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/CodeGen/SelectionDAG.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/Support/MathExtras.h"
36 #include "llvm/Target/TargetLowering.h"
37 #include "llvm/Support/Compiler.h"
38 #include "llvm/Support/CommandLine.h"
46 static Statistic<> NodesCombined ("dagcombiner",
47 "Number of dag nodes combined");
51 CombinerAA("combiner-alias-analysis", cl::Hidden,
52 cl::desc("Turn on alias analysis turning testing"));
54 class VISIBILITY_HIDDEN DAGCombiner {
59 // Worklist of all of the nodes that need to be simplified.
60 std::vector<SDNode*> WorkList;
62 /// AddUsersToWorkList - When an instruction is simplified, add all users of
63 /// the instruction to the work lists because they might get more simplified
66 void AddUsersToWorkList(SDNode *N) {
67 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
69 WorkList.push_back(*UI);
72 /// removeFromWorkList - remove all instances of N from the worklist.
74 void removeFromWorkList(SDNode *N) {
75 WorkList.erase(std::remove(WorkList.begin(), WorkList.end(), N),
80 void AddToWorkList(SDNode *N) {
81 WorkList.push_back(N);
84 SDOperand CombineTo(SDNode *N, const SDOperand *To, unsigned NumTo) {
85 assert(N->getNumValues() == NumTo && "Broken CombineTo call!");
87 DEBUG(std::cerr << "\nReplacing "; N->dump();
88 std::cerr << "\nWith: "; To[0].Val->dump(&DAG);
89 std::cerr << " and " << NumTo-1 << " other values\n");
90 std::vector<SDNode*> NowDead;
91 DAG.ReplaceAllUsesWith(N, To, &NowDead);
93 // Push the new nodes and any users onto the worklist
94 for (unsigned i = 0, e = NumTo; i != e; ++i) {
95 WorkList.push_back(To[i].Val);
96 AddUsersToWorkList(To[i].Val);
99 // Nodes can end up on the worklist more than once. Make sure we do
100 // not process a node that has been replaced.
101 removeFromWorkList(N);
102 for (unsigned i = 0, e = NowDead.size(); i != e; ++i)
103 removeFromWorkList(NowDead[i]);
105 // Finally, since the node is now dead, remove it from the graph.
107 return SDOperand(N, 0);
110 SDOperand CombineTo(SDNode *N, SDOperand Res) {
111 return CombineTo(N, &Res, 1);
114 SDOperand CombineTo(SDNode *N, SDOperand Res0, SDOperand Res1) {
115 SDOperand To[] = { Res0, Res1 };
116 return CombineTo(N, To, 2);
120 /// SimplifyDemandedBits - Check the specified integer node value to see if
121 /// it can be simplified or if things it uses can be simplified by bit
122 /// propagation. If so, return true.
123 bool SimplifyDemandedBits(SDOperand Op) {
124 TargetLowering::TargetLoweringOpt TLO(DAG);
125 uint64_t KnownZero, KnownOne;
126 uint64_t Demanded = MVT::getIntVTBitMask(Op.getValueType());
127 if (!TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
131 WorkList.push_back(Op.Val);
133 // Replace the old value with the new one.
135 DEBUG(std::cerr << "\nReplacing "; TLO.Old.Val->dump();
136 std::cerr << "\nWith: "; TLO.New.Val->dump(&DAG);
139 std::vector<SDNode*> NowDead;
140 DAG.ReplaceAllUsesOfValueWith(TLO.Old, TLO.New, NowDead);
142 // Push the new node and any (possibly new) users onto the worklist.
143 WorkList.push_back(TLO.New.Val);
144 AddUsersToWorkList(TLO.New.Val);
146 // Nodes can end up on the worklist more than once. Make sure we do
147 // not process a node that has been replaced.
148 for (unsigned i = 0, e = NowDead.size(); i != e; ++i)
149 removeFromWorkList(NowDead[i]);
151 // Finally, if the node is now dead, remove it from the graph. The node
152 // may not be dead if the replacement process recursively simplified to
153 // something else needing this node.
154 if (TLO.Old.Val->use_empty()) {
155 removeFromWorkList(TLO.Old.Val);
156 DAG.DeleteNode(TLO.Old.Val);
161 /// visit - call the node-specific routine that knows how to fold each
162 /// particular type of node.
163 SDOperand visit(SDNode *N);
165 // Visitation implementation - Implement dag node combining for different
166 // node types. The semantics are as follows:
168 // SDOperand.Val == 0 - No change was made
169 // SDOperand.Val == N - N was replaced, is dead, and is already handled.
170 // otherwise - N should be replaced by the returned Operand.
172 SDOperand visitTokenFactor(SDNode *N);
173 SDOperand visitADD(SDNode *N);
174 SDOperand visitSUB(SDNode *N);
175 SDOperand visitMUL(SDNode *N);
176 SDOperand visitSDIV(SDNode *N);
177 SDOperand visitUDIV(SDNode *N);
178 SDOperand visitSREM(SDNode *N);
179 SDOperand visitUREM(SDNode *N);
180 SDOperand visitMULHU(SDNode *N);
181 SDOperand visitMULHS(SDNode *N);
182 SDOperand visitAND(SDNode *N);
183 SDOperand visitOR(SDNode *N);
184 SDOperand visitXOR(SDNode *N);
185 SDOperand visitVBinOp(SDNode *N, ISD::NodeType IntOp, ISD::NodeType FPOp);
186 SDOperand visitSHL(SDNode *N);
187 SDOperand visitSRA(SDNode *N);
188 SDOperand visitSRL(SDNode *N);
189 SDOperand visitCTLZ(SDNode *N);
190 SDOperand visitCTTZ(SDNode *N);
191 SDOperand visitCTPOP(SDNode *N);
192 SDOperand visitSELECT(SDNode *N);
193 SDOperand visitSELECT_CC(SDNode *N);
194 SDOperand visitSETCC(SDNode *N);
195 SDOperand visitSIGN_EXTEND(SDNode *N);
196 SDOperand visitZERO_EXTEND(SDNode *N);
197 SDOperand visitANY_EXTEND(SDNode *N);
198 SDOperand visitSIGN_EXTEND_INREG(SDNode *N);
199 SDOperand visitTRUNCATE(SDNode *N);
200 SDOperand visitBIT_CONVERT(SDNode *N);
201 SDOperand visitVBIT_CONVERT(SDNode *N);
202 SDOperand visitFADD(SDNode *N);
203 SDOperand visitFSUB(SDNode *N);
204 SDOperand visitFMUL(SDNode *N);
205 SDOperand visitFDIV(SDNode *N);
206 SDOperand visitFREM(SDNode *N);
207 SDOperand visitFCOPYSIGN(SDNode *N);
208 SDOperand visitSINT_TO_FP(SDNode *N);
209 SDOperand visitUINT_TO_FP(SDNode *N);
210 SDOperand visitFP_TO_SINT(SDNode *N);
211 SDOperand visitFP_TO_UINT(SDNode *N);
212 SDOperand visitFP_ROUND(SDNode *N);
213 SDOperand visitFP_ROUND_INREG(SDNode *N);
214 SDOperand visitFP_EXTEND(SDNode *N);
215 SDOperand visitFNEG(SDNode *N);
216 SDOperand visitFABS(SDNode *N);
217 SDOperand visitBRCOND(SDNode *N);
218 SDOperand visitBR_CC(SDNode *N);
219 SDOperand visitLOAD(SDNode *N);
220 SDOperand visitXEXTLOAD(SDNode *N);
221 SDOperand visitSTORE(SDNode *N);
222 SDOperand visitINSERT_VECTOR_ELT(SDNode *N);
223 SDOperand visitVINSERT_VECTOR_ELT(SDNode *N);
224 SDOperand visitVBUILD_VECTOR(SDNode *N);
225 SDOperand visitVECTOR_SHUFFLE(SDNode *N);
226 SDOperand visitVVECTOR_SHUFFLE(SDNode *N);
228 SDOperand XformToShuffleWithZero(SDNode *N);
229 SDOperand ReassociateOps(unsigned Opc, SDOperand LHS, SDOperand RHS);
231 bool SimplifySelectOps(SDNode *SELECT, SDOperand LHS, SDOperand RHS);
232 SDOperand SimplifyBinOpWithSameOpcodeHands(SDNode *N);
233 SDOperand SimplifySelect(SDOperand N0, SDOperand N1, SDOperand N2);
234 SDOperand SimplifySelectCC(SDOperand N0, SDOperand N1, SDOperand N2,
235 SDOperand N3, ISD::CondCode CC);
236 SDOperand SimplifySetCC(MVT::ValueType VT, SDOperand N0, SDOperand N1,
237 ISD::CondCode Cond, bool foldBooleans = true);
238 SDOperand ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(SDNode *, MVT::ValueType);
239 SDOperand BuildSDIV(SDNode *N);
240 SDOperand BuildUDIV(SDNode *N);
241 SDNode *MatchRotate(SDOperand LHS, SDOperand RHS);
243 /// FindBaseOffset - Return true if we can determine base and offset
244 /// information from a given pointer operand. Provides base and offset as a
246 static bool FindBaseOffset(SDOperand Ptr,
247 SDOperand &Object, int64_t &Offset);
249 /// isAlias - Return true if there is the possibility that the two addresses
251 static bool isAlias(SDOperand Ptr1, int64_t Size1, SDOperand SrcValue1,
252 SDOperand Ptr2, int64_t Size2, SDOperand SrcValue2);
254 /// FindAliasInfo - Extracts the relevant alias information from the memory
256 static void FindAliasInfo(SDNode *N,
257 SDOperand &Ptr, int64_t &Size, SDOperand &SrcValue);
259 /// hasChain - Return true if Op has a chain. Provides chain if present.
261 static bool hasChain(SDOperand Op, SDOperand &Chain);
263 /// FindBetterChain - Walk up chain skipping non-aliasing memory nodes,
264 /// looking for a better chain.
265 SDOperand FindBetterChain(SDNode *N, SDOperand Chain);
268 DAGCombiner(SelectionDAG &D)
269 : DAG(D), TLI(D.getTargetLoweringInfo()), AfterLegalize(false) {}
271 /// Run - runs the dag combiner on all nodes in the work list
272 void Run(bool RunningAfterLegalize);
276 //===----------------------------------------------------------------------===//
277 // TargetLowering::DAGCombinerInfo implementation
278 //===----------------------------------------------------------------------===//
280 void TargetLowering::DAGCombinerInfo::AddToWorklist(SDNode *N) {
281 ((DAGCombiner*)DC)->AddToWorkList(N);
284 SDOperand TargetLowering::DAGCombinerInfo::
285 CombineTo(SDNode *N, const std::vector<SDOperand> &To) {
286 return ((DAGCombiner*)DC)->CombineTo(N, &To[0], To.size());
289 SDOperand TargetLowering::DAGCombinerInfo::
290 CombineTo(SDNode *N, SDOperand Res) {
291 return ((DAGCombiner*)DC)->CombineTo(N, Res);
295 SDOperand TargetLowering::DAGCombinerInfo::
296 CombineTo(SDNode *N, SDOperand Res0, SDOperand Res1) {
297 return ((DAGCombiner*)DC)->CombineTo(N, Res0, Res1);
303 //===----------------------------------------------------------------------===//
306 // isSetCCEquivalent - Return true if this node is a setcc, or is a select_cc
307 // that selects between the values 1 and 0, making it equivalent to a setcc.
308 // Also, set the incoming LHS, RHS, and CC references to the appropriate
309 // nodes based on the type of node we are checking. This simplifies life a
310 // bit for the callers.
311 static bool isSetCCEquivalent(SDOperand N, SDOperand &LHS, SDOperand &RHS,
313 if (N.getOpcode() == ISD::SETCC) {
314 LHS = N.getOperand(0);
315 RHS = N.getOperand(1);
316 CC = N.getOperand(2);
319 if (N.getOpcode() == ISD::SELECT_CC &&
320 N.getOperand(2).getOpcode() == ISD::Constant &&
321 N.getOperand(3).getOpcode() == ISD::Constant &&
322 cast<ConstantSDNode>(N.getOperand(2))->getValue() == 1 &&
323 cast<ConstantSDNode>(N.getOperand(3))->isNullValue()) {
324 LHS = N.getOperand(0);
325 RHS = N.getOperand(1);
326 CC = N.getOperand(4);
332 // isOneUseSetCC - Return true if this is a SetCC-equivalent operation with only
333 // one use. If this is true, it allows the users to invert the operation for
334 // free when it is profitable to do so.
335 static bool isOneUseSetCC(SDOperand N) {
336 SDOperand N0, N1, N2;
337 if (isSetCCEquivalent(N, N0, N1, N2) && N.Val->hasOneUse())
342 SDOperand DAGCombiner::ReassociateOps(unsigned Opc, SDOperand N0, SDOperand N1){
343 MVT::ValueType VT = N0.getValueType();
344 // reassoc. (op (op x, c1), y) -> (op (op x, y), c1) iff x+c1 has one use
345 // reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2))
346 if (N0.getOpcode() == Opc && isa<ConstantSDNode>(N0.getOperand(1))) {
347 if (isa<ConstantSDNode>(N1)) {
348 SDOperand OpNode = DAG.getNode(Opc, VT, N0.getOperand(1), N1);
349 AddToWorkList(OpNode.Val);
350 return DAG.getNode(Opc, VT, OpNode, N0.getOperand(0));
351 } else if (N0.hasOneUse()) {
352 SDOperand OpNode = DAG.getNode(Opc, VT, N0.getOperand(0), N1);
353 AddToWorkList(OpNode.Val);
354 return DAG.getNode(Opc, VT, OpNode, N0.getOperand(1));
357 // reassoc. (op y, (op x, c1)) -> (op (op x, y), c1) iff x+c1 has one use
358 // reassoc. (op c2, (op x, c1)) -> (op x, (op c1, c2))
359 if (N1.getOpcode() == Opc && isa<ConstantSDNode>(N1.getOperand(1))) {
360 if (isa<ConstantSDNode>(N0)) {
361 SDOperand OpNode = DAG.getNode(Opc, VT, N1.getOperand(1), N0);
362 AddToWorkList(OpNode.Val);
363 return DAG.getNode(Opc, VT, OpNode, N1.getOperand(0));
364 } else if (N1.hasOneUse()) {
365 SDOperand OpNode = DAG.getNode(Opc, VT, N1.getOperand(0), N0);
366 AddToWorkList(OpNode.Val);
367 return DAG.getNode(Opc, VT, OpNode, N1.getOperand(1));
373 void DAGCombiner::Run(bool RunningAfterLegalize) {
374 // set the instance variable, so that the various visit routines may use it.
375 AfterLegalize = RunningAfterLegalize;
377 // Add all the dag nodes to the worklist.
378 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
379 E = DAG.allnodes_end(); I != E; ++I)
380 WorkList.push_back(I);
382 // Create a dummy node (which is not added to allnodes), that adds a reference
383 // to the root node, preventing it from being deleted, and tracking any
384 // changes of the root.
385 HandleSDNode Dummy(DAG.getRoot());
388 /// DagCombineInfo - Expose the DAG combiner to the target combiner impls.
389 TargetLowering::DAGCombinerInfo
390 DagCombineInfo(DAG, !RunningAfterLegalize, this);
392 // while the worklist isn't empty, inspect the node on the end of it and
393 // try and combine it.
394 while (!WorkList.empty()) {
395 SDNode *N = WorkList.back();
398 // If N has no uses, it is dead. Make sure to revisit all N's operands once
399 // N is deleted from the DAG, since they too may now be dead or may have a
400 // reduced number of uses, allowing other xforms.
401 if (N->use_empty() && N != &Dummy) {
402 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
403 WorkList.push_back(N->getOperand(i).Val);
405 removeFromWorkList(N);
410 SDOperand RV = visit(N);
412 // If nothing happened, try a target-specific DAG combine.
414 assert(N->getOpcode() != ISD::DELETED_NODE &&
415 "Node was deleted but visit returned NULL!");
416 if (N->getOpcode() >= ISD::BUILTIN_OP_END ||
417 TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode()))
418 RV = TLI.PerformDAGCombine(N, DagCombineInfo);
423 // If we get back the same node we passed in, rather than a new node or
424 // zero, we know that the node must have defined multiple values and
425 // CombineTo was used. Since CombineTo takes care of the worklist
426 // mechanics for us, we have no work to do in this case.
428 assert(N->getOpcode() != ISD::DELETED_NODE &&
429 RV.Val->getOpcode() != ISD::DELETED_NODE &&
430 "Node was deleted but visit returned new node!");
432 DEBUG(std::cerr << "\nReplacing "; N->dump();
433 std::cerr << "\nWith: "; RV.Val->dump(&DAG);
435 std::vector<SDNode*> NowDead;
436 if (N->getNumValues() == RV.Val->getNumValues())
437 DAG.ReplaceAllUsesWith(N, RV.Val, &NowDead);
439 assert(N->getValueType(0) == RV.getValueType() && "Type mismatch");
441 DAG.ReplaceAllUsesWith(N, &OpV, &NowDead);
444 // Push the new node and any users onto the worklist
445 WorkList.push_back(RV.Val);
446 AddUsersToWorkList(RV.Val);
448 // Nodes can end up on the worklist more than once. Make sure we do
449 // not process a node that has been replaced.
450 removeFromWorkList(N);
451 for (unsigned i = 0, e = NowDead.size(); i != e; ++i)
452 removeFromWorkList(NowDead[i]);
454 // Finally, since the node is now dead, remove it from the graph.
460 // If the root changed (e.g. it was a dead load, update the root).
461 DAG.setRoot(Dummy.getValue());
464 SDOperand DAGCombiner::visit(SDNode *N) {
465 switch(N->getOpcode()) {
467 case ISD::TokenFactor: return visitTokenFactor(N);
468 case ISD::ADD: return visitADD(N);
469 case ISD::SUB: return visitSUB(N);
470 case ISD::MUL: return visitMUL(N);
471 case ISD::SDIV: return visitSDIV(N);
472 case ISD::UDIV: return visitUDIV(N);
473 case ISD::SREM: return visitSREM(N);
474 case ISD::UREM: return visitUREM(N);
475 case ISD::MULHU: return visitMULHU(N);
476 case ISD::MULHS: return visitMULHS(N);
477 case ISD::AND: return visitAND(N);
478 case ISD::OR: return visitOR(N);
479 case ISD::XOR: return visitXOR(N);
480 case ISD::SHL: return visitSHL(N);
481 case ISD::SRA: return visitSRA(N);
482 case ISD::SRL: return visitSRL(N);
483 case ISD::CTLZ: return visitCTLZ(N);
484 case ISD::CTTZ: return visitCTTZ(N);
485 case ISD::CTPOP: return visitCTPOP(N);
486 case ISD::SELECT: return visitSELECT(N);
487 case ISD::SELECT_CC: return visitSELECT_CC(N);
488 case ISD::SETCC: return visitSETCC(N);
489 case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N);
490 case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N);
491 case ISD::ANY_EXTEND: return visitANY_EXTEND(N);
492 case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N);
493 case ISD::TRUNCATE: return visitTRUNCATE(N);
494 case ISD::BIT_CONVERT: return visitBIT_CONVERT(N);
495 case ISD::VBIT_CONVERT: return visitVBIT_CONVERT(N);
496 case ISD::FADD: return visitFADD(N);
497 case ISD::FSUB: return visitFSUB(N);
498 case ISD::FMUL: return visitFMUL(N);
499 case ISD::FDIV: return visitFDIV(N);
500 case ISD::FREM: return visitFREM(N);
501 case ISD::FCOPYSIGN: return visitFCOPYSIGN(N);
502 case ISD::SINT_TO_FP: return visitSINT_TO_FP(N);
503 case ISD::UINT_TO_FP: return visitUINT_TO_FP(N);
504 case ISD::FP_TO_SINT: return visitFP_TO_SINT(N);
505 case ISD::FP_TO_UINT: return visitFP_TO_UINT(N);
506 case ISD::FP_ROUND: return visitFP_ROUND(N);
507 case ISD::FP_ROUND_INREG: return visitFP_ROUND_INREG(N);
508 case ISD::FP_EXTEND: return visitFP_EXTEND(N);
509 case ISD::FNEG: return visitFNEG(N);
510 case ISD::FABS: return visitFABS(N);
511 case ISD::BRCOND: return visitBRCOND(N);
512 case ISD::BR_CC: return visitBR_CC(N);
513 case ISD::LOAD: return visitLOAD(N);
516 case ISD::ZEXTLOAD: return visitXEXTLOAD(N);
517 case ISD::STORE: return visitSTORE(N);
518 case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N);
519 case ISD::VINSERT_VECTOR_ELT: return visitVINSERT_VECTOR_ELT(N);
520 case ISD::VBUILD_VECTOR: return visitVBUILD_VECTOR(N);
521 case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N);
522 case ISD::VVECTOR_SHUFFLE: return visitVVECTOR_SHUFFLE(N);
523 case ISD::VADD: return visitVBinOp(N, ISD::ADD , ISD::FADD);
524 case ISD::VSUB: return visitVBinOp(N, ISD::SUB , ISD::FSUB);
525 case ISD::VMUL: return visitVBinOp(N, ISD::MUL , ISD::FMUL);
526 case ISD::VSDIV: return visitVBinOp(N, ISD::SDIV, ISD::FDIV);
527 case ISD::VUDIV: return visitVBinOp(N, ISD::UDIV, ISD::UDIV);
528 case ISD::VAND: return visitVBinOp(N, ISD::AND , ISD::AND);
529 case ISD::VOR: return visitVBinOp(N, ISD::OR , ISD::OR);
530 case ISD::VXOR: return visitVBinOp(N, ISD::XOR , ISD::XOR);
535 SDOperand DAGCombiner::visitTokenFactor(SDNode *N) {
536 // If the token factor has two operands and one is the entry token, replace
537 // the token factor with the other operand.
538 if (N->getNumOperands() == 2) {
539 if (N->getOperand(0).getOpcode() == ISD::EntryToken ||
540 N->getOperand(0) == N->getOperand(1))
541 return N->getOperand(1);
542 if (N->getOperand(1).getOpcode() == ISD::EntryToken)
543 return N->getOperand(0);
546 SmallVector<SDNode *, 8> TFs; // Set of token factor nodes.
547 SmallVector<SDOperand, 8> Ops; // Ops for replacing token factor.
549 // Add this ndoe to the token factor set.
552 // Separate token factors from other operands.
553 for (unsigned i = 0, ie = N->getNumOperands(); i != ie; ++i) {
554 SDOperand Op = N->getOperand(i);
555 if (Op.getOpcode() == ISD::TokenFactor)
556 TFs.push_back(Op.Val);
557 else if (Op.getOpcode() != ISD::EntryToken)
561 // If there are token factor operands.
562 if (TFs.size() > 1) {
563 bool Changed = false; // If we should replace this token factor.
565 // For each token factor.
566 for (unsigned j = 1, je = TFs.size(); j != je; ++j) {
568 bool CanMerge = true; // Can we merge this token factor.
571 if (!TF->hasOneUse()) {
572 // Check to see if all users point to members of the token factor set.
573 for (SDNode::use_iterator UI = TF->use_begin(), UE = TF->use_end();
574 CanMerge && UI != UE; ++UI) {
576 CanMerge = User->getOpcode() == ISD::TokenFactor &&
577 std::find(TFs.begin(), TFs.end(), User) != TFs.end();
581 CanMerge = TF->hasOneUse();
584 // If it's valid to merge.
586 // Remove dead token factor node.
589 // Make sure we don't duplicate operands.
590 unsigned m = Ops.size(); // Number of prior operands.
591 for (unsigned l = 0, le = TF->getNumOperands(); l != le; ++l) {
592 SDOperand Op = TF->getOperand(l);
593 if (std::find(Ops.begin(), Ops.end(), Op) == Ops.end())
598 // Can't merge this token factor.
599 Ops.push_back(SDOperand(TF, 0));
603 // If we've change things around then replace token factor.
605 return DAG.getNode(ISD::TokenFactor, MVT::Other, &Ops[0], Ops.size());
612 SDOperand DAGCombiner::visitADD(SDNode *N) {
613 SDOperand N0 = N->getOperand(0);
614 SDOperand N1 = N->getOperand(1);
615 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
616 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
617 MVT::ValueType VT = N0.getValueType();
619 // fold (add c1, c2) -> c1+c2
621 return DAG.getNode(ISD::ADD, VT, N0, N1);
622 // canonicalize constant to RHS
624 return DAG.getNode(ISD::ADD, VT, N1, N0);
625 // fold (add x, 0) -> x
626 if (N1C && N1C->isNullValue())
628 // fold ((c1-A)+c2) -> (c1+c2)-A
629 if (N1C && N0.getOpcode() == ISD::SUB)
630 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getOperand(0)))
631 return DAG.getNode(ISD::SUB, VT,
632 DAG.getConstant(N1C->getValue()+N0C->getValue(), VT),
635 SDOperand RADD = ReassociateOps(ISD::ADD, N0, N1);
638 // fold ((0-A) + B) -> B-A
639 if (N0.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N0.getOperand(0)) &&
640 cast<ConstantSDNode>(N0.getOperand(0))->isNullValue())
641 return DAG.getNode(ISD::SUB, VT, N1, N0.getOperand(1));
642 // fold (A + (0-B)) -> A-B
643 if (N1.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N1.getOperand(0)) &&
644 cast<ConstantSDNode>(N1.getOperand(0))->isNullValue())
645 return DAG.getNode(ISD::SUB, VT, N0, N1.getOperand(1));
646 // fold (A+(B-A)) -> B
647 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1))
648 return N1.getOperand(0);
650 if (!MVT::isVector(VT) && SimplifyDemandedBits(SDOperand(N, 0)))
651 return SDOperand(N, 0);
653 // fold (a+b) -> (a|b) iff a and b share no bits.
654 if (MVT::isInteger(VT) && !MVT::isVector(VT)) {
655 uint64_t LHSZero, LHSOne;
656 uint64_t RHSZero, RHSOne;
657 uint64_t Mask = MVT::getIntVTBitMask(VT);
658 TLI.ComputeMaskedBits(N0, Mask, LHSZero, LHSOne);
660 TLI.ComputeMaskedBits(N1, Mask, RHSZero, RHSOne);
662 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
663 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
664 if ((RHSZero & (~LHSZero & Mask)) == (~LHSZero & Mask) ||
665 (LHSZero & (~RHSZero & Mask)) == (~RHSZero & Mask))
666 return DAG.getNode(ISD::OR, VT, N0, N1);
673 SDOperand DAGCombiner::visitSUB(SDNode *N) {
674 SDOperand N0 = N->getOperand(0);
675 SDOperand N1 = N->getOperand(1);
676 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val);
677 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val);
678 MVT::ValueType VT = N0.getValueType();
680 // fold (sub x, x) -> 0
682 return DAG.getConstant(0, N->getValueType(0));
683 // fold (sub c1, c2) -> c1-c2
685 return DAG.getNode(ISD::SUB, VT, N0, N1);
686 // fold (sub x, c) -> (add x, -c)
688 return DAG.getNode(ISD::ADD, VT, N0, DAG.getConstant(-N1C->getValue(), VT));
690 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1)
691 return N0.getOperand(1);
693 if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1)
694 return N0.getOperand(0);
698 SDOperand DAGCombiner::visitMUL(SDNode *N) {
699 SDOperand N0 = N->getOperand(0);
700 SDOperand N1 = N->getOperand(1);
701 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
702 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
703 MVT::ValueType VT = N0.getValueType();
705 // fold (mul c1, c2) -> c1*c2
707 return DAG.getNode(ISD::MUL, VT, N0, N1);
708 // canonicalize constant to RHS
710 return DAG.getNode(ISD::MUL, VT, N1, N0);
711 // fold (mul x, 0) -> 0
712 if (N1C && N1C->isNullValue())
714 // fold (mul x, -1) -> 0-x
715 if (N1C && N1C->isAllOnesValue())
716 return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT), N0);
717 // fold (mul x, (1 << c)) -> x << c
718 if (N1C && isPowerOf2_64(N1C->getValue()))
719 return DAG.getNode(ISD::SHL, VT, N0,
720 DAG.getConstant(Log2_64(N1C->getValue()),
721 TLI.getShiftAmountTy()));
722 // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c
723 if (N1C && isPowerOf2_64(-N1C->getSignExtended())) {
724 // FIXME: If the input is something that is easily negated (e.g. a
725 // single-use add), we should put the negate there.
726 return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT),
727 DAG.getNode(ISD::SHL, VT, N0,
728 DAG.getConstant(Log2_64(-N1C->getSignExtended()),
729 TLI.getShiftAmountTy())));
732 // (mul (shl X, c1), c2) -> (mul X, c2 << c1)
733 if (N1C && N0.getOpcode() == ISD::SHL &&
734 isa<ConstantSDNode>(N0.getOperand(1))) {
735 SDOperand C3 = DAG.getNode(ISD::SHL, VT, N1, N0.getOperand(1));
736 AddToWorkList(C3.Val);
737 return DAG.getNode(ISD::MUL, VT, N0.getOperand(0), C3);
740 // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one
743 SDOperand Sh(0,0), Y(0,0);
744 // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)).
745 if (N0.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N0.getOperand(1)) &&
746 N0.Val->hasOneUse()) {
748 } else if (N1.getOpcode() == ISD::SHL &&
749 isa<ConstantSDNode>(N1.getOperand(1)) && N1.Val->hasOneUse()) {
753 SDOperand Mul = DAG.getNode(ISD::MUL, VT, Sh.getOperand(0), Y);
754 return DAG.getNode(ISD::SHL, VT, Mul, Sh.getOperand(1));
757 // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2)
758 if (N1C && N0.getOpcode() == ISD::ADD && N0.Val->hasOneUse() &&
759 isa<ConstantSDNode>(N0.getOperand(1))) {
760 return DAG.getNode(ISD::ADD, VT,
761 DAG.getNode(ISD::MUL, VT, N0.getOperand(0), N1),
762 DAG.getNode(ISD::MUL, VT, N0.getOperand(1), N1));
766 SDOperand RMUL = ReassociateOps(ISD::MUL, N0, N1);
772 SDOperand DAGCombiner::visitSDIV(SDNode *N) {
773 SDOperand N0 = N->getOperand(0);
774 SDOperand N1 = N->getOperand(1);
775 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val);
776 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val);
777 MVT::ValueType VT = N->getValueType(0);
779 // fold (sdiv c1, c2) -> c1/c2
780 if (N0C && N1C && !N1C->isNullValue())
781 return DAG.getNode(ISD::SDIV, VT, N0, N1);
782 // fold (sdiv X, 1) -> X
783 if (N1C && N1C->getSignExtended() == 1LL)
785 // fold (sdiv X, -1) -> 0-X
786 if (N1C && N1C->isAllOnesValue())
787 return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT), N0);
788 // If we know the sign bits of both operands are zero, strength reduce to a
789 // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2
790 uint64_t SignBit = 1ULL << (MVT::getSizeInBits(VT)-1);
791 if (TLI.MaskedValueIsZero(N1, SignBit) &&
792 TLI.MaskedValueIsZero(N0, SignBit))
793 return DAG.getNode(ISD::UDIV, N1.getValueType(), N0, N1);
794 // fold (sdiv X, pow2) -> simple ops after legalize
795 if (N1C && N1C->getValue() && !TLI.isIntDivCheap() &&
796 (isPowerOf2_64(N1C->getSignExtended()) ||
797 isPowerOf2_64(-N1C->getSignExtended()))) {
798 // If dividing by powers of two is cheap, then don't perform the following
800 if (TLI.isPow2DivCheap())
802 int64_t pow2 = N1C->getSignExtended();
803 int64_t abs2 = pow2 > 0 ? pow2 : -pow2;
804 unsigned lg2 = Log2_64(abs2);
805 // Splat the sign bit into the register
806 SDOperand SGN = DAG.getNode(ISD::SRA, VT, N0,
807 DAG.getConstant(MVT::getSizeInBits(VT)-1,
808 TLI.getShiftAmountTy()));
809 AddToWorkList(SGN.Val);
810 // Add (N0 < 0) ? abs2 - 1 : 0;
811 SDOperand SRL = DAG.getNode(ISD::SRL, VT, SGN,
812 DAG.getConstant(MVT::getSizeInBits(VT)-lg2,
813 TLI.getShiftAmountTy()));
814 SDOperand ADD = DAG.getNode(ISD::ADD, VT, N0, SRL);
815 AddToWorkList(SRL.Val);
816 AddToWorkList(ADD.Val); // Divide by pow2
817 SDOperand SRA = DAG.getNode(ISD::SRA, VT, ADD,
818 DAG.getConstant(lg2, TLI.getShiftAmountTy()));
819 // If we're dividing by a positive value, we're done. Otherwise, we must
820 // negate the result.
823 AddToWorkList(SRA.Val);
824 return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT), SRA);
826 // if integer divide is expensive and we satisfy the requirements, emit an
827 // alternate sequence.
828 if (N1C && (N1C->getSignExtended() < -1 || N1C->getSignExtended() > 1) &&
829 !TLI.isIntDivCheap()) {
830 SDOperand Op = BuildSDIV(N);
831 if (Op.Val) return Op;
836 SDOperand DAGCombiner::visitUDIV(SDNode *N) {
837 SDOperand N0 = N->getOperand(0);
838 SDOperand N1 = N->getOperand(1);
839 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val);
840 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val);
841 MVT::ValueType VT = N->getValueType(0);
843 // fold (udiv c1, c2) -> c1/c2
844 if (N0C && N1C && !N1C->isNullValue())
845 return DAG.getNode(ISD::UDIV, VT, N0, N1);
846 // fold (udiv x, (1 << c)) -> x >>u c
847 if (N1C && isPowerOf2_64(N1C->getValue()))
848 return DAG.getNode(ISD::SRL, VT, N0,
849 DAG.getConstant(Log2_64(N1C->getValue()),
850 TLI.getShiftAmountTy()));
851 // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2
852 if (N1.getOpcode() == ISD::SHL) {
853 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
854 if (isPowerOf2_64(SHC->getValue())) {
855 MVT::ValueType ADDVT = N1.getOperand(1).getValueType();
856 SDOperand Add = DAG.getNode(ISD::ADD, ADDVT, N1.getOperand(1),
857 DAG.getConstant(Log2_64(SHC->getValue()),
859 AddToWorkList(Add.Val);
860 return DAG.getNode(ISD::SRL, VT, N0, Add);
864 // fold (udiv x, c) -> alternate
865 if (N1C && N1C->getValue() && !TLI.isIntDivCheap()) {
866 SDOperand Op = BuildUDIV(N);
867 if (Op.Val) return Op;
872 SDOperand DAGCombiner::visitSREM(SDNode *N) {
873 SDOperand N0 = N->getOperand(0);
874 SDOperand N1 = N->getOperand(1);
875 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
876 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
877 MVT::ValueType VT = N->getValueType(0);
879 // fold (srem c1, c2) -> c1%c2
880 if (N0C && N1C && !N1C->isNullValue())
881 return DAG.getNode(ISD::SREM, VT, N0, N1);
882 // If we know the sign bits of both operands are zero, strength reduce to a
883 // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15
884 uint64_t SignBit = 1ULL << (MVT::getSizeInBits(VT)-1);
885 if (TLI.MaskedValueIsZero(N1, SignBit) &&
886 TLI.MaskedValueIsZero(N0, SignBit))
887 return DAG.getNode(ISD::UREM, VT, N0, N1);
891 SDOperand DAGCombiner::visitUREM(SDNode *N) {
892 SDOperand N0 = N->getOperand(0);
893 SDOperand N1 = N->getOperand(1);
894 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
895 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
896 MVT::ValueType VT = N->getValueType(0);
898 // fold (urem c1, c2) -> c1%c2
899 if (N0C && N1C && !N1C->isNullValue())
900 return DAG.getNode(ISD::UREM, VT, N0, N1);
901 // fold (urem x, pow2) -> (and x, pow2-1)
902 if (N1C && !N1C->isNullValue() && isPowerOf2_64(N1C->getValue()))
903 return DAG.getNode(ISD::AND, VT, N0, DAG.getConstant(N1C->getValue()-1,VT));
904 // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1))
905 if (N1.getOpcode() == ISD::SHL) {
906 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
907 if (isPowerOf2_64(SHC->getValue())) {
908 SDOperand Add = DAG.getNode(ISD::ADD, VT, N1,DAG.getConstant(~0ULL,VT));
909 AddToWorkList(Add.Val);
910 return DAG.getNode(ISD::AND, VT, N0, Add);
917 SDOperand DAGCombiner::visitMULHS(SDNode *N) {
918 SDOperand N0 = N->getOperand(0);
919 SDOperand N1 = N->getOperand(1);
920 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
922 // fold (mulhs x, 0) -> 0
923 if (N1C && N1C->isNullValue())
925 // fold (mulhs x, 1) -> (sra x, size(x)-1)
926 if (N1C && N1C->getValue() == 1)
927 return DAG.getNode(ISD::SRA, N0.getValueType(), N0,
928 DAG.getConstant(MVT::getSizeInBits(N0.getValueType())-1,
929 TLI.getShiftAmountTy()));
933 SDOperand DAGCombiner::visitMULHU(SDNode *N) {
934 SDOperand N0 = N->getOperand(0);
935 SDOperand N1 = N->getOperand(1);
936 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
938 // fold (mulhu x, 0) -> 0
939 if (N1C && N1C->isNullValue())
941 // fold (mulhu x, 1) -> 0
942 if (N1C && N1C->getValue() == 1)
943 return DAG.getConstant(0, N0.getValueType());
947 /// SimplifyBinOpWithSameOpcodeHands - If this is a binary operator with
948 /// two operands of the same opcode, try to simplify it.
949 SDOperand DAGCombiner::SimplifyBinOpWithSameOpcodeHands(SDNode *N) {
950 SDOperand N0 = N->getOperand(0), N1 = N->getOperand(1);
951 MVT::ValueType VT = N0.getValueType();
952 assert(N0.getOpcode() == N1.getOpcode() && "Bad input!");
954 // For each of OP in AND/OR/XOR:
955 // fold (OP (zext x), (zext y)) -> (zext (OP x, y))
956 // fold (OP (sext x), (sext y)) -> (sext (OP x, y))
957 // fold (OP (aext x), (aext y)) -> (aext (OP x, y))
958 // fold (OP (trunc x), (trunc y)) -> (trunc (OP x, y))
959 if ((N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND||
960 N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::TRUNCATE) &&
961 N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()) {
962 SDOperand ORNode = DAG.getNode(N->getOpcode(),
963 N0.getOperand(0).getValueType(),
964 N0.getOperand(0), N1.getOperand(0));
965 AddToWorkList(ORNode.Val);
966 return DAG.getNode(N0.getOpcode(), VT, ORNode);
969 // For each of OP in SHL/SRL/SRA/AND...
970 // fold (and (OP x, z), (OP y, z)) -> (OP (and x, y), z)
971 // fold (or (OP x, z), (OP y, z)) -> (OP (or x, y), z)
972 // fold (xor (OP x, z), (OP y, z)) -> (OP (xor x, y), z)
973 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL ||
974 N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::AND) &&
975 N0.getOperand(1) == N1.getOperand(1)) {
976 SDOperand ORNode = DAG.getNode(N->getOpcode(),
977 N0.getOperand(0).getValueType(),
978 N0.getOperand(0), N1.getOperand(0));
979 AddToWorkList(ORNode.Val);
980 return DAG.getNode(N0.getOpcode(), VT, ORNode, N0.getOperand(1));
986 SDOperand DAGCombiner::visitAND(SDNode *N) {
987 SDOperand N0 = N->getOperand(0);
988 SDOperand N1 = N->getOperand(1);
989 SDOperand LL, LR, RL, RR, CC0, CC1;
990 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
991 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
992 MVT::ValueType VT = N1.getValueType();
993 unsigned OpSizeInBits = MVT::getSizeInBits(VT);
995 // fold (and c1, c2) -> c1&c2
997 return DAG.getNode(ISD::AND, VT, N0, N1);
998 // canonicalize constant to RHS
1000 return DAG.getNode(ISD::AND, VT, N1, N0);
1001 // fold (and x, -1) -> x
1002 if (N1C && N1C->isAllOnesValue())
1004 // if (and x, c) is known to be zero, return 0
1005 if (N1C && TLI.MaskedValueIsZero(SDOperand(N, 0), MVT::getIntVTBitMask(VT)))
1006 return DAG.getConstant(0, VT);
1008 SDOperand RAND = ReassociateOps(ISD::AND, N0, N1);
1011 // fold (and (or x, 0xFFFF), 0xFF) -> 0xFF
1012 if (N1C && N0.getOpcode() == ISD::OR)
1013 if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
1014 if ((ORI->getValue() & N1C->getValue()) == N1C->getValue())
1016 // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits.
1017 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
1018 unsigned InMask = MVT::getIntVTBitMask(N0.getOperand(0).getValueType());
1019 if (TLI.MaskedValueIsZero(N0.getOperand(0),
1020 ~N1C->getValue() & InMask)) {
1021 SDOperand Zext = DAG.getNode(ISD::ZERO_EXTEND, N0.getValueType(),
1024 // Replace uses of the AND with uses of the Zero extend node.
1027 // We actually want to replace all uses of the any_extend with the
1028 // zero_extend, to avoid duplicating things. This will later cause this
1029 // AND to be folded.
1030 CombineTo(N0.Val, Zext);
1031 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
1034 // fold (and (setcc x), (setcc y)) -> (setcc (and x, y))
1035 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
1036 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
1037 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
1039 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
1040 MVT::isInteger(LL.getValueType())) {
1041 // fold (X == 0) & (Y == 0) -> (X|Y == 0)
1042 if (cast<ConstantSDNode>(LR)->getValue() == 0 && Op1 == ISD::SETEQ) {
1043 SDOperand ORNode = DAG.getNode(ISD::OR, LR.getValueType(), LL, RL);
1044 AddToWorkList(ORNode.Val);
1045 return DAG.getSetCC(VT, ORNode, LR, Op1);
1047 // fold (X == -1) & (Y == -1) -> (X&Y == -1)
1048 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETEQ) {
1049 SDOperand ANDNode = DAG.getNode(ISD::AND, LR.getValueType(), LL, RL);
1050 AddToWorkList(ANDNode.Val);
1051 return DAG.getSetCC(VT, ANDNode, LR, Op1);
1053 // fold (X > -1) & (Y > -1) -> (X|Y > -1)
1054 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETGT) {
1055 SDOperand ORNode = DAG.getNode(ISD::OR, LR.getValueType(), LL, RL);
1056 AddToWorkList(ORNode.Val);
1057 return DAG.getSetCC(VT, ORNode, LR, Op1);
1060 // canonicalize equivalent to ll == rl
1061 if (LL == RR && LR == RL) {
1062 Op1 = ISD::getSetCCSwappedOperands(Op1);
1065 if (LL == RL && LR == RR) {
1066 bool isInteger = MVT::isInteger(LL.getValueType());
1067 ISD::CondCode Result = ISD::getSetCCAndOperation(Op0, Op1, isInteger);
1068 if (Result != ISD::SETCC_INVALID)
1069 return DAG.getSetCC(N0.getValueType(), LL, LR, Result);
1073 // Simplify: and (op x...), (op y...) -> (op (and x, y))
1074 if (N0.getOpcode() == N1.getOpcode()) {
1075 SDOperand Tmp = SimplifyBinOpWithSameOpcodeHands(N);
1076 if (Tmp.Val) return Tmp;
1079 // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1)
1080 // fold (and (sra)) -> (and (srl)) when possible.
1081 if (!MVT::isVector(VT) &&
1082 SimplifyDemandedBits(SDOperand(N, 0)))
1083 return SDOperand(N, 0);
1084 // fold (zext_inreg (extload x)) -> (zextload x)
1085 if (N0.getOpcode() == ISD::EXTLOAD) {
1086 MVT::ValueType EVT = cast<VTSDNode>(N0.getOperand(3))->getVT();
1087 // If we zero all the possible extended bits, then we can turn this into
1088 // a zextload if we are running before legalize or the operation is legal.
1089 if (TLI.MaskedValueIsZero(N1, ~0ULL << MVT::getSizeInBits(EVT)) &&
1090 (!AfterLegalize || TLI.isOperationLegal(ISD::ZEXTLOAD, EVT))) {
1091 SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, N0.getOperand(0),
1092 N0.getOperand(1), N0.getOperand(2),
1095 CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1));
1096 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
1099 // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use
1100 if (N0.getOpcode() == ISD::SEXTLOAD && N0.hasOneUse()) {
1101 MVT::ValueType EVT = cast<VTSDNode>(N0.getOperand(3))->getVT();
1102 // If we zero all the possible extended bits, then we can turn this into
1103 // a zextload if we are running before legalize or the operation is legal.
1104 if (TLI.MaskedValueIsZero(N1, ~0ULL << MVT::getSizeInBits(EVT)) &&
1105 (!AfterLegalize || TLI.isOperationLegal(ISD::ZEXTLOAD, EVT))) {
1106 SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, N0.getOperand(0),
1107 N0.getOperand(1), N0.getOperand(2),
1110 CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1));
1111 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
1115 // fold (and (load x), 255) -> (zextload x, i8)
1116 // fold (and (extload x, i16), 255) -> (zextload x, i8)
1118 (N0.getOpcode() == ISD::LOAD || N0.getOpcode() == ISD::EXTLOAD ||
1119 N0.getOpcode() == ISD::ZEXTLOAD) &&
1121 MVT::ValueType EVT, LoadedVT;
1122 if (N1C->getValue() == 255)
1124 else if (N1C->getValue() == 65535)
1126 else if (N1C->getValue() == ~0U)
1131 LoadedVT = N0.getOpcode() == ISD::LOAD ? VT :
1132 cast<VTSDNode>(N0.getOperand(3))->getVT();
1133 if (EVT != MVT::Other && LoadedVT > EVT &&
1134 (!AfterLegalize || TLI.isOperationLegal(ISD::ZEXTLOAD, EVT))) {
1135 MVT::ValueType PtrType = N0.getOperand(1).getValueType();
1136 // For big endian targets, we need to add an offset to the pointer to load
1137 // the correct bytes. For little endian systems, we merely need to read
1138 // fewer bytes from the same pointer.
1140 (MVT::getSizeInBits(LoadedVT) - MVT::getSizeInBits(EVT)) / 8;
1141 SDOperand NewPtr = N0.getOperand(1);
1142 if (!TLI.isLittleEndian())
1143 NewPtr = DAG.getNode(ISD::ADD, PtrType, NewPtr,
1144 DAG.getConstant(PtrOff, PtrType));
1145 AddToWorkList(NewPtr.Val);
1147 DAG.getExtLoad(ISD::ZEXTLOAD, VT, N0.getOperand(0), NewPtr,
1148 N0.getOperand(2), EVT);
1150 CombineTo(N0.Val, Load, Load.getValue(1));
1151 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
1158 SDOperand DAGCombiner::visitOR(SDNode *N) {
1159 SDOperand N0 = N->getOperand(0);
1160 SDOperand N1 = N->getOperand(1);
1161 SDOperand LL, LR, RL, RR, CC0, CC1;
1162 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1163 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1164 MVT::ValueType VT = N1.getValueType();
1165 unsigned OpSizeInBits = MVT::getSizeInBits(VT);
1167 // fold (or c1, c2) -> c1|c2
1169 return DAG.getNode(ISD::OR, VT, N0, N1);
1170 // canonicalize constant to RHS
1172 return DAG.getNode(ISD::OR, VT, N1, N0);
1173 // fold (or x, 0) -> x
1174 if (N1C && N1C->isNullValue())
1176 // fold (or x, -1) -> -1
1177 if (N1C && N1C->isAllOnesValue())
1179 // fold (or x, c) -> c iff (x & ~c) == 0
1181 TLI.MaskedValueIsZero(N0,~N1C->getValue() & (~0ULL>>(64-OpSizeInBits))))
1184 SDOperand ROR = ReassociateOps(ISD::OR, N0, N1);
1187 // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2)
1188 if (N1C && N0.getOpcode() == ISD::AND && N0.Val->hasOneUse() &&
1189 isa<ConstantSDNode>(N0.getOperand(1))) {
1190 ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1));
1191 return DAG.getNode(ISD::AND, VT, DAG.getNode(ISD::OR, VT, N0.getOperand(0),
1193 DAG.getConstant(N1C->getValue() | C1->getValue(), VT));
1195 // fold (or (setcc x), (setcc y)) -> (setcc (or x, y))
1196 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
1197 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
1198 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
1200 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
1201 MVT::isInteger(LL.getValueType())) {
1202 // fold (X != 0) | (Y != 0) -> (X|Y != 0)
1203 // fold (X < 0) | (Y < 0) -> (X|Y < 0)
1204 if (cast<ConstantSDNode>(LR)->getValue() == 0 &&
1205 (Op1 == ISD::SETNE || Op1 == ISD::SETLT)) {
1206 SDOperand ORNode = DAG.getNode(ISD::OR, LR.getValueType(), LL, RL);
1207 AddToWorkList(ORNode.Val);
1208 return DAG.getSetCC(VT, ORNode, LR, Op1);
1210 // fold (X != -1) | (Y != -1) -> (X&Y != -1)
1211 // fold (X > -1) | (Y > -1) -> (X&Y > -1)
1212 if (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
1213 (Op1 == ISD::SETNE || Op1 == ISD::SETGT)) {
1214 SDOperand ANDNode = DAG.getNode(ISD::AND, LR.getValueType(), LL, RL);
1215 AddToWorkList(ANDNode.Val);
1216 return DAG.getSetCC(VT, ANDNode, LR, Op1);
1219 // canonicalize equivalent to ll == rl
1220 if (LL == RR && LR == RL) {
1221 Op1 = ISD::getSetCCSwappedOperands(Op1);
1224 if (LL == RL && LR == RR) {
1225 bool isInteger = MVT::isInteger(LL.getValueType());
1226 ISD::CondCode Result = ISD::getSetCCOrOperation(Op0, Op1, isInteger);
1227 if (Result != ISD::SETCC_INVALID)
1228 return DAG.getSetCC(N0.getValueType(), LL, LR, Result);
1232 // Simplify: or (op x...), (op y...) -> (op (or x, y))
1233 if (N0.getOpcode() == N1.getOpcode()) {
1234 SDOperand Tmp = SimplifyBinOpWithSameOpcodeHands(N);
1235 if (Tmp.Val) return Tmp;
1238 // (X & C1) | (Y & C2) -> (X|Y) & C3 if possible.
1239 if (N0.getOpcode() == ISD::AND &&
1240 N1.getOpcode() == ISD::AND &&
1241 N0.getOperand(1).getOpcode() == ISD::Constant &&
1242 N1.getOperand(1).getOpcode() == ISD::Constant &&
1243 // Don't increase # computations.
1244 (N0.Val->hasOneUse() || N1.Val->hasOneUse())) {
1245 // We can only do this xform if we know that bits from X that are set in C2
1246 // but not in C1 are already zero. Likewise for Y.
1247 uint64_t LHSMask = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
1248 uint64_t RHSMask = cast<ConstantSDNode>(N1.getOperand(1))->getValue();
1250 if (TLI.MaskedValueIsZero(N0.getOperand(0), RHSMask&~LHSMask) &&
1251 TLI.MaskedValueIsZero(N1.getOperand(0), LHSMask&~RHSMask)) {
1252 SDOperand X =DAG.getNode(ISD::OR, VT, N0.getOperand(0), N1.getOperand(0));
1253 return DAG.getNode(ISD::AND, VT, X, DAG.getConstant(LHSMask|RHSMask, VT));
1258 // See if this is some rotate idiom.
1259 if (SDNode *Rot = MatchRotate(N0, N1))
1260 return SDOperand(Rot, 0);
1266 /// MatchRotateHalf - Match "(X shl/srl V1) & V2" where V2 may not be present.
1267 static bool MatchRotateHalf(SDOperand Op, SDOperand &Shift, SDOperand &Mask) {
1268 if (Op.getOpcode() == ISD::AND) {
1269 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1270 Mask = Op.getOperand(1);
1271 Op = Op.getOperand(0);
1277 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SHL) {
1285 // MatchRotate - Handle an 'or' of two operands. If this is one of the many
1286 // idioms for rotate, and if the target supports rotation instructions, generate
1288 SDNode *DAGCombiner::MatchRotate(SDOperand LHS, SDOperand RHS) {
1289 // Must be a legal type. Expanded an promoted things won't work with rotates.
1290 MVT::ValueType VT = LHS.getValueType();
1291 if (!TLI.isTypeLegal(VT)) return 0;
1293 // The target must have at least one rotate flavor.
1294 bool HasROTL = TLI.isOperationLegal(ISD::ROTL, VT);
1295 bool HasROTR = TLI.isOperationLegal(ISD::ROTR, VT);
1296 if (!HasROTL && !HasROTR) return 0;
1298 // Match "(X shl/srl V1) & V2" where V2 may not be present.
1299 SDOperand LHSShift; // The shift.
1300 SDOperand LHSMask; // AND value if any.
1301 if (!MatchRotateHalf(LHS, LHSShift, LHSMask))
1302 return 0; // Not part of a rotate.
1304 SDOperand RHSShift; // The shift.
1305 SDOperand RHSMask; // AND value if any.
1306 if (!MatchRotateHalf(RHS, RHSShift, RHSMask))
1307 return 0; // Not part of a rotate.
1309 if (LHSShift.getOperand(0) != RHSShift.getOperand(0))
1310 return 0; // Not shifting the same value.
1312 if (LHSShift.getOpcode() == RHSShift.getOpcode())
1313 return 0; // Shifts must disagree.
1315 // Canonicalize shl to left side in a shl/srl pair.
1316 if (RHSShift.getOpcode() == ISD::SHL) {
1317 std::swap(LHS, RHS);
1318 std::swap(LHSShift, RHSShift);
1319 std::swap(LHSMask , RHSMask );
1322 unsigned OpSizeInBits = MVT::getSizeInBits(VT);
1324 // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1)
1325 // fold (or (shl x, C1), (srl x, C2)) -> (rotr x, C2)
1326 if (LHSShift.getOperand(1).getOpcode() == ISD::Constant &&
1327 RHSShift.getOperand(1).getOpcode() == ISD::Constant) {
1328 uint64_t LShVal = cast<ConstantSDNode>(LHSShift.getOperand(1))->getValue();
1329 uint64_t RShVal = cast<ConstantSDNode>(RHSShift.getOperand(1))->getValue();
1330 if ((LShVal + RShVal) != OpSizeInBits)
1335 Rot = DAG.getNode(ISD::ROTL, VT, LHSShift.getOperand(0),
1336 LHSShift.getOperand(1));
1338 Rot = DAG.getNode(ISD::ROTR, VT, LHSShift.getOperand(0),
1339 RHSShift.getOperand(1));
1341 // If there is an AND of either shifted operand, apply it to the result.
1342 if (LHSMask.Val || RHSMask.Val) {
1343 uint64_t Mask = MVT::getIntVTBitMask(VT);
1346 uint64_t RHSBits = (1ULL << LShVal)-1;
1347 Mask &= cast<ConstantSDNode>(LHSMask)->getValue() | RHSBits;
1350 uint64_t LHSBits = ~((1ULL << (OpSizeInBits-RShVal))-1);
1351 Mask &= cast<ConstantSDNode>(RHSMask)->getValue() | LHSBits;
1354 Rot = DAG.getNode(ISD::AND, VT, Rot, DAG.getConstant(Mask, VT));
1360 // If there is a mask here, and we have a variable shift, we can't be sure
1361 // that we're masking out the right stuff.
1362 if (LHSMask.Val || RHSMask.Val)
1365 // fold (or (shl x, y), (srl x, (sub 32, y))) -> (rotl x, y)
1366 // fold (or (shl x, y), (srl x, (sub 32, y))) -> (rotr x, (sub 32, y))
1367 if (RHSShift.getOperand(1).getOpcode() == ISD::SUB &&
1368 LHSShift.getOperand(1) == RHSShift.getOperand(1).getOperand(1)) {
1369 if (ConstantSDNode *SUBC =
1370 dyn_cast<ConstantSDNode>(RHSShift.getOperand(1).getOperand(0))) {
1371 if (SUBC->getValue() == OpSizeInBits)
1373 return DAG.getNode(ISD::ROTL, VT, LHSShift.getOperand(0),
1374 LHSShift.getOperand(1)).Val;
1376 return DAG.getNode(ISD::ROTR, VT, LHSShift.getOperand(0),
1377 LHSShift.getOperand(1)).Val;
1381 // fold (or (shl x, (sub 32, y)), (srl x, r)) -> (rotr x, y)
1382 // fold (or (shl x, (sub 32, y)), (srl x, r)) -> (rotl x, (sub 32, y))
1383 if (LHSShift.getOperand(1).getOpcode() == ISD::SUB &&
1384 RHSShift.getOperand(1) == LHSShift.getOperand(1).getOperand(1)) {
1385 if (ConstantSDNode *SUBC =
1386 dyn_cast<ConstantSDNode>(LHSShift.getOperand(1).getOperand(0))) {
1387 if (SUBC->getValue() == OpSizeInBits)
1389 return DAG.getNode(ISD::ROTL, VT, LHSShift.getOperand(0),
1390 LHSShift.getOperand(1)).Val;
1392 return DAG.getNode(ISD::ROTR, VT, LHSShift.getOperand(0),
1393 RHSShift.getOperand(1)).Val;
1401 SDOperand DAGCombiner::visitXOR(SDNode *N) {
1402 SDOperand N0 = N->getOperand(0);
1403 SDOperand N1 = N->getOperand(1);
1404 SDOperand LHS, RHS, CC;
1405 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1406 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1407 MVT::ValueType VT = N0.getValueType();
1409 // fold (xor c1, c2) -> c1^c2
1411 return DAG.getNode(ISD::XOR, VT, N0, N1);
1412 // canonicalize constant to RHS
1414 return DAG.getNode(ISD::XOR, VT, N1, N0);
1415 // fold (xor x, 0) -> x
1416 if (N1C && N1C->isNullValue())
1419 SDOperand RXOR = ReassociateOps(ISD::XOR, N0, N1);
1422 // fold !(x cc y) -> (x !cc y)
1423 if (N1C && N1C->getValue() == 1 && isSetCCEquivalent(N0, LHS, RHS, CC)) {
1424 bool isInt = MVT::isInteger(LHS.getValueType());
1425 ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
1427 if (N0.getOpcode() == ISD::SETCC)
1428 return DAG.getSetCC(VT, LHS, RHS, NotCC);
1429 if (N0.getOpcode() == ISD::SELECT_CC)
1430 return DAG.getSelectCC(LHS, RHS, N0.getOperand(2),N0.getOperand(3),NotCC);
1431 assert(0 && "Unhandled SetCC Equivalent!");
1434 // fold !(x or y) -> (!x and !y) iff x or y are setcc
1435 if (N1C && N1C->getValue() == 1 &&
1436 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
1437 SDOperand LHS = N0.getOperand(0), RHS = N0.getOperand(1);
1438 if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) {
1439 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
1440 LHS = DAG.getNode(ISD::XOR, VT, LHS, N1); // RHS = ~LHS
1441 RHS = DAG.getNode(ISD::XOR, VT, RHS, N1); // RHS = ~RHS
1442 AddToWorkList(LHS.Val); AddToWorkList(RHS.Val);
1443 return DAG.getNode(NewOpcode, VT, LHS, RHS);
1446 // fold !(x or y) -> (!x and !y) iff x or y are constants
1447 if (N1C && N1C->isAllOnesValue() &&
1448 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
1449 SDOperand LHS = N0.getOperand(0), RHS = N0.getOperand(1);
1450 if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) {
1451 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
1452 LHS = DAG.getNode(ISD::XOR, VT, LHS, N1); // RHS = ~LHS
1453 RHS = DAG.getNode(ISD::XOR, VT, RHS, N1); // RHS = ~RHS
1454 AddToWorkList(LHS.Val); AddToWorkList(RHS.Val);
1455 return DAG.getNode(NewOpcode, VT, LHS, RHS);
1458 // fold (xor (xor x, c1), c2) -> (xor x, c1^c2)
1459 if (N1C && N0.getOpcode() == ISD::XOR) {
1460 ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0));
1461 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
1463 return DAG.getNode(ISD::XOR, VT, N0.getOperand(1),
1464 DAG.getConstant(N1C->getValue()^N00C->getValue(), VT));
1466 return DAG.getNode(ISD::XOR, VT, N0.getOperand(0),
1467 DAG.getConstant(N1C->getValue()^N01C->getValue(), VT));
1469 // fold (xor x, x) -> 0
1471 if (!MVT::isVector(VT)) {
1472 return DAG.getConstant(0, VT);
1473 } else if (!AfterLegalize || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT)) {
1474 // Produce a vector of zeros.
1475 SDOperand El = DAG.getConstant(0, MVT::getVectorBaseType(VT));
1476 std::vector<SDOperand> Ops(MVT::getVectorNumElements(VT), El);
1477 return DAG.getNode(ISD::BUILD_VECTOR, VT, &Ops[0], Ops.size());
1481 // Simplify: xor (op x...), (op y...) -> (op (xor x, y))
1482 if (N0.getOpcode() == N1.getOpcode()) {
1483 SDOperand Tmp = SimplifyBinOpWithSameOpcodeHands(N);
1484 if (Tmp.Val) return Tmp;
1487 // Simplify the expression using non-local knowledge.
1488 if (!MVT::isVector(VT) &&
1489 SimplifyDemandedBits(SDOperand(N, 0)))
1490 return SDOperand(N, 0);
1495 SDOperand DAGCombiner::visitSHL(SDNode *N) {
1496 SDOperand N0 = N->getOperand(0);
1497 SDOperand N1 = N->getOperand(1);
1498 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1499 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1500 MVT::ValueType VT = N0.getValueType();
1501 unsigned OpSizeInBits = MVT::getSizeInBits(VT);
1503 // fold (shl c1, c2) -> c1<<c2
1505 return DAG.getNode(ISD::SHL, VT, N0, N1);
1506 // fold (shl 0, x) -> 0
1507 if (N0C && N0C->isNullValue())
1509 // fold (shl x, c >= size(x)) -> undef
1510 if (N1C && N1C->getValue() >= OpSizeInBits)
1511 return DAG.getNode(ISD::UNDEF, VT);
1512 // fold (shl x, 0) -> x
1513 if (N1C && N1C->isNullValue())
1515 // if (shl x, c) is known to be zero, return 0
1516 if (TLI.MaskedValueIsZero(SDOperand(N, 0), MVT::getIntVTBitMask(VT)))
1517 return DAG.getConstant(0, VT);
1518 if (SimplifyDemandedBits(SDOperand(N, 0)))
1519 return SDOperand(N, 0);
1520 // fold (shl (shl x, c1), c2) -> 0 or (shl x, c1+c2)
1521 if (N1C && N0.getOpcode() == ISD::SHL &&
1522 N0.getOperand(1).getOpcode() == ISD::Constant) {
1523 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
1524 uint64_t c2 = N1C->getValue();
1525 if (c1 + c2 > OpSizeInBits)
1526 return DAG.getConstant(0, VT);
1527 return DAG.getNode(ISD::SHL, VT, N0.getOperand(0),
1528 DAG.getConstant(c1 + c2, N1.getValueType()));
1530 // fold (shl (srl x, c1), c2) -> (shl (and x, -1 << c1), c2-c1) or
1531 // (srl (and x, -1 << c1), c1-c2)
1532 if (N1C && N0.getOpcode() == ISD::SRL &&
1533 N0.getOperand(1).getOpcode() == ISD::Constant) {
1534 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
1535 uint64_t c2 = N1C->getValue();
1536 SDOperand Mask = DAG.getNode(ISD::AND, VT, N0.getOperand(0),
1537 DAG.getConstant(~0ULL << c1, VT));
1539 return DAG.getNode(ISD::SHL, VT, Mask,
1540 DAG.getConstant(c2-c1, N1.getValueType()));
1542 return DAG.getNode(ISD::SRL, VT, Mask,
1543 DAG.getConstant(c1-c2, N1.getValueType()));
1545 // fold (shl (sra x, c1), c1) -> (and x, -1 << c1)
1546 if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1))
1547 return DAG.getNode(ISD::AND, VT, N0.getOperand(0),
1548 DAG.getConstant(~0ULL << N1C->getValue(), VT));
1549 // fold (shl (add x, c1), c2) -> (add (shl x, c2), c1<<c2)
1550 if (N1C && N0.getOpcode() == ISD::ADD && N0.Val->hasOneUse() &&
1551 isa<ConstantSDNode>(N0.getOperand(1))) {
1552 return DAG.getNode(ISD::ADD, VT,
1553 DAG.getNode(ISD::SHL, VT, N0.getOperand(0), N1),
1554 DAG.getNode(ISD::SHL, VT, N0.getOperand(1), N1));
1559 SDOperand DAGCombiner::visitSRA(SDNode *N) {
1560 SDOperand N0 = N->getOperand(0);
1561 SDOperand N1 = N->getOperand(1);
1562 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1563 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1564 MVT::ValueType VT = N0.getValueType();
1566 // fold (sra c1, c2) -> c1>>c2
1568 return DAG.getNode(ISD::SRA, VT, N0, N1);
1569 // fold (sra 0, x) -> 0
1570 if (N0C && N0C->isNullValue())
1572 // fold (sra -1, x) -> -1
1573 if (N0C && N0C->isAllOnesValue())
1575 // fold (sra x, c >= size(x)) -> undef
1576 if (N1C && N1C->getValue() >= MVT::getSizeInBits(VT))
1577 return DAG.getNode(ISD::UNDEF, VT);
1578 // fold (sra x, 0) -> x
1579 if (N1C && N1C->isNullValue())
1581 // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports
1583 if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) {
1584 unsigned LowBits = MVT::getSizeInBits(VT) - (unsigned)N1C->getValue();
1587 default: EVT = MVT::Other; break;
1588 case 1: EVT = MVT::i1; break;
1589 case 8: EVT = MVT::i8; break;
1590 case 16: EVT = MVT::i16; break;
1591 case 32: EVT = MVT::i32; break;
1593 if (EVT > MVT::Other && TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, EVT))
1594 return DAG.getNode(ISD::SIGN_EXTEND_INREG, VT, N0.getOperand(0),
1595 DAG.getValueType(EVT));
1598 // fold (sra (sra x, c1), c2) -> (sra x, c1+c2)
1599 if (N1C && N0.getOpcode() == ISD::SRA) {
1600 if (ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1601 unsigned Sum = N1C->getValue() + C1->getValue();
1602 if (Sum >= MVT::getSizeInBits(VT)) Sum = MVT::getSizeInBits(VT)-1;
1603 return DAG.getNode(ISD::SRA, VT, N0.getOperand(0),
1604 DAG.getConstant(Sum, N1C->getValueType(0)));
1608 // Simplify, based on bits shifted out of the LHS.
1609 if (N1C && SimplifyDemandedBits(SDOperand(N, 0)))
1610 return SDOperand(N, 0);
1613 // If the sign bit is known to be zero, switch this to a SRL.
1614 if (TLI.MaskedValueIsZero(N0, MVT::getIntVTSignBit(VT)))
1615 return DAG.getNode(ISD::SRL, VT, N0, N1);
1619 SDOperand DAGCombiner::visitSRL(SDNode *N) {
1620 SDOperand N0 = N->getOperand(0);
1621 SDOperand N1 = N->getOperand(1);
1622 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1623 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1624 MVT::ValueType VT = N0.getValueType();
1625 unsigned OpSizeInBits = MVT::getSizeInBits(VT);
1627 // fold (srl c1, c2) -> c1 >>u c2
1629 return DAG.getNode(ISD::SRL, VT, N0, N1);
1630 // fold (srl 0, x) -> 0
1631 if (N0C && N0C->isNullValue())
1633 // fold (srl x, c >= size(x)) -> undef
1634 if (N1C && N1C->getValue() >= OpSizeInBits)
1635 return DAG.getNode(ISD::UNDEF, VT);
1636 // fold (srl x, 0) -> x
1637 if (N1C && N1C->isNullValue())
1639 // if (srl x, c) is known to be zero, return 0
1640 if (N1C && TLI.MaskedValueIsZero(SDOperand(N, 0), ~0ULL >> (64-OpSizeInBits)))
1641 return DAG.getConstant(0, VT);
1642 // fold (srl (srl x, c1), c2) -> 0 or (srl x, c1+c2)
1643 if (N1C && N0.getOpcode() == ISD::SRL &&
1644 N0.getOperand(1).getOpcode() == ISD::Constant) {
1645 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
1646 uint64_t c2 = N1C->getValue();
1647 if (c1 + c2 > OpSizeInBits)
1648 return DAG.getConstant(0, VT);
1649 return DAG.getNode(ISD::SRL, VT, N0.getOperand(0),
1650 DAG.getConstant(c1 + c2, N1.getValueType()));
1653 // fold (srl (anyextend x), c) -> (anyextend (srl x, c))
1654 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
1655 // Shifting in all undef bits?
1656 MVT::ValueType SmallVT = N0.getOperand(0).getValueType();
1657 if (N1C->getValue() >= MVT::getSizeInBits(SmallVT))
1658 return DAG.getNode(ISD::UNDEF, VT);
1660 SDOperand SmallShift = DAG.getNode(ISD::SRL, SmallVT, N0.getOperand(0), N1);
1661 AddToWorkList(SmallShift.Val);
1662 return DAG.getNode(ISD::ANY_EXTEND, VT, SmallShift);
1665 // fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit).
1666 if (N1C && N0.getOpcode() == ISD::CTLZ &&
1667 N1C->getValue() == Log2_32(MVT::getSizeInBits(VT))) {
1668 uint64_t KnownZero, KnownOne, Mask = MVT::getIntVTBitMask(VT);
1669 TLI.ComputeMaskedBits(N0.getOperand(0), Mask, KnownZero, KnownOne);
1671 // If any of the input bits are KnownOne, then the input couldn't be all
1672 // zeros, thus the result of the srl will always be zero.
1673 if (KnownOne) return DAG.getConstant(0, VT);
1675 // If all of the bits input the to ctlz node are known to be zero, then
1676 // the result of the ctlz is "32" and the result of the shift is one.
1677 uint64_t UnknownBits = ~KnownZero & Mask;
1678 if (UnknownBits == 0) return DAG.getConstant(1, VT);
1680 // Otherwise, check to see if there is exactly one bit input to the ctlz.
1681 if ((UnknownBits & (UnknownBits-1)) == 0) {
1682 // Okay, we know that only that the single bit specified by UnknownBits
1683 // could be set on input to the CTLZ node. If this bit is set, the SRL
1684 // will return 0, if it is clear, it returns 1. Change the CTLZ/SRL pair
1685 // to an SRL,XOR pair, which is likely to simplify more.
1686 unsigned ShAmt = CountTrailingZeros_64(UnknownBits);
1687 SDOperand Op = N0.getOperand(0);
1689 Op = DAG.getNode(ISD::SRL, VT, Op,
1690 DAG.getConstant(ShAmt, TLI.getShiftAmountTy()));
1691 AddToWorkList(Op.Val);
1693 return DAG.getNode(ISD::XOR, VT, Op, DAG.getConstant(1, VT));
1700 SDOperand DAGCombiner::visitCTLZ(SDNode *N) {
1701 SDOperand N0 = N->getOperand(0);
1702 MVT::ValueType VT = N->getValueType(0);
1704 // fold (ctlz c1) -> c2
1705 if (isa<ConstantSDNode>(N0))
1706 return DAG.getNode(ISD::CTLZ, VT, N0);
1710 SDOperand DAGCombiner::visitCTTZ(SDNode *N) {
1711 SDOperand N0 = N->getOperand(0);
1712 MVT::ValueType VT = N->getValueType(0);
1714 // fold (cttz c1) -> c2
1715 if (isa<ConstantSDNode>(N0))
1716 return DAG.getNode(ISD::CTTZ, VT, N0);
1720 SDOperand DAGCombiner::visitCTPOP(SDNode *N) {
1721 SDOperand N0 = N->getOperand(0);
1722 MVT::ValueType VT = N->getValueType(0);
1724 // fold (ctpop c1) -> c2
1725 if (isa<ConstantSDNode>(N0))
1726 return DAG.getNode(ISD::CTPOP, VT, N0);
1730 SDOperand DAGCombiner::visitSELECT(SDNode *N) {
1731 SDOperand N0 = N->getOperand(0);
1732 SDOperand N1 = N->getOperand(1);
1733 SDOperand N2 = N->getOperand(2);
1734 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1735 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1736 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
1737 MVT::ValueType VT = N->getValueType(0);
1739 // fold select C, X, X -> X
1742 // fold select true, X, Y -> X
1743 if (N0C && !N0C->isNullValue())
1745 // fold select false, X, Y -> Y
1746 if (N0C && N0C->isNullValue())
1748 // fold select C, 1, X -> C | X
1749 if (MVT::i1 == VT && N1C && N1C->getValue() == 1)
1750 return DAG.getNode(ISD::OR, VT, N0, N2);
1751 // fold select C, 0, X -> ~C & X
1752 // FIXME: this should check for C type == X type, not i1?
1753 if (MVT::i1 == VT && N1C && N1C->isNullValue()) {
1754 SDOperand XORNode = DAG.getNode(ISD::XOR, VT, N0, DAG.getConstant(1, VT));
1755 AddToWorkList(XORNode.Val);
1756 return DAG.getNode(ISD::AND, VT, XORNode, N2);
1758 // fold select C, X, 1 -> ~C | X
1759 if (MVT::i1 == VT && N2C && N2C->getValue() == 1) {
1760 SDOperand XORNode = DAG.getNode(ISD::XOR, VT, N0, DAG.getConstant(1, VT));
1761 AddToWorkList(XORNode.Val);
1762 return DAG.getNode(ISD::OR, VT, XORNode, N1);
1764 // fold select C, X, 0 -> C & X
1765 // FIXME: this should check for C type == X type, not i1?
1766 if (MVT::i1 == VT && N2C && N2C->isNullValue())
1767 return DAG.getNode(ISD::AND, VT, N0, N1);
1768 // fold X ? X : Y --> X ? 1 : Y --> X | Y
1769 if (MVT::i1 == VT && N0 == N1)
1770 return DAG.getNode(ISD::OR, VT, N0, N2);
1771 // fold X ? Y : X --> X ? Y : 0 --> X & Y
1772 if (MVT::i1 == VT && N0 == N2)
1773 return DAG.getNode(ISD::AND, VT, N0, N1);
1775 // If we can fold this based on the true/false value, do so.
1776 if (SimplifySelectOps(N, N1, N2))
1777 return SDOperand(N, 0); // Don't revisit N.
1779 // fold selects based on a setcc into other things, such as min/max/abs
1780 if (N0.getOpcode() == ISD::SETCC)
1782 // Check against MVT::Other for SELECT_CC, which is a workaround for targets
1783 // having to say they don't support SELECT_CC on every type the DAG knows
1784 // about, since there is no way to mark an opcode illegal at all value types
1785 if (TLI.isOperationLegal(ISD::SELECT_CC, MVT::Other))
1786 return DAG.getNode(ISD::SELECT_CC, VT, N0.getOperand(0), N0.getOperand(1),
1787 N1, N2, N0.getOperand(2));
1789 return SimplifySelect(N0, N1, N2);
1793 SDOperand DAGCombiner::visitSELECT_CC(SDNode *N) {
1794 SDOperand N0 = N->getOperand(0);
1795 SDOperand N1 = N->getOperand(1);
1796 SDOperand N2 = N->getOperand(2);
1797 SDOperand N3 = N->getOperand(3);
1798 SDOperand N4 = N->getOperand(4);
1799 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1800 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1801 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
1802 ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get();
1804 // fold select_cc lhs, rhs, x, x, cc -> x
1808 // Determine if the condition we're dealing with is constant
1809 SDOperand SCC = SimplifySetCC(TLI.getSetCCResultTy(), N0, N1, CC, false);
1811 if (ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.Val)) {
1812 if (SCCC->getValue())
1813 return N2; // cond always true -> true val
1815 return N3; // cond always false -> false val
1818 // Fold to a simpler select_cc
1819 if (SCC.Val && SCC.getOpcode() == ISD::SETCC)
1820 return DAG.getNode(ISD::SELECT_CC, N2.getValueType(),
1821 SCC.getOperand(0), SCC.getOperand(1), N2, N3,
1824 // If we can fold this based on the true/false value, do so.
1825 if (SimplifySelectOps(N, N2, N3))
1826 return SDOperand(N, 0); // Don't revisit N.
1828 // fold select_cc into other things, such as min/max/abs
1829 return SimplifySelectCC(N0, N1, N2, N3, CC);
1832 SDOperand DAGCombiner::visitSETCC(SDNode *N) {
1833 return SimplifySetCC(N->getValueType(0), N->getOperand(0), N->getOperand(1),
1834 cast<CondCodeSDNode>(N->getOperand(2))->get());
1837 SDOperand DAGCombiner::visitSIGN_EXTEND(SDNode *N) {
1838 SDOperand N0 = N->getOperand(0);
1839 MVT::ValueType VT = N->getValueType(0);
1841 // fold (sext c1) -> c1
1842 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0))
1843 return DAG.getNode(ISD::SIGN_EXTEND, VT, N0);
1845 // fold (sext (sext x)) -> (sext x)
1846 // fold (sext (aext x)) -> (sext x)
1847 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
1848 return DAG.getNode(ISD::SIGN_EXTEND, VT, N0.getOperand(0));
1850 // fold (sext (truncate x)) -> (sextinreg x).
1851 if (N0.getOpcode() == ISD::TRUNCATE &&
1852 (!AfterLegalize || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG,
1853 N0.getValueType()))) {
1854 SDOperand Op = N0.getOperand(0);
1855 if (Op.getValueType() < VT) {
1856 Op = DAG.getNode(ISD::ANY_EXTEND, VT, Op);
1857 } else if (Op.getValueType() > VT) {
1858 Op = DAG.getNode(ISD::TRUNCATE, VT, Op);
1860 return DAG.getNode(ISD::SIGN_EXTEND_INREG, VT, Op,
1861 DAG.getValueType(N0.getValueType()));
1864 // fold (sext (load x)) -> (sext (truncate (sextload x)))
1865 if (N0.getOpcode() == ISD::LOAD && N0.hasOneUse() &&
1866 (!AfterLegalize||TLI.isOperationLegal(ISD::SEXTLOAD, N0.getValueType()))){
1867 SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, N0.getOperand(0),
1868 N0.getOperand(1), N0.getOperand(2),
1870 CombineTo(N, ExtLoad);
1871 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
1872 ExtLoad.getValue(1));
1873 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
1876 // fold (sext (sextload x)) -> (sext (truncate (sextload x)))
1877 // fold (sext ( extload x)) -> (sext (truncate (sextload x)))
1878 if ((N0.getOpcode() == ISD::SEXTLOAD || N0.getOpcode() == ISD::EXTLOAD) &&
1880 MVT::ValueType EVT = cast<VTSDNode>(N0.getOperand(3))->getVT();
1881 SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, N0.getOperand(0),
1882 N0.getOperand(1), N0.getOperand(2), EVT);
1883 CombineTo(N, ExtLoad);
1884 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
1885 ExtLoad.getValue(1));
1886 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
1892 SDOperand DAGCombiner::visitZERO_EXTEND(SDNode *N) {
1893 SDOperand N0 = N->getOperand(0);
1894 MVT::ValueType VT = N->getValueType(0);
1896 // fold (zext c1) -> c1
1897 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0))
1898 return DAG.getNode(ISD::ZERO_EXTEND, VT, N0);
1899 // fold (zext (zext x)) -> (zext x)
1900 // fold (zext (aext x)) -> (zext x)
1901 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
1902 return DAG.getNode(ISD::ZERO_EXTEND, VT, N0.getOperand(0));
1904 // fold (zext (truncate x)) -> (and x, mask)
1905 if (N0.getOpcode() == ISD::TRUNCATE &&
1906 (!AfterLegalize || TLI.isOperationLegal(ISD::AND, VT))) {
1907 SDOperand Op = N0.getOperand(0);
1908 if (Op.getValueType() < VT) {
1909 Op = DAG.getNode(ISD::ANY_EXTEND, VT, Op);
1910 } else if (Op.getValueType() > VT) {
1911 Op = DAG.getNode(ISD::TRUNCATE, VT, Op);
1913 return DAG.getZeroExtendInReg(Op, N0.getValueType());
1916 // fold (zext (and (trunc x), cst)) -> (and x, cst).
1917 if (N0.getOpcode() == ISD::AND &&
1918 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
1919 N0.getOperand(1).getOpcode() == ISD::Constant) {
1920 SDOperand X = N0.getOperand(0).getOperand(0);
1921 if (X.getValueType() < VT) {
1922 X = DAG.getNode(ISD::ANY_EXTEND, VT, X);
1923 } else if (X.getValueType() > VT) {
1924 X = DAG.getNode(ISD::TRUNCATE, VT, X);
1926 uint64_t Mask = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
1927 return DAG.getNode(ISD::AND, VT, X, DAG.getConstant(Mask, VT));
1930 // fold (zext (load x)) -> (zext (truncate (zextload x)))
1931 if (N0.getOpcode() == ISD::LOAD && N0.hasOneUse() &&
1932 (!AfterLegalize||TLI.isOperationLegal(ISD::ZEXTLOAD, N0.getValueType()))){
1933 SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, N0.getOperand(0),
1934 N0.getOperand(1), N0.getOperand(2),
1936 CombineTo(N, ExtLoad);
1937 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
1938 ExtLoad.getValue(1));
1939 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
1942 // fold (zext (zextload x)) -> (zext (truncate (zextload x)))
1943 // fold (zext ( extload x)) -> (zext (truncate (zextload x)))
1944 if ((N0.getOpcode() == ISD::ZEXTLOAD || N0.getOpcode() == ISD::EXTLOAD) &&
1946 MVT::ValueType EVT = cast<VTSDNode>(N0.getOperand(3))->getVT();
1947 SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, N0.getOperand(0),
1948 N0.getOperand(1), N0.getOperand(2), EVT);
1949 CombineTo(N, ExtLoad);
1950 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
1951 ExtLoad.getValue(1));
1952 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
1957 SDOperand DAGCombiner::visitANY_EXTEND(SDNode *N) {
1958 SDOperand N0 = N->getOperand(0);
1959 MVT::ValueType VT = N->getValueType(0);
1961 // fold (aext c1) -> c1
1962 if (isa<ConstantSDNode>(N0))
1963 return DAG.getNode(ISD::ANY_EXTEND, VT, N0);
1964 // fold (aext (aext x)) -> (aext x)
1965 // fold (aext (zext x)) -> (zext x)
1966 // fold (aext (sext x)) -> (sext x)
1967 if (N0.getOpcode() == ISD::ANY_EXTEND ||
1968 N0.getOpcode() == ISD::ZERO_EXTEND ||
1969 N0.getOpcode() == ISD::SIGN_EXTEND)
1970 return DAG.getNode(N0.getOpcode(), VT, N0.getOperand(0));
1972 // fold (aext (truncate x))
1973 if (N0.getOpcode() == ISD::TRUNCATE) {
1974 SDOperand TruncOp = N0.getOperand(0);
1975 if (TruncOp.getValueType() == VT)
1976 return TruncOp; // x iff x size == zext size.
1977 if (TruncOp.getValueType() > VT)
1978 return DAG.getNode(ISD::TRUNCATE, VT, TruncOp);
1979 return DAG.getNode(ISD::ANY_EXTEND, VT, TruncOp);
1982 // fold (aext (and (trunc x), cst)) -> (and x, cst).
1983 if (N0.getOpcode() == ISD::AND &&
1984 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
1985 N0.getOperand(1).getOpcode() == ISD::Constant) {
1986 SDOperand X = N0.getOperand(0).getOperand(0);
1987 if (X.getValueType() < VT) {
1988 X = DAG.getNode(ISD::ANY_EXTEND, VT, X);
1989 } else if (X.getValueType() > VT) {
1990 X = DAG.getNode(ISD::TRUNCATE, VT, X);
1992 uint64_t Mask = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
1993 return DAG.getNode(ISD::AND, VT, X, DAG.getConstant(Mask, VT));
1996 // fold (aext (load x)) -> (aext (truncate (extload x)))
1997 if (N0.getOpcode() == ISD::LOAD && N0.hasOneUse() &&
1998 (!AfterLegalize||TLI.isOperationLegal(ISD::EXTLOAD, N0.getValueType()))) {
1999 SDOperand ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, VT, N0.getOperand(0),
2000 N0.getOperand(1), N0.getOperand(2),
2002 CombineTo(N, ExtLoad);
2003 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
2004 ExtLoad.getValue(1));
2005 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
2008 // fold (aext (zextload x)) -> (aext (truncate (zextload x)))
2009 // fold (aext (sextload x)) -> (aext (truncate (sextload x)))
2010 // fold (aext ( extload x)) -> (aext (truncate (extload x)))
2011 if ((N0.getOpcode() == ISD::ZEXTLOAD || N0.getOpcode() == ISD::EXTLOAD ||
2012 N0.getOpcode() == ISD::SEXTLOAD) &&
2014 MVT::ValueType EVT = cast<VTSDNode>(N0.getOperand(3))->getVT();
2015 SDOperand ExtLoad = DAG.getExtLoad(N0.getOpcode(), VT, N0.getOperand(0),
2016 N0.getOperand(1), N0.getOperand(2), EVT);
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!
2026 SDOperand DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
2027 SDOperand N0 = N->getOperand(0);
2028 SDOperand N1 = N->getOperand(1);
2029 MVT::ValueType VT = N->getValueType(0);
2030 MVT::ValueType EVT = cast<VTSDNode>(N1)->getVT();
2031 unsigned EVTBits = MVT::getSizeInBits(EVT);
2033 // fold (sext_in_reg c1) -> c1
2034 if (isa<ConstantSDNode>(N0) || N0.getOpcode() == ISD::UNDEF)
2035 return DAG.getNode(ISD::SIGN_EXTEND_INREG, VT, N0, N1);
2037 // If the input is already sign extended, just drop the extension.
2038 if (TLI.ComputeNumSignBits(N0) >= MVT::getSizeInBits(VT)-EVTBits+1)
2041 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2
2042 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
2043 EVT < cast<VTSDNode>(N0.getOperand(1))->getVT()) {
2044 return DAG.getNode(ISD::SIGN_EXTEND_INREG, VT, N0.getOperand(0), N1);
2047 // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is zero
2048 if (TLI.MaskedValueIsZero(N0, 1ULL << (EVTBits-1)))
2049 return DAG.getZeroExtendInReg(N0, EVT);
2051 // fold (sext_in_reg (srl X, 24), i8) -> sra X, 24
2052 // fold (sext_in_reg (srl X, 23), i8) -> sra X, 23 iff possible.
2053 // We already fold "(sext_in_reg (srl X, 25), i8) -> srl X, 25" above.
2054 if (N0.getOpcode() == ISD::SRL) {
2055 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
2056 if (ShAmt->getValue()+EVTBits <= MVT::getSizeInBits(VT)) {
2057 // We can turn this into an SRA iff the input to the SRL is already sign
2059 unsigned InSignBits = TLI.ComputeNumSignBits(N0.getOperand(0));
2060 if (MVT::getSizeInBits(VT)-(ShAmt->getValue()+EVTBits) < InSignBits)
2061 return DAG.getNode(ISD::SRA, VT, N0.getOperand(0), N0.getOperand(1));
2065 // fold (sext_inreg (extload x)) -> (sextload x)
2066 if (N0.getOpcode() == ISD::EXTLOAD &&
2067 EVT == cast<VTSDNode>(N0.getOperand(3))->getVT() &&
2068 (!AfterLegalize || TLI.isOperationLegal(ISD::SEXTLOAD, EVT))) {
2069 SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, N0.getOperand(0),
2070 N0.getOperand(1), N0.getOperand(2),
2072 CombineTo(N, ExtLoad);
2073 CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1));
2074 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
2076 // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use
2077 if (N0.getOpcode() == ISD::ZEXTLOAD && N0.hasOneUse() &&
2078 EVT == cast<VTSDNode>(N0.getOperand(3))->getVT() &&
2079 (!AfterLegalize || TLI.isOperationLegal(ISD::SEXTLOAD, EVT))) {
2080 SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, N0.getOperand(0),
2081 N0.getOperand(1), N0.getOperand(2),
2083 CombineTo(N, ExtLoad);
2084 CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1));
2085 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
2090 SDOperand DAGCombiner::visitTRUNCATE(SDNode *N) {
2091 SDOperand N0 = N->getOperand(0);
2092 MVT::ValueType VT = N->getValueType(0);
2095 if (N0.getValueType() == N->getValueType(0))
2097 // fold (truncate c1) -> c1
2098 if (isa<ConstantSDNode>(N0))
2099 return DAG.getNode(ISD::TRUNCATE, VT, N0);
2100 // fold (truncate (truncate x)) -> (truncate x)
2101 if (N0.getOpcode() == ISD::TRUNCATE)
2102 return DAG.getNode(ISD::TRUNCATE, VT, N0.getOperand(0));
2103 // fold (truncate (ext x)) -> (ext x) or (truncate x) or x
2104 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::SIGN_EXTEND||
2105 N0.getOpcode() == ISD::ANY_EXTEND) {
2106 if (N0.getValueType() < VT)
2107 // if the source is smaller than the dest, we still need an extend
2108 return DAG.getNode(N0.getOpcode(), VT, N0.getOperand(0));
2109 else if (N0.getValueType() > VT)
2110 // if the source is larger than the dest, than we just need the truncate
2111 return DAG.getNode(ISD::TRUNCATE, VT, N0.getOperand(0));
2113 // if the source and dest are the same type, we can drop both the extend
2115 return N0.getOperand(0);
2117 // fold (truncate (load x)) -> (smaller load x)
2118 if (N0.getOpcode() == ISD::LOAD && N0.hasOneUse()) {
2119 assert(MVT::getSizeInBits(N0.getValueType()) > MVT::getSizeInBits(VT) &&
2120 "Cannot truncate to larger type!");
2121 MVT::ValueType PtrType = N0.getOperand(1).getValueType();
2122 // For big endian targets, we need to add an offset to the pointer to load
2123 // the correct bytes. For little endian systems, we merely need to read
2124 // fewer bytes from the same pointer.
2126 (MVT::getSizeInBits(N0.getValueType()) - MVT::getSizeInBits(VT)) / 8;
2127 SDOperand NewPtr = TLI.isLittleEndian() ? N0.getOperand(1) :
2128 DAG.getNode(ISD::ADD, PtrType, N0.getOperand(1),
2129 DAG.getConstant(PtrOff, PtrType));
2130 AddToWorkList(NewPtr.Val);
2131 SDOperand Load = DAG.getLoad(VT, N0.getOperand(0), NewPtr,N0.getOperand(2));
2133 CombineTo(N0.Val, Load, Load.getValue(1));
2134 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
2139 SDOperand DAGCombiner::visitBIT_CONVERT(SDNode *N) {
2140 SDOperand N0 = N->getOperand(0);
2141 MVT::ValueType VT = N->getValueType(0);
2143 // If the input is a constant, let getNode() fold it.
2144 if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) {
2145 SDOperand Res = DAG.getNode(ISD::BIT_CONVERT, VT, N0);
2146 if (Res.Val != N) return Res;
2149 if (N0.getOpcode() == ISD::BIT_CONVERT) // conv(conv(x,t1),t2) -> conv(x,t2)
2150 return DAG.getNode(ISD::BIT_CONVERT, VT, N0.getOperand(0));
2152 // fold (conv (load x)) -> (load (conv*)x)
2153 // FIXME: These xforms need to know that the resultant load doesn't need a
2154 // higher alignment than the original!
2155 if (0 && N0.getOpcode() == ISD::LOAD && N0.hasOneUse()) {
2156 SDOperand Load = DAG.getLoad(VT, N0.getOperand(0), N0.getOperand(1),
2159 CombineTo(N0.Val, DAG.getNode(ISD::BIT_CONVERT, N0.getValueType(), Load),
2167 SDOperand DAGCombiner::visitVBIT_CONVERT(SDNode *N) {
2168 SDOperand N0 = N->getOperand(0);
2169 MVT::ValueType VT = N->getValueType(0);
2171 // If the input is a VBUILD_VECTOR with all constant elements, fold this now.
2172 // First check to see if this is all constant.
2173 if (N0.getOpcode() == ISD::VBUILD_VECTOR && N0.Val->hasOneUse() &&
2174 VT == MVT::Vector) {
2175 bool isSimple = true;
2176 for (unsigned i = 0, e = N0.getNumOperands()-2; i != e; ++i)
2177 if (N0.getOperand(i).getOpcode() != ISD::UNDEF &&
2178 N0.getOperand(i).getOpcode() != ISD::Constant &&
2179 N0.getOperand(i).getOpcode() != ISD::ConstantFP) {
2184 MVT::ValueType DestEltVT = cast<VTSDNode>(N->getOperand(2))->getVT();
2185 if (isSimple && !MVT::isVector(DestEltVT)) {
2186 return ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(N0.Val, DestEltVT);
2193 /// ConstantFoldVBIT_CONVERTofVBUILD_VECTOR - We know that BV is a vbuild_vector
2194 /// node with Constant, ConstantFP or Undef operands. DstEltVT indicates the
2195 /// destination element value type.
2196 SDOperand DAGCombiner::
2197 ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(SDNode *BV, MVT::ValueType DstEltVT) {
2198 MVT::ValueType SrcEltVT = BV->getOperand(0).getValueType();
2200 // If this is already the right type, we're done.
2201 if (SrcEltVT == DstEltVT) return SDOperand(BV, 0);
2203 unsigned SrcBitSize = MVT::getSizeInBits(SrcEltVT);
2204 unsigned DstBitSize = MVT::getSizeInBits(DstEltVT);
2206 // If this is a conversion of N elements of one type to N elements of another
2207 // type, convert each element. This handles FP<->INT cases.
2208 if (SrcBitSize == DstBitSize) {
2209 SmallVector<SDOperand, 8> Ops;
2210 for (unsigned i = 0, e = BV->getNumOperands()-2; i != e; ++i) {
2211 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, DstEltVT, BV->getOperand(i)));
2212 AddToWorkList(Ops.back().Val);
2214 Ops.push_back(*(BV->op_end()-2)); // Add num elements.
2215 Ops.push_back(DAG.getValueType(DstEltVT));
2216 return DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, &Ops[0], Ops.size());
2219 // Otherwise, we're growing or shrinking the elements. To avoid having to
2220 // handle annoying details of growing/shrinking FP values, we convert them to
2222 if (MVT::isFloatingPoint(SrcEltVT)) {
2223 // Convert the input float vector to a int vector where the elements are the
2225 assert((SrcEltVT == MVT::f32 || SrcEltVT == MVT::f64) && "Unknown FP VT!");
2226 MVT::ValueType IntVT = SrcEltVT == MVT::f32 ? MVT::i32 : MVT::i64;
2227 BV = ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(BV, IntVT).Val;
2231 // Now we know the input is an integer vector. If the output is a FP type,
2232 // convert to integer first, then to FP of the right size.
2233 if (MVT::isFloatingPoint(DstEltVT)) {
2234 assert((DstEltVT == MVT::f32 || DstEltVT == MVT::f64) && "Unknown FP VT!");
2235 MVT::ValueType TmpVT = DstEltVT == MVT::f32 ? MVT::i32 : MVT::i64;
2236 SDNode *Tmp = ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(BV, TmpVT).Val;
2238 // Next, convert to FP elements of the same size.
2239 return ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(Tmp, DstEltVT);
2242 // Okay, we know the src/dst types are both integers of differing types.
2243 // Handling growing first.
2244 assert(MVT::isInteger(SrcEltVT) && MVT::isInteger(DstEltVT));
2245 if (SrcBitSize < DstBitSize) {
2246 unsigned NumInputsPerOutput = DstBitSize/SrcBitSize;
2248 SmallVector<SDOperand, 8> Ops;
2249 for (unsigned i = 0, e = BV->getNumOperands()-2; i != e;
2250 i += NumInputsPerOutput) {
2251 bool isLE = TLI.isLittleEndian();
2252 uint64_t NewBits = 0;
2253 bool EltIsUndef = true;
2254 for (unsigned j = 0; j != NumInputsPerOutput; ++j) {
2255 // Shift the previously computed bits over.
2256 NewBits <<= SrcBitSize;
2257 SDOperand Op = BV->getOperand(i+ (isLE ? (NumInputsPerOutput-j-1) : j));
2258 if (Op.getOpcode() == ISD::UNDEF) continue;
2261 NewBits |= cast<ConstantSDNode>(Op)->getValue();
2265 Ops.push_back(DAG.getNode(ISD::UNDEF, DstEltVT));
2267 Ops.push_back(DAG.getConstant(NewBits, DstEltVT));
2270 Ops.push_back(DAG.getConstant(Ops.size(), MVT::i32)); // Add num elements.
2271 Ops.push_back(DAG.getValueType(DstEltVT)); // Add element size.
2272 return DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, &Ops[0], Ops.size());
2275 // Finally, this must be the case where we are shrinking elements: each input
2276 // turns into multiple outputs.
2277 unsigned NumOutputsPerInput = SrcBitSize/DstBitSize;
2278 SmallVector<SDOperand, 8> Ops;
2279 for (unsigned i = 0, e = BV->getNumOperands()-2; i != e; ++i) {
2280 if (BV->getOperand(i).getOpcode() == ISD::UNDEF) {
2281 for (unsigned j = 0; j != NumOutputsPerInput; ++j)
2282 Ops.push_back(DAG.getNode(ISD::UNDEF, DstEltVT));
2285 uint64_t OpVal = cast<ConstantSDNode>(BV->getOperand(i))->getValue();
2287 for (unsigned j = 0; j != NumOutputsPerInput; ++j) {
2288 unsigned ThisVal = OpVal & ((1ULL << DstBitSize)-1);
2289 OpVal >>= DstBitSize;
2290 Ops.push_back(DAG.getConstant(ThisVal, DstEltVT));
2293 // For big endian targets, swap the order of the pieces of each element.
2294 if (!TLI.isLittleEndian())
2295 std::reverse(Ops.end()-NumOutputsPerInput, Ops.end());
2297 Ops.push_back(DAG.getConstant(Ops.size(), MVT::i32)); // Add num elements.
2298 Ops.push_back(DAG.getValueType(DstEltVT)); // Add element size.
2299 return DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, &Ops[0], Ops.size());
2304 SDOperand DAGCombiner::visitFADD(SDNode *N) {
2305 SDOperand N0 = N->getOperand(0);
2306 SDOperand N1 = N->getOperand(1);
2307 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2308 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
2309 MVT::ValueType VT = N->getValueType(0);
2311 // fold (fadd c1, c2) -> c1+c2
2313 return DAG.getNode(ISD::FADD, VT, N0, N1);
2314 // canonicalize constant to RHS
2315 if (N0CFP && !N1CFP)
2316 return DAG.getNode(ISD::FADD, VT, N1, N0);
2317 // fold (A + (-B)) -> A-B
2318 if (N1.getOpcode() == ISD::FNEG)
2319 return DAG.getNode(ISD::FSUB, VT, N0, N1.getOperand(0));
2320 // fold ((-A) + B) -> B-A
2321 if (N0.getOpcode() == ISD::FNEG)
2322 return DAG.getNode(ISD::FSUB, VT, N1, N0.getOperand(0));
2326 SDOperand DAGCombiner::visitFSUB(SDNode *N) {
2327 SDOperand N0 = N->getOperand(0);
2328 SDOperand N1 = N->getOperand(1);
2329 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2330 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
2331 MVT::ValueType VT = N->getValueType(0);
2333 // fold (fsub c1, c2) -> c1-c2
2335 return DAG.getNode(ISD::FSUB, VT, N0, N1);
2336 // fold (A-(-B)) -> A+B
2337 if (N1.getOpcode() == ISD::FNEG)
2338 return DAG.getNode(ISD::FADD, VT, N0, N1.getOperand(0));
2342 SDOperand DAGCombiner::visitFMUL(SDNode *N) {
2343 SDOperand N0 = N->getOperand(0);
2344 SDOperand N1 = N->getOperand(1);
2345 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2346 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
2347 MVT::ValueType VT = N->getValueType(0);
2349 // fold (fmul c1, c2) -> c1*c2
2351 return DAG.getNode(ISD::FMUL, VT, N0, N1);
2352 // canonicalize constant to RHS
2353 if (N0CFP && !N1CFP)
2354 return DAG.getNode(ISD::FMUL, VT, N1, N0);
2355 // fold (fmul X, 2.0) -> (fadd X, X)
2356 if (N1CFP && N1CFP->isExactlyValue(+2.0))
2357 return DAG.getNode(ISD::FADD, VT, N0, N0);
2361 SDOperand DAGCombiner::visitFDIV(SDNode *N) {
2362 SDOperand N0 = N->getOperand(0);
2363 SDOperand N1 = N->getOperand(1);
2364 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2365 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
2366 MVT::ValueType VT = N->getValueType(0);
2368 // fold (fdiv c1, c2) -> c1/c2
2370 return DAG.getNode(ISD::FDIV, VT, N0, N1);
2374 SDOperand DAGCombiner::visitFREM(SDNode *N) {
2375 SDOperand N0 = N->getOperand(0);
2376 SDOperand N1 = N->getOperand(1);
2377 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2378 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
2379 MVT::ValueType VT = N->getValueType(0);
2381 // fold (frem c1, c2) -> fmod(c1,c2)
2383 return DAG.getNode(ISD::FREM, VT, N0, N1);
2387 SDOperand DAGCombiner::visitFCOPYSIGN(SDNode *N) {
2388 SDOperand N0 = N->getOperand(0);
2389 SDOperand N1 = N->getOperand(1);
2390 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2391 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
2392 MVT::ValueType VT = N->getValueType(0);
2394 if (N0CFP && N1CFP) // Constant fold
2395 return DAG.getNode(ISD::FCOPYSIGN, VT, N0, N1);
2398 // copysign(x, c1) -> fabs(x) iff ispos(c1)
2399 // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1)
2404 u.d = N1CFP->getValue();
2406 return DAG.getNode(ISD::FABS, VT, N0);
2408 return DAG.getNode(ISD::FNEG, VT, DAG.getNode(ISD::FABS, VT, N0));
2411 // copysign(fabs(x), y) -> copysign(x, y)
2412 // copysign(fneg(x), y) -> copysign(x, y)
2413 // copysign(copysign(x,z), y) -> copysign(x, y)
2414 if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG ||
2415 N0.getOpcode() == ISD::FCOPYSIGN)
2416 return DAG.getNode(ISD::FCOPYSIGN, VT, N0.getOperand(0), N1);
2418 // copysign(x, abs(y)) -> abs(x)
2419 if (N1.getOpcode() == ISD::FABS)
2420 return DAG.getNode(ISD::FABS, VT, N0);
2422 // copysign(x, copysign(y,z)) -> copysign(x, z)
2423 if (N1.getOpcode() == ISD::FCOPYSIGN)
2424 return DAG.getNode(ISD::FCOPYSIGN, VT, N0, N1.getOperand(1));
2426 // copysign(x, fp_extend(y)) -> copysign(x, y)
2427 // copysign(x, fp_round(y)) -> copysign(x, y)
2428 if (N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND)
2429 return DAG.getNode(ISD::FCOPYSIGN, VT, N0, N1.getOperand(0));
2436 SDOperand DAGCombiner::visitSINT_TO_FP(SDNode *N) {
2437 SDOperand N0 = N->getOperand(0);
2438 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2439 MVT::ValueType VT = N->getValueType(0);
2441 // fold (sint_to_fp c1) -> c1fp
2443 return DAG.getNode(ISD::SINT_TO_FP, VT, N0);
2447 SDOperand DAGCombiner::visitUINT_TO_FP(SDNode *N) {
2448 SDOperand N0 = N->getOperand(0);
2449 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2450 MVT::ValueType VT = N->getValueType(0);
2452 // fold (uint_to_fp c1) -> c1fp
2454 return DAG.getNode(ISD::UINT_TO_FP, VT, N0);
2458 SDOperand DAGCombiner::visitFP_TO_SINT(SDNode *N) {
2459 SDOperand N0 = N->getOperand(0);
2460 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2461 MVT::ValueType VT = N->getValueType(0);
2463 // fold (fp_to_sint c1fp) -> c1
2465 return DAG.getNode(ISD::FP_TO_SINT, VT, N0);
2469 SDOperand DAGCombiner::visitFP_TO_UINT(SDNode *N) {
2470 SDOperand N0 = N->getOperand(0);
2471 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2472 MVT::ValueType VT = N->getValueType(0);
2474 // fold (fp_to_uint c1fp) -> c1
2476 return DAG.getNode(ISD::FP_TO_UINT, VT, N0);
2480 SDOperand DAGCombiner::visitFP_ROUND(SDNode *N) {
2481 SDOperand N0 = N->getOperand(0);
2482 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2483 MVT::ValueType VT = N->getValueType(0);
2485 // fold (fp_round c1fp) -> c1fp
2487 return DAG.getNode(ISD::FP_ROUND, VT, N0);
2489 // fold (fp_round (fp_extend x)) -> x
2490 if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType())
2491 return N0.getOperand(0);
2493 // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y)
2494 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.Val->hasOneUse()) {
2495 SDOperand Tmp = DAG.getNode(ISD::FP_ROUND, VT, N0.getOperand(0));
2496 AddToWorkList(Tmp.Val);
2497 return DAG.getNode(ISD::FCOPYSIGN, VT, Tmp, N0.getOperand(1));
2503 SDOperand DAGCombiner::visitFP_ROUND_INREG(SDNode *N) {
2504 SDOperand N0 = N->getOperand(0);
2505 MVT::ValueType VT = N->getValueType(0);
2506 MVT::ValueType EVT = cast<VTSDNode>(N->getOperand(1))->getVT();
2507 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2509 // fold (fp_round_inreg c1fp) -> c1fp
2511 SDOperand Round = DAG.getConstantFP(N0CFP->getValue(), EVT);
2512 return DAG.getNode(ISD::FP_EXTEND, VT, Round);
2517 SDOperand DAGCombiner::visitFP_EXTEND(SDNode *N) {
2518 SDOperand N0 = N->getOperand(0);
2519 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2520 MVT::ValueType VT = N->getValueType(0);
2522 // fold (fp_extend c1fp) -> c1fp
2524 return DAG.getNode(ISD::FP_EXTEND, VT, N0);
2526 // fold (fpext (load x)) -> (fpext (fpround (extload x)))
2527 if (N0.getOpcode() == ISD::LOAD && N0.hasOneUse() &&
2528 (!AfterLegalize||TLI.isOperationLegal(ISD::EXTLOAD, N0.getValueType()))) {
2529 SDOperand ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, VT, N0.getOperand(0),
2530 N0.getOperand(1), N0.getOperand(2),
2532 CombineTo(N, ExtLoad);
2533 CombineTo(N0.Val, DAG.getNode(ISD::FP_ROUND, N0.getValueType(), ExtLoad),
2534 ExtLoad.getValue(1));
2535 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
2542 SDOperand DAGCombiner::visitFNEG(SDNode *N) {
2543 SDOperand N0 = N->getOperand(0);
2544 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2545 MVT::ValueType VT = N->getValueType(0);
2547 // fold (fneg c1) -> -c1
2549 return DAG.getNode(ISD::FNEG, VT, N0);
2550 // fold (fneg (sub x, y)) -> (sub y, x)
2551 if (N0.getOpcode() == ISD::SUB)
2552 return DAG.getNode(ISD::SUB, VT, N0.getOperand(1), N0.getOperand(0));
2553 // fold (fneg (fneg x)) -> x
2554 if (N0.getOpcode() == ISD::FNEG)
2555 return N0.getOperand(0);
2559 SDOperand DAGCombiner::visitFABS(SDNode *N) {
2560 SDOperand N0 = N->getOperand(0);
2561 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2562 MVT::ValueType VT = N->getValueType(0);
2564 // fold (fabs c1) -> fabs(c1)
2566 return DAG.getNode(ISD::FABS, VT, N0);
2567 // fold (fabs (fabs x)) -> (fabs x)
2568 if (N0.getOpcode() == ISD::FABS)
2569 return N->getOperand(0);
2570 // fold (fabs (fneg x)) -> (fabs x)
2571 // fold (fabs (fcopysign x, y)) -> (fabs x)
2572 if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN)
2573 return DAG.getNode(ISD::FABS, VT, N0.getOperand(0));
2578 SDOperand DAGCombiner::visitBRCOND(SDNode *N) {
2579 SDOperand Chain = N->getOperand(0);
2580 SDOperand N1 = N->getOperand(1);
2581 SDOperand N2 = N->getOperand(2);
2582 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2584 // never taken branch, fold to chain
2585 if (N1C && N1C->isNullValue())
2587 // unconditional branch
2588 if (N1C && N1C->getValue() == 1)
2589 return DAG.getNode(ISD::BR, MVT::Other, Chain, N2);
2590 // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal
2592 if (N1.getOpcode() == ISD::SETCC &&
2593 TLI.isOperationLegal(ISD::BR_CC, MVT::Other)) {
2594 return DAG.getNode(ISD::BR_CC, MVT::Other, Chain, N1.getOperand(2),
2595 N1.getOperand(0), N1.getOperand(1), N2);
2600 // Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB.
2602 SDOperand DAGCombiner::visitBR_CC(SDNode *N) {
2603 CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1));
2604 SDOperand CondLHS = N->getOperand(2), CondRHS = N->getOperand(3);
2606 // Use SimplifySetCC to simplify SETCC's.
2607 SDOperand Simp = SimplifySetCC(MVT::i1, CondLHS, CondRHS, CC->get(), false);
2608 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(Simp.Val);
2610 // fold br_cc true, dest -> br dest (unconditional branch)
2611 if (SCCC && SCCC->getValue())
2612 return DAG.getNode(ISD::BR, MVT::Other, N->getOperand(0),
2614 // fold br_cc false, dest -> unconditional fall through
2615 if (SCCC && SCCC->isNullValue())
2616 return N->getOperand(0);
2617 // fold to a simpler setcc
2618 if (Simp.Val && Simp.getOpcode() == ISD::SETCC)
2619 return DAG.getNode(ISD::BR_CC, MVT::Other, N->getOperand(0),
2620 Simp.getOperand(2), Simp.getOperand(0),
2621 Simp.getOperand(1), N->getOperand(4));
2625 SDOperand DAGCombiner::visitLOAD(SDNode *N) {
2626 SDOperand Chain = N->getOperand(0);
2627 SDOperand Ptr = N->getOperand(1);
2628 SDOperand SrcValue = N->getOperand(2);
2630 // If there are no uses of the loaded value, change uses of the chain value
2631 // into uses of the chain input (i.e. delete the dead load).
2632 if (N->hasNUsesOfValue(0, 0))
2633 return CombineTo(N, DAG.getNode(ISD::UNDEF, N->getValueType(0)), Chain);
2635 // If this load is directly stored, replace the load value with the stored
2637 // TODO: Handle store large -> read small portion.
2638 // TODO: Handle TRUNCSTORE/EXTLOAD
2639 if (Chain.getOpcode() == ISD::STORE && Chain.getOperand(2) == Ptr &&
2640 Chain.getOperand(1).getValueType() == N->getValueType(0))
2641 return CombineTo(N, Chain.getOperand(1), Chain);
2644 // Walk up chain skipping non-aliasing memory nodes.
2645 SDOperand BetterChain = FindBetterChain(N, Chain);
2647 // If the there is a better chain.
2648 if (Chain != BetterChain) {
2649 // Replace the chain to void dependency.
2650 SDOperand ReplLoad = DAG.getLoad(N->getValueType(0), BetterChain, Ptr,
2653 // Create token factor to keep chain around.
2654 SDOperand Token = DAG.getNode(ISD::TokenFactor, MVT::Other,
2655 Chain, ReplLoad.getValue(1));
2657 // Replace uses with load and token factor.
2658 CombineTo(N, ReplLoad.getValue(0), Token);
2660 return SDOperand(N, 0);
2667 /// visitXEXTLOAD - Handle EXTLOAD/ZEXTLOAD/SEXTLOAD.
2668 SDOperand DAGCombiner::visitXEXTLOAD(SDNode *N) {
2669 SDOperand Chain = N->getOperand(0);
2670 SDOperand Ptr = N->getOperand(1);
2671 SDOperand SrcValue = N->getOperand(2);
2672 SDOperand EVT = N->getOperand(3);
2674 // If there are no uses of the loaded value, change uses of the chain value
2675 // into uses of the chain input (i.e. delete the dead load).
2676 if (N->hasNUsesOfValue(0, 0))
2677 return CombineTo(N, DAG.getNode(ISD::UNDEF, N->getValueType(0)), Chain);
2682 SDOperand DAGCombiner::visitSTORE(SDNode *N) {
2683 SDOperand Chain = N->getOperand(0);
2684 SDOperand Value = N->getOperand(1);
2685 SDOperand Ptr = N->getOperand(2);
2686 SDOperand SrcValue = N->getOperand(3);
2688 // If this is a store that kills a previous store, remove the previous store.
2689 if (Chain.getOpcode() == ISD::STORE && Chain.getOperand(2) == Ptr &&
2690 Chain.Val->hasOneUse() /* Avoid introducing DAG cycles */ &&
2691 // Make sure that these stores are the same value type:
2692 // FIXME: we really care that the second store is >= size of the first.
2693 Value.getValueType() == Chain.getOperand(1).getValueType()) {
2694 // Create a new store of Value that replaces both stores.
2695 SDNode *PrevStore = Chain.Val;
2696 if (PrevStore->getOperand(1) == Value) // Same value multiply stored.
2698 SDOperand NewStore = DAG.getNode(ISD::STORE, MVT::Other,
2699 PrevStore->getOperand(0), Value, Ptr,
2701 CombineTo(N, NewStore); // Nuke this store.
2702 CombineTo(PrevStore, NewStore); // Nuke the previous store.
2703 return SDOperand(N, 0);
2706 // If this is a store of a bit convert, store the input value.
2707 // FIXME: This needs to know that the resultant store does not need a
2708 // higher alignment than the original.
2709 if (Value.getOpcode() == ISD::BIT_CONVERT) {
2710 return DAG.getNode(ISD::STORE, MVT::Other, Chain, Value.getOperand(0),
2715 // If the store ptr is a frame index and the frame index has a use of one
2716 // and this is a return block, then the store is redundant.
2717 if (Ptr.hasOneUse() && isa<FrameIndexSDNode>(Ptr) &&
2718 DAG.getRoot().getOpcode() == ISD::RET) {
2722 // Walk up chain skipping non-aliasing memory nodes.
2723 SDOperand BetterChain = FindBetterChain(N, Chain);
2725 // If the there is a better chain.
2726 if (Chain != BetterChain) {
2727 // Replace the chain to void dependency.
2728 SDOperand ReplStore = DAG.getNode(ISD::STORE, MVT::Other,
2729 BetterChain, Value, Ptr,
2731 // Create token to keep both nodes around.
2732 SDOperand Token = DAG.getNode(ISD::TokenFactor, MVT::Other,
2735 // Make sure we merge token factors.
2736 AddUsersToWorkList(N);
2738 // Old chain needs to be cleaned up.
2739 AddToWorkList(Chain.Val);
2748 SDOperand DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) {
2749 SDOperand InVec = N->getOperand(0);
2750 SDOperand InVal = N->getOperand(1);
2751 SDOperand EltNo = N->getOperand(2);
2753 // If the invec is a BUILD_VECTOR and if EltNo is a constant, build a new
2754 // vector with the inserted element.
2755 if (InVec.getOpcode() == ISD::BUILD_VECTOR && isa<ConstantSDNode>(EltNo)) {
2756 unsigned Elt = cast<ConstantSDNode>(EltNo)->getValue();
2757 SmallVector<SDOperand, 8> Ops(InVec.Val->op_begin(), InVec.Val->op_end());
2758 if (Elt < Ops.size())
2760 return DAG.getNode(ISD::BUILD_VECTOR, InVec.getValueType(),
2761 &Ops[0], Ops.size());
2767 SDOperand DAGCombiner::visitVINSERT_VECTOR_ELT(SDNode *N) {
2768 SDOperand InVec = N->getOperand(0);
2769 SDOperand InVal = N->getOperand(1);
2770 SDOperand EltNo = N->getOperand(2);
2771 SDOperand NumElts = N->getOperand(3);
2772 SDOperand EltType = N->getOperand(4);
2774 // If the invec is a VBUILD_VECTOR and if EltNo is a constant, build a new
2775 // vector with the inserted element.
2776 if (InVec.getOpcode() == ISD::VBUILD_VECTOR && isa<ConstantSDNode>(EltNo)) {
2777 unsigned Elt = cast<ConstantSDNode>(EltNo)->getValue();
2778 SmallVector<SDOperand, 8> Ops(InVec.Val->op_begin(), InVec.Val->op_end());
2779 if (Elt < Ops.size()-2)
2781 return DAG.getNode(ISD::VBUILD_VECTOR, InVec.getValueType(),
2782 &Ops[0], Ops.size());
2788 SDOperand DAGCombiner::visitVBUILD_VECTOR(SDNode *N) {
2789 unsigned NumInScalars = N->getNumOperands()-2;
2790 SDOperand NumElts = N->getOperand(NumInScalars);
2791 SDOperand EltType = N->getOperand(NumInScalars+1);
2793 // Check to see if this is a VBUILD_VECTOR of a bunch of VEXTRACT_VECTOR_ELT
2794 // operations. If so, and if the EXTRACT_ELT vector inputs come from at most
2795 // two distinct vectors, turn this into a shuffle node.
2796 SDOperand VecIn1, VecIn2;
2797 for (unsigned i = 0; i != NumInScalars; ++i) {
2798 // Ignore undef inputs.
2799 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
2801 // If this input is something other than a VEXTRACT_VECTOR_ELT with a
2802 // constant index, bail out.
2803 if (N->getOperand(i).getOpcode() != ISD::VEXTRACT_VECTOR_ELT ||
2804 !isa<ConstantSDNode>(N->getOperand(i).getOperand(1))) {
2805 VecIn1 = VecIn2 = SDOperand(0, 0);
2809 // If the input vector type disagrees with the result of the vbuild_vector,
2810 // we can't make a shuffle.
2811 SDOperand ExtractedFromVec = N->getOperand(i).getOperand(0);
2812 if (*(ExtractedFromVec.Val->op_end()-2) != NumElts ||
2813 *(ExtractedFromVec.Val->op_end()-1) != EltType) {
2814 VecIn1 = VecIn2 = SDOperand(0, 0);
2818 // Otherwise, remember this. We allow up to two distinct input vectors.
2819 if (ExtractedFromVec == VecIn1 || ExtractedFromVec == VecIn2)
2822 if (VecIn1.Val == 0) {
2823 VecIn1 = ExtractedFromVec;
2824 } else if (VecIn2.Val == 0) {
2825 VecIn2 = ExtractedFromVec;
2828 VecIn1 = VecIn2 = SDOperand(0, 0);
2833 // If everything is good, we can make a shuffle operation.
2835 SmallVector<SDOperand, 8> BuildVecIndices;
2836 for (unsigned i = 0; i != NumInScalars; ++i) {
2837 if (N->getOperand(i).getOpcode() == ISD::UNDEF) {
2838 BuildVecIndices.push_back(DAG.getNode(ISD::UNDEF, MVT::i32));
2842 SDOperand Extract = N->getOperand(i);
2844 // If extracting from the first vector, just use the index directly.
2845 if (Extract.getOperand(0) == VecIn1) {
2846 BuildVecIndices.push_back(Extract.getOperand(1));
2850 // Otherwise, use InIdx + VecSize
2851 unsigned Idx = cast<ConstantSDNode>(Extract.getOperand(1))->getValue();
2852 BuildVecIndices.push_back(DAG.getConstant(Idx+NumInScalars, MVT::i32));
2855 // Add count and size info.
2856 BuildVecIndices.push_back(NumElts);
2857 BuildVecIndices.push_back(DAG.getValueType(MVT::i32));
2859 // Return the new VVECTOR_SHUFFLE node.
2865 // Use an undef vbuild_vector as input for the second operand.
2866 std::vector<SDOperand> UnOps(NumInScalars,
2867 DAG.getNode(ISD::UNDEF,
2868 cast<VTSDNode>(EltType)->getVT()));
2869 UnOps.push_back(NumElts);
2870 UnOps.push_back(EltType);
2871 Ops[1] = DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector,
2872 &UnOps[0], UnOps.size());
2873 AddToWorkList(Ops[1].Val);
2875 Ops[2] = DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector,
2876 &BuildVecIndices[0], BuildVecIndices.size());
2879 return DAG.getNode(ISD::VVECTOR_SHUFFLE, MVT::Vector, Ops, 5);
2885 SDOperand DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
2886 SDOperand ShufMask = N->getOperand(2);
2887 unsigned NumElts = ShufMask.getNumOperands();
2889 // If the shuffle mask is an identity operation on the LHS, return the LHS.
2890 bool isIdentity = true;
2891 for (unsigned i = 0; i != NumElts; ++i) {
2892 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF &&
2893 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() != i) {
2898 if (isIdentity) return N->getOperand(0);
2900 // If the shuffle mask is an identity operation on the RHS, return the RHS.
2902 for (unsigned i = 0; i != NumElts; ++i) {
2903 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF &&
2904 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() != i+NumElts) {
2909 if (isIdentity) return N->getOperand(1);
2911 // Check if the shuffle is a unary shuffle, i.e. one of the vectors is not
2913 bool isUnary = true;
2914 bool isSplat = true;
2916 unsigned BaseIdx = 0;
2917 for (unsigned i = 0; i != NumElts; ++i)
2918 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF) {
2919 unsigned Idx = cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue();
2920 int V = (Idx < NumElts) ? 0 : 1;
2934 SDOperand N0 = N->getOperand(0);
2935 SDOperand N1 = N->getOperand(1);
2936 // Normalize unary shuffle so the RHS is undef.
2937 if (isUnary && VecNum == 1)
2940 // If it is a splat, check if the argument vector is a build_vector with
2941 // all scalar elements the same.
2944 if (V->getOpcode() == ISD::BIT_CONVERT)
2945 V = V->getOperand(0).Val;
2946 if (V->getOpcode() == ISD::BUILD_VECTOR) {
2947 unsigned NumElems = V->getNumOperands()-2;
2948 if (NumElems > BaseIdx) {
2950 bool AllSame = true;
2951 for (unsigned i = 0; i != NumElems; ++i) {
2952 if (V->getOperand(i).getOpcode() != ISD::UNDEF) {
2953 Base = V->getOperand(i);
2957 // Splat of <u, u, u, u>, return <u, u, u, u>
2960 for (unsigned i = 0; i != NumElems; ++i) {
2961 if (V->getOperand(i).getOpcode() != ISD::UNDEF &&
2962 V->getOperand(i) != Base) {
2967 // Splat of <x, x, x, x>, return <x, x, x, x>
2974 // If it is a unary or the LHS and the RHS are the same node, turn the RHS
2976 if (isUnary || N0 == N1) {
2977 if (N0.getOpcode() == ISD::UNDEF)
2978 return DAG.getNode(ISD::UNDEF, N->getValueType(0));
2979 // Check the SHUFFLE mask, mapping any inputs from the 2nd operand into the
2981 SmallVector<SDOperand, 8> MappedOps;
2982 for (unsigned i = 0, e = ShufMask.getNumOperands(); i != e; ++i) {
2983 if (ShufMask.getOperand(i).getOpcode() == ISD::UNDEF ||
2984 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() < NumElts) {
2985 MappedOps.push_back(ShufMask.getOperand(i));
2988 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() - NumElts;
2989 MappedOps.push_back(DAG.getConstant(NewIdx, MVT::i32));
2992 ShufMask = DAG.getNode(ISD::BUILD_VECTOR, ShufMask.getValueType(),
2993 &MappedOps[0], MappedOps.size());
2994 AddToWorkList(ShufMask.Val);
2995 return DAG.getNode(ISD::VECTOR_SHUFFLE, N->getValueType(0),
2997 DAG.getNode(ISD::UNDEF, N->getValueType(0)),
3004 SDOperand DAGCombiner::visitVVECTOR_SHUFFLE(SDNode *N) {
3005 SDOperand ShufMask = N->getOperand(2);
3006 unsigned NumElts = ShufMask.getNumOperands()-2;
3008 // If the shuffle mask is an identity operation on the LHS, return the LHS.
3009 bool isIdentity = true;
3010 for (unsigned i = 0; i != NumElts; ++i) {
3011 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF &&
3012 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() != i) {
3017 if (isIdentity) return N->getOperand(0);
3019 // If the shuffle mask is an identity operation on the RHS, return the RHS.
3021 for (unsigned i = 0; i != NumElts; ++i) {
3022 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF &&
3023 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() != i+NumElts) {
3028 if (isIdentity) return N->getOperand(1);
3030 // Check if the shuffle is a unary shuffle, i.e. one of the vectors is not
3032 bool isUnary = true;
3033 bool isSplat = true;
3035 unsigned BaseIdx = 0;
3036 for (unsigned i = 0; i != NumElts; ++i)
3037 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF) {
3038 unsigned Idx = cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue();
3039 int V = (Idx < NumElts) ? 0 : 1;
3053 SDOperand N0 = N->getOperand(0);
3054 SDOperand N1 = N->getOperand(1);
3055 // Normalize unary shuffle so the RHS is undef.
3056 if (isUnary && VecNum == 1)
3059 // If it is a splat, check if the argument vector is a build_vector with
3060 // all scalar elements the same.
3063 if (V->getOpcode() == ISD::VBIT_CONVERT)
3064 V = V->getOperand(0).Val;
3065 if (V->getOpcode() == ISD::VBUILD_VECTOR) {
3066 unsigned NumElems = V->getNumOperands()-2;
3067 if (NumElems > BaseIdx) {
3069 bool AllSame = true;
3070 for (unsigned i = 0; i != NumElems; ++i) {
3071 if (V->getOperand(i).getOpcode() != ISD::UNDEF) {
3072 Base = V->getOperand(i);
3076 // Splat of <u, u, u, u>, return <u, u, u, u>
3079 for (unsigned i = 0; i != NumElems; ++i) {
3080 if (V->getOperand(i).getOpcode() != ISD::UNDEF &&
3081 V->getOperand(i) != Base) {
3086 // Splat of <x, x, x, x>, return <x, x, x, x>
3093 // If it is a unary or the LHS and the RHS are the same node, turn the RHS
3095 if (isUnary || N0 == N1) {
3096 // Check the SHUFFLE mask, mapping any inputs from the 2nd operand into the
3098 SmallVector<SDOperand, 8> MappedOps;
3099 for (unsigned i = 0; i != NumElts; ++i) {
3100 if (ShufMask.getOperand(i).getOpcode() == ISD::UNDEF ||
3101 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() < NumElts) {
3102 MappedOps.push_back(ShufMask.getOperand(i));
3105 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() - NumElts;
3106 MappedOps.push_back(DAG.getConstant(NewIdx, MVT::i32));
3109 // Add the type/#elts values.
3110 MappedOps.push_back(ShufMask.getOperand(NumElts));
3111 MappedOps.push_back(ShufMask.getOperand(NumElts+1));
3113 ShufMask = DAG.getNode(ISD::VBUILD_VECTOR, ShufMask.getValueType(),
3114 &MappedOps[0], MappedOps.size());
3115 AddToWorkList(ShufMask.Val);
3117 // Build the undef vector.
3118 SDOperand UDVal = DAG.getNode(ISD::UNDEF, MappedOps[0].getValueType());
3119 for (unsigned i = 0; i != NumElts; ++i)
3120 MappedOps[i] = UDVal;
3121 MappedOps[NumElts ] = *(N0.Val->op_end()-2);
3122 MappedOps[NumElts+1] = *(N0.Val->op_end()-1);
3123 UDVal = DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector,
3124 &MappedOps[0], MappedOps.size());
3126 return DAG.getNode(ISD::VVECTOR_SHUFFLE, MVT::Vector,
3127 N0, UDVal, ShufMask,
3128 MappedOps[NumElts], MappedOps[NumElts+1]);
3134 /// XformToShuffleWithZero - Returns a vector_shuffle if it able to transform
3135 /// a VAND to a vector_shuffle with the destination vector and a zero vector.
3136 /// e.g. VAND V, <0xffffffff, 0, 0xffffffff, 0>. ==>
3137 /// vector_shuffle V, Zero, <0, 4, 2, 4>
3138 SDOperand DAGCombiner::XformToShuffleWithZero(SDNode *N) {
3139 SDOperand LHS = N->getOperand(0);
3140 SDOperand RHS = N->getOperand(1);
3141 if (N->getOpcode() == ISD::VAND) {
3142 SDOperand DstVecSize = *(LHS.Val->op_end()-2);
3143 SDOperand DstVecEVT = *(LHS.Val->op_end()-1);
3144 if (RHS.getOpcode() == ISD::VBIT_CONVERT)
3145 RHS = RHS.getOperand(0);
3146 if (RHS.getOpcode() == ISD::VBUILD_VECTOR) {
3147 std::vector<SDOperand> IdxOps;
3148 unsigned NumOps = RHS.getNumOperands();
3149 unsigned NumElts = NumOps-2;
3150 MVT::ValueType EVT = cast<VTSDNode>(RHS.getOperand(NumOps-1))->getVT();
3151 for (unsigned i = 0; i != NumElts; ++i) {
3152 SDOperand Elt = RHS.getOperand(i);
3153 if (!isa<ConstantSDNode>(Elt))
3155 else if (cast<ConstantSDNode>(Elt)->isAllOnesValue())
3156 IdxOps.push_back(DAG.getConstant(i, EVT));
3157 else if (cast<ConstantSDNode>(Elt)->isNullValue())
3158 IdxOps.push_back(DAG.getConstant(NumElts, EVT));
3163 // Let's see if the target supports this vector_shuffle.
3164 if (!TLI.isVectorClearMaskLegal(IdxOps, EVT, DAG))
3167 // Return the new VVECTOR_SHUFFLE node.
3168 SDOperand NumEltsNode = DAG.getConstant(NumElts, MVT::i32);
3169 SDOperand EVTNode = DAG.getValueType(EVT);
3170 std::vector<SDOperand> Ops;
3171 LHS = DAG.getNode(ISD::VBIT_CONVERT, MVT::Vector, LHS, NumEltsNode,
3174 AddToWorkList(LHS.Val);
3175 std::vector<SDOperand> ZeroOps(NumElts, DAG.getConstant(0, EVT));
3176 ZeroOps.push_back(NumEltsNode);
3177 ZeroOps.push_back(EVTNode);
3178 Ops.push_back(DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector,
3179 &ZeroOps[0], ZeroOps.size()));
3180 IdxOps.push_back(NumEltsNode);
3181 IdxOps.push_back(EVTNode);
3182 Ops.push_back(DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector,
3183 &IdxOps[0], IdxOps.size()));
3184 Ops.push_back(NumEltsNode);
3185 Ops.push_back(EVTNode);
3186 SDOperand Result = DAG.getNode(ISD::VVECTOR_SHUFFLE, MVT::Vector,
3187 &Ops[0], Ops.size());
3188 if (NumEltsNode != DstVecSize || EVTNode != DstVecEVT) {
3189 Result = DAG.getNode(ISD::VBIT_CONVERT, MVT::Vector, Result,
3190 DstVecSize, DstVecEVT);
3198 /// visitVBinOp - Visit a binary vector operation, like VADD. IntOp indicates
3199 /// the scalar operation of the vop if it is operating on an integer vector
3200 /// (e.g. ADD) and FPOp indicates the FP version (e.g. FADD).
3201 SDOperand DAGCombiner::visitVBinOp(SDNode *N, ISD::NodeType IntOp,
3202 ISD::NodeType FPOp) {
3203 MVT::ValueType EltType = cast<VTSDNode>(*(N->op_end()-1))->getVT();
3204 ISD::NodeType ScalarOp = MVT::isInteger(EltType) ? IntOp : FPOp;
3205 SDOperand LHS = N->getOperand(0);
3206 SDOperand RHS = N->getOperand(1);
3207 SDOperand Shuffle = XformToShuffleWithZero(N);
3208 if (Shuffle.Val) return Shuffle;
3210 // If the LHS and RHS are VBUILD_VECTOR nodes, see if we can constant fold
3212 if (LHS.getOpcode() == ISD::VBUILD_VECTOR &&
3213 RHS.getOpcode() == ISD::VBUILD_VECTOR) {
3214 SmallVector<SDOperand, 8> Ops;
3215 for (unsigned i = 0, e = LHS.getNumOperands()-2; i != e; ++i) {
3216 SDOperand LHSOp = LHS.getOperand(i);
3217 SDOperand RHSOp = RHS.getOperand(i);
3218 // If these two elements can't be folded, bail out.
3219 if ((LHSOp.getOpcode() != ISD::UNDEF &&
3220 LHSOp.getOpcode() != ISD::Constant &&
3221 LHSOp.getOpcode() != ISD::ConstantFP) ||
3222 (RHSOp.getOpcode() != ISD::UNDEF &&
3223 RHSOp.getOpcode() != ISD::Constant &&
3224 RHSOp.getOpcode() != ISD::ConstantFP))
3226 // Can't fold divide by zero.
3227 if (N->getOpcode() == ISD::VSDIV || N->getOpcode() == ISD::VUDIV) {
3228 if ((RHSOp.getOpcode() == ISD::Constant &&
3229 cast<ConstantSDNode>(RHSOp.Val)->isNullValue()) ||
3230 (RHSOp.getOpcode() == ISD::ConstantFP &&
3231 !cast<ConstantFPSDNode>(RHSOp.Val)->getValue()))
3234 Ops.push_back(DAG.getNode(ScalarOp, EltType, LHSOp, RHSOp));
3235 AddToWorkList(Ops.back().Val);
3236 assert((Ops.back().getOpcode() == ISD::UNDEF ||
3237 Ops.back().getOpcode() == ISD::Constant ||
3238 Ops.back().getOpcode() == ISD::ConstantFP) &&
3239 "Scalar binop didn't fold!");
3242 if (Ops.size() == LHS.getNumOperands()-2) {
3243 Ops.push_back(*(LHS.Val->op_end()-2));
3244 Ops.push_back(*(LHS.Val->op_end()-1));
3245 return DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, &Ops[0], Ops.size());
3252 SDOperand DAGCombiner::SimplifySelect(SDOperand N0, SDOperand N1, SDOperand N2){
3253 assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!");
3255 SDOperand SCC = SimplifySelectCC(N0.getOperand(0), N0.getOperand(1), N1, N2,
3256 cast<CondCodeSDNode>(N0.getOperand(2))->get());
3257 // If we got a simplified select_cc node back from SimplifySelectCC, then
3258 // break it down into a new SETCC node, and a new SELECT node, and then return
3259 // the SELECT node, since we were called with a SELECT node.
3261 // Check to see if we got a select_cc back (to turn into setcc/select).
3262 // Otherwise, just return whatever node we got back, like fabs.
3263 if (SCC.getOpcode() == ISD::SELECT_CC) {
3264 SDOperand SETCC = DAG.getNode(ISD::SETCC, N0.getValueType(),
3265 SCC.getOperand(0), SCC.getOperand(1),
3267 AddToWorkList(SETCC.Val);
3268 return DAG.getNode(ISD::SELECT, SCC.getValueType(), SCC.getOperand(2),
3269 SCC.getOperand(3), SETCC);
3276 /// SimplifySelectOps - Given a SELECT or a SELECT_CC node, where LHS and RHS
3277 /// are the two values being selected between, see if we can simplify the
3278 /// select. Callers of this should assume that TheSelect is deleted if this
3279 /// returns true. As such, they should return the appropriate thing (e.g. the
3280 /// node) back to the top-level of the DAG combiner loop to avoid it being
3283 bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDOperand LHS,
3286 // If this is a select from two identical things, try to pull the operation
3287 // through the select.
3288 if (LHS.getOpcode() == RHS.getOpcode() && LHS.hasOneUse() && RHS.hasOneUse()){
3290 std::cerr << "SELECT: ["; LHS.Val->dump();
3291 std::cerr << "] ["; RHS.Val->dump();
3295 // If this is a load and the token chain is identical, replace the select
3296 // of two loads with a load through a select of the address to load from.
3297 // This triggers in things like "select bool X, 10.0, 123.0" after the FP
3298 // constants have been dropped into the constant pool.
3299 if ((LHS.getOpcode() == ISD::LOAD ||
3300 LHS.getOpcode() == ISD::EXTLOAD ||
3301 LHS.getOpcode() == ISD::ZEXTLOAD ||
3302 LHS.getOpcode() == ISD::SEXTLOAD) &&
3303 // Token chains must be identical.
3304 LHS.getOperand(0) == RHS.getOperand(0) &&
3305 // If this is an EXTLOAD, the VT's must match.
3306 (LHS.getOpcode() == ISD::LOAD ||
3307 LHS.getOperand(3) == RHS.getOperand(3))) {
3308 // FIXME: this conflates two src values, discarding one. This is not
3309 // the right thing to do, but nothing uses srcvalues now. When they do,
3310 // turn SrcValue into a list of locations.
3312 if (TheSelect->getOpcode() == ISD::SELECT)
3313 Addr = DAG.getNode(ISD::SELECT, LHS.getOperand(1).getValueType(),
3314 TheSelect->getOperand(0), LHS.getOperand(1),
3317 Addr = DAG.getNode(ISD::SELECT_CC, LHS.getOperand(1).getValueType(),
3318 TheSelect->getOperand(0),
3319 TheSelect->getOperand(1),
3320 LHS.getOperand(1), RHS.getOperand(1),
3321 TheSelect->getOperand(4));
3324 if (LHS.getOpcode() == ISD::LOAD)
3325 Load = DAG.getLoad(TheSelect->getValueType(0), LHS.getOperand(0),
3326 Addr, LHS.getOperand(2));
3328 Load = DAG.getExtLoad(LHS.getOpcode(), TheSelect->getValueType(0),
3329 LHS.getOperand(0), Addr, LHS.getOperand(2),
3330 cast<VTSDNode>(LHS.getOperand(3))->getVT());
3331 // Users of the select now use the result of the load.
3332 CombineTo(TheSelect, Load);
3334 // Users of the old loads now use the new load's chain. We know the
3335 // old-load value is dead now.
3336 CombineTo(LHS.Val, Load.getValue(0), Load.getValue(1));
3337 CombineTo(RHS.Val, Load.getValue(0), Load.getValue(1));
3345 SDOperand DAGCombiner::SimplifySelectCC(SDOperand N0, SDOperand N1,
3346 SDOperand N2, SDOperand N3,
3349 MVT::ValueType VT = N2.getValueType();
3350 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val);
3351 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.Val);
3352 ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N3.Val);
3354 // Determine if the condition we're dealing with is constant
3355 SDOperand SCC = SimplifySetCC(TLI.getSetCCResultTy(), N0, N1, CC, false);
3356 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.Val);
3358 // fold select_cc true, x, y -> x
3359 if (SCCC && SCCC->getValue())
3361 // fold select_cc false, x, y -> y
3362 if (SCCC && SCCC->getValue() == 0)
3365 // Check to see if we can simplify the select into an fabs node
3366 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) {
3367 // Allow either -0.0 or 0.0
3368 if (CFP->getValue() == 0.0) {
3369 // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs
3370 if ((CC == ISD::SETGE || CC == ISD::SETGT) &&
3371 N0 == N2 && N3.getOpcode() == ISD::FNEG &&
3372 N2 == N3.getOperand(0))
3373 return DAG.getNode(ISD::FABS, VT, N0);
3375 // select (setl[te] X, +/-0.0), fneg(X), X -> fabs
3376 if ((CC == ISD::SETLT || CC == ISD::SETLE) &&
3377 N0 == N3 && N2.getOpcode() == ISD::FNEG &&
3378 N2.getOperand(0) == N3)
3379 return DAG.getNode(ISD::FABS, VT, N3);
3383 // Check to see if we can perform the "gzip trick", transforming
3384 // select_cc setlt X, 0, A, 0 -> and (sra X, size(X)-1), A
3385 if (N1C && N3C && N3C->isNullValue() && CC == ISD::SETLT &&
3386 MVT::isInteger(N0.getValueType()) &&
3387 MVT::isInteger(N2.getValueType()) &&
3388 (N1C->isNullValue() || // (a < 0) ? b : 0
3389 (N1C->getValue() == 1 && N0 == N2))) { // (a < 1) ? a : 0
3390 MVT::ValueType XType = N0.getValueType();
3391 MVT::ValueType AType = N2.getValueType();
3392 if (XType >= AType) {
3393 // and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a
3394 // single-bit constant.
3395 if (N2C && ((N2C->getValue() & (N2C->getValue()-1)) == 0)) {
3396 unsigned ShCtV = Log2_64(N2C->getValue());
3397 ShCtV = MVT::getSizeInBits(XType)-ShCtV-1;
3398 SDOperand ShCt = DAG.getConstant(ShCtV, TLI.getShiftAmountTy());
3399 SDOperand Shift = DAG.getNode(ISD::SRL, XType, N0, ShCt);
3400 AddToWorkList(Shift.Val);
3401 if (XType > AType) {
3402 Shift = DAG.getNode(ISD::TRUNCATE, AType, Shift);
3403 AddToWorkList(Shift.Val);
3405 return DAG.getNode(ISD::AND, AType, Shift, N2);
3407 SDOperand Shift = DAG.getNode(ISD::SRA, XType, N0,
3408 DAG.getConstant(MVT::getSizeInBits(XType)-1,
3409 TLI.getShiftAmountTy()));
3410 AddToWorkList(Shift.Val);
3411 if (XType > AType) {
3412 Shift = DAG.getNode(ISD::TRUNCATE, AType, Shift);
3413 AddToWorkList(Shift.Val);
3415 return DAG.getNode(ISD::AND, AType, Shift, N2);
3419 // fold select C, 16, 0 -> shl C, 4
3420 if (N2C && N3C && N3C->isNullValue() && isPowerOf2_64(N2C->getValue()) &&
3421 TLI.getSetCCResultContents() == TargetLowering::ZeroOrOneSetCCResult) {
3422 // Get a SetCC of the condition
3423 // FIXME: Should probably make sure that setcc is legal if we ever have a
3424 // target where it isn't.
3425 SDOperand Temp, SCC;
3426 // cast from setcc result type to select result type
3427 if (AfterLegalize) {
3428 SCC = DAG.getSetCC(TLI.getSetCCResultTy(), N0, N1, CC);
3429 Temp = DAG.getZeroExtendInReg(SCC, N2.getValueType());
3431 SCC = DAG.getSetCC(MVT::i1, N0, N1, CC);
3432 Temp = DAG.getNode(ISD::ZERO_EXTEND, N2.getValueType(), SCC);
3434 AddToWorkList(SCC.Val);
3435 AddToWorkList(Temp.Val);
3436 // shl setcc result by log2 n2c
3437 return DAG.getNode(ISD::SHL, N2.getValueType(), Temp,
3438 DAG.getConstant(Log2_64(N2C->getValue()),
3439 TLI.getShiftAmountTy()));
3442 // Check to see if this is the equivalent of setcc
3443 // FIXME: Turn all of these into setcc if setcc if setcc is legal
3444 // otherwise, go ahead with the folds.
3445 if (0 && N3C && N3C->isNullValue() && N2C && (N2C->getValue() == 1ULL)) {
3446 MVT::ValueType XType = N0.getValueType();
3447 if (TLI.isOperationLegal(ISD::SETCC, TLI.getSetCCResultTy())) {
3448 SDOperand Res = DAG.getSetCC(TLI.getSetCCResultTy(), N0, N1, CC);
3449 if (Res.getValueType() != VT)
3450 Res = DAG.getNode(ISD::ZERO_EXTEND, VT, Res);
3454 // seteq X, 0 -> srl (ctlz X, log2(size(X)))
3455 if (N1C && N1C->isNullValue() && CC == ISD::SETEQ &&
3456 TLI.isOperationLegal(ISD::CTLZ, XType)) {
3457 SDOperand Ctlz = DAG.getNode(ISD::CTLZ, XType, N0);
3458 return DAG.getNode(ISD::SRL, XType, Ctlz,
3459 DAG.getConstant(Log2_32(MVT::getSizeInBits(XType)),
3460 TLI.getShiftAmountTy()));
3462 // setgt X, 0 -> srl (and (-X, ~X), size(X)-1)
3463 if (N1C && N1C->isNullValue() && CC == ISD::SETGT) {
3464 SDOperand NegN0 = DAG.getNode(ISD::SUB, XType, DAG.getConstant(0, XType),
3466 SDOperand NotN0 = DAG.getNode(ISD::XOR, XType, N0,
3467 DAG.getConstant(~0ULL, XType));
3468 return DAG.getNode(ISD::SRL, XType,
3469 DAG.getNode(ISD::AND, XType, NegN0, NotN0),
3470 DAG.getConstant(MVT::getSizeInBits(XType)-1,
3471 TLI.getShiftAmountTy()));
3473 // setgt X, -1 -> xor (srl (X, size(X)-1), 1)
3474 if (N1C && N1C->isAllOnesValue() && CC == ISD::SETGT) {
3475 SDOperand Sign = DAG.getNode(ISD::SRL, XType, N0,
3476 DAG.getConstant(MVT::getSizeInBits(XType)-1,
3477 TLI.getShiftAmountTy()));
3478 return DAG.getNode(ISD::XOR, XType, Sign, DAG.getConstant(1, XType));
3482 // Check to see if this is an integer abs. select_cc setl[te] X, 0, -X, X ->
3483 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
3484 if (N1C && N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE) &&
3485 N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1)) {
3486 if (ConstantSDNode *SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0))) {
3487 MVT::ValueType XType = N0.getValueType();
3488 if (SubC->isNullValue() && MVT::isInteger(XType)) {
3489 SDOperand Shift = DAG.getNode(ISD::SRA, XType, N0,
3490 DAG.getConstant(MVT::getSizeInBits(XType)-1,
3491 TLI.getShiftAmountTy()));
3492 SDOperand Add = DAG.getNode(ISD::ADD, XType, N0, Shift);
3493 AddToWorkList(Shift.Val);
3494 AddToWorkList(Add.Val);
3495 return DAG.getNode(ISD::XOR, XType, Add, Shift);
3503 SDOperand DAGCombiner::SimplifySetCC(MVT::ValueType VT, SDOperand N0,
3504 SDOperand N1, ISD::CondCode Cond,
3505 bool foldBooleans) {
3506 // These setcc operations always fold.
3510 case ISD::SETFALSE2: return DAG.getConstant(0, VT);
3512 case ISD::SETTRUE2: return DAG.getConstant(1, VT);
3515 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val)) {
3516 uint64_t C1 = N1C->getValue();
3517 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val)) {
3518 uint64_t C0 = N0C->getValue();
3520 // Sign extend the operands if required
3521 if (ISD::isSignedIntSetCC(Cond)) {
3522 C0 = N0C->getSignExtended();
3523 C1 = N1C->getSignExtended();
3527 default: assert(0 && "Unknown integer setcc!");
3528 case ISD::SETEQ: return DAG.getConstant(C0 == C1, VT);
3529 case ISD::SETNE: return DAG.getConstant(C0 != C1, VT);
3530 case ISD::SETULT: return DAG.getConstant(C0 < C1, VT);
3531 case ISD::SETUGT: return DAG.getConstant(C0 > C1, VT);
3532 case ISD::SETULE: return DAG.getConstant(C0 <= C1, VT);
3533 case ISD::SETUGE: return DAG.getConstant(C0 >= C1, VT);
3534 case ISD::SETLT: return DAG.getConstant((int64_t)C0 < (int64_t)C1, VT);
3535 case ISD::SETGT: return DAG.getConstant((int64_t)C0 > (int64_t)C1, VT);
3536 case ISD::SETLE: return DAG.getConstant((int64_t)C0 <= (int64_t)C1, VT);
3537 case ISD::SETGE: return DAG.getConstant((int64_t)C0 >= (int64_t)C1, VT);
3540 // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an
3541 // equality comparison, then we're just comparing whether X itself is
3543 if (N0.getOpcode() == ISD::SRL && (C1 == 0 || C1 == 1) &&
3544 N0.getOperand(0).getOpcode() == ISD::CTLZ &&
3545 N0.getOperand(1).getOpcode() == ISD::Constant) {
3546 unsigned ShAmt = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
3547 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
3548 ShAmt == Log2_32(MVT::getSizeInBits(N0.getValueType()))) {
3549 if ((C1 == 0) == (Cond == ISD::SETEQ)) {
3550 // (srl (ctlz x), 5) == 0 -> X != 0
3551 // (srl (ctlz x), 5) != 1 -> X != 0
3554 // (srl (ctlz x), 5) != 0 -> X == 0
3555 // (srl (ctlz x), 5) == 1 -> X == 0
3558 SDOperand Zero = DAG.getConstant(0, N0.getValueType());
3559 return DAG.getSetCC(VT, N0.getOperand(0).getOperand(0),
3564 // If the LHS is a ZERO_EXTEND, perform the comparison on the input.
3565 if (N0.getOpcode() == ISD::ZERO_EXTEND) {
3566 unsigned InSize = MVT::getSizeInBits(N0.getOperand(0).getValueType());
3568 // If the comparison constant has bits in the upper part, the
3569 // zero-extended value could never match.
3570 if (C1 & (~0ULL << InSize)) {
3571 unsigned VSize = MVT::getSizeInBits(N0.getValueType());
3575 case ISD::SETEQ: return DAG.getConstant(0, VT);
3578 case ISD::SETNE: return DAG.getConstant(1, VT);
3581 // True if the sign bit of C1 is set.
3582 return DAG.getConstant((C1 & (1ULL << VSize)) != 0, VT);
3585 // True if the sign bit of C1 isn't set.
3586 return DAG.getConstant((C1 & (1ULL << VSize)) == 0, VT);
3592 // Otherwise, we can perform the comparison with the low bits.
3600 return DAG.getSetCC(VT, N0.getOperand(0),
3601 DAG.getConstant(C1, N0.getOperand(0).getValueType()),
3604 break; // todo, be more careful with signed comparisons
3606 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
3607 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
3608 MVT::ValueType ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT();
3609 unsigned ExtSrcTyBits = MVT::getSizeInBits(ExtSrcTy);
3610 MVT::ValueType ExtDstTy = N0.getValueType();
3611 unsigned ExtDstTyBits = MVT::getSizeInBits(ExtDstTy);
3613 // If the extended part has any inconsistent bits, it cannot ever
3614 // compare equal. In other words, they have to be all ones or all
3617 (~0ULL >> (64-ExtSrcTyBits)) & (~0ULL << (ExtDstTyBits-1));
3618 if ((C1 & ExtBits) != 0 && (C1 & ExtBits) != ExtBits)
3619 return DAG.getConstant(Cond == ISD::SETNE, VT);
3622 MVT::ValueType Op0Ty = N0.getOperand(0).getValueType();
3623 if (Op0Ty == ExtSrcTy) {
3624 ZextOp = N0.getOperand(0);
3626 int64_t Imm = ~0ULL >> (64-ExtSrcTyBits);
3627 ZextOp = DAG.getNode(ISD::AND, Op0Ty, N0.getOperand(0),
3628 DAG.getConstant(Imm, Op0Ty));
3630 AddToWorkList(ZextOp.Val);
3631 // Otherwise, make this a use of a zext.
3632 return DAG.getSetCC(VT, ZextOp,
3633 DAG.getConstant(C1 & (~0ULL>>(64-ExtSrcTyBits)),
3636 } else if ((N1C->getValue() == 0 || N1C->getValue() == 1) &&
3637 (Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
3638 (N0.getOpcode() == ISD::XOR ||
3639 (N0.getOpcode() == ISD::AND &&
3640 N0.getOperand(0).getOpcode() == ISD::XOR &&
3641 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) &&
3642 isa<ConstantSDNode>(N0.getOperand(1)) &&
3643 cast<ConstantSDNode>(N0.getOperand(1))->getValue() == 1) {
3644 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We can
3645 // only do this if the top bits are known zero.
3646 if (TLI.MaskedValueIsZero(N1,
3647 MVT::getIntVTBitMask(N0.getValueType())-1)) {
3648 // Okay, get the un-inverted input value.
3650 if (N0.getOpcode() == ISD::XOR)
3651 Val = N0.getOperand(0);
3653 assert(N0.getOpcode() == ISD::AND &&
3654 N0.getOperand(0).getOpcode() == ISD::XOR);
3655 // ((X^1)&1)^1 -> X & 1
3656 Val = DAG.getNode(ISD::AND, N0.getValueType(),
3657 N0.getOperand(0).getOperand(0), N0.getOperand(1));
3659 return DAG.getSetCC(VT, Val, N1,
3660 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
3664 uint64_t MinVal, MaxVal;
3665 unsigned OperandBitSize = MVT::getSizeInBits(N1C->getValueType(0));
3666 if (ISD::isSignedIntSetCC(Cond)) {
3667 MinVal = 1ULL << (OperandBitSize-1);
3668 if (OperandBitSize != 1) // Avoid X >> 64, which is undefined.
3669 MaxVal = ~0ULL >> (65-OperandBitSize);
3674 MaxVal = ~0ULL >> (64-OperandBitSize);
3677 // Canonicalize GE/LE comparisons to use GT/LT comparisons.
3678 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) {
3679 if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true
3680 --C1; // X >= C0 --> X > (C0-1)
3681 return DAG.getSetCC(VT, N0, DAG.getConstant(C1, N1.getValueType()),
3682 (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT);
3685 if (Cond == ISD::SETLE || Cond == ISD::SETULE) {
3686 if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true
3687 ++C1; // X <= C0 --> X < (C0+1)
3688 return DAG.getSetCC(VT, N0, DAG.getConstant(C1, N1.getValueType()),
3689 (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT);
3692 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal)
3693 return DAG.getConstant(0, VT); // X < MIN --> false
3695 // Canonicalize setgt X, Min --> setne X, Min
3696 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal)
3697 return DAG.getSetCC(VT, N0, N1, ISD::SETNE);
3698 // Canonicalize setlt X, Max --> setne X, Max
3699 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal)
3700 return DAG.getSetCC(VT, N0, N1, ISD::SETNE);
3702 // If we have setult X, 1, turn it into seteq X, 0
3703 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1)
3704 return DAG.getSetCC(VT, N0, DAG.getConstant(MinVal, N0.getValueType()),
3706 // If we have setugt X, Max-1, turn it into seteq X, Max
3707 else if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1)
3708 return DAG.getSetCC(VT, N0, DAG.getConstant(MaxVal, N0.getValueType()),
3711 // If we have "setcc X, C0", check to see if we can shrink the immediate
3714 // SETUGT X, SINTMAX -> SETLT X, 0
3715 if (Cond == ISD::SETUGT && OperandBitSize != 1 &&
3716 C1 == (~0ULL >> (65-OperandBitSize)))
3717 return DAG.getSetCC(VT, N0, DAG.getConstant(0, N1.getValueType()),
3720 // FIXME: Implement the rest of these.
3722 // Fold bit comparisons when we can.
3723 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
3724 VT == N0.getValueType() && N0.getOpcode() == ISD::AND)
3725 if (ConstantSDNode *AndRHS =
3726 dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
3727 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3
3728 // Perform the xform if the AND RHS is a single bit.
3729 if ((AndRHS->getValue() & (AndRHS->getValue()-1)) == 0) {
3730 return DAG.getNode(ISD::SRL, VT, N0,
3731 DAG.getConstant(Log2_64(AndRHS->getValue()),
3732 TLI.getShiftAmountTy()));
3734 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getValue()) {
3735 // (X & 8) == 8 --> (X & 8) >> 3
3736 // Perform the xform if C1 is a single bit.
3737 if ((C1 & (C1-1)) == 0) {
3738 return DAG.getNode(ISD::SRL, VT, N0,
3739 DAG.getConstant(Log2_64(C1),TLI.getShiftAmountTy()));
3744 } else if (isa<ConstantSDNode>(N0.Val)) {
3745 // Ensure that the constant occurs on the RHS.
3746 return DAG.getSetCC(VT, N1, N0, ISD::getSetCCSwappedOperands(Cond));
3749 if (ConstantFPSDNode *N0C = dyn_cast<ConstantFPSDNode>(N0.Val))
3750 if (ConstantFPSDNode *N1C = dyn_cast<ConstantFPSDNode>(N1.Val)) {
3751 double C0 = N0C->getValue(), C1 = N1C->getValue();
3754 default: break; // FIXME: Implement the rest of these!
3755 case ISD::SETEQ: return DAG.getConstant(C0 == C1, VT);
3756 case ISD::SETNE: return DAG.getConstant(C0 != C1, VT);
3757 case ISD::SETLT: return DAG.getConstant(C0 < C1, VT);
3758 case ISD::SETGT: return DAG.getConstant(C0 > C1, VT);
3759 case ISD::SETLE: return DAG.getConstant(C0 <= C1, VT);
3760 case ISD::SETGE: return DAG.getConstant(C0 >= C1, VT);
3763 // Ensure that the constant occurs on the RHS.
3764 return DAG.getSetCC(VT, N1, N0, ISD::getSetCCSwappedOperands(Cond));
3768 // We can always fold X == Y for integer setcc's.
3769 if (MVT::isInteger(N0.getValueType()))
3770 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
3771 unsigned UOF = ISD::getUnorderedFlavor(Cond);
3772 if (UOF == 2) // FP operators that are undefined on NaNs.
3773 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
3774 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond)))
3775 return DAG.getConstant(UOF, VT);
3776 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO
3777 // if it is not already.
3778 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO;
3779 if (NewCond != Cond)
3780 return DAG.getSetCC(VT, N0, N1, NewCond);
3783 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
3784 MVT::isInteger(N0.getValueType())) {
3785 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB ||
3786 N0.getOpcode() == ISD::XOR) {
3787 // Simplify (X+Y) == (X+Z) --> Y == Z
3788 if (N0.getOpcode() == N1.getOpcode()) {
3789 if (N0.getOperand(0) == N1.getOperand(0))
3790 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(1), Cond);
3791 if (N0.getOperand(1) == N1.getOperand(1))
3792 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(0), Cond);
3793 if (DAG.isCommutativeBinOp(N0.getOpcode())) {
3794 // If X op Y == Y op X, try other combinations.
3795 if (N0.getOperand(0) == N1.getOperand(1))
3796 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(0), Cond);
3797 if (N0.getOperand(1) == N1.getOperand(0))
3798 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(1), Cond);
3802 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) {
3803 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
3804 // Turn (X+C1) == C2 --> X == C2-C1
3805 if (N0.getOpcode() == ISD::ADD && N0.Val->hasOneUse()) {
3806 return DAG.getSetCC(VT, N0.getOperand(0),
3807 DAG.getConstant(RHSC->getValue()-LHSR->getValue(),
3808 N0.getValueType()), Cond);
3811 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0.
3812 if (N0.getOpcode() == ISD::XOR)
3813 // If we know that all of the inverted bits are zero, don't bother
3814 // performing the inversion.
3815 if (TLI.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getValue()))
3816 return DAG.getSetCC(VT, N0.getOperand(0),
3817 DAG.getConstant(LHSR->getValue()^RHSC->getValue(),
3818 N0.getValueType()), Cond);
3821 // Turn (C1-X) == C2 --> X == C1-C2
3822 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) {
3823 if (N0.getOpcode() == ISD::SUB && N0.Val->hasOneUse()) {
3824 return DAG.getSetCC(VT, N0.getOperand(1),
3825 DAG.getConstant(SUBC->getValue()-RHSC->getValue(),
3826 N0.getValueType()), Cond);
3831 // Simplify (X+Z) == X --> Z == 0
3832 if (N0.getOperand(0) == N1)
3833 return DAG.getSetCC(VT, N0.getOperand(1),
3834 DAG.getConstant(0, N0.getValueType()), Cond);
3835 if (N0.getOperand(1) == N1) {
3836 if (DAG.isCommutativeBinOp(N0.getOpcode()))
3837 return DAG.getSetCC(VT, N0.getOperand(0),
3838 DAG.getConstant(0, N0.getValueType()), Cond);
3840 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!");
3841 // (Z-X) == X --> Z == X<<1
3842 SDOperand SH = DAG.getNode(ISD::SHL, N1.getValueType(),
3844 DAG.getConstant(1,TLI.getShiftAmountTy()));
3845 AddToWorkList(SH.Val);
3846 return DAG.getSetCC(VT, N0.getOperand(0), SH, Cond);
3851 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB ||
3852 N1.getOpcode() == ISD::XOR) {
3853 // Simplify X == (X+Z) --> Z == 0
3854 if (N1.getOperand(0) == N0) {
3855 return DAG.getSetCC(VT, N1.getOperand(1),
3856 DAG.getConstant(0, N1.getValueType()), Cond);
3857 } else if (N1.getOperand(1) == N0) {
3858 if (DAG.isCommutativeBinOp(N1.getOpcode())) {
3859 return DAG.getSetCC(VT, N1.getOperand(0),
3860 DAG.getConstant(0, N1.getValueType()), Cond);
3862 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!");
3863 // X == (Z-X) --> X<<1 == Z
3864 SDOperand SH = DAG.getNode(ISD::SHL, N1.getValueType(), N0,
3865 DAG.getConstant(1,TLI.getShiftAmountTy()));
3866 AddToWorkList(SH.Val);
3867 return DAG.getSetCC(VT, SH, N1.getOperand(0), Cond);
3873 // Fold away ALL boolean setcc's.
3875 if (N0.getValueType() == MVT::i1 && foldBooleans) {
3877 default: assert(0 && "Unknown integer setcc!");
3878 case ISD::SETEQ: // X == Y -> (X^Y)^1
3879 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, N1);
3880 N0 = DAG.getNode(ISD::XOR, MVT::i1, Temp, DAG.getConstant(1, MVT::i1));
3881 AddToWorkList(Temp.Val);
3883 case ISD::SETNE: // X != Y --> (X^Y)
3884 N0 = DAG.getNode(ISD::XOR, MVT::i1, N0, N1);
3886 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> X^1 & Y
3887 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> X^1 & Y
3888 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1));
3889 N0 = DAG.getNode(ISD::AND, MVT::i1, N1, Temp);
3890 AddToWorkList(Temp.Val);
3892 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> Y^1 & X
3893 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> Y^1 & X
3894 Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1));
3895 N0 = DAG.getNode(ISD::AND, MVT::i1, N0, Temp);
3896 AddToWorkList(Temp.Val);
3898 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> X^1 | Y
3899 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> X^1 | Y
3900 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1));
3901 N0 = DAG.getNode(ISD::OR, MVT::i1, N1, Temp);
3902 AddToWorkList(Temp.Val);
3904 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> Y^1 | X
3905 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> Y^1 | X
3906 Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1));
3907 N0 = DAG.getNode(ISD::OR, MVT::i1, N0, Temp);
3910 if (VT != MVT::i1) {
3911 AddToWorkList(N0.Val);
3912 // FIXME: If running after legalize, we probably can't do this.
3913 N0 = DAG.getNode(ISD::ZERO_EXTEND, VT, N0);
3918 // Could not fold it.
3922 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
3923 /// return a DAG expression to select that will generate the same value by
3924 /// multiplying by a magic number. See:
3925 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
3926 SDOperand DAGCombiner::BuildSDIV(SDNode *N) {
3927 std::vector<SDNode*> Built;
3928 SDOperand S = TLI.BuildSDIV(N, DAG, &Built);
3930 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end();
3936 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
3937 /// return a DAG expression to select that will generate the same value by
3938 /// multiplying by a magic number. See:
3939 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
3940 SDOperand DAGCombiner::BuildUDIV(SDNode *N) {
3941 std::vector<SDNode*> Built;
3942 SDOperand S = TLI.BuildUDIV(N, DAG, &Built);
3944 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end();
3950 /// FindBaseOffset - Return true if we can determine base and offset information
3951 /// from a given pointer operand. Provides base and offset as a result.
3952 bool DAGCombiner::FindBaseOffset(SDOperand Ptr,
3953 SDOperand &Object, int64_t &Offset) {
3955 // Is it a frame variable, global or constant.
3956 if (isa<FrameIndexSDNode>(Ptr) ||
3957 isa<ConstantPoolSDNode>(Ptr) ||
3958 isa<GlobalAddressSDNode>(Ptr)) {
3959 Object = Ptr; Offset = 0;
3961 } else if (Ptr.getOpcode() == ISD::ADD &&
3962 FindBaseOffset(Ptr.getOperand(0), Object, Offset)) {
3963 // If it's an add of an simple constant then include it in the offset.
3964 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Ptr.getOperand(1))) {
3965 Offset += C->getValue();
3973 /// isAlias - Return true if there is the possibility that the two addresses
3975 bool DAGCombiner::isAlias(SDOperand Ptr1, int64_t Size1,
3976 SDOperand SrcValue1,
3977 SDOperand Ptr2, int64_t Size2,
3978 SDOperand SrcValue2) {
3979 // If they are the same then they must be aliases.
3980 if (Ptr1 == Ptr2) return true;
3982 // Gather base offset information. Objects can be frame variables, globals
3984 SDOperand Object1, Object2;
3985 int64_t Offset1, Offset2;
3986 if (FindBaseOffset(Ptr1, Object1, Offset1) &&
3987 FindBaseOffset(Ptr2, Object2, Offset2)) {
3988 // If they have a different base address, then they can't alias.
3989 if (Object1 != Object2) return false;
3991 // Check to see if the addresses overlap.
3992 if ((Offset1 + Size1) <= Offset2 || (Offset2 + Size2) <= Offset1)
3996 // Otherwise we don't know and have to play it safe.
4000 /// FindAliasInfo - Extracts the relevant alias information from the memory
4002 void DAGCombiner::FindAliasInfo(SDNode *N,
4003 SDOperand &Ptr, int64_t &Size, SDOperand &SrcValue) {
4004 switch (N->getOpcode()) {
4006 Ptr = N->getOperand(1);
4007 Size = MVT::getSizeInBits(N->getOperand(1).getValueType()) >> 3;
4008 SrcValue = N->getOperand(2);
4011 Ptr = N->getOperand(2);
4012 Size = MVT::getSizeInBits(N->getOperand(1).getValueType()) >> 3;
4013 SrcValue = N->getOperand(3);
4016 assert(0 && "getAliasInfo expected a memory op");
4020 /// hasChain - Return true if Op has a chain. Provides chain if present.
4022 bool DAGCombiner::hasChain(SDOperand Op, SDOperand &Chain) {
4023 if (Op.getNumOperands() == 0) return false;
4024 Chain = Op.getOperand(0);
4025 return Chain.getValueType() == MVT::Other;
4028 /// FindBetterChain - Walk up chain skipping non-aliasing memory nodes, looking
4029 /// for a better chain.
4030 SDOperand DAGCombiner::FindBetterChain(SDNode *N, SDOperand Chain) {
4031 // Get alias information for node.
4035 FindAliasInfo(N, Ptr, Size, SrcValue);
4037 // While we don't encounter any aliasing memory nodes walk up chain.
4039 switch (Chain.getOpcode()) {
4040 case ISD::EntryToken:
4041 // Entry token is ideal chain operand.
4045 // Get alias information for chain.
4048 SDOperand ChainSrcValue;
4049 FindAliasInfo(Chain.Val, ChainPtr, ChainSize, ChainSrcValue);
4051 // If chain is alias then stop here, otherwise continue up chain.
4052 if (isAlias(Ptr, Size, SrcValue, ChainPtr, ChainSize, ChainSrcValue))
4055 Chain = Chain.getOperand(0);
4059 case ISD::TokenFactor: {
4060 // Continue up each of token factor operand and accumulate results in
4061 // a new token factor. CSE will handle duplicate elimination.
4062 SmallVector<SDOperand, 8> Ops; // Ops for replacing token factor.
4063 bool Change = false;
4065 // For each token factor operand.
4066 for (unsigned i = 0, e = Chain.getNumOperands(); i != e; ++i) {
4067 SDOperand Op = Chain.getOperand(i);
4068 SDOperand OpChain = FindBetterChain(N, Op);
4070 // Make sure we don't duplicate an operand.
4071 if (OpChain.getOpcode() != ISD::EntryToken &&
4072 std::find(Ops.begin(), Ops.end(), OpChain) == Ops.end()) {
4073 Ops.push_back(OpChain);
4076 // If we added a new operand.
4077 Change = Change || Op != OpChain;
4080 // If we have new operands.
4082 // Create a specialized token factor for this chain. getNode CSE will
4083 // handle duplicates. If it's a single operand, getNode will just
4084 // return the opernand instead of a new token factor.
4085 return DAG.getNode(ISD::TokenFactor, MVT::Other, &Ops[0], Ops.size());
4088 // Leave things alone.
4091 // For all other instructions we will just have to take what we can get.
4092 default: return Chain;
4099 // SelectionDAG::Combine - This is the entry point for the file.
4101 void SelectionDAG::Combine(bool RunningAfterLegalize) {
4102 /// run - This is the main entry point to this class.
4104 DAGCombiner(*this).Run(RunningAfterLegalize);