744 return DAG.getNode(ISD::SRA, VT, N0, N1);
1745 // fold (sra 0, x) -> 0
1746 if (N0C && N0C->isNullValue())
1748 // fold (sra -1, x) -> -1
1749 if (N0C && N0C->isAllOnesValue())
1751 // fold (sra x, c >= size(x)) -> undef
1752 if (N1C && N1C->getValue() >= MVT::getSizeInBits(VT))
1753 return DAG.getNode(ISD::UNDEF, VT);
1754 // fold (sra x, 0) -> x
1755 if (N1C && N1C->isNullValue())
1757 // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports
1759 if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) {
1760 unsigned LowBits = MVT::getSizeInBits(VT) - (unsigned)N1C->getValue();
1763 default: EVT = MVT::Other; break;
1764 case 1: EVT = MVT::i1; break;
1765 case 8: EVT = MVT::i8; break;
1766 case 16: EVT = MVT::i16; break;
1767 case 32: EVT = MVT::i32; break;
1769 if (EVT > MVT::Other && TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, EVT))
1770 return DAG.getNode(ISD::SIGN_EXTEND_INREG, VT, N0.getOperand(0),
1771 DAG.getValueType(EVT));
1774 // fold (sra (sra x, c1), c2) -> (sra x, c1+c2)
1775 if (N1C && N0.getOpcode() == ISD::SRA) {
1776 if (ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1777 unsigned Sum = N1C->getValue() + C1->getValue();
1778 if (Sum >= MVT::getSizeInBits(VT)) Sum = MVT::getSizeInBits(VT)-1;
1779 return DAG.getNode(ISD::SRA, VT, N0.getOperand(0),
1780 DAG.getConstant(Sum, N1C->getValueType(0)));
1784 // Simplify, based on bits shifted out of the LHS.
1785 if (N1C && SimplifyDemandedBits(SDOperand(N, 0)))
1786 return SDOperand(N, 0);
1789 // If the sign bit is known to be zero, switch this to a SRL.
1790 if (TLI.MaskedValueIsZero(N0, MVT::getIntVTSignBit(VT)))
1791 return DAG.getNode(ISD::SRL, VT, N0, N1);
1795 SDOperand DAGCombiner::visitSRL(SDNode *N) {
1796 SDOperand N0 = N->getOperand(0);
1797 SDOperand N1 = N->getOperand(1);
1798 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1799 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1800 MVT::ValueType VT = N0.getValueType();
1801 unsigned OpSizeInBits = MVT::getSizeInBits(VT);
1803 // fold (srl c1, c2) -> c1 >>u c2
1805 return DAG.getNode(ISD::SRL, VT, N0, N1);
1806 // fold (srl 0, x) -> 0
1807 if (N0C && N0C->isNullValue())
1809 // fold (srl x, c >= size(x)) -> undef
1810 if (N1C && N1C->getValue() >= OpSizeInBits)
1811 return DAG.getNode(ISD::UNDEF, VT);
1812 // fold (srl x, 0) -> x
1813 if (N1C && N1C->isNullValue())
1815 // if (srl x, c) is known to be zero, return 0
1816 if (N1C && TLI.MaskedValueIsZero(SDOperand(N, 0), ~0ULL >> (64-OpSizeInBits)))
1817 return DAG.getConstant(0, VT);
1818 // fold (srl (srl x, c1), c2) -> 0 or (srl x, c1+c2)
1819 if (N1C && N0.getOpcode() == ISD::SRL &&
1820 N0.getOperand(1).getOpcode() == ISD::Constant) {
1821 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
1822 uint64_t c2 = N1C->getValue();
1823 if (c1 + c2 > OpSizeInBits)
1824 return DAG.getConstant(0, VT);
1825 return DAG.getNode(ISD::SRL, VT, N0.getOperand(0),
1826 DAG.getConstant(c1 + c2, N1.getValueType()));
1829 // fold (srl (anyextend x), c) -> (anyextend (srl x, c))
1830 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
1831 // Shifting in all undef bits?
1832 MVT::ValueType SmallVT = N0.getOperand(0).getValueType();
1833 if (N1C->getValue() >= MVT::getSizeInBits(SmallVT))
1834 return DAG.getNode(ISD::UNDEF, VT);
1836 SDOperand SmallShift = DAG.getNode(ISD::SRL, SmallVT, N0.getOperand(0), N1);
1837 AddToWorkList(SmallShift.Val);
1838 return DAG.getNode(ISD::ANY_EXTEND, VT, SmallShift);
1841 // fold (srl (sra X, Y), 31) -> (srl X, 31). This srl only looks at the sign
1842 // bit, which is unmodified by sra.
1843 if (N1C && N1C->getValue()+1 == MVT::getSizeInBits(VT)) {
1844 if (N0.getOpcode() == ISD::SRA)
1845 return DAG.getNode(ISD::SRL, VT, N0.getOperand(0), N1);
1848 // fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit).
1849 if (N1C && N0.getOpcode() == ISD::CTLZ &&
1850 N1C->getValue() == Log2_32(MVT::getSizeInBits(VT))) {
1851 uint64_t KnownZero, KnownOne, Mask = MVT::getIntVTBitMask(VT);
1852 TLI.ComputeMaskedBits(N0.getOperand(0), Mask, KnownZero, KnownOne);
1854 // If any of the input bits are KnownOne, then the input couldn't be all
1855 // zeros, thus the result of the srl will always be zero.
1856 if (KnownOne) return DAG.getConstant(0, VT);
1858 // If all of the bits input the to ctlz node are known to be zero, then
1859 // the result of the ctlz is "32" and the result of the shift is one.
1860 uint64_t UnknownBits = ~KnownZero & Mask;
1861 if (UnknownBits == 0) return DAG.getConstant(1, VT);
1863 // Otherwise, check to see if there is exactly one bit input to the ctlz.
1864 if ((UnknownBits & (UnknownBits-1)) == 0) {
1865 // Okay, we know that only that the single bit specified by UnknownBits
1866 // could be set on input to the CTLZ node. If this bit is set, the SRL
1867 // will return 0, if it is clear, it returns 1. Change the CTLZ/SRL pair
1868 // to an SRL,XOR pair, which is likely to simplify more.
1869 unsigned ShAmt = CountTrailingZeros_64(UnknownBits);
1870 SDOperand Op = N0.getOperand(0);
1872 Op = DAG.getNode(ISD::SRL, VT, Op,
1873 DAG.getConstant(ShAmt, TLI.getShiftAmountTy()));
1874 AddToWorkList(Op.Val);
1876 return DAG.getNode(ISD::XOR, VT, Op, DAG.getConstant(1, VT));
1883 SDOperand DAGCombiner::visitCTLZ(SDNode *N) {
1884 SDOperand N0 = N->getOperand(0);
1885 MVT::ValueType VT = N->getValueType(0);
1887 // fold (ctlz c1) -> c2
1888 if (isa<ConstantSDNode>(N0))
1889 return DAG.getNode(ISD::CTLZ, VT, N0);
1893 SDOperand DAGCombiner::visitCTTZ(SDNode *N) {
1894 SDOperand N0 = N->getOperand(0);
1895 MVT::ValueType VT = N->getValueType(0);
1897 // fold (cttz c1) -> c2
1898 if (isa<ConstantSDNode>(N0))
1899 return DAG.getNode(ISD::CTTZ, VT, N0);
1903 SDOperand DAGCombiner::visitCTPOP(SDNode *N) {
1904 SDOperand N0 = N->getOperand(0);
1905 MVT::ValueType VT = N->getValueType(0);
1907 // fold (ctpop c1) -> c2
1908 if (isa<ConstantSDNode>(N0))
1909 return DAG.getNode(ISD::CTPOP, VT, N0);
1913 SDOperand DAGCombiner::visitSELECT(SDNode *N) {
1914 SDOperand N0 = N->getOperand(0);
1915 SDOperand N1 = N->getOperand(1);
1916 SDOperand N2 = N->getOperand(2);
1917 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1918 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1919 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
1920 MVT::ValueType VT = N->getValueType(0);
1922 // fold select C, X, X -> X
1925 // fold select true, X, Y -> X
1926 if (N0C && !N0C->isNullValue())
1928 // fold select false, X, Y -> Y
1929 if (N0C && N0C->isNullValue())
1931 // fold select C, 1, X -> C | X
1932 if (MVT::i1 == VT && N1C && N1C->getValue() == 1)
1933 return DAG.getNode(ISD::OR, VT, N0, N2);
1934 // fold select C, 0, X -> ~C & X
1935 // FIXME: this should check for C type == X type, not i1?
1936 if (MVT::i1 == VT && N1C && N1C->isNullValue()) {
1937 SDOperand XORNode = DAG.getNode(ISD::XOR, VT, N0, DAG.getConstant(1, VT));
1938 AddToWorkList(XORNode.Val);
1939 return DAG.getNode(ISD::AND, VT, XORNode, N2);
1941 // fold select C, X, 1 -> ~C | X
1942 if (MVT::i1 == VT && N2C && N2C->getValue() == 1) {
1943 SDOperand XORNode = DAG.getNode(ISD::XOR, VT, N0, DAG.getConstant(1, VT));
1944 AddToWorkList(XORNode.Val);
1945 return DAG.getNode(ISD::OR, VT, XORNode, N1);
1947 // fold select C, X, 0 -> C & X
1948 // FIXME: this should check for C type == X type, not i1?
1949 if (MVT::i1 == VT && N2C && N2C->isNullValue())
1950 return DAG.getNode(ISD::AND, VT, N0, N1);
1951 // fold X ? X : Y --> X ? 1 : Y --> X | Y
1952 if (MVT::i1 == VT && N0 == N1)
1953 return DAG.getNode(ISD::OR, VT, N0, N2);
1954 // fold X ? Y : X --> X ? Y : 0 --> X & Y
1955 if (MVT::i1 == VT && N0 == N2)
1956 return DAG.getNode(ISD::AND, VT, N0, N1);
1958 // If we can fold this based on the true/false value, do so.
1959 if (SimplifySelectOps(N, N1, N2))
1960 return SDOperand(N, 0); // Don't revisit N.
1962 // fold selects based on a setcc into other things, such as min/max/abs
1963 if (N0.getOpcode() == ISD::SETCC)
1965 // Check against MVT::Other for SELECT_CC, which is a workaround for targets
1966 // having to say they don't support SELECT_CC on every type the DAG knows
1967 // about, since there is no way to mark an opcode illegal at all value types
1968 if (TLI.isOperationLegal(ISD::SELECT_CC, MVT::Other))
1969 return DAG.getNode(ISD::SELECT_CC, VT, N0.getOperand(0), N0.getOperand(1),
1970 N1, N2, N0.getOperand(2));
1972 return SimplifySelect(N0, N1, N2);
1976 SDOperand DAGCombiner::visitSELECT_CC(SDNode *N) {
1977 SDOperand N0 = N->getOperand(0);
1978 SDOperand N1 = N->getOperand(1);
1979 SDOperand N2 = N->getOperand(2);
1980 SDOperand N3 = N->getOperand(3);
1981 SDOperand N4 = N->getOperand(4);
1982 ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get();
1984 // fold select_cc lhs, rhs, x, x, cc -> x
1988 // Determine if the condition we're dealing with is constant
1989 SDOperand SCC = SimplifySetCC(TLI.getSetCCResultTy(), N0, N1, CC, false);
1990 if (SCC.Val) AddToWorkList(SCC.Val);
1992 if (ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.Val)) {
1993 if (SCCC->getValue())
1994 return N2; // cond always true -> true val
1996 return N3; // cond always false -> false val
1999 // Fold to a simpler select_cc
2000 if (SCC.Val && SCC.getOpcode() == ISD::SETCC)
2001 return DAG.getNode(ISD::SELECT_CC, N2.getValueType(),
2002 SCC.getOperand(0), SCC.getOperand(1), N2, N3,
2005 // If we can fold this based on the true/false value, do so.
2006 if (SimplifySelectOps(N, N2, N3))
2007 return SDOperand(N, 0); // Don't revisit N.
2009 // fold select_cc into other things, such as min/max/abs
2010 return SimplifySelectCC(N0, N1, N2, N3, CC);
2013 SDOperand DAGCombiner::visitSETCC(SDNode *N) {
2014 return SimplifySetCC(N->getValueType(0), N->getOperand(0), N->getOperand(1),
2015 cast<CondCodeSDNode>(N->getOperand(2))->get());
2018 SDOperand DAGCombiner::visitSIGN_EXTEND(SDNode *N) {
2019 SDOperand N0 = N->getOperand(0);
2020 MVT::ValueType VT = N->getValueType(0);
2022 // fold (sext c1) -> c1
2023 if (isa<ConstantSDNode>(N0))
2024 return DAG.getNode(ISD::SIGN_EXTEND, VT, N0);
2026 // fold (sext (sext x)) -> (sext x)
2027 // fold (sext (aext x)) -> (sext x)
2028 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
2029 return DAG.getNode(ISD::SIGN_EXTEND, VT, N0.getOperand(0));
2031 // fold (sext (truncate x)) -> (sextinreg x).
2032 if (N0.getOpcode() == ISD::TRUNCATE &&
2033 (!AfterLegalize || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG,
2034 N0.getValueType()))) {
2035 SDOperand Op = N0.getOperand(0);
2036 if (Op.getValueType() < VT) {
2037 Op = DAG.getNode(ISD::ANY_EXTEND, VT, Op);
2038 } else if (Op.getValueType() > VT) {
2039 Op = DAG.getNode(ISD::TRUNCATE, VT, Op);
2041 return DAG.getNode(ISD::SIGN_EXTEND_INREG, VT, Op,
2042 DAG.getValueType(N0.getValueType()));
2045 // fold (sext (load x)) -> (sext (truncate (sextload x)))
2046 if (ISD::isNON_EXTLoad(N0.Val) && N0.hasOneUse() &&
2047 (!AfterLegalize||TLI.isLoadXLegal(ISD::SEXTLOAD, N0.getValueType()))){
2048 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2049 SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, LN0->getChain(),
2050 LN0->getBasePtr(), LN0->getSrcValue(),
2051 LN0->getSrcValueOffset(),
2053 CombineTo(N, ExtLoad);
2054 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
2055 ExtLoad.getValue(1));
2056 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
2059 // fold (sext (sextload x)) -> (sext (truncate (sextload x)))
2060 // fold (sext ( extload x)) -> (sext (truncate (sextload x)))
2061 if ((ISD::isSEXTLoad(N0.Val) || ISD::isEXTLoad(N0.Val)) && N0.hasOneUse()) {
2062 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2063 MVT::ValueType EVT = LN0->getLoadedVT();
2064 SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, LN0->getChain(),
2065 LN0->getBasePtr(), LN0->getSrcValue(),
2066 LN0->getSrcValueOffset(), EVT);
2067 CombineTo(N, ExtLoad);
2068 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
2069 ExtLoad.getValue(1));
2070 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
2076 SDOperand DAGCombiner::visitZERO_EXTEND(SDNode *N) {
2077 SDOperand N0 = N->getOperand(0);
2078 MVT::ValueType VT = N->getValueType(0);
2080 // fold (zext c1) -> c1
2081 if (isa<ConstantSDNode>(N0))
2082 return DAG.getNode(ISD::ZERO_EXTEND, VT, N0);
2083 // fold (zext (zext x)) -> (zext x)
2084 // fold (zext (aext x)) -> (zext x)
2085 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
2086 return DAG.getNode(ISD::ZERO_EXTEND, VT, N0.getOperand(0));
2088 // fold (zext (truncate x)) -> (and x, mask)
2089 if (N0.getOpcode() == ISD::TRUNCATE &&
2090 (!AfterLegalize || TLI.isOperationLegal(ISD::AND, VT))) {
2091 SDOperand Op = N0.getOperand(0);
2092 if (Op.getValueType() < VT) {
2093 Op = DAG.getNode(ISD::ANY_EXTEND, VT, Op);
2094 } else if (Op.getValueType() > VT) {
2095 Op = DAG.getNode(ISD::TRUNCATE, VT, Op);
2097 return DAG.getZeroExtendInReg(Op, N0.getValueType());
2100 // fold (zext (and (trunc x), cst)) -> (and x, cst).
2101 if (N0.getOpcode() == ISD::AND &&
2102 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
2103 N0.getOperand(1).getOpcode() == ISD::Constant) {
2104 SDOperand X = N0.getOperand(0).getOperand(0);
2105 if (X.getValueType() < VT) {
2106 X = DAG.getNode(ISD::ANY_EXTEND, VT, X);
2107 } else if (X.getValueType() > VT) {
2108 X = DAG.getNode(ISD::TRUNCATE, VT, X);
2110 uint64_t Mask = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
2111 return DAG.getNode(ISD::AND, VT, X, DAG.getConstant(Mask, VT));
2114 // fold (zext (load x)) -> (zext (truncate (zextload x)))
2115 if (ISD::isNON_EXTLoad(N0.Val) && N0.hasOneUse() &&
2116 (!AfterLegalize||TLI.isLoadXLegal(ISD::ZEXTLOAD, N0.getValueType()))) {
2117 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2118 SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, LN0->getChain(),
2119 LN0->getBasePtr(), LN0->getSrcValue(),
2120 LN0->getSrcValueOffset(),
2122 CombineTo(N, ExtLoad);
2123 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
2124 ExtLoad.getValue(1));
2125 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
2128 // fold (zext (zextload x)) -> (zext (truncate (zextload x)))
2129 // fold (zext ( extload x)) -> (zext (truncate (zextload x)))
2130 if ((ISD::isZEXTLoad(N0.Val) || ISD::isEXTLoad(N0.Val)) && N0.hasOneUse()) {
2131 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2132 MVT::ValueType EVT = LN0->getLoadedVT();
2133 SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, LN0->getChain(),
2134 LN0->getBasePtr(), LN0->getSrcValue(),
2135 LN0->getSrcValueOffset(), EVT);
2136 CombineTo(N, ExtLoad);
2137 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
2138 ExtLoad.getValue(1));
2139 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
2144 SDOperand DAGCombiner::visitANY_EXTEND(SDNode *N) {
2145 SDOperand N0 = N->getOperand(0);
2146 MVT::ValueType VT = N->getValueType(0);
2148 // fold (aext c1) -> c1
2149 if (isa<ConstantSDNode>(N0))
2150 return DAG.getNode(ISD::ANY_EXTEND, VT, N0);
2151 // fold (aext (aext x)) -> (aext x)
2152 // fold (aext (zext x)) -> (zext x)
2153 // fold (aext (sext x)) -> (sext x)
2154 if (N0.getOpcode() == ISD::ANY_EXTEND ||
2155 N0.getOpcode() == ISD::ZERO_EXTEND ||
2156 N0.getOpcode() == ISD::SIGN_EXTEND)
2157 return DAG.getNode(N0.getOpcode(), VT, N0.getOperand(0));
2159 // fold (aext (truncate x))
2160 if (N0.getOpcode() == ISD::TRUNCATE) {
2161 SDOperand TruncOp = N0.getOperand(0);
2162 if (TruncOp.getValueType() == VT)
2163 return TruncOp; // x iff x size == zext size.
2164 if (TruncOp.getValueType() > VT)
2165 return DAG.getNode(ISD::TRUNCATE, VT, TruncOp);
2166 return DAG.getNode(ISD::ANY_EXTEND, VT, TruncOp);
2169 // fold (aext (and (trunc x), cst)) -> (and x, cst).
2170 if (N0.getOpcode() == ISD::AND &&
2171 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
2172 N0.getOperand(1).getOpcode() == ISD::Constant) {
2173 SDOperand X = N0.getOperand(0).getOperand(0);
2174 if (X.getValueType() < VT) {
2175 X = DAG.getNode(ISD::ANY_EXTEND, VT, X);
2176 } else if (X.getValueType() > VT) {
2177 X = DAG.getNode(ISD::TRUNCATE, VT, X);
2179 uint64_t Mask = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
2180 return DAG.getNode(ISD::AND, VT, X, DAG.getConstant(Mask, VT));
2183 // fold (aext (load x)) -> (aext (truncate (extload x)))
2184 if (ISD::isNON_EXTLoad(N0.Val) && N0.hasOneUse() &&
2185 (!AfterLegalize||TLI.isLoadXLegal(ISD::EXTLOAD, N0.getValueType()))) {
2186 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2187 SDOperand ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, VT, LN0->getChain(),
2188 LN0->getBasePtr(), LN0->getSrcValue(),
2189 LN0->getSrcValueOffset(),
2191 CombineTo(N, ExtLoad);
2192 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
2193 ExtLoad.getValue(1));
2194 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
2197 // fold (aext (zextload x)) -> (aext (truncate (zextload x)))
2198 // fold (aext (sextload x)) -> (aext (truncate (sextload x)))
2199 // fold (aext ( extload x)) -> (aext (truncate (extload x)))
2200 if (N0.getOpcode() == ISD::LOAD && !ISD::isNON_EXTLoad(N0.Val) &&
2202 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2203 MVT::ValueType EVT = LN0->getLoadedVT();
2204 SDOperand ExtLoad = DAG.getExtLoad(LN0->getExtensionType(), VT,
2205 LN0->getChain(), LN0->getBasePtr(),
2207 LN0->getSrcValueOffset(), EVT);
2208 CombineTo(N, ExtLoad);
2209 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
2210 ExtLoad.getValue(1));
2211 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
2217 SDOperand DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
2218 SDOperand N0 = N->getOperand(0);
2219 SDOperand N1 = N->getOperand(1);
2220 MVT::ValueType VT = N->getValueType(0);
2221 MVT::ValueType EVT = cast<VTSDNode>(N1)->getVT();
2222 unsigned EVTBits = MVT::getSizeInBits(EVT);
2224 // fold (sext_in_reg c1) -> c1
2225 if (isa<ConstantSDNode>(N0) || N0.getOpcode() == ISD::UNDEF)
2226 return DAG.getNode(ISD::SIGN_EXTEND_INREG, VT, N0, N1);
2228 // If the input is already sign extended, just drop the extension.
2229 if (TLI.ComputeNumSignBits(N0) >= MVT::getSizeInBits(VT)-EVTBits+1)
2232 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2
2233 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
2234 EVT < cast<VTSDNode>(N0.getOperand(1))->getVT()) {
2235 return DAG.getNode(ISD::SIGN_EXTEND_INREG, VT, N0.getOperand(0), N1);
2238 // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is zero
2239 if (TLI.MaskedValueIsZero(N0, 1ULL << (EVTBits-1)))
2240 return DAG.getZeroExtendInReg(N0, EVT);
2242 // fold (sext_in_reg (srl X, 24), i8) -> sra X, 24
2243 // fold (sext_in_reg (srl X, 23), i8) -> sra X, 23 iff possible.
2244 // We already fold "(sext_in_reg (srl X, 25), i8) -> srl X, 25" above.
2245 if (N0.getOpcode() == ISD::SRL) {
2246 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
2247 if (ShAmt->getValue()+EVTBits <= MVT::getSizeInBits(VT)) {
2248 // We can turn this into an SRA iff the input to the SRL is already sign
2250 unsigned InSignBits = TLI.ComputeNumSignBits(N0.getOperand(0));
2251 if (MVT::getSizeInBits(VT)-(ShAmt->getValue()+EVTBits) < InSignBits)
2252 return DAG.getNode(ISD::SRA, VT, N0.getOperand(0), N0.getOperand(1));
2256 // fold (sext_inreg (extload x)) -> (sextload x)
2257 if (ISD::isEXTLoad(N0.Val) &&
2258 EVT == cast<LoadSDNode>(N0)->getLoadedVT() &&
2259 (!AfterLegalize || TLI.isLoadXLegal(ISD::SEXTLOAD, EVT))) {
2260 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2261 SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, LN0->getChain(),
2262 LN0->getBasePtr(), LN0->getSrcValue(),
2263 LN0->getSrcValueOffset(), EVT);
2264 CombineTo(N, ExtLoad);
2265 CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1));
2266 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
2268 // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use
2269 if (ISD::isZEXTLoad(N0.Val) && N0.hasOneUse() &&
2270 EVT == cast<LoadSDNode>(N0)->getLoadedVT() &&
2271 (!AfterLegalize || TLI.isLoadXLegal(ISD::SEXTLOAD, EVT))) {
2272 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2273 SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, LN0->getChain(),
2274 LN0->getBasePtr(), LN0->getSrcValue(),
2275 LN0->getSrcValueOffset(), EVT);
2276 CombineTo(N, ExtLoad);
2277 CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1));
2278 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
2283 SDOperand DAGCombiner::visitTRUNCATE(SDNode *N) {
2284 SDOperand N0 = N->getOperand(0);
2285 MVT::ValueType VT = N->getValueType(0);
2288 if (N0.getValueType() == N->getValueType(0))
2290 // fold (truncate c1) -> c1
2291 if (isa<ConstantSDNode>(N0))
2292 return DAG.getNode(ISD::TRUNCATE, VT, N0);
2293 // fold (truncate (truncate x)) -> (truncate x)
2294 if (N0.getOpcode() == ISD::TRUNCATE)
2295 return DAG.getNode(ISD::TRUNCATE, VT, N0.getOperand(0));
2296 // fold (truncate (ext x)) -> (ext x) or (truncate x) or x
2297 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::SIGN_EXTEND||
2298 N0.getOpcode() == ISD::ANY_EXTEND) {
2299 if (N0.getValueType() < VT)
2300 // if the source is smaller than the dest, we still need an extend
2301 return DAG.getNode(N0.getOpcode(), VT, N0.getOperand(0));
2302 else if (N0.getValueType() > VT)
2303 // if the source is larger than the dest, than we just need the truncate
2304 return DAG.getNode(ISD::TRUNCATE, VT, N0.getOperand(0));
2306 // if the source and dest are the same type, we can drop both the extend
2308 return N0.getOperand(0);
2310 // fold (truncate (load x)) -> (smaller load x)
2311 if (ISD::isNON_EXTLoad(N0.Val) && N0.hasOneUse()) {
2312 assert(MVT::getSizeInBits(N0.getValueType()) > MVT::getSizeInBits(VT) &&
2313 "Cannot truncate to larger type!");
2314 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2315 MVT::ValueType PtrType = N0.getOperand(1).getValueType();
2316 // For big endian targets, we need to add an offset to the pointer to load
2317 // the correct bytes. For little endian systems, we merely need to read
2318 // fewer bytes from the same pointer.
2320 (MVT::getSizeInBits(N0.getValueType()) - MVT::getSizeInBits(VT)) / 8;
2321 SDOperand NewPtr = TLI.isLittleEndian() ? LN0->getBasePtr() :
2322 DAG.getNode(ISD::ADD, PtrType, LN0->getBasePtr(),
2323 DAG.getConstant(PtrOff, PtrType));
2324 AddToWorkList(NewPtr.Val);
2325 SDOperand Load = DAG.getLoad(VT, LN0->getChain(), NewPtr,
2326 LN0->getSrcValue(), LN0->getSrcValueOffset());
2328 CombineTo(N0.Val, Load, Load.getValue(1));
2329 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
2334 SDOperand DAGCombiner::visitBIT_CONVERT(SDNode *N) {
2335 SDOperand N0 = N->getOperand(0);
2336 MVT::ValueType VT = N->getValueType(0);
2338 // If the input is a constant, let getNode() fold it.
2339 if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) {
2340 SDOperand Res = DAG.getNode(ISD::BIT_CONVERT, VT, N0);
2341 if (Res.Val != N) return Res;
2344 if (N0.getOpcode() == ISD::BIT_CONVERT) // conv(conv(x,t1),t2) -> conv(x,t2)
2345 return DAG.getNode(ISD::BIT_CONVERT, VT, N0.getOperand(0));
2347 // fold (conv (load x)) -> (load (conv*)x)
2348 // FIXME: These xforms need to know that the resultant load doesn't need a
2349 // higher alignment than the original!
2350 if (0 && ISD::isNON_EXTLoad(N0.Val) && N0.hasOneUse()) {
2351 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2352 SDOperand Load = DAG.getLoad(VT, LN0->getChain(), LN0->getBasePtr(),
2353 LN0->getSrcValue(), LN0->getSrcValueOffset());
2355 CombineTo(N0.Val, DAG.getNode(ISD::BIT_CONVERT, N0.getValueType(), Load),
2363 SDOperand DAGCombiner::visitVBIT_CONVERT(SDNode *N) {
2364 SDOperand N0 = N->getOperand(0);
2365 MVT::ValueType VT = N->getValueType(0);
2367 // If the input is a VBUILD_VECTOR with all constant elements, fold this now.
2368 // First check to see if this is all constant.
2369 if (N0.getOpcode() == ISD::VBUILD_VECTOR && N0.Val->hasOneUse() &&
2370 VT == MVT::Vector) {
2371 bool isSimple = true;
2372 for (unsigned i = 0, e = N0.getNumOperands()-2; i != e; ++i)
2373 if (N0.getOperand(i).getOpcode() != ISD::UNDEF &&
2374 N0.getOperand(i).getOpcode() != ISD::Constant &&
2375 N0.getOperand(i).getOpcode() != ISD::ConstantFP) {
2380 MVT::ValueType DestEltVT = cast<VTSDNode>(N->getOperand(2))->getVT();
2381 if (isSimple && !MVT::isVector(DestEltVT)) {
2382 return ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(N0.Val, DestEltVT);
2389 /// ConstantFoldVBIT_CONVERTofVBUILD_VECTOR - We know that BV is a vbuild_vector
2390 /// node with Constant, ConstantFP or Undef operands. DstEltVT indicates the
2391 /// destination element value type.
2392 SDOperand DAGCombiner::
2393 ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(SDNode *BV, MVT::ValueType DstEltVT) {
2394 MVT::ValueType SrcEltVT = BV->getOperand(0).getValueType();
2396 // If this is already the right type, we're done.
2397 if (SrcEltVT == DstEltVT) return SDOperand(BV, 0);
2399 unsigned SrcBitSize = MVT::getSizeInBits(SrcEltVT);
2400 unsigned DstBitSize = MVT::getSizeInBits(DstEltVT);
2402 // If this is a conversion of N elements of one type to N elements of another
2403 // type, convert each element. This handles FP<->INT cases.
2404 if (SrcBitSize == DstBitSize) {
2405 SmallVector<SDOperand, 8> Ops;
2406 for (unsigned i = 0, e = BV->getNumOperands()-2; i != e; ++i) {
2407 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, DstEltVT, BV->getOperand(i)));
2408 AddToWorkList(Ops.back().Val);
2410 Ops.push_back(*(BV->op_end()-2)); // Add num elements.
2411 Ops.push_back(DAG.getValueType(DstEltVT));
2412 return DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, &Ops[0], Ops.size());
2415 // Otherwise, we're growing or shrinking the elements. To avoid having to
2416 // handle annoying details of growing/shrinking FP values, we convert them to
2418 if (MVT::isFloatingPoint(SrcEltVT)) {
2419 // Convert the input float vector to a int vector where the elements are the
2421 assert((SrcEltVT == MVT::f32 || SrcEltVT == MVT::f64) && "Unknown FP VT!");
2422 MVT::ValueType IntVT = SrcEltVT == MVT::f32 ? MVT::i32 : MVT::i64;
2423 BV = ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(BV, IntVT).Val;
2427 // Now we know the input is an integer vector. If the output is a FP type,
2428 // convert to integer first, then to FP of the right size.
2429 if (MVT::isFloatingPoint(DstEltVT)) {
2430 assert((DstEltVT == MVT::f32 || DstEltVT == MVT::f64) && "Unknown FP VT!");
2431 MVT::ValueType TmpVT = DstEltVT == MVT::f32 ? MVT::i32 : MVT::i64;
2432 SDNode *Tmp = ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(BV, TmpVT).Val;
2434 // Next, convert to FP elements of the same size.
2435 return ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(Tmp, DstEltVT);
2438 // Okay, we know the src/dst types are both integers of differing types.
2439 // Handling growing first.
2440 assert(MVT::isInteger(SrcEltVT) && MVT::isInteger(DstEltVT));
2441 if (SrcBitSize < DstBitSize) {
2442 unsigned NumInputsPerOutput = DstBitSize/SrcBitSize;
2444 SmallVector<SDOperand, 8> Ops;
2445 for (unsigned i = 0, e = BV->getNumOperands()-2; i != e;
2446 i += NumInputsPerOutput) {
2447 bool isLE = TLI.isLittleEndian();
2448 uint64_t NewBits = 0;
2449 bool EltIsUndef = true;
2450 for (unsigned j = 0; j != NumInputsPerOutput; ++j) {
2451 // Shift the previously computed bits over.
2452 NewBits <<= SrcBitSize;
2453 SDOperand Op = BV->getOperand(i+ (isLE ? (NumInputsPerOutput-j-1) : j));
2454 if (Op.getOpcode() == ISD::UNDEF) continue;
2457 NewBits |= cast<ConstantSDNode>(Op)->getValue();
2461 Ops.push_back(DAG.getNode(ISD::UNDEF, DstEltVT));
2463 Ops.push_back(DAG.getConstant(NewBits, DstEltVT));
2466 Ops.push_back(DAG.getConstant(Ops.size(), MVT::i32)); // Add num elements.
2467 Ops.push_back(DAG.getValueType(DstEltVT)); // Add element size.
2468 return DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, &Ops[0], Ops.size());
2471 // Finally, this must be the case where we are shrinking elements: each input
2472 // turns into multiple outputs.
2473 unsigned NumOutputsPerInput = SrcBitSize/DstBitSize;
2474 SmallVector<SDOperand, 8> Ops;
2475 for (unsigned i = 0, e = BV->getNumOperands()-2; i != e; ++i) {
2476 if (BV->getOperand(i).getOpcode() == ISD::UNDEF) {
2477 for (unsigned j = 0; j != NumOutputsPerInput; ++j)
2478 Ops.push_back(DAG.getNode(ISD::UNDEF, DstEltVT));
2481 uint64_t OpVal = cast<ConstantSDNode>(BV->getOperand(i))->getValue();
2483 for (unsigned j = 0; j != NumOutputsPerInput; ++j) {
2484 unsigned ThisVal = OpVal & ((1ULL << DstBitSize)-1);
2485 OpVal >>= DstBitSize;
2486 Ops.push_back(DAG.getConstant(ThisVal, DstEltVT));
2489 // For big endian targets, swap the order of the pieces of each element.
2490 if (!TLI.isLittleEndian())
2491 std::reverse(Ops.end()-NumOutputsPerInput, Ops.end());
2493 Ops.push_back(DAG.getConstant(Ops.size(), MVT::i32)); // Add num elements.
2494 Ops.push_back(DAG.getValueType(DstEltVT)); // Add element size.
2495 return DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, &Ops[0], Ops.size());
2500 SDOperand DAGCombiner::visitFADD(SDNode *N) {
2501 SDOperand N0 = N->getOperand(0);
2502 SDOperand N1 = N->getOperand(1);
2503 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2504 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
2505 MVT::ValueType VT = N->getValueType(0);
2507 // fold (fadd c1, c2) -> c1+c2
2509 return DAG.getNode(ISD::FADD, VT, N0, N1);
2510 // canonicalize constant to RHS
2511 if (N0CFP && !N1CFP)
2512 return DAG.getNode(ISD::FADD, VT, N1, N0);
2513 // fold (A + (-B)) -> A-B
2514 if (N1.getOpcode() == ISD::FNEG)
2515 return DAG.getNode(ISD::FSUB, VT, N0, N1.getOperand(0));
2516 // fold ((-A) + B) -> B-A
2517 if (N0.getOpcode() == ISD::FNEG)
2518 return DAG.getNode(ISD::FSUB, VT, N1, N0.getOperand(0));
2522 SDOperand DAGCombiner::visitFSUB(SDNode *N) {
2523 SDOperand N0 = N->getOperand(0);
2524 SDOperand N1 = N->getOperand(1);
2525 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2526 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
2527 MVT::ValueType VT = N->getValueType(0);
2529 // fold (fsub c1, c2) -> c1-c2
2531 return DAG.getNode(ISD::FSUB, VT, N0, N1);
2532 // fold (A-(-B)) -> A+B
2533 if (N1.getOpcode() == ISD::FNEG)
2534 return DAG.getNode(ISD::FADD, VT, N0, N1.getOperand(0));
2538 SDOperand DAGCombiner::visitFMUL(SDNode *N) {
2539 SDOperand N0 = N->getOperand(0);
2540 SDOperand N1 = N->getOperand(1);
2541 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2542 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
2543 MVT::ValueType VT = N->getValueType(0);
2545 // fold (fmul c1, c2) -> c1*c2
2547 return DAG.getNode(ISD::FMUL, VT, N0, N1);
2548 // canonicalize constant to RHS
2549 if (N0CFP && !N1CFP)
2550 return DAG.getNode(ISD::FMUL, VT, N1, N0);
2551 // fold (fmul X, 2.0) -> (fadd X, X)
2552 if (N1CFP && N1CFP->isExactlyValue(+2.0))
2553 return DAG.getNode(ISD::FADD, VT, N0, N0);
2557 SDOperand DAGCombiner::visitFDIV(SDNode *N) {
2558 SDOperand N0 = N->getOperand(0);
2559 SDOperand N1 = N->getOperand(1);
2560 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2561 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
2562 MVT::ValueType VT = N->getValueType(0);
2564 // fold (fdiv c1, c2) -> c1/c2
2566 return DAG.getNode(ISD::FDIV, VT, N0, N1);
2570 SDOperand DAGCombiner::visitFREM(SDNode *N) {
2571 SDOperand N0 = N->getOperand(0);
2572 SDOperand N1 = N->getOperand(1);
2573 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2574 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
2575 MVT::ValueType VT = N->getValueType(0);
2577 // fold (frem c1, c2) -> fmod(c1,c2)
2579 return DAG.getNode(ISD::FREM, VT, N0, N1);
2583 SDOperand DAGCombiner::visitFCOPYSIGN(SDNode *N) {
2584 SDOperand N0 = N->getOperand(0);
2585 SDOperand N1 = N->getOperand(1);
2586 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2587 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
2588 MVT::ValueType VT = N->getValueType(0);
2590 if (N0CFP && N1CFP) // Constant fold
2591 return DAG.getNode(ISD::FCOPYSIGN, VT, N0, N1);
2594 // copysign(x, c1) -> fabs(x) iff ispos(c1)
2595 // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1)
2600 u.d = N1CFP->getValue();
2602 return DAG.getNode(ISD::FABS, VT, N0);
2604 return DAG.getNode(ISD::FNEG, VT, DAG.getNode(ISD::FABS, VT, N0));
2607 // copysign(fabs(x), y) -> copysign(x, y)
2608 // copysign(fneg(x), y) -> copysign(x, y)
2609 // copysign(copysign(x,z), y) -> copysign(x, y)
2610 if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG ||
2611 N0.getOpcode() == ISD::FCOPYSIGN)
2612 return DAG.getNode(ISD::FCOPYSIGN, VT, N0.getOperand(0), N1);
2614 // copysign(x, abs(y)) -> abs(x)
2615 if (N1.getOpcode() == ISD::FABS)
2616 return DAG.getNode(ISD::FABS, VT, N0);
2618 // copysign(x, copysign(y,z)) -> copysign(x, z)
2619 if (N1.getOpcode() == ISD::FCOPYSIGN)
2620 return DAG.getNode(ISD::FCOPYSIGN, VT, N0, N1.getOperand(1));
2622 // copysign(x, fp_extend(y)) -> copysign(x, y)
2623 // copysign(x, fp_round(y)) -> copysign(x, y)
2624 if (N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND)
2625 return DAG.getNode(ISD::FCOPYSIGN, VT, N0, N1.getOperand(0));
2632 SDOperand DAGCombiner::visitSINT_TO_FP(SDNode *N) {
2633 SDOperand N0 = N->getOperand(0);
2634 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2635 MVT::ValueType VT = N->getValueType(0);
2637 // fold (sint_to_fp c1) -> c1fp
2639 return DAG.getNode(ISD::SINT_TO_FP, VT, N0);
2643 SDOperand DAGCombiner::visitUINT_TO_FP(SDNode *N) {
2644 SDOperand N0 = N->getOperand(0);
2645 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2646 MVT::ValueType VT = N->getValueType(0);
2648 // fold (uint_to_fp c1) -> c1fp
2650 return DAG.getNode(ISD::UINT_TO_FP, VT, N0);
2654 SDOperand DAGCombiner::visitFP_TO_SINT(SDNode *N) {
2655 SDOperand N0 = N->getOperand(0);
2656 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2657 MVT::ValueType VT = N->getValueType(0);
2659 // fold (fp_to_sint c1fp) -> c1
2661 return DAG.getNode(ISD::FP_TO_SINT, VT, N0);
2665 SDOperand DAGCombiner::visitFP_TO_UINT(SDNode *N) {
2666 SDOperand N0 = N->getOperand(0);
2667 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2668 MVT::ValueType VT = N->getValueType(0);
2670 // fold (fp_to_uint c1fp) -> c1
2672 return DAG.getNode(ISD::FP_TO_UINT, VT, N0);
2676 SDOperand DAGCombiner::visitFP_ROUND(SDNode *N) {
2677 SDOperand N0 = N->getOperand(0);
2678 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2679 MVT::ValueType VT = N->getValueType(0);
2681 // fold (fp_round c1fp) -> c1fp
2683 return DAG.getNode(ISD::FP_ROUND, VT, N0);
2685 // fold (fp_round (fp_extend x)) -> x
2686 if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType())
2687 return N0.getOperand(0);
2689 // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y)
2690 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.Val->hasOneUse()) {
2691 SDOperand Tmp = DAG.getNode(ISD::FP_ROUND, VT, N0.getOperand(0));
2692 AddToWorkList(Tmp.Val);
2693 return DAG.getNode(ISD::FCOPYSIGN, VT, Tmp, N0.getOperand(1));
2699 SDOperand DAGCombiner::visitFP_ROUND_INREG(SDNode *N) {
2700 SDOperand N0 = N->getOperand(0);
2701 MVT::ValueType VT = N->getValueType(0);
2702 MVT::ValueType EVT = cast<VTSDNode>(N->getOperand(1))->getVT();
2703 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2705 // fold (fp_round_inreg c1fp) -> c1fp
2707 SDOperand Round = DAG.getConstantFP(N0CFP->getValue(), EVT);
2708 return DAG.getNode(ISD::FP_EXTEND, VT, Round);
2713 SDOperand DAGCombiner::visitFP_EXTEND(SDNode *N) {
2714 SDOperand N0 = N->getOperand(0);
2715 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2716 MVT::ValueType VT = N->getValueType(0);
2718 // fold (fp_extend c1fp) -> c1fp
2720 return DAG.getNode(ISD::FP_EXTEND, VT, N0);
2722 // fold (fpext (load x)) -> (fpext (fpround (extload x)))
2723 if (ISD::isNON_EXTLoad(N0.Val) && N0.hasOneUse() &&
2724 (!AfterLegalize||TLI.isLoadXLegal(ISD::EXTLOAD, N0.getValueType()))) {
2725 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2726 SDOperand ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, VT, LN0->getChain(),
2727 LN0->getBasePtr(), LN0->getSrcValue(),
2728 LN0->getSrcValueOffset(),
2730 CombineTo(N, ExtLoad);
2731 CombineTo(N0.Val, DAG.getNode(ISD::FP_ROUND, N0.getValueType(), ExtLoad),
2732 ExtLoad.getValue(1));
2733 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
2740 SDOperand DAGCombiner::visitFNEG(SDNode *N) {
2741 SDOperand N0 = N->getOperand(0);
2742 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2743 MVT::ValueType VT = N->getValueType(0);
2745 // fold (fneg c1) -> -c1
2747 return DAG.getNode(ISD::FNEG, VT, N0);
2748 // fold (fneg (sub x, y)) -> (sub y, x)
2749 if (N0.getOpcode() == ISD::SUB)
2750 return DAG.getNode(ISD::SUB, VT, N0.getOperand(1), N0.getOperand(0));
2751 // fold (fneg (fneg x)) -> x
2752 if (N0.getOpcode() == ISD::FNEG)
2753 return N0.getOperand(0);
2757 SDOperand DAGCombiner::visitFABS(SDNode *N) {
2758 SDOperand N0 = N->getOperand(0);
2759 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2760 MVT::ValueType VT = N->getValueType(0);
2762 // fold (fabs c1) -> fabs(c1)
2764 return DAG.getNode(ISD::FABS, VT, N0);
2765 // fold (fabs (fabs x)) -> (fabs x)
2766 if (N0.getOpcode() == ISD::FABS)
2767 return N->getOperand(0);
2768 // fold (fabs (fneg x)) -> (fabs x)
2769 // fold (fabs (fcopysign x, y)) -> (fabs x)
2770 if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN)
2771 return DAG.getNode(ISD::FABS, VT, N0.getOperand(0));
2776 SDOperand DAGCombiner::visitBRCOND(SDNode *N) {
2777 SDOperand Chain = N->getOperand(0);
2778 SDOperand N1 = N->getOperand(1);
2779 SDOperand N2 = N->getOperand(2);
2780 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2782 // never taken branch, fold to chain
2783 if (N1C && N1C->isNullValue())
2785 // unconditional branch
2786 if (N1C && N1C->getValue() == 1)
2787 return DAG.getNode(ISD::BR, MVT::Other, Chain, N2);
2788 // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal
2790 if (N1.getOpcode() == ISD::SETCC &&
2791 TLI.isOperationLegal(ISD::BR_CC, MVT::Other)) {
2792 return DAG.getNode(ISD::BR_CC, MVT::Other, Chain, N1.getOperand(2),
2793 N1.getOperand(0), N1.getOperand(1), N2);
2798 // Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB.
2800 SDOperand DAGCombiner::visitBR_CC(SDNode *N) {
2801 CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1));
2802 SDOperand CondLHS = N->getOperand(2), CondRHS = N->getOperand(3);
2804 // Use SimplifySetCC to simplify SETCC's.
2805 SDOperand Simp = SimplifySetCC(MVT::i1, CondLHS, CondRHS, CC->get(), false);
2806 if (Simp.Val) AddToWorkList(Simp.Val);
2808 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(Simp.Val);
2810 // fold br_cc true, dest -> br dest (unconditional branch)
2811 if (SCCC && SCCC->getValue())
2812 return DAG.getNode(ISD::BR, MVT::Other, N->getOperand(0),
2814 // fold br_cc false, dest -> unconditional fall through
2815 if (SCCC && SCCC->isNullValue())
2816 return N->getOperand(0);
2818 // fold to a simpler setcc
2819 if (Simp.Val && Simp.getOpcode() == ISD::SETCC)
2820 return DAG.getNode(ISD::BR_CC, MVT::Other, N->getOperand(0),
2821 Simp.getOperand(2), Simp.getOperand(0),
2822 Simp.getOperand(1), N->getOperand(4));
2826 SDOperand DAGCombiner::visitLOAD(SDNode *N) {
2827 LoadSDNode *LD = cast<LoadSDNode>(N);
2828 SDOperand Chain = LD->getChain();
2829 SDOperand Ptr = LD->getBasePtr();
2831 // If there are no uses of the loaded value, change uses of the chain value
2832 // into uses of the chain input (i.e. delete the dead load).
2833 if (N->hasNUsesOfValue(0, 0))
2834 return CombineTo(N, DAG.getNode(ISD::UNDEF, N->getValueType(0)), Chain);
2836 // If this load is directly stored, replace the load value with the stored
2838 // TODO: Handle store large -> read small portion.
2839 // TODO: Handle TRUNCSTORE/LOADEXT
2840 if (LD->getExtensionType() == ISD::NON_EXTLOAD) {
2841 if (ISD::isNON_TRUNCStore(Chain.Val)) {
2842 StoreSDNode *PrevST = cast<StoreSDNode>(Chain);
2843 if (PrevST->getBasePtr() == Ptr &&
2844 PrevST->getValue().getValueType() == N->getValueType(0))
2845 return CombineTo(N, Chain.getOperand(1), Chain);
2850 // Walk up chain skipping non-aliasing memory nodes.
2851 SDOperand BetterChain = FindBetterChain(N, Chain);
2853 // If there is a better chain.
2854 if (Chain != BetterChain) {
2857 // Replace the chain to void dependency.
2858 if (LD->getExtensionType() == ISD::NON_EXTLOAD) {
2859 ReplLoad = DAG.getLoad(N->getValueType(0), BetterChain, Ptr,
2860 LD->getSrcValue(), LD->getSrcValueOffset());
2862 ReplLoad = DAG.getExtLoad(LD->getExtensionType(),
2863 LD->getValueType(0),
2864 BetterChain, Ptr, LD->getSrcValue(),
2865 LD->getSrcValueOffset(),
2869 // Create token factor to keep old chain connected.
2870 SDOperand Token = DAG.getNode(ISD::TokenFactor, MVT::Other,
2871 Chain, ReplLoad.getValue(1));
2873 // Replace uses with load result and token factor. Don't add users
2875 return CombineTo(N, ReplLoad.getValue(0), Token, false);
2879 // Try transforming N to an indexed load.
2880 if (CombineToPreIndexedLoadStore(N))
2881 return SDOperand(N, 0);
2886 SDOperand DAGCombiner::visitSTORE(SDNode *N) {
2887 StoreSDNode *ST = cast<StoreSDNode>(N);
2888 SDOperand Chain = ST->getChain();
2889 SDOperand Value = ST->getValue();
2890 SDOperand Ptr = ST->getBasePtr();
2892 // If this is a store of a bit convert, store the input value.
2893 // FIXME: This needs to know that the resultant store does not need a
2894 // higher alignment than the original.
2895 if (0 && Value.getOpcode() == ISD::BIT_CONVERT) {
2896 return DAG.getStore(Chain, Value.getOperand(0), Ptr, ST->getSrcValue(),
2897 ST->getSrcValueOffset());
2901 // Walk up chain skipping non-aliasing memory nodes.
2902 SDOperand BetterChain = FindBetterChain(N, Chain);
2904 // If there is a better chain.
2905 if (Chain != BetterChain) {
2906 // Replace the chain to avoid dependency.
2907 SDOperand ReplStore;
2908 if (ST->isTruncatingStore()) {
2909 ReplStore = DAG.getTruncStore(BetterChain, Value, Ptr,
2910 ST->getSrcValue(),ST->getSrcValueOffset(), ST->getStoredVT());
2912 ReplStore = DAG.getStore(BetterChain, Value, Ptr,
2913 ST->getSrcValue(), ST->getSrcValueOffset());
2916 // Create token to keep both nodes around.
2918 DAG.getNode(ISD::TokenFactor, MVT::Other, Chain, ReplStore);
2920 // Don't add users to work list.
2921 return CombineTo(N, Token, false);
2925 // Try transforming N to an indexed store.
2926 if (CombineToPreIndexedLoadStore(N))
2927 return SDOperand(N, 0);
2932 SDOperand DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) {
2933 SDOperand InVec = N->getOperand(0);
2934 SDOperand InVal = N->getOperand(1);
2935 SDOperand EltNo = N->getOperand(2);
2937 // If the invec is a BUILD_VECTOR and if EltNo is a constant, build a new
2938 // vector with the inserted element.
2939 if (InVec.getOpcode() == ISD::BUILD_VECTOR && isa<ConstantSDNode>(EltNo)) {
2940 unsigned Elt = cast<ConstantSDNode>(EltNo)->getValue();
2941 SmallVector<SDOperand, 8> Ops(InVec.Val->op_begin(), InVec.Val->op_end());
2942 if (Elt < Ops.size())
2944 return DAG.getNode(ISD::BUILD_VECTOR, InVec.getValueType(),
2945 &Ops[0], Ops.size());
2951 SDOperand DAGCombiner::visitVINSERT_VECTOR_ELT(SDNode *N) {
2952 SDOperand InVec = N->getOperand(0);
2953 SDOperand InVal = N->getOperand(1);
2954 SDOperand EltNo = N->getOperand(2);
2955 SDOperand NumElts = N->getOperand(3);
2956 SDOperand EltType = N->getOperand(4);
2958 // If the invec is a VBUILD_VECTOR and if EltNo is a constant, build a new
2959 // vector with the inserted element.
2960 if (InVec.getOpcode() == ISD::VBUILD_VECTOR && isa<ConstantSDNode>(EltNo)) {
2961 unsigned Elt = cast<ConstantSDNode>(EltNo)->getValue();
2962 SmallVector<SDOperand, 8> Ops(InVec.Val->op_begin(), InVec.Val->op_end());
2963 if (Elt < Ops.size()-2)
2965 return DAG.getNode(ISD::VBUILD_VECTOR, InVec.getValueType(),
2966 &Ops[0], Ops.size());
2972 SDOperand DAGCombiner::visitVBUILD_VECTOR(SDNode *N) {
2973 unsigned NumInScalars = N->getNumOperands()-2;
2974 SDOperand NumElts = N->getOperand(NumInScalars);
2975 SDOperand EltType = N->getOperand(NumInScalars+1);
2977 // Check to see if this is a VBUILD_VECTOR of a bunch of VEXTRACT_VECTOR_ELT
2978 // operations. If so, and if the EXTRACT_ELT vector inputs come from at most
2979 // two distinct vectors, turn this into a shuffle node.
2980 SDOperand VecIn1, VecIn2;
2981 for (unsigned i = 0; i != NumInScalars; ++i) {
2982 // Ignore undef inputs.
2983 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
2985 // If this input is something other than a VEXTRACT_VECTOR_ELT with a
2986 // constant index, bail out.
2987 if (N->getOperand(i).getOpcode() != ISD::VEXTRACT_VECTOR_ELT ||
2988 !isa<ConstantSDNode>(N->getOperand(i).getOperand(1))) {
2989 VecIn1 = VecIn2 = SDOperand(0, 0);
2993 // If the input vector type disagrees with the result of the vbuild_vector,
2994 // we can't make a shuffle.
2995 SDOperand ExtractedFromVec = N->getOperand(i).getOperand(0);
2996 if (*(ExtractedFromVec.Val->op_end()-2) != NumElts ||
2997 *(ExtractedFromVec.Val->op_end()-1) != EltType) {
2998 VecIn1 = VecIn2 = SDOperand(0, 0);
3002 // Otherwise, remember this. We allow up to two distinct input vectors.
3003 if (ExtractedFromVec == VecIn1 || ExtractedFromVec == VecIn2)
3006 if (VecIn1.Val == 0) {
3007 VecIn1 = ExtractedFromVec;
3008 } else if (VecIn2.Val == 0) {
3009 VecIn2 = ExtractedFromVec;
3012 VecIn1 = VecIn2 = SDOperand(0, 0);
3017 // If everything is good, we can make a shuffle operation.
3019 SmallVector<SDOperand, 8> BuildVecIndices;
3020 for (unsigned i = 0; i != NumInScalars; ++i) {
3021 if (N->getOperand(i).getOpcode() == ISD::UNDEF) {
3022 BuildVecIndices.push_back(DAG.getNode(ISD::UNDEF, MVT::i32));
3026 SDOperand Extract = N->getOperand(i);
3028 // If extracting from the first vector, just use the index directly.
3029 if (Extract.getOperand(0) == VecIn1) {
3030 BuildVecIndices.push_back(Extract.getOperand(1));
3034 // Otherwise, use InIdx + VecSize
3035 unsigned Idx = cast<ConstantSDNode>(Extract.getOperand(1))->getValue();
3036 BuildVecIndices.push_back(DAG.getConstant(Idx+NumInScalars, MVT::i32));
3039 // Add count and size info.
3040 BuildVecIndices.push_back(NumElts);
3041 BuildVecIndices.push_back(DAG.getValueType(MVT::i32));
3043 // Return the new VVECTOR_SHUFFLE node.
3049 // Use an undef vbuild_vector as input for the second operand.
3050 std::vector<SDOperand> UnOps(NumInScalars,
3051 DAG.getNode(ISD::UNDEF,
3052 cast<VTSDNode>(EltType)->getVT()));
3053 UnOps.push_back(NumElts);
3054 UnOps.push_back(EltType);
3055 Ops[1] = DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector,
3056 &UnOps[0], UnOps.size());
3057 AddToWorkList(Ops[1].Val);
3059 Ops[2] = DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector,
3060 &BuildVecIndices[0], BuildVecIndices.size());
3063 return DAG.getNode(ISD::VVECTOR_SHUFFLE, MVT::Vector, Ops, 5);
3069 SDOperand DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
3070 SDOperand ShufMask = N->getOperand(2);
3071 unsigned NumElts = ShufMask.getNumOperands();
3073 // If the shuffle mask is an identity operation on the LHS, return the LHS.
3074 bool isIdentity = true;
3075 for (unsigned i = 0; i != NumElts; ++i) {
3076 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF &&
3077 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() != i) {
3082 if (isIdentity) return N->getOperand(0);
3084 // If the shuffle mask is an identity operation on the RHS, return the RHS.
3086 for (unsigned i = 0; i != NumElts; ++i) {
3087 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF &&
3088 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() != i+NumElts) {
3093 if (isIdentity) return N->getOperand(1);
3095 // Check if the shuffle is a unary shuffle, i.e. one of the vectors is not
3097 bool isUnary = true;
3098 bool isSplat = true;
3100 unsigned BaseIdx = 0;
3101 for (unsigned i = 0; i != NumElts; ++i)
3102 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF) {
3103 unsigned Idx = cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue();
3104 int V = (Idx < NumElts) ? 0 : 1;
3118 SDOperand N0 = N->getOperand(0);
3119 SDOperand N1 = N->getOperand(1);
3120 // Normalize unary shuffle so the RHS is undef.
3121 if (isUnary && VecNum == 1)
3124 // If it is a splat, check if the argument vector is a build_vector with
3125 // all scalar elements the same.
3128 if (V->getOpcode() == ISD::BIT_CONVERT)
3129 V = V->getOperand(0).Val;
3130 if (V->getOpcode() == ISD::BUILD_VECTOR) {
3131 unsigned NumElems = V->getNumOperands()-2;
3132 if (NumElems > BaseIdx) {
3134 bool AllSame = true;
3135 for (unsigned i = 0; i != NumElems; ++i) {
3136 if (V->getOperand(i).getOpcode() != ISD::UNDEF) {
3137 Base = V->getOperand(i);
3141 // Splat of <u, u, u, u>, return <u, u, u, u>
3144 for (unsigned i = 0; i != NumElems; ++i) {
3145 if (V->getOperand(i).getOpcode() != ISD::UNDEF &&
3146 V->getOperand(i) != Base) {
3151 // Splat of <x, x, x, x>, return <x, x, x, x>
3158 // If it is a unary or the LHS and the RHS are the same node, turn the RHS
3160 if (isUnary || N0 == N1) {
3161 if (N0.getOpcode() == ISD::UNDEF)
3162 return DAG.getNode(ISD::UNDEF, N->getValueType(0));
3163 // Check the SHUFFLE mask, mapping any inputs from the 2nd operand into the
3165 SmallVector<SDOperand, 8> MappedOps;
3166 for (unsigned i = 0, e = ShufMask.getNumOperands(); i != e; ++i) {
3167 if (ShufMask.getOperand(i).getOpcode() == ISD::UNDEF ||
3168 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() < NumElts) {
3169 MappedOps.push_back(ShufMask.getOperand(i));
3172 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() - NumElts;
3173 MappedOps.push_back(DAG.getConstant(NewIdx, MVT::i32));
3176 ShufMask = DAG.getNode(ISD::BUILD_VECTOR, ShufMask.getValueType(),
3177 &MappedOps[0], MappedOps.size());
3178 AddToWorkList(ShufMask.Val);
3179 return DAG.getNode(ISD::VECTOR_SHUFFLE, N->getValueType(0),
3181 DAG.getNode(ISD::UNDEF, N->getValueType(0)),
3188 SDOperand DAGCombiner::visitVVECTOR_SHUFFLE(SDNode *N) {
3189 SDOperand ShufMask = N->getOperand(2);
3190 unsigned NumElts = ShufMask.getNumOperands()-2;
3192 // If the shuffle mask is an identity operation on the LHS, return the LHS.
3193 bool isIdentity = true;
3194 for (unsigned i = 0; i != NumElts; ++i) {
3195 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF &&
3196 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() != i) {
3201 if (isIdentity) return N->getOperand(0);
3203 // If the shuffle mask is an identity operation on the RHS, return the RHS.
3205 for (unsigned i = 0; i != NumElts; ++i) {
3206 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF &&
3207 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() != i+NumElts) {
3212 if (isIdentity) return N->getOperand(1);
3214 // Check if the shuffle is a unary shuffle, i.e. one of the vectors is not
3216 bool isUnary = true;
3217 bool isSplat = true;
3219 unsigned BaseIdx = 0;
3220 for (unsigned i = 0; i != NumElts; ++i)
3221 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF) {
3222 unsigned Idx = cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue();
3223 int V = (Idx < NumElts) ? 0 : 1;
3237 SDOperand N0 = N->getOperand(0);
3238 SDOperand N1 = N->getOperand(1);
3239 // Normalize unary shuffle so the RHS is undef.
3240 if (isUnary && VecNum == 1)
3243 // If it is a splat, check if the argument vector is a build_vector with
3244 // all scalar elements the same.
3248 // If this is a vbit convert that changes the element type of the vector but
3249 // not the number of vector elements, look through it. Be careful not to
3250 // look though conversions that change things like v4f32 to v2f64.
3251 if (V->getOpcode() == ISD::VBIT_CONVERT) {
3252 SDOperand ConvInput = V->getOperand(0);
3253 if (ConvInput.getValueType() == MVT::Vector &&
3255 ConvInput.getConstantOperandVal(ConvInput.getNumOperands()-2))
3259 if (V->getOpcode() == ISD::VBUILD_VECTOR) {
3260 unsigned NumElems = V->getNumOperands()-2;
3261 if (NumElems > BaseIdx) {
3263 bool AllSame = true;
3264 for (unsigned i = 0; i != NumElems; ++i) {
3265 if (V->getOperand(i).getOpcode() != ISD::UNDEF) {
3266 Base = V->getOperand(i);
3270 // Splat of <u, u, u, u>, return <u, u, u, u>
3273 for (unsigned i = 0; i != NumElems; ++i) {
3274 if (V->getOperand(i).getOpcode() != ISD::UNDEF &&
3275 V->getOperand(i) != Base) {
3280 // Splat of <x, x, x, x>, return <x, x, x, x>
3287 // If it is a unary or the LHS and the RHS are the same node, turn the RHS
3289 if (isUnary || N0 == N1) {
3290 // Check the SHUFFLE mask, mapping any inputs from the 2nd operand into the
3292 SmallVector<SDOperand, 8> MappedOps;
3293 for (unsigned i = 0; i != NumElts; ++i) {
3294 if (ShufMask.getOperand(i).getOpcode() == ISD::UNDEF ||
3295 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() < NumElts) {
3296 MappedOps.push_back(ShufMask.getOperand(i));
3299 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() - NumElts;
3300 MappedOps.push_back(DAG.getConstant(NewIdx, MVT::i32));
3303 // Add the type/#elts values.
3304 MappedOps.push_back(ShufMask.getOperand(NumElts));
3305 MappedOps.push_back(ShufMask.getOperand(NumElts+1));
3307 ShufMask = DAG.getNode(ISD::VBUILD_VECTOR, ShufMask.getValueType(),
3308 &MappedOps[0], MappedOps.size());
3309 AddToWorkList(ShufMask.Val);
3311 // Build the undef vector.
3312 SDOperand UDVal = DAG.getNode(ISD::UNDEF, MappedOps[0].getValueType());
3313 for (unsigned i = 0; i != NumElts; ++i)
3314 MappedOps[i] = UDVal;
3315 MappedOps[NumElts ] = *(N0.Val->op_end()-2);
3316 MappedOps[NumElts+1] = *(N0.Val->op_end()-1);
3317 UDVal = DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector,
3318 &MappedOps[0], MappedOps.size());
3320 return DAG.getNode(ISD::VVECTOR_SHUFFLE, MVT::Vector,
3321 N0, UDVal, ShufMask,
3322 MappedOps[NumElts], MappedOps[NumElts+1]);
3328 /// XformToShuffleWithZero - Returns a vector_shuffle if it able to transform
3329 /// a VAND to a vector_shuffle with the destination vector and a zero vector.
3330 /// e.g. VAND V, <0xffffffff, 0, 0xffffffff, 0>. ==>
3331 /// vector_shuffle V, Zero, <0, 4, 2, 4>
3332 SDOperand DAGCombiner::XformToShuffleWithZero(SDNode *N) {
3333 SDOperand LHS = N->getOperand(0);
3334 SDOperand RHS = N->getOperand(1);
3335 if (N->getOpcode() == ISD::VAND) {
3336 SDOperand DstVecSize = *(LHS.Val->op_end()-2);
3337 SDOperand DstVecEVT = *(LHS.Val->op_end()-1);
3338 if (RHS.getOpcode() == ISD::VBIT_CONVERT)
3339 RHS = RHS.getOperand(0);
3340 if (RHS.getOpcode() == ISD::VBUILD_VECTOR) {
3341 std::vector<SDOperand> IdxOps;
3342 unsigned NumOps = RHS.getNumOperands();
3343 unsigned NumElts = NumOps-2;
3344 MVT::ValueType EVT = cast<VTSDNode>(RHS.getOperand(NumOps-1))->getVT();
3345 for (unsigned i = 0; i != NumElts; ++i) {
3346 SDOperand Elt = RHS.getOperand(i);
3347 if (!isa<ConstantSDNode>(Elt))
3349 else if (cast<ConstantSDNode>(Elt)->isAllOnesValue())
3350 IdxOps.push_back(DAG.getConstant(i, EVT));
3351 else if (cast<ConstantSDNode>(Elt)->isNullValue())
3352 IdxOps.push_back(DAG.getConstant(NumElts, EVT));
3357 // Let's see if the target supports this vector_shuffle.
3358 if (!TLI.isVectorClearMaskLegal(IdxOps, EVT, DAG))
3361 // Return the new VVECTOR_SHUFFLE node.
3362 SDOperand NumEltsNode = DAG.getConstant(NumElts, MVT::i32);
3363 SDOperand EVTNode = DAG.getValueType(EVT);
3364 std::vector<SDOperand> Ops;
3365 LHS = DAG.getNode(ISD::VBIT_CONVERT, MVT::Vector, LHS, NumEltsNode,
3368 AddToWorkList(LHS.Val);
3369 std::vector<SDOperand> ZeroOps(NumElts, DAG.getConstant(0, EVT));
3370 ZeroOps.push_back(NumEltsNode);
3371 ZeroOps.push_back(EVTNode);
3372 Ops.push_back(DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector,
3373 &ZeroOps[0], ZeroOps.size()));
3374 IdxOps.push_back(NumEltsNode);
3375 IdxOps.push_back(EVTNode);
3376 Ops.push_back(DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector,
3377 &IdxOps[0], IdxOps.size()));
3378 Ops.push_back(NumEltsNode);
3379 Ops.push_back(EVTNode);
3380 SDOperand Result = DAG.getNode(ISD::VVECTOR_SHUFFLE, MVT::Vector,
3381 &Ops[0], Ops.size());
3382 if (NumEltsNode != DstVecSize || EVTNode != DstVecEVT) {
3383 Result = DAG.getNode(ISD::VBIT_CONVERT, MVT::Vector, Result,
3384 DstVecSize, DstVecEVT);
3392 /// visitVBinOp - Visit a binary vector operation, like VADD. IntOp indicates
3393 /// the scalar operation of the vop if it is operating on an integer vector
3394 /// (e.g. ADD) and FPOp indicates the FP version (e.g. FADD).
3395 SDOperand DAGCombiner::visitVBinOp(SDNode *N, ISD::NodeType IntOp,
3396 ISD::NodeType FPOp) {
3397 MVT::ValueType EltType = cast<VTSDNode>(*(N->op_end()-1))->getVT();
3398 ISD::NodeType ScalarOp = MVT::isInteger(EltType) ? IntOp : FPOp;
3399 SDOperand LHS = N->getOperand(0);
3400 SDOperand RHS = N->getOperand(1);
3401 SDOperand Shuffle = XformToShuffleWithZero(N);
3402 if (Shuffle.Val) return Shuffle;
3404 // If the LHS and RHS are VBUILD_VECTOR nodes, see if we can constant fold
3406 if (LHS.getOpcode() == ISD::VBUILD_VECTOR &&
3407 RHS.getOpcode() == ISD::VBUILD_VECTOR) {
3408 SmallVector<SDOperand, 8> Ops;
3409 for (unsigned i = 0, e = LHS.getNumOperands()-2; i != e; ++i) {
3410 SDOperand LHSOp = LHS.getOperand(i);
3411 SDOperand RHSOp = RHS.getOperand(i);
3412 // If these two elements can't be folded, bail out.
3413 if ((LHSOp.getOpcode() != ISD::UNDEF &&
3414 LHSOp.getOpcode() != ISD::Constant &&
3415 LHSOp.getOpcode() != ISD::ConstantFP) ||
3416 (RHSOp.getOpcode() != ISD::UNDEF &&
3417 RHSOp.getOpcode() != ISD::Constant &&
3418 RHSOp.getOpcode() != ISD::ConstantFP))
3420 // Can't fold divide by zero.
3421 if (N->getOpcode() == ISD::VSDIV || N->getOpcode() == ISD::VUDIV) {
3422 if ((RHSOp.getOpcode() == ISD::Constant &&
3423 cast<ConstantSDNode>(RHSOp.Val)->isNullValue()) ||
3424 (RHSOp.getOpcode() == ISD::ConstantFP &&
3425 !cast<ConstantFPSDNode>(RHSOp.Val)->getValue()))
3428 Ops.push_back(DAG.getNode(ScalarOp, EltType, LHSOp, RHSOp));
3429 AddToWorkList(Ops.back().Val);
3430 assert((Ops.back().getOpcode() == ISD::UNDEF ||
3431 Ops.back().getOpcode() == ISD::Constant ||
3432 Ops.back().getOpcode() == ISD::ConstantFP) &&
3433 "Scalar binop didn't fold!");
3436 if (Ops.size() == LHS.getNumOperands()-2) {
3437 Ops.push_back(*(LHS.Val->op_end()-2));
3438 Ops.push_back(*(LHS.Val->op_end()-1));
3439 return DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, &Ops[0], Ops.size());
3446 SDOperand DAGCombiner::SimplifySelect(SDOperand N0, SDOperand N1, SDOperand N2){
3447 assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!");
3449 SDOperand SCC = SimplifySelectCC(N0.getOperand(0), N0.getOperand(1), N1, N2,
3450 cast<CondCodeSDNode>(N0.getOperand(2))->get());
3451 // If we got a simplified select_cc node back from SimplifySelectCC, then
3452 // break it down into a new SETCC node, and a new SELECT node, and then return
3453 // the SELECT node, since we were called with a SELECT node.
3455 // Check to see if we got a select_cc back (to turn into setcc/select).
3456 // Otherwise, just return whatever node we got back, like fabs.
3457 if (SCC.getOpcode() == ISD::SELECT_CC) {
3458 SDOperand SETCC = DAG.getNode(ISD::SETCC, N0.getValueType(),
3459 SCC.getOperand(0), SCC.getOperand(1),
3461 AddToWorkList(SETCC.Val);
3462 return DAG.getNode(ISD::SELECT, SCC.getValueType(), SCC.getOperand(2),
3463 SCC.getOperand(3), SETCC);
3470 /// SimplifySelectOps - Given a SELECT or a SELECT_CC node, where LHS and RHS
3471 /// are the two values being selected between, see if we can simplify the
3472 /// select. Callers of this should assume that TheSelect is deleted if this
3473 /// returns true. As such, they should return the appropriate thing (e.g. the
3474 /// node) back to the top-level of the DAG combiner loop to avoid it being
3477 bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDOperand LHS,
3480 // If this is a select from two identical things, try to pull the operation
3481 // through the select.
3482 if (LHS.getOpcode() == RHS.getOpcode() && LHS.hasOneUse() && RHS.hasOneUse()){
3483 // If this is a load and the token chain is identical, replace the select
3484 // of two loads with a load through a select of the address to load from.
3485 // This triggers in things like "select bool X, 10.0, 123.0" after the FP
3486 // constants have been dropped into the constant pool.
3487 if (LHS.getOpcode() == ISD::LOAD &&
3488 // Token chains must be identical.
3489 LHS.getOperand(0) == RHS.getOperand(0)) {
3490 LoadSDNode *LLD = cast<LoadSDNode>(LHS);
3491 LoadSDNode *RLD = cast<LoadSDNode>(RHS);
3493 // If this is an EXTLOAD, the VT's must match.
3494 if (LLD->getLoadedVT() == RLD->getLoadedVT()) {
3495 // FIXME: this conflates two src values, discarding one. This is not
3496 // the right thing to do, but nothing uses srcvalues now. When they do,
3497 // turn SrcValue into a list of locations.
3499 if (TheSelect->getOpcode() == ISD::SELECT)
3500 Addr = DAG.getNode(ISD::SELECT, LLD->getBasePtr().getValueType(),
3501 TheSelect->getOperand(0), LLD->getBasePtr(),
3504 Addr = DAG.getNode(ISD::SELECT_CC, LLD->getBasePtr().getValueType(),
3505 TheSelect->getOperand(0),
3506 TheSelect->getOperand(1),
3507 LLD->getBasePtr(), RLD->getBasePtr(),
3508 TheSelect->getOperand(4));
3511 if (LLD->getExtensionType() == ISD::NON_EXTLOAD)
3512 Load = DAG.getLoad(TheSelect->getValueType(0), LLD->getChain(),
3513 Addr,LLD->getSrcValue(), LLD->getSrcValueOffset());
3515 Load = DAG.getExtLoad(LLD->getExtensionType(),
3516 TheSelect->getValueType(0),
3517 LLD->getChain(), Addr, LLD->getSrcValue(),
3518 LLD->getSrcValueOffset(),
3519 LLD->getLoadedVT());
3521 // Users of the select now use the result of the load.
3522 CombineTo(TheSelect, Load);
3524 // Users of the old loads now use the new load's chain. We know the
3525 // old-load value is dead now.
3526 CombineTo(LHS.Val, Load.getValue(0), Load.getValue(1));
3527 CombineTo(RHS.Val, Load.getValue(0), Load.getValue(1));
3536 SDOperand DAGCombiner::SimplifySelectCC(SDOperand N0, SDOperand N1,
3537 SDOperand N2, SDOperand N3,
3540 MVT::ValueType VT = N2.getValueType();
3541 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val);
3542 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.Val);
3543 ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N3.Val);
3545 // Determine if the condition we're dealing with is constant
3546 SDOperand SCC = SimplifySetCC(TLI.getSetCCResultTy(), N0, N1, CC, false);
3547 if (SCC.Val) AddToWorkList(SCC.Val);
3548 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.Val);
3550 // fold select_cc true, x, y -> x
3551 if (SCCC && SCCC->getValue())
3553 // fold select_cc false, x, y -> y
3554 if (SCCC && SCCC->getValue() == 0)
3557 // Check to see if we can simplify the select into an fabs node
3558 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) {
3559 // Allow either -0.0 or 0.0
3560 if (CFP->getValue() == 0.0) {
3561 // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs
3562 if ((CC == ISD::SETGE || CC == ISD::SETGT) &&
3563 N0 == N2 && N3.getOpcode() == ISD::FNEG &&
3564 N2 == N3.getOperand(0))
3565 return DAG.getNode(ISD::FABS, VT, N0);
3567 // select (setl[te] X, +/-0.0), fneg(X), X -> fabs
3568 if ((CC == ISD::SETLT || CC == ISD::SETLE) &&
3569 N0 == N3 && N2.getOpcode() == ISD::FNEG &&
3570 N2.getOperand(0) == N3)
3571 return DAG.getNode(ISD::FABS, VT, N3);
3575 // Check to see if we can perform the "gzip trick", transforming
3576 // select_cc setlt X, 0, A, 0 -> and (sra X, size(X)-1), A
3577 if (N1C && N3C && N3C->isNullValue() && CC == ISD::SETLT &&
3578 MVT::isInteger(N0.getValueType()) &&
3579 MVT::isInteger(N2.getValueType()) &&
3580 (N1C->isNullValue() || // (a < 0) ? b : 0
3581 (N1C->getValue() == 1 && N0 == N2))) { // (a < 1) ? a : 0
3582 MVT::ValueType XType = N0.getValueType();
3583 MVT::ValueType AType = N2.getValueType();
3584 if (XType >= AType) {
3585 // and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a
3586 // single-bit constant.
3587 if (N2C && ((N2C->getValue() & (N2C->getValue()-1)) == 0)) {
3588 unsigned ShCtV = Log2_64(N2C->getValue());
3589 ShCtV = MVT::getSizeInBits(XType)-ShCtV-1;
3590 SDOperand ShCt = DAG.getConstant(ShCtV, TLI.getShiftAmountTy());
3591 SDOperand Shift = DAG.getNode(ISD::SRL, XType, N0, ShCt);
3592 AddToWorkList(Shift.Val);
3593 if (XType > AType) {
3594 Shift = DAG.getNode(ISD::TRUNCATE, AType, Shift);
3595 AddToWorkList(Shift.Val);
3597 return DAG.getNode(ISD::AND, AType, Shift, N2);
3599 SDOperand Shift = DAG.getNode(ISD::SRA, XType, N0,
3600 DAG.getConstant(MVT::getSizeInBits(XType)-1,
3601 TLI.getShiftAmountTy()));
3602 AddToWorkList(Shift.Val);
3603 if (XType > AType) {
3604 Shift = DAG.getNode(ISD::TRUNCATE, AType, Shift);
3605 AddToWorkList(Shift.Val);
3607 return DAG.getNode(ISD::AND, AType, Shift, N2);
3611 // fold select C, 16, 0 -> shl C, 4
3612 if (N2C && N3C && N3C->isNullValue() && isPowerOf2_64(N2C->getValue()) &&
3613 TLI.getSetCCResultContents() == TargetLowering::ZeroOrOneSetCCResult) {
3614 // Get a SetCC of the condition
3615 // FIXME: Should probably make sure that setcc is legal if we ever have a
3616 // target where it isn't.
3617 SDOperand Temp, SCC;
3618 // cast from setcc result type to select result type
3619 if (AfterLegalize) {
3620 SCC = DAG.getSetCC(TLI.getSetCCResultTy(), N0, N1, CC);
3621 Temp = DAG.getZeroExtendInReg(SCC, N2.getValueType());
3623 SCC = DAG.getSetCC(MVT::i1, N0, N1, CC);
3624 Temp = DAG.getNode(ISD::ZERO_EXTEND, N2.getValueType(), SCC);
3626 AddToWorkList(SCC.Val);
3627 AddToWorkList(Temp.Val);
3628 // shl setcc result by log2 n2c
3629 return DAG.getNode(ISD::SHL, N2.getValueType(), Temp,
3630 DAG.getConstant(Log2_64(N2C->getValue()),
3631 TLI.getShiftAmountTy()));
3634 // Check to see if this is the equivalent of setcc
3635 // FIXME: Turn all of these into setcc if setcc if setcc is legal
3636 // otherwise, go ahead with the folds.
3637 if (0 && N3C && N3C->isNullValue() && N2C && (N2C->getValue() == 1ULL)) {
3638 MVT::ValueType XType = N0.getValueType();
3639 if (TLI.isOperationLegal(ISD::SETCC, TLI.getSetCCResultTy())) {
3640 SDOperand Res = DAG.getSetCC(TLI.getSetCCResultTy(), N0, N1, CC);
3641 if (Res.getValueType() != VT)
3642 Res = DAG.getNode(ISD::ZERO_EXTEND, VT, Res);
3646 // seteq X, 0 -> srl (ctlz X, log2(size(X)))
3647 if (N1C && N1C->isNullValue() && CC == ISD::SETEQ &&
3648 TLI.isOperationLegal(ISD::CTLZ, XType)) {
3649 SDOperand Ctlz = DAG.getNode(ISD::CTLZ, XType, N0);
3650 return DAG.getNode(ISD::SRL, XType, Ctlz,
3651 DAG.getConstant(Log2_32(MVT::getSizeInBits(XType)),
3652 TLI.getShiftAmountTy()));
3654 // setgt X, 0 -> srl (and (-X, ~X), size(X)-1)
3655 if (N1C && N1C->isNullValue() && CC == ISD::SETGT) {
3656 SDOperand NegN0 = DAG.getNode(ISD::SUB, XType, DAG.getConstant(0, XType),
3658 SDOperand NotN0 = DAG.getNode(ISD::XOR, XType, N0,
3659 DAG.getConstant(~0ULL, XType));
3660 return DAG.getNode(ISD::SRL, XType,
3661 DAG.getNode(ISD::AND, XType, NegN0, NotN0),
3662 DAG.getConstant(MVT::getSizeInBits(XType)-1,
3663 TLI.getShiftAmountTy()));
3665 // setgt X, -1 -> xor (srl (X, size(X)-1), 1)
3666 if (N1C && N1C->isAllOnesValue() && CC == ISD::SETGT) {
3667 SDOperand Sign = DAG.getNode(ISD::SRL, XType, N0,
3668 DAG.getConstant(MVT::getSizeInBits(XType)-1,
3669 TLI.getShiftAmountTy()));
3670 return DAG.getNode(ISD::XOR, XType, Sign, DAG.getConstant(1, XType));
3674 // Check to see if this is an integer abs. select_cc setl[te] X, 0, -X, X ->
3675 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
3676 if (N1C && N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE) &&
3677 N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1)) {
3678 if (ConstantSDNode *SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0))) {
3679 MVT::ValueType XType = N0.getValueType();
3680 if (SubC->isNullValue() && MVT::isInteger(XType)) {
3681 SDOperand Shift = DAG.getNode(ISD::SRA, XType, N0,
3682 DAG.getConstant(MVT::getSizeInBits(XType)-1,
3683 TLI.getShiftAmountTy()));
3684 SDOperand Add = DAG.getNode(ISD::ADD, XType, N0, Shift);
3685 AddToWorkList(Shift.Val);
3686 AddToWorkList(Add.Val);
3687 return DAG.getNode(ISD::XOR, XType, Add, Shift);
3695 SDOperand DAGCombiner::SimplifySetCC(MVT::ValueType VT, SDOperand N0,
3696 SDOperand N1, ISD::CondCode Cond,
3697 bool foldBooleans) {
3698 // These setcc operations always fold.
3702 case ISD::SETFALSE2: return DAG.getConstant(0, VT);
3704 case ISD::SETTRUE2: return DAG.getConstant(1, VT);
3707 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val)) {
3708 uint64_t C1 = N1C->getValue();
3709 if (isa<ConstantSDNode>(N0.Val)) {
3710 return DAG.FoldSetCC(VT, N0, N1, Cond);
3712 // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an
3713 // equality comparison, then we're just comparing whether X itself is
3715 if (N0.getOpcode() == ISD::SRL && (C1 == 0 || C1 == 1) &&
3716 N0.getOperand(0).getOpcode() == ISD::CTLZ &&
3717 N0.getOperand(1).getOpcode() == ISD::Constant) {
3718 unsigned ShAmt = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
3719 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
3720 ShAmt == Log2_32(MVT::getSizeInBits(N0.getValueType()))) {
3721 if ((C1 == 0) == (Cond == ISD::SETEQ)) {
3722 // (srl (ctlz x), 5) == 0 -> X != 0
3723 // (srl (ctlz x), 5) != 1 -> X != 0
3726 // (srl (ctlz x), 5) != 0 -> X == 0
3727 // (srl (ctlz x), 5) == 1 -> X == 0
3730 SDOperand Zero = DAG.getConstant(0, N0.getValueType());
3731 return DAG.getSetCC(VT, N0.getOperand(0).getOperand(0),
3736 // If the LHS is a ZERO_EXTEND, perform the comparison on the input.
3737 if (N0.getOpcode() == ISD::ZERO_EXTEND) {
3738 unsigned InSize = MVT::getSizeInBits(N0.getOperand(0).getValueType());
3740 // If the comparison constant has bits in the upper part, the
3741 // zero-extended value could never match.
3742 if (C1 & (~0ULL << InSize)) {
3743 unsigned VSize = MVT::getSizeInBits(N0.getValueType());
3747 case ISD::SETEQ: return DAG.getConstant(0, VT);
3750 case ISD::SETNE: return DAG.getConstant(1, VT);
3753 // True if the sign bit of C1 is set.
3754 return DAG.getConstant((C1 & (1ULL << VSize)) != 0, VT);
3757 // True if the sign bit of C1 isn't set.
3758 return DAG.getConstant((C1 & (1ULL << VSize)) == 0, VT);
3764 // Otherwise, we can perform the comparison with the low bits.
3772 return DAG.getSetCC(VT, N0.getOperand(0),
3773 DAG.getConstant(C1, N0.getOperand(0).getValueType()),
3776 break; // todo, be more careful with signed comparisons
3778 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
3779 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
3780 MVT::ValueType ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT();
3781 unsigned ExtSrcTyBits = MVT::getSizeInBits(ExtSrcTy);
3782 MVT::ValueType ExtDstTy = N0.getValueType();
3783 unsigned ExtDstTyBits = MVT::getSizeInBits(ExtDstTy);
3785 // If the extended part has any inconsistent bits, it cannot ever
3786 // compare equal. In other words, they have to be all ones or all
3789 (~0ULL >> (64-ExtSrcTyBits)) & (~0ULL << (ExtDstTyBits-1));
3790 if ((C1 & ExtBits) != 0 && (C1 & ExtBits) != ExtBits)
3791 return DAG.getConstant(Cond == ISD::SETNE, VT);
3794 MVT::ValueType Op0Ty = N0.getOperand(0).getValueType();
3795 if (Op0Ty == ExtSrcTy) {
3796 ZextOp = N0.getOperand(0);
3798 int64_t Imm = ~0ULL >> (64-ExtSrcTyBits);
3799 ZextOp = DAG.getNode(ISD::AND, Op0Ty, N0.getOperand(0),
3800 DAG.getConstant(Imm, Op0Ty));
3802 AddToWorkList(ZextOp.Val);
3803 // Otherwise, make this a use of a zext.
3804 return DAG.getSetCC(VT, ZextOp,
3805 DAG.getConstant(C1 & (~0ULL>>(64-ExtSrcTyBits)),
3808 } else if ((N1C->getValue() == 0 || N1C->getValue() == 1) &&
3809 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
3811 // SETCC (SETCC), [0|1], [EQ|NE] -> SETCC
3812 if (N0.getOpcode() == ISD::SETCC) {
3813 bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (N1C->getValue() != 1);
3817 // Invert the condition.
3818 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
3819 CC = ISD::getSetCCInverse(CC,
3820 MVT::isInteger(N0.getOperand(0).getValueType()));
3821 return DAG.getSetCC(VT, N0.getOperand(0), N0.getOperand(1), CC);
3824 if ((N0.getOpcode() == ISD::XOR ||
3825 (N0.getOpcode() == ISD::AND &&
3826 N0.getOperand(0).getOpcode() == ISD::XOR &&
3827 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) &&
3828 isa<ConstantSDNode>(N0.getOperand(1)) &&
3829 cast<ConstantSDNode>(N0.getOperand(1))->getValue() == 1) {
3830 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We
3831 // can only do this if the top bits are known zero.
3832 if (TLI.MaskedValueIsZero(N0,
3833 MVT::getIntVTBitMask(N0.getValueType())-1)){
3834 // Okay, get the un-inverted input value.
3836 if (N0.getOpcode() == ISD::XOR)
3837 Val = N0.getOperand(0);
3839 assert(N0.getOpcode() == ISD::AND &&
3840 N0.getOperand(0).getOpcode() == ISD::XOR);
3841 // ((X^1)&1)^1 -> X & 1
3842 Val = DAG.getNode(ISD::AND, N0.getValueType(),
3843 N0.getOperand(0).getOperand(0),
3846 return DAG.getSetCC(VT, Val, N1,
3847 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
3852 uint64_t MinVal, MaxVal;
3853 unsigned OperandBitSize = MVT::getSizeInBits(N1C->getValueType(0));
3854 if (ISD::isSignedIntSetCC(Cond)) {
3855 MinVal = 1ULL << (OperandBitSize-1);
3856 if (OperandBitSize != 1) // Avoid X >> 64, which is undefined.
3857 MaxVal = ~0ULL >> (65-OperandBitSize);
3862 MaxVal = ~0ULL >> (64-OperandBitSize);
3865 // Canonicalize GE/LE comparisons to use GT/LT comparisons.
3866 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) {
3867 if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true
3868 --C1; // X >= C0 --> X > (C0-1)
3869 return DAG.getSetCC(VT, N0, DAG.getConstant(C1, N1.getValueType()),
3870 (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT);
3873 if (Cond == ISD::SETLE || Cond == ISD::SETULE) {
3874 if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true
3875 ++C1; // X <= C0 --> X < (C0+1)
3876 return DAG.getSetCC(VT, N0, DAG.getConstant(C1, N1.getValueType()),
3877 (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT);
3880 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal)
3881 return DAG.getConstant(0, VT); // X < MIN --> false
3883 // Canonicalize setgt X, Min --> setne X, Min
3884 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal)
3885 return DAG.getSetCC(VT, N0, N1, ISD::SETNE);
3886 // Canonicalize setlt X, Max --> setne X, Max
3887 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal)
3888 return DAG.getSetCC(VT, N0, N1, ISD::SETNE);
3890 // If we have setult X, 1, turn it into seteq X, 0
3891 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1)
3892 return DAG.getSetCC(VT, N0, DAG.getConstant(MinVal, N0.getValueType()),
3894 // If we have setugt X, Max-1, turn it into seteq X, Max
3895 else if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1)
3896 return DAG.getSetCC(VT, N0, DAG.getConstant(MaxVal, N0.getValueType()),
3899 // If we have "setcc X, C0", check to see if we can shrink the immediate
3902 // SETUGT X, SINTMAX -> SETLT X, 0
3903 if (Cond == ISD::SETUGT && OperandBitSize != 1 &&
3904 C1 == (~0ULL >> (65-OperandBitSize)))
3905 return DAG.getSetCC(VT, N0, DAG.getConstant(0, N1.getValueType()),
3908 // FIXME: Implement the rest of these.
3910 // Fold bit comparisons when we can.
3911 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
3912 VT == N0.getValueType() && N0.getOpcode() == ISD::AND)
3913 if (ConstantSDNode *AndRHS =
3914 dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
3915 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3
3916 // Perform the xform if the AND RHS is a single bit.
3917 if (isPowerOf2_64(AndRHS->getValue())) {
3918 return DAG.getNode(ISD::SRL, VT, N0,
3919 DAG.getConstant(Log2_64(AndRHS->getValue()),
3920 TLI.getShiftAmountTy()));
3922 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getValue()) {
3923 // (X & 8) == 8 --> (X & 8) >> 3
3924 // Perform the xform if C1 is a single bit.
3925 if (isPowerOf2_64(C1)) {
3926 return DAG.getNode(ISD::SRL, VT, N0,
3927 DAG.getConstant(Log2_64(C1),TLI.getShiftAmountTy()));
3932 } else if (isa<ConstantSDNode>(N0.Val)) {
3933 // Ensure that the constant occurs on the RHS.
3934 return DAG.getSetCC(VT, N1, N0, ISD::getSetCCSwappedOperands(Cond));
3937 if (isa<ConstantFPSDNode>(N0.Val)) {
3938 // Constant fold or commute setcc.
3939 SDOperand O = DAG.FoldSetCC(VT, N0, N1, Cond);
3940 if (O.Val) return O;
3944 // We can always fold X == X for integer setcc's.
3945 if (MVT::isInteger(N0.getValueType()))
3946 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
3947 unsigned UOF = ISD::getUnorderedFlavor(Cond);
3948 if (UOF == 2) // FP operators that are undefined on NaNs.
3949 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
3950 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond)))
3951 return DAG.getConstant(UOF, VT);
3952 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO
3953 // if it is not already.
3954 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO;
3955 if (NewCond != Cond)
3956 return DAG.getSetCC(VT, N0, N1, NewCond);
3959 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
3960 MVT::isInteger(N0.getValueType())) {
3961 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB ||
3962 N0.getOpcode() == ISD::XOR) {
3963 // Simplify (X+Y) == (X+Z) --> Y == Z
3964 if (N0.getOpcode() == N1.getOpcode()) {
3965 if (N0.getOperand(0) == N1.getOperand(0))
3966 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(1), Cond);
3967 if (N0.getOperand(1) == N1.getOperand(1))
3968 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(0), Cond);
3969 if (DAG.isCommutativeBinOp(N0.getOpcode())) {
3970 // If X op Y == Y op X, try other combinations.
3971 if (N0.getOperand(0) == N1.getOperand(1))
3972 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(0), Cond);
3973 if (N0.getOperand(1) == N1.getOperand(0))
3974 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(1), Cond);
3978 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) {
3979 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
3980 // Turn (X+C1) == C2 --> X == C2-C1
3981 if (N0.getOpcode() == ISD::ADD && N0.Val->hasOneUse()) {
3982 return DAG.getSetCC(VT, N0.getOperand(0),
3983 DAG.getConstant(RHSC->getValue()-LHSR->getValue(),
3984 N0.getValueType()), Cond);
3987 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0.
3988 if (N0.getOpcode() == ISD::XOR)
3989 // If we know that all of the inverted bits are zero, don't bother
3990 // performing the inversion.
3991 if (TLI.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getValue()))
3992 return DAG.getSetCC(VT, N0.getOperand(0),
3993 DAG.getConstant(LHSR->getValue()^RHSC->getValue(),
3994 N0.getValueType()), Cond);
3997 // Turn (C1-X) == C2 --> X == C1-C2
3998 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) {
3999 if (N0.getOpcode() == ISD::SUB && N0.Val->hasOneUse()) {
4000 return DAG.getSetCC(VT, N0.getOperand(1),
4001 DAG.getConstant(SUBC->getValue()-RHSC->getValue(),
4002 N0.getValueType()), Cond);
4007 // Simplify (X+Z) == X --> Z == 0
4008 if (N0.getOperand(0) == N1)
4009 return DAG.getSetCC(VT, N0.getOperand(1),
4010 DAG.getConstant(0, N0.getValueType()), Cond);
4011 if (N0.getOperand(1) == N1) {
4012 if (DAG.isCommutativeBinOp(N0.getOpcode()))
4013 return DAG.getSetCC(VT, N0.getOperand(0),
4014 DAG.getConstant(0, N0.getValueType()), Cond);
4016 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!");
4017 // (Z-X) == X --> Z == X<<1
4018 SDOperand SH = DAG.getNode(ISD::SHL, N1.getValueType(),
4020 DAG.getConstant(1,TLI.getShiftAmountTy()));
4021 AddToWorkList(SH.Val);
4022 return DAG.getSetCC(VT, N0.getOperand(0), SH, Cond);
4027 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB ||
4028 N1.getOpcode() == ISD::XOR) {
4029 // Simplify X == (X+Z) --> Z == 0
4030 if (N1.getOperand(0) == N0) {
4031 return DAG.getSetCC(VT, N1.getOperand(1),
4032 DAG.getConstant(0, N1.getValueType()), Cond);
4033 } else if (N1.getOperand(1) == N0) {
4034 if (DAG.isCommutativeBinOp(N1.getOpcode())) {
4035 return DAG.getSetCC(VT, N1.getOperand(0),
4036 DAG.getConstant(0, N1.getValueType()), Cond);
4038 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!");
4039 // X == (Z-X) --> X<<1 == Z
4040 SDOperand SH = DAG.getNode(ISD::SHL, N1.getValueType(), N0,
4041 DAG.getConstant(1,TLI.getShiftAmountTy()));
4042 AddToWorkList(SH.Val);
4043 return DAG.getSetCC(VT, SH, N1.getOperand(0), Cond);
4049 // Fold away ALL boolean setcc's.
4051 if (N0.getValueType() == MVT::i1 && foldBooleans) {
4053 default: assert(0 && "Unknown integer setcc!");
4054 case ISD::SETEQ: // X == Y -> (X^Y)^1
4055 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, N1);
4056 N0 = DAG.getNode(ISD::XOR, MVT::i1, Temp, DAG.getConstant(1, MVT::i1));
4057 AddToWorkList(Temp.Val);
4059 case ISD::SETNE: // X != Y --> (X^Y)
4060 N0 = DAG.getNode(ISD::XOR, MVT::i1, N0, N1);
4062 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> X^1 & Y
4063 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> X^1 & Y
4064 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1));
4065 N0 = DAG.getNode(ISD::AND, MVT::i1, N1, Temp);
4066 AddToWorkList(Temp.Val);
4068 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> Y^1 & X
4069 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> Y^1 & X
4070 Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1));
4071 N0 = DAG.getNode(ISD::AND, MVT::i1, N0, Temp);
4072 AddToWorkList(Temp.Val);
4074 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> X^1 | Y
4075 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> X^1 | Y
4076 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1));
4077 N0 = DAG.getNode(ISD::OR, MVT::i1, N1, Temp);
4078 AddToWorkList(Temp.Val);
4080 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> Y^1 | X
4081 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> Y^1 | X
4082 Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1));
4083 N0 = DAG.getNode(ISD::OR, MVT::i1, N0, Temp);
4086 if (VT != MVT::i1) {
4087 AddToWorkList(N0.Val);
4088 // FIXME: If running after legalize, we probably can't do this.
4089 N0 = DAG.getNode(ISD::ZERO_EXTEND, VT, N0);
4094 // Could not fold it.
4098 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
4099 /// return a DAG expression to select that will generate the same value by
4100 /// multiplying by a magic number. See:
4101 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
4102 SDOperand DAGCombiner::BuildSDIV(SDNode *N) {
4103 std::vector<SDNode*> Built;
4104 SDOperand S = TLI.BuildSDIV(N, DAG, &Built);
4106 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end();
4112 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
4113 /// return a DAG expression to select that will generate the same value by
4114 /// multiplying by a magic number. See:
4115 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
4116 SDOperand DAGCombiner::BuildUDIV(SDNode *N) {
4117 std::vector<SDNode*> Built;
4118 SDOperand S = TLI.BuildUDIV(N, DAG, &Built);
4120 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end();
4126 /// FindBaseOffset - Return true if base is known not to alias with anything
4127 /// but itself. Provides base object and offset as results.
4128 static bool FindBaseOffset(SDOperand Ptr, SDOperand &Base, int64_t &Offset) {
4129 // Assume it is a primitive operation.
4130 Base = Ptr; Offset = 0;
4132 // If it's an adding a simple constant then integrate the offset.
4133 if (Base.getOpcode() == ISD::ADD) {
4134 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Base.getOperand(1))) {
4135 Base = Base.getOperand(0);
4136 Offset += C->getValue();
4140 // If it's any of the following then it can't alias with anything but itself.
4141 return isa<FrameIndexSDNode>(Base) ||
4142 isa<ConstantPoolSDNode>(Base) ||
4143 isa<GlobalAddressSDNode>(Base);
4146 /// isAlias - Return true if there is any possibility that the two addresses
4148 bool DAGCombiner::isAlias(SDOperand Ptr1, int64_t Size1,
4149 const Value *SrcValue1, int SrcValueOffset1,
4150 SDOperand Ptr2, int64_t Size2,
4151 const Value *SrcValue2, int SrcValueOffset2)
4153 // If they are the same then they must be aliases.
4154 if (Ptr1 == Ptr2) return true;
4156 // Gather base node and offset information.
4157 SDOperand Base1, Base2;
4158 int64_t Offset1, Offset2;
4159 bool KnownBase1 = FindBaseOffset(Ptr1, Base1, Offset1);
4160 bool KnownBase2 = FindBaseOffset(Ptr2, Base2, Offset2);
4162 // If they have a same base address then...
4163 if (Base1 == Base2) {
4164 // Check to see if the addresses overlap.
4165 return!((Offset1 + Size1) <= Offset2 || (Offset2 + Size2) <= Offset1);
4168 // If we know both bases then they can't alias.
4169 if (KnownBase1 && KnownBase2) return false;
4171 if (CombinerGlobalAA) {
4172 // Use alias analysis information.
4173 int Overlap1 = Size1 + SrcValueOffset1 + Offset1;
4174 int Overlap2 = Size2 + SrcValueOffset2 + Offset2;
4175 AliasAnalysis::AliasResult AAResult =
4176 AA.alias(SrcValue1, Overlap1, SrcValue2, Overlap2);
4177 if (AAResult == AliasAnalysis::NoAlias)
4181 // Otherwise we have to assume they alias.
4185 /// FindAliasInfo - Extracts the relevant alias information from the memory
4186 /// node. Returns true if the operand was a load.
4187 bool DAGCombiner::FindAliasInfo(SDNode *N,
4188 SDOperand &Ptr, int64_t &Size,
4189 const Value *&SrcValue, int &SrcValueOffset) {
4190 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4191 Ptr = LD->getBasePtr();
4192 Size = MVT::getSizeInBits(LD->getLoadedVT()) >> 3;
4193 SrcValue = LD->getSrcValue();
4194 SrcValueOffset = LD->getSrcValueOffset();
4196 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4197 Ptr = ST->getBasePtr();
4198 Size = MVT::getSizeInBits(ST->getStoredVT()) >> 3;
4199 SrcValue = ST->getSrcValue();
4200 SrcValueOffset = ST->getSrcValueOffset();
4202 assert(0 && "FindAliasInfo expected a memory operand");
4208 /// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes,
4209 /// looking for aliasing nodes and adding them to the Aliases vector.
4210 void DAGCombiner::GatherAllAliases(SDNode *N, SDOperand OriginalChain,
4211 SmallVector<SDOperand, 8> &Aliases) {
4212 SmallVector<SDOperand, 8> Chains; // List of chains to visit.
4213 std::set<SDNode *> Visited; // Visited node set.
4215 // Get alias information for node.
4218 const Value *SrcValue;
4220 bool IsLoad = FindAliasInfo(N, Ptr, Size, SrcValue, SrcValueOffset);
4223 Chains.push_back(OriginalChain);
4225 // Look at each chain and determine if it is an alias. If so, add it to the
4226 // aliases list. If not, then continue up the chain looking for the next
4228 while (!Chains.empty()) {
4229 SDOperand Chain = Chains.back();
4232 // Don't bother if we've been before.
4233 if (Visited.find(Chain.Val) != Visited.end()) continue;
4234 Visited.insert(Chain.Val);
4236 switch (Chain.getOpcode()) {
4237 case ISD::EntryToken:
4238 // Entry token is ideal chain operand, but handled in FindBetterChain.
4243 // Get alias information for Chain.
4246 const Value *OpSrcValue;
4247 int OpSrcValueOffset;
4248 bool IsOpLoad = FindAliasInfo(Chain.Val, OpPtr, OpSize,
4249 OpSrcValue, OpSrcValueOffset);
4251 // If chain is alias then stop here.
4252 if (!(IsLoad && IsOpLoad) &&
4253 isAlias(Ptr, Size, SrcValue, SrcValueOffset,
4254 OpPtr, OpSize, OpSrcValue, OpSrcValueOffset)) {
4255 Aliases.push_back(Chain);
4257 // Look further up the chain.
4258 Chains.push_back(Chain.getOperand(0));
4259 // Clean up old chain.
4260 AddToWorkList(Chain.Val);
4265 case ISD::TokenFactor:
4266 // We have to check each of the operands of the token factor, so we queue
4267 // then up. Adding the operands to the queue (stack) in reverse order
4268 // maintains the original order and increases the likelihood that getNode
4269 // will find a matching token factor (CSE.)
4270 for (unsigned n = Chain.getNumOperands(); n;)
4271 Chains.push_back(Chain.getOperand(--n));
4272 // Eliminate the token factor if we can.
4273 AddToWorkList(Chain.Val);
4277 // For all other instructions we will just have to take what we can get.
4278 Aliases.push_back(Chain);
4284 /// FindBetterChain - Walk up chain skipping non-aliasing memory nodes, looking
4285 /// for a better chain (aliasing node.)
4286 SDOperand DAGCombiner::FindBetterChain(SDNode *N, SDOperand OldChain) {
4287 SmallVector<SDOperand, 8> Aliases; // Ops for replacing token factor.
4289 // Accumulate all the aliases to this node.
4290 GatherAllAliases(N, OldChain, Aliases);
4292 if (Aliases.size() == 0) {
4293 // If no operands then chain to entry token.
4294 return DAG.getEntryNode();
4295 } else if (Aliases.size() == 1) {
4296 // If a single operand then chain to it. We don't need to revisit it.
4300 // Construct a custom tailored token factor.
4301 SDOperand NewChain = DAG.getNode(ISD::TokenFactor, MVT::Other,
4302 &Aliases[0], Aliases.size());
4304 // Make sure the old chain gets cleaned up.
4305 if (NewChain != OldChain) AddToWorkList(OldChain.Val);
4310 // SelectionDAG::Combine - This is the entry point for the file.
4312 void SelectionDAG::Combine(bool RunningAfterLegalize, AliasAnalysis &AA) {
4313 /// run - This is the main entry point to this class.
4315 DAGCombiner(*this, AA).Run(RunningAfterLegalize);