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/Target/TargetOptions.h"
39 #include "llvm/Support/Compiler.h"
40 #include "llvm/Support/CommandLine.h"
44 STATISTIC(NodesCombined , "Number of dag nodes combined");
45 STATISTIC(PreIndexedNodes , "Number of pre-indexed nodes created");
46 STATISTIC(PostIndexedNodes, "Number of post-indexed nodes created");
51 ViewDAGCombine1("view-dag-combine1-dags", cl::Hidden,
52 cl::desc("Pop up a window to show dags before the first "
55 ViewDAGCombine2("view-dag-combine2-dags", cl::Hidden,
56 cl::desc("Pop up a window to show dags before the second "
59 static const bool ViewDAGCombine1 = false;
60 static const bool ViewDAGCombine2 = false;
64 CombinerAA("combiner-alias-analysis", cl::Hidden,
65 cl::desc("Turn on alias analysis during testing"));
68 CombinerGlobalAA("combiner-global-alias-analysis", cl::Hidden,
69 cl::desc("Include global information in alias analysis"));
71 //------------------------------ DAGCombiner ---------------------------------//
73 class VISIBILITY_HIDDEN DAGCombiner {
78 // Worklist of all of the nodes that need to be simplified.
79 std::vector<SDNode*> WorkList;
81 // AA - Used for DAG load/store alias analysis.
84 /// AddUsersToWorkList - When an instruction is simplified, add all users of
85 /// the instruction to the work lists because they might get more simplified
88 void AddUsersToWorkList(SDNode *N) {
89 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
94 /// removeFromWorkList - remove all instances of N from the worklist.
96 void removeFromWorkList(SDNode *N) {
97 WorkList.erase(std::remove(WorkList.begin(), WorkList.end(), N),
102 /// AddToWorkList - Add to the work list making sure it's instance is at the
103 /// the back (next to be processed.)
104 void AddToWorkList(SDNode *N) {
105 removeFromWorkList(N);
106 WorkList.push_back(N);
109 SDOperand CombineTo(SDNode *N, const SDOperand *To, unsigned NumTo,
111 assert(N->getNumValues() == NumTo && "Broken CombineTo call!");
113 DOUT << "\nReplacing.1 "; DEBUG(N->dump());
114 DOUT << "\nWith: "; DEBUG(To[0].Val->dump(&DAG));
115 DOUT << " and " << NumTo-1 << " other values\n";
116 std::vector<SDNode*> NowDead;
117 DAG.ReplaceAllUsesWith(N, To, &NowDead);
120 // Push the new nodes and any users onto the worklist
121 for (unsigned i = 0, e = NumTo; i != e; ++i) {
122 AddToWorkList(To[i].Val);
123 AddUsersToWorkList(To[i].Val);
127 // Nodes can be reintroduced into the worklist. Make sure we do not
128 // process a node that has been replaced.
129 removeFromWorkList(N);
130 for (unsigned i = 0, e = NowDead.size(); i != e; ++i)
131 removeFromWorkList(NowDead[i]);
133 // Finally, since the node is now dead, remove it from the graph.
135 return SDOperand(N, 0);
138 SDOperand CombineTo(SDNode *N, SDOperand Res, bool AddTo = true) {
139 return CombineTo(N, &Res, 1, AddTo);
142 SDOperand CombineTo(SDNode *N, SDOperand Res0, SDOperand Res1,
144 SDOperand To[] = { Res0, Res1 };
145 return CombineTo(N, To, 2, AddTo);
149 /// SimplifyDemandedBits - Check the specified integer node value to see if
150 /// it can be simplified or if things it uses can be simplified by bit
151 /// propagation. If so, return true.
152 bool SimplifyDemandedBits(SDOperand Op) {
153 TargetLowering::TargetLoweringOpt TLO(DAG);
154 uint64_t KnownZero, KnownOne;
155 uint64_t Demanded = MVT::getIntVTBitMask(Op.getValueType());
156 if (!TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
160 AddToWorkList(Op.Val);
162 // Replace the old value with the new one.
164 DOUT << "\nReplacing.2 "; DEBUG(TLO.Old.Val->dump());
165 DOUT << "\nWith: "; DEBUG(TLO.New.Val->dump(&DAG));
168 std::vector<SDNode*> NowDead;
169 DAG.ReplaceAllUsesOfValueWith(TLO.Old, TLO.New, NowDead);
171 // Push the new node and any (possibly new) users onto the worklist.
172 AddToWorkList(TLO.New.Val);
173 AddUsersToWorkList(TLO.New.Val);
175 // Nodes can end up on the worklist more than once. Make sure we do
176 // not process a node that has been replaced.
177 for (unsigned i = 0, e = NowDead.size(); i != e; ++i)
178 removeFromWorkList(NowDead[i]);
180 // Finally, if the node is now dead, remove it from the graph. The node
181 // may not be dead if the replacement process recursively simplified to
182 // something else needing this node.
183 if (TLO.Old.Val->use_empty()) {
184 removeFromWorkList(TLO.Old.Val);
185 DAG.DeleteNode(TLO.Old.Val);
190 bool CombineToPreIndexedLoadStore(SDNode *N);
191 bool CombineToPostIndexedLoadStore(SDNode *N);
194 /// visit - call the node-specific routine that knows how to fold each
195 /// particular type of node.
196 SDOperand visit(SDNode *N);
198 // Visitation implementation - Implement dag node combining for different
199 // node types. The semantics are as follows:
201 // SDOperand.Val == 0 - No change was made
202 // SDOperand.Val == N - N was replaced, is dead, and is already handled.
203 // otherwise - N should be replaced by the returned Operand.
205 SDOperand visitTokenFactor(SDNode *N);
206 SDOperand visitADD(SDNode *N);
207 SDOperand visitSUB(SDNode *N);
208 SDOperand visitMUL(SDNode *N);
209 SDOperand visitSDIV(SDNode *N);
210 SDOperand visitUDIV(SDNode *N);
211 SDOperand visitSREM(SDNode *N);
212 SDOperand visitUREM(SDNode *N);
213 SDOperand visitMULHU(SDNode *N);
214 SDOperand visitMULHS(SDNode *N);
215 SDOperand visitAND(SDNode *N);
216 SDOperand visitOR(SDNode *N);
217 SDOperand visitXOR(SDNode *N);
218 SDOperand visitVBinOp(SDNode *N, ISD::NodeType IntOp, ISD::NodeType FPOp);
219 SDOperand visitSHL(SDNode *N);
220 SDOperand visitSRA(SDNode *N);
221 SDOperand visitSRL(SDNode *N);
222 SDOperand visitCTLZ(SDNode *N);
223 SDOperand visitCTTZ(SDNode *N);
224 SDOperand visitCTPOP(SDNode *N);
225 SDOperand visitSELECT(SDNode *N);
226 SDOperand visitSELECT_CC(SDNode *N);
227 SDOperand visitSETCC(SDNode *N);
228 SDOperand visitSIGN_EXTEND(SDNode *N);
229 SDOperand visitZERO_EXTEND(SDNode *N);
230 SDOperand visitANY_EXTEND(SDNode *N);
231 SDOperand visitSIGN_EXTEND_INREG(SDNode *N);
232 SDOperand visitTRUNCATE(SDNode *N);
233 SDOperand visitBIT_CONVERT(SDNode *N);
234 SDOperand visitVBIT_CONVERT(SDNode *N);
235 SDOperand visitFADD(SDNode *N);
236 SDOperand visitFSUB(SDNode *N);
237 SDOperand visitFMUL(SDNode *N);
238 SDOperand visitFDIV(SDNode *N);
239 SDOperand visitFREM(SDNode *N);
240 SDOperand visitFCOPYSIGN(SDNode *N);
241 SDOperand visitSINT_TO_FP(SDNode *N);
242 SDOperand visitUINT_TO_FP(SDNode *N);
243 SDOperand visitFP_TO_SINT(SDNode *N);
244 SDOperand visitFP_TO_UINT(SDNode *N);
245 SDOperand visitFP_ROUND(SDNode *N);
246 SDOperand visitFP_ROUND_INREG(SDNode *N);
247 SDOperand visitFP_EXTEND(SDNode *N);
248 SDOperand visitFNEG(SDNode *N);
249 SDOperand visitFABS(SDNode *N);
250 SDOperand visitBRCOND(SDNode *N);
251 SDOperand visitBR_CC(SDNode *N);
252 SDOperand visitLOAD(SDNode *N);
253 SDOperand visitSTORE(SDNode *N);
254 SDOperand visitINSERT_VECTOR_ELT(SDNode *N);
255 SDOperand visitVINSERT_VECTOR_ELT(SDNode *N);
256 SDOperand visitVBUILD_VECTOR(SDNode *N);
257 SDOperand visitVECTOR_SHUFFLE(SDNode *N);
258 SDOperand visitVVECTOR_SHUFFLE(SDNode *N);
260 SDOperand XformToShuffleWithZero(SDNode *N);
261 SDOperand ReassociateOps(unsigned Opc, SDOperand LHS, SDOperand RHS);
263 bool SimplifySelectOps(SDNode *SELECT, SDOperand LHS, SDOperand RHS);
264 SDOperand SimplifyBinOpWithSameOpcodeHands(SDNode *N);
265 SDOperand SimplifySelect(SDOperand N0, SDOperand N1, SDOperand N2);
266 SDOperand SimplifySelectCC(SDOperand N0, SDOperand N1, SDOperand N2,
267 SDOperand N3, ISD::CondCode CC);
268 SDOperand SimplifySetCC(MVT::ValueType VT, SDOperand N0, SDOperand N1,
269 ISD::CondCode Cond, bool foldBooleans = true);
270 SDOperand ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(SDNode *, MVT::ValueType);
271 SDOperand BuildSDIV(SDNode *N);
272 SDOperand BuildUDIV(SDNode *N);
273 SDNode *MatchRotate(SDOperand LHS, SDOperand RHS);
275 /// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes,
276 /// looking for aliasing nodes and adding them to the Aliases vector.
277 void GatherAllAliases(SDNode *N, SDOperand OriginalChain,
278 SmallVector<SDOperand, 8> &Aliases);
280 /// isAlias - Return true if there is any possibility that the two addresses
282 bool isAlias(SDOperand Ptr1, int64_t Size1,
283 const Value *SrcValue1, int SrcValueOffset1,
284 SDOperand Ptr2, int64_t Size2,
285 const Value *SrcValue2, int SrcValueOffset2);
287 /// FindAliasInfo - Extracts the relevant alias information from the memory
288 /// node. Returns true if the operand was a load.
289 bool FindAliasInfo(SDNode *N,
290 SDOperand &Ptr, int64_t &Size,
291 const Value *&SrcValue, int &SrcValueOffset);
293 /// FindBetterChain - Walk up chain skipping non-aliasing memory nodes,
294 /// looking for a better chain (aliasing node.)
295 SDOperand FindBetterChain(SDNode *N, SDOperand Chain);
298 DAGCombiner(SelectionDAG &D, AliasAnalysis &A)
300 TLI(D.getTargetLoweringInfo()),
301 AfterLegalize(false),
304 /// Run - runs the dag combiner on all nodes in the work list
305 void Run(bool RunningAfterLegalize);
309 //===----------------------------------------------------------------------===//
310 // TargetLowering::DAGCombinerInfo implementation
311 //===----------------------------------------------------------------------===//
313 void TargetLowering::DAGCombinerInfo::AddToWorklist(SDNode *N) {
314 ((DAGCombiner*)DC)->AddToWorkList(N);
317 SDOperand TargetLowering::DAGCombinerInfo::
318 CombineTo(SDNode *N, const std::vector<SDOperand> &To) {
319 return ((DAGCombiner*)DC)->CombineTo(N, &To[0], To.size());
322 SDOperand TargetLowering::DAGCombinerInfo::
323 CombineTo(SDNode *N, SDOperand Res) {
324 return ((DAGCombiner*)DC)->CombineTo(N, Res);
328 SDOperand TargetLowering::DAGCombinerInfo::
329 CombineTo(SDNode *N, SDOperand Res0, SDOperand Res1) {
330 return ((DAGCombiner*)DC)->CombineTo(N, Res0, Res1);
336 //===----------------------------------------------------------------------===//
339 // isSetCCEquivalent - Return true if this node is a setcc, or is a select_cc
340 // that selects between the values 1 and 0, making it equivalent to a setcc.
341 // Also, set the incoming LHS, RHS, and CC references to the appropriate
342 // nodes based on the type of node we are checking. This simplifies life a
343 // bit for the callers.
344 static bool isSetCCEquivalent(SDOperand N, SDOperand &LHS, SDOperand &RHS,
346 if (N.getOpcode() == ISD::SETCC) {
347 LHS = N.getOperand(0);
348 RHS = N.getOperand(1);
349 CC = N.getOperand(2);
352 if (N.getOpcode() == ISD::SELECT_CC &&
353 N.getOperand(2).getOpcode() == ISD::Constant &&
354 N.getOperand(3).getOpcode() == ISD::Constant &&
355 cast<ConstantSDNode>(N.getOperand(2))->getValue() == 1 &&
356 cast<ConstantSDNode>(N.getOperand(3))->isNullValue()) {
357 LHS = N.getOperand(0);
358 RHS = N.getOperand(1);
359 CC = N.getOperand(4);
365 // isOneUseSetCC - Return true if this is a SetCC-equivalent operation with only
366 // one use. If this is true, it allows the users to invert the operation for
367 // free when it is profitable to do so.
368 static bool isOneUseSetCC(SDOperand N) {
369 SDOperand N0, N1, N2;
370 if (isSetCCEquivalent(N, N0, N1, N2) && N.Val->hasOneUse())
375 SDOperand DAGCombiner::ReassociateOps(unsigned Opc, SDOperand N0, SDOperand N1){
376 MVT::ValueType VT = N0.getValueType();
377 // reassoc. (op (op x, c1), y) -> (op (op x, y), c1) iff x+c1 has one use
378 // reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2))
379 if (N0.getOpcode() == Opc && isa<ConstantSDNode>(N0.getOperand(1))) {
380 if (isa<ConstantSDNode>(N1)) {
381 SDOperand OpNode = DAG.getNode(Opc, VT, N0.getOperand(1), N1);
382 AddToWorkList(OpNode.Val);
383 return DAG.getNode(Opc, VT, OpNode, N0.getOperand(0));
384 } else if (N0.hasOneUse()) {
385 SDOperand OpNode = DAG.getNode(Opc, VT, N0.getOperand(0), N1);
386 AddToWorkList(OpNode.Val);
387 return DAG.getNode(Opc, VT, OpNode, N0.getOperand(1));
390 // reassoc. (op y, (op x, c1)) -> (op (op x, y), c1) iff x+c1 has one use
391 // reassoc. (op c2, (op x, c1)) -> (op x, (op c1, c2))
392 if (N1.getOpcode() == Opc && isa<ConstantSDNode>(N1.getOperand(1))) {
393 if (isa<ConstantSDNode>(N0)) {
394 SDOperand OpNode = DAG.getNode(Opc, VT, N1.getOperand(1), N0);
395 AddToWorkList(OpNode.Val);
396 return DAG.getNode(Opc, VT, OpNode, N1.getOperand(0));
397 } else if (N1.hasOneUse()) {
398 SDOperand OpNode = DAG.getNode(Opc, VT, N1.getOperand(0), N0);
399 AddToWorkList(OpNode.Val);
400 return DAG.getNode(Opc, VT, OpNode, N1.getOperand(1));
406 void DAGCombiner::Run(bool RunningAfterLegalize) {
407 // set the instance variable, so that the various visit routines may use it.
408 AfterLegalize = RunningAfterLegalize;
410 // Add all the dag nodes to the worklist.
411 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
412 E = DAG.allnodes_end(); I != E; ++I)
413 WorkList.push_back(I);
415 // Create a dummy node (which is not added to allnodes), that adds a reference
416 // to the root node, preventing it from being deleted, and tracking any
417 // changes of the root.
418 HandleSDNode Dummy(DAG.getRoot());
420 // The root of the dag may dangle to deleted nodes until the dag combiner is
421 // done. Set it to null to avoid confusion.
422 DAG.setRoot(SDOperand());
424 /// DagCombineInfo - Expose the DAG combiner to the target combiner impls.
425 TargetLowering::DAGCombinerInfo
426 DagCombineInfo(DAG, !RunningAfterLegalize, this);
428 // while the worklist isn't empty, inspect the node on the end of it and
429 // try and combine it.
430 while (!WorkList.empty()) {
431 SDNode *N = WorkList.back();
434 // If N has no uses, it is dead. Make sure to revisit all N's operands once
435 // N is deleted from the DAG, since they too may now be dead or may have a
436 // reduced number of uses, allowing other xforms.
437 if (N->use_empty() && N != &Dummy) {
438 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
439 AddToWorkList(N->getOperand(i).Val);
445 SDOperand RV = visit(N);
447 // If nothing happened, try a target-specific DAG combine.
449 assert(N->getOpcode() != ISD::DELETED_NODE &&
450 "Node was deleted but visit returned NULL!");
451 if (N->getOpcode() >= ISD::BUILTIN_OP_END ||
452 TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode()))
453 RV = TLI.PerformDAGCombine(N, DagCombineInfo);
458 // If we get back the same node we passed in, rather than a new node or
459 // zero, we know that the node must have defined multiple values and
460 // CombineTo was used. Since CombineTo takes care of the worklist
461 // mechanics for us, we have no work to do in this case.
463 assert(N->getOpcode() != ISD::DELETED_NODE &&
464 RV.Val->getOpcode() != ISD::DELETED_NODE &&
465 "Node was deleted but visit returned new node!");
467 DOUT << "\nReplacing.3 "; DEBUG(N->dump());
468 DOUT << "\nWith: "; DEBUG(RV.Val->dump(&DAG));
470 std::vector<SDNode*> NowDead;
471 if (N->getNumValues() == RV.Val->getNumValues())
472 DAG.ReplaceAllUsesWith(N, RV.Val, &NowDead);
474 assert(N->getValueType(0) == RV.getValueType() && "Type mismatch");
476 DAG.ReplaceAllUsesWith(N, &OpV, &NowDead);
479 // Push the new node and any users onto the worklist
480 AddToWorkList(RV.Val);
481 AddUsersToWorkList(RV.Val);
483 // Nodes can be reintroduced into the worklist. Make sure we do not
484 // process a node that has been replaced.
485 removeFromWorkList(N);
486 for (unsigned i = 0, e = NowDead.size(); i != e; ++i)
487 removeFromWorkList(NowDead[i]);
489 // Finally, since the node is now dead, remove it from the graph.
495 // If the root changed (e.g. it was a dead load, update the root).
496 DAG.setRoot(Dummy.getValue());
499 SDOperand DAGCombiner::visit(SDNode *N) {
500 switch(N->getOpcode()) {
502 case ISD::TokenFactor: return visitTokenFactor(N);
503 case ISD::ADD: return visitADD(N);
504 case ISD::SUB: return visitSUB(N);
505 case ISD::MUL: return visitMUL(N);
506 case ISD::SDIV: return visitSDIV(N);
507 case ISD::UDIV: return visitUDIV(N);
508 case ISD::SREM: return visitSREM(N);
509 case ISD::UREM: return visitUREM(N);
510 case ISD::MULHU: return visitMULHU(N);
511 case ISD::MULHS: return visitMULHS(N);
512 case ISD::AND: return visitAND(N);
513 case ISD::OR: return visitOR(N);
514 case ISD::XOR: return visitXOR(N);
515 case ISD::SHL: return visitSHL(N);
516 case ISD::SRA: return visitSRA(N);
517 case ISD::SRL: return visitSRL(N);
518 case ISD::CTLZ: return visitCTLZ(N);
519 case ISD::CTTZ: return visitCTTZ(N);
520 case ISD::CTPOP: return visitCTPOP(N);
521 case ISD::SELECT: return visitSELECT(N);
522 case ISD::SELECT_CC: return visitSELECT_CC(N);
523 case ISD::SETCC: return visitSETCC(N);
524 case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N);
525 case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N);
526 case ISD::ANY_EXTEND: return visitANY_EXTEND(N);
527 case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N);
528 case ISD::TRUNCATE: return visitTRUNCATE(N);
529 case ISD::BIT_CONVERT: return visitBIT_CONVERT(N);
530 case ISD::VBIT_CONVERT: return visitVBIT_CONVERT(N);
531 case ISD::FADD: return visitFADD(N);
532 case ISD::FSUB: return visitFSUB(N);
533 case ISD::FMUL: return visitFMUL(N);
534 case ISD::FDIV: return visitFDIV(N);
535 case ISD::FREM: return visitFREM(N);
536 case ISD::FCOPYSIGN: return visitFCOPYSIGN(N);
537 case ISD::SINT_TO_FP: return visitSINT_TO_FP(N);
538 case ISD::UINT_TO_FP: return visitUINT_TO_FP(N);
539 case ISD::FP_TO_SINT: return visitFP_TO_SINT(N);
540 case ISD::FP_TO_UINT: return visitFP_TO_UINT(N);
541 case ISD::FP_ROUND: return visitFP_ROUND(N);
542 case ISD::FP_ROUND_INREG: return visitFP_ROUND_INREG(N);
543 case ISD::FP_EXTEND: return visitFP_EXTEND(N);
544 case ISD::FNEG: return visitFNEG(N);
545 case ISD::FABS: return visitFABS(N);
546 case ISD::BRCOND: return visitBRCOND(N);
547 case ISD::BR_CC: return visitBR_CC(N);
548 case ISD::LOAD: return visitLOAD(N);
549 case ISD::STORE: return visitSTORE(N);
550 case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N);
551 case ISD::VINSERT_VECTOR_ELT: return visitVINSERT_VECTOR_ELT(N);
552 case ISD::VBUILD_VECTOR: return visitVBUILD_VECTOR(N);
553 case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N);
554 case ISD::VVECTOR_SHUFFLE: return visitVVECTOR_SHUFFLE(N);
555 case ISD::VADD: return visitVBinOp(N, ISD::ADD , ISD::FADD);
556 case ISD::VSUB: return visitVBinOp(N, ISD::SUB , ISD::FSUB);
557 case ISD::VMUL: return visitVBinOp(N, ISD::MUL , ISD::FMUL);
558 case ISD::VSDIV: return visitVBinOp(N, ISD::SDIV, ISD::FDIV);
559 case ISD::VUDIV: return visitVBinOp(N, ISD::UDIV, ISD::UDIV);
560 case ISD::VAND: return visitVBinOp(N, ISD::AND , ISD::AND);
561 case ISD::VOR: return visitVBinOp(N, ISD::OR , ISD::OR);
562 case ISD::VXOR: return visitVBinOp(N, ISD::XOR , ISD::XOR);
567 /// getInputChainForNode - Given a node, return its input chain if it has one,
568 /// otherwise return a null sd operand.
569 static SDOperand getInputChainForNode(SDNode *N) {
570 if (unsigned NumOps = N->getNumOperands()) {
571 if (N->getOperand(0).getValueType() == MVT::Other)
572 return N->getOperand(0);
573 else if (N->getOperand(NumOps-1).getValueType() == MVT::Other)
574 return N->getOperand(NumOps-1);
575 for (unsigned i = 1; i < NumOps-1; ++i)
576 if (N->getOperand(i).getValueType() == MVT::Other)
577 return N->getOperand(i);
579 return SDOperand(0, 0);
582 SDOperand DAGCombiner::visitTokenFactor(SDNode *N) {
583 // If N has two operands, where one has an input chain equal to the other,
584 // the 'other' chain is redundant.
585 if (N->getNumOperands() == 2) {
586 if (getInputChainForNode(N->getOperand(0).Val) == N->getOperand(1))
587 return N->getOperand(0);
588 if (getInputChainForNode(N->getOperand(1).Val) == N->getOperand(0))
589 return N->getOperand(1);
593 SmallVector<SDNode *, 8> TFs; // List of token factors to visit.
594 SmallVector<SDOperand, 8> Ops; // Ops for replacing token factor.
595 bool Changed = false; // If we should replace this token factor.
597 // Start out with this token factor.
600 // Iterate through token factors. The TFs grows when new token factors are
602 for (unsigned i = 0; i < TFs.size(); ++i) {
605 // Check each of the operands.
606 for (unsigned i = 0, ie = TF->getNumOperands(); i != ie; ++i) {
607 SDOperand Op = TF->getOperand(i);
609 switch (Op.getOpcode()) {
610 case ISD::EntryToken:
611 // Entry tokens don't need to be added to the list. They are
616 case ISD::TokenFactor:
617 if ((CombinerAA || Op.hasOneUse()) &&
618 std::find(TFs.begin(), TFs.end(), Op.Val) == TFs.end()) {
619 // Queue up for processing.
620 TFs.push_back(Op.Val);
621 // Clean up in case the token factor is removed.
622 AddToWorkList(Op.Val);
629 // Only add if not there prior.
630 if (std::find(Ops.begin(), Ops.end(), Op) == Ops.end())
639 // If we've change things around then replace token factor.
641 if (Ops.size() == 0) {
642 // The entry token is the only possible outcome.
643 Result = DAG.getEntryNode();
645 // New and improved token factor.
646 Result = DAG.getNode(ISD::TokenFactor, MVT::Other, &Ops[0], Ops.size());
649 // Don't add users to work list.
650 return CombineTo(N, Result, false);
656 SDOperand DAGCombiner::visitADD(SDNode *N) {
657 SDOperand N0 = N->getOperand(0);
658 SDOperand N1 = N->getOperand(1);
659 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
660 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
661 MVT::ValueType VT = N0.getValueType();
663 // fold (add c1, c2) -> c1+c2
665 return DAG.getNode(ISD::ADD, VT, N0, N1);
666 // canonicalize constant to RHS
668 return DAG.getNode(ISD::ADD, VT, N1, N0);
669 // fold (add x, 0) -> x
670 if (N1C && N1C->isNullValue())
672 // fold ((c1-A)+c2) -> (c1+c2)-A
673 if (N1C && N0.getOpcode() == ISD::SUB)
674 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getOperand(0)))
675 return DAG.getNode(ISD::SUB, VT,
676 DAG.getConstant(N1C->getValue()+N0C->getValue(), VT),
679 SDOperand RADD = ReassociateOps(ISD::ADD, N0, N1);
682 // fold ((0-A) + B) -> B-A
683 if (N0.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N0.getOperand(0)) &&
684 cast<ConstantSDNode>(N0.getOperand(0))->isNullValue())
685 return DAG.getNode(ISD::SUB, VT, N1, N0.getOperand(1));
686 // fold (A + (0-B)) -> A-B
687 if (N1.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N1.getOperand(0)) &&
688 cast<ConstantSDNode>(N1.getOperand(0))->isNullValue())
689 return DAG.getNode(ISD::SUB, VT, N0, N1.getOperand(1));
690 // fold (A+(B-A)) -> B
691 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1))
692 return N1.getOperand(0);
694 if (!MVT::isVector(VT) && SimplifyDemandedBits(SDOperand(N, 0)))
695 return SDOperand(N, 0);
697 // fold (a+b) -> (a|b) iff a and b share no bits.
698 if (MVT::isInteger(VT) && !MVT::isVector(VT)) {
699 uint64_t LHSZero, LHSOne;
700 uint64_t RHSZero, RHSOne;
701 uint64_t Mask = MVT::getIntVTBitMask(VT);
702 TLI.ComputeMaskedBits(N0, Mask, LHSZero, LHSOne);
704 TLI.ComputeMaskedBits(N1, Mask, RHSZero, RHSOne);
706 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
707 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
708 if ((RHSZero & (~LHSZero & Mask)) == (~LHSZero & Mask) ||
709 (LHSZero & (~RHSZero & Mask)) == (~RHSZero & Mask))
710 return DAG.getNode(ISD::OR, VT, N0, N1);
717 SDOperand DAGCombiner::visitSUB(SDNode *N) {
718 SDOperand N0 = N->getOperand(0);
719 SDOperand N1 = N->getOperand(1);
720 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val);
721 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val);
722 MVT::ValueType VT = N0.getValueType();
724 // fold (sub x, x) -> 0
726 return DAG.getConstant(0, N->getValueType(0));
727 // fold (sub c1, c2) -> c1-c2
729 return DAG.getNode(ISD::SUB, VT, N0, N1);
730 // fold (sub x, c) -> (add x, -c)
732 return DAG.getNode(ISD::ADD, VT, N0, DAG.getConstant(-N1C->getValue(), VT));
734 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1)
735 return N0.getOperand(1);
737 if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1)
738 return N0.getOperand(0);
742 SDOperand DAGCombiner::visitMUL(SDNode *N) {
743 SDOperand N0 = N->getOperand(0);
744 SDOperand N1 = N->getOperand(1);
745 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
746 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
747 MVT::ValueType VT = N0.getValueType();
749 // fold (mul c1, c2) -> c1*c2
751 return DAG.getNode(ISD::MUL, VT, N0, N1);
752 // canonicalize constant to RHS
754 return DAG.getNode(ISD::MUL, VT, N1, N0);
755 // fold (mul x, 0) -> 0
756 if (N1C && N1C->isNullValue())
758 // fold (mul x, -1) -> 0-x
759 if (N1C && N1C->isAllOnesValue())
760 return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT), N0);
761 // fold (mul x, (1 << c)) -> x << c
762 if (N1C && isPowerOf2_64(N1C->getValue()))
763 return DAG.getNode(ISD::SHL, VT, N0,
764 DAG.getConstant(Log2_64(N1C->getValue()),
765 TLI.getShiftAmountTy()));
766 // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c
767 if (N1C && isPowerOf2_64(-N1C->getSignExtended())) {
768 // FIXME: If the input is something that is easily negated (e.g. a
769 // single-use add), we should put the negate there.
770 return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT),
771 DAG.getNode(ISD::SHL, VT, N0,
772 DAG.getConstant(Log2_64(-N1C->getSignExtended()),
773 TLI.getShiftAmountTy())));
776 // (mul (shl X, c1), c2) -> (mul X, c2 << c1)
777 if (N1C && N0.getOpcode() == ISD::SHL &&
778 isa<ConstantSDNode>(N0.getOperand(1))) {
779 SDOperand C3 = DAG.getNode(ISD::SHL, VT, N1, N0.getOperand(1));
780 AddToWorkList(C3.Val);
781 return DAG.getNode(ISD::MUL, VT, N0.getOperand(0), C3);
784 // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one
787 SDOperand Sh(0,0), Y(0,0);
788 // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)).
789 if (N0.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N0.getOperand(1)) &&
790 N0.Val->hasOneUse()) {
792 } else if (N1.getOpcode() == ISD::SHL &&
793 isa<ConstantSDNode>(N1.getOperand(1)) && N1.Val->hasOneUse()) {
797 SDOperand Mul = DAG.getNode(ISD::MUL, VT, Sh.getOperand(0), Y);
798 return DAG.getNode(ISD::SHL, VT, Mul, Sh.getOperand(1));
801 // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2)
802 if (N1C && N0.getOpcode() == ISD::ADD && N0.Val->hasOneUse() &&
803 isa<ConstantSDNode>(N0.getOperand(1))) {
804 return DAG.getNode(ISD::ADD, VT,
805 DAG.getNode(ISD::MUL, VT, N0.getOperand(0), N1),
806 DAG.getNode(ISD::MUL, VT, N0.getOperand(1), N1));
810 SDOperand RMUL = ReassociateOps(ISD::MUL, N0, N1);
816 SDOperand DAGCombiner::visitSDIV(SDNode *N) {
817 SDOperand N0 = N->getOperand(0);
818 SDOperand N1 = N->getOperand(1);
819 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val);
820 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val);
821 MVT::ValueType VT = N->getValueType(0);
823 // fold (sdiv c1, c2) -> c1/c2
824 if (N0C && N1C && !N1C->isNullValue())
825 return DAG.getNode(ISD::SDIV, VT, N0, N1);
826 // fold (sdiv X, 1) -> X
827 if (N1C && N1C->getSignExtended() == 1LL)
829 // fold (sdiv X, -1) -> 0-X
830 if (N1C && N1C->isAllOnesValue())
831 return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT), N0);
832 // If we know the sign bits of both operands are zero, strength reduce to a
833 // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2
834 uint64_t SignBit = 1ULL << (MVT::getSizeInBits(VT)-1);
835 if (TLI.MaskedValueIsZero(N1, SignBit) &&
836 TLI.MaskedValueIsZero(N0, SignBit))
837 return DAG.getNode(ISD::UDIV, N1.getValueType(), N0, N1);
838 // fold (sdiv X, pow2) -> simple ops after legalize
839 if (N1C && N1C->getValue() && !TLI.isIntDivCheap() &&
840 (isPowerOf2_64(N1C->getSignExtended()) ||
841 isPowerOf2_64(-N1C->getSignExtended()))) {
842 // If dividing by powers of two is cheap, then don't perform the following
844 if (TLI.isPow2DivCheap())
846 int64_t pow2 = N1C->getSignExtended();
847 int64_t abs2 = pow2 > 0 ? pow2 : -pow2;
848 unsigned lg2 = Log2_64(abs2);
849 // Splat the sign bit into the register
850 SDOperand SGN = DAG.getNode(ISD::SRA, VT, N0,
851 DAG.getConstant(MVT::getSizeInBits(VT)-1,
852 TLI.getShiftAmountTy()));
853 AddToWorkList(SGN.Val);
854 // Add (N0 < 0) ? abs2 - 1 : 0;
855 SDOperand SRL = DAG.getNode(ISD::SRL, VT, SGN,
856 DAG.getConstant(MVT::getSizeInBits(VT)-lg2,
857 TLI.getShiftAmountTy()));
858 SDOperand ADD = DAG.getNode(ISD::ADD, VT, N0, SRL);
859 AddToWorkList(SRL.Val);
860 AddToWorkList(ADD.Val); // Divide by pow2
861 SDOperand SRA = DAG.getNode(ISD::SRA, VT, ADD,
862 DAG.getConstant(lg2, TLI.getShiftAmountTy()));
863 // If we're dividing by a positive value, we're done. Otherwise, we must
864 // negate the result.
867 AddToWorkList(SRA.Val);
868 return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT), SRA);
870 // if integer divide is expensive and we satisfy the requirements, emit an
871 // alternate sequence.
872 if (N1C && (N1C->getSignExtended() < -1 || N1C->getSignExtended() > 1) &&
873 !TLI.isIntDivCheap()) {
874 SDOperand Op = BuildSDIV(N);
875 if (Op.Val) return Op;
880 SDOperand DAGCombiner::visitUDIV(SDNode *N) {
881 SDOperand N0 = N->getOperand(0);
882 SDOperand N1 = N->getOperand(1);
883 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val);
884 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val);
885 MVT::ValueType VT = N->getValueType(0);
887 // fold (udiv c1, c2) -> c1/c2
888 if (N0C && N1C && !N1C->isNullValue())
889 return DAG.getNode(ISD::UDIV, VT, N0, N1);
890 // fold (udiv x, (1 << c)) -> x >>u c
891 if (N1C && isPowerOf2_64(N1C->getValue()))
892 return DAG.getNode(ISD::SRL, VT, N0,
893 DAG.getConstant(Log2_64(N1C->getValue()),
894 TLI.getShiftAmountTy()));
895 // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2
896 if (N1.getOpcode() == ISD::SHL) {
897 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
898 if (isPowerOf2_64(SHC->getValue())) {
899 MVT::ValueType ADDVT = N1.getOperand(1).getValueType();
900 SDOperand Add = DAG.getNode(ISD::ADD, ADDVT, N1.getOperand(1),
901 DAG.getConstant(Log2_64(SHC->getValue()),
903 AddToWorkList(Add.Val);
904 return DAG.getNode(ISD::SRL, VT, N0, Add);
908 // fold (udiv x, c) -> alternate
909 if (N1C && N1C->getValue() && !TLI.isIntDivCheap()) {
910 SDOperand Op = BuildUDIV(N);
911 if (Op.Val) return Op;
916 SDOperand DAGCombiner::visitSREM(SDNode *N) {
917 SDOperand N0 = N->getOperand(0);
918 SDOperand N1 = N->getOperand(1);
919 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
920 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
921 MVT::ValueType VT = N->getValueType(0);
923 // fold (srem c1, c2) -> c1%c2
924 if (N0C && N1C && !N1C->isNullValue())
925 return DAG.getNode(ISD::SREM, VT, N0, N1);
926 // If we know the sign bits of both operands are zero, strength reduce to a
927 // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15
928 uint64_t SignBit = 1ULL << (MVT::getSizeInBits(VT)-1);
929 if (TLI.MaskedValueIsZero(N1, SignBit) &&
930 TLI.MaskedValueIsZero(N0, SignBit))
931 return DAG.getNode(ISD::UREM, VT, N0, N1);
933 // Unconditionally lower X%C -> X-X/C*C. This allows the X/C logic to hack on
934 // the remainder operation.
935 if (N1C && !N1C->isNullValue()) {
936 SDOperand Div = DAG.getNode(ISD::SDIV, VT, N0, N1);
937 SDOperand Mul = DAG.getNode(ISD::MUL, VT, Div, N1);
938 SDOperand Sub = DAG.getNode(ISD::SUB, VT, N0, Mul);
939 AddToWorkList(Div.Val);
940 AddToWorkList(Mul.Val);
947 SDOperand DAGCombiner::visitUREM(SDNode *N) {
948 SDOperand N0 = N->getOperand(0);
949 SDOperand N1 = N->getOperand(1);
950 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
951 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
952 MVT::ValueType VT = N->getValueType(0);
954 // fold (urem c1, c2) -> c1%c2
955 if (N0C && N1C && !N1C->isNullValue())
956 return DAG.getNode(ISD::UREM, VT, N0, N1);
957 // fold (urem x, pow2) -> (and x, pow2-1)
958 if (N1C && !N1C->isNullValue() && isPowerOf2_64(N1C->getValue()))
959 return DAG.getNode(ISD::AND, VT, N0, DAG.getConstant(N1C->getValue()-1,VT));
960 // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1))
961 if (N1.getOpcode() == ISD::SHL) {
962 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
963 if (isPowerOf2_64(SHC->getValue())) {
964 SDOperand Add = DAG.getNode(ISD::ADD, VT, N1,DAG.getConstant(~0ULL,VT));
965 AddToWorkList(Add.Val);
966 return DAG.getNode(ISD::AND, VT, N0, Add);
971 // Unconditionally lower X%C -> X-X/C*C. This allows the X/C logic to hack on
972 // the remainder operation.
973 if (N1C && !N1C->isNullValue()) {
974 SDOperand Div = DAG.getNode(ISD::UDIV, VT, N0, N1);
975 SDOperand Mul = DAG.getNode(ISD::MUL, VT, Div, N1);
976 SDOperand Sub = DAG.getNode(ISD::SUB, VT, N0, Mul);
977 AddToWorkList(Div.Val);
978 AddToWorkList(Mul.Val);
985 SDOperand DAGCombiner::visitMULHS(SDNode *N) {
986 SDOperand N0 = N->getOperand(0);
987 SDOperand N1 = N->getOperand(1);
988 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
990 // fold (mulhs x, 0) -> 0
991 if (N1C && N1C->isNullValue())
993 // fold (mulhs x, 1) -> (sra x, size(x)-1)
994 if (N1C && N1C->getValue() == 1)
995 return DAG.getNode(ISD::SRA, N0.getValueType(), N0,
996 DAG.getConstant(MVT::getSizeInBits(N0.getValueType())-1,
997 TLI.getShiftAmountTy()));
1001 SDOperand DAGCombiner::visitMULHU(SDNode *N) {
1002 SDOperand N0 = N->getOperand(0);
1003 SDOperand N1 = N->getOperand(1);
1004 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1006 // fold (mulhu x, 0) -> 0
1007 if (N1C && N1C->isNullValue())
1009 // fold (mulhu x, 1) -> 0
1010 if (N1C && N1C->getValue() == 1)
1011 return DAG.getConstant(0, N0.getValueType());
1015 /// SimplifyBinOpWithSameOpcodeHands - If this is a binary operator with
1016 /// two operands of the same opcode, try to simplify it.
1017 SDOperand DAGCombiner::SimplifyBinOpWithSameOpcodeHands(SDNode *N) {
1018 SDOperand N0 = N->getOperand(0), N1 = N->getOperand(1);
1019 MVT::ValueType VT = N0.getValueType();
1020 assert(N0.getOpcode() == N1.getOpcode() && "Bad input!");
1022 // For each of OP in AND/OR/XOR:
1023 // fold (OP (zext x), (zext y)) -> (zext (OP x, y))
1024 // fold (OP (sext x), (sext y)) -> (sext (OP x, y))
1025 // fold (OP (aext x), (aext y)) -> (aext (OP x, y))
1026 // fold (OP (trunc x), (trunc y)) -> (trunc (OP x, y))
1027 if ((N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND||
1028 N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::TRUNCATE) &&
1029 N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()) {
1030 SDOperand ORNode = DAG.getNode(N->getOpcode(),
1031 N0.getOperand(0).getValueType(),
1032 N0.getOperand(0), N1.getOperand(0));
1033 AddToWorkList(ORNode.Val);
1034 return DAG.getNode(N0.getOpcode(), VT, ORNode);
1037 // For each of OP in SHL/SRL/SRA/AND...
1038 // fold (and (OP x, z), (OP y, z)) -> (OP (and x, y), z)
1039 // fold (or (OP x, z), (OP y, z)) -> (OP (or x, y), z)
1040 // fold (xor (OP x, z), (OP y, z)) -> (OP (xor x, y), z)
1041 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL ||
1042 N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::AND) &&
1043 N0.getOperand(1) == N1.getOperand(1)) {
1044 SDOperand ORNode = DAG.getNode(N->getOpcode(),
1045 N0.getOperand(0).getValueType(),
1046 N0.getOperand(0), N1.getOperand(0));
1047 AddToWorkList(ORNode.Val);
1048 return DAG.getNode(N0.getOpcode(), VT, ORNode, N0.getOperand(1));
1054 SDOperand DAGCombiner::visitAND(SDNode *N) {
1055 SDOperand N0 = N->getOperand(0);
1056 SDOperand N1 = N->getOperand(1);
1057 SDOperand LL, LR, RL, RR, CC0, CC1;
1058 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1059 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1060 MVT::ValueType VT = N1.getValueType();
1062 // fold (and c1, c2) -> c1&c2
1064 return DAG.getNode(ISD::AND, VT, N0, N1);
1065 // canonicalize constant to RHS
1067 return DAG.getNode(ISD::AND, VT, N1, N0);
1068 // fold (and x, -1) -> x
1069 if (N1C && N1C->isAllOnesValue())
1071 // if (and x, c) is known to be zero, return 0
1072 if (N1C && TLI.MaskedValueIsZero(SDOperand(N, 0), MVT::getIntVTBitMask(VT)))
1073 return DAG.getConstant(0, VT);
1075 SDOperand RAND = ReassociateOps(ISD::AND, N0, N1);
1078 // fold (and (or x, 0xFFFF), 0xFF) -> 0xFF
1079 if (N1C && N0.getOpcode() == ISD::OR)
1080 if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
1081 if ((ORI->getValue() & N1C->getValue()) == N1C->getValue())
1083 // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits.
1084 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
1085 unsigned InMask = MVT::getIntVTBitMask(N0.getOperand(0).getValueType());
1086 if (TLI.MaskedValueIsZero(N0.getOperand(0),
1087 ~N1C->getValue() & InMask)) {
1088 SDOperand Zext = DAG.getNode(ISD::ZERO_EXTEND, N0.getValueType(),
1091 // Replace uses of the AND with uses of the Zero extend node.
1094 // We actually want to replace all uses of the any_extend with the
1095 // zero_extend, to avoid duplicating things. This will later cause this
1096 // AND to be folded.
1097 CombineTo(N0.Val, Zext);
1098 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
1101 // fold (and (setcc x), (setcc y)) -> (setcc (and x, y))
1102 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
1103 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
1104 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
1106 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
1107 MVT::isInteger(LL.getValueType())) {
1108 // fold (X == 0) & (Y == 0) -> (X|Y == 0)
1109 if (cast<ConstantSDNode>(LR)->getValue() == 0 && Op1 == ISD::SETEQ) {
1110 SDOperand ORNode = DAG.getNode(ISD::OR, LR.getValueType(), LL, RL);
1111 AddToWorkList(ORNode.Val);
1112 return DAG.getSetCC(VT, ORNode, LR, Op1);
1114 // fold (X == -1) & (Y == -1) -> (X&Y == -1)
1115 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETEQ) {
1116 SDOperand ANDNode = DAG.getNode(ISD::AND, LR.getValueType(), LL, RL);
1117 AddToWorkList(ANDNode.Val);
1118 return DAG.getSetCC(VT, ANDNode, LR, Op1);
1120 // fold (X > -1) & (Y > -1) -> (X|Y > -1)
1121 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETGT) {
1122 SDOperand ORNode = DAG.getNode(ISD::OR, LR.getValueType(), LL, RL);
1123 AddToWorkList(ORNode.Val);
1124 return DAG.getSetCC(VT, ORNode, LR, Op1);
1127 // canonicalize equivalent to ll == rl
1128 if (LL == RR && LR == RL) {
1129 Op1 = ISD::getSetCCSwappedOperands(Op1);
1132 if (LL == RL && LR == RR) {
1133 bool isInteger = MVT::isInteger(LL.getValueType());
1134 ISD::CondCode Result = ISD::getSetCCAndOperation(Op0, Op1, isInteger);
1135 if (Result != ISD::SETCC_INVALID)
1136 return DAG.getSetCC(N0.getValueType(), LL, LR, Result);
1140 // Simplify: and (op x...), (op y...) -> (op (and x, y))
1141 if (N0.getOpcode() == N1.getOpcode()) {
1142 SDOperand Tmp = SimplifyBinOpWithSameOpcodeHands(N);
1143 if (Tmp.Val) return Tmp;
1146 // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1)
1147 // fold (and (sra)) -> (and (srl)) when possible.
1148 if (!MVT::isVector(VT) &&
1149 SimplifyDemandedBits(SDOperand(N, 0)))
1150 return SDOperand(N, 0);
1151 // fold (zext_inreg (extload x)) -> (zextload x)
1152 if (ISD::isEXTLoad(N0.Val)) {
1153 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
1154 MVT::ValueType EVT = LN0->getLoadedVT();
1155 // If we zero all the possible extended bits, then we can turn this into
1156 // a zextload if we are running before legalize or the operation is legal.
1157 if (TLI.MaskedValueIsZero(N1, ~0ULL << MVT::getSizeInBits(EVT)) &&
1158 (!AfterLegalize || TLI.isLoadXLegal(ISD::ZEXTLOAD, EVT))) {
1159 SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, LN0->getChain(),
1160 LN0->getBasePtr(), LN0->getSrcValue(),
1161 LN0->getSrcValueOffset(), EVT);
1163 CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1));
1164 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
1167 // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use
1168 if (ISD::isSEXTLoad(N0.Val) && N0.hasOneUse()) {
1169 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
1170 MVT::ValueType EVT = LN0->getLoadedVT();
1171 // If we zero all the possible extended bits, then we can turn this into
1172 // a zextload if we are running before legalize or the operation is legal.
1173 if (TLI.MaskedValueIsZero(N1, ~0ULL << MVT::getSizeInBits(EVT)) &&
1174 (!AfterLegalize || TLI.isLoadXLegal(ISD::ZEXTLOAD, EVT))) {
1175 SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, LN0->getChain(),
1176 LN0->getBasePtr(), LN0->getSrcValue(),
1177 LN0->getSrcValueOffset(), EVT);
1179 CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1));
1180 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
1184 // fold (and (load x), 255) -> (zextload x, i8)
1185 // fold (and (extload x, i16), 255) -> (zextload x, i8)
1186 if (N1C && N0.getOpcode() == ISD::LOAD) {
1187 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
1188 if (LN0->getExtensionType() != ISD::SEXTLOAD &&
1190 MVT::ValueType EVT, LoadedVT;
1191 if (N1C->getValue() == 255)
1193 else if (N1C->getValue() == 65535)
1195 else if (N1C->getValue() == ~0U)
1200 LoadedVT = LN0->getLoadedVT();
1201 if (EVT != MVT::Other && LoadedVT > EVT &&
1202 (!AfterLegalize || TLI.isLoadXLegal(ISD::ZEXTLOAD, EVT))) {
1203 MVT::ValueType PtrType = N0.getOperand(1).getValueType();
1204 // For big endian targets, we need to add an offset to the pointer to
1205 // load the correct bytes. For little endian systems, we merely need to
1206 // read fewer bytes from the same pointer.
1208 (MVT::getSizeInBits(LoadedVT) - MVT::getSizeInBits(EVT)) / 8;
1209 SDOperand NewPtr = LN0->getBasePtr();
1210 if (!TLI.isLittleEndian())
1211 NewPtr = DAG.getNode(ISD::ADD, PtrType, NewPtr,
1212 DAG.getConstant(PtrOff, PtrType));
1213 AddToWorkList(NewPtr.Val);
1215 DAG.getExtLoad(ISD::ZEXTLOAD, VT, LN0->getChain(), NewPtr,
1216 LN0->getSrcValue(), LN0->getSrcValueOffset(), EVT);
1218 CombineTo(N0.Val, Load, Load.getValue(1));
1219 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
1227 SDOperand DAGCombiner::visitOR(SDNode *N) {
1228 SDOperand N0 = N->getOperand(0);
1229 SDOperand N1 = N->getOperand(1);
1230 SDOperand LL, LR, RL, RR, CC0, CC1;
1231 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1232 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1233 MVT::ValueType VT = N1.getValueType();
1234 unsigned OpSizeInBits = MVT::getSizeInBits(VT);
1236 // fold (or c1, c2) -> c1|c2
1238 return DAG.getNode(ISD::OR, VT, N0, N1);
1239 // canonicalize constant to RHS
1241 return DAG.getNode(ISD::OR, VT, N1, N0);
1242 // fold (or x, 0) -> x
1243 if (N1C && N1C->isNullValue())
1245 // fold (or x, -1) -> -1
1246 if (N1C && N1C->isAllOnesValue())
1248 // fold (or x, c) -> c iff (x & ~c) == 0
1250 TLI.MaskedValueIsZero(N0,~N1C->getValue() & (~0ULL>>(64-OpSizeInBits))))
1253 SDOperand ROR = ReassociateOps(ISD::OR, N0, N1);
1256 // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2)
1257 if (N1C && N0.getOpcode() == ISD::AND && N0.Val->hasOneUse() &&
1258 isa<ConstantSDNode>(N0.getOperand(1))) {
1259 ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1));
1260 return DAG.getNode(ISD::AND, VT, DAG.getNode(ISD::OR, VT, N0.getOperand(0),
1262 DAG.getConstant(N1C->getValue() | C1->getValue(), VT));
1264 // fold (or (setcc x), (setcc y)) -> (setcc (or x, y))
1265 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
1266 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
1267 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
1269 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
1270 MVT::isInteger(LL.getValueType())) {
1271 // fold (X != 0) | (Y != 0) -> (X|Y != 0)
1272 // fold (X < 0) | (Y < 0) -> (X|Y < 0)
1273 if (cast<ConstantSDNode>(LR)->getValue() == 0 &&
1274 (Op1 == ISD::SETNE || Op1 == ISD::SETLT)) {
1275 SDOperand ORNode = DAG.getNode(ISD::OR, LR.getValueType(), LL, RL);
1276 AddToWorkList(ORNode.Val);
1277 return DAG.getSetCC(VT, ORNode, LR, Op1);
1279 // fold (X != -1) | (Y != -1) -> (X&Y != -1)
1280 // fold (X > -1) | (Y > -1) -> (X&Y > -1)
1281 if (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
1282 (Op1 == ISD::SETNE || Op1 == ISD::SETGT)) {
1283 SDOperand ANDNode = DAG.getNode(ISD::AND, LR.getValueType(), LL, RL);
1284 AddToWorkList(ANDNode.Val);
1285 return DAG.getSetCC(VT, ANDNode, LR, Op1);
1288 // canonicalize equivalent to ll == rl
1289 if (LL == RR && LR == RL) {
1290 Op1 = ISD::getSetCCSwappedOperands(Op1);
1293 if (LL == RL && LR == RR) {
1294 bool isInteger = MVT::isInteger(LL.getValueType());
1295 ISD::CondCode Result = ISD::getSetCCOrOperation(Op0, Op1, isInteger);
1296 if (Result != ISD::SETCC_INVALID)
1297 return DAG.getSetCC(N0.getValueType(), LL, LR, Result);
1301 // Simplify: or (op x...), (op y...) -> (op (or x, y))
1302 if (N0.getOpcode() == N1.getOpcode()) {
1303 SDOperand Tmp = SimplifyBinOpWithSameOpcodeHands(N);
1304 if (Tmp.Val) return Tmp;
1307 // (X & C1) | (Y & C2) -> (X|Y) & C3 if possible.
1308 if (N0.getOpcode() == ISD::AND &&
1309 N1.getOpcode() == ISD::AND &&
1310 N0.getOperand(1).getOpcode() == ISD::Constant &&
1311 N1.getOperand(1).getOpcode() == ISD::Constant &&
1312 // Don't increase # computations.
1313 (N0.Val->hasOneUse() || N1.Val->hasOneUse())) {
1314 // We can only do this xform if we know that bits from X that are set in C2
1315 // but not in C1 are already zero. Likewise for Y.
1316 uint64_t LHSMask = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
1317 uint64_t RHSMask = cast<ConstantSDNode>(N1.getOperand(1))->getValue();
1319 if (TLI.MaskedValueIsZero(N0.getOperand(0), RHSMask&~LHSMask) &&
1320 TLI.MaskedValueIsZero(N1.getOperand(0), LHSMask&~RHSMask)) {
1321 SDOperand X =DAG.getNode(ISD::OR, VT, N0.getOperand(0), N1.getOperand(0));
1322 return DAG.getNode(ISD::AND, VT, X, DAG.getConstant(LHSMask|RHSMask, VT));
1327 // See if this is some rotate idiom.
1328 if (SDNode *Rot = MatchRotate(N0, N1))
1329 return SDOperand(Rot, 0);
1335 /// MatchRotateHalf - Match "(X shl/srl V1) & V2" where V2 may not be present.
1336 static bool MatchRotateHalf(SDOperand Op, SDOperand &Shift, SDOperand &Mask) {
1337 if (Op.getOpcode() == ISD::AND) {
1338 if (isa<ConstantSDNode>(Op.getOperand(1))) {
1339 Mask = Op.getOperand(1);
1340 Op = Op.getOperand(0);
1346 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SHL) {
1354 // MatchRotate - Handle an 'or' of two operands. If this is one of the many
1355 // idioms for rotate, and if the target supports rotation instructions, generate
1357 SDNode *DAGCombiner::MatchRotate(SDOperand LHS, SDOperand RHS) {
1358 // Must be a legal type. Expanded an promoted things won't work with rotates.
1359 MVT::ValueType VT = LHS.getValueType();
1360 if (!TLI.isTypeLegal(VT)) return 0;
1362 // The target must have at least one rotate flavor.
1363 bool HasROTL = TLI.isOperationLegal(ISD::ROTL, VT);
1364 bool HasROTR = TLI.isOperationLegal(ISD::ROTR, VT);
1365 if (!HasROTL && !HasROTR) return 0;
1367 // Match "(X shl/srl V1) & V2" where V2 may not be present.
1368 SDOperand LHSShift; // The shift.
1369 SDOperand LHSMask; // AND value if any.
1370 if (!MatchRotateHalf(LHS, LHSShift, LHSMask))
1371 return 0; // Not part of a rotate.
1373 SDOperand RHSShift; // The shift.
1374 SDOperand RHSMask; // AND value if any.
1375 if (!MatchRotateHalf(RHS, RHSShift, RHSMask))
1376 return 0; // Not part of a rotate.
1378 if (LHSShift.getOperand(0) != RHSShift.getOperand(0))
1379 return 0; // Not shifting the same value.
1381 if (LHSShift.getOpcode() == RHSShift.getOpcode())
1382 return 0; // Shifts must disagree.
1384 // Canonicalize shl to left side in a shl/srl pair.
1385 if (RHSShift.getOpcode() == ISD::SHL) {
1386 std::swap(LHS, RHS);
1387 std::swap(LHSShift, RHSShift);
1388 std::swap(LHSMask , RHSMask );
1391 unsigned OpSizeInBits = MVT::getSizeInBits(VT);
1393 // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1)
1394 // fold (or (shl x, C1), (srl x, C2)) -> (rotr x, C2)
1395 if (LHSShift.getOperand(1).getOpcode() == ISD::Constant &&
1396 RHSShift.getOperand(1).getOpcode() == ISD::Constant) {
1397 uint64_t LShVal = cast<ConstantSDNode>(LHSShift.getOperand(1))->getValue();
1398 uint64_t RShVal = cast<ConstantSDNode>(RHSShift.getOperand(1))->getValue();
1399 if ((LShVal + RShVal) != OpSizeInBits)
1404 Rot = DAG.getNode(ISD::ROTL, VT, LHSShift.getOperand(0),
1405 LHSShift.getOperand(1));
1407 Rot = DAG.getNode(ISD::ROTR, VT, LHSShift.getOperand(0),
1408 RHSShift.getOperand(1));
1410 // If there is an AND of either shifted operand, apply it to the result.
1411 if (LHSMask.Val || RHSMask.Val) {
1412 uint64_t Mask = MVT::getIntVTBitMask(VT);
1415 uint64_t RHSBits = (1ULL << LShVal)-1;
1416 Mask &= cast<ConstantSDNode>(LHSMask)->getValue() | RHSBits;
1419 uint64_t LHSBits = ~((1ULL << (OpSizeInBits-RShVal))-1);
1420 Mask &= cast<ConstantSDNode>(RHSMask)->getValue() | LHSBits;
1423 Rot = DAG.getNode(ISD::AND, VT, Rot, DAG.getConstant(Mask, VT));
1429 // If there is a mask here, and we have a variable shift, we can't be sure
1430 // that we're masking out the right stuff.
1431 if (LHSMask.Val || RHSMask.Val)
1434 // fold (or (shl x, y), (srl x, (sub 32, y))) -> (rotl x, y)
1435 // fold (or (shl x, y), (srl x, (sub 32, y))) -> (rotr x, (sub 32, y))
1436 if (RHSShift.getOperand(1).getOpcode() == ISD::SUB &&
1437 LHSShift.getOperand(1) == RHSShift.getOperand(1).getOperand(1)) {
1438 if (ConstantSDNode *SUBC =
1439 dyn_cast<ConstantSDNode>(RHSShift.getOperand(1).getOperand(0))) {
1440 if (SUBC->getValue() == OpSizeInBits)
1442 return DAG.getNode(ISD::ROTL, VT, LHSShift.getOperand(0),
1443 LHSShift.getOperand(1)).Val;
1445 return DAG.getNode(ISD::ROTR, VT, LHSShift.getOperand(0),
1446 LHSShift.getOperand(1)).Val;
1450 // fold (or (shl x, (sub 32, y)), (srl x, r)) -> (rotr x, y)
1451 // fold (or (shl x, (sub 32, y)), (srl x, r)) -> (rotl x, (sub 32, y))
1452 if (LHSShift.getOperand(1).getOpcode() == ISD::SUB &&
1453 RHSShift.getOperand(1) == LHSShift.getOperand(1).getOperand(1)) {
1454 if (ConstantSDNode *SUBC =
1455 dyn_cast<ConstantSDNode>(LHSShift.getOperand(1).getOperand(0))) {
1456 if (SUBC->getValue() == OpSizeInBits)
1458 return DAG.getNode(ISD::ROTL, VT, LHSShift.getOperand(0),
1459 LHSShift.getOperand(1)).Val;
1461 return DAG.getNode(ISD::ROTR, VT, LHSShift.getOperand(0),
1462 RHSShift.getOperand(1)).Val;
1470 SDOperand DAGCombiner::visitXOR(SDNode *N) {
1471 SDOperand N0 = N->getOperand(0);
1472 SDOperand N1 = N->getOperand(1);
1473 SDOperand LHS, RHS, CC;
1474 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1475 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1476 MVT::ValueType VT = N0.getValueType();
1478 // fold (xor c1, c2) -> c1^c2
1480 return DAG.getNode(ISD::XOR, VT, N0, N1);
1481 // canonicalize constant to RHS
1483 return DAG.getNode(ISD::XOR, VT, N1, N0);
1484 // fold (xor x, 0) -> x
1485 if (N1C && N1C->isNullValue())
1488 SDOperand RXOR = ReassociateOps(ISD::XOR, N0, N1);
1491 // fold !(x cc y) -> (x !cc y)
1492 if (N1C && N1C->getValue() == 1 && isSetCCEquivalent(N0, LHS, RHS, CC)) {
1493 bool isInt = MVT::isInteger(LHS.getValueType());
1494 ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
1496 if (N0.getOpcode() == ISD::SETCC)
1497 return DAG.getSetCC(VT, LHS, RHS, NotCC);
1498 if (N0.getOpcode() == ISD::SELECT_CC)
1499 return DAG.getSelectCC(LHS, RHS, N0.getOperand(2),N0.getOperand(3),NotCC);
1500 assert(0 && "Unhandled SetCC Equivalent!");
1503 // fold !(x or y) -> (!x and !y) iff x or y are setcc
1504 if (N1C && N1C->getValue() == 1 && VT == MVT::i1 &&
1505 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
1506 SDOperand LHS = N0.getOperand(0), RHS = N0.getOperand(1);
1507 if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) {
1508 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
1509 LHS = DAG.getNode(ISD::XOR, VT, LHS, N1); // RHS = ~LHS
1510 RHS = DAG.getNode(ISD::XOR, VT, RHS, N1); // RHS = ~RHS
1511 AddToWorkList(LHS.Val); AddToWorkList(RHS.Val);
1512 return DAG.getNode(NewOpcode, VT, LHS, RHS);
1515 // fold !(x or y) -> (!x and !y) iff x or y are constants
1516 if (N1C && N1C->isAllOnesValue() &&
1517 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
1518 SDOperand LHS = N0.getOperand(0), RHS = N0.getOperand(1);
1519 if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) {
1520 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
1521 LHS = DAG.getNode(ISD::XOR, VT, LHS, N1); // RHS = ~LHS
1522 RHS = DAG.getNode(ISD::XOR, VT, RHS, N1); // RHS = ~RHS
1523 AddToWorkList(LHS.Val); AddToWorkList(RHS.Val);
1524 return DAG.getNode(NewOpcode, VT, LHS, RHS);
1527 // fold (xor (xor x, c1), c2) -> (xor x, c1^c2)
1528 if (N1C && N0.getOpcode() == ISD::XOR) {
1529 ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0));
1530 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
1532 return DAG.getNode(ISD::XOR, VT, N0.getOperand(1),
1533 DAG.getConstant(N1C->getValue()^N00C->getValue(), VT));
1535 return DAG.getNode(ISD::XOR, VT, N0.getOperand(0),
1536 DAG.getConstant(N1C->getValue()^N01C->getValue(), VT));
1538 // fold (xor x, x) -> 0
1540 if (!MVT::isVector(VT)) {
1541 return DAG.getConstant(0, VT);
1542 } else if (!AfterLegalize || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT)) {
1543 // Produce a vector of zeros.
1544 SDOperand El = DAG.getConstant(0, MVT::getVectorBaseType(VT));
1545 std::vector<SDOperand> Ops(MVT::getVectorNumElements(VT), El);
1546 return DAG.getNode(ISD::BUILD_VECTOR, VT, &Ops[0], Ops.size());
1550 // Simplify: xor (op x...), (op y...) -> (op (xor x, y))
1551 if (N0.getOpcode() == N1.getOpcode()) {
1552 SDOperand Tmp = SimplifyBinOpWithSameOpcodeHands(N);
1553 if (Tmp.Val) return Tmp;
1556 // Simplify the expression using non-local knowledge.
1557 if (!MVT::isVector(VT) &&
1558 SimplifyDemandedBits(SDOperand(N, 0)))
1559 return SDOperand(N, 0);
1564 SDOperand DAGCombiner::visitSHL(SDNode *N) {
1565 SDOperand N0 = N->getOperand(0);
1566 SDOperand N1 = N->getOperand(1);
1567 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1568 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1569 MVT::ValueType VT = N0.getValueType();
1570 unsigned OpSizeInBits = MVT::getSizeInBits(VT);
1572 // fold (shl c1, c2) -> c1<<c2
1574 return DAG.getNode(ISD::SHL, VT, N0, N1);
1575 // fold (shl 0, x) -> 0
1576 if (N0C && N0C->isNullValue())
1578 // fold (shl x, c >= size(x)) -> undef
1579 if (N1C && N1C->getValue() >= OpSizeInBits)
1580 return DAG.getNode(ISD::UNDEF, VT);
1581 // fold (shl x, 0) -> x
1582 if (N1C && N1C->isNullValue())
1584 // if (shl x, c) is known to be zero, return 0
1585 if (TLI.MaskedValueIsZero(SDOperand(N, 0), MVT::getIntVTBitMask(VT)))
1586 return DAG.getConstant(0, VT);
1587 if (SimplifyDemandedBits(SDOperand(N, 0)))
1588 return SDOperand(N, 0);
1589 // fold (shl (shl x, c1), c2) -> 0 or (shl x, c1+c2)
1590 if (N1C && N0.getOpcode() == ISD::SHL &&
1591 N0.getOperand(1).getOpcode() == ISD::Constant) {
1592 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
1593 uint64_t c2 = N1C->getValue();
1594 if (c1 + c2 > OpSizeInBits)
1595 return DAG.getConstant(0, VT);
1596 return DAG.getNode(ISD::SHL, VT, N0.getOperand(0),
1597 DAG.getConstant(c1 + c2, N1.getValueType()));
1599 // fold (shl (srl x, c1), c2) -> (shl (and x, -1 << c1), c2-c1) or
1600 // (srl (and x, -1 << c1), c1-c2)
1601 if (N1C && N0.getOpcode() == ISD::SRL &&
1602 N0.getOperand(1).getOpcode() == ISD::Constant) {
1603 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
1604 uint64_t c2 = N1C->getValue();
1605 SDOperand Mask = DAG.getNode(ISD::AND, VT, N0.getOperand(0),
1606 DAG.getConstant(~0ULL << c1, VT));
1608 return DAG.getNode(ISD::SHL, VT, Mask,
1609 DAG.getConstant(c2-c1, N1.getValueType()));
1611 return DAG.getNode(ISD::SRL, VT, Mask,
1612 DAG.getConstant(c1-c2, N1.getValueType()));
1614 // fold (shl (sra x, c1), c1) -> (and x, -1 << c1)
1615 if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1))
1616 return DAG.getNode(ISD::AND, VT, N0.getOperand(0),
1617 DAG.getConstant(~0ULL << N1C->getValue(), VT));
1618 // fold (shl (add x, c1), c2) -> (add (shl x, c2), c1<<c2)
1619 if (N1C && N0.getOpcode() == ISD::ADD && N0.Val->hasOneUse() &&
1620 isa<ConstantSDNode>(N0.getOperand(1))) {
1621 return DAG.getNode(ISD::ADD, VT,
1622 DAG.getNode(ISD::SHL, VT, N0.getOperand(0), N1),
1623 DAG.getNode(ISD::SHL, VT, N0.getOperand(1), N1));
1628 SDOperand DAGCombiner::visitSRA(SDNode *N) {
1629 SDOperand N0 = N->getOperand(0);
1630 SDOperand N1 = N->getOperand(1);
1631 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1632 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1633 MVT::ValueType VT = N0.getValueType();
1635 // fold (sra c1, c2) -> c1>>c2
1637 return DAG.getNode(ISD::SRA, VT, N0, N1);
1638 // fold (sra 0, x) -> 0
1639 if (N0C && N0C->isNullValue())
1641 // fold (sra -1, x) -> -1
1642 if (N0C && N0C->isAllOnesValue())
1644 // fold (sra x, c >= size(x)) -> undef
1645 if (N1C && N1C->getValue() >= MVT::getSizeInBits(VT))
1646 return DAG.getNode(ISD::UNDEF, VT);
1647 // fold (sra x, 0) -> x
1648 if (N1C && N1C->isNullValue())
1650 // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports
1652 if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) {
1653 unsigned LowBits = MVT::getSizeInBits(VT) - (unsigned)N1C->getValue();
1656 default: EVT = MVT::Other; break;
1657 case 1: EVT = MVT::i1; break;
1658 case 8: EVT = MVT::i8; break;
1659 case 16: EVT = MVT::i16; break;
1660 case 32: EVT = MVT::i32; break;
1662 if (EVT > MVT::Other && TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, EVT))
1663 return DAG.getNode(ISD::SIGN_EXTEND_INREG, VT, N0.getOperand(0),
1664 DAG.getValueType(EVT));
1667 // fold (sra (sra x, c1), c2) -> (sra x, c1+c2)
1668 if (N1C && N0.getOpcode() == ISD::SRA) {
1669 if (ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1670 unsigned Sum = N1C->getValue() + C1->getValue();
1671 if (Sum >= MVT::getSizeInBits(VT)) Sum = MVT::getSizeInBits(VT)-1;
1672 return DAG.getNode(ISD::SRA, VT, N0.getOperand(0),
1673 DAG.getConstant(Sum, N1C->getValueType(0)));
1677 // Simplify, based on bits shifted out of the LHS.
1678 if (N1C && SimplifyDemandedBits(SDOperand(N, 0)))
1679 return SDOperand(N, 0);
1682 // If the sign bit is known to be zero, switch this to a SRL.
1683 if (TLI.MaskedValueIsZero(N0, MVT::getIntVTSignBit(VT)))
1684 return DAG.getNode(ISD::SRL, VT, N0, N1);
1688 SDOperand DAGCombiner::visitSRL(SDNode *N) {
1689 SDOperand N0 = N->getOperand(0);
1690 SDOperand N1 = N->getOperand(1);
1691 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1692 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1693 MVT::ValueType VT = N0.getValueType();
1694 unsigned OpSizeInBits = MVT::getSizeInBits(VT);
1696 // fold (srl c1, c2) -> c1 >>u c2
1698 return DAG.getNode(ISD::SRL, VT, N0, N1);
1699 // fold (srl 0, x) -> 0
1700 if (N0C && N0C->isNullValue())
1702 // fold (srl x, c >= size(x)) -> undef
1703 if (N1C && N1C->getValue() >= OpSizeInBits)
1704 return DAG.getNode(ISD::UNDEF, VT);
1705 // fold (srl x, 0) -> x
1706 if (N1C && N1C->isNullValue())
1708 // if (srl x, c) is known to be zero, return 0
1709 if (N1C && TLI.MaskedValueIsZero(SDOperand(N, 0), ~0ULL >> (64-OpSizeInBits)))
1710 return DAG.getConstant(0, VT);
1711 // fold (srl (srl x, c1), c2) -> 0 or (srl x, c1+c2)
1712 if (N1C && N0.getOpcode() == ISD::SRL &&
1713 N0.getOperand(1).getOpcode() == ISD::Constant) {
1714 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
1715 uint64_t c2 = N1C->getValue();
1716 if (c1 + c2 > OpSizeInBits)
1717 return DAG.getConstant(0, VT);
1718 return DAG.getNode(ISD::SRL, VT, N0.getOperand(0),
1719 DAG.getConstant(c1 + c2, N1.getValueType()));
1722 // fold (srl (anyextend x), c) -> (anyextend (srl x, c))
1723 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
1724 // Shifting in all undef bits?
1725 MVT::ValueType SmallVT = N0.getOperand(0).getValueType();
1726 if (N1C->getValue() >= MVT::getSizeInBits(SmallVT))
1727 return DAG.getNode(ISD::UNDEF, VT);
1729 SDOperand SmallShift = DAG.getNode(ISD::SRL, SmallVT, N0.getOperand(0), N1);
1730 AddToWorkList(SmallShift.Val);
1731 return DAG.getNode(ISD::ANY_EXTEND, VT, SmallShift);
1734 // fold (srl (sra X, Y), 31) -> (srl X, 31). This srl only looks at the sign
1735 // bit, which is unmodified by sra.
1736 if (N1C && N1C->getValue()+1 == MVT::getSizeInBits(VT)) {
1737 if (N0.getOpcode() == ISD::SRA)
1738 return DAG.getNode(ISD::SRL, VT, N0.getOperand(0), N1);
1741 // fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit).
1742 if (N1C && N0.getOpcode() == ISD::CTLZ &&
1743 N1C->getValue() == Log2_32(MVT::getSizeInBits(VT))) {
1744 uint64_t KnownZero, KnownOne, Mask = MVT::getIntVTBitMask(VT);
1745 TLI.ComputeMaskedBits(N0.getOperand(0), Mask, KnownZero, KnownOne);
1747 // If any of the input bits are KnownOne, then the input couldn't be all
1748 // zeros, thus the result of the srl will always be zero.
1749 if (KnownOne) return DAG.getConstant(0, VT);
1751 // If all of the bits input the to ctlz node are known to be zero, then
1752 // the result of the ctlz is "32" and the result of the shift is one.
1753 uint64_t UnknownBits = ~KnownZero & Mask;
1754 if (UnknownBits == 0) return DAG.getConstant(1, VT);
1756 // Otherwise, check to see if there is exactly one bit input to the ctlz.
1757 if ((UnknownBits & (UnknownBits-1)) == 0) {
1758 // Okay, we know that only that the single bit specified by UnknownBits
1759 // could be set on input to the CTLZ node. If this bit is set, the SRL
1760 // will return 0, if it is clear, it returns 1. Change the CTLZ/SRL pair
1761 // to an SRL,XOR pair, which is likely to simplify more.
1762 unsigned ShAmt = CountTrailingZeros_64(UnknownBits);
1763 SDOperand Op = N0.getOperand(0);
1765 Op = DAG.getNode(ISD::SRL, VT, Op,
1766 DAG.getConstant(ShAmt, TLI.getShiftAmountTy()));
1767 AddToWorkList(Op.Val);
1769 return DAG.getNode(ISD::XOR, VT, Op, DAG.getConstant(1, VT));
1776 SDOperand DAGCombiner::visitCTLZ(SDNode *N) {
1777 SDOperand N0 = N->getOperand(0);
1778 MVT::ValueType VT = N->getValueType(0);
1780 // fold (ctlz c1) -> c2
1781 if (isa<ConstantSDNode>(N0))
1782 return DAG.getNode(ISD::CTLZ, VT, N0);
1786 SDOperand DAGCombiner::visitCTTZ(SDNode *N) {
1787 SDOperand N0 = N->getOperand(0);
1788 MVT::ValueType VT = N->getValueType(0);
1790 // fold (cttz c1) -> c2
1791 if (isa<ConstantSDNode>(N0))
1792 return DAG.getNode(ISD::CTTZ, VT, N0);
1796 SDOperand DAGCombiner::visitCTPOP(SDNode *N) {
1797 SDOperand N0 = N->getOperand(0);
1798 MVT::ValueType VT = N->getValueType(0);
1800 // fold (ctpop c1) -> c2
1801 if (isa<ConstantSDNode>(N0))
1802 return DAG.getNode(ISD::CTPOP, VT, N0);
1806 SDOperand DAGCombiner::visitSELECT(SDNode *N) {
1807 SDOperand N0 = N->getOperand(0);
1808 SDOperand N1 = N->getOperand(1);
1809 SDOperand N2 = N->getOperand(2);
1810 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1811 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1812 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
1813 MVT::ValueType VT = N->getValueType(0);
1815 // fold select C, X, X -> X
1818 // fold select true, X, Y -> X
1819 if (N0C && !N0C->isNullValue())
1821 // fold select false, X, Y -> Y
1822 if (N0C && N0C->isNullValue())
1824 // fold select C, 1, X -> C | X
1825 if (MVT::i1 == VT && N1C && N1C->getValue() == 1)
1826 return DAG.getNode(ISD::OR, VT, N0, N2);
1827 // fold select C, 0, X -> ~C & X
1828 // FIXME: this should check for C type == X type, not i1?
1829 if (MVT::i1 == VT && N1C && N1C->isNullValue()) {
1830 SDOperand XORNode = DAG.getNode(ISD::XOR, VT, N0, DAG.getConstant(1, VT));
1831 AddToWorkList(XORNode.Val);
1832 return DAG.getNode(ISD::AND, VT, XORNode, N2);
1834 // fold select C, X, 1 -> ~C | X
1835 if (MVT::i1 == VT && N2C && N2C->getValue() == 1) {
1836 SDOperand XORNode = DAG.getNode(ISD::XOR, VT, N0, DAG.getConstant(1, VT));
1837 AddToWorkList(XORNode.Val);
1838 return DAG.getNode(ISD::OR, VT, XORNode, N1);
1840 // fold select C, X, 0 -> C & X
1841 // FIXME: this should check for C type == X type, not i1?
1842 if (MVT::i1 == VT && N2C && N2C->isNullValue())
1843 return DAG.getNode(ISD::AND, VT, N0, N1);
1844 // fold X ? X : Y --> X ? 1 : Y --> X | Y
1845 if (MVT::i1 == VT && N0 == N1)
1846 return DAG.getNode(ISD::OR, VT, N0, N2);
1847 // fold X ? Y : X --> X ? Y : 0 --> X & Y
1848 if (MVT::i1 == VT && N0 == N2)
1849 return DAG.getNode(ISD::AND, VT, N0, N1);
1851 // If we can fold this based on the true/false value, do so.
1852 if (SimplifySelectOps(N, N1, N2))
1853 return SDOperand(N, 0); // Don't revisit N.
1855 // fold selects based on a setcc into other things, such as min/max/abs
1856 if (N0.getOpcode() == ISD::SETCC)
1858 // Check against MVT::Other for SELECT_CC, which is a workaround for targets
1859 // having to say they don't support SELECT_CC on every type the DAG knows
1860 // about, since there is no way to mark an opcode illegal at all value types
1861 if (TLI.isOperationLegal(ISD::SELECT_CC, MVT::Other))
1862 return DAG.getNode(ISD::SELECT_CC, VT, N0.getOperand(0), N0.getOperand(1),
1863 N1, N2, N0.getOperand(2));
1865 return SimplifySelect(N0, N1, N2);
1869 SDOperand DAGCombiner::visitSELECT_CC(SDNode *N) {
1870 SDOperand N0 = N->getOperand(0);
1871 SDOperand N1 = N->getOperand(1);
1872 SDOperand N2 = N->getOperand(2);
1873 SDOperand N3 = N->getOperand(3);
1874 SDOperand N4 = N->getOperand(4);
1875 ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get();
1877 // fold select_cc lhs, rhs, x, x, cc -> x
1881 // Determine if the condition we're dealing with is constant
1882 SDOperand SCC = SimplifySetCC(TLI.getSetCCResultTy(), N0, N1, CC, false);
1883 if (SCC.Val) AddToWorkList(SCC.Val);
1885 if (ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.Val)) {
1886 if (SCCC->getValue())
1887 return N2; // cond always true -> true val
1889 return N3; // cond always false -> false val
1892 // Fold to a simpler select_cc
1893 if (SCC.Val && SCC.getOpcode() == ISD::SETCC)
1894 return DAG.getNode(ISD::SELECT_CC, N2.getValueType(),
1895 SCC.getOperand(0), SCC.getOperand(1), N2, N3,
1898 // If we can fold this based on the true/false value, do so.
1899 if (SimplifySelectOps(N, N2, N3))
1900 return SDOperand(N, 0); // Don't revisit N.
1902 // fold select_cc into other things, such as min/max/abs
1903 return SimplifySelectCC(N0, N1, N2, N3, CC);
1906 SDOperand DAGCombiner::visitSETCC(SDNode *N) {
1907 return SimplifySetCC(N->getValueType(0), N->getOperand(0), N->getOperand(1),
1908 cast<CondCodeSDNode>(N->getOperand(2))->get());
1911 SDOperand DAGCombiner::visitSIGN_EXTEND(SDNode *N) {
1912 SDOperand N0 = N->getOperand(0);
1913 MVT::ValueType VT = N->getValueType(0);
1915 // fold (sext c1) -> c1
1916 if (isa<ConstantSDNode>(N0))
1917 return DAG.getNode(ISD::SIGN_EXTEND, VT, N0);
1919 // fold (sext (sext x)) -> (sext x)
1920 // fold (sext (aext x)) -> (sext x)
1921 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
1922 return DAG.getNode(ISD::SIGN_EXTEND, VT, N0.getOperand(0));
1924 // fold (sext (truncate x)) -> (sextinreg x).
1925 if (N0.getOpcode() == ISD::TRUNCATE &&
1926 (!AfterLegalize || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG,
1927 N0.getValueType()))) {
1928 SDOperand Op = N0.getOperand(0);
1929 if (Op.getValueType() < VT) {
1930 Op = DAG.getNode(ISD::ANY_EXTEND, VT, Op);
1931 } else if (Op.getValueType() > VT) {
1932 Op = DAG.getNode(ISD::TRUNCATE, VT, Op);
1934 return DAG.getNode(ISD::SIGN_EXTEND_INREG, VT, Op,
1935 DAG.getValueType(N0.getValueType()));
1938 // fold (sext (load x)) -> (sext (truncate (sextload x)))
1939 if (ISD::isNON_EXTLoad(N0.Val) && N0.hasOneUse() &&
1940 (!AfterLegalize||TLI.isLoadXLegal(ISD::SEXTLOAD, N0.getValueType()))){
1941 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
1942 SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, LN0->getChain(),
1943 LN0->getBasePtr(), LN0->getSrcValue(),
1944 LN0->getSrcValueOffset(),
1946 CombineTo(N, ExtLoad);
1947 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
1948 ExtLoad.getValue(1));
1949 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
1952 // fold (sext (sextload x)) -> (sext (truncate (sextload x)))
1953 // fold (sext ( extload x)) -> (sext (truncate (sextload x)))
1954 if ((ISD::isSEXTLoad(N0.Val) || ISD::isEXTLoad(N0.Val)) && N0.hasOneUse()) {
1955 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
1956 MVT::ValueType EVT = LN0->getLoadedVT();
1957 if (!AfterLegalize || TLI.isLoadXLegal(ISD::SEXTLOAD, EVT)) {
1958 SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, LN0->getChain(),
1959 LN0->getBasePtr(), LN0->getSrcValue(),
1960 LN0->getSrcValueOffset(), EVT);
1961 CombineTo(N, ExtLoad);
1962 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
1963 ExtLoad.getValue(1));
1964 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
1971 SDOperand DAGCombiner::visitZERO_EXTEND(SDNode *N) {
1972 SDOperand N0 = N->getOperand(0);
1973 MVT::ValueType VT = N->getValueType(0);
1975 // fold (zext c1) -> c1
1976 if (isa<ConstantSDNode>(N0))
1977 return DAG.getNode(ISD::ZERO_EXTEND, VT, N0);
1978 // fold (zext (zext x)) -> (zext x)
1979 // fold (zext (aext x)) -> (zext x)
1980 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
1981 return DAG.getNode(ISD::ZERO_EXTEND, VT, N0.getOperand(0));
1983 // fold (zext (truncate x)) -> (and x, mask)
1984 if (N0.getOpcode() == ISD::TRUNCATE &&
1985 (!AfterLegalize || TLI.isOperationLegal(ISD::AND, VT))) {
1986 SDOperand Op = N0.getOperand(0);
1987 if (Op.getValueType() < VT) {
1988 Op = DAG.getNode(ISD::ANY_EXTEND, VT, Op);
1989 } else if (Op.getValueType() > VT) {
1990 Op = DAG.getNode(ISD::TRUNCATE, VT, Op);
1992 return DAG.getZeroExtendInReg(Op, N0.getValueType());
1995 // fold (zext (and (trunc x), cst)) -> (and x, cst).
1996 if (N0.getOpcode() == ISD::AND &&
1997 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
1998 N0.getOperand(1).getOpcode() == ISD::Constant) {
1999 SDOperand X = N0.getOperand(0).getOperand(0);
2000 if (X.getValueType() < VT) {
2001 X = DAG.getNode(ISD::ANY_EXTEND, VT, X);
2002 } else if (X.getValueType() > VT) {
2003 X = DAG.getNode(ISD::TRUNCATE, VT, X);
2005 uint64_t Mask = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
2006 return DAG.getNode(ISD::AND, VT, X, DAG.getConstant(Mask, VT));
2009 // fold (zext (load x)) -> (zext (truncate (zextload x)))
2010 if (ISD::isNON_EXTLoad(N0.Val) && N0.hasOneUse() &&
2011 (!AfterLegalize||TLI.isLoadXLegal(ISD::ZEXTLOAD, N0.getValueType()))) {
2012 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2013 SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, LN0->getChain(),
2014 LN0->getBasePtr(), LN0->getSrcValue(),
2015 LN0->getSrcValueOffset(),
2017 CombineTo(N, ExtLoad);
2018 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
2019 ExtLoad.getValue(1));
2020 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
2023 // fold (zext (zextload x)) -> (zext (truncate (zextload x)))
2024 // fold (zext ( extload x)) -> (zext (truncate (zextload x)))
2025 if ((ISD::isZEXTLoad(N0.Val) || ISD::isEXTLoad(N0.Val)) && N0.hasOneUse()) {
2026 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2027 MVT::ValueType EVT = LN0->getLoadedVT();
2028 SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, LN0->getChain(),
2029 LN0->getBasePtr(), LN0->getSrcValue(),
2030 LN0->getSrcValueOffset(), EVT);
2031 CombineTo(N, ExtLoad);
2032 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
2033 ExtLoad.getValue(1));
2034 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
2039 SDOperand DAGCombiner::visitANY_EXTEND(SDNode *N) {
2040 SDOperand N0 = N->getOperand(0);
2041 MVT::ValueType VT = N->getValueType(0);
2043 // fold (aext c1) -> c1
2044 if (isa<ConstantSDNode>(N0))
2045 return DAG.getNode(ISD::ANY_EXTEND, VT, N0);
2046 // fold (aext (aext x)) -> (aext x)
2047 // fold (aext (zext x)) -> (zext x)
2048 // fold (aext (sext x)) -> (sext x)
2049 if (N0.getOpcode() == ISD::ANY_EXTEND ||
2050 N0.getOpcode() == ISD::ZERO_EXTEND ||
2051 N0.getOpcode() == ISD::SIGN_EXTEND)
2052 return DAG.getNode(N0.getOpcode(), VT, N0.getOperand(0));
2054 // fold (aext (truncate x))
2055 if (N0.getOpcode() == ISD::TRUNCATE) {
2056 SDOperand TruncOp = N0.getOperand(0);
2057 if (TruncOp.getValueType() == VT)
2058 return TruncOp; // x iff x size == zext size.
2059 if (TruncOp.getValueType() > VT)
2060 return DAG.getNode(ISD::TRUNCATE, VT, TruncOp);
2061 return DAG.getNode(ISD::ANY_EXTEND, VT, TruncOp);
2064 // fold (aext (and (trunc x), cst)) -> (and x, cst).
2065 if (N0.getOpcode() == ISD::AND &&
2066 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
2067 N0.getOperand(1).getOpcode() == ISD::Constant) {
2068 SDOperand X = N0.getOperand(0).getOperand(0);
2069 if (X.getValueType() < VT) {
2070 X = DAG.getNode(ISD::ANY_EXTEND, VT, X);
2071 } else if (X.getValueType() > VT) {
2072 X = DAG.getNode(ISD::TRUNCATE, VT, X);
2074 uint64_t Mask = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
2075 return DAG.getNode(ISD::AND, VT, X, DAG.getConstant(Mask, VT));
2078 // fold (aext (load x)) -> (aext (truncate (extload x)))
2079 if (ISD::isNON_EXTLoad(N0.Val) && N0.hasOneUse() &&
2080 (!AfterLegalize||TLI.isLoadXLegal(ISD::EXTLOAD, N0.getValueType()))) {
2081 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2082 SDOperand ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, VT, LN0->getChain(),
2083 LN0->getBasePtr(), LN0->getSrcValue(),
2084 LN0->getSrcValueOffset(),
2086 CombineTo(N, ExtLoad);
2087 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
2088 ExtLoad.getValue(1));
2089 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
2092 // fold (aext (zextload x)) -> (aext (truncate (zextload x)))
2093 // fold (aext (sextload x)) -> (aext (truncate (sextload x)))
2094 // fold (aext ( extload x)) -> (aext (truncate (extload x)))
2095 if (N0.getOpcode() == ISD::LOAD && !ISD::isNON_EXTLoad(N0.Val) &&
2097 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2098 MVT::ValueType EVT = LN0->getLoadedVT();
2099 SDOperand ExtLoad = DAG.getExtLoad(LN0->getExtensionType(), VT,
2100 LN0->getChain(), LN0->getBasePtr(),
2102 LN0->getSrcValueOffset(), EVT);
2103 CombineTo(N, ExtLoad);
2104 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
2105 ExtLoad.getValue(1));
2106 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
2112 SDOperand DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
2113 SDOperand N0 = N->getOperand(0);
2114 SDOperand N1 = N->getOperand(1);
2115 MVT::ValueType VT = N->getValueType(0);
2116 MVT::ValueType EVT = cast<VTSDNode>(N1)->getVT();
2117 unsigned EVTBits = MVT::getSizeInBits(EVT);
2119 // fold (sext_in_reg c1) -> c1
2120 if (isa<ConstantSDNode>(N0) || N0.getOpcode() == ISD::UNDEF)
2121 return DAG.getNode(ISD::SIGN_EXTEND_INREG, VT, N0, N1);
2123 // If the input is already sign extended, just drop the extension.
2124 if (TLI.ComputeNumSignBits(N0) >= MVT::getSizeInBits(VT)-EVTBits+1)
2127 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2
2128 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
2129 EVT < cast<VTSDNode>(N0.getOperand(1))->getVT()) {
2130 return DAG.getNode(ISD::SIGN_EXTEND_INREG, VT, N0.getOperand(0), N1);
2133 // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is zero
2134 if (TLI.MaskedValueIsZero(N0, 1ULL << (EVTBits-1)))
2135 return DAG.getZeroExtendInReg(N0, EVT);
2137 // fold (sext_in_reg (srl X, 24), i8) -> sra X, 24
2138 // fold (sext_in_reg (srl X, 23), i8) -> sra X, 23 iff possible.
2139 // We already fold "(sext_in_reg (srl X, 25), i8) -> srl X, 25" above.
2140 if (N0.getOpcode() == ISD::SRL) {
2141 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
2142 if (ShAmt->getValue()+EVTBits <= MVT::getSizeInBits(VT)) {
2143 // We can turn this into an SRA iff the input to the SRL is already sign
2145 unsigned InSignBits = TLI.ComputeNumSignBits(N0.getOperand(0));
2146 if (MVT::getSizeInBits(VT)-(ShAmt->getValue()+EVTBits) < InSignBits)
2147 return DAG.getNode(ISD::SRA, VT, N0.getOperand(0), N0.getOperand(1));
2151 // fold (sext_inreg (extload x)) -> (sextload x)
2152 if (ISD::isEXTLoad(N0.Val) &&
2153 EVT == cast<LoadSDNode>(N0)->getLoadedVT() &&
2154 (!AfterLegalize || TLI.isLoadXLegal(ISD::SEXTLOAD, EVT))) {
2155 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2156 SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, LN0->getChain(),
2157 LN0->getBasePtr(), LN0->getSrcValue(),
2158 LN0->getSrcValueOffset(), EVT);
2159 CombineTo(N, ExtLoad);
2160 CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1));
2161 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
2163 // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use
2164 if (ISD::isZEXTLoad(N0.Val) && N0.hasOneUse() &&
2165 EVT == cast<LoadSDNode>(N0)->getLoadedVT() &&
2166 (!AfterLegalize || TLI.isLoadXLegal(ISD::SEXTLOAD, EVT))) {
2167 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2168 SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, LN0->getChain(),
2169 LN0->getBasePtr(), LN0->getSrcValue(),
2170 LN0->getSrcValueOffset(), EVT);
2171 CombineTo(N, ExtLoad);
2172 CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1));
2173 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
2178 SDOperand DAGCombiner::visitTRUNCATE(SDNode *N) {
2179 SDOperand N0 = N->getOperand(0);
2180 MVT::ValueType VT = N->getValueType(0);
2183 if (N0.getValueType() == N->getValueType(0))
2185 // fold (truncate c1) -> c1
2186 if (isa<ConstantSDNode>(N0))
2187 return DAG.getNode(ISD::TRUNCATE, VT, N0);
2188 // fold (truncate (truncate x)) -> (truncate x)
2189 if (N0.getOpcode() == ISD::TRUNCATE)
2190 return DAG.getNode(ISD::TRUNCATE, VT, N0.getOperand(0));
2191 // fold (truncate (ext x)) -> (ext x) or (truncate x) or x
2192 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::SIGN_EXTEND||
2193 N0.getOpcode() == ISD::ANY_EXTEND) {
2194 if (N0.getOperand(0).getValueType() < VT)
2195 // if the source is smaller than the dest, we still need an extend
2196 return DAG.getNode(N0.getOpcode(), VT, N0.getOperand(0));
2197 else if (N0.getOperand(0).getValueType() > VT)
2198 // if the source is larger than the dest, than we just need the truncate
2199 return DAG.getNode(ISD::TRUNCATE, VT, N0.getOperand(0));
2201 // if the source and dest are the same type, we can drop both the extend
2203 return N0.getOperand(0);
2205 // fold (truncate (load x)) -> (smaller load x)
2206 if (ISD::isNON_EXTLoad(N0.Val) && N0.hasOneUse() &&
2207 // Do not allow folding to i1 here. i1 is implicitly stored in memory in
2208 // zero extended form: by shrinking the load, we lose track of the fact
2209 // that it is already zero extended.
2210 // FIXME: This should be reevaluated.
2212 assert(MVT::getSizeInBits(N0.getValueType()) > MVT::getSizeInBits(VT) &&
2213 "Cannot truncate to larger type!");
2214 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2215 MVT::ValueType PtrType = N0.getOperand(1).getValueType();
2216 // For big endian targets, we need to add an offset to the pointer to load
2217 // the correct bytes. For little endian systems, we merely need to read
2218 // fewer bytes from the same pointer.
2220 (MVT::getSizeInBits(N0.getValueType()) - MVT::getSizeInBits(VT)) / 8;
2221 SDOperand NewPtr = TLI.isLittleEndian() ? LN0->getBasePtr() :
2222 DAG.getNode(ISD::ADD, PtrType, LN0->getBasePtr(),
2223 DAG.getConstant(PtrOff, PtrType));
2224 AddToWorkList(NewPtr.Val);
2225 SDOperand Load = DAG.getLoad(VT, LN0->getChain(), NewPtr,
2226 LN0->getSrcValue(), LN0->getSrcValueOffset());
2228 CombineTo(N0.Val, Load, Load.getValue(1));
2229 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
2234 SDOperand DAGCombiner::visitBIT_CONVERT(SDNode *N) {
2235 SDOperand N0 = N->getOperand(0);
2236 MVT::ValueType VT = N->getValueType(0);
2238 // If the input is a constant, let getNode() fold it.
2239 if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) {
2240 SDOperand Res = DAG.getNode(ISD::BIT_CONVERT, VT, N0);
2241 if (Res.Val != N) return Res;
2244 if (N0.getOpcode() == ISD::BIT_CONVERT) // conv(conv(x,t1),t2) -> conv(x,t2)
2245 return DAG.getNode(ISD::BIT_CONVERT, VT, N0.getOperand(0));
2247 // fold (conv (load x)) -> (load (conv*)x)
2248 // FIXME: These xforms need to know that the resultant load doesn't need a
2249 // higher alignment than the original!
2250 if (0 && ISD::isNON_EXTLoad(N0.Val) && N0.hasOneUse()) {
2251 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2252 SDOperand Load = DAG.getLoad(VT, LN0->getChain(), LN0->getBasePtr(),
2253 LN0->getSrcValue(), LN0->getSrcValueOffset());
2255 CombineTo(N0.Val, DAG.getNode(ISD::BIT_CONVERT, N0.getValueType(), Load),
2263 SDOperand DAGCombiner::visitVBIT_CONVERT(SDNode *N) {
2264 SDOperand N0 = N->getOperand(0);
2265 MVT::ValueType VT = N->getValueType(0);
2267 // If the input is a VBUILD_VECTOR with all constant elements, fold this now.
2268 // First check to see if this is all constant.
2269 if (N0.getOpcode() == ISD::VBUILD_VECTOR && N0.Val->hasOneUse() &&
2270 VT == MVT::Vector) {
2271 bool isSimple = true;
2272 for (unsigned i = 0, e = N0.getNumOperands()-2; i != e; ++i)
2273 if (N0.getOperand(i).getOpcode() != ISD::UNDEF &&
2274 N0.getOperand(i).getOpcode() != ISD::Constant &&
2275 N0.getOperand(i).getOpcode() != ISD::ConstantFP) {
2280 MVT::ValueType DestEltVT = cast<VTSDNode>(N->getOperand(2))->getVT();
2281 if (isSimple && !MVT::isVector(DestEltVT)) {
2282 return ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(N0.Val, DestEltVT);
2289 /// ConstantFoldVBIT_CONVERTofVBUILD_VECTOR - We know that BV is a vbuild_vector
2290 /// node with Constant, ConstantFP or Undef operands. DstEltVT indicates the
2291 /// destination element value type.
2292 SDOperand DAGCombiner::
2293 ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(SDNode *BV, MVT::ValueType DstEltVT) {
2294 MVT::ValueType SrcEltVT = BV->getOperand(0).getValueType();
2296 // If this is already the right type, we're done.
2297 if (SrcEltVT == DstEltVT) return SDOperand(BV, 0);
2299 unsigned SrcBitSize = MVT::getSizeInBits(SrcEltVT);
2300 unsigned DstBitSize = MVT::getSizeInBits(DstEltVT);
2302 // If this is a conversion of N elements of one type to N elements of another
2303 // type, convert each element. This handles FP<->INT cases.
2304 if (SrcBitSize == DstBitSize) {
2305 SmallVector<SDOperand, 8> Ops;
2306 for (unsigned i = 0, e = BV->getNumOperands()-2; i != e; ++i) {
2307 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, DstEltVT, BV->getOperand(i)));
2308 AddToWorkList(Ops.back().Val);
2310 Ops.push_back(*(BV->op_end()-2)); // Add num elements.
2311 Ops.push_back(DAG.getValueType(DstEltVT));
2312 return DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, &Ops[0], Ops.size());
2315 // Otherwise, we're growing or shrinking the elements. To avoid having to
2316 // handle annoying details of growing/shrinking FP values, we convert them to
2318 if (MVT::isFloatingPoint(SrcEltVT)) {
2319 // Convert the input float vector to a int vector where the elements are the
2321 assert((SrcEltVT == MVT::f32 || SrcEltVT == MVT::f64) && "Unknown FP VT!");
2322 MVT::ValueType IntVT = SrcEltVT == MVT::f32 ? MVT::i32 : MVT::i64;
2323 BV = ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(BV, IntVT).Val;
2327 // Now we know the input is an integer vector. If the output is a FP type,
2328 // convert to integer first, then to FP of the right size.
2329 if (MVT::isFloatingPoint(DstEltVT)) {
2330 assert((DstEltVT == MVT::f32 || DstEltVT == MVT::f64) && "Unknown FP VT!");
2331 MVT::ValueType TmpVT = DstEltVT == MVT::f32 ? MVT::i32 : MVT::i64;
2332 SDNode *Tmp = ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(BV, TmpVT).Val;
2334 // Next, convert to FP elements of the same size.
2335 return ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(Tmp, DstEltVT);
2338 // Okay, we know the src/dst types are both integers of differing types.
2339 // Handling growing first.
2340 assert(MVT::isInteger(SrcEltVT) && MVT::isInteger(DstEltVT));
2341 if (SrcBitSize < DstBitSize) {
2342 unsigned NumInputsPerOutput = DstBitSize/SrcBitSize;
2344 SmallVector<SDOperand, 8> Ops;
2345 for (unsigned i = 0, e = BV->getNumOperands()-2; i != e;
2346 i += NumInputsPerOutput) {
2347 bool isLE = TLI.isLittleEndian();
2348 uint64_t NewBits = 0;
2349 bool EltIsUndef = true;
2350 for (unsigned j = 0; j != NumInputsPerOutput; ++j) {
2351 // Shift the previously computed bits over.
2352 NewBits <<= SrcBitSize;
2353 SDOperand Op = BV->getOperand(i+ (isLE ? (NumInputsPerOutput-j-1) : j));
2354 if (Op.getOpcode() == ISD::UNDEF) continue;
2357 NewBits |= cast<ConstantSDNode>(Op)->getValue();
2361 Ops.push_back(DAG.getNode(ISD::UNDEF, DstEltVT));
2363 Ops.push_back(DAG.getConstant(NewBits, DstEltVT));
2366 Ops.push_back(DAG.getConstant(Ops.size(), MVT::i32)); // Add num elements.
2367 Ops.push_back(DAG.getValueType(DstEltVT)); // Add element size.
2368 return DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, &Ops[0], Ops.size());
2371 // Finally, this must be the case where we are shrinking elements: each input
2372 // turns into multiple outputs.
2373 unsigned NumOutputsPerInput = SrcBitSize/DstBitSize;
2374 SmallVector<SDOperand, 8> Ops;
2375 for (unsigned i = 0, e = BV->getNumOperands()-2; i != e; ++i) {
2376 if (BV->getOperand(i).getOpcode() == ISD::UNDEF) {
2377 for (unsigned j = 0; j != NumOutputsPerInput; ++j)
2378 Ops.push_back(DAG.getNode(ISD::UNDEF, DstEltVT));
2381 uint64_t OpVal = cast<ConstantSDNode>(BV->getOperand(i))->getValue();
2383 for (unsigned j = 0; j != NumOutputsPerInput; ++j) {
2384 unsigned ThisVal = OpVal & ((1ULL << DstBitSize)-1);
2385 OpVal >>= DstBitSize;
2386 Ops.push_back(DAG.getConstant(ThisVal, DstEltVT));
2389 // For big endian targets, swap the order of the pieces of each element.
2390 if (!TLI.isLittleEndian())
2391 std::reverse(Ops.end()-NumOutputsPerInput, Ops.end());
2393 Ops.push_back(DAG.getConstant(Ops.size(), MVT::i32)); // Add num elements.
2394 Ops.push_back(DAG.getValueType(DstEltVT)); // Add element size.
2395 return DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, &Ops[0], Ops.size());
2400 SDOperand DAGCombiner::visitFADD(SDNode *N) {
2401 SDOperand N0 = N->getOperand(0);
2402 SDOperand N1 = N->getOperand(1);
2403 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2404 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
2405 MVT::ValueType VT = N->getValueType(0);
2407 // fold (fadd c1, c2) -> c1+c2
2409 return DAG.getNode(ISD::FADD, VT, N0, N1);
2410 // canonicalize constant to RHS
2411 if (N0CFP && !N1CFP)
2412 return DAG.getNode(ISD::FADD, VT, N1, N0);
2413 // fold (A + (-B)) -> A-B
2414 if (N1.getOpcode() == ISD::FNEG)
2415 return DAG.getNode(ISD::FSUB, VT, N0, N1.getOperand(0));
2416 // fold ((-A) + B) -> B-A
2417 if (N0.getOpcode() == ISD::FNEG)
2418 return DAG.getNode(ISD::FSUB, VT, N1, N0.getOperand(0));
2420 // If allowed, fold (fadd (fadd x, c1), c2) -> (fadd x, (fadd c1, c2))
2421 if (UnsafeFPMath && N1CFP && N0.getOpcode() == ISD::FADD &&
2422 N0.Val->hasOneUse() && isa<ConstantFPSDNode>(N0.getOperand(1)))
2423 return DAG.getNode(ISD::FADD, VT, N0.getOperand(0),
2424 DAG.getNode(ISD::FADD, VT, N0.getOperand(1), N1));
2429 SDOperand DAGCombiner::visitFSUB(SDNode *N) {
2430 SDOperand N0 = N->getOperand(0);
2431 SDOperand N1 = N->getOperand(1);
2432 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2433 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
2434 MVT::ValueType VT = N->getValueType(0);
2436 // fold (fsub c1, c2) -> c1-c2
2438 return DAG.getNode(ISD::FSUB, VT, N0, N1);
2439 // fold (A-(-B)) -> A+B
2440 if (N1.getOpcode() == ISD::FNEG)
2441 return DAG.getNode(ISD::FADD, VT, N0, N1.getOperand(0));
2445 SDOperand DAGCombiner::visitFMUL(SDNode *N) {
2446 SDOperand N0 = N->getOperand(0);
2447 SDOperand N1 = N->getOperand(1);
2448 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2449 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
2450 MVT::ValueType VT = N->getValueType(0);
2452 // fold (fmul c1, c2) -> c1*c2
2454 return DAG.getNode(ISD::FMUL, VT, N0, N1);
2455 // canonicalize constant to RHS
2456 if (N0CFP && !N1CFP)
2457 return DAG.getNode(ISD::FMUL, VT, N1, N0);
2458 // fold (fmul X, 2.0) -> (fadd X, X)
2459 if (N1CFP && N1CFP->isExactlyValue(+2.0))
2460 return DAG.getNode(ISD::FADD, VT, N0, N0);
2462 // If allowed, fold (fmul (fmul x, c1), c2) -> (fmul x, (fmul c1, c2))
2463 if (UnsafeFPMath && N1CFP && N0.getOpcode() == ISD::FMUL &&
2464 N0.Val->hasOneUse() && isa<ConstantFPSDNode>(N0.getOperand(1)))
2465 return DAG.getNode(ISD::FMUL, VT, N0.getOperand(0),
2466 DAG.getNode(ISD::FMUL, VT, N0.getOperand(1), N1));
2471 SDOperand DAGCombiner::visitFDIV(SDNode *N) {
2472 SDOperand N0 = N->getOperand(0);
2473 SDOperand N1 = N->getOperand(1);
2474 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2475 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
2476 MVT::ValueType VT = N->getValueType(0);
2478 // fold (fdiv c1, c2) -> c1/c2
2480 return DAG.getNode(ISD::FDIV, VT, N0, N1);
2484 SDOperand DAGCombiner::visitFREM(SDNode *N) {
2485 SDOperand N0 = N->getOperand(0);
2486 SDOperand N1 = N->getOperand(1);
2487 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2488 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
2489 MVT::ValueType VT = N->getValueType(0);
2491 // fold (frem c1, c2) -> fmod(c1,c2)
2493 return DAG.getNode(ISD::FREM, VT, N0, N1);
2497 SDOperand DAGCombiner::visitFCOPYSIGN(SDNode *N) {
2498 SDOperand N0 = N->getOperand(0);
2499 SDOperand N1 = N->getOperand(1);
2500 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2501 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
2502 MVT::ValueType VT = N->getValueType(0);
2504 if (N0CFP && N1CFP) // Constant fold
2505 return DAG.getNode(ISD::FCOPYSIGN, VT, N0, N1);
2508 // copysign(x, c1) -> fabs(x) iff ispos(c1)
2509 // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1)
2514 u.d = N1CFP->getValue();
2516 return DAG.getNode(ISD::FABS, VT, N0);
2518 return DAG.getNode(ISD::FNEG, VT, DAG.getNode(ISD::FABS, VT, N0));
2521 // copysign(fabs(x), y) -> copysign(x, y)
2522 // copysign(fneg(x), y) -> copysign(x, y)
2523 // copysign(copysign(x,z), y) -> copysign(x, y)
2524 if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG ||
2525 N0.getOpcode() == ISD::FCOPYSIGN)
2526 return DAG.getNode(ISD::FCOPYSIGN, VT, N0.getOperand(0), N1);
2528 // copysign(x, abs(y)) -> abs(x)
2529 if (N1.getOpcode() == ISD::FABS)
2530 return DAG.getNode(ISD::FABS, VT, N0);
2532 // copysign(x, copysign(y,z)) -> copysign(x, z)
2533 if (N1.getOpcode() == ISD::FCOPYSIGN)
2534 return DAG.getNode(ISD::FCOPYSIGN, VT, N0, N1.getOperand(1));
2536 // copysign(x, fp_extend(y)) -> copysign(x, y)
2537 // copysign(x, fp_round(y)) -> copysign(x, y)
2538 if (N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND)
2539 return DAG.getNode(ISD::FCOPYSIGN, VT, N0, N1.getOperand(0));
2546 SDOperand DAGCombiner::visitSINT_TO_FP(SDNode *N) {
2547 SDOperand N0 = N->getOperand(0);
2548 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2549 MVT::ValueType VT = N->getValueType(0);
2551 // fold (sint_to_fp c1) -> c1fp
2553 return DAG.getNode(ISD::SINT_TO_FP, VT, N0);
2557 SDOperand DAGCombiner::visitUINT_TO_FP(SDNode *N) {
2558 SDOperand N0 = N->getOperand(0);
2559 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2560 MVT::ValueType VT = N->getValueType(0);
2562 // fold (uint_to_fp c1) -> c1fp
2564 return DAG.getNode(ISD::UINT_TO_FP, VT, N0);
2568 SDOperand DAGCombiner::visitFP_TO_SINT(SDNode *N) {
2569 SDOperand N0 = N->getOperand(0);
2570 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2571 MVT::ValueType VT = N->getValueType(0);
2573 // fold (fp_to_sint c1fp) -> c1
2575 return DAG.getNode(ISD::FP_TO_SINT, VT, N0);
2579 SDOperand DAGCombiner::visitFP_TO_UINT(SDNode *N) {
2580 SDOperand N0 = N->getOperand(0);
2581 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2582 MVT::ValueType VT = N->getValueType(0);
2584 // fold (fp_to_uint c1fp) -> c1
2586 return DAG.getNode(ISD::FP_TO_UINT, VT, N0);
2590 SDOperand DAGCombiner::visitFP_ROUND(SDNode *N) {
2591 SDOperand N0 = N->getOperand(0);
2592 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2593 MVT::ValueType VT = N->getValueType(0);
2595 // fold (fp_round c1fp) -> c1fp
2597 return DAG.getNode(ISD::FP_ROUND, VT, N0);
2599 // fold (fp_round (fp_extend x)) -> x
2600 if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType())
2601 return N0.getOperand(0);
2603 // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y)
2604 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.Val->hasOneUse()) {
2605 SDOperand Tmp = DAG.getNode(ISD::FP_ROUND, VT, N0.getOperand(0));
2606 AddToWorkList(Tmp.Val);
2607 return DAG.getNode(ISD::FCOPYSIGN, VT, Tmp, N0.getOperand(1));
2613 SDOperand DAGCombiner::visitFP_ROUND_INREG(SDNode *N) {
2614 SDOperand N0 = N->getOperand(0);
2615 MVT::ValueType VT = N->getValueType(0);
2616 MVT::ValueType EVT = cast<VTSDNode>(N->getOperand(1))->getVT();
2617 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2619 // fold (fp_round_inreg c1fp) -> c1fp
2621 SDOperand Round = DAG.getConstantFP(N0CFP->getValue(), EVT);
2622 return DAG.getNode(ISD::FP_EXTEND, VT, Round);
2627 SDOperand DAGCombiner::visitFP_EXTEND(SDNode *N) {
2628 SDOperand N0 = N->getOperand(0);
2629 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2630 MVT::ValueType VT = N->getValueType(0);
2632 // fold (fp_extend c1fp) -> c1fp
2634 return DAG.getNode(ISD::FP_EXTEND, VT, N0);
2636 // fold (fpext (load x)) -> (fpext (fpround (extload x)))
2637 if (ISD::isNON_EXTLoad(N0.Val) && N0.hasOneUse() &&
2638 (!AfterLegalize||TLI.isLoadXLegal(ISD::EXTLOAD, N0.getValueType()))) {
2639 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2640 SDOperand ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, VT, LN0->getChain(),
2641 LN0->getBasePtr(), LN0->getSrcValue(),
2642 LN0->getSrcValueOffset(),
2644 CombineTo(N, ExtLoad);
2645 CombineTo(N0.Val, DAG.getNode(ISD::FP_ROUND, N0.getValueType(), ExtLoad),
2646 ExtLoad.getValue(1));
2647 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
2654 SDOperand DAGCombiner::visitFNEG(SDNode *N) {
2655 SDOperand N0 = N->getOperand(0);
2656 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2657 MVT::ValueType VT = N->getValueType(0);
2659 // fold (fneg c1) -> -c1
2661 return DAG.getNode(ISD::FNEG, VT, N0);
2662 // fold (fneg (sub x, y)) -> (sub y, x)
2663 if (N0.getOpcode() == ISD::SUB)
2664 return DAG.getNode(ISD::SUB, VT, N0.getOperand(1), N0.getOperand(0));
2665 // fold (fneg (fneg x)) -> x
2666 if (N0.getOpcode() == ISD::FNEG)
2667 return N0.getOperand(0);
2671 SDOperand DAGCombiner::visitFABS(SDNode *N) {
2672 SDOperand N0 = N->getOperand(0);
2673 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2674 MVT::ValueType VT = N->getValueType(0);
2676 // fold (fabs c1) -> fabs(c1)
2678 return DAG.getNode(ISD::FABS, VT, N0);
2679 // fold (fabs (fabs x)) -> (fabs x)
2680 if (N0.getOpcode() == ISD::FABS)
2681 return N->getOperand(0);
2682 // fold (fabs (fneg x)) -> (fabs x)
2683 // fold (fabs (fcopysign x, y)) -> (fabs x)
2684 if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN)
2685 return DAG.getNode(ISD::FABS, VT, N0.getOperand(0));
2690 SDOperand DAGCombiner::visitBRCOND(SDNode *N) {
2691 SDOperand Chain = N->getOperand(0);
2692 SDOperand N1 = N->getOperand(1);
2693 SDOperand N2 = N->getOperand(2);
2694 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2696 // never taken branch, fold to chain
2697 if (N1C && N1C->isNullValue())
2699 // unconditional branch
2700 if (N1C && N1C->getValue() == 1)
2701 return DAG.getNode(ISD::BR, MVT::Other, Chain, N2);
2702 // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal
2704 if (N1.getOpcode() == ISD::SETCC &&
2705 TLI.isOperationLegal(ISD::BR_CC, MVT::Other)) {
2706 return DAG.getNode(ISD::BR_CC, MVT::Other, Chain, N1.getOperand(2),
2707 N1.getOperand(0), N1.getOperand(1), N2);
2712 // Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB.
2714 SDOperand DAGCombiner::visitBR_CC(SDNode *N) {
2715 CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1));
2716 SDOperand CondLHS = N->getOperand(2), CondRHS = N->getOperand(3);
2718 // Use SimplifySetCC to simplify SETCC's.
2719 SDOperand Simp = SimplifySetCC(MVT::i1, CondLHS, CondRHS, CC->get(), false);
2720 if (Simp.Val) AddToWorkList(Simp.Val);
2722 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(Simp.Val);
2724 // fold br_cc true, dest -> br dest (unconditional branch)
2725 if (SCCC && SCCC->getValue())
2726 return DAG.getNode(ISD::BR, MVT::Other, N->getOperand(0),
2728 // fold br_cc false, dest -> unconditional fall through
2729 if (SCCC && SCCC->isNullValue())
2730 return N->getOperand(0);
2732 // fold to a simpler setcc
2733 if (Simp.Val && Simp.getOpcode() == ISD::SETCC)
2734 return DAG.getNode(ISD::BR_CC, MVT::Other, N->getOperand(0),
2735 Simp.getOperand(2), Simp.getOperand(0),
2736 Simp.getOperand(1), N->getOperand(4));
2741 /// CombineToPreIndexedLoadStore - Try turning a load / store and a
2742 /// pre-indexed load / store when the base pointer is a add or subtract
2743 /// and it has other uses besides the load / store. After the
2744 /// transformation, the new indexed load / store has effectively folded
2745 /// the add / subtract in and all of its other uses are redirected to the
2746 /// new load / store.
2747 bool DAGCombiner::CombineToPreIndexedLoadStore(SDNode *N) {
2754 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
2755 if (LD->getAddressingMode() != ISD::UNINDEXED)
2757 VT = LD->getLoadedVT();
2758 if (LD->getAddressingMode() != ISD::UNINDEXED &&
2759 !TLI.isIndexedLoadLegal(ISD::PRE_INC, VT) &&
2760 !TLI.isIndexedLoadLegal(ISD::PRE_DEC, VT))
2762 Ptr = LD->getBasePtr();
2763 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
2764 if (ST->getAddressingMode() != ISD::UNINDEXED)
2766 VT = ST->getStoredVT();
2767 if (!TLI.isIndexedStoreLegal(ISD::PRE_INC, VT) &&
2768 !TLI.isIndexedStoreLegal(ISD::PRE_DEC, VT))
2770 Ptr = ST->getBasePtr();
2775 // If the pointer is not an add/sub, or if it doesn't have multiple uses, bail
2776 // out. There is no reason to make this a preinc/predec.
2777 if ((Ptr.getOpcode() != ISD::ADD && Ptr.getOpcode() != ISD::SUB) ||
2778 Ptr.Val->hasOneUse())
2781 // Ask the target to do addressing mode selection.
2784 ISD::MemIndexedMode AM = ISD::UNINDEXED;
2785 if (!TLI.getPreIndexedAddressParts(N, BasePtr, Offset, AM, DAG))
2788 // Try turning it into a pre-indexed load / store except when:
2789 // 1) The base is a frame index.
2790 // 2) If N is a store and the ptr is either the same as or is a
2791 // predecessor of the value being stored.
2792 // 3) Another use of base ptr is a predecessor of N. If ptr is folded
2793 // that would create a cycle.
2794 // 4) All uses are load / store ops that use it as base ptr.
2796 // Check #1. Preinc'ing a frame index would require copying the stack pointer
2797 // (plus the implicit offset) to a register to preinc anyway.
2798 if (isa<FrameIndexSDNode>(BasePtr))
2803 SDOperand Val = cast<StoreSDNode>(N)->getValue();
2804 if (Val == Ptr || Ptr.Val->isPredecessor(Val.Val))
2808 // Now check for #2 and #3.
2809 bool RealUse = false;
2810 for (SDNode::use_iterator I = Ptr.Val->use_begin(),
2811 E = Ptr.Val->use_end(); I != E; ++I) {
2815 if (Use->isPredecessor(N))
2818 if (!((Use->getOpcode() == ISD::LOAD &&
2819 cast<LoadSDNode>(Use)->getBasePtr() == Ptr) ||
2820 (Use->getOpcode() == ISD::STORE) &&
2821 cast<StoreSDNode>(Use)->getBasePtr() == Ptr))
2829 Result = DAG.getIndexedLoad(SDOperand(N,0), BasePtr, Offset, AM);
2831 Result = DAG.getIndexedStore(SDOperand(N,0), BasePtr, Offset, AM);
2834 DOUT << "\nReplacing.4 "; DEBUG(N->dump());
2835 DOUT << "\nWith: "; DEBUG(Result.Val->dump(&DAG));
2837 std::vector<SDNode*> NowDead;
2839 DAG.ReplaceAllUsesOfValueWith(SDOperand(N, 0), Result.getValue(0),
2841 DAG.ReplaceAllUsesOfValueWith(SDOperand(N, 1), Result.getValue(2),
2844 DAG.ReplaceAllUsesOfValueWith(SDOperand(N, 0), Result.getValue(1),
2848 // Nodes can end up on the worklist more than once. Make sure we do
2849 // not process a node that has been replaced.
2850 for (unsigned i = 0, e = NowDead.size(); i != e; ++i)
2851 removeFromWorkList(NowDead[i]);
2852 // Finally, since the node is now dead, remove it from the graph.
2855 // Replace the uses of Ptr with uses of the updated base value.
2856 DAG.ReplaceAllUsesOfValueWith(Ptr, Result.getValue(isLoad ? 1 : 0),
2858 removeFromWorkList(Ptr.Val);
2859 for (unsigned i = 0, e = NowDead.size(); i != e; ++i)
2860 removeFromWorkList(NowDead[i]);
2861 DAG.DeleteNode(Ptr.Val);
2866 /// CombineToPostIndexedLoadStore - Try combine a load / store with a
2867 /// add / sub of the base pointer node into a post-indexed load / store.
2868 /// The transformation folded the add / subtract into the new indexed
2869 /// load / store effectively and all of its uses are redirected to the
2870 /// new load / store.
2871 bool DAGCombiner::CombineToPostIndexedLoadStore(SDNode *N) {
2878 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
2879 if (LD->getAddressingMode() != ISD::UNINDEXED)
2881 VT = LD->getLoadedVT();
2882 if (!TLI.isIndexedLoadLegal(ISD::POST_INC, VT) &&
2883 !TLI.isIndexedLoadLegal(ISD::POST_DEC, VT))
2885 Ptr = LD->getBasePtr();
2886 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
2887 if (ST->getAddressingMode() != ISD::UNINDEXED)
2889 VT = ST->getStoredVT();
2890 if (!TLI.isIndexedStoreLegal(ISD::POST_INC, VT) &&
2891 !TLI.isIndexedStoreLegal(ISD::POST_DEC, VT))
2893 Ptr = ST->getBasePtr();
2898 if (Ptr.Val->hasOneUse())
2901 for (SDNode::use_iterator I = Ptr.Val->use_begin(),
2902 E = Ptr.Val->use_end(); I != E; ++I) {
2905 (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB))
2910 ISD::MemIndexedMode AM = ISD::UNINDEXED;
2911 if (TLI.getPostIndexedAddressParts(N, Op, BasePtr, Offset, AM, DAG)) {
2913 std::swap(BasePtr, Offset);
2917 // Try turning it into a post-indexed load / store except when
2918 // 1) All uses are load / store ops that use it as base ptr.
2919 // 2) Op must be independent of N, i.e. Op is neither a predecessor
2920 // nor a successor of N. Otherwise, if Op is folded that would
2924 bool TryNext = false;
2925 for (SDNode::use_iterator II = BasePtr.Val->use_begin(),
2926 EE = BasePtr.Val->use_end(); II != EE; ++II) {
2931 // If all the uses are load / store addresses, then don't do the
2933 if (Use->getOpcode() == ISD::ADD || Use->getOpcode() == ISD::SUB){
2934 bool RealUse = false;
2935 for (SDNode::use_iterator III = Use->use_begin(),
2936 EEE = Use->use_end(); III != EEE; ++III) {
2937 SDNode *UseUse = *III;
2938 if (!((UseUse->getOpcode() == ISD::LOAD &&
2939 cast<LoadSDNode>(UseUse)->getBasePtr().Val == Use) ||
2940 (UseUse->getOpcode() == ISD::STORE) &&
2941 cast<StoreSDNode>(UseUse)->getBasePtr().Val == Use))
2955 if (!Op->isPredecessor(N) && !N->isPredecessor(Op)) {
2956 SDOperand Result = isLoad
2957 ? DAG.getIndexedLoad(SDOperand(N,0), BasePtr, Offset, AM)
2958 : DAG.getIndexedStore(SDOperand(N,0), BasePtr, Offset, AM);
2961 DOUT << "\nReplacing.5 "; DEBUG(N->dump());
2962 DOUT << "\nWith: "; DEBUG(Result.Val->dump(&DAG));
2964 std::vector<SDNode*> NowDead;
2966 DAG.ReplaceAllUsesOfValueWith(SDOperand(N, 0), Result.getValue(0),
2968 DAG.ReplaceAllUsesOfValueWith(SDOperand(N, 1), Result.getValue(2),
2971 DAG.ReplaceAllUsesOfValueWith(SDOperand(N, 0), Result.getValue(1),
2975 // Nodes can end up on the worklist more than once. Make sure we do
2976 // not process a node that has been replaced.
2977 for (unsigned i = 0, e = NowDead.size(); i != e; ++i)
2978 removeFromWorkList(NowDead[i]);
2979 // Finally, since the node is now dead, remove it from the graph.
2982 // Replace the uses of Use with uses of the updated base value.
2983 DAG.ReplaceAllUsesOfValueWith(SDOperand(Op, 0),
2984 Result.getValue(isLoad ? 1 : 0),
2986 removeFromWorkList(Op);
2987 for (unsigned i = 0, e = NowDead.size(); i != e; ++i)
2988 removeFromWorkList(NowDead[i]);
2999 SDOperand DAGCombiner::visitLOAD(SDNode *N) {
3000 LoadSDNode *LD = cast<LoadSDNode>(N);
3001 SDOperand Chain = LD->getChain();
3002 SDOperand Ptr = LD->getBasePtr();
3004 // If there are no uses of the loaded value, change uses of the chain value
3005 // into uses of the chain input (i.e. delete the dead load).
3006 if (N->hasNUsesOfValue(0, 0))
3007 return CombineTo(N, DAG.getNode(ISD::UNDEF, N->getValueType(0)), Chain);
3009 // If this load is directly stored, replace the load value with the stored
3011 // TODO: Handle store large -> read small portion.
3012 // TODO: Handle TRUNCSTORE/LOADEXT
3013 if (LD->getExtensionType() == ISD::NON_EXTLOAD) {
3014 if (ISD::isNON_TRUNCStore(Chain.Val)) {
3015 StoreSDNode *PrevST = cast<StoreSDNode>(Chain);
3016 if (PrevST->getBasePtr() == Ptr &&
3017 PrevST->getValue().getValueType() == N->getValueType(0))
3018 return CombineTo(N, Chain.getOperand(1), Chain);
3023 // Walk up chain skipping non-aliasing memory nodes.
3024 SDOperand BetterChain = FindBetterChain(N, Chain);
3026 // If there is a better chain.
3027 if (Chain != BetterChain) {
3030 // Replace the chain to void dependency.
3031 if (LD->getExtensionType() == ISD::NON_EXTLOAD) {
3032 ReplLoad = DAG.getLoad(N->getValueType(0), BetterChain, Ptr,
3033 LD->getSrcValue(), LD->getSrcValueOffset());
3035 ReplLoad = DAG.getExtLoad(LD->getExtensionType(),
3036 LD->getValueType(0),
3037 BetterChain, Ptr, LD->getSrcValue(),
3038 LD->getSrcValueOffset(),
3042 // Create token factor to keep old chain connected.
3043 SDOperand Token = DAG.getNode(ISD::TokenFactor, MVT::Other,
3044 Chain, ReplLoad.getValue(1));
3046 // Replace uses with load result and token factor. Don't add users
3048 return CombineTo(N, ReplLoad.getValue(0), Token, false);
3052 // Try transforming N to an indexed load.
3053 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
3054 return SDOperand(N, 0);
3059 SDOperand DAGCombiner::visitSTORE(SDNode *N) {
3060 StoreSDNode *ST = cast<StoreSDNode>(N);
3061 SDOperand Chain = ST->getChain();
3062 SDOperand Value = ST->getValue();
3063 SDOperand Ptr = ST->getBasePtr();
3065 // If this is a store of a bit convert, store the input value.
3066 // FIXME: This needs to know that the resultant store does not need a
3067 // higher alignment than the original.
3068 if (0 && Value.getOpcode() == ISD::BIT_CONVERT) {
3069 return DAG.getStore(Chain, Value.getOperand(0), Ptr, ST->getSrcValue(),
3070 ST->getSrcValueOffset());
3073 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
3074 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Value)) {
3075 if (Value.getOpcode() != ISD::TargetConstantFP) {
3077 switch (CFP->getValueType(0)) {
3078 default: assert(0 && "Unknown FP type");
3080 if (!AfterLegalize || TLI.isTypeLegal(MVT::i32)) {
3081 Tmp = DAG.getConstant(FloatToBits(CFP->getValue()), MVT::i32);
3082 return DAG.getStore(Chain, Tmp, Ptr, ST->getSrcValue(),
3083 ST->getSrcValueOffset());
3087 if (!AfterLegalize || TLI.isTypeLegal(MVT::i64)) {
3088 Tmp = DAG.getConstant(DoubleToBits(CFP->getValue()), MVT::i64);
3089 return DAG.getStore(Chain, Tmp, Ptr, ST->getSrcValue(),
3090 ST->getSrcValueOffset());
3091 } else if (TLI.isTypeLegal(MVT::i32)) {
3092 // Many FP stores are not make apparent until after legalize, e.g. for
3093 // argument passing. Since this is so common, custom legalize the
3094 // 64-bit integer store into two 32-bit stores.
3095 uint64_t Val = DoubleToBits(CFP->getValue());
3096 SDOperand Lo = DAG.getConstant(Val & 0xFFFFFFFF, MVT::i32);
3097 SDOperand Hi = DAG.getConstant(Val >> 32, MVT::i32);
3098 if (!TLI.isLittleEndian()) std::swap(Lo, Hi);
3100 SDOperand St0 = DAG.getStore(Chain, Lo, Ptr, ST->getSrcValue(),
3101 ST->getSrcValueOffset());
3102 Ptr = DAG.getNode(ISD::ADD, Ptr.getValueType(), Ptr,
3103 DAG.getConstant(4, Ptr.getValueType()));
3104 SDOperand St1 = DAG.getStore(Chain, Hi, Ptr, ST->getSrcValue(),
3105 ST->getSrcValueOffset()+4);
3106 return DAG.getNode(ISD::TokenFactor, MVT::Other, St0, St1);
3114 // Walk up chain skipping non-aliasing memory nodes.
3115 SDOperand BetterChain = FindBetterChain(N, Chain);
3117 // If there is a better chain.
3118 if (Chain != BetterChain) {
3119 // Replace the chain to avoid dependency.
3120 SDOperand ReplStore;
3121 if (ST->isTruncatingStore()) {
3122 ReplStore = DAG.getTruncStore(BetterChain, Value, Ptr,
3123 ST->getSrcValue(),ST->getSrcValueOffset(), ST->getStoredVT());
3125 ReplStore = DAG.getStore(BetterChain, Value, Ptr,
3126 ST->getSrcValue(), ST->getSrcValueOffset());
3129 // Create token to keep both nodes around.
3131 DAG.getNode(ISD::TokenFactor, MVT::Other, Chain, ReplStore);
3133 // Don't add users to work list.
3134 return CombineTo(N, Token, false);
3138 // Try transforming N to an indexed store.
3139 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
3140 return SDOperand(N, 0);
3145 SDOperand DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) {
3146 SDOperand InVec = N->getOperand(0);
3147 SDOperand InVal = N->getOperand(1);
3148 SDOperand EltNo = N->getOperand(2);
3150 // If the invec is a BUILD_VECTOR and if EltNo is a constant, build a new
3151 // vector with the inserted element.
3152 if (InVec.getOpcode() == ISD::BUILD_VECTOR && isa<ConstantSDNode>(EltNo)) {
3153 unsigned Elt = cast<ConstantSDNode>(EltNo)->getValue();
3154 SmallVector<SDOperand, 8> Ops(InVec.Val->op_begin(), InVec.Val->op_end());
3155 if (Elt < Ops.size())
3157 return DAG.getNode(ISD::BUILD_VECTOR, InVec.getValueType(),
3158 &Ops[0], Ops.size());
3164 SDOperand DAGCombiner::visitVINSERT_VECTOR_ELT(SDNode *N) {
3165 SDOperand InVec = N->getOperand(0);
3166 SDOperand InVal = N->getOperand(1);
3167 SDOperand EltNo = N->getOperand(2);
3168 SDOperand NumElts = N->getOperand(3);
3169 SDOperand EltType = N->getOperand(4);
3171 // If the invec is a VBUILD_VECTOR and if EltNo is a constant, build a new
3172 // vector with the inserted element.
3173 if (InVec.getOpcode() == ISD::VBUILD_VECTOR && isa<ConstantSDNode>(EltNo)) {
3174 unsigned Elt = cast<ConstantSDNode>(EltNo)->getValue();
3175 SmallVector<SDOperand, 8> Ops(InVec.Val->op_begin(), InVec.Val->op_end());
3176 if (Elt < Ops.size()-2)
3178 return DAG.getNode(ISD::VBUILD_VECTOR, InVec.getValueType(),
3179 &Ops[0], Ops.size());
3185 SDOperand DAGCombiner::visitVBUILD_VECTOR(SDNode *N) {
3186 unsigned NumInScalars = N->getNumOperands()-2;
3187 SDOperand NumElts = N->getOperand(NumInScalars);
3188 SDOperand EltType = N->getOperand(NumInScalars+1);
3190 // Check to see if this is a VBUILD_VECTOR of a bunch of VEXTRACT_VECTOR_ELT
3191 // operations. If so, and if the EXTRACT_ELT vector inputs come from at most
3192 // two distinct vectors, turn this into a shuffle node.
3193 SDOperand VecIn1, VecIn2;
3194 for (unsigned i = 0; i != NumInScalars; ++i) {
3195 // Ignore undef inputs.
3196 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
3198 // If this input is something other than a VEXTRACT_VECTOR_ELT with a
3199 // constant index, bail out.
3200 if (N->getOperand(i).getOpcode() != ISD::VEXTRACT_VECTOR_ELT ||
3201 !isa<ConstantSDNode>(N->getOperand(i).getOperand(1))) {
3202 VecIn1 = VecIn2 = SDOperand(0, 0);
3206 // If the input vector type disagrees with the result of the vbuild_vector,
3207 // we can't make a shuffle.
3208 SDOperand ExtractedFromVec = N->getOperand(i).getOperand(0);
3209 if (*(ExtractedFromVec.Val->op_end()-2) != NumElts ||
3210 *(ExtractedFromVec.Val->op_end()-1) != EltType) {
3211 VecIn1 = VecIn2 = SDOperand(0, 0);
3215 // Otherwise, remember this. We allow up to two distinct input vectors.
3216 if (ExtractedFromVec == VecIn1 || ExtractedFromVec == VecIn2)
3219 if (VecIn1.Val == 0) {
3220 VecIn1 = ExtractedFromVec;
3221 } else if (VecIn2.Val == 0) {
3222 VecIn2 = ExtractedFromVec;
3225 VecIn1 = VecIn2 = SDOperand(0, 0);
3230 // If everything is good, we can make a shuffle operation.
3232 SmallVector<SDOperand, 8> BuildVecIndices;
3233 for (unsigned i = 0; i != NumInScalars; ++i) {
3234 if (N->getOperand(i).getOpcode() == ISD::UNDEF) {
3235 BuildVecIndices.push_back(DAG.getNode(ISD::UNDEF, MVT::i32));
3239 SDOperand Extract = N->getOperand(i);
3241 // If extracting from the first vector, just use the index directly.
3242 if (Extract.getOperand(0) == VecIn1) {
3243 BuildVecIndices.push_back(Extract.getOperand(1));
3247 // Otherwise, use InIdx + VecSize
3248 unsigned Idx = cast<ConstantSDNode>(Extract.getOperand(1))->getValue();
3249 BuildVecIndices.push_back(DAG.getConstant(Idx+NumInScalars, MVT::i32));
3252 // Add count and size info.
3253 BuildVecIndices.push_back(NumElts);
3254 BuildVecIndices.push_back(DAG.getValueType(MVT::i32));
3256 // Return the new VVECTOR_SHUFFLE node.
3262 // Use an undef vbuild_vector as input for the second operand.
3263 std::vector<SDOperand> UnOps(NumInScalars,
3264 DAG.getNode(ISD::UNDEF,
3265 cast<VTSDNode>(EltType)->getVT()));
3266 UnOps.push_back(NumElts);
3267 UnOps.push_back(EltType);
3268 Ops[1] = DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector,
3269 &UnOps[0], UnOps.size());
3270 AddToWorkList(Ops[1].Val);
3272 Ops[2] = DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector,
3273 &BuildVecIndices[0], BuildVecIndices.size());
3276 return DAG.getNode(ISD::VVECTOR_SHUFFLE, MVT::Vector, Ops, 5);
3282 SDOperand DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
3283 SDOperand ShufMask = N->getOperand(2);
3284 unsigned NumElts = ShufMask.getNumOperands();
3286 // If the shuffle mask is an identity operation on the LHS, return the LHS.
3287 bool isIdentity = true;
3288 for (unsigned i = 0; i != NumElts; ++i) {
3289 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF &&
3290 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() != i) {
3295 if (isIdentity) return N->getOperand(0);
3297 // If the shuffle mask is an identity operation on the RHS, return the RHS.
3299 for (unsigned i = 0; i != NumElts; ++i) {
3300 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF &&
3301 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() != i+NumElts) {
3306 if (isIdentity) return N->getOperand(1);
3308 // Check if the shuffle is a unary shuffle, i.e. one of the vectors is not
3310 bool isUnary = true;
3311 bool isSplat = true;
3313 unsigned BaseIdx = 0;
3314 for (unsigned i = 0; i != NumElts; ++i)
3315 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF) {
3316 unsigned Idx = cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue();
3317 int V = (Idx < NumElts) ? 0 : 1;
3331 SDOperand N0 = N->getOperand(0);
3332 SDOperand N1 = N->getOperand(1);
3333 // Normalize unary shuffle so the RHS is undef.
3334 if (isUnary && VecNum == 1)
3337 // If it is a splat, check if the argument vector is a build_vector with
3338 // all scalar elements the same.
3341 if (V->getOpcode() == ISD::BIT_CONVERT)
3342 V = V->getOperand(0).Val;
3343 if (V->getOpcode() == ISD::BUILD_VECTOR) {
3344 unsigned NumElems = V->getNumOperands()-2;
3345 if (NumElems > BaseIdx) {
3347 bool AllSame = true;
3348 for (unsigned i = 0; i != NumElems; ++i) {
3349 if (V->getOperand(i).getOpcode() != ISD::UNDEF) {
3350 Base = V->getOperand(i);
3354 // Splat of <u, u, u, u>, return <u, u, u, u>
3357 for (unsigned i = 0; i != NumElems; ++i) {
3358 if (V->getOperand(i).getOpcode() != ISD::UNDEF &&
3359 V->getOperand(i) != Base) {
3364 // Splat of <x, x, x, x>, return <x, x, x, x>
3371 // If it is a unary or the LHS and the RHS are the same node, turn the RHS
3373 if (isUnary || N0 == N1) {
3374 if (N0.getOpcode() == ISD::UNDEF)
3375 return DAG.getNode(ISD::UNDEF, N->getValueType(0));
3376 // Check the SHUFFLE mask, mapping any inputs from the 2nd operand into the
3378 SmallVector<SDOperand, 8> MappedOps;
3379 for (unsigned i = 0, e = ShufMask.getNumOperands(); i != e; ++i) {
3380 if (ShufMask.getOperand(i).getOpcode() == ISD::UNDEF ||
3381 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() < NumElts) {
3382 MappedOps.push_back(ShufMask.getOperand(i));
3385 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() - NumElts;
3386 MappedOps.push_back(DAG.getConstant(NewIdx, MVT::i32));
3389 ShufMask = DAG.getNode(ISD::BUILD_VECTOR, ShufMask.getValueType(),
3390 &MappedOps[0], MappedOps.size());
3391 AddToWorkList(ShufMask.Val);
3392 return DAG.getNode(ISD::VECTOR_SHUFFLE, N->getValueType(0),
3394 DAG.getNode(ISD::UNDEF, N->getValueType(0)),
3401 SDOperand DAGCombiner::visitVVECTOR_SHUFFLE(SDNode *N) {
3402 SDOperand ShufMask = N->getOperand(2);
3403 unsigned NumElts = ShufMask.getNumOperands()-2;
3405 // If the shuffle mask is an identity operation on the LHS, return the LHS.
3406 bool isIdentity = true;
3407 for (unsigned i = 0; i != NumElts; ++i) {
3408 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF &&
3409 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() != i) {
3414 if (isIdentity) return N->getOperand(0);
3416 // If the shuffle mask is an identity operation on the RHS, return the RHS.
3418 for (unsigned i = 0; i != NumElts; ++i) {
3419 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF &&
3420 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() != i+NumElts) {
3425 if (isIdentity) return N->getOperand(1);
3427 // Check if the shuffle is a unary shuffle, i.e. one of the vectors is not
3429 bool isUnary = true;
3430 bool isSplat = true;
3432 unsigned BaseIdx = 0;
3433 for (unsigned i = 0; i != NumElts; ++i)
3434 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF) {
3435 unsigned Idx = cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue();
3436 int V = (Idx < NumElts) ? 0 : 1;
3450 SDOperand N0 = N->getOperand(0);
3451 SDOperand N1 = N->getOperand(1);
3452 // Normalize unary shuffle so the RHS is undef.
3453 if (isUnary && VecNum == 1)
3456 // If it is a splat, check if the argument vector is a build_vector with
3457 // all scalar elements the same.
3461 // If this is a vbit convert that changes the element type of the vector but
3462 // not the number of vector elements, look through it. Be careful not to
3463 // look though conversions that change things like v4f32 to v2f64.
3464 if (V->getOpcode() == ISD::VBIT_CONVERT) {
3465 SDOperand ConvInput = V->getOperand(0);
3466 if (ConvInput.getValueType() == MVT::Vector &&
3468 ConvInput.getConstantOperandVal(ConvInput.getNumOperands()-2))
3472 if (V->getOpcode() == ISD::VBUILD_VECTOR) {
3473 unsigned NumElems = V->getNumOperands()-2;
3474 if (NumElems > BaseIdx) {
3476 bool AllSame = true;
3477 for (unsigned i = 0; i != NumElems; ++i) {
3478 if (V->getOperand(i).getOpcode() != ISD::UNDEF) {
3479 Base = V->getOperand(i);
3483 // Splat of <u, u, u, u>, return <u, u, u, u>
3486 for (unsigned i = 0; i != NumElems; ++i) {
3487 if (V->getOperand(i).getOpcode() != ISD::UNDEF &&
3488 V->getOperand(i) != Base) {
3493 // Splat of <x, x, x, x>, return <x, x, x, x>
3500 // If it is a unary or the LHS and the RHS are the same node, turn the RHS
3502 if (isUnary || N0 == N1) {
3503 // Check the SHUFFLE mask, mapping any inputs from the 2nd operand into the
3505 SmallVector<SDOperand, 8> MappedOps;
3506 for (unsigned i = 0; i != NumElts; ++i) {
3507 if (ShufMask.getOperand(i).getOpcode() == ISD::UNDEF ||
3508 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() < NumElts) {
3509 MappedOps.push_back(ShufMask.getOperand(i));
3512 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() - NumElts;
3513 MappedOps.push_back(DAG.getConstant(NewIdx, MVT::i32));
3516 // Add the type/#elts values.
3517 MappedOps.push_back(ShufMask.getOperand(NumElts));
3518 MappedOps.push_back(ShufMask.getOperand(NumElts+1));
3520 ShufMask = DAG.getNode(ISD::VBUILD_VECTOR, ShufMask.getValueType(),
3521 &MappedOps[0], MappedOps.size());
3522 AddToWorkList(ShufMask.Val);
3524 // Build the undef vector.
3525 SDOperand UDVal = DAG.getNode(ISD::UNDEF, MappedOps[0].getValueType());
3526 for (unsigned i = 0; i != NumElts; ++i)
3527 MappedOps[i] = UDVal;
3528 MappedOps[NumElts ] = *(N0.Val->op_end()-2);
3529 MappedOps[NumElts+1] = *(N0.Val->op_end()-1);
3530 UDVal = DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector,
3531 &MappedOps[0], MappedOps.size());
3533 return DAG.getNode(ISD::VVECTOR_SHUFFLE, MVT::Vector,
3534 N0, UDVal, ShufMask,
3535 MappedOps[NumElts], MappedOps[NumElts+1]);
3541 /// XformToShuffleWithZero - Returns a vector_shuffle if it able to transform
3542 /// a VAND to a vector_shuffle with the destination vector and a zero vector.
3543 /// e.g. VAND V, <0xffffffff, 0, 0xffffffff, 0>. ==>
3544 /// vector_shuffle V, Zero, <0, 4, 2, 4>
3545 SDOperand DAGCombiner::XformToShuffleWithZero(SDNode *N) {
3546 SDOperand LHS = N->getOperand(0);
3547 SDOperand RHS = N->getOperand(1);
3548 if (N->getOpcode() == ISD::VAND) {
3549 SDOperand DstVecSize = *(LHS.Val->op_end()-2);
3550 SDOperand DstVecEVT = *(LHS.Val->op_end()-1);
3551 if (RHS.getOpcode() == ISD::VBIT_CONVERT)
3552 RHS = RHS.getOperand(0);
3553 if (RHS.getOpcode() == ISD::VBUILD_VECTOR) {
3554 std::vector<SDOperand> IdxOps;
3555 unsigned NumOps = RHS.getNumOperands();
3556 unsigned NumElts = NumOps-2;
3557 MVT::ValueType EVT = cast<VTSDNode>(RHS.getOperand(NumOps-1))->getVT();
3558 for (unsigned i = 0; i != NumElts; ++i) {
3559 SDOperand Elt = RHS.getOperand(i);
3560 if (!isa<ConstantSDNode>(Elt))
3562 else if (cast<ConstantSDNode>(Elt)->isAllOnesValue())
3563 IdxOps.push_back(DAG.getConstant(i, EVT));
3564 else if (cast<ConstantSDNode>(Elt)->isNullValue())
3565 IdxOps.push_back(DAG.getConstant(NumElts, EVT));
3570 // Let's see if the target supports this vector_shuffle.
3571 if (!TLI.isVectorClearMaskLegal(IdxOps, EVT, DAG))
3574 // Return the new VVECTOR_SHUFFLE node.
3575 SDOperand NumEltsNode = DAG.getConstant(NumElts, MVT::i32);
3576 SDOperand EVTNode = DAG.getValueType(EVT);
3577 std::vector<SDOperand> Ops;
3578 LHS = DAG.getNode(ISD::VBIT_CONVERT, MVT::Vector, LHS, NumEltsNode,
3581 AddToWorkList(LHS.Val);
3582 std::vector<SDOperand> ZeroOps(NumElts, DAG.getConstant(0, EVT));
3583 ZeroOps.push_back(NumEltsNode);
3584 ZeroOps.push_back(EVTNode);
3585 Ops.push_back(DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector,
3586 &ZeroOps[0], ZeroOps.size()));
3587 IdxOps.push_back(NumEltsNode);
3588 IdxOps.push_back(EVTNode);
3589 Ops.push_back(DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector,
3590 &IdxOps[0], IdxOps.size()));
3591 Ops.push_back(NumEltsNode);
3592 Ops.push_back(EVTNode);
3593 SDOperand Result = DAG.getNode(ISD::VVECTOR_SHUFFLE, MVT::Vector,
3594 &Ops[0], Ops.size());
3595 if (NumEltsNode != DstVecSize || EVTNode != DstVecEVT) {
3596 Result = DAG.getNode(ISD::VBIT_CONVERT, MVT::Vector, Result,
3597 DstVecSize, DstVecEVT);
3605 /// visitVBinOp - Visit a binary vector operation, like VADD. IntOp indicates
3606 /// the scalar operation of the vop if it is operating on an integer vector
3607 /// (e.g. ADD) and FPOp indicates the FP version (e.g. FADD).
3608 SDOperand DAGCombiner::visitVBinOp(SDNode *N, ISD::NodeType IntOp,
3609 ISD::NodeType FPOp) {
3610 MVT::ValueType EltType = cast<VTSDNode>(*(N->op_end()-1))->getVT();
3611 ISD::NodeType ScalarOp = MVT::isInteger(EltType) ? IntOp : FPOp;
3612 SDOperand LHS = N->getOperand(0);
3613 SDOperand RHS = N->getOperand(1);
3614 SDOperand Shuffle = XformToShuffleWithZero(N);
3615 if (Shuffle.Val) return Shuffle;
3617 // If the LHS and RHS are VBUILD_VECTOR nodes, see if we can constant fold
3619 if (LHS.getOpcode() == ISD::VBUILD_VECTOR &&
3620 RHS.getOpcode() == ISD::VBUILD_VECTOR) {
3621 SmallVector<SDOperand, 8> Ops;
3622 for (unsigned i = 0, e = LHS.getNumOperands()-2; i != e; ++i) {
3623 SDOperand LHSOp = LHS.getOperand(i);
3624 SDOperand RHSOp = RHS.getOperand(i);
3625 // If these two elements can't be folded, bail out.
3626 if ((LHSOp.getOpcode() != ISD::UNDEF &&
3627 LHSOp.getOpcode() != ISD::Constant &&
3628 LHSOp.getOpcode() != ISD::ConstantFP) ||
3629 (RHSOp.getOpcode() != ISD::UNDEF &&
3630 RHSOp.getOpcode() != ISD::Constant &&
3631 RHSOp.getOpcode() != ISD::ConstantFP))
3633 // Can't fold divide by zero.
3634 if (N->getOpcode() == ISD::VSDIV || N->getOpcode() == ISD::VUDIV) {
3635 if ((RHSOp.getOpcode() == ISD::Constant &&
3636 cast<ConstantSDNode>(RHSOp.Val)->isNullValue()) ||
3637 (RHSOp.getOpcode() == ISD::ConstantFP &&
3638 !cast<ConstantFPSDNode>(RHSOp.Val)->getValue()))
3641 Ops.push_back(DAG.getNode(ScalarOp, EltType, LHSOp, RHSOp));
3642 AddToWorkList(Ops.back().Val);
3643 assert((Ops.back().getOpcode() == ISD::UNDEF ||
3644 Ops.back().getOpcode() == ISD::Constant ||
3645 Ops.back().getOpcode() == ISD::ConstantFP) &&
3646 "Scalar binop didn't fold!");
3649 if (Ops.size() == LHS.getNumOperands()-2) {
3650 Ops.push_back(*(LHS.Val->op_end()-2));
3651 Ops.push_back(*(LHS.Val->op_end()-1));
3652 return DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, &Ops[0], Ops.size());
3659 SDOperand DAGCombiner::SimplifySelect(SDOperand N0, SDOperand N1, SDOperand N2){
3660 assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!");
3662 SDOperand SCC = SimplifySelectCC(N0.getOperand(0), N0.getOperand(1), N1, N2,
3663 cast<CondCodeSDNode>(N0.getOperand(2))->get());
3664 // If we got a simplified select_cc node back from SimplifySelectCC, then
3665 // break it down into a new SETCC node, and a new SELECT node, and then return
3666 // the SELECT node, since we were called with a SELECT node.
3668 // Check to see if we got a select_cc back (to turn into setcc/select).
3669 // Otherwise, just return whatever node we got back, like fabs.
3670 if (SCC.getOpcode() == ISD::SELECT_CC) {
3671 SDOperand SETCC = DAG.getNode(ISD::SETCC, N0.getValueType(),
3672 SCC.getOperand(0), SCC.getOperand(1),
3674 AddToWorkList(SETCC.Val);
3675 return DAG.getNode(ISD::SELECT, SCC.getValueType(), SCC.getOperand(2),
3676 SCC.getOperand(3), SETCC);
3683 /// SimplifySelectOps - Given a SELECT or a SELECT_CC node, where LHS and RHS
3684 /// are the two values being selected between, see if we can simplify the
3685 /// select. Callers of this should assume that TheSelect is deleted if this
3686 /// returns true. As such, they should return the appropriate thing (e.g. the
3687 /// node) back to the top-level of the DAG combiner loop to avoid it being
3690 bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDOperand LHS,
3693 // If this is a select from two identical things, try to pull the operation
3694 // through the select.
3695 if (LHS.getOpcode() == RHS.getOpcode() && LHS.hasOneUse() && RHS.hasOneUse()){
3696 // If this is a load and the token chain is identical, replace the select
3697 // of two loads with a load through a select of the address to load from.
3698 // This triggers in things like "select bool X, 10.0, 123.0" after the FP
3699 // constants have been dropped into the constant pool.
3700 if (LHS.getOpcode() == ISD::LOAD &&
3701 // Token chains must be identical.
3702 LHS.getOperand(0) == RHS.getOperand(0)) {
3703 LoadSDNode *LLD = cast<LoadSDNode>(LHS);
3704 LoadSDNode *RLD = cast<LoadSDNode>(RHS);
3706 // If this is an EXTLOAD, the VT's must match.
3707 if (LLD->getLoadedVT() == RLD->getLoadedVT()) {
3708 // FIXME: this conflates two src values, discarding one. This is not
3709 // the right thing to do, but nothing uses srcvalues now. When they do,
3710 // turn SrcValue into a list of locations.
3712 if (TheSelect->getOpcode() == ISD::SELECT)
3713 Addr = DAG.getNode(ISD::SELECT, LLD->getBasePtr().getValueType(),
3714 TheSelect->getOperand(0), LLD->getBasePtr(),
3717 Addr = DAG.getNode(ISD::SELECT_CC, LLD->getBasePtr().getValueType(),
3718 TheSelect->getOperand(0),
3719 TheSelect->getOperand(1),
3720 LLD->getBasePtr(), RLD->getBasePtr(),
3721 TheSelect->getOperand(4));
3724 if (LLD->getExtensionType() == ISD::NON_EXTLOAD)
3725 Load = DAG.getLoad(TheSelect->getValueType(0), LLD->getChain(),
3726 Addr,LLD->getSrcValue(), LLD->getSrcValueOffset());
3728 Load = DAG.getExtLoad(LLD->getExtensionType(),
3729 TheSelect->getValueType(0),
3730 LLD->getChain(), Addr, LLD->getSrcValue(),
3731 LLD->getSrcValueOffset(),
3732 LLD->getLoadedVT());
3734 // Users of the select now use the result of the load.
3735 CombineTo(TheSelect, Load);
3737 // Users of the old loads now use the new load's chain. We know the
3738 // old-load value is dead now.
3739 CombineTo(LHS.Val, Load.getValue(0), Load.getValue(1));
3740 CombineTo(RHS.Val, Load.getValue(0), Load.getValue(1));
3749 SDOperand DAGCombiner::SimplifySelectCC(SDOperand N0, SDOperand N1,
3750 SDOperand N2, SDOperand N3,
3753 MVT::ValueType VT = N2.getValueType();
3754 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val);
3755 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.Val);
3756 ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N3.Val);
3758 // Determine if the condition we're dealing with is constant
3759 SDOperand SCC = SimplifySetCC(TLI.getSetCCResultTy(), N0, N1, CC, false);
3760 if (SCC.Val) AddToWorkList(SCC.Val);
3761 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.Val);
3763 // fold select_cc true, x, y -> x
3764 if (SCCC && SCCC->getValue())
3766 // fold select_cc false, x, y -> y
3767 if (SCCC && SCCC->getValue() == 0)
3770 // Check to see if we can simplify the select into an fabs node
3771 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) {
3772 // Allow either -0.0 or 0.0
3773 if (CFP->getValue() == 0.0) {
3774 // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs
3775 if ((CC == ISD::SETGE || CC == ISD::SETGT) &&
3776 N0 == N2 && N3.getOpcode() == ISD::FNEG &&
3777 N2 == N3.getOperand(0))
3778 return DAG.getNode(ISD::FABS, VT, N0);
3780 // select (setl[te] X, +/-0.0), fneg(X), X -> fabs
3781 if ((CC == ISD::SETLT || CC == ISD::SETLE) &&
3782 N0 == N3 && N2.getOpcode() == ISD::FNEG &&
3783 N2.getOperand(0) == N3)
3784 return DAG.getNode(ISD::FABS, VT, N3);
3788 // Check to see if we can perform the "gzip trick", transforming
3789 // select_cc setlt X, 0, A, 0 -> and (sra X, size(X)-1), A
3790 if (N1C && N3C && N3C->isNullValue() && CC == ISD::SETLT &&
3791 MVT::isInteger(N0.getValueType()) &&
3792 MVT::isInteger(N2.getValueType()) &&
3793 (N1C->isNullValue() || // (a < 0) ? b : 0
3794 (N1C->getValue() == 1 && N0 == N2))) { // (a < 1) ? a : 0
3795 MVT::ValueType XType = N0.getValueType();
3796 MVT::ValueType AType = N2.getValueType();
3797 if (XType >= AType) {
3798 // and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a
3799 // single-bit constant.
3800 if (N2C && ((N2C->getValue() & (N2C->getValue()-1)) == 0)) {
3801 unsigned ShCtV = Log2_64(N2C->getValue());
3802 ShCtV = MVT::getSizeInBits(XType)-ShCtV-1;
3803 SDOperand ShCt = DAG.getConstant(ShCtV, TLI.getShiftAmountTy());
3804 SDOperand Shift = DAG.getNode(ISD::SRL, XType, N0, ShCt);
3805 AddToWorkList(Shift.Val);
3806 if (XType > AType) {
3807 Shift = DAG.getNode(ISD::TRUNCATE, AType, Shift);
3808 AddToWorkList(Shift.Val);
3810 return DAG.getNode(ISD::AND, AType, Shift, N2);
3812 SDOperand Shift = DAG.getNode(ISD::SRA, XType, N0,
3813 DAG.getConstant(MVT::getSizeInBits(XType)-1,
3814 TLI.getShiftAmountTy()));
3815 AddToWorkList(Shift.Val);
3816 if (XType > AType) {
3817 Shift = DAG.getNode(ISD::TRUNCATE, AType, Shift);
3818 AddToWorkList(Shift.Val);
3820 return DAG.getNode(ISD::AND, AType, Shift, N2);
3824 // fold select C, 16, 0 -> shl C, 4
3825 if (N2C && N3C && N3C->isNullValue() && isPowerOf2_64(N2C->getValue()) &&
3826 TLI.getSetCCResultContents() == TargetLowering::ZeroOrOneSetCCResult) {
3827 // Get a SetCC of the condition
3828 // FIXME: Should probably make sure that setcc is legal if we ever have a
3829 // target where it isn't.
3830 SDOperand Temp, SCC;
3831 // cast from setcc result type to select result type
3832 if (AfterLegalize) {
3833 SCC = DAG.getSetCC(TLI.getSetCCResultTy(), N0, N1, CC);
3834 if (N2.getValueType() < SCC.getValueType())
3835 Temp = DAG.getZeroExtendInReg(SCC, N2.getValueType());
3837 Temp = DAG.getNode(ISD::ZERO_EXTEND, N2.getValueType(), SCC);
3839 SCC = DAG.getSetCC(MVT::i1, N0, N1, CC);
3840 Temp = DAG.getNode(ISD::ZERO_EXTEND, N2.getValueType(), SCC);
3842 AddToWorkList(SCC.Val);
3843 AddToWorkList(Temp.Val);
3844 // shl setcc result by log2 n2c
3845 return DAG.getNode(ISD::SHL, N2.getValueType(), Temp,
3846 DAG.getConstant(Log2_64(N2C->getValue()),
3847 TLI.getShiftAmountTy()));
3850 // Check to see if this is the equivalent of setcc
3851 // FIXME: Turn all of these into setcc if setcc if setcc is legal
3852 // otherwise, go ahead with the folds.
3853 if (0 && N3C && N3C->isNullValue() && N2C && (N2C->getValue() == 1ULL)) {
3854 MVT::ValueType XType = N0.getValueType();
3855 if (TLI.isOperationLegal(ISD::SETCC, TLI.getSetCCResultTy())) {
3856 SDOperand Res = DAG.getSetCC(TLI.getSetCCResultTy(), N0, N1, CC);
3857 if (Res.getValueType() != VT)
3858 Res = DAG.getNode(ISD::ZERO_EXTEND, VT, Res);
3862 // seteq X, 0 -> srl (ctlz X, log2(size(X)))
3863 if (N1C && N1C->isNullValue() && CC == ISD::SETEQ &&
3864 TLI.isOperationLegal(ISD::CTLZ, XType)) {
3865 SDOperand Ctlz = DAG.getNode(ISD::CTLZ, XType, N0);
3866 return DAG.getNode(ISD::SRL, XType, Ctlz,
3867 DAG.getConstant(Log2_32(MVT::getSizeInBits(XType)),
3868 TLI.getShiftAmountTy()));
3870 // setgt X, 0 -> srl (and (-X, ~X), size(X)-1)
3871 if (N1C && N1C->isNullValue() && CC == ISD::SETGT) {
3872 SDOperand NegN0 = DAG.getNode(ISD::SUB, XType, DAG.getConstant(0, XType),
3874 SDOperand NotN0 = DAG.getNode(ISD::XOR, XType, N0,
3875 DAG.getConstant(~0ULL, XType));
3876 return DAG.getNode(ISD::SRL, XType,
3877 DAG.getNode(ISD::AND, XType, NegN0, NotN0),
3878 DAG.getConstant(MVT::getSizeInBits(XType)-1,
3879 TLI.getShiftAmountTy()));
3881 // setgt X, -1 -> xor (srl (X, size(X)-1), 1)
3882 if (N1C && N1C->isAllOnesValue() && CC == ISD::SETGT) {
3883 SDOperand Sign = DAG.getNode(ISD::SRL, XType, N0,
3884 DAG.getConstant(MVT::getSizeInBits(XType)-1,
3885 TLI.getShiftAmountTy()));
3886 return DAG.getNode(ISD::XOR, XType, Sign, DAG.getConstant(1, XType));
3890 // Check to see if this is an integer abs. select_cc setl[te] X, 0, -X, X ->
3891 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
3892 if (N1C && N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE) &&
3893 N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1)) {
3894 if (ConstantSDNode *SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0))) {
3895 MVT::ValueType XType = N0.getValueType();
3896 if (SubC->isNullValue() && MVT::isInteger(XType)) {
3897 SDOperand Shift = DAG.getNode(ISD::SRA, XType, N0,
3898 DAG.getConstant(MVT::getSizeInBits(XType)-1,
3899 TLI.getShiftAmountTy()));
3900 SDOperand Add = DAG.getNode(ISD::ADD, XType, N0, Shift);
3901 AddToWorkList(Shift.Val);
3902 AddToWorkList(Add.Val);
3903 return DAG.getNode(ISD::XOR, XType, Add, Shift);
3911 SDOperand DAGCombiner::SimplifySetCC(MVT::ValueType VT, SDOperand N0,
3912 SDOperand N1, ISD::CondCode Cond,
3913 bool foldBooleans) {
3914 // These setcc operations always fold.
3918 case ISD::SETFALSE2: return DAG.getConstant(0, VT);
3920 case ISD::SETTRUE2: return DAG.getConstant(1, VT);
3923 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val)) {
3924 uint64_t C1 = N1C->getValue();
3925 if (isa<ConstantSDNode>(N0.Val)) {
3926 return DAG.FoldSetCC(VT, N0, N1, Cond);
3928 // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an
3929 // equality comparison, then we're just comparing whether X itself is
3931 if (N0.getOpcode() == ISD::SRL && (C1 == 0 || C1 == 1) &&
3932 N0.getOperand(0).getOpcode() == ISD::CTLZ &&
3933 N0.getOperand(1).getOpcode() == ISD::Constant) {
3934 unsigned ShAmt = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
3935 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
3936 ShAmt == Log2_32(MVT::getSizeInBits(N0.getValueType()))) {
3937 if ((C1 == 0) == (Cond == ISD::SETEQ)) {
3938 // (srl (ctlz x), 5) == 0 -> X != 0
3939 // (srl (ctlz x), 5) != 1 -> X != 0
3942 // (srl (ctlz x), 5) != 0 -> X == 0
3943 // (srl (ctlz x), 5) == 1 -> X == 0
3946 SDOperand Zero = DAG.getConstant(0, N0.getValueType());
3947 return DAG.getSetCC(VT, N0.getOperand(0).getOperand(0),
3952 // If the LHS is a ZERO_EXTEND, perform the comparison on the input.
3953 if (N0.getOpcode() == ISD::ZERO_EXTEND) {
3954 unsigned InSize = MVT::getSizeInBits(N0.getOperand(0).getValueType());
3956 // If the comparison constant has bits in the upper part, the
3957 // zero-extended value could never match.
3958 if (C1 & (~0ULL << InSize)) {
3959 unsigned VSize = MVT::getSizeInBits(N0.getValueType());
3963 case ISD::SETEQ: return DAG.getConstant(0, VT);
3966 case ISD::SETNE: return DAG.getConstant(1, VT);
3969 // True if the sign bit of C1 is set.
3970 return DAG.getConstant((C1 & (1ULL << VSize)) != 0, VT);
3973 // True if the sign bit of C1 isn't set.
3974 return DAG.getConstant((C1 & (1ULL << VSize)) == 0, VT);
3980 // Otherwise, we can perform the comparison with the low bits.
3988 return DAG.getSetCC(VT, N0.getOperand(0),
3989 DAG.getConstant(C1, N0.getOperand(0).getValueType()),
3992 break; // todo, be more careful with signed comparisons
3994 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
3995 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
3996 MVT::ValueType ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT();
3997 unsigned ExtSrcTyBits = MVT::getSizeInBits(ExtSrcTy);
3998 MVT::ValueType ExtDstTy = N0.getValueType();
3999 unsigned ExtDstTyBits = MVT::getSizeInBits(ExtDstTy);
4001 // If the extended part has any inconsistent bits, it cannot ever
4002 // compare equal. In other words, they have to be all ones or all
4005 (~0ULL >> (64-ExtSrcTyBits)) & (~0ULL << (ExtDstTyBits-1));
4006 if ((C1 & ExtBits) != 0 && (C1 & ExtBits) != ExtBits)
4007 return DAG.getConstant(Cond == ISD::SETNE, VT);
4010 MVT::ValueType Op0Ty = N0.getOperand(0).getValueType();
4011 if (Op0Ty == ExtSrcTy) {
4012 ZextOp = N0.getOperand(0);
4014 int64_t Imm = ~0ULL >> (64-ExtSrcTyBits);
4015 ZextOp = DAG.getNode(ISD::AND, Op0Ty, N0.getOperand(0),
4016 DAG.getConstant(Imm, Op0Ty));
4018 AddToWorkList(ZextOp.Val);
4019 // Otherwise, make this a use of a zext.
4020 return DAG.getSetCC(VT, ZextOp,
4021 DAG.getConstant(C1 & (~0ULL>>(64-ExtSrcTyBits)),
4024 } else if ((N1C->getValue() == 0 || N1C->getValue() == 1) &&
4025 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
4027 // SETCC (SETCC), [0|1], [EQ|NE] -> SETCC
4028 if (N0.getOpcode() == ISD::SETCC) {
4029 bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (N1C->getValue() != 1);
4033 // Invert the condition.
4034 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
4035 CC = ISD::getSetCCInverse(CC,
4036 MVT::isInteger(N0.getOperand(0).getValueType()));
4037 return DAG.getSetCC(VT, N0.getOperand(0), N0.getOperand(1), CC);
4040 if ((N0.getOpcode() == ISD::XOR ||
4041 (N0.getOpcode() == ISD::AND &&
4042 N0.getOperand(0).getOpcode() == ISD::XOR &&
4043 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) &&
4044 isa<ConstantSDNode>(N0.getOperand(1)) &&
4045 cast<ConstantSDNode>(N0.getOperand(1))->getValue() == 1) {
4046 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We
4047 // can only do this if the top bits are known zero.
4048 if (TLI.MaskedValueIsZero(N0,
4049 MVT::getIntVTBitMask(N0.getValueType())-1)){
4050 // Okay, get the un-inverted input value.
4052 if (N0.getOpcode() == ISD::XOR)
4053 Val = N0.getOperand(0);
4055 assert(N0.getOpcode() == ISD::AND &&
4056 N0.getOperand(0).getOpcode() == ISD::XOR);
4057 // ((X^1)&1)^1 -> X & 1
4058 Val = DAG.getNode(ISD::AND, N0.getValueType(),
4059 N0.getOperand(0).getOperand(0),
4062 return DAG.getSetCC(VT, Val, N1,
4063 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
4068 uint64_t MinVal, MaxVal;
4069 unsigned OperandBitSize = MVT::getSizeInBits(N1C->getValueType(0));
4070 if (ISD::isSignedIntSetCC(Cond)) {
4071 MinVal = 1ULL << (OperandBitSize-1);
4072 if (OperandBitSize != 1) // Avoid X >> 64, which is undefined.
4073 MaxVal = ~0ULL >> (65-OperandBitSize);
4078 MaxVal = ~0ULL >> (64-OperandBitSize);
4081 // Canonicalize GE/LE comparisons to use GT/LT comparisons.
4082 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) {
4083 if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true
4084 --C1; // X >= C0 --> X > (C0-1)
4085 return DAG.getSetCC(VT, N0, DAG.getConstant(C1, N1.getValueType()),
4086 (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT);
4089 if (Cond == ISD::SETLE || Cond == ISD::SETULE) {
4090 if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true
4091 ++C1; // X <= C0 --> X < (C0+1)
4092 return DAG.getSetCC(VT, N0, DAG.getConstant(C1, N1.getValueType()),
4093 (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT);
4096 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal)
4097 return DAG.getConstant(0, VT); // X < MIN --> false
4099 // Canonicalize setgt X, Min --> setne X, Min
4100 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal)
4101 return DAG.getSetCC(VT, N0, N1, ISD::SETNE);
4102 // Canonicalize setlt X, Max --> setne X, Max
4103 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal)
4104 return DAG.getSetCC(VT, N0, N1, ISD::SETNE);
4106 // If we have setult X, 1, turn it into seteq X, 0
4107 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1)
4108 return DAG.getSetCC(VT, N0, DAG.getConstant(MinVal, N0.getValueType()),
4110 // If we have setugt X, Max-1, turn it into seteq X, Max
4111 else if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1)
4112 return DAG.getSetCC(VT, N0, DAG.getConstant(MaxVal, N0.getValueType()),
4115 // If we have "setcc X, C0", check to see if we can shrink the immediate
4118 // SETUGT X, SINTMAX -> SETLT X, 0
4119 if (Cond == ISD::SETUGT && OperandBitSize != 1 &&
4120 C1 == (~0ULL >> (65-OperandBitSize)))
4121 return DAG.getSetCC(VT, N0, DAG.getConstant(0, N1.getValueType()),
4124 // FIXME: Implement the rest of these.
4126 // Fold bit comparisons when we can.
4127 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
4128 VT == N0.getValueType() && N0.getOpcode() == ISD::AND)
4129 if (ConstantSDNode *AndRHS =
4130 dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
4131 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3
4132 // Perform the xform if the AND RHS is a single bit.
4133 if (isPowerOf2_64(AndRHS->getValue())) {
4134 return DAG.getNode(ISD::SRL, VT, N0,
4135 DAG.getConstant(Log2_64(AndRHS->getValue()),
4136 TLI.getShiftAmountTy()));
4138 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getValue()) {
4139 // (X & 8) == 8 --> (X & 8) >> 3
4140 // Perform the xform if C1 is a single bit.
4141 if (isPowerOf2_64(C1)) {
4142 return DAG.getNode(ISD::SRL, VT, N0,
4143 DAG.getConstant(Log2_64(C1),TLI.getShiftAmountTy()));
4148 } else if (isa<ConstantSDNode>(N0.Val)) {
4149 // Ensure that the constant occurs on the RHS.
4150 return DAG.getSetCC(VT, N1, N0, ISD::getSetCCSwappedOperands(Cond));
4153 if (isa<ConstantFPSDNode>(N0.Val)) {
4154 // Constant fold or commute setcc.
4155 SDOperand O = DAG.FoldSetCC(VT, N0, N1, Cond);
4156 if (O.Val) return O;
4160 // We can always fold X == X for integer setcc's.
4161 if (MVT::isInteger(N0.getValueType()))
4162 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
4163 unsigned UOF = ISD::getUnorderedFlavor(Cond);
4164 if (UOF == 2) // FP operators that are undefined on NaNs.
4165 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
4166 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond)))
4167 return DAG.getConstant(UOF, VT);
4168 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO
4169 // if it is not already.
4170 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO;
4171 if (NewCond != Cond)
4172 return DAG.getSetCC(VT, N0, N1, NewCond);
4175 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
4176 MVT::isInteger(N0.getValueType())) {
4177 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB ||
4178 N0.getOpcode() == ISD::XOR) {
4179 // Simplify (X+Y) == (X+Z) --> Y == Z
4180 if (N0.getOpcode() == N1.getOpcode()) {
4181 if (N0.getOperand(0) == N1.getOperand(0))
4182 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(1), Cond);
4183 if (N0.getOperand(1) == N1.getOperand(1))
4184 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(0), Cond);
4185 if (DAG.isCommutativeBinOp(N0.getOpcode())) {
4186 // If X op Y == Y op X, try other combinations.
4187 if (N0.getOperand(0) == N1.getOperand(1))
4188 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(0), Cond);
4189 if (N0.getOperand(1) == N1.getOperand(0))
4190 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(1), Cond);
4194 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) {
4195 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
4196 // Turn (X+C1) == C2 --> X == C2-C1
4197 if (N0.getOpcode() == ISD::ADD && N0.Val->hasOneUse()) {
4198 return DAG.getSetCC(VT, N0.getOperand(0),
4199 DAG.getConstant(RHSC->getValue()-LHSR->getValue(),
4200 N0.getValueType()), Cond);
4203 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0.
4204 if (N0.getOpcode() == ISD::XOR)
4205 // If we know that all of the inverted bits are zero, don't bother
4206 // performing the inversion.
4207 if (TLI.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getValue()))
4208 return DAG.getSetCC(VT, N0.getOperand(0),
4209 DAG.getConstant(LHSR->getValue()^RHSC->getValue(),
4210 N0.getValueType()), Cond);
4213 // Turn (C1-X) == C2 --> X == C1-C2
4214 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) {
4215 if (N0.getOpcode() == ISD::SUB && N0.Val->hasOneUse()) {
4216 return DAG.getSetCC(VT, N0.getOperand(1),
4217 DAG.getConstant(SUBC->getValue()-RHSC->getValue(),
4218 N0.getValueType()), Cond);
4223 // Simplify (X+Z) == X --> Z == 0
4224 if (N0.getOperand(0) == N1)
4225 return DAG.getSetCC(VT, N0.getOperand(1),
4226 DAG.getConstant(0, N0.getValueType()), Cond);
4227 if (N0.getOperand(1) == N1) {
4228 if (DAG.isCommutativeBinOp(N0.getOpcode()))
4229 return DAG.getSetCC(VT, N0.getOperand(0),
4230 DAG.getConstant(0, N0.getValueType()), Cond);
4232 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!");
4233 // (Z-X) == X --> Z == X<<1
4234 SDOperand SH = DAG.getNode(ISD::SHL, N1.getValueType(),
4236 DAG.getConstant(1,TLI.getShiftAmountTy()));
4237 AddToWorkList(SH.Val);
4238 return DAG.getSetCC(VT, N0.getOperand(0), SH, Cond);
4243 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB ||
4244 N1.getOpcode() == ISD::XOR) {
4245 // Simplify X == (X+Z) --> Z == 0
4246 if (N1.getOperand(0) == N0) {
4247 return DAG.getSetCC(VT, N1.getOperand(1),
4248 DAG.getConstant(0, N1.getValueType()), Cond);
4249 } else if (N1.getOperand(1) == N0) {
4250 if (DAG.isCommutativeBinOp(N1.getOpcode())) {
4251 return DAG.getSetCC(VT, N1.getOperand(0),
4252 DAG.getConstant(0, N1.getValueType()), Cond);
4254 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!");
4255 // X == (Z-X) --> X<<1 == Z
4256 SDOperand SH = DAG.getNode(ISD::SHL, N1.getValueType(), N0,
4257 DAG.getConstant(1,TLI.getShiftAmountTy()));
4258 AddToWorkList(SH.Val);
4259 return DAG.getSetCC(VT, SH, N1.getOperand(0), Cond);
4265 // Fold away ALL boolean setcc's.
4267 if (N0.getValueType() == MVT::i1 && foldBooleans) {
4269 default: assert(0 && "Unknown integer setcc!");
4270 case ISD::SETEQ: // X == Y -> (X^Y)^1
4271 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, N1);
4272 N0 = DAG.getNode(ISD::XOR, MVT::i1, Temp, DAG.getConstant(1, MVT::i1));
4273 AddToWorkList(Temp.Val);
4275 case ISD::SETNE: // X != Y --> (X^Y)
4276 N0 = DAG.getNode(ISD::XOR, MVT::i1, N0, N1);
4278 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> X^1 & Y
4279 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> X^1 & Y
4280 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1));
4281 N0 = DAG.getNode(ISD::AND, MVT::i1, N1, Temp);
4282 AddToWorkList(Temp.Val);
4284 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> Y^1 & X
4285 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> Y^1 & X
4286 Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1));
4287 N0 = DAG.getNode(ISD::AND, MVT::i1, N0, Temp);
4288 AddToWorkList(Temp.Val);
4290 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> X^1 | Y
4291 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> X^1 | Y
4292 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1));
4293 N0 = DAG.getNode(ISD::OR, MVT::i1, N1, Temp);
4294 AddToWorkList(Temp.Val);
4296 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> Y^1 | X
4297 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> Y^1 | X
4298 Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1));
4299 N0 = DAG.getNode(ISD::OR, MVT::i1, N0, Temp);
4302 if (VT != MVT::i1) {
4303 AddToWorkList(N0.Val);
4304 // FIXME: If running after legalize, we probably can't do this.
4305 N0 = DAG.getNode(ISD::ZERO_EXTEND, VT, N0);
4310 // Could not fold it.
4314 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
4315 /// return a DAG expression to select that will generate the same value by
4316 /// multiplying by a magic number. See:
4317 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
4318 SDOperand DAGCombiner::BuildSDIV(SDNode *N) {
4319 std::vector<SDNode*> Built;
4320 SDOperand S = TLI.BuildSDIV(N, DAG, &Built);
4322 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end();
4328 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
4329 /// return a DAG expression to select that will generate the same value by
4330 /// multiplying by a magic number. See:
4331 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
4332 SDOperand DAGCombiner::BuildUDIV(SDNode *N) {
4333 std::vector<SDNode*> Built;
4334 SDOperand S = TLI.BuildUDIV(N, DAG, &Built);
4336 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end();
4342 /// FindBaseOffset - Return true if base is known not to alias with anything
4343 /// but itself. Provides base object and offset as results.
4344 static bool FindBaseOffset(SDOperand Ptr, SDOperand &Base, int64_t &Offset) {
4345 // Assume it is a primitive operation.
4346 Base = Ptr; Offset = 0;
4348 // If it's an adding a simple constant then integrate the offset.
4349 if (Base.getOpcode() == ISD::ADD) {
4350 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Base.getOperand(1))) {
4351 Base = Base.getOperand(0);
4352 Offset += C->getValue();
4356 // If it's any of the following then it can't alias with anything but itself.
4357 return isa<FrameIndexSDNode>(Base) ||
4358 isa<ConstantPoolSDNode>(Base) ||
4359 isa<GlobalAddressSDNode>(Base);
4362 /// isAlias - Return true if there is any possibility that the two addresses
4364 bool DAGCombiner::isAlias(SDOperand Ptr1, int64_t Size1,
4365 const Value *SrcValue1, int SrcValueOffset1,
4366 SDOperand Ptr2, int64_t Size2,
4367 const Value *SrcValue2, int SrcValueOffset2)
4369 // If they are the same then they must be aliases.
4370 if (Ptr1 == Ptr2) return true;
4372 // Gather base node and offset information.
4373 SDOperand Base1, Base2;
4374 int64_t Offset1, Offset2;
4375 bool KnownBase1 = FindBaseOffset(Ptr1, Base1, Offset1);
4376 bool KnownBase2 = FindBaseOffset(Ptr2, Base2, Offset2);
4378 // If they have a same base address then...
4379 if (Base1 == Base2) {
4380 // Check to see if the addresses overlap.
4381 return!((Offset1 + Size1) <= Offset2 || (Offset2 + Size2) <= Offset1);
4384 // If we know both bases then they can't alias.
4385 if (KnownBase1 && KnownBase2) return false;
4387 if (CombinerGlobalAA) {
4388 // Use alias analysis information.
4389 int Overlap1 = Size1 + SrcValueOffset1 + Offset1;
4390 int Overlap2 = Size2 + SrcValueOffset2 + Offset2;
4391 AliasAnalysis::AliasResult AAResult =
4392 AA.alias(SrcValue1, Overlap1, SrcValue2, Overlap2);
4393 if (AAResult == AliasAnalysis::NoAlias)
4397 // Otherwise we have to assume they alias.
4401 /// FindAliasInfo - Extracts the relevant alias information from the memory
4402 /// node. Returns true if the operand was a load.
4403 bool DAGCombiner::FindAliasInfo(SDNode *N,
4404 SDOperand &Ptr, int64_t &Size,
4405 const Value *&SrcValue, int &SrcValueOffset) {
4406 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4407 Ptr = LD->getBasePtr();
4408 Size = MVT::getSizeInBits(LD->getLoadedVT()) >> 3;
4409 SrcValue = LD->getSrcValue();
4410 SrcValueOffset = LD->getSrcValueOffset();
4412 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4413 Ptr = ST->getBasePtr();
4414 Size = MVT::getSizeInBits(ST->getStoredVT()) >> 3;
4415 SrcValue = ST->getSrcValue();
4416 SrcValueOffset = ST->getSrcValueOffset();
4418 assert(0 && "FindAliasInfo expected a memory operand");
4424 /// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes,
4425 /// looking for aliasing nodes and adding them to the Aliases vector.
4426 void DAGCombiner::GatherAllAliases(SDNode *N, SDOperand OriginalChain,
4427 SmallVector<SDOperand, 8> &Aliases) {
4428 SmallVector<SDOperand, 8> Chains; // List of chains to visit.
4429 std::set<SDNode *> Visited; // Visited node set.
4431 // Get alias information for node.
4434 const Value *SrcValue;
4436 bool IsLoad = FindAliasInfo(N, Ptr, Size, SrcValue, SrcValueOffset);
4439 Chains.push_back(OriginalChain);
4441 // Look at each chain and determine if it is an alias. If so, add it to the
4442 // aliases list. If not, then continue up the chain looking for the next
4444 while (!Chains.empty()) {
4445 SDOperand Chain = Chains.back();
4448 // Don't bother if we've been before.
4449 if (Visited.find(Chain.Val) != Visited.end()) continue;
4450 Visited.insert(Chain.Val);
4452 switch (Chain.getOpcode()) {
4453 case ISD::EntryToken:
4454 // Entry token is ideal chain operand, but handled in FindBetterChain.
4459 // Get alias information for Chain.
4462 const Value *OpSrcValue;
4463 int OpSrcValueOffset;
4464 bool IsOpLoad = FindAliasInfo(Chain.Val, OpPtr, OpSize,
4465 OpSrcValue, OpSrcValueOffset);
4467 // If chain is alias then stop here.
4468 if (!(IsLoad && IsOpLoad) &&
4469 isAlias(Ptr, Size, SrcValue, SrcValueOffset,
4470 OpPtr, OpSize, OpSrcValue, OpSrcValueOffset)) {
4471 Aliases.push_back(Chain);
4473 // Look further up the chain.
4474 Chains.push_back(Chain.getOperand(0));
4475 // Clean up old chain.
4476 AddToWorkList(Chain.Val);
4481 case ISD::TokenFactor:
4482 // We have to check each of the operands of the token factor, so we queue
4483 // then up. Adding the operands to the queue (stack) in reverse order
4484 // maintains the original order and increases the likelihood that getNode
4485 // will find a matching token factor (CSE.)
4486 for (unsigned n = Chain.getNumOperands(); n;)
4487 Chains.push_back(Chain.getOperand(--n));
4488 // Eliminate the token factor if we can.
4489 AddToWorkList(Chain.Val);
4493 // For all other instructions we will just have to take what we can get.
4494 Aliases.push_back(Chain);
4500 /// FindBetterChain - Walk up chain skipping non-aliasing memory nodes, looking
4501 /// for a better chain (aliasing node.)
4502 SDOperand DAGCombiner::FindBetterChain(SDNode *N, SDOperand OldChain) {
4503 SmallVector<SDOperand, 8> Aliases; // Ops for replacing token factor.
4505 // Accumulate all the aliases to this node.
4506 GatherAllAliases(N, OldChain, Aliases);
4508 if (Aliases.size() == 0) {
4509 // If no operands then chain to entry token.
4510 return DAG.getEntryNode();
4511 } else if (Aliases.size() == 1) {
4512 // If a single operand then chain to it. We don't need to revisit it.
4516 // Construct a custom tailored token factor.
4517 SDOperand NewChain = DAG.getNode(ISD::TokenFactor, MVT::Other,
4518 &Aliases[0], Aliases.size());
4520 // Make sure the old chain gets cleaned up.
4521 if (NewChain != OldChain) AddToWorkList(OldChain.Val);
4526 // SelectionDAG::Combine - This is the entry point for the file.
4528 void SelectionDAG::Combine(bool RunningAfterLegalize, AliasAnalysis &AA) {
4529 if (!RunningAfterLegalize && ViewDAGCombine1)
4531 if (RunningAfterLegalize && ViewDAGCombine2)
4533 /// run - This is the main entry point to this class.
4535 DAGCombiner(*this, AA).Run(RunningAfterLegalize);