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: make truncate see through SIGN_EXTEND and AND
26 // FIXME: divide by zero is currently left unfolded. do we want to turn this
28 // FIXME: select ne (select cc, 1, 0), 0, true, false -> select cc, true, false
30 //===----------------------------------------------------------------------===//
32 #define DEBUG_TYPE "dagcombine"
33 #include "llvm/ADT/Statistic.h"
34 #include "llvm/CodeGen/SelectionDAG.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/MathExtras.h"
37 #include "llvm/Target/TargetLowering.h"
38 #include "llvm/Support/Compiler.h"
45 static Statistic<> NodesCombined ("dagcombiner",
46 "Number of dag nodes combined");
48 class VISIBILITY_HIDDEN DAGCombiner {
53 // Worklist of all of the nodes that need to be simplified.
54 std::vector<SDNode*> WorkList;
56 /// AddUsersToWorkList - When an instruction is simplified, add all users of
57 /// the instruction to the work lists because they might get more simplified
60 void AddUsersToWorkList(SDNode *N) {
61 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
63 WorkList.push_back(*UI);
66 /// removeFromWorkList - remove all instances of N from the worklist.
68 void removeFromWorkList(SDNode *N) {
69 WorkList.erase(std::remove(WorkList.begin(), WorkList.end(), N),
74 void AddToWorkList(SDNode *N) {
75 WorkList.push_back(N);
78 SDOperand CombineTo(SDNode *N, const SDOperand *To, unsigned NumTo) {
79 assert(N->getNumValues() == NumTo && "Broken CombineTo call!");
81 DEBUG(std::cerr << "\nReplacing "; N->dump();
82 std::cerr << "\nWith: "; To[0].Val->dump(&DAG);
83 std::cerr << " and " << NumTo-1 << " other values\n");
84 std::vector<SDNode*> NowDead;
85 DAG.ReplaceAllUsesWith(N, To, &NowDead);
87 // Push the new nodes and any users onto the worklist
88 for (unsigned i = 0, e = NumTo; i != e; ++i) {
89 WorkList.push_back(To[i].Val);
90 AddUsersToWorkList(To[i].Val);
93 // Nodes can end up on the worklist more than once. Make sure we do
94 // not process a node that has been replaced.
95 removeFromWorkList(N);
96 for (unsigned i = 0, e = NowDead.size(); i != e; ++i)
97 removeFromWorkList(NowDead[i]);
99 // Finally, since the node is now dead, remove it from the graph.
101 return SDOperand(N, 0);
104 SDOperand CombineTo(SDNode *N, SDOperand Res) {
105 return CombineTo(N, &Res, 1);
108 SDOperand CombineTo(SDNode *N, SDOperand Res0, SDOperand Res1) {
109 SDOperand To[] = { Res0, Res1 };
110 return CombineTo(N, To, 2);
114 /// SimplifyDemandedBits - Check the specified integer node value to see if
115 /// it can be simplified or if things it uses can be simplified by bit
116 /// propagation. If so, return true.
117 bool SimplifyDemandedBits(SDOperand Op) {
118 TargetLowering::TargetLoweringOpt TLO(DAG);
119 uint64_t KnownZero, KnownOne;
120 uint64_t Demanded = MVT::getIntVTBitMask(Op.getValueType());
121 if (!TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
125 WorkList.push_back(Op.Val);
127 // Replace the old value with the new one.
129 DEBUG(std::cerr << "\nReplacing "; TLO.Old.Val->dump();
130 std::cerr << "\nWith: "; TLO.New.Val->dump(&DAG));
132 std::vector<SDNode*> NowDead;
133 DAG.ReplaceAllUsesOfValueWith(TLO.Old, TLO.New, NowDead);
135 // Push the new node and any (possibly new) users onto the worklist.
136 WorkList.push_back(TLO.New.Val);
137 AddUsersToWorkList(TLO.New.Val);
139 // Nodes can end up on the worklist more than once. Make sure we do
140 // not process a node that has been replaced.
141 for (unsigned i = 0, e = NowDead.size(); i != e; ++i)
142 removeFromWorkList(NowDead[i]);
144 // Finally, if the node is now dead, remove it from the graph. The node
145 // may not be dead if the replacement process recursively simplified to
146 // something else needing this node.
147 if (TLO.Old.Val->use_empty()) {
148 removeFromWorkList(TLO.Old.Val);
149 DAG.DeleteNode(TLO.Old.Val);
154 /// visit - call the node-specific routine that knows how to fold each
155 /// particular type of node.
156 SDOperand visit(SDNode *N);
158 // Visitation implementation - Implement dag node combining for different
159 // node types. The semantics are as follows:
161 // SDOperand.Val == 0 - No change was made
162 // SDOperand.Val == N - N was replaced, is dead, and is already handled.
163 // otherwise - N should be replaced by the returned Operand.
165 SDOperand visitTokenFactor(SDNode *N);
166 SDOperand visitADD(SDNode *N);
167 SDOperand visitSUB(SDNode *N);
168 SDOperand visitMUL(SDNode *N);
169 SDOperand visitSDIV(SDNode *N);
170 SDOperand visitUDIV(SDNode *N);
171 SDOperand visitSREM(SDNode *N);
172 SDOperand visitUREM(SDNode *N);
173 SDOperand visitMULHU(SDNode *N);
174 SDOperand visitMULHS(SDNode *N);
175 SDOperand visitAND(SDNode *N);
176 SDOperand visitOR(SDNode *N);
177 SDOperand visitXOR(SDNode *N);
178 SDOperand visitVBinOp(SDNode *N, ISD::NodeType IntOp, ISD::NodeType FPOp);
179 SDOperand visitSHL(SDNode *N);
180 SDOperand visitSRA(SDNode *N);
181 SDOperand visitSRL(SDNode *N);
182 SDOperand visitCTLZ(SDNode *N);
183 SDOperand visitCTTZ(SDNode *N);
184 SDOperand visitCTPOP(SDNode *N);
185 SDOperand visitSELECT(SDNode *N);
186 SDOperand visitSELECT_CC(SDNode *N);
187 SDOperand visitSETCC(SDNode *N);
188 SDOperand visitSIGN_EXTEND(SDNode *N);
189 SDOperand visitZERO_EXTEND(SDNode *N);
190 SDOperand visitANY_EXTEND(SDNode *N);
191 SDOperand visitSIGN_EXTEND_INREG(SDNode *N);
192 SDOperand visitTRUNCATE(SDNode *N);
193 SDOperand visitBIT_CONVERT(SDNode *N);
194 SDOperand visitVBIT_CONVERT(SDNode *N);
195 SDOperand visitFADD(SDNode *N);
196 SDOperand visitFSUB(SDNode *N);
197 SDOperand visitFMUL(SDNode *N);
198 SDOperand visitFDIV(SDNode *N);
199 SDOperand visitFREM(SDNode *N);
200 SDOperand visitFCOPYSIGN(SDNode *N);
201 SDOperand visitSINT_TO_FP(SDNode *N);
202 SDOperand visitUINT_TO_FP(SDNode *N);
203 SDOperand visitFP_TO_SINT(SDNode *N);
204 SDOperand visitFP_TO_UINT(SDNode *N);
205 SDOperand visitFP_ROUND(SDNode *N);
206 SDOperand visitFP_ROUND_INREG(SDNode *N);
207 SDOperand visitFP_EXTEND(SDNode *N);
208 SDOperand visitFNEG(SDNode *N);
209 SDOperand visitFABS(SDNode *N);
210 SDOperand visitBRCOND(SDNode *N);
211 SDOperand visitBR_CC(SDNode *N);
212 SDOperand visitLOAD(SDNode *N);
213 SDOperand visitXEXTLOAD(SDNode *N);
214 SDOperand visitSTORE(SDNode *N);
215 SDOperand visitINSERT_VECTOR_ELT(SDNode *N);
216 SDOperand visitVINSERT_VECTOR_ELT(SDNode *N);
217 SDOperand visitVBUILD_VECTOR(SDNode *N);
218 SDOperand visitVECTOR_SHUFFLE(SDNode *N);
219 SDOperand visitVVECTOR_SHUFFLE(SDNode *N);
221 SDOperand XformToShuffleWithZero(SDNode *N);
222 SDOperand ReassociateOps(unsigned Opc, SDOperand LHS, SDOperand RHS);
224 bool SimplifySelectOps(SDNode *SELECT, SDOperand LHS, SDOperand RHS);
225 SDOperand SimplifyBinOpWithSameOpcodeHands(SDNode *N);
226 SDOperand SimplifySelect(SDOperand N0, SDOperand N1, SDOperand N2);
227 SDOperand SimplifySelectCC(SDOperand N0, SDOperand N1, SDOperand N2,
228 SDOperand N3, ISD::CondCode CC);
229 SDOperand SimplifySetCC(MVT::ValueType VT, SDOperand N0, SDOperand N1,
230 ISD::CondCode Cond, bool foldBooleans = true);
231 SDOperand ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(SDNode *, MVT::ValueType);
232 SDOperand BuildSDIV(SDNode *N);
233 SDOperand BuildUDIV(SDNode *N);
234 SDNode *MatchRotate(SDOperand LHS, SDOperand RHS);
236 DAGCombiner(SelectionDAG &D)
237 : DAG(D), TLI(D.getTargetLoweringInfo()), AfterLegalize(false) {}
239 /// Run - runs the dag combiner on all nodes in the work list
240 void Run(bool RunningAfterLegalize);
244 //===----------------------------------------------------------------------===//
245 // TargetLowering::DAGCombinerInfo implementation
246 //===----------------------------------------------------------------------===//
248 void TargetLowering::DAGCombinerInfo::AddToWorklist(SDNode *N) {
249 ((DAGCombiner*)DC)->AddToWorkList(N);
252 SDOperand TargetLowering::DAGCombinerInfo::
253 CombineTo(SDNode *N, const std::vector<SDOperand> &To) {
254 return ((DAGCombiner*)DC)->CombineTo(N, &To[0], To.size());
257 SDOperand TargetLowering::DAGCombinerInfo::
258 CombineTo(SDNode *N, SDOperand Res) {
259 return ((DAGCombiner*)DC)->CombineTo(N, Res);
263 SDOperand TargetLowering::DAGCombinerInfo::
264 CombineTo(SDNode *N, SDOperand Res0, SDOperand Res1) {
265 return ((DAGCombiner*)DC)->CombineTo(N, Res0, Res1);
271 //===----------------------------------------------------------------------===//
274 // isSetCCEquivalent - Return true if this node is a setcc, or is a select_cc
275 // that selects between the values 1 and 0, making it equivalent to a setcc.
276 // Also, set the incoming LHS, RHS, and CC references to the appropriate
277 // nodes based on the type of node we are checking. This simplifies life a
278 // bit for the callers.
279 static bool isSetCCEquivalent(SDOperand N, SDOperand &LHS, SDOperand &RHS,
281 if (N.getOpcode() == ISD::SETCC) {
282 LHS = N.getOperand(0);
283 RHS = N.getOperand(1);
284 CC = N.getOperand(2);
287 if (N.getOpcode() == ISD::SELECT_CC &&
288 N.getOperand(2).getOpcode() == ISD::Constant &&
289 N.getOperand(3).getOpcode() == ISD::Constant &&
290 cast<ConstantSDNode>(N.getOperand(2))->getValue() == 1 &&
291 cast<ConstantSDNode>(N.getOperand(3))->isNullValue()) {
292 LHS = N.getOperand(0);
293 RHS = N.getOperand(1);
294 CC = N.getOperand(4);
300 // isOneUseSetCC - Return true if this is a SetCC-equivalent operation with only
301 // one use. If this is true, it allows the users to invert the operation for
302 // free when it is profitable to do so.
303 static bool isOneUseSetCC(SDOperand N) {
304 SDOperand N0, N1, N2;
305 if (isSetCCEquivalent(N, N0, N1, N2) && N.Val->hasOneUse())
310 SDOperand DAGCombiner::ReassociateOps(unsigned Opc, SDOperand N0, SDOperand N1){
311 MVT::ValueType VT = N0.getValueType();
312 // reassoc. (op (op x, c1), y) -> (op (op x, y), c1) iff x+c1 has one use
313 // reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2))
314 if (N0.getOpcode() == Opc && isa<ConstantSDNode>(N0.getOperand(1))) {
315 if (isa<ConstantSDNode>(N1)) {
316 SDOperand OpNode = DAG.getNode(Opc, VT, N0.getOperand(1), N1);
317 AddToWorkList(OpNode.Val);
318 return DAG.getNode(Opc, VT, OpNode, N0.getOperand(0));
319 } else if (N0.hasOneUse()) {
320 SDOperand OpNode = DAG.getNode(Opc, VT, N0.getOperand(0), N1);
321 AddToWorkList(OpNode.Val);
322 return DAG.getNode(Opc, VT, OpNode, N0.getOperand(1));
325 // reassoc. (op y, (op x, c1)) -> (op (op x, y), c1) iff x+c1 has one use
326 // reassoc. (op c2, (op x, c1)) -> (op x, (op c1, c2))
327 if (N1.getOpcode() == Opc && isa<ConstantSDNode>(N1.getOperand(1))) {
328 if (isa<ConstantSDNode>(N0)) {
329 SDOperand OpNode = DAG.getNode(Opc, VT, N1.getOperand(1), N0);
330 AddToWorkList(OpNode.Val);
331 return DAG.getNode(Opc, VT, OpNode, N1.getOperand(0));
332 } else if (N1.hasOneUse()) {
333 SDOperand OpNode = DAG.getNode(Opc, VT, N1.getOperand(0), N0);
334 AddToWorkList(OpNode.Val);
335 return DAG.getNode(Opc, VT, OpNode, N1.getOperand(1));
341 void DAGCombiner::Run(bool RunningAfterLegalize) {
342 // set the instance variable, so that the various visit routines may use it.
343 AfterLegalize = RunningAfterLegalize;
345 // Add all the dag nodes to the worklist.
346 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
347 E = DAG.allnodes_end(); I != E; ++I)
348 WorkList.push_back(I);
350 // Create a dummy node (which is not added to allnodes), that adds a reference
351 // to the root node, preventing it from being deleted, and tracking any
352 // changes of the root.
353 HandleSDNode Dummy(DAG.getRoot());
356 /// DagCombineInfo - Expose the DAG combiner to the target combiner impls.
357 TargetLowering::DAGCombinerInfo
358 DagCombineInfo(DAG, !RunningAfterLegalize, this);
360 // while the worklist isn't empty, inspect the node on the end of it and
361 // try and combine it.
362 while (!WorkList.empty()) {
363 SDNode *N = WorkList.back();
366 // If N has no uses, it is dead. Make sure to revisit all N's operands once
367 // N is deleted from the DAG, since they too may now be dead or may have a
368 // reduced number of uses, allowing other xforms.
369 if (N->use_empty() && N != &Dummy) {
370 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
371 WorkList.push_back(N->getOperand(i).Val);
373 removeFromWorkList(N);
378 SDOperand RV = visit(N);
380 // If nothing happened, try a target-specific DAG combine.
382 assert(N->getOpcode() != ISD::DELETED_NODE &&
383 "Node was deleted but visit returned NULL!");
384 if (N->getOpcode() >= ISD::BUILTIN_OP_END ||
385 TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode()))
386 RV = TLI.PerformDAGCombine(N, DagCombineInfo);
391 // If we get back the same node we passed in, rather than a new node or
392 // zero, we know that the node must have defined multiple values and
393 // CombineTo was used. Since CombineTo takes care of the worklist
394 // mechanics for us, we have no work to do in this case.
396 assert(N->getOpcode() != ISD::DELETED_NODE &&
397 RV.Val->getOpcode() != ISD::DELETED_NODE &&
398 "Node was deleted but visit returned new node!");
400 DEBUG(std::cerr << "\nReplacing "; N->dump();
401 std::cerr << "\nWith: "; RV.Val->dump(&DAG);
403 std::vector<SDNode*> NowDead;
405 DAG.ReplaceAllUsesWith(N, &OpV, &NowDead);
407 // Push the new node and any users onto the worklist
408 WorkList.push_back(RV.Val);
409 AddUsersToWorkList(RV.Val);
411 // Nodes can end up on the worklist more than once. Make sure we do
412 // not process a node that has been replaced.
413 removeFromWorkList(N);
414 for (unsigned i = 0, e = NowDead.size(); i != e; ++i)
415 removeFromWorkList(NowDead[i]);
417 // Finally, since the node is now dead, remove it from the graph.
423 // If the root changed (e.g. it was a dead load, update the root).
424 DAG.setRoot(Dummy.getValue());
427 SDOperand DAGCombiner::visit(SDNode *N) {
428 switch(N->getOpcode()) {
430 case ISD::TokenFactor: return visitTokenFactor(N);
431 case ISD::ADD: return visitADD(N);
432 case ISD::SUB: return visitSUB(N);
433 case ISD::MUL: return visitMUL(N);
434 case ISD::SDIV: return visitSDIV(N);
435 case ISD::UDIV: return visitUDIV(N);
436 case ISD::SREM: return visitSREM(N);
437 case ISD::UREM: return visitUREM(N);
438 case ISD::MULHU: return visitMULHU(N);
439 case ISD::MULHS: return visitMULHS(N);
440 case ISD::AND: return visitAND(N);
441 case ISD::OR: return visitOR(N);
442 case ISD::XOR: return visitXOR(N);
443 case ISD::SHL: return visitSHL(N);
444 case ISD::SRA: return visitSRA(N);
445 case ISD::SRL: return visitSRL(N);
446 case ISD::CTLZ: return visitCTLZ(N);
447 case ISD::CTTZ: return visitCTTZ(N);
448 case ISD::CTPOP: return visitCTPOP(N);
449 case ISD::SELECT: return visitSELECT(N);
450 case ISD::SELECT_CC: return visitSELECT_CC(N);
451 case ISD::SETCC: return visitSETCC(N);
452 case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N);
453 case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N);
454 case ISD::ANY_EXTEND: return visitANY_EXTEND(N);
455 case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N);
456 case ISD::TRUNCATE: return visitTRUNCATE(N);
457 case ISD::BIT_CONVERT: return visitBIT_CONVERT(N);
458 case ISD::VBIT_CONVERT: return visitVBIT_CONVERT(N);
459 case ISD::FADD: return visitFADD(N);
460 case ISD::FSUB: return visitFSUB(N);
461 case ISD::FMUL: return visitFMUL(N);
462 case ISD::FDIV: return visitFDIV(N);
463 case ISD::FREM: return visitFREM(N);
464 case ISD::FCOPYSIGN: return visitFCOPYSIGN(N);
465 case ISD::SINT_TO_FP: return visitSINT_TO_FP(N);
466 case ISD::UINT_TO_FP: return visitUINT_TO_FP(N);
467 case ISD::FP_TO_SINT: return visitFP_TO_SINT(N);
468 case ISD::FP_TO_UINT: return visitFP_TO_UINT(N);
469 case ISD::FP_ROUND: return visitFP_ROUND(N);
470 case ISD::FP_ROUND_INREG: return visitFP_ROUND_INREG(N);
471 case ISD::FP_EXTEND: return visitFP_EXTEND(N);
472 case ISD::FNEG: return visitFNEG(N);
473 case ISD::FABS: return visitFABS(N);
474 case ISD::BRCOND: return visitBRCOND(N);
475 case ISD::BR_CC: return visitBR_CC(N);
476 case ISD::LOAD: return visitLOAD(N);
479 case ISD::ZEXTLOAD: return visitXEXTLOAD(N);
480 case ISD::STORE: return visitSTORE(N);
481 case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N);
482 case ISD::VINSERT_VECTOR_ELT: return visitVINSERT_VECTOR_ELT(N);
483 case ISD::VBUILD_VECTOR: return visitVBUILD_VECTOR(N);
484 case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N);
485 case ISD::VVECTOR_SHUFFLE: return visitVVECTOR_SHUFFLE(N);
486 case ISD::VADD: return visitVBinOp(N, ISD::ADD , ISD::FADD);
487 case ISD::VSUB: return visitVBinOp(N, ISD::SUB , ISD::FSUB);
488 case ISD::VMUL: return visitVBinOp(N, ISD::MUL , ISD::FMUL);
489 case ISD::VSDIV: return visitVBinOp(N, ISD::SDIV, ISD::FDIV);
490 case ISD::VUDIV: return visitVBinOp(N, ISD::UDIV, ISD::UDIV);
491 case ISD::VAND: return visitVBinOp(N, ISD::AND , ISD::AND);
492 case ISD::VOR: return visitVBinOp(N, ISD::OR , ISD::OR);
493 case ISD::VXOR: return visitVBinOp(N, ISD::XOR , ISD::XOR);
498 SDOperand DAGCombiner::visitTokenFactor(SDNode *N) {
499 SmallVector<SDOperand, 8> Ops;
500 bool Changed = false;
502 // If the token factor has two operands and one is the entry token, replace
503 // the token factor with the other operand.
504 if (N->getNumOperands() == 2) {
505 if (N->getOperand(0).getOpcode() == ISD::EntryToken ||
506 N->getOperand(0) == N->getOperand(1))
507 return N->getOperand(1);
508 if (N->getOperand(1).getOpcode() == ISD::EntryToken)
509 return N->getOperand(0);
512 // fold (tokenfactor (tokenfactor)) -> tokenfactor
513 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
514 SDOperand Op = N->getOperand(i);
515 if (Op.getOpcode() == ISD::TokenFactor && Op.hasOneUse()) {
516 AddToWorkList(Op.Val); // Remove dead node.
518 for (unsigned j = 0, e = Op.getNumOperands(); j != e; ++j)
519 Ops.push_back(Op.getOperand(j));
520 } else if (i == 0 || N->getOperand(i) != N->getOperand(i-1)) {
523 // Deleted an operand that was the same as the last one.
528 return DAG.getNode(ISD::TokenFactor, MVT::Other, &Ops[0], Ops.size());
532 SDOperand DAGCombiner::visitADD(SDNode *N) {
533 SDOperand N0 = N->getOperand(0);
534 SDOperand N1 = N->getOperand(1);
535 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
536 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
537 MVT::ValueType VT = N0.getValueType();
539 // fold (add c1, c2) -> c1+c2
541 return DAG.getNode(ISD::ADD, VT, N0, N1);
542 // canonicalize constant to RHS
544 return DAG.getNode(ISD::ADD, VT, N1, N0);
545 // fold (add x, 0) -> x
546 if (N1C && N1C->isNullValue())
548 // fold ((c1-A)+c2) -> (c1+c2)-A
549 if (N1C && N0.getOpcode() == ISD::SUB)
550 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getOperand(0)))
551 return DAG.getNode(ISD::SUB, VT,
552 DAG.getConstant(N1C->getValue()+N0C->getValue(), VT),
555 SDOperand RADD = ReassociateOps(ISD::ADD, N0, N1);
558 // fold ((0-A) + B) -> B-A
559 if (N0.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N0.getOperand(0)) &&
560 cast<ConstantSDNode>(N0.getOperand(0))->isNullValue())
561 return DAG.getNode(ISD::SUB, VT, N1, N0.getOperand(1));
562 // fold (A + (0-B)) -> A-B
563 if (N1.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N1.getOperand(0)) &&
564 cast<ConstantSDNode>(N1.getOperand(0))->isNullValue())
565 return DAG.getNode(ISD::SUB, VT, N0, N1.getOperand(1));
566 // fold (A+(B-A)) -> B
567 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1))
568 return N1.getOperand(0);
570 if (!MVT::isVector(VT) && SimplifyDemandedBits(SDOperand(N, 0)))
571 return SDOperand(N, 0);
573 // fold (a+b) -> (a|b) iff a and b share no bits.
574 if (MVT::isInteger(VT) && !MVT::isVector(VT)) {
575 uint64_t LHSZero, LHSOne;
576 uint64_t RHSZero, RHSOne;
577 uint64_t Mask = MVT::getIntVTBitMask(VT);
578 TLI.ComputeMaskedBits(N0, Mask, LHSZero, LHSOne);
580 TLI.ComputeMaskedBits(N1, Mask, RHSZero, RHSOne);
582 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
583 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
584 if ((RHSZero & (~LHSZero & Mask)) == (~LHSZero & Mask) ||
585 (LHSZero & (~RHSZero & Mask)) == (~RHSZero & Mask))
586 return DAG.getNode(ISD::OR, VT, N0, N1);
593 SDOperand DAGCombiner::visitSUB(SDNode *N) {
594 SDOperand N0 = N->getOperand(0);
595 SDOperand N1 = N->getOperand(1);
596 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val);
597 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val);
598 MVT::ValueType VT = N0.getValueType();
600 // fold (sub x, x) -> 0
602 return DAG.getConstant(0, N->getValueType(0));
603 // fold (sub c1, c2) -> c1-c2
605 return DAG.getNode(ISD::SUB, VT, N0, N1);
606 // fold (sub x, c) -> (add x, -c)
608 return DAG.getNode(ISD::ADD, VT, N0, DAG.getConstant(-N1C->getValue(), VT));
610 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1)
611 return N0.getOperand(1);
613 if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1)
614 return N0.getOperand(0);
618 SDOperand DAGCombiner::visitMUL(SDNode *N) {
619 SDOperand N0 = N->getOperand(0);
620 SDOperand N1 = N->getOperand(1);
621 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
622 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
623 MVT::ValueType VT = N0.getValueType();
625 // fold (mul c1, c2) -> c1*c2
627 return DAG.getNode(ISD::MUL, VT, N0, N1);
628 // canonicalize constant to RHS
630 return DAG.getNode(ISD::MUL, VT, N1, N0);
631 // fold (mul x, 0) -> 0
632 if (N1C && N1C->isNullValue())
634 // fold (mul x, -1) -> 0-x
635 if (N1C && N1C->isAllOnesValue())
636 return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT), N0);
637 // fold (mul x, (1 << c)) -> x << c
638 if (N1C && isPowerOf2_64(N1C->getValue()))
639 return DAG.getNode(ISD::SHL, VT, N0,
640 DAG.getConstant(Log2_64(N1C->getValue()),
641 TLI.getShiftAmountTy()));
642 // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c
643 if (N1C && isPowerOf2_64(-N1C->getSignExtended())) {
644 // FIXME: If the input is something that is easily negated (e.g. a
645 // single-use add), we should put the negate there.
646 return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT),
647 DAG.getNode(ISD::SHL, VT, N0,
648 DAG.getConstant(Log2_64(-N1C->getSignExtended()),
649 TLI.getShiftAmountTy())));
652 // (mul (shl X, c1), c2) -> (mul X, c2 << c1)
653 if (N1C && N0.getOpcode() == ISD::SHL &&
654 isa<ConstantSDNode>(N0.getOperand(1))) {
655 SDOperand C3 = DAG.getNode(ISD::SHL, VT, N1, N0.getOperand(1));
656 AddToWorkList(C3.Val);
657 return DAG.getNode(ISD::MUL, VT, N0.getOperand(0), C3);
660 // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one
663 SDOperand Sh(0,0), Y(0,0);
664 // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)).
665 if (N0.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N0.getOperand(1)) &&
666 N0.Val->hasOneUse()) {
668 } else if (N1.getOpcode() == ISD::SHL &&
669 isa<ConstantSDNode>(N1.getOperand(1)) && N1.Val->hasOneUse()) {
673 SDOperand Mul = DAG.getNode(ISD::MUL, VT, Sh.getOperand(0), Y);
674 return DAG.getNode(ISD::SHL, VT, Mul, Sh.getOperand(1));
677 // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2)
678 if (N1C && N0.getOpcode() == ISD::ADD && N0.Val->hasOneUse() &&
679 isa<ConstantSDNode>(N0.getOperand(1))) {
680 return DAG.getNode(ISD::ADD, VT,
681 DAG.getNode(ISD::MUL, VT, N0.getOperand(0), N1),
682 DAG.getNode(ISD::MUL, VT, N0.getOperand(1), N1));
686 SDOperand RMUL = ReassociateOps(ISD::MUL, N0, N1);
692 SDOperand DAGCombiner::visitSDIV(SDNode *N) {
693 SDOperand N0 = N->getOperand(0);
694 SDOperand N1 = N->getOperand(1);
695 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val);
696 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val);
697 MVT::ValueType VT = N->getValueType(0);
699 // fold (sdiv c1, c2) -> c1/c2
700 if (N0C && N1C && !N1C->isNullValue())
701 return DAG.getNode(ISD::SDIV, VT, N0, N1);
702 // fold (sdiv X, 1) -> X
703 if (N1C && N1C->getSignExtended() == 1LL)
705 // fold (sdiv X, -1) -> 0-X
706 if (N1C && N1C->isAllOnesValue())
707 return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT), N0);
708 // If we know the sign bits of both operands are zero, strength reduce to a
709 // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2
710 uint64_t SignBit = 1ULL << (MVT::getSizeInBits(VT)-1);
711 if (TLI.MaskedValueIsZero(N1, SignBit) &&
712 TLI.MaskedValueIsZero(N0, SignBit))
713 return DAG.getNode(ISD::UDIV, N1.getValueType(), N0, N1);
714 // fold (sdiv X, pow2) -> simple ops after legalize
715 if (N1C && N1C->getValue() && !TLI.isIntDivCheap() &&
716 (isPowerOf2_64(N1C->getSignExtended()) ||
717 isPowerOf2_64(-N1C->getSignExtended()))) {
718 // If dividing by powers of two is cheap, then don't perform the following
720 if (TLI.isPow2DivCheap())
722 int64_t pow2 = N1C->getSignExtended();
723 int64_t abs2 = pow2 > 0 ? pow2 : -pow2;
724 unsigned lg2 = Log2_64(abs2);
725 // Splat the sign bit into the register
726 SDOperand SGN = DAG.getNode(ISD::SRA, VT, N0,
727 DAG.getConstant(MVT::getSizeInBits(VT)-1,
728 TLI.getShiftAmountTy()));
729 AddToWorkList(SGN.Val);
730 // Add (N0 < 0) ? abs2 - 1 : 0;
731 SDOperand SRL = DAG.getNode(ISD::SRL, VT, SGN,
732 DAG.getConstant(MVT::getSizeInBits(VT)-lg2,
733 TLI.getShiftAmountTy()));
734 SDOperand ADD = DAG.getNode(ISD::ADD, VT, N0, SRL);
735 AddToWorkList(SRL.Val);
736 AddToWorkList(ADD.Val); // Divide by pow2
737 SDOperand SRA = DAG.getNode(ISD::SRA, VT, ADD,
738 DAG.getConstant(lg2, TLI.getShiftAmountTy()));
739 // If we're dividing by a positive value, we're done. Otherwise, we must
740 // negate the result.
743 AddToWorkList(SRA.Val);
744 return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT), SRA);
746 // if integer divide is expensive and we satisfy the requirements, emit an
747 // alternate sequence.
748 if (N1C && (N1C->getSignExtended() < -1 || N1C->getSignExtended() > 1) &&
749 !TLI.isIntDivCheap()) {
750 SDOperand Op = BuildSDIV(N);
751 if (Op.Val) return Op;
756 SDOperand DAGCombiner::visitUDIV(SDNode *N) {
757 SDOperand N0 = N->getOperand(0);
758 SDOperand N1 = N->getOperand(1);
759 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val);
760 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val);
761 MVT::ValueType VT = N->getValueType(0);
763 // fold (udiv c1, c2) -> c1/c2
764 if (N0C && N1C && !N1C->isNullValue())
765 return DAG.getNode(ISD::UDIV, VT, N0, N1);
766 // fold (udiv x, (1 << c)) -> x >>u c
767 if (N1C && isPowerOf2_64(N1C->getValue()))
768 return DAG.getNode(ISD::SRL, VT, N0,
769 DAG.getConstant(Log2_64(N1C->getValue()),
770 TLI.getShiftAmountTy()));
771 // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2
772 if (N1.getOpcode() == ISD::SHL) {
773 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
774 if (isPowerOf2_64(SHC->getValue())) {
775 MVT::ValueType ADDVT = N1.getOperand(1).getValueType();
776 SDOperand Add = DAG.getNode(ISD::ADD, ADDVT, N1.getOperand(1),
777 DAG.getConstant(Log2_64(SHC->getValue()),
779 AddToWorkList(Add.Val);
780 return DAG.getNode(ISD::SRL, VT, N0, Add);
784 // fold (udiv x, c) -> alternate
785 if (N1C && N1C->getValue() && !TLI.isIntDivCheap()) {
786 SDOperand Op = BuildUDIV(N);
787 if (Op.Val) return Op;
792 SDOperand DAGCombiner::visitSREM(SDNode *N) {
793 SDOperand N0 = N->getOperand(0);
794 SDOperand N1 = N->getOperand(1);
795 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
796 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
797 MVT::ValueType VT = N->getValueType(0);
799 // fold (srem c1, c2) -> c1%c2
800 if (N0C && N1C && !N1C->isNullValue())
801 return DAG.getNode(ISD::SREM, VT, N0, N1);
802 // If we know the sign bits of both operands are zero, strength reduce to a
803 // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15
804 uint64_t SignBit = 1ULL << (MVT::getSizeInBits(VT)-1);
805 if (TLI.MaskedValueIsZero(N1, SignBit) &&
806 TLI.MaskedValueIsZero(N0, SignBit))
807 return DAG.getNode(ISD::UREM, VT, N0, N1);
811 SDOperand DAGCombiner::visitUREM(SDNode *N) {
812 SDOperand N0 = N->getOperand(0);
813 SDOperand N1 = N->getOperand(1);
814 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
815 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
816 MVT::ValueType VT = N->getValueType(0);
818 // fold (urem c1, c2) -> c1%c2
819 if (N0C && N1C && !N1C->isNullValue())
820 return DAG.getNode(ISD::UREM, VT, N0, N1);
821 // fold (urem x, pow2) -> (and x, pow2-1)
822 if (N1C && !N1C->isNullValue() && isPowerOf2_64(N1C->getValue()))
823 return DAG.getNode(ISD::AND, VT, N0, DAG.getConstant(N1C->getValue()-1,VT));
824 // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1))
825 if (N1.getOpcode() == ISD::SHL) {
826 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
827 if (isPowerOf2_64(SHC->getValue())) {
828 SDOperand Add = DAG.getNode(ISD::ADD, VT, N1,DAG.getConstant(~0ULL,VT));
829 AddToWorkList(Add.Val);
830 return DAG.getNode(ISD::AND, VT, N0, Add);
837 SDOperand DAGCombiner::visitMULHS(SDNode *N) {
838 SDOperand N0 = N->getOperand(0);
839 SDOperand N1 = N->getOperand(1);
840 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
842 // fold (mulhs x, 0) -> 0
843 if (N1C && N1C->isNullValue())
845 // fold (mulhs x, 1) -> (sra x, size(x)-1)
846 if (N1C && N1C->getValue() == 1)
847 return DAG.getNode(ISD::SRA, N0.getValueType(), N0,
848 DAG.getConstant(MVT::getSizeInBits(N0.getValueType())-1,
849 TLI.getShiftAmountTy()));
853 SDOperand DAGCombiner::visitMULHU(SDNode *N) {
854 SDOperand N0 = N->getOperand(0);
855 SDOperand N1 = N->getOperand(1);
856 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
858 // fold (mulhu x, 0) -> 0
859 if (N1C && N1C->isNullValue())
861 // fold (mulhu x, 1) -> 0
862 if (N1C && N1C->getValue() == 1)
863 return DAG.getConstant(0, N0.getValueType());
867 /// SimplifyBinOpWithSameOpcodeHands - If this is a binary operator with
868 /// two operands of the same opcode, try to simplify it.
869 SDOperand DAGCombiner::SimplifyBinOpWithSameOpcodeHands(SDNode *N) {
870 SDOperand N0 = N->getOperand(0), N1 = N->getOperand(1);
871 MVT::ValueType VT = N0.getValueType();
872 assert(N0.getOpcode() == N1.getOpcode() && "Bad input!");
874 // For each of OP in AND/OR/XOR:
875 // fold (OP (zext x), (zext y)) -> (zext (OP x, y))
876 // fold (OP (sext x), (sext y)) -> (sext (OP x, y))
877 // fold (OP (aext x), (aext y)) -> (aext (OP x, y))
878 // fold (OP (trunc x), (trunc y)) -> (trunc (OP x, y))
879 if ((N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND||
880 N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::TRUNCATE) &&
881 N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()) {
882 SDOperand ORNode = DAG.getNode(N->getOpcode(),
883 N0.getOperand(0).getValueType(),
884 N0.getOperand(0), N1.getOperand(0));
885 AddToWorkList(ORNode.Val);
886 return DAG.getNode(N0.getOpcode(), VT, ORNode);
889 // For each of OP in SHL/SRL/SRA/AND...
890 // fold (and (OP x, z), (OP y, z)) -> (OP (and x, y), z)
891 // fold (or (OP x, z), (OP y, z)) -> (OP (or x, y), z)
892 // fold (xor (OP x, z), (OP y, z)) -> (OP (xor x, y), z)
893 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL ||
894 N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::AND) &&
895 N0.getOperand(1) == N1.getOperand(1)) {
896 SDOperand ORNode = DAG.getNode(N->getOpcode(),
897 N0.getOperand(0).getValueType(),
898 N0.getOperand(0), N1.getOperand(0));
899 AddToWorkList(ORNode.Val);
900 return DAG.getNode(N0.getOpcode(), VT, ORNode, N0.getOperand(1));
906 SDOperand DAGCombiner::visitAND(SDNode *N) {
907 SDOperand N0 = N->getOperand(0);
908 SDOperand N1 = N->getOperand(1);
909 SDOperand LL, LR, RL, RR, CC0, CC1;
910 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
911 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
912 MVT::ValueType VT = N1.getValueType();
913 unsigned OpSizeInBits = MVT::getSizeInBits(VT);
915 // fold (and c1, c2) -> c1&c2
917 return DAG.getNode(ISD::AND, VT, N0, N1);
918 // canonicalize constant to RHS
920 return DAG.getNode(ISD::AND, VT, N1, N0);
921 // fold (and x, -1) -> x
922 if (N1C && N1C->isAllOnesValue())
924 // if (and x, c) is known to be zero, return 0
925 if (N1C && TLI.MaskedValueIsZero(SDOperand(N, 0), MVT::getIntVTBitMask(VT)))
926 return DAG.getConstant(0, VT);
928 SDOperand RAND = ReassociateOps(ISD::AND, N0, N1);
931 // fold (and (or x, 0xFFFF), 0xFF) -> 0xFF
932 if (N1C && N0.getOpcode() == ISD::OR)
933 if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
934 if ((ORI->getValue() & N1C->getValue()) == N1C->getValue())
936 // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits.
937 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
938 unsigned InMask = MVT::getIntVTBitMask(N0.getOperand(0).getValueType());
939 if (TLI.MaskedValueIsZero(N0.getOperand(0),
940 ~N1C->getValue() & InMask)) {
941 SDOperand Zext = DAG.getNode(ISD::ZERO_EXTEND, N0.getValueType(),
944 // Replace uses of the AND with uses of the Zero extend node.
947 // We actually want to replace all uses of the any_extend with the
948 // zero_extend, to avoid duplicating things. This will later cause this
950 CombineTo(N0.Val, Zext);
951 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
954 // fold (and (setcc x), (setcc y)) -> (setcc (and x, y))
955 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
956 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
957 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
959 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
960 MVT::isInteger(LL.getValueType())) {
961 // fold (X == 0) & (Y == 0) -> (X|Y == 0)
962 if (cast<ConstantSDNode>(LR)->getValue() == 0 && Op1 == ISD::SETEQ) {
963 SDOperand ORNode = DAG.getNode(ISD::OR, LR.getValueType(), LL, RL);
964 AddToWorkList(ORNode.Val);
965 return DAG.getSetCC(VT, ORNode, LR, Op1);
967 // fold (X == -1) & (Y == -1) -> (X&Y == -1)
968 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETEQ) {
969 SDOperand ANDNode = DAG.getNode(ISD::AND, LR.getValueType(), LL, RL);
970 AddToWorkList(ANDNode.Val);
971 return DAG.getSetCC(VT, ANDNode, LR, Op1);
973 // fold (X > -1) & (Y > -1) -> (X|Y > -1)
974 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETGT) {
975 SDOperand ORNode = DAG.getNode(ISD::OR, LR.getValueType(), LL, RL);
976 AddToWorkList(ORNode.Val);
977 return DAG.getSetCC(VT, ORNode, LR, Op1);
980 // canonicalize equivalent to ll == rl
981 if (LL == RR && LR == RL) {
982 Op1 = ISD::getSetCCSwappedOperands(Op1);
985 if (LL == RL && LR == RR) {
986 bool isInteger = MVT::isInteger(LL.getValueType());
987 ISD::CondCode Result = ISD::getSetCCAndOperation(Op0, Op1, isInteger);
988 if (Result != ISD::SETCC_INVALID)
989 return DAG.getSetCC(N0.getValueType(), LL, LR, Result);
993 // Simplify: and (op x...), (op y...) -> (op (and x, y))
994 if (N0.getOpcode() == N1.getOpcode()) {
995 SDOperand Tmp = SimplifyBinOpWithSameOpcodeHands(N);
996 if (Tmp.Val) return Tmp;
999 // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1)
1000 // fold (and (sra)) -> (and (srl)) when possible.
1001 if (!MVT::isVector(VT) &&
1002 SimplifyDemandedBits(SDOperand(N, 0)))
1003 return SDOperand(N, 0);
1004 // fold (zext_inreg (extload x)) -> (zextload x)
1005 if (N0.getOpcode() == ISD::EXTLOAD) {
1006 MVT::ValueType EVT = cast<VTSDNode>(N0.getOperand(3))->getVT();
1007 // If we zero all the possible extended bits, then we can turn this into
1008 // a zextload if we are running before legalize or the operation is legal.
1009 if (TLI.MaskedValueIsZero(N1, ~0ULL << MVT::getSizeInBits(EVT)) &&
1010 (!AfterLegalize || TLI.isOperationLegal(ISD::ZEXTLOAD, EVT))) {
1011 SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, N0.getOperand(0),
1012 N0.getOperand(1), N0.getOperand(2),
1015 CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1));
1016 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
1019 // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use
1020 if (N0.getOpcode() == ISD::SEXTLOAD && N0.hasOneUse()) {
1021 MVT::ValueType EVT = cast<VTSDNode>(N0.getOperand(3))->getVT();
1022 // If we zero all the possible extended bits, then we can turn this into
1023 // a zextload if we are running before legalize or the operation is legal.
1024 if (TLI.MaskedValueIsZero(N1, ~0ULL << MVT::getSizeInBits(EVT)) &&
1025 (!AfterLegalize || TLI.isOperationLegal(ISD::ZEXTLOAD, EVT))) {
1026 SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, N0.getOperand(0),
1027 N0.getOperand(1), N0.getOperand(2),
1030 CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1));
1031 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
1035 // fold (and (load x), 255) -> (zextload x, i8)
1036 // fold (and (extload x, i16), 255) -> (zextload x, i8)
1038 (N0.getOpcode() == ISD::LOAD || N0.getOpcode() == ISD::EXTLOAD ||
1039 N0.getOpcode() == ISD::ZEXTLOAD) &&
1041 MVT::ValueType EVT, LoadedVT;
1042 if (N1C->getValue() == 255)
1044 else if (N1C->getValue() == 65535)
1046 else if (N1C->getValue() == ~0U)
1051 LoadedVT = N0.getOpcode() == ISD::LOAD ? VT :
1052 cast<VTSDNode>(N0.getOperand(3))->getVT();
1053 if (EVT != MVT::Other && LoadedVT > EVT &&
1054 (!AfterLegalize || TLI.isOperationLegal(ISD::ZEXTLOAD, EVT))) {
1055 MVT::ValueType PtrType = N0.getOperand(1).getValueType();
1056 // For big endian targets, we need to add an offset to the pointer to load
1057 // the correct bytes. For little endian systems, we merely need to read
1058 // fewer bytes from the same pointer.
1060 (MVT::getSizeInBits(LoadedVT) - MVT::getSizeInBits(EVT)) / 8;
1061 SDOperand NewPtr = N0.getOperand(1);
1062 if (!TLI.isLittleEndian())
1063 NewPtr = DAG.getNode(ISD::ADD, PtrType, NewPtr,
1064 DAG.getConstant(PtrOff, PtrType));
1065 AddToWorkList(NewPtr.Val);
1067 DAG.getExtLoad(ISD::ZEXTLOAD, VT, N0.getOperand(0), NewPtr,
1068 N0.getOperand(2), EVT);
1070 CombineTo(N0.Val, Load, Load.getValue(1));
1071 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
1078 SDOperand DAGCombiner::visitOR(SDNode *N) {
1079 SDOperand N0 = N->getOperand(0);
1080 SDOperand N1 = N->getOperand(1);
1081 SDOperand LL, LR, RL, RR, CC0, CC1;
1082 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1083 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1084 MVT::ValueType VT = N1.getValueType();
1085 unsigned OpSizeInBits = MVT::getSizeInBits(VT);
1087 // fold (or c1, c2) -> c1|c2
1089 return DAG.getNode(ISD::OR, VT, N0, N1);
1090 // canonicalize constant to RHS
1092 return DAG.getNode(ISD::OR, VT, N1, N0);
1093 // fold (or x, 0) -> x
1094 if (N1C && N1C->isNullValue())
1096 // fold (or x, -1) -> -1
1097 if (N1C && N1C->isAllOnesValue())
1099 // fold (or x, c) -> c iff (x & ~c) == 0
1101 TLI.MaskedValueIsZero(N0,~N1C->getValue() & (~0ULL>>(64-OpSizeInBits))))
1104 SDOperand ROR = ReassociateOps(ISD::OR, N0, N1);
1107 // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2)
1108 if (N1C && N0.getOpcode() == ISD::AND && N0.Val->hasOneUse() &&
1109 isa<ConstantSDNode>(N0.getOperand(1))) {
1110 ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1));
1111 return DAG.getNode(ISD::AND, VT, DAG.getNode(ISD::OR, VT, N0.getOperand(0),
1113 DAG.getConstant(N1C->getValue() | C1->getValue(), VT));
1115 // fold (or (setcc x), (setcc y)) -> (setcc (or x, y))
1116 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
1117 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
1118 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
1120 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
1121 MVT::isInteger(LL.getValueType())) {
1122 // fold (X != 0) | (Y != 0) -> (X|Y != 0)
1123 // fold (X < 0) | (Y < 0) -> (X|Y < 0)
1124 if (cast<ConstantSDNode>(LR)->getValue() == 0 &&
1125 (Op1 == ISD::SETNE || Op1 == ISD::SETLT)) {
1126 SDOperand ORNode = DAG.getNode(ISD::OR, LR.getValueType(), LL, RL);
1127 AddToWorkList(ORNode.Val);
1128 return DAG.getSetCC(VT, ORNode, LR, Op1);
1130 // fold (X != -1) | (Y != -1) -> (X&Y != -1)
1131 // fold (X > -1) | (Y > -1) -> (X&Y > -1)
1132 if (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
1133 (Op1 == ISD::SETNE || Op1 == ISD::SETGT)) {
1134 SDOperand ANDNode = DAG.getNode(ISD::AND, LR.getValueType(), LL, RL);
1135 AddToWorkList(ANDNode.Val);
1136 return DAG.getSetCC(VT, ANDNode, LR, Op1);
1139 // canonicalize equivalent to ll == rl
1140 if (LL == RR && LR == RL) {
1141 Op1 = ISD::getSetCCSwappedOperands(Op1);
1144 if (LL == RL && LR == RR) {
1145 bool isInteger = MVT::isInteger(LL.getValueType());
1146 ISD::CondCode Result = ISD::getSetCCOrOperation(Op0, Op1, isInteger);
1147 if (Result != ISD::SETCC_INVALID)
1148 return DAG.getSetCC(N0.getValueType(), LL, LR, Result);
1152 // Simplify: or (op x...), (op y...) -> (op (or x, y))
1153 if (N0.getOpcode() == N1.getOpcode()) {
1154 SDOperand Tmp = SimplifyBinOpWithSameOpcodeHands(N);
1155 if (Tmp.Val) return Tmp;
1158 // See if this is some rotate idiom.
1159 if (SDNode *Rot = MatchRotate(N0, N1))
1160 return SDOperand(Rot, 0);
1166 /// MatchRotateHalf - Match "(X shl/srl V1) & V2" where V2 may not be present.
1167 static bool MatchRotateHalf(SDOperand Op, SDOperand &Shift, SDOperand &Mask) {
1168 if (Op.getOpcode() == ISD::AND) {
1169 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1170 Mask = Op.getOperand(1);
1171 Op = Op.getOperand(0);
1177 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SHL) {
1185 // MatchRotate - Handle an 'or' of two operands. If this is one of the many
1186 // idioms for rotate, and if the target supports rotation instructions, generate
1188 SDNode *DAGCombiner::MatchRotate(SDOperand LHS, SDOperand RHS) {
1189 // Must be a legal type. Expanded an promoted things won't work with rotates.
1190 MVT::ValueType VT = LHS.getValueType();
1191 if (!TLI.isTypeLegal(VT)) return 0;
1193 // The target must have at least one rotate flavor.
1194 bool HasROTL = TLI.isOperationLegal(ISD::ROTL, VT);
1195 bool HasROTR = TLI.isOperationLegal(ISD::ROTR, VT);
1196 if (!HasROTL && !HasROTR) return 0;
1198 // Match "(X shl/srl V1) & V2" where V2 may not be present.
1199 SDOperand LHSShift; // The shift.
1200 SDOperand LHSMask; // AND value if any.
1201 if (!MatchRotateHalf(LHS, LHSShift, LHSMask))
1202 return 0; // Not part of a rotate.
1204 SDOperand RHSShift; // The shift.
1205 SDOperand RHSMask; // AND value if any.
1206 if (!MatchRotateHalf(RHS, RHSShift, RHSMask))
1207 return 0; // Not part of a rotate.
1209 if (LHSShift.getOperand(0) != RHSShift.getOperand(0))
1210 return 0; // Not shifting the same value.
1212 if (LHSShift.getOpcode() == RHSShift.getOpcode())
1213 return 0; // Shifts must disagree.
1215 // Canonicalize shl to left side in a shl/srl pair.
1216 if (RHSShift.getOpcode() == ISD::SHL) {
1217 std::swap(LHS, RHS);
1218 std::swap(LHSShift, RHSShift);
1219 std::swap(LHSMask , RHSMask );
1222 unsigned OpSizeInBits = MVT::getSizeInBits(VT);
1224 // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1)
1225 // fold (or (shl x, C1), (srl x, C2)) -> (rotr x, C2)
1226 if (LHSShift.getOperand(1).getOpcode() == ISD::Constant &&
1227 RHSShift.getOperand(1).getOpcode() == ISD::Constant) {
1228 uint64_t LShVal = cast<ConstantSDNode>(LHSShift.getOperand(1))->getValue();
1229 uint64_t RShVal = cast<ConstantSDNode>(RHSShift.getOperand(1))->getValue();
1230 if ((LShVal + RShVal) != OpSizeInBits)
1235 Rot = DAG.getNode(ISD::ROTL, VT, LHSShift.getOperand(0),
1236 LHSShift.getOperand(1));
1238 Rot = DAG.getNode(ISD::ROTR, VT, LHSShift.getOperand(0),
1239 RHSShift.getOperand(1));
1241 // If there is an AND of either shifted operand, apply it to the result.
1242 if (LHSMask.Val || RHSMask.Val) {
1243 uint64_t Mask = MVT::getIntVTBitMask(VT);
1246 uint64_t RHSBits = (1ULL << LShVal)-1;
1247 Mask &= cast<ConstantSDNode>(LHSMask)->getValue() | RHSBits;
1250 uint64_t LHSBits = ~((1ULL << (OpSizeInBits-RShVal))-1);
1251 Mask &= cast<ConstantSDNode>(RHSMask)->getValue() | LHSBits;
1254 Rot = DAG.getNode(ISD::AND, VT, Rot, DAG.getConstant(Mask, VT));
1260 // If there is a mask here, and we have a variable shift, we can't be sure
1261 // that we're masking out the right stuff.
1262 if (LHSMask.Val || RHSMask.Val)
1265 // fold (or (shl x, y), (srl x, (sub 32, y))) -> (rotl x, y)
1266 // fold (or (shl x, y), (srl x, (sub 32, y))) -> (rotr x, (sub 32, y))
1267 if (RHSShift.getOperand(1).getOpcode() == ISD::SUB &&
1268 LHSShift.getOperand(1) == RHSShift.getOperand(1).getOperand(1)) {
1269 if (ConstantSDNode *SUBC =
1270 dyn_cast<ConstantSDNode>(RHSShift.getOperand(1).getOperand(0))) {
1271 if (SUBC->getValue() == OpSizeInBits)
1273 return DAG.getNode(ISD::ROTL, VT, LHSShift.getOperand(0),
1274 LHSShift.getOperand(1)).Val;
1276 return DAG.getNode(ISD::ROTR, VT, LHSShift.getOperand(0),
1277 LHSShift.getOperand(1)).Val;
1281 // fold (or (shl x, (sub 32, y)), (srl x, r)) -> (rotr x, y)
1282 // fold (or (shl x, (sub 32, y)), (srl x, r)) -> (rotl x, (sub 32, y))
1283 if (LHSShift.getOperand(1).getOpcode() == ISD::SUB &&
1284 RHSShift.getOperand(1) == LHSShift.getOperand(1).getOperand(1)) {
1285 if (ConstantSDNode *SUBC =
1286 dyn_cast<ConstantSDNode>(LHSShift.getOperand(1).getOperand(0))) {
1287 if (SUBC->getValue() == OpSizeInBits)
1289 return DAG.getNode(ISD::ROTL, VT, LHSShift.getOperand(0),
1290 LHSShift.getOperand(1)).Val;
1292 return DAG.getNode(ISD::ROTR, VT, LHSShift.getOperand(0),
1293 RHSShift.getOperand(1)).Val;
1301 SDOperand DAGCombiner::visitXOR(SDNode *N) {
1302 SDOperand N0 = N->getOperand(0);
1303 SDOperand N1 = N->getOperand(1);
1304 SDOperand LHS, RHS, CC;
1305 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1306 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1307 MVT::ValueType VT = N0.getValueType();
1309 // fold (xor c1, c2) -> c1^c2
1311 return DAG.getNode(ISD::XOR, VT, N0, N1);
1312 // canonicalize constant to RHS
1314 return DAG.getNode(ISD::XOR, VT, N1, N0);
1315 // fold (xor x, 0) -> x
1316 if (N1C && N1C->isNullValue())
1319 SDOperand RXOR = ReassociateOps(ISD::XOR, N0, N1);
1322 // fold !(x cc y) -> (x !cc y)
1323 if (N1C && N1C->getValue() == 1 && isSetCCEquivalent(N0, LHS, RHS, CC)) {
1324 bool isInt = MVT::isInteger(LHS.getValueType());
1325 ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
1327 if (N0.getOpcode() == ISD::SETCC)
1328 return DAG.getSetCC(VT, LHS, RHS, NotCC);
1329 if (N0.getOpcode() == ISD::SELECT_CC)
1330 return DAG.getSelectCC(LHS, RHS, N0.getOperand(2),N0.getOperand(3),NotCC);
1331 assert(0 && "Unhandled SetCC Equivalent!");
1334 // fold !(x or y) -> (!x and !y) iff x or y are setcc
1335 if (N1C && N1C->getValue() == 1 &&
1336 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
1337 SDOperand LHS = N0.getOperand(0), RHS = N0.getOperand(1);
1338 if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) {
1339 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
1340 LHS = DAG.getNode(ISD::XOR, VT, LHS, N1); // RHS = ~LHS
1341 RHS = DAG.getNode(ISD::XOR, VT, RHS, N1); // RHS = ~RHS
1342 AddToWorkList(LHS.Val); AddToWorkList(RHS.Val);
1343 return DAG.getNode(NewOpcode, VT, LHS, RHS);
1346 // fold !(x or y) -> (!x and !y) iff x or y are constants
1347 if (N1C && N1C->isAllOnesValue() &&
1348 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
1349 SDOperand LHS = N0.getOperand(0), RHS = N0.getOperand(1);
1350 if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) {
1351 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
1352 LHS = DAG.getNode(ISD::XOR, VT, LHS, N1); // RHS = ~LHS
1353 RHS = DAG.getNode(ISD::XOR, VT, RHS, N1); // RHS = ~RHS
1354 AddToWorkList(LHS.Val); AddToWorkList(RHS.Val);
1355 return DAG.getNode(NewOpcode, VT, LHS, RHS);
1358 // fold (xor (xor x, c1), c2) -> (xor x, c1^c2)
1359 if (N1C && N0.getOpcode() == ISD::XOR) {
1360 ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0));
1361 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
1363 return DAG.getNode(ISD::XOR, VT, N0.getOperand(1),
1364 DAG.getConstant(N1C->getValue()^N00C->getValue(), VT));
1366 return DAG.getNode(ISD::XOR, VT, N0.getOperand(0),
1367 DAG.getConstant(N1C->getValue()^N01C->getValue(), VT));
1369 // fold (xor x, x) -> 0
1371 if (!MVT::isVector(VT)) {
1372 return DAG.getConstant(0, VT);
1373 } else if (!AfterLegalize || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT)) {
1374 // Produce a vector of zeros.
1375 SDOperand El = DAG.getConstant(0, MVT::getVectorBaseType(VT));
1376 std::vector<SDOperand> Ops(MVT::getVectorNumElements(VT), El);
1377 return DAG.getNode(ISD::BUILD_VECTOR, VT, &Ops[0], Ops.size());
1381 // Simplify: xor (op x...), (op y...) -> (op (xor x, y))
1382 if (N0.getOpcode() == N1.getOpcode()) {
1383 SDOperand Tmp = SimplifyBinOpWithSameOpcodeHands(N);
1384 if (Tmp.Val) return Tmp;
1387 // Simplify the expression using non-local knowledge.
1388 if (!MVT::isVector(VT) &&
1389 SimplifyDemandedBits(SDOperand(N, 0)))
1390 return SDOperand(N, 0);
1395 SDOperand DAGCombiner::visitSHL(SDNode *N) {
1396 SDOperand N0 = N->getOperand(0);
1397 SDOperand N1 = N->getOperand(1);
1398 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1399 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1400 MVT::ValueType VT = N0.getValueType();
1401 unsigned OpSizeInBits = MVT::getSizeInBits(VT);
1403 // fold (shl c1, c2) -> c1<<c2
1405 return DAG.getNode(ISD::SHL, VT, N0, N1);
1406 // fold (shl 0, x) -> 0
1407 if (N0C && N0C->isNullValue())
1409 // fold (shl x, c >= size(x)) -> undef
1410 if (N1C && N1C->getValue() >= OpSizeInBits)
1411 return DAG.getNode(ISD::UNDEF, VT);
1412 // fold (shl x, 0) -> x
1413 if (N1C && N1C->isNullValue())
1415 // if (shl x, c) is known to be zero, return 0
1416 if (TLI.MaskedValueIsZero(SDOperand(N, 0), MVT::getIntVTBitMask(VT)))
1417 return DAG.getConstant(0, VT);
1418 if (SimplifyDemandedBits(SDOperand(N, 0)))
1419 return SDOperand(N, 0);
1420 // fold (shl (shl x, c1), c2) -> 0 or (shl x, c1+c2)
1421 if (N1C && N0.getOpcode() == ISD::SHL &&
1422 N0.getOperand(1).getOpcode() == ISD::Constant) {
1423 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
1424 uint64_t c2 = N1C->getValue();
1425 if (c1 + c2 > OpSizeInBits)
1426 return DAG.getConstant(0, VT);
1427 return DAG.getNode(ISD::SHL, VT, N0.getOperand(0),
1428 DAG.getConstant(c1 + c2, N1.getValueType()));
1430 // fold (shl (srl x, c1), c2) -> (shl (and x, -1 << c1), c2-c1) or
1431 // (srl (and x, -1 << c1), c1-c2)
1432 if (N1C && N0.getOpcode() == ISD::SRL &&
1433 N0.getOperand(1).getOpcode() == ISD::Constant) {
1434 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
1435 uint64_t c2 = N1C->getValue();
1436 SDOperand Mask = DAG.getNode(ISD::AND, VT, N0.getOperand(0),
1437 DAG.getConstant(~0ULL << c1, VT));
1439 return DAG.getNode(ISD::SHL, VT, Mask,
1440 DAG.getConstant(c2-c1, N1.getValueType()));
1442 return DAG.getNode(ISD::SRL, VT, Mask,
1443 DAG.getConstant(c1-c2, N1.getValueType()));
1445 // fold (shl (sra x, c1), c1) -> (and x, -1 << c1)
1446 if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1))
1447 return DAG.getNode(ISD::AND, VT, N0.getOperand(0),
1448 DAG.getConstant(~0ULL << N1C->getValue(), VT));
1449 // fold (shl (add x, c1), c2) -> (add (shl x, c2), c1<<c2)
1450 if (N1C && N0.getOpcode() == ISD::ADD && N0.Val->hasOneUse() &&
1451 isa<ConstantSDNode>(N0.getOperand(1))) {
1452 return DAG.getNode(ISD::ADD, VT,
1453 DAG.getNode(ISD::SHL, VT, N0.getOperand(0), N1),
1454 DAG.getNode(ISD::SHL, VT, N0.getOperand(1), N1));
1459 SDOperand DAGCombiner::visitSRA(SDNode *N) {
1460 SDOperand N0 = N->getOperand(0);
1461 SDOperand N1 = N->getOperand(1);
1462 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1463 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1464 MVT::ValueType VT = N0.getValueType();
1466 // fold (sra c1, c2) -> c1>>c2
1468 return DAG.getNode(ISD::SRA, VT, N0, N1);
1469 // fold (sra 0, x) -> 0
1470 if (N0C && N0C->isNullValue())
1472 // fold (sra -1, x) -> -1
1473 if (N0C && N0C->isAllOnesValue())
1475 // fold (sra x, c >= size(x)) -> undef
1476 if (N1C && N1C->getValue() >= MVT::getSizeInBits(VT))
1477 return DAG.getNode(ISD::UNDEF, VT);
1478 // fold (sra x, 0) -> x
1479 if (N1C && N1C->isNullValue())
1481 // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports
1483 if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) {
1484 unsigned LowBits = MVT::getSizeInBits(VT) - (unsigned)N1C->getValue();
1487 default: EVT = MVT::Other; break;
1488 case 1: EVT = MVT::i1; break;
1489 case 8: EVT = MVT::i8; break;
1490 case 16: EVT = MVT::i16; break;
1491 case 32: EVT = MVT::i32; break;
1493 if (EVT > MVT::Other && TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, EVT))
1494 return DAG.getNode(ISD::SIGN_EXTEND_INREG, VT, N0.getOperand(0),
1495 DAG.getValueType(EVT));
1498 // fold (sra (sra x, c1), c2) -> (sra x, c1+c2)
1499 if (N1C && N0.getOpcode() == ISD::SRA) {
1500 if (ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1501 unsigned Sum = N1C->getValue() + C1->getValue();
1502 if (Sum >= MVT::getSizeInBits(VT)) Sum = MVT::getSizeInBits(VT)-1;
1503 return DAG.getNode(ISD::SRA, VT, N0.getOperand(0),
1504 DAG.getConstant(Sum, N1C->getValueType(0)));
1508 // Simplify, based on bits shifted out of the LHS.
1509 if (N1C && SimplifyDemandedBits(SDOperand(N, 0)))
1510 return SDOperand(N, 0);
1513 // If the sign bit is known to be zero, switch this to a SRL.
1514 if (TLI.MaskedValueIsZero(N0, MVT::getIntVTSignBit(VT)))
1515 return DAG.getNode(ISD::SRL, VT, N0, N1);
1519 SDOperand DAGCombiner::visitSRL(SDNode *N) {
1520 SDOperand N0 = N->getOperand(0);
1521 SDOperand N1 = N->getOperand(1);
1522 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1523 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1524 MVT::ValueType VT = N0.getValueType();
1525 unsigned OpSizeInBits = MVT::getSizeInBits(VT);
1527 // fold (srl c1, c2) -> c1 >>u c2
1529 return DAG.getNode(ISD::SRL, VT, N0, N1);
1530 // fold (srl 0, x) -> 0
1531 if (N0C && N0C->isNullValue())
1533 // fold (srl x, c >= size(x)) -> undef
1534 if (N1C && N1C->getValue() >= OpSizeInBits)
1535 return DAG.getNode(ISD::UNDEF, VT);
1536 // fold (srl x, 0) -> x
1537 if (N1C && N1C->isNullValue())
1539 // if (srl x, c) is known to be zero, return 0
1540 if (N1C && TLI.MaskedValueIsZero(SDOperand(N, 0), ~0ULL >> (64-OpSizeInBits)))
1541 return DAG.getConstant(0, VT);
1542 // fold (srl (srl x, c1), c2) -> 0 or (srl x, c1+c2)
1543 if (N1C && N0.getOpcode() == ISD::SRL &&
1544 N0.getOperand(1).getOpcode() == ISD::Constant) {
1545 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
1546 uint64_t c2 = N1C->getValue();
1547 if (c1 + c2 > OpSizeInBits)
1548 return DAG.getConstant(0, VT);
1549 return DAG.getNode(ISD::SRL, VT, N0.getOperand(0),
1550 DAG.getConstant(c1 + c2, N1.getValueType()));
1553 // fold (srl (anyextend x), c) -> (anyextend (srl x, c))
1554 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
1555 // Shifting in all undef bits?
1556 MVT::ValueType SmallVT = N0.getOperand(0).getValueType();
1557 if (N1C->getValue() >= MVT::getSizeInBits(SmallVT))
1558 return DAG.getNode(ISD::UNDEF, VT);
1560 SDOperand SmallShift = DAG.getNode(ISD::SRL, SmallVT, N0.getOperand(0), N1);
1561 AddToWorkList(SmallShift.Val);
1562 return DAG.getNode(ISD::ANY_EXTEND, VT, SmallShift);
1565 // fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit).
1566 if (N1C && N0.getOpcode() == ISD::CTLZ &&
1567 N1C->getValue() == Log2_32(MVT::getSizeInBits(VT))) {
1568 uint64_t KnownZero, KnownOne, Mask = MVT::getIntVTBitMask(VT);
1569 TLI.ComputeMaskedBits(N0.getOperand(0), Mask, KnownZero, KnownOne);
1571 // If any of the input bits are KnownOne, then the input couldn't be all
1572 // zeros, thus the result of the srl will always be zero.
1573 if (KnownOne) return DAG.getConstant(0, VT);
1575 // If all of the bits input the to ctlz node are known to be zero, then
1576 // the result of the ctlz is "32" and the result of the shift is one.
1577 uint64_t UnknownBits = ~KnownZero & Mask;
1578 if (UnknownBits == 0) return DAG.getConstant(1, VT);
1580 // Otherwise, check to see if there is exactly one bit input to the ctlz.
1581 if ((UnknownBits & (UnknownBits-1)) == 0) {
1582 // Okay, we know that only that the single bit specified by UnknownBits
1583 // could be set on input to the CTLZ node. If this bit is set, the SRL
1584 // will return 0, if it is clear, it returns 1. Change the CTLZ/SRL pair
1585 // to an SRL,XOR pair, which is likely to simplify more.
1586 unsigned ShAmt = CountTrailingZeros_64(UnknownBits);
1587 SDOperand Op = N0.getOperand(0);
1589 Op = DAG.getNode(ISD::SRL, VT, Op,
1590 DAG.getConstant(ShAmt, TLI.getShiftAmountTy()));
1591 AddToWorkList(Op.Val);
1593 return DAG.getNode(ISD::XOR, VT, Op, DAG.getConstant(1, VT));
1600 SDOperand DAGCombiner::visitCTLZ(SDNode *N) {
1601 SDOperand N0 = N->getOperand(0);
1602 MVT::ValueType VT = N->getValueType(0);
1604 // fold (ctlz c1) -> c2
1605 if (isa<ConstantSDNode>(N0))
1606 return DAG.getNode(ISD::CTLZ, VT, N0);
1610 SDOperand DAGCombiner::visitCTTZ(SDNode *N) {
1611 SDOperand N0 = N->getOperand(0);
1612 MVT::ValueType VT = N->getValueType(0);
1614 // fold (cttz c1) -> c2
1615 if (isa<ConstantSDNode>(N0))
1616 return DAG.getNode(ISD::CTTZ, VT, N0);
1620 SDOperand DAGCombiner::visitCTPOP(SDNode *N) {
1621 SDOperand N0 = N->getOperand(0);
1622 MVT::ValueType VT = N->getValueType(0);
1624 // fold (ctpop c1) -> c2
1625 if (isa<ConstantSDNode>(N0))
1626 return DAG.getNode(ISD::CTPOP, VT, N0);
1630 SDOperand DAGCombiner::visitSELECT(SDNode *N) {
1631 SDOperand N0 = N->getOperand(0);
1632 SDOperand N1 = N->getOperand(1);
1633 SDOperand N2 = N->getOperand(2);
1634 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1635 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1636 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
1637 MVT::ValueType VT = N->getValueType(0);
1639 // fold select C, X, X -> X
1642 // fold select true, X, Y -> X
1643 if (N0C && !N0C->isNullValue())
1645 // fold select false, X, Y -> Y
1646 if (N0C && N0C->isNullValue())
1648 // fold select C, 1, X -> C | X
1649 if (MVT::i1 == VT && N1C && N1C->getValue() == 1)
1650 return DAG.getNode(ISD::OR, VT, N0, N2);
1651 // fold select C, 0, X -> ~C & X
1652 // FIXME: this should check for C type == X type, not i1?
1653 if (MVT::i1 == VT && N1C && N1C->isNullValue()) {
1654 SDOperand XORNode = DAG.getNode(ISD::XOR, VT, N0, DAG.getConstant(1, VT));
1655 AddToWorkList(XORNode.Val);
1656 return DAG.getNode(ISD::AND, VT, XORNode, N2);
1658 // fold select C, X, 1 -> ~C | X
1659 if (MVT::i1 == VT && N2C && N2C->getValue() == 1) {
1660 SDOperand XORNode = DAG.getNode(ISD::XOR, VT, N0, DAG.getConstant(1, VT));
1661 AddToWorkList(XORNode.Val);
1662 return DAG.getNode(ISD::OR, VT, XORNode, N1);
1664 // fold select C, X, 0 -> C & X
1665 // FIXME: this should check for C type == X type, not i1?
1666 if (MVT::i1 == VT && N2C && N2C->isNullValue())
1667 return DAG.getNode(ISD::AND, VT, N0, N1);
1668 // fold X ? X : Y --> X ? 1 : Y --> X | Y
1669 if (MVT::i1 == VT && N0 == N1)
1670 return DAG.getNode(ISD::OR, VT, N0, N2);
1671 // fold X ? Y : X --> X ? Y : 0 --> X & Y
1672 if (MVT::i1 == VT && N0 == N2)
1673 return DAG.getNode(ISD::AND, VT, N0, N1);
1675 // If we can fold this based on the true/false value, do so.
1676 if (SimplifySelectOps(N, N1, N2))
1677 return SDOperand(N, 0); // Don't revisit N.
1679 // fold selects based on a setcc into other things, such as min/max/abs
1680 if (N0.getOpcode() == ISD::SETCC)
1682 // Check against MVT::Other for SELECT_CC, which is a workaround for targets
1683 // having to say they don't support SELECT_CC on every type the DAG knows
1684 // about, since there is no way to mark an opcode illegal at all value types
1685 if (TLI.isOperationLegal(ISD::SELECT_CC, MVT::Other))
1686 return DAG.getNode(ISD::SELECT_CC, VT, N0.getOperand(0), N0.getOperand(1),
1687 N1, N2, N0.getOperand(2));
1689 return SimplifySelect(N0, N1, N2);
1693 SDOperand DAGCombiner::visitSELECT_CC(SDNode *N) {
1694 SDOperand N0 = N->getOperand(0);
1695 SDOperand N1 = N->getOperand(1);
1696 SDOperand N2 = N->getOperand(2);
1697 SDOperand N3 = N->getOperand(3);
1698 SDOperand N4 = N->getOperand(4);
1699 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1700 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1701 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
1702 ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get();
1704 // Determine if the condition we're dealing with is constant
1705 SDOperand SCC = SimplifySetCC(TLI.getSetCCResultTy(), N0, N1, CC, false);
1706 //ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.Val);
1708 // fold select_cc lhs, rhs, x, x, cc -> x
1712 // If we can fold this based on the true/false value, do so.
1713 if (SimplifySelectOps(N, N2, N3))
1714 return SDOperand(N, 0); // Don't revisit N.
1716 // fold select_cc into other things, such as min/max/abs
1717 return SimplifySelectCC(N0, N1, N2, N3, CC);
1720 SDOperand DAGCombiner::visitSETCC(SDNode *N) {
1721 return SimplifySetCC(N->getValueType(0), N->getOperand(0), N->getOperand(1),
1722 cast<CondCodeSDNode>(N->getOperand(2))->get());
1725 SDOperand DAGCombiner::visitSIGN_EXTEND(SDNode *N) {
1726 SDOperand N0 = N->getOperand(0);
1727 MVT::ValueType VT = N->getValueType(0);
1729 // fold (sext c1) -> c1
1730 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0))
1731 return DAG.getNode(ISD::SIGN_EXTEND, VT, N0);
1733 // fold (sext (sext x)) -> (sext x)
1734 // fold (sext (aext x)) -> (sext x)
1735 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
1736 return DAG.getNode(ISD::SIGN_EXTEND, VT, N0.getOperand(0));
1738 // fold (sext (truncate x)) -> (sextinreg x) iff x size == sext size.
1739 if (N0.getOpcode() == ISD::TRUNCATE && N0.getOperand(0).getValueType() == VT&&
1741 TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, N0.getValueType())))
1742 return DAG.getNode(ISD::SIGN_EXTEND_INREG, VT, N0.getOperand(0),
1743 DAG.getValueType(N0.getValueType()));
1745 // fold (sext (load x)) -> (sext (truncate (sextload x)))
1746 if (N0.getOpcode() == ISD::LOAD && N0.hasOneUse() &&
1747 (!AfterLegalize||TLI.isOperationLegal(ISD::SEXTLOAD, N0.getValueType()))){
1748 SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, N0.getOperand(0),
1749 N0.getOperand(1), N0.getOperand(2),
1751 CombineTo(N, ExtLoad);
1752 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
1753 ExtLoad.getValue(1));
1754 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
1757 // fold (sext (sextload x)) -> (sext (truncate (sextload x)))
1758 // fold (sext ( extload x)) -> (sext (truncate (sextload x)))
1759 if ((N0.getOpcode() == ISD::SEXTLOAD || N0.getOpcode() == ISD::EXTLOAD) &&
1761 MVT::ValueType EVT = cast<VTSDNode>(N0.getOperand(3))->getVT();
1762 SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, N0.getOperand(0),
1763 N0.getOperand(1), N0.getOperand(2), EVT);
1764 CombineTo(N, ExtLoad);
1765 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
1766 ExtLoad.getValue(1));
1767 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
1773 SDOperand DAGCombiner::visitZERO_EXTEND(SDNode *N) {
1774 SDOperand N0 = N->getOperand(0);
1775 MVT::ValueType VT = N->getValueType(0);
1777 // fold (zext c1) -> c1
1778 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0))
1779 return DAG.getNode(ISD::ZERO_EXTEND, VT, N0);
1780 // fold (zext (zext x)) -> (zext x)
1781 // fold (zext (aext x)) -> (zext x)
1782 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
1783 return DAG.getNode(ISD::ZERO_EXTEND, VT, N0.getOperand(0));
1784 // fold (zext (truncate x)) -> (zextinreg x) iff x size == zext size.
1785 if (N0.getOpcode() == ISD::TRUNCATE && N0.getOperand(0).getValueType() == VT&&
1786 (!AfterLegalize || TLI.isOperationLegal(ISD::AND, N0.getValueType())))
1787 return DAG.getZeroExtendInReg(N0.getOperand(0), N0.getValueType());
1788 // fold (zext (load x)) -> (zext (truncate (zextload x)))
1789 if (N0.getOpcode() == ISD::LOAD && N0.hasOneUse() &&
1790 (!AfterLegalize||TLI.isOperationLegal(ISD::ZEXTLOAD, N0.getValueType()))){
1791 SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, N0.getOperand(0),
1792 N0.getOperand(1), N0.getOperand(2),
1794 CombineTo(N, ExtLoad);
1795 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
1796 ExtLoad.getValue(1));
1797 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
1800 // fold (zext (zextload x)) -> (zext (truncate (zextload x)))
1801 // fold (zext ( extload x)) -> (zext (truncate (zextload x)))
1802 if ((N0.getOpcode() == ISD::ZEXTLOAD || N0.getOpcode() == ISD::EXTLOAD) &&
1804 MVT::ValueType EVT = cast<VTSDNode>(N0.getOperand(3))->getVT();
1805 SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, N0.getOperand(0),
1806 N0.getOperand(1), N0.getOperand(2), EVT);
1807 CombineTo(N, ExtLoad);
1808 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
1809 ExtLoad.getValue(1));
1810 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
1815 SDOperand DAGCombiner::visitANY_EXTEND(SDNode *N) {
1816 SDOperand N0 = N->getOperand(0);
1817 MVT::ValueType VT = N->getValueType(0);
1819 // fold (aext c1) -> c1
1820 if (isa<ConstantSDNode>(N0))
1821 return DAG.getNode(ISD::ANY_EXTEND, VT, N0);
1822 // fold (aext (aext x)) -> (aext x)
1823 // fold (aext (zext x)) -> (zext x)
1824 // fold (aext (sext x)) -> (sext x)
1825 if (N0.getOpcode() == ISD::ANY_EXTEND ||
1826 N0.getOpcode() == ISD::ZERO_EXTEND ||
1827 N0.getOpcode() == ISD::SIGN_EXTEND)
1828 return DAG.getNode(N0.getOpcode(), VT, N0.getOperand(0));
1830 // fold (aext (truncate x)) -> x iff x size == zext size.
1831 if (N0.getOpcode() == ISD::TRUNCATE && N0.getOperand(0).getValueType() == VT)
1832 return N0.getOperand(0);
1833 // fold (aext (load x)) -> (aext (truncate (extload x)))
1834 if (N0.getOpcode() == ISD::LOAD && N0.hasOneUse() &&
1835 (!AfterLegalize||TLI.isOperationLegal(ISD::EXTLOAD, N0.getValueType()))) {
1836 SDOperand ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, VT, N0.getOperand(0),
1837 N0.getOperand(1), N0.getOperand(2),
1839 CombineTo(N, ExtLoad);
1840 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
1841 ExtLoad.getValue(1));
1842 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
1845 // fold (aext (zextload x)) -> (aext (truncate (zextload x)))
1846 // fold (aext (sextload x)) -> (aext (truncate (sextload x)))
1847 // fold (aext ( extload x)) -> (aext (truncate (extload x)))
1848 if ((N0.getOpcode() == ISD::ZEXTLOAD || N0.getOpcode() == ISD::EXTLOAD ||
1849 N0.getOpcode() == ISD::SEXTLOAD) &&
1851 MVT::ValueType EVT = cast<VTSDNode>(N0.getOperand(3))->getVT();
1852 SDOperand ExtLoad = DAG.getExtLoad(N0.getOpcode(), VT, N0.getOperand(0),
1853 N0.getOperand(1), N0.getOperand(2), EVT);
1854 CombineTo(N, ExtLoad);
1855 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
1856 ExtLoad.getValue(1));
1857 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
1863 SDOperand DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
1864 SDOperand N0 = N->getOperand(0);
1865 SDOperand N1 = N->getOperand(1);
1866 MVT::ValueType VT = N->getValueType(0);
1867 MVT::ValueType EVT = cast<VTSDNode>(N1)->getVT();
1868 unsigned EVTBits = MVT::getSizeInBits(EVT);
1870 // fold (sext_in_reg c1) -> c1
1871 if (isa<ConstantSDNode>(N0) || N0.getOpcode() == ISD::UNDEF)
1872 return DAG.getNode(ISD::SIGN_EXTEND_INREG, VT, N0, N1);
1874 // If the input is already sign extended, just drop the extension.
1875 if (TLI.ComputeNumSignBits(N0) >= MVT::getSizeInBits(VT)-EVTBits+1)
1878 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2
1879 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
1880 EVT < cast<VTSDNode>(N0.getOperand(1))->getVT()) {
1881 return DAG.getNode(ISD::SIGN_EXTEND_INREG, VT, N0.getOperand(0), N1);
1884 // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is zero
1885 if (TLI.MaskedValueIsZero(N0, 1ULL << (EVTBits-1)))
1886 return DAG.getZeroExtendInReg(N0, EVT);
1888 // fold (sext_in_reg (srl X, 24), i8) -> sra X, 24
1889 // fold (sext_in_reg (srl X, 23), i8) -> sra X, 23 iff possible.
1890 // We already fold "(sext_in_reg (srl X, 25), i8) -> srl X, 25" above.
1891 if (N0.getOpcode() == ISD::SRL) {
1892 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
1893 if (ShAmt->getValue()+EVTBits <= MVT::getSizeInBits(VT)) {
1894 // We can turn this into an SRA iff the input to the SRL is already sign
1896 unsigned InSignBits = TLI.ComputeNumSignBits(N0.getOperand(0));
1897 if (MVT::getSizeInBits(VT)-(ShAmt->getValue()+EVTBits) < InSignBits)
1898 return DAG.getNode(ISD::SRA, VT, N0.getOperand(0), N0.getOperand(1));
1902 // fold (sext_inreg (extload x)) -> (sextload x)
1903 if (N0.getOpcode() == ISD::EXTLOAD &&
1904 EVT == cast<VTSDNode>(N0.getOperand(3))->getVT() &&
1905 (!AfterLegalize || TLI.isOperationLegal(ISD::SEXTLOAD, EVT))) {
1906 SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, N0.getOperand(0),
1907 N0.getOperand(1), N0.getOperand(2),
1909 CombineTo(N, ExtLoad);
1910 CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1));
1911 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
1913 // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use
1914 if (N0.getOpcode() == ISD::ZEXTLOAD && N0.hasOneUse() &&
1915 EVT == cast<VTSDNode>(N0.getOperand(3))->getVT() &&
1916 (!AfterLegalize || TLI.isOperationLegal(ISD::SEXTLOAD, EVT))) {
1917 SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, N0.getOperand(0),
1918 N0.getOperand(1), N0.getOperand(2),
1920 CombineTo(N, ExtLoad);
1921 CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1));
1922 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
1927 SDOperand DAGCombiner::visitTRUNCATE(SDNode *N) {
1928 SDOperand N0 = N->getOperand(0);
1929 MVT::ValueType VT = N->getValueType(0);
1932 if (N0.getValueType() == N->getValueType(0))
1934 // fold (truncate c1) -> c1
1935 if (isa<ConstantSDNode>(N0))
1936 return DAG.getNode(ISD::TRUNCATE, VT, N0);
1937 // fold (truncate (truncate x)) -> (truncate x)
1938 if (N0.getOpcode() == ISD::TRUNCATE)
1939 return DAG.getNode(ISD::TRUNCATE, VT, N0.getOperand(0));
1940 // fold (truncate (ext x)) -> (ext x) or (truncate x) or x
1941 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::SIGN_EXTEND||
1942 N0.getOpcode() == ISD::ANY_EXTEND) {
1943 if (N0.getValueType() < VT)
1944 // if the source is smaller than the dest, we still need an extend
1945 return DAG.getNode(N0.getOpcode(), VT, N0.getOperand(0));
1946 else if (N0.getValueType() > VT)
1947 // if the source is larger than the dest, than we just need the truncate
1948 return DAG.getNode(ISD::TRUNCATE, VT, N0.getOperand(0));
1950 // if the source and dest are the same type, we can drop both the extend
1952 return N0.getOperand(0);
1954 // fold (truncate (load x)) -> (smaller load x)
1955 if (N0.getOpcode() == ISD::LOAD && N0.hasOneUse()) {
1956 assert(MVT::getSizeInBits(N0.getValueType()) > MVT::getSizeInBits(VT) &&
1957 "Cannot truncate to larger type!");
1958 MVT::ValueType PtrType = N0.getOperand(1).getValueType();
1959 // For big endian targets, we need to add an offset to the pointer to load
1960 // the correct bytes. For little endian systems, we merely need to read
1961 // fewer bytes from the same pointer.
1963 (MVT::getSizeInBits(N0.getValueType()) - MVT::getSizeInBits(VT)) / 8;
1964 SDOperand NewPtr = TLI.isLittleEndian() ? N0.getOperand(1) :
1965 DAG.getNode(ISD::ADD, PtrType, N0.getOperand(1),
1966 DAG.getConstant(PtrOff, PtrType));
1967 AddToWorkList(NewPtr.Val);
1968 SDOperand Load = DAG.getLoad(VT, N0.getOperand(0), NewPtr,N0.getOperand(2));
1970 CombineTo(N0.Val, Load, Load.getValue(1));
1971 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
1976 SDOperand DAGCombiner::visitBIT_CONVERT(SDNode *N) {
1977 SDOperand N0 = N->getOperand(0);
1978 MVT::ValueType VT = N->getValueType(0);
1980 // If the input is a constant, let getNode() fold it.
1981 if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) {
1982 SDOperand Res = DAG.getNode(ISD::BIT_CONVERT, VT, N0);
1983 if (Res.Val != N) return Res;
1986 if (N0.getOpcode() == ISD::BIT_CONVERT) // conv(conv(x,t1),t2) -> conv(x,t2)
1987 return DAG.getNode(ISD::BIT_CONVERT, VT, N0.getOperand(0));
1989 // fold (conv (load x)) -> (load (conv*)x)
1990 // FIXME: These xforms need to know that the resultant load doesn't need a
1991 // higher alignment than the original!
1992 if (0 && N0.getOpcode() == ISD::LOAD && N0.hasOneUse()) {
1993 SDOperand Load = DAG.getLoad(VT, N0.getOperand(0), N0.getOperand(1),
1996 CombineTo(N0.Val, DAG.getNode(ISD::BIT_CONVERT, N0.getValueType(), Load),
2004 SDOperand DAGCombiner::visitVBIT_CONVERT(SDNode *N) {
2005 SDOperand N0 = N->getOperand(0);
2006 MVT::ValueType VT = N->getValueType(0);
2008 // If the input is a VBUILD_VECTOR with all constant elements, fold this now.
2009 // First check to see if this is all constant.
2010 if (N0.getOpcode() == ISD::VBUILD_VECTOR && N0.Val->hasOneUse() &&
2011 VT == MVT::Vector) {
2012 bool isSimple = true;
2013 for (unsigned i = 0, e = N0.getNumOperands()-2; i != e; ++i)
2014 if (N0.getOperand(i).getOpcode() != ISD::UNDEF &&
2015 N0.getOperand(i).getOpcode() != ISD::Constant &&
2016 N0.getOperand(i).getOpcode() != ISD::ConstantFP) {
2021 MVT::ValueType DestEltVT = cast<VTSDNode>(N->getOperand(2))->getVT();
2022 if (isSimple && !MVT::isVector(DestEltVT)) {
2023 return ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(N0.Val, DestEltVT);
2030 /// ConstantFoldVBIT_CONVERTofVBUILD_VECTOR - We know that BV is a vbuild_vector
2031 /// node with Constant, ConstantFP or Undef operands. DstEltVT indicates the
2032 /// destination element value type.
2033 SDOperand DAGCombiner::
2034 ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(SDNode *BV, MVT::ValueType DstEltVT) {
2035 MVT::ValueType SrcEltVT = BV->getOperand(0).getValueType();
2037 // If this is already the right type, we're done.
2038 if (SrcEltVT == DstEltVT) return SDOperand(BV, 0);
2040 unsigned SrcBitSize = MVT::getSizeInBits(SrcEltVT);
2041 unsigned DstBitSize = MVT::getSizeInBits(DstEltVT);
2043 // If this is a conversion of N elements of one type to N elements of another
2044 // type, convert each element. This handles FP<->INT cases.
2045 if (SrcBitSize == DstBitSize) {
2046 SmallVector<SDOperand, 8> Ops;
2047 for (unsigned i = 0, e = BV->getNumOperands()-2; i != e; ++i) {
2048 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, DstEltVT, BV->getOperand(i)));
2049 AddToWorkList(Ops.back().Val);
2051 Ops.push_back(*(BV->op_end()-2)); // Add num elements.
2052 Ops.push_back(DAG.getValueType(DstEltVT));
2053 return DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, &Ops[0], Ops.size());
2056 // Otherwise, we're growing or shrinking the elements. To avoid having to
2057 // handle annoying details of growing/shrinking FP values, we convert them to
2059 if (MVT::isFloatingPoint(SrcEltVT)) {
2060 // Convert the input float vector to a int vector where the elements are the
2062 assert((SrcEltVT == MVT::f32 || SrcEltVT == MVT::f64) && "Unknown FP VT!");
2063 MVT::ValueType IntVT = SrcEltVT == MVT::f32 ? MVT::i32 : MVT::i64;
2064 BV = ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(BV, IntVT).Val;
2068 // Now we know the input is an integer vector. If the output is a FP type,
2069 // convert to integer first, then to FP of the right size.
2070 if (MVT::isFloatingPoint(DstEltVT)) {
2071 assert((DstEltVT == MVT::f32 || DstEltVT == MVT::f64) && "Unknown FP VT!");
2072 MVT::ValueType TmpVT = DstEltVT == MVT::f32 ? MVT::i32 : MVT::i64;
2073 SDNode *Tmp = ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(BV, TmpVT).Val;
2075 // Next, convert to FP elements of the same size.
2076 return ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(Tmp, DstEltVT);
2079 // Okay, we know the src/dst types are both integers of differing types.
2080 // Handling growing first.
2081 assert(MVT::isInteger(SrcEltVT) && MVT::isInteger(DstEltVT));
2082 if (SrcBitSize < DstBitSize) {
2083 unsigned NumInputsPerOutput = DstBitSize/SrcBitSize;
2085 SmallVector<SDOperand, 8> Ops;
2086 for (unsigned i = 0, e = BV->getNumOperands()-2; i != e;
2087 i += NumInputsPerOutput) {
2088 bool isLE = TLI.isLittleEndian();
2089 uint64_t NewBits = 0;
2090 bool EltIsUndef = true;
2091 for (unsigned j = 0; j != NumInputsPerOutput; ++j) {
2092 // Shift the previously computed bits over.
2093 NewBits <<= SrcBitSize;
2094 SDOperand Op = BV->getOperand(i+ (isLE ? (NumInputsPerOutput-j-1) : j));
2095 if (Op.getOpcode() == ISD::UNDEF) continue;
2098 NewBits |= cast<ConstantSDNode>(Op)->getValue();
2102 Ops.push_back(DAG.getNode(ISD::UNDEF, DstEltVT));
2104 Ops.push_back(DAG.getConstant(NewBits, DstEltVT));
2107 Ops.push_back(DAG.getConstant(Ops.size(), MVT::i32)); // Add num elements.
2108 Ops.push_back(DAG.getValueType(DstEltVT)); // Add element size.
2109 return DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, &Ops[0], Ops.size());
2112 // Finally, this must be the case where we are shrinking elements: each input
2113 // turns into multiple outputs.
2114 unsigned NumOutputsPerInput = SrcBitSize/DstBitSize;
2115 SmallVector<SDOperand, 8> Ops;
2116 for (unsigned i = 0, e = BV->getNumOperands()-2; i != e; ++i) {
2117 if (BV->getOperand(i).getOpcode() == ISD::UNDEF) {
2118 for (unsigned j = 0; j != NumOutputsPerInput; ++j)
2119 Ops.push_back(DAG.getNode(ISD::UNDEF, DstEltVT));
2122 uint64_t OpVal = cast<ConstantSDNode>(BV->getOperand(i))->getValue();
2124 for (unsigned j = 0; j != NumOutputsPerInput; ++j) {
2125 unsigned ThisVal = OpVal & ((1ULL << DstBitSize)-1);
2126 OpVal >>= DstBitSize;
2127 Ops.push_back(DAG.getConstant(ThisVal, DstEltVT));
2130 // For big endian targets, swap the order of the pieces of each element.
2131 if (!TLI.isLittleEndian())
2132 std::reverse(Ops.end()-NumOutputsPerInput, Ops.end());
2134 Ops.push_back(DAG.getConstant(Ops.size(), MVT::i32)); // Add num elements.
2135 Ops.push_back(DAG.getValueType(DstEltVT)); // Add element size.
2136 return DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, &Ops[0], Ops.size());
2141 SDOperand DAGCombiner::visitFADD(SDNode *N) {
2142 SDOperand N0 = N->getOperand(0);
2143 SDOperand N1 = N->getOperand(1);
2144 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2145 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
2146 MVT::ValueType VT = N->getValueType(0);
2148 // fold (fadd c1, c2) -> c1+c2
2150 return DAG.getNode(ISD::FADD, VT, N0, N1);
2151 // canonicalize constant to RHS
2152 if (N0CFP && !N1CFP)
2153 return DAG.getNode(ISD::FADD, VT, N1, N0);
2154 // fold (A + (-B)) -> A-B
2155 if (N1.getOpcode() == ISD::FNEG)
2156 return DAG.getNode(ISD::FSUB, VT, N0, N1.getOperand(0));
2157 // fold ((-A) + B) -> B-A
2158 if (N0.getOpcode() == ISD::FNEG)
2159 return DAG.getNode(ISD::FSUB, VT, N1, N0.getOperand(0));
2163 SDOperand DAGCombiner::visitFSUB(SDNode *N) {
2164 SDOperand N0 = N->getOperand(0);
2165 SDOperand N1 = N->getOperand(1);
2166 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2167 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
2168 MVT::ValueType VT = N->getValueType(0);
2170 // fold (fsub c1, c2) -> c1-c2
2172 return DAG.getNode(ISD::FSUB, VT, N0, N1);
2173 // fold (A-(-B)) -> A+B
2174 if (N1.getOpcode() == ISD::FNEG)
2175 return DAG.getNode(ISD::FADD, VT, N0, N1.getOperand(0));
2179 SDOperand DAGCombiner::visitFMUL(SDNode *N) {
2180 SDOperand N0 = N->getOperand(0);
2181 SDOperand N1 = N->getOperand(1);
2182 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2183 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
2184 MVT::ValueType VT = N->getValueType(0);
2186 // fold (fmul c1, c2) -> c1*c2
2188 return DAG.getNode(ISD::FMUL, VT, N0, N1);
2189 // canonicalize constant to RHS
2190 if (N0CFP && !N1CFP)
2191 return DAG.getNode(ISD::FMUL, VT, N1, N0);
2192 // fold (fmul X, 2.0) -> (fadd X, X)
2193 if (N1CFP && N1CFP->isExactlyValue(+2.0))
2194 return DAG.getNode(ISD::FADD, VT, N0, N0);
2198 SDOperand DAGCombiner::visitFDIV(SDNode *N) {
2199 SDOperand N0 = N->getOperand(0);
2200 SDOperand N1 = N->getOperand(1);
2201 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2202 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
2203 MVT::ValueType VT = N->getValueType(0);
2205 // fold (fdiv c1, c2) -> c1/c2
2207 return DAG.getNode(ISD::FDIV, VT, N0, N1);
2211 SDOperand DAGCombiner::visitFREM(SDNode *N) {
2212 SDOperand N0 = N->getOperand(0);
2213 SDOperand N1 = N->getOperand(1);
2214 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2215 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
2216 MVT::ValueType VT = N->getValueType(0);
2218 // fold (frem c1, c2) -> fmod(c1,c2)
2220 return DAG.getNode(ISD::FREM, VT, N0, N1);
2224 SDOperand DAGCombiner::visitFCOPYSIGN(SDNode *N) {
2225 SDOperand N0 = N->getOperand(0);
2226 SDOperand N1 = N->getOperand(1);
2227 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2228 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
2229 MVT::ValueType VT = N->getValueType(0);
2231 if (N0CFP && N1CFP) // Constant fold
2232 return DAG.getNode(ISD::FCOPYSIGN, VT, N0, N1);
2235 // copysign(x, c1) -> fabs(x) iff ispos(c1)
2236 // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1)
2241 u.d = N1CFP->getValue();
2243 return DAG.getNode(ISD::FABS, VT, N0);
2245 return DAG.getNode(ISD::FNEG, VT, DAG.getNode(ISD::FABS, VT, N0));
2248 // copysign(fabs(x), y) -> copysign(x, y)
2249 // copysign(fneg(x), y) -> copysign(x, y)
2250 // copysign(copysign(x,z), y) -> copysign(x, y)
2251 if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG ||
2252 N0.getOpcode() == ISD::FCOPYSIGN)
2253 return DAG.getNode(ISD::FCOPYSIGN, VT, N0.getOperand(0), N1);
2255 // copysign(x, abs(y)) -> abs(x)
2256 if (N1.getOpcode() == ISD::FABS)
2257 return DAG.getNode(ISD::FABS, VT, N0);
2259 // copysign(x, copysign(y,z)) -> copysign(x, z)
2260 if (N1.getOpcode() == ISD::FCOPYSIGN)
2261 return DAG.getNode(ISD::FCOPYSIGN, VT, N0, N1.getOperand(1));
2263 // copysign(x, fp_extend(y)) -> copysign(x, y)
2264 // copysign(x, fp_round(y)) -> copysign(x, y)
2265 if (N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND)
2266 return DAG.getNode(ISD::FCOPYSIGN, VT, N0, N1.getOperand(0));
2273 SDOperand DAGCombiner::visitSINT_TO_FP(SDNode *N) {
2274 SDOperand N0 = N->getOperand(0);
2275 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2276 MVT::ValueType VT = N->getValueType(0);
2278 // fold (sint_to_fp c1) -> c1fp
2280 return DAG.getNode(ISD::SINT_TO_FP, VT, N0);
2284 SDOperand DAGCombiner::visitUINT_TO_FP(SDNode *N) {
2285 SDOperand N0 = N->getOperand(0);
2286 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2287 MVT::ValueType VT = N->getValueType(0);
2289 // fold (uint_to_fp c1) -> c1fp
2291 return DAG.getNode(ISD::UINT_TO_FP, VT, N0);
2295 SDOperand DAGCombiner::visitFP_TO_SINT(SDNode *N) {
2296 SDOperand N0 = N->getOperand(0);
2297 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2298 MVT::ValueType VT = N->getValueType(0);
2300 // fold (fp_to_sint c1fp) -> c1
2302 return DAG.getNode(ISD::FP_TO_SINT, VT, N0);
2306 SDOperand DAGCombiner::visitFP_TO_UINT(SDNode *N) {
2307 SDOperand N0 = N->getOperand(0);
2308 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2309 MVT::ValueType VT = N->getValueType(0);
2311 // fold (fp_to_uint c1fp) -> c1
2313 return DAG.getNode(ISD::FP_TO_UINT, VT, N0);
2317 SDOperand DAGCombiner::visitFP_ROUND(SDNode *N) {
2318 SDOperand N0 = N->getOperand(0);
2319 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2320 MVT::ValueType VT = N->getValueType(0);
2322 // fold (fp_round c1fp) -> c1fp
2324 return DAG.getNode(ISD::FP_ROUND, VT, N0);
2326 // fold (fp_round (fp_extend x)) -> x
2327 if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType())
2328 return N0.getOperand(0);
2330 // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y)
2331 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.Val->hasOneUse()) {
2332 SDOperand Tmp = DAG.getNode(ISD::FP_ROUND, VT, N0.getOperand(0));
2333 AddToWorkList(Tmp.Val);
2334 return DAG.getNode(ISD::FCOPYSIGN, VT, Tmp, N0.getOperand(1));
2340 SDOperand DAGCombiner::visitFP_ROUND_INREG(SDNode *N) {
2341 SDOperand N0 = N->getOperand(0);
2342 MVT::ValueType VT = N->getValueType(0);
2343 MVT::ValueType EVT = cast<VTSDNode>(N->getOperand(1))->getVT();
2344 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2346 // fold (fp_round_inreg c1fp) -> c1fp
2348 SDOperand Round = DAG.getConstantFP(N0CFP->getValue(), EVT);
2349 return DAG.getNode(ISD::FP_EXTEND, VT, Round);
2354 SDOperand DAGCombiner::visitFP_EXTEND(SDNode *N) {
2355 SDOperand N0 = N->getOperand(0);
2356 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2357 MVT::ValueType VT = N->getValueType(0);
2359 // fold (fp_extend c1fp) -> c1fp
2361 return DAG.getNode(ISD::FP_EXTEND, VT, N0);
2363 // fold (fpext (load x)) -> (fpext (fpround (extload x)))
2364 if (N0.getOpcode() == ISD::LOAD && N0.hasOneUse() &&
2365 (!AfterLegalize||TLI.isOperationLegal(ISD::EXTLOAD, N0.getValueType()))) {
2366 SDOperand ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, VT, N0.getOperand(0),
2367 N0.getOperand(1), N0.getOperand(2),
2369 CombineTo(N, ExtLoad);
2370 CombineTo(N0.Val, DAG.getNode(ISD::FP_ROUND, N0.getValueType(), ExtLoad),
2371 ExtLoad.getValue(1));
2372 return SDOperand(N, 0); // Return N so it doesn't get rechecked!
2379 SDOperand DAGCombiner::visitFNEG(SDNode *N) {
2380 SDOperand N0 = N->getOperand(0);
2381 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2382 MVT::ValueType VT = N->getValueType(0);
2384 // fold (fneg c1) -> -c1
2386 return DAG.getNode(ISD::FNEG, VT, N0);
2387 // fold (fneg (sub x, y)) -> (sub y, x)
2388 if (N0.getOpcode() == ISD::SUB)
2389 return DAG.getNode(ISD::SUB, VT, N0.getOperand(1), N0.getOperand(0));
2390 // fold (fneg (fneg x)) -> x
2391 if (N0.getOpcode() == ISD::FNEG)
2392 return N0.getOperand(0);
2396 SDOperand DAGCombiner::visitFABS(SDNode *N) {
2397 SDOperand N0 = N->getOperand(0);
2398 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2399 MVT::ValueType VT = N->getValueType(0);
2401 // fold (fabs c1) -> fabs(c1)
2403 return DAG.getNode(ISD::FABS, VT, N0);
2404 // fold (fabs (fabs x)) -> (fabs x)
2405 if (N0.getOpcode() == ISD::FABS)
2406 return N->getOperand(0);
2407 // fold (fabs (fneg x)) -> (fabs x)
2408 // fold (fabs (fcopysign x, y)) -> (fabs x)
2409 if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN)
2410 return DAG.getNode(ISD::FABS, VT, N0.getOperand(0));
2415 SDOperand DAGCombiner::visitBRCOND(SDNode *N) {
2416 SDOperand Chain = N->getOperand(0);
2417 SDOperand N1 = N->getOperand(1);
2418 SDOperand N2 = N->getOperand(2);
2419 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2421 // never taken branch, fold to chain
2422 if (N1C && N1C->isNullValue())
2424 // unconditional branch
2425 if (N1C && N1C->getValue() == 1)
2426 return DAG.getNode(ISD::BR, MVT::Other, Chain, N2);
2427 // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal
2429 if (N1.getOpcode() == ISD::SETCC &&
2430 TLI.isOperationLegal(ISD::BR_CC, MVT::Other)) {
2431 return DAG.getNode(ISD::BR_CC, MVT::Other, Chain, N1.getOperand(2),
2432 N1.getOperand(0), N1.getOperand(1), N2);
2437 // Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB.
2439 SDOperand DAGCombiner::visitBR_CC(SDNode *N) {
2440 CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1));
2441 SDOperand CondLHS = N->getOperand(2), CondRHS = N->getOperand(3);
2443 // Use SimplifySetCC to simplify SETCC's.
2444 SDOperand Simp = SimplifySetCC(MVT::i1, CondLHS, CondRHS, CC->get(), false);
2445 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(Simp.Val);
2447 // fold br_cc true, dest -> br dest (unconditional branch)
2448 if (SCCC && SCCC->getValue())
2449 return DAG.getNode(ISD::BR, MVT::Other, N->getOperand(0),
2451 // fold br_cc false, dest -> unconditional fall through
2452 if (SCCC && SCCC->isNullValue())
2453 return N->getOperand(0);
2454 // fold to a simpler setcc
2455 if (Simp.Val && Simp.getOpcode() == ISD::SETCC)
2456 return DAG.getNode(ISD::BR_CC, MVT::Other, N->getOperand(0),
2457 Simp.getOperand(2), Simp.getOperand(0),
2458 Simp.getOperand(1), N->getOperand(4));
2462 SDOperand DAGCombiner::visitLOAD(SDNode *N) {
2463 SDOperand Chain = N->getOperand(0);
2464 SDOperand Ptr = N->getOperand(1);
2465 SDOperand SrcValue = N->getOperand(2);
2467 // If there are no uses of the loaded value, change uses of the chain value
2468 // into uses of the chain input (i.e. delete the dead load).
2469 if (N->hasNUsesOfValue(0, 0))
2470 return CombineTo(N, DAG.getNode(ISD::UNDEF, N->getValueType(0)), Chain);
2472 // If this load is directly stored, replace the load value with the stored
2474 // TODO: Handle store large -> read small portion.
2475 // TODO: Handle TRUNCSTORE/EXTLOAD
2476 if (Chain.getOpcode() == ISD::STORE && Chain.getOperand(2) == Ptr &&
2477 Chain.getOperand(1).getValueType() == N->getValueType(0))
2478 return CombineTo(N, Chain.getOperand(1), Chain);
2483 /// visitXEXTLOAD - Handle EXTLOAD/ZEXTLOAD/SEXTLOAD.
2484 SDOperand DAGCombiner::visitXEXTLOAD(SDNode *N) {
2485 SDOperand Chain = N->getOperand(0);
2486 SDOperand Ptr = N->getOperand(1);
2487 SDOperand SrcValue = N->getOperand(2);
2488 SDOperand EVT = N->getOperand(3);
2490 // If there are no uses of the loaded value, change uses of the chain value
2491 // into uses of the chain input (i.e. delete the dead load).
2492 if (N->hasNUsesOfValue(0, 0))
2493 return CombineTo(N, DAG.getNode(ISD::UNDEF, N->getValueType(0)), Chain);
2498 SDOperand DAGCombiner::visitSTORE(SDNode *N) {
2499 SDOperand Chain = N->getOperand(0);
2500 SDOperand Value = N->getOperand(1);
2501 SDOperand Ptr = N->getOperand(2);
2502 SDOperand SrcValue = N->getOperand(3);
2504 // If this is a store that kills a previous store, remove the previous store.
2505 if (Chain.getOpcode() == ISD::STORE && Chain.getOperand(2) == Ptr &&
2506 Chain.Val->hasOneUse() /* Avoid introducing DAG cycles */ &&
2507 // Make sure that these stores are the same value type:
2508 // FIXME: we really care that the second store is >= size of the first.
2509 Value.getValueType() == Chain.getOperand(1).getValueType()) {
2510 // Create a new store of Value that replaces both stores.
2511 SDNode *PrevStore = Chain.Val;
2512 if (PrevStore->getOperand(1) == Value) // Same value multiply stored.
2514 SDOperand NewStore = DAG.getNode(ISD::STORE, MVT::Other,
2515 PrevStore->getOperand(0), Value, Ptr,
2517 CombineTo(N, NewStore); // Nuke this store.
2518 CombineTo(PrevStore, NewStore); // Nuke the previous store.
2519 return SDOperand(N, 0);
2522 // If this is a store of a bit convert, store the input value.
2523 // FIXME: This needs to know that the resultant store does not need a
2524 // higher alignment than the original.
2525 if (0 && Value.getOpcode() == ISD::BIT_CONVERT)
2526 return DAG.getNode(ISD::STORE, MVT::Other, Chain, Value.getOperand(0),
2532 SDOperand DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) {
2533 SDOperand InVec = N->getOperand(0);
2534 SDOperand InVal = N->getOperand(1);
2535 SDOperand EltNo = N->getOperand(2);
2537 // If the invec is a BUILD_VECTOR and if EltNo is a constant, build a new
2538 // vector with the inserted element.
2539 if (InVec.getOpcode() == ISD::BUILD_VECTOR && isa<ConstantSDNode>(EltNo)) {
2540 unsigned Elt = cast<ConstantSDNode>(EltNo)->getValue();
2541 SmallVector<SDOperand, 8> Ops(InVec.Val->op_begin(), InVec.Val->op_end());
2542 if (Elt < Ops.size())
2544 return DAG.getNode(ISD::BUILD_VECTOR, InVec.getValueType(),
2545 &Ops[0], Ops.size());
2551 SDOperand DAGCombiner::visitVINSERT_VECTOR_ELT(SDNode *N) {
2552 SDOperand InVec = N->getOperand(0);
2553 SDOperand InVal = N->getOperand(1);
2554 SDOperand EltNo = N->getOperand(2);
2555 SDOperand NumElts = N->getOperand(3);
2556 SDOperand EltType = N->getOperand(4);
2558 // If the invec is a VBUILD_VECTOR and if EltNo is a constant, build a new
2559 // vector with the inserted element.
2560 if (InVec.getOpcode() == ISD::VBUILD_VECTOR && isa<ConstantSDNode>(EltNo)) {
2561 unsigned Elt = cast<ConstantSDNode>(EltNo)->getValue();
2562 SmallVector<SDOperand, 8> Ops(InVec.Val->op_begin(), InVec.Val->op_end());
2563 if (Elt < Ops.size()-2)
2565 return DAG.getNode(ISD::VBUILD_VECTOR, InVec.getValueType(),
2566 &Ops[0], Ops.size());
2572 SDOperand DAGCombiner::visitVBUILD_VECTOR(SDNode *N) {
2573 unsigned NumInScalars = N->getNumOperands()-2;
2574 SDOperand NumElts = N->getOperand(NumInScalars);
2575 SDOperand EltType = N->getOperand(NumInScalars+1);
2577 // Check to see if this is a VBUILD_VECTOR of a bunch of VEXTRACT_VECTOR_ELT
2578 // operations. If so, and if the EXTRACT_ELT vector inputs come from at most
2579 // two distinct vectors, turn this into a shuffle node.
2580 SDOperand VecIn1, VecIn2;
2581 for (unsigned i = 0; i != NumInScalars; ++i) {
2582 // Ignore undef inputs.
2583 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
2585 // If this input is something other than a VEXTRACT_VECTOR_ELT with a
2586 // constant index, bail out.
2587 if (N->getOperand(i).getOpcode() != ISD::VEXTRACT_VECTOR_ELT ||
2588 !isa<ConstantSDNode>(N->getOperand(i).getOperand(1))) {
2589 VecIn1 = VecIn2 = SDOperand(0, 0);
2593 // If the input vector type disagrees with the result of the vbuild_vector,
2594 // we can't make a shuffle.
2595 SDOperand ExtractedFromVec = N->getOperand(i).getOperand(0);
2596 if (*(ExtractedFromVec.Val->op_end()-2) != NumElts ||
2597 *(ExtractedFromVec.Val->op_end()-1) != EltType) {
2598 VecIn1 = VecIn2 = SDOperand(0, 0);
2602 // Otherwise, remember this. We allow up to two distinct input vectors.
2603 if (ExtractedFromVec == VecIn1 || ExtractedFromVec == VecIn2)
2606 if (VecIn1.Val == 0) {
2607 VecIn1 = ExtractedFromVec;
2608 } else if (VecIn2.Val == 0) {
2609 VecIn2 = ExtractedFromVec;
2612 VecIn1 = VecIn2 = SDOperand(0, 0);
2617 // If everything is good, we can make a shuffle operation.
2619 SmallVector<SDOperand, 8> BuildVecIndices;
2620 for (unsigned i = 0; i != NumInScalars; ++i) {
2621 if (N->getOperand(i).getOpcode() == ISD::UNDEF) {
2622 BuildVecIndices.push_back(DAG.getNode(ISD::UNDEF, MVT::i32));
2626 SDOperand Extract = N->getOperand(i);
2628 // If extracting from the first vector, just use the index directly.
2629 if (Extract.getOperand(0) == VecIn1) {
2630 BuildVecIndices.push_back(Extract.getOperand(1));
2634 // Otherwise, use InIdx + VecSize
2635 unsigned Idx = cast<ConstantSDNode>(Extract.getOperand(1))->getValue();
2636 BuildVecIndices.push_back(DAG.getConstant(Idx+NumInScalars, MVT::i32));
2639 // Add count and size info.
2640 BuildVecIndices.push_back(NumElts);
2641 BuildVecIndices.push_back(DAG.getValueType(MVT::i32));
2643 // Return the new VVECTOR_SHUFFLE node.
2649 // Use an undef vbuild_vector as input for the second operand.
2650 std::vector<SDOperand> UnOps(NumInScalars,
2651 DAG.getNode(ISD::UNDEF,
2652 cast<VTSDNode>(EltType)->getVT()));
2653 UnOps.push_back(NumElts);
2654 UnOps.push_back(EltType);
2655 Ops[1] = DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector,
2656 &UnOps[0], UnOps.size());
2657 AddToWorkList(Ops[1].Val);
2659 Ops[2] = DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector,
2660 &BuildVecIndices[0], BuildVecIndices.size());
2663 return DAG.getNode(ISD::VVECTOR_SHUFFLE, MVT::Vector, Ops, 5);
2669 SDOperand DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
2670 SDOperand ShufMask = N->getOperand(2);
2671 unsigned NumElts = ShufMask.getNumOperands();
2673 // If the shuffle mask is an identity operation on the LHS, return the LHS.
2674 bool isIdentity = true;
2675 for (unsigned i = 0; i != NumElts; ++i) {
2676 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF &&
2677 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() != i) {
2682 if (isIdentity) return N->getOperand(0);
2684 // If the shuffle mask is an identity operation on the RHS, return the RHS.
2686 for (unsigned i = 0; i != NumElts; ++i) {
2687 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF &&
2688 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() != i+NumElts) {
2693 if (isIdentity) return N->getOperand(1);
2695 // Check if the shuffle is a unary shuffle, i.e. one of the vectors is not
2697 bool isUnary = true;
2698 bool isSplat = true;
2700 unsigned BaseIdx = 0;
2701 for (unsigned i = 0; i != NumElts; ++i)
2702 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF) {
2703 unsigned Idx = cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue();
2704 int V = (Idx < NumElts) ? 0 : 1;
2718 SDOperand N0 = N->getOperand(0);
2719 SDOperand N1 = N->getOperand(1);
2720 // Normalize unary shuffle so the RHS is undef.
2721 if (isUnary && VecNum == 1)
2724 // If it is a splat, check if the argument vector is a build_vector with
2725 // all scalar elements the same.
2728 if (V->getOpcode() == ISD::BIT_CONVERT)
2729 V = V->getOperand(0).Val;
2730 if (V->getOpcode() == ISD::BUILD_VECTOR) {
2731 unsigned NumElems = V->getNumOperands()-2;
2732 if (NumElems > BaseIdx) {
2734 bool AllSame = true;
2735 for (unsigned i = 0; i != NumElems; ++i) {
2736 if (V->getOperand(i).getOpcode() != ISD::UNDEF) {
2737 Base = V->getOperand(i);
2741 // Splat of <u, u, u, u>, return <u, u, u, u>
2744 for (unsigned i = 0; i != NumElems; ++i) {
2745 if (V->getOperand(i).getOpcode() != ISD::UNDEF &&
2746 V->getOperand(i) != Base) {
2751 // Splat of <x, x, x, x>, return <x, x, x, x>
2758 // If it is a unary or the LHS and the RHS are the same node, turn the RHS
2760 if (isUnary || N0 == N1) {
2761 if (N0.getOpcode() == ISD::UNDEF)
2762 return DAG.getNode(ISD::UNDEF, N->getValueType(0));
2763 // Check the SHUFFLE mask, mapping any inputs from the 2nd operand into the
2765 SmallVector<SDOperand, 8> MappedOps;
2766 for (unsigned i = 0, e = ShufMask.getNumOperands(); i != e; ++i) {
2767 if (ShufMask.getOperand(i).getOpcode() == ISD::UNDEF ||
2768 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() < NumElts) {
2769 MappedOps.push_back(ShufMask.getOperand(i));
2772 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() - NumElts;
2773 MappedOps.push_back(DAG.getConstant(NewIdx, MVT::i32));
2776 ShufMask = DAG.getNode(ISD::BUILD_VECTOR, ShufMask.getValueType(),
2777 &MappedOps[0], MappedOps.size());
2778 AddToWorkList(ShufMask.Val);
2779 return DAG.getNode(ISD::VECTOR_SHUFFLE, N->getValueType(0),
2781 DAG.getNode(ISD::UNDEF, N->getValueType(0)),
2788 SDOperand DAGCombiner::visitVVECTOR_SHUFFLE(SDNode *N) {
2789 SDOperand ShufMask = N->getOperand(2);
2790 unsigned NumElts = ShufMask.getNumOperands()-2;
2792 // If the shuffle mask is an identity operation on the LHS, return the LHS.
2793 bool isIdentity = true;
2794 for (unsigned i = 0; i != NumElts; ++i) {
2795 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF &&
2796 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() != i) {
2801 if (isIdentity) return N->getOperand(0);
2803 // If the shuffle mask is an identity operation on the RHS, return the RHS.
2805 for (unsigned i = 0; i != NumElts; ++i) {
2806 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF &&
2807 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() != i+NumElts) {
2812 if (isIdentity) return N->getOperand(1);
2814 // Check if the shuffle is a unary shuffle, i.e. one of the vectors is not
2816 bool isUnary = true;
2817 bool isSplat = true;
2819 unsigned BaseIdx = 0;
2820 for (unsigned i = 0; i != NumElts; ++i)
2821 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF) {
2822 unsigned Idx = cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue();
2823 int V = (Idx < NumElts) ? 0 : 1;
2837 SDOperand N0 = N->getOperand(0);
2838 SDOperand N1 = N->getOperand(1);
2839 // Normalize unary shuffle so the RHS is undef.
2840 if (isUnary && VecNum == 1)
2843 // If it is a splat, check if the argument vector is a build_vector with
2844 // all scalar elements the same.
2847 if (V->getOpcode() == ISD::VBIT_CONVERT)
2848 V = V->getOperand(0).Val;
2849 if (V->getOpcode() == ISD::VBUILD_VECTOR) {
2850 unsigned NumElems = V->getNumOperands()-2;
2851 if (NumElems > BaseIdx) {
2853 bool AllSame = true;
2854 for (unsigned i = 0; i != NumElems; ++i) {
2855 if (V->getOperand(i).getOpcode() != ISD::UNDEF) {
2856 Base = V->getOperand(i);
2860 // Splat of <u, u, u, u>, return <u, u, u, u>
2863 for (unsigned i = 0; i != NumElems; ++i) {
2864 if (V->getOperand(i).getOpcode() != ISD::UNDEF &&
2865 V->getOperand(i) != Base) {
2870 // Splat of <x, x, x, x>, return <x, x, x, x>
2877 // If it is a unary or the LHS and the RHS are the same node, turn the RHS
2879 if (isUnary || N0 == N1) {
2880 // Check the SHUFFLE mask, mapping any inputs from the 2nd operand into the
2882 SmallVector<SDOperand, 8> MappedOps;
2883 for (unsigned i = 0; i != NumElts; ++i) {
2884 if (ShufMask.getOperand(i).getOpcode() == ISD::UNDEF ||
2885 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() < NumElts) {
2886 MappedOps.push_back(ShufMask.getOperand(i));
2889 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() - NumElts;
2890 MappedOps.push_back(DAG.getConstant(NewIdx, MVT::i32));
2893 // Add the type/#elts values.
2894 MappedOps.push_back(ShufMask.getOperand(NumElts));
2895 MappedOps.push_back(ShufMask.getOperand(NumElts+1));
2897 ShufMask = DAG.getNode(ISD::VBUILD_VECTOR, ShufMask.getValueType(),
2898 &MappedOps[0], MappedOps.size());
2899 AddToWorkList(ShufMask.Val);
2901 // Build the undef vector.
2902 SDOperand UDVal = DAG.getNode(ISD::UNDEF, MappedOps[0].getValueType());
2903 for (unsigned i = 0; i != NumElts; ++i)
2904 MappedOps[i] = UDVal;
2905 MappedOps[NumElts ] = *(N0.Val->op_end()-2);
2906 MappedOps[NumElts+1] = *(N0.Val->op_end()-1);
2907 UDVal = DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector,
2908 &MappedOps[0], MappedOps.size());
2910 return DAG.getNode(ISD::VVECTOR_SHUFFLE, MVT::Vector,
2911 N0, UDVal, ShufMask,
2912 MappedOps[NumElts], MappedOps[NumElts+1]);
2918 /// XformToShuffleWithZero - Returns a vector_shuffle if it able to transform
2919 /// a VAND to a vector_shuffle with the destination vector and a zero vector.
2920 /// e.g. VAND V, <0xffffffff, 0, 0xffffffff, 0>. ==>
2921 /// vector_shuffle V, Zero, <0, 4, 2, 4>
2922 SDOperand DAGCombiner::XformToShuffleWithZero(SDNode *N) {
2923 SDOperand LHS = N->getOperand(0);
2924 SDOperand RHS = N->getOperand(1);
2925 if (N->getOpcode() == ISD::VAND) {
2926 SDOperand DstVecSize = *(LHS.Val->op_end()-2);
2927 SDOperand DstVecEVT = *(LHS.Val->op_end()-1);
2928 if (RHS.getOpcode() == ISD::VBIT_CONVERT)
2929 RHS = RHS.getOperand(0);
2930 if (RHS.getOpcode() == ISD::VBUILD_VECTOR) {
2931 std::vector<SDOperand> IdxOps;
2932 unsigned NumOps = RHS.getNumOperands();
2933 unsigned NumElts = NumOps-2;
2934 MVT::ValueType EVT = cast<VTSDNode>(RHS.getOperand(NumOps-1))->getVT();
2935 for (unsigned i = 0; i != NumElts; ++i) {
2936 SDOperand Elt = RHS.getOperand(i);
2937 if (!isa<ConstantSDNode>(Elt))
2939 else if (cast<ConstantSDNode>(Elt)->isAllOnesValue())
2940 IdxOps.push_back(DAG.getConstant(i, EVT));
2941 else if (cast<ConstantSDNode>(Elt)->isNullValue())
2942 IdxOps.push_back(DAG.getConstant(NumElts, EVT));
2947 // Let's see if the target supports this vector_shuffle.
2948 if (!TLI.isVectorClearMaskLegal(IdxOps, EVT, DAG))
2951 // Return the new VVECTOR_SHUFFLE node.
2952 SDOperand NumEltsNode = DAG.getConstant(NumElts, MVT::i32);
2953 SDOperand EVTNode = DAG.getValueType(EVT);
2954 std::vector<SDOperand> Ops;
2955 LHS = DAG.getNode(ISD::VBIT_CONVERT, MVT::Vector, LHS, NumEltsNode,
2958 AddToWorkList(LHS.Val);
2959 std::vector<SDOperand> ZeroOps(NumElts, DAG.getConstant(0, EVT));
2960 ZeroOps.push_back(NumEltsNode);
2961 ZeroOps.push_back(EVTNode);
2962 Ops.push_back(DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector,
2963 &ZeroOps[0], ZeroOps.size()));
2964 IdxOps.push_back(NumEltsNode);
2965 IdxOps.push_back(EVTNode);
2966 Ops.push_back(DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector,
2967 &IdxOps[0], IdxOps.size()));
2968 Ops.push_back(NumEltsNode);
2969 Ops.push_back(EVTNode);
2970 SDOperand Result = DAG.getNode(ISD::VVECTOR_SHUFFLE, MVT::Vector,
2971 &Ops[0], Ops.size());
2972 if (NumEltsNode != DstVecSize || EVTNode != DstVecEVT) {
2973 Result = DAG.getNode(ISD::VBIT_CONVERT, MVT::Vector, Result,
2974 DstVecSize, DstVecEVT);
2982 /// visitVBinOp - Visit a binary vector operation, like VADD. IntOp indicates
2983 /// the scalar operation of the vop if it is operating on an integer vector
2984 /// (e.g. ADD) and FPOp indicates the FP version (e.g. FADD).
2985 SDOperand DAGCombiner::visitVBinOp(SDNode *N, ISD::NodeType IntOp,
2986 ISD::NodeType FPOp) {
2987 MVT::ValueType EltType = cast<VTSDNode>(*(N->op_end()-1))->getVT();
2988 ISD::NodeType ScalarOp = MVT::isInteger(EltType) ? IntOp : FPOp;
2989 SDOperand LHS = N->getOperand(0);
2990 SDOperand RHS = N->getOperand(1);
2991 SDOperand Shuffle = XformToShuffleWithZero(N);
2992 if (Shuffle.Val) return Shuffle;
2994 // If the LHS and RHS are VBUILD_VECTOR nodes, see if we can constant fold
2996 if (LHS.getOpcode() == ISD::VBUILD_VECTOR &&
2997 RHS.getOpcode() == ISD::VBUILD_VECTOR) {
2998 SmallVector<SDOperand, 8> Ops;
2999 for (unsigned i = 0, e = LHS.getNumOperands()-2; i != e; ++i) {
3000 SDOperand LHSOp = LHS.getOperand(i);
3001 SDOperand RHSOp = RHS.getOperand(i);
3002 // If these two elements can't be folded, bail out.
3003 if ((LHSOp.getOpcode() != ISD::UNDEF &&
3004 LHSOp.getOpcode() != ISD::Constant &&
3005 LHSOp.getOpcode() != ISD::ConstantFP) ||
3006 (RHSOp.getOpcode() != ISD::UNDEF &&
3007 RHSOp.getOpcode() != ISD::Constant &&
3008 RHSOp.getOpcode() != ISD::ConstantFP))
3010 // Can't fold divide by zero.
3011 if (N->getOpcode() == ISD::VSDIV || N->getOpcode() == ISD::VUDIV) {
3012 if ((RHSOp.getOpcode() == ISD::Constant &&
3013 cast<ConstantSDNode>(RHSOp.Val)->isNullValue()) ||
3014 (RHSOp.getOpcode() == ISD::ConstantFP &&
3015 !cast<ConstantFPSDNode>(RHSOp.Val)->getValue()))
3018 Ops.push_back(DAG.getNode(ScalarOp, EltType, LHSOp, RHSOp));
3019 AddToWorkList(Ops.back().Val);
3020 assert((Ops.back().getOpcode() == ISD::UNDEF ||
3021 Ops.back().getOpcode() == ISD::Constant ||
3022 Ops.back().getOpcode() == ISD::ConstantFP) &&
3023 "Scalar binop didn't fold!");
3026 if (Ops.size() == LHS.getNumOperands()-2) {
3027 Ops.push_back(*(LHS.Val->op_end()-2));
3028 Ops.push_back(*(LHS.Val->op_end()-1));
3029 return DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, &Ops[0], Ops.size());
3036 SDOperand DAGCombiner::SimplifySelect(SDOperand N0, SDOperand N1, SDOperand N2){
3037 assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!");
3039 SDOperand SCC = SimplifySelectCC(N0.getOperand(0), N0.getOperand(1), N1, N2,
3040 cast<CondCodeSDNode>(N0.getOperand(2))->get());
3041 // If we got a simplified select_cc node back from SimplifySelectCC, then
3042 // break it down into a new SETCC node, and a new SELECT node, and then return
3043 // the SELECT node, since we were called with a SELECT node.
3045 // Check to see if we got a select_cc back (to turn into setcc/select).
3046 // Otherwise, just return whatever node we got back, like fabs.
3047 if (SCC.getOpcode() == ISD::SELECT_CC) {
3048 SDOperand SETCC = DAG.getNode(ISD::SETCC, N0.getValueType(),
3049 SCC.getOperand(0), SCC.getOperand(1),
3051 AddToWorkList(SETCC.Val);
3052 return DAG.getNode(ISD::SELECT, SCC.getValueType(), SCC.getOperand(2),
3053 SCC.getOperand(3), SETCC);
3060 /// SimplifySelectOps - Given a SELECT or a SELECT_CC node, where LHS and RHS
3061 /// are the two values being selected between, see if we can simplify the
3062 /// select. Callers of this should assume that TheSelect is deleted if this
3063 /// returns true. As such, they should return the appropriate thing (e.g. the
3064 /// node) back to the top-level of the DAG combiner loop to avoid it being
3067 bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDOperand LHS,
3070 // If this is a select from two identical things, try to pull the operation
3071 // through the select.
3072 if (LHS.getOpcode() == RHS.getOpcode() && LHS.hasOneUse() && RHS.hasOneUse()){
3074 std::cerr << "SELECT: ["; LHS.Val->dump();
3075 std::cerr << "] ["; RHS.Val->dump();
3079 // If this is a load and the token chain is identical, replace the select
3080 // of two loads with a load through a select of the address to load from.
3081 // This triggers in things like "select bool X, 10.0, 123.0" after the FP
3082 // constants have been dropped into the constant pool.
3083 if ((LHS.getOpcode() == ISD::LOAD ||
3084 LHS.getOpcode() == ISD::EXTLOAD ||
3085 LHS.getOpcode() == ISD::ZEXTLOAD ||
3086 LHS.getOpcode() == ISD::SEXTLOAD) &&
3087 // Token chains must be identical.
3088 LHS.getOperand(0) == RHS.getOperand(0) &&
3089 // If this is an EXTLOAD, the VT's must match.
3090 (LHS.getOpcode() == ISD::LOAD ||
3091 LHS.getOperand(3) == RHS.getOperand(3))) {
3092 // FIXME: this conflates two src values, discarding one. This is not
3093 // the right thing to do, but nothing uses srcvalues now. When they do,
3094 // turn SrcValue into a list of locations.
3096 if (TheSelect->getOpcode() == ISD::SELECT)
3097 Addr = DAG.getNode(ISD::SELECT, LHS.getOperand(1).getValueType(),
3098 TheSelect->getOperand(0), LHS.getOperand(1),
3101 Addr = DAG.getNode(ISD::SELECT_CC, LHS.getOperand(1).getValueType(),
3102 TheSelect->getOperand(0),
3103 TheSelect->getOperand(1),
3104 LHS.getOperand(1), RHS.getOperand(1),
3105 TheSelect->getOperand(4));
3108 if (LHS.getOpcode() == ISD::LOAD)
3109 Load = DAG.getLoad(TheSelect->getValueType(0), LHS.getOperand(0),
3110 Addr, LHS.getOperand(2));
3112 Load = DAG.getExtLoad(LHS.getOpcode(), TheSelect->getValueType(0),
3113 LHS.getOperand(0), Addr, LHS.getOperand(2),
3114 cast<VTSDNode>(LHS.getOperand(3))->getVT());
3115 // Users of the select now use the result of the load.
3116 CombineTo(TheSelect, Load);
3118 // Users of the old loads now use the new load's chain. We know the
3119 // old-load value is dead now.
3120 CombineTo(LHS.Val, Load.getValue(0), Load.getValue(1));
3121 CombineTo(RHS.Val, Load.getValue(0), Load.getValue(1));
3129 SDOperand DAGCombiner::SimplifySelectCC(SDOperand N0, SDOperand N1,
3130 SDOperand N2, SDOperand N3,
3133 MVT::ValueType VT = N2.getValueType();
3134 //ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val);
3135 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val);
3136 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.Val);
3137 ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N3.Val);
3139 // Determine if the condition we're dealing with is constant
3140 SDOperand SCC = SimplifySetCC(TLI.getSetCCResultTy(), N0, N1, CC, false);
3141 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.Val);
3143 // fold select_cc true, x, y -> x
3144 if (SCCC && SCCC->getValue())
3146 // fold select_cc false, x, y -> y
3147 if (SCCC && SCCC->getValue() == 0)
3150 // Check to see if we can simplify the select into an fabs node
3151 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) {
3152 // Allow either -0.0 or 0.0
3153 if (CFP->getValue() == 0.0) {
3154 // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs
3155 if ((CC == ISD::SETGE || CC == ISD::SETGT) &&
3156 N0 == N2 && N3.getOpcode() == ISD::FNEG &&
3157 N2 == N3.getOperand(0))
3158 return DAG.getNode(ISD::FABS, VT, N0);
3160 // select (setl[te] X, +/-0.0), fneg(X), X -> fabs
3161 if ((CC == ISD::SETLT || CC == ISD::SETLE) &&
3162 N0 == N3 && N2.getOpcode() == ISD::FNEG &&
3163 N2.getOperand(0) == N3)
3164 return DAG.getNode(ISD::FABS, VT, N3);
3168 // Check to see if we can perform the "gzip trick", transforming
3169 // select_cc setlt X, 0, A, 0 -> and (sra X, size(X)-1), A
3170 if (N1C && N1C->isNullValue() && N3C && N3C->isNullValue() &&
3171 MVT::isInteger(N0.getValueType()) &&
3172 MVT::isInteger(N2.getValueType()) && CC == ISD::SETLT) {
3173 MVT::ValueType XType = N0.getValueType();
3174 MVT::ValueType AType = N2.getValueType();
3175 if (XType >= AType) {
3176 // and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a
3177 // single-bit constant.
3178 if (N2C && ((N2C->getValue() & (N2C->getValue()-1)) == 0)) {
3179 unsigned ShCtV = Log2_64(N2C->getValue());
3180 ShCtV = MVT::getSizeInBits(XType)-ShCtV-1;
3181 SDOperand ShCt = DAG.getConstant(ShCtV, TLI.getShiftAmountTy());
3182 SDOperand Shift = DAG.getNode(ISD::SRL, XType, N0, ShCt);
3183 AddToWorkList(Shift.Val);
3184 if (XType > AType) {
3185 Shift = DAG.getNode(ISD::TRUNCATE, AType, Shift);
3186 AddToWorkList(Shift.Val);
3188 return DAG.getNode(ISD::AND, AType, Shift, N2);
3190 SDOperand Shift = DAG.getNode(ISD::SRA, XType, N0,
3191 DAG.getConstant(MVT::getSizeInBits(XType)-1,
3192 TLI.getShiftAmountTy()));
3193 AddToWorkList(Shift.Val);
3194 if (XType > AType) {
3195 Shift = DAG.getNode(ISD::TRUNCATE, AType, Shift);
3196 AddToWorkList(Shift.Val);
3198 return DAG.getNode(ISD::AND, AType, Shift, N2);
3202 // fold select C, 16, 0 -> shl C, 4
3203 if (N2C && N3C && N3C->isNullValue() && isPowerOf2_64(N2C->getValue()) &&
3204 TLI.getSetCCResultContents() == TargetLowering::ZeroOrOneSetCCResult) {
3205 // Get a SetCC of the condition
3206 // FIXME: Should probably make sure that setcc is legal if we ever have a
3207 // target where it isn't.
3208 SDOperand Temp, SCC;
3209 // cast from setcc result type to select result type
3210 if (AfterLegalize) {
3211 SCC = DAG.getSetCC(TLI.getSetCCResultTy(), N0, N1, CC);
3212 Temp = DAG.getZeroExtendInReg(SCC, N2.getValueType());
3214 SCC = DAG.getSetCC(MVT::i1, N0, N1, CC);
3215 Temp = DAG.getNode(ISD::ZERO_EXTEND, N2.getValueType(), SCC);
3217 AddToWorkList(SCC.Val);
3218 AddToWorkList(Temp.Val);
3219 // shl setcc result by log2 n2c
3220 return DAG.getNode(ISD::SHL, N2.getValueType(), Temp,
3221 DAG.getConstant(Log2_64(N2C->getValue()),
3222 TLI.getShiftAmountTy()));
3225 // Check to see if this is the equivalent of setcc
3226 // FIXME: Turn all of these into setcc if setcc if setcc is legal
3227 // otherwise, go ahead with the folds.
3228 if (0 && N3C && N3C->isNullValue() && N2C && (N2C->getValue() == 1ULL)) {
3229 MVT::ValueType XType = N0.getValueType();
3230 if (TLI.isOperationLegal(ISD::SETCC, TLI.getSetCCResultTy())) {
3231 SDOperand Res = DAG.getSetCC(TLI.getSetCCResultTy(), N0, N1, CC);
3232 if (Res.getValueType() != VT)
3233 Res = DAG.getNode(ISD::ZERO_EXTEND, VT, Res);
3237 // seteq X, 0 -> srl (ctlz X, log2(size(X)))
3238 if (N1C && N1C->isNullValue() && CC == ISD::SETEQ &&
3239 TLI.isOperationLegal(ISD::CTLZ, XType)) {
3240 SDOperand Ctlz = DAG.getNode(ISD::CTLZ, XType, N0);
3241 return DAG.getNode(ISD::SRL, XType, Ctlz,
3242 DAG.getConstant(Log2_32(MVT::getSizeInBits(XType)),
3243 TLI.getShiftAmountTy()));
3245 // setgt X, 0 -> srl (and (-X, ~X), size(X)-1)
3246 if (N1C && N1C->isNullValue() && CC == ISD::SETGT) {
3247 SDOperand NegN0 = DAG.getNode(ISD::SUB, XType, DAG.getConstant(0, XType),
3249 SDOperand NotN0 = DAG.getNode(ISD::XOR, XType, N0,
3250 DAG.getConstant(~0ULL, XType));
3251 return DAG.getNode(ISD::SRL, XType,
3252 DAG.getNode(ISD::AND, XType, NegN0, NotN0),
3253 DAG.getConstant(MVT::getSizeInBits(XType)-1,
3254 TLI.getShiftAmountTy()));
3256 // setgt X, -1 -> xor (srl (X, size(X)-1), 1)
3257 if (N1C && N1C->isAllOnesValue() && CC == ISD::SETGT) {
3258 SDOperand Sign = DAG.getNode(ISD::SRL, XType, N0,
3259 DAG.getConstant(MVT::getSizeInBits(XType)-1,
3260 TLI.getShiftAmountTy()));
3261 return DAG.getNode(ISD::XOR, XType, Sign, DAG.getConstant(1, XType));
3265 // Check to see if this is an integer abs. select_cc setl[te] X, 0, -X, X ->
3266 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
3267 if (N1C && N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE) &&
3268 N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1)) {
3269 if (ConstantSDNode *SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0))) {
3270 MVT::ValueType XType = N0.getValueType();
3271 if (SubC->isNullValue() && MVT::isInteger(XType)) {
3272 SDOperand Shift = DAG.getNode(ISD::SRA, XType, N0,
3273 DAG.getConstant(MVT::getSizeInBits(XType)-1,
3274 TLI.getShiftAmountTy()));
3275 SDOperand Add = DAG.getNode(ISD::ADD, XType, N0, Shift);
3276 AddToWorkList(Shift.Val);
3277 AddToWorkList(Add.Val);
3278 return DAG.getNode(ISD::XOR, XType, Add, Shift);
3286 SDOperand DAGCombiner::SimplifySetCC(MVT::ValueType VT, SDOperand N0,
3287 SDOperand N1, ISD::CondCode Cond,
3288 bool foldBooleans) {
3289 // These setcc operations always fold.
3293 case ISD::SETFALSE2: return DAG.getConstant(0, VT);
3295 case ISD::SETTRUE2: return DAG.getConstant(1, VT);
3298 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val)) {
3299 uint64_t C1 = N1C->getValue();
3300 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val)) {
3301 uint64_t C0 = N0C->getValue();
3303 // Sign extend the operands if required
3304 if (ISD::isSignedIntSetCC(Cond)) {
3305 C0 = N0C->getSignExtended();
3306 C1 = N1C->getSignExtended();
3310 default: assert(0 && "Unknown integer setcc!");
3311 case ISD::SETEQ: return DAG.getConstant(C0 == C1, VT);
3312 case ISD::SETNE: return DAG.getConstant(C0 != C1, VT);
3313 case ISD::SETULT: return DAG.getConstant(C0 < C1, VT);
3314 case ISD::SETUGT: return DAG.getConstant(C0 > C1, VT);
3315 case ISD::SETULE: return DAG.getConstant(C0 <= C1, VT);
3316 case ISD::SETUGE: return DAG.getConstant(C0 >= C1, VT);
3317 case ISD::SETLT: return DAG.getConstant((int64_t)C0 < (int64_t)C1, VT);
3318 case ISD::SETGT: return DAG.getConstant((int64_t)C0 > (int64_t)C1, VT);
3319 case ISD::SETLE: return DAG.getConstant((int64_t)C0 <= (int64_t)C1, VT);
3320 case ISD::SETGE: return DAG.getConstant((int64_t)C0 >= (int64_t)C1, VT);
3323 // If the LHS is a ZERO_EXTEND, perform the comparison on the input.
3324 if (N0.getOpcode() == ISD::ZERO_EXTEND) {
3325 unsigned InSize = MVT::getSizeInBits(N0.getOperand(0).getValueType());
3327 // If the comparison constant has bits in the upper part, the
3328 // zero-extended value could never match.
3329 if (C1 & (~0ULL << InSize)) {
3330 unsigned VSize = MVT::getSizeInBits(N0.getValueType());
3334 case ISD::SETEQ: return DAG.getConstant(0, VT);
3337 case ISD::SETNE: return DAG.getConstant(1, VT);
3340 // True if the sign bit of C1 is set.
3341 return DAG.getConstant((C1 & (1ULL << VSize)) != 0, VT);
3344 // True if the sign bit of C1 isn't set.
3345 return DAG.getConstant((C1 & (1ULL << VSize)) == 0, VT);
3351 // Otherwise, we can perform the comparison with the low bits.
3359 return DAG.getSetCC(VT, N0.getOperand(0),
3360 DAG.getConstant(C1, N0.getOperand(0).getValueType()),
3363 break; // todo, be more careful with signed comparisons
3365 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
3366 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
3367 MVT::ValueType ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT();
3368 unsigned ExtSrcTyBits = MVT::getSizeInBits(ExtSrcTy);
3369 MVT::ValueType ExtDstTy = N0.getValueType();
3370 unsigned ExtDstTyBits = MVT::getSizeInBits(ExtDstTy);
3372 // If the extended part has any inconsistent bits, it cannot ever
3373 // compare equal. In other words, they have to be all ones or all
3376 (~0ULL >> (64-ExtSrcTyBits)) & (~0ULL << (ExtDstTyBits-1));
3377 if ((C1 & ExtBits) != 0 && (C1 & ExtBits) != ExtBits)
3378 return DAG.getConstant(Cond == ISD::SETNE, VT);
3381 MVT::ValueType Op0Ty = N0.getOperand(0).getValueType();
3382 if (Op0Ty == ExtSrcTy) {
3383 ZextOp = N0.getOperand(0);
3385 int64_t Imm = ~0ULL >> (64-ExtSrcTyBits);
3386 ZextOp = DAG.getNode(ISD::AND, Op0Ty, N0.getOperand(0),
3387 DAG.getConstant(Imm, Op0Ty));
3389 AddToWorkList(ZextOp.Val);
3390 // Otherwise, make this a use of a zext.
3391 return DAG.getSetCC(VT, ZextOp,
3392 DAG.getConstant(C1 & (~0ULL>>(64-ExtSrcTyBits)),
3395 } else if ((N1C->getValue() == 0 || N1C->getValue() == 1) &&
3396 (Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
3397 (N0.getOpcode() == ISD::XOR ||
3398 (N0.getOpcode() == ISD::AND &&
3399 N0.getOperand(0).getOpcode() == ISD::XOR &&
3400 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) &&
3401 isa<ConstantSDNode>(N0.getOperand(1)) &&
3402 cast<ConstantSDNode>(N0.getOperand(1))->getValue() == 1) {
3403 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We can
3404 // only do this if the top bits are known zero.
3405 if (TLI.MaskedValueIsZero(N1,
3406 MVT::getIntVTBitMask(N0.getValueType())-1)) {
3407 // Okay, get the un-inverted input value.
3409 if (N0.getOpcode() == ISD::XOR)
3410 Val = N0.getOperand(0);
3412 assert(N0.getOpcode() == ISD::AND &&
3413 N0.getOperand(0).getOpcode() == ISD::XOR);
3414 // ((X^1)&1)^1 -> X & 1
3415 Val = DAG.getNode(ISD::AND, N0.getValueType(),
3416 N0.getOperand(0).getOperand(0), N0.getOperand(1));
3418 return DAG.getSetCC(VT, Val, N1,
3419 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
3423 uint64_t MinVal, MaxVal;
3424 unsigned OperandBitSize = MVT::getSizeInBits(N1C->getValueType(0));
3425 if (ISD::isSignedIntSetCC(Cond)) {
3426 MinVal = 1ULL << (OperandBitSize-1);
3427 if (OperandBitSize != 1) // Avoid X >> 64, which is undefined.
3428 MaxVal = ~0ULL >> (65-OperandBitSize);
3433 MaxVal = ~0ULL >> (64-OperandBitSize);
3436 // Canonicalize GE/LE comparisons to use GT/LT comparisons.
3437 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) {
3438 if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true
3439 --C1; // X >= C0 --> X > (C0-1)
3440 return DAG.getSetCC(VT, N0, DAG.getConstant(C1, N1.getValueType()),
3441 (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT);
3444 if (Cond == ISD::SETLE || Cond == ISD::SETULE) {
3445 if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true
3446 ++C1; // X <= C0 --> X < (C0+1)
3447 return DAG.getSetCC(VT, N0, DAG.getConstant(C1, N1.getValueType()),
3448 (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT);
3451 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal)
3452 return DAG.getConstant(0, VT); // X < MIN --> false
3454 // Canonicalize setgt X, Min --> setne X, Min
3455 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal)
3456 return DAG.getSetCC(VT, N0, N1, ISD::SETNE);
3457 // Canonicalize setlt X, Max --> setne X, Max
3458 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal)
3459 return DAG.getSetCC(VT, N0, N1, ISD::SETNE);
3461 // If we have setult X, 1, turn it into seteq X, 0
3462 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1)
3463 return DAG.getSetCC(VT, N0, DAG.getConstant(MinVal, N0.getValueType()),
3465 // If we have setugt X, Max-1, turn it into seteq X, Max
3466 else if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1)
3467 return DAG.getSetCC(VT, N0, DAG.getConstant(MaxVal, N0.getValueType()),
3470 // If we have "setcc X, C0", check to see if we can shrink the immediate
3473 // SETUGT X, SINTMAX -> SETLT X, 0
3474 if (Cond == ISD::SETUGT && OperandBitSize != 1 &&
3475 C1 == (~0ULL >> (65-OperandBitSize)))
3476 return DAG.getSetCC(VT, N0, DAG.getConstant(0, N1.getValueType()),
3479 // FIXME: Implement the rest of these.
3481 // Fold bit comparisons when we can.
3482 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
3483 VT == N0.getValueType() && N0.getOpcode() == ISD::AND)
3484 if (ConstantSDNode *AndRHS =
3485 dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
3486 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3
3487 // Perform the xform if the AND RHS is a single bit.
3488 if ((AndRHS->getValue() & (AndRHS->getValue()-1)) == 0) {
3489 return DAG.getNode(ISD::SRL, VT, N0,
3490 DAG.getConstant(Log2_64(AndRHS->getValue()),
3491 TLI.getShiftAmountTy()));
3493 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getValue()) {
3494 // (X & 8) == 8 --> (X & 8) >> 3
3495 // Perform the xform if C1 is a single bit.
3496 if ((C1 & (C1-1)) == 0) {
3497 return DAG.getNode(ISD::SRL, VT, N0,
3498 DAG.getConstant(Log2_64(C1),TLI.getShiftAmountTy()));
3503 } else if (isa<ConstantSDNode>(N0.Val)) {
3504 // Ensure that the constant occurs on the RHS.
3505 return DAG.getSetCC(VT, N1, N0, ISD::getSetCCSwappedOperands(Cond));
3508 if (ConstantFPSDNode *N0C = dyn_cast<ConstantFPSDNode>(N0.Val))
3509 if (ConstantFPSDNode *N1C = dyn_cast<ConstantFPSDNode>(N1.Val)) {
3510 double C0 = N0C->getValue(), C1 = N1C->getValue();
3513 default: break; // FIXME: Implement the rest of these!
3514 case ISD::SETEQ: return DAG.getConstant(C0 == C1, VT);
3515 case ISD::SETNE: return DAG.getConstant(C0 != C1, VT);
3516 case ISD::SETLT: return DAG.getConstant(C0 < C1, VT);
3517 case ISD::SETGT: return DAG.getConstant(C0 > C1, VT);
3518 case ISD::SETLE: return DAG.getConstant(C0 <= C1, VT);
3519 case ISD::SETGE: return DAG.getConstant(C0 >= C1, VT);
3522 // Ensure that the constant occurs on the RHS.
3523 return DAG.getSetCC(VT, N1, N0, ISD::getSetCCSwappedOperands(Cond));
3527 // We can always fold X == Y for integer setcc's.
3528 if (MVT::isInteger(N0.getValueType()))
3529 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
3530 unsigned UOF = ISD::getUnorderedFlavor(Cond);
3531 if (UOF == 2) // FP operators that are undefined on NaNs.
3532 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
3533 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond)))
3534 return DAG.getConstant(UOF, VT);
3535 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO
3536 // if it is not already.
3537 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO;
3538 if (NewCond != Cond)
3539 return DAG.getSetCC(VT, N0, N1, NewCond);
3542 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
3543 MVT::isInteger(N0.getValueType())) {
3544 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB ||
3545 N0.getOpcode() == ISD::XOR) {
3546 // Simplify (X+Y) == (X+Z) --> Y == Z
3547 if (N0.getOpcode() == N1.getOpcode()) {
3548 if (N0.getOperand(0) == N1.getOperand(0))
3549 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(1), Cond);
3550 if (N0.getOperand(1) == N1.getOperand(1))
3551 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(0), Cond);
3552 if (DAG.isCommutativeBinOp(N0.getOpcode())) {
3553 // If X op Y == Y op X, try other combinations.
3554 if (N0.getOperand(0) == N1.getOperand(1))
3555 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(0), Cond);
3556 if (N0.getOperand(1) == N1.getOperand(0))
3557 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(1), Cond);
3561 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) {
3562 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
3563 // Turn (X+C1) == C2 --> X == C2-C1
3564 if (N0.getOpcode() == ISD::ADD && N0.Val->hasOneUse()) {
3565 return DAG.getSetCC(VT, N0.getOperand(0),
3566 DAG.getConstant(RHSC->getValue()-LHSR->getValue(),
3567 N0.getValueType()), Cond);
3570 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0.
3571 if (N0.getOpcode() == ISD::XOR)
3572 // If we know that all of the inverted bits are zero, don't bother
3573 // performing the inversion.
3574 if (TLI.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getValue()))
3575 return DAG.getSetCC(VT, N0.getOperand(0),
3576 DAG.getConstant(LHSR->getValue()^RHSC->getValue(),
3577 N0.getValueType()), Cond);
3580 // Turn (C1-X) == C2 --> X == C1-C2
3581 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) {
3582 if (N0.getOpcode() == ISD::SUB && N0.Val->hasOneUse()) {
3583 return DAG.getSetCC(VT, N0.getOperand(1),
3584 DAG.getConstant(SUBC->getValue()-RHSC->getValue(),
3585 N0.getValueType()), Cond);
3590 // Simplify (X+Z) == X --> Z == 0
3591 if (N0.getOperand(0) == N1)
3592 return DAG.getSetCC(VT, N0.getOperand(1),
3593 DAG.getConstant(0, N0.getValueType()), Cond);
3594 if (N0.getOperand(1) == N1) {
3595 if (DAG.isCommutativeBinOp(N0.getOpcode()))
3596 return DAG.getSetCC(VT, N0.getOperand(0),
3597 DAG.getConstant(0, N0.getValueType()), Cond);
3599 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!");
3600 // (Z-X) == X --> Z == X<<1
3601 SDOperand SH = DAG.getNode(ISD::SHL, N1.getValueType(),
3603 DAG.getConstant(1,TLI.getShiftAmountTy()));
3604 AddToWorkList(SH.Val);
3605 return DAG.getSetCC(VT, N0.getOperand(0), SH, Cond);
3610 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB ||
3611 N1.getOpcode() == ISD::XOR) {
3612 // Simplify X == (X+Z) --> Z == 0
3613 if (N1.getOperand(0) == N0) {
3614 return DAG.getSetCC(VT, N1.getOperand(1),
3615 DAG.getConstant(0, N1.getValueType()), Cond);
3616 } else if (N1.getOperand(1) == N0) {
3617 if (DAG.isCommutativeBinOp(N1.getOpcode())) {
3618 return DAG.getSetCC(VT, N1.getOperand(0),
3619 DAG.getConstant(0, N1.getValueType()), Cond);
3621 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!");
3622 // X == (Z-X) --> X<<1 == Z
3623 SDOperand SH = DAG.getNode(ISD::SHL, N1.getValueType(), N0,
3624 DAG.getConstant(1,TLI.getShiftAmountTy()));
3625 AddToWorkList(SH.Val);
3626 return DAG.getSetCC(VT, SH, N1.getOperand(0), Cond);
3632 // Fold away ALL boolean setcc's.
3634 if (N0.getValueType() == MVT::i1 && foldBooleans) {
3636 default: assert(0 && "Unknown integer setcc!");
3637 case ISD::SETEQ: // X == Y -> (X^Y)^1
3638 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, N1);
3639 N0 = DAG.getNode(ISD::XOR, MVT::i1, Temp, DAG.getConstant(1, MVT::i1));
3640 AddToWorkList(Temp.Val);
3642 case ISD::SETNE: // X != Y --> (X^Y)
3643 N0 = DAG.getNode(ISD::XOR, MVT::i1, N0, N1);
3645 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> X^1 & Y
3646 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> X^1 & Y
3647 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1));
3648 N0 = DAG.getNode(ISD::AND, MVT::i1, N1, Temp);
3649 AddToWorkList(Temp.Val);
3651 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> Y^1 & X
3652 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> Y^1 & X
3653 Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1));
3654 N0 = DAG.getNode(ISD::AND, MVT::i1, N0, Temp);
3655 AddToWorkList(Temp.Val);
3657 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> X^1 | Y
3658 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> X^1 | Y
3659 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1));
3660 N0 = DAG.getNode(ISD::OR, MVT::i1, N1, Temp);
3661 AddToWorkList(Temp.Val);
3663 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> Y^1 | X
3664 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> Y^1 | X
3665 Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1));
3666 N0 = DAG.getNode(ISD::OR, MVT::i1, N0, Temp);
3669 if (VT != MVT::i1) {
3670 AddToWorkList(N0.Val);
3671 // FIXME: If running after legalize, we probably can't do this.
3672 N0 = DAG.getNode(ISD::ZERO_EXTEND, VT, N0);
3677 // Could not fold it.
3681 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
3682 /// return a DAG expression to select that will generate the same value by
3683 /// multiplying by a magic number. See:
3684 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
3685 SDOperand DAGCombiner::BuildSDIV(SDNode *N) {
3686 std::vector<SDNode*> Built;
3687 SDOperand S = TLI.BuildSDIV(N, DAG, &Built);
3689 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end();
3695 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
3696 /// return a DAG expression to select that will generate the same value by
3697 /// multiplying by a magic number. See:
3698 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
3699 SDOperand DAGCombiner::BuildUDIV(SDNode *N) {
3700 std::vector<SDNode*> Built;
3701 SDOperand S = TLI.BuildUDIV(N, DAG, &Built);
3703 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end();
3709 // SelectionDAG::Combine - This is the entry point for the file.
3711 void SelectionDAG::Combine(bool RunningAfterLegalize) {
3712 /// run - This is the main entry point to this class.
3714 DAGCombiner(*this).Run(RunningAfterLegalize);