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"
44 Statistic<> NodesCombined ("dagcombiner", "Number of dag nodes combined");
51 // Worklist of all of the nodes that need to be simplified.
52 std::vector<SDNode*> WorkList;
54 /// AddUsersToWorkList - When an instruction is simplified, add all users of
55 /// the instruction to the work lists because they might get more simplified
58 void AddUsersToWorkList(SDNode *N) {
59 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
61 WorkList.push_back(*UI);
64 /// removeFromWorkList - remove all instances of N from the worklist.
66 void removeFromWorkList(SDNode *N) {
67 WorkList.erase(std::remove(WorkList.begin(), WorkList.end(), N),
72 void AddToWorkList(SDNode *N) {
73 WorkList.push_back(N);
76 SDOperand CombineTo(SDNode *N, const std::vector<SDOperand> &To) {
78 DEBUG(std::cerr << "\nReplacing "; N->dump();
79 std::cerr << "\nWith: "; To[0].Val->dump();
80 std::cerr << " and " << To.size()-1 << " other values\n");
81 std::vector<SDNode*> NowDead;
82 DAG.ReplaceAllUsesWith(N, To, &NowDead);
84 // Push the new nodes and any users onto the worklist
85 for (unsigned i = 0, e = To.size(); i != e; ++i) {
86 WorkList.push_back(To[i].Val);
87 AddUsersToWorkList(To[i].Val);
90 // Nodes can end up on the worklist more than once. Make sure we do
91 // not process a node that has been replaced.
92 removeFromWorkList(N);
93 for (unsigned i = 0, e = NowDead.size(); i != e; ++i)
94 removeFromWorkList(NowDead[i]);
96 // Finally, since the node is now dead, remove it from the graph.
98 return SDOperand(N, 0);
101 SDOperand CombineTo(SDNode *N, SDOperand Res) {
102 std::vector<SDOperand> To;
104 return CombineTo(N, To);
107 SDOperand CombineTo(SDNode *N, SDOperand Res0, SDOperand Res1) {
108 std::vector<SDOperand> To;
111 return CombineTo(N, To);
115 /// SimplifyDemandedBits - Check the specified integer node value to see if
116 /// it can be simplified or if things it uses can be simplified by bit
117 /// propagation. If so, return true.
118 bool SimplifyDemandedBits(SDOperand Op) {
119 TargetLowering::TargetLoweringOpt TLO(DAG);
120 uint64_t KnownZero, KnownOne;
121 uint64_t Demanded = MVT::getIntVTBitMask(Op.getValueType());
122 if (!TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
126 WorkList.push_back(Op.Val);
128 // Replace the old value with the new one.
130 DEBUG(std::cerr << "\nReplacing "; TLO.Old.Val->dump();
131 std::cerr << "\nWith: "; TLO.New.Val->dump());
133 std::vector<SDNode*> NowDead;
134 DAG.ReplaceAllUsesOfValueWith(TLO.Old, TLO.New, NowDead);
136 // Push the new node and any (possibly new) users onto the worklist.
137 WorkList.push_back(TLO.New.Val);
138 AddUsersToWorkList(TLO.New.Val);
140 // Nodes can end up on the worklist more than once. Make sure we do
141 // not process a node that has been replaced.
142 for (unsigned i = 0, e = NowDead.size(); i != e; ++i)
143 removeFromWorkList(NowDead[i]);
145 // Finally, if the node is now dead, remove it from the graph. The node
146 // may not be dead if the replacement process recursively simplified to
147 // something else needing this node.
148 if (TLO.Old.Val->use_empty()) {
149 removeFromWorkList(TLO.Old.Val);
150 DAG.DeleteNode(TLO.Old.Val);
155 /// visit - call the node-specific routine that knows how to fold each
156 /// particular type of node.
157 SDOperand visit(SDNode *N);
159 // Visitation implementation - Implement dag node combining for different
160 // node types. The semantics are as follows:
162 // SDOperand.Val == 0 - No change was made
163 // SDOperand.Val == N - N was replaced, is dead, and is already handled.
164 // otherwise - N should be replaced by the returned Operand.
166 SDOperand visitTokenFactor(SDNode *N);
167 SDOperand visitADD(SDNode *N);
168 SDOperand visitSUB(SDNode *N);
169 SDOperand visitMUL(SDNode *N);
170 SDOperand visitSDIV(SDNode *N);
171 SDOperand visitUDIV(SDNode *N);
172 SDOperand visitSREM(SDNode *N);
173 SDOperand visitUREM(SDNode *N);
174 SDOperand visitMULHU(SDNode *N);
175 SDOperand visitMULHS(SDNode *N);
176 SDOperand visitAND(SDNode *N);
177 SDOperand visitOR(SDNode *N);
178 SDOperand visitXOR(SDNode *N);
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 visitSIGN_EXTEND_INREG(SDNode *N);
191 SDOperand visitTRUNCATE(SDNode *N);
192 SDOperand visitBIT_CONVERT(SDNode *N);
193 SDOperand visitFADD(SDNode *N);
194 SDOperand visitFSUB(SDNode *N);
195 SDOperand visitFMUL(SDNode *N);
196 SDOperand visitFDIV(SDNode *N);
197 SDOperand visitFREM(SDNode *N);
198 SDOperand visitFCOPYSIGN(SDNode *N);
199 SDOperand visitSINT_TO_FP(SDNode *N);
200 SDOperand visitUINT_TO_FP(SDNode *N);
201 SDOperand visitFP_TO_SINT(SDNode *N);
202 SDOperand visitFP_TO_UINT(SDNode *N);
203 SDOperand visitFP_ROUND(SDNode *N);
204 SDOperand visitFP_ROUND_INREG(SDNode *N);
205 SDOperand visitFP_EXTEND(SDNode *N);
206 SDOperand visitFNEG(SDNode *N);
207 SDOperand visitFABS(SDNode *N);
208 SDOperand visitBRCOND(SDNode *N);
209 SDOperand visitBR_CC(SDNode *N);
210 SDOperand visitLOAD(SDNode *N);
211 SDOperand visitXEXTLOAD(SDNode *N);
212 SDOperand visitSTORE(SDNode *N);
213 SDOperand visitINSERT_VECTOR_ELT(SDNode *N);
214 SDOperand visitVINSERT_VECTOR_ELT(SDNode *N);
215 SDOperand visitVBUILD_VECTOR(SDNode *N);
216 SDOperand visitVECTOR_SHUFFLE(SDNode *N);
217 SDOperand visitVVECTOR_SHUFFLE(SDNode *N);
219 SDOperand ReassociateOps(unsigned Opc, SDOperand LHS, SDOperand RHS);
221 bool SimplifySelectOps(SDNode *SELECT, SDOperand LHS, SDOperand RHS);
222 SDOperand SimplifySelect(SDOperand N0, SDOperand N1, SDOperand N2);
223 SDOperand SimplifySelectCC(SDOperand N0, SDOperand N1, SDOperand N2,
224 SDOperand N3, ISD::CondCode CC);
225 SDOperand SimplifySetCC(MVT::ValueType VT, SDOperand N0, SDOperand N1,
226 ISD::CondCode Cond, bool foldBooleans = true);
228 SDOperand BuildSDIV(SDNode *N);
229 SDOperand BuildUDIV(SDNode *N);
231 DAGCombiner(SelectionDAG &D)
232 : DAG(D), TLI(D.getTargetLoweringInfo()), AfterLegalize(false) {}
234 /// Run - runs the dag combiner on all nodes in the work list
235 void Run(bool RunningAfterLegalize);
239 //===----------------------------------------------------------------------===//
240 // TargetLowering::DAGCombinerInfo implementation
241 //===----------------------------------------------------------------------===//
243 void TargetLowering::DAGCombinerInfo::AddToWorklist(SDNode *N) {
244 ((DAGCombiner*)DC)->AddToWorkList(N);
247 SDOperand TargetLowering::DAGCombinerInfo::
248 CombineTo(SDNode *N, const std::vector<SDOperand> &To) {
249 return ((DAGCombiner*)DC)->CombineTo(N, To);
252 SDOperand TargetLowering::DAGCombinerInfo::
253 CombineTo(SDNode *N, SDOperand Res) {
254 return ((DAGCombiner*)DC)->CombineTo(N, Res);
258 SDOperand TargetLowering::DAGCombinerInfo::
259 CombineTo(SDNode *N, SDOperand Res0, SDOperand Res1) {
260 return ((DAGCombiner*)DC)->CombineTo(N, Res0, Res1);
266 //===----------------------------------------------------------------------===//
270 int64_t m; // magic number
271 int64_t s; // shift amount
275 uint64_t m; // magic number
276 int64_t a; // add indicator
277 int64_t s; // shift amount
280 /// magic - calculate the magic numbers required to codegen an integer sdiv as
281 /// a sequence of multiply and shifts. Requires that the divisor not be 0, 1,
283 static ms magic32(int32_t d) {
285 uint32_t ad, anc, delta, q1, r1, q2, r2, t;
286 const uint32_t two31 = 0x80000000U;
290 t = two31 + ((uint32_t)d >> 31);
291 anc = t - 1 - t%ad; // absolute value of nc
292 p = 31; // initialize p
293 q1 = two31/anc; // initialize q1 = 2p/abs(nc)
294 r1 = two31 - q1*anc; // initialize r1 = rem(2p,abs(nc))
295 q2 = two31/ad; // initialize q2 = 2p/abs(d)
296 r2 = two31 - q2*ad; // initialize r2 = rem(2p,abs(d))
299 q1 = 2*q1; // update q1 = 2p/abs(nc)
300 r1 = 2*r1; // update r1 = rem(2p/abs(nc))
301 if (r1 >= anc) { // must be unsigned comparison
305 q2 = 2*q2; // update q2 = 2p/abs(d)
306 r2 = 2*r2; // update r2 = rem(2p/abs(d))
307 if (r2 >= ad) { // must be unsigned comparison
312 } while (q1 < delta || (q1 == delta && r1 == 0));
314 mag.m = (int32_t)(q2 + 1); // make sure to sign extend
315 if (d < 0) mag.m = -mag.m; // resulting magic number
316 mag.s = p - 32; // resulting shift
320 /// magicu - calculate the magic numbers required to codegen an integer udiv as
321 /// a sequence of multiply, add and shifts. Requires that the divisor not be 0.
322 static mu magicu32(uint32_t d) {
324 uint32_t nc, delta, q1, r1, q2, r2;
326 magu.a = 0; // initialize "add" indicator
328 p = 31; // initialize p
329 q1 = 0x80000000/nc; // initialize q1 = 2p/nc
330 r1 = 0x80000000 - q1*nc; // initialize r1 = rem(2p,nc)
331 q2 = 0x7FFFFFFF/d; // initialize q2 = (2p-1)/d
332 r2 = 0x7FFFFFFF - q2*d; // initialize r2 = rem((2p-1),d)
335 if (r1 >= nc - r1 ) {
336 q1 = 2*q1 + 1; // update q1
337 r1 = 2*r1 - nc; // update r1
340 q1 = 2*q1; // update q1
341 r1 = 2*r1; // update r1
343 if (r2 + 1 >= d - r2) {
344 if (q2 >= 0x7FFFFFFF) magu.a = 1;
345 q2 = 2*q2 + 1; // update q2
346 r2 = 2*r2 + 1 - d; // update r2
349 if (q2 >= 0x80000000) magu.a = 1;
350 q2 = 2*q2; // update q2
351 r2 = 2*r2 + 1; // update r2
354 } while (p < 64 && (q1 < delta || (q1 == delta && r1 == 0)));
355 magu.m = q2 + 1; // resulting magic number
356 magu.s = p - 32; // resulting shift
360 /// magic - calculate the magic numbers required to codegen an integer sdiv as
361 /// a sequence of multiply and shifts. Requires that the divisor not be 0, 1,
363 static ms magic64(int64_t d) {
365 uint64_t ad, anc, delta, q1, r1, q2, r2, t;
366 const uint64_t two63 = 9223372036854775808ULL; // 2^63
369 ad = d >= 0 ? d : -d;
370 t = two63 + ((uint64_t)d >> 63);
371 anc = t - 1 - t%ad; // absolute value of nc
372 p = 63; // initialize p
373 q1 = two63/anc; // initialize q1 = 2p/abs(nc)
374 r1 = two63 - q1*anc; // initialize r1 = rem(2p,abs(nc))
375 q2 = two63/ad; // initialize q2 = 2p/abs(d)
376 r2 = two63 - q2*ad; // initialize r2 = rem(2p,abs(d))
379 q1 = 2*q1; // update q1 = 2p/abs(nc)
380 r1 = 2*r1; // update r1 = rem(2p/abs(nc))
381 if (r1 >= anc) { // must be unsigned comparison
385 q2 = 2*q2; // update q2 = 2p/abs(d)
386 r2 = 2*r2; // update r2 = rem(2p/abs(d))
387 if (r2 >= ad) { // must be unsigned comparison
392 } while (q1 < delta || (q1 == delta && r1 == 0));
395 if (d < 0) mag.m = -mag.m; // resulting magic number
396 mag.s = p - 64; // resulting shift
400 /// magicu - calculate the magic numbers required to codegen an integer udiv as
401 /// a sequence of multiply, add and shifts. Requires that the divisor not be 0.
402 static mu magicu64(uint64_t d)
405 uint64_t nc, delta, q1, r1, q2, r2;
407 magu.a = 0; // initialize "add" indicator
409 p = 63; // initialize p
410 q1 = 0x8000000000000000ull/nc; // initialize q1 = 2p/nc
411 r1 = 0x8000000000000000ull - q1*nc; // initialize r1 = rem(2p,nc)
412 q2 = 0x7FFFFFFFFFFFFFFFull/d; // initialize q2 = (2p-1)/d
413 r2 = 0x7FFFFFFFFFFFFFFFull - q2*d; // initialize r2 = rem((2p-1),d)
416 if (r1 >= nc - r1 ) {
417 q1 = 2*q1 + 1; // update q1
418 r1 = 2*r1 - nc; // update r1
421 q1 = 2*q1; // update q1
422 r1 = 2*r1; // update r1
424 if (r2 + 1 >= d - r2) {
425 if (q2 >= 0x7FFFFFFFFFFFFFFFull) magu.a = 1;
426 q2 = 2*q2 + 1; // update q2
427 r2 = 2*r2 + 1 - d; // update r2
430 if (q2 >= 0x8000000000000000ull) magu.a = 1;
431 q2 = 2*q2; // update q2
432 r2 = 2*r2 + 1; // update r2
435 } while (p < 64 && (q1 < delta || (q1 == delta && r1 == 0)));
436 magu.m = q2 + 1; // resulting magic number
437 magu.s = p - 64; // resulting shift
441 // isSetCCEquivalent - Return true if this node is a setcc, or is a select_cc
442 // that selects between the values 1 and 0, making it equivalent to a setcc.
443 // Also, set the incoming LHS, RHS, and CC references to the appropriate
444 // nodes based on the type of node we are checking. This simplifies life a
445 // bit for the callers.
446 static bool isSetCCEquivalent(SDOperand N, SDOperand &LHS, SDOperand &RHS,
448 if (N.getOpcode() == ISD::SETCC) {
449 LHS = N.getOperand(0);
450 RHS = N.getOperand(1);
451 CC = N.getOperand(2);
454 if (N.getOpcode() == ISD::SELECT_CC &&
455 N.getOperand(2).getOpcode() == ISD::Constant &&
456 N.getOperand(3).getOpcode() == ISD::Constant &&
457 cast<ConstantSDNode>(N.getOperand(2))->getValue() == 1 &&
458 cast<ConstantSDNode>(N.getOperand(3))->isNullValue()) {
459 LHS = N.getOperand(0);
460 RHS = N.getOperand(1);
461 CC = N.getOperand(4);
467 // isOneUseSetCC - Return true if this is a SetCC-equivalent operation with only
468 // one use. If this is true, it allows the users to invert the operation for
469 // free when it is profitable to do so.
470 static bool isOneUseSetCC(SDOperand N) {
471 SDOperand N0, N1, N2;
472 if (isSetCCEquivalent(N, N0, N1, N2) && N.Val->hasOneUse())
477 // FIXME: This should probably go in the ISD class rather than being duplicated
479 static bool isCommutativeBinOp(unsigned Opcode) {
485 case ISD::XOR: return true;
486 default: return false; // FIXME: Need commutative info for user ops!
490 SDOperand DAGCombiner::ReassociateOps(unsigned Opc, SDOperand N0, SDOperand N1){
491 MVT::ValueType VT = N0.getValueType();
492 // reassoc. (op (op x, c1), y) -> (op (op x, y), c1) iff x+c1 has one use
493 // reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2))
494 if (N0.getOpcode() == Opc && isa<ConstantSDNode>(N0.getOperand(1))) {
495 if (isa<ConstantSDNode>(N1)) {
496 SDOperand OpNode = DAG.getNode(Opc, VT, N0.getOperand(1), N1);
497 AddToWorkList(OpNode.Val);
498 return DAG.getNode(Opc, VT, OpNode, N0.getOperand(0));
499 } else if (N0.hasOneUse()) {
500 SDOperand OpNode = DAG.getNode(Opc, VT, N0.getOperand(0), N1);
501 AddToWorkList(OpNode.Val);
502 return DAG.getNode(Opc, VT, OpNode, N0.getOperand(1));
505 // reassoc. (op y, (op x, c1)) -> (op (op x, y), c1) iff x+c1 has one use
506 // reassoc. (op c2, (op x, c1)) -> (op x, (op c1, c2))
507 if (N1.getOpcode() == Opc && isa<ConstantSDNode>(N1.getOperand(1))) {
508 if (isa<ConstantSDNode>(N0)) {
509 SDOperand OpNode = DAG.getNode(Opc, VT, N1.getOperand(1), N0);
510 AddToWorkList(OpNode.Val);
511 return DAG.getNode(Opc, VT, OpNode, N1.getOperand(0));
512 } else if (N1.hasOneUse()) {
513 SDOperand OpNode = DAG.getNode(Opc, VT, N1.getOperand(0), N0);
514 AddToWorkList(OpNode.Val);
515 return DAG.getNode(Opc, VT, OpNode, N1.getOperand(1));
521 void DAGCombiner::Run(bool RunningAfterLegalize) {
522 // set the instance variable, so that the various visit routines may use it.
523 AfterLegalize = RunningAfterLegalize;
525 // Add all the dag nodes to the worklist.
526 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
527 E = DAG.allnodes_end(); I != E; ++I)
528 WorkList.push_back(I);
530 // Create a dummy node (which is not added to allnodes), that adds a reference
531 // to the root node, preventing it from being deleted, and tracking any
532 // changes of the root.
533 HandleSDNode Dummy(DAG.getRoot());
536 /// DagCombineInfo - Expose the DAG combiner to the target combiner impls.
537 TargetLowering::DAGCombinerInfo
538 DagCombineInfo(DAG, !RunningAfterLegalize, this);
540 // while the worklist isn't empty, inspect the node on the end of it and
541 // try and combine it.
542 while (!WorkList.empty()) {
543 SDNode *N = WorkList.back();
546 // If N has no uses, it is dead. Make sure to revisit all N's operands once
547 // N is deleted from the DAG, since they too may now be dead or may have a
548 // reduced number of uses, allowing other xforms.
549 if (N->use_empty() && N != &Dummy) {
550 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
551 WorkList.push_back(N->getOperand(i).Val);
553 removeFromWorkList(N);
558 SDOperand RV = visit(N);
560 // If nothing happened, try a target-specific DAG combine.
562 if (N->getOpcode() >= ISD::BUILTIN_OP_END ||
563 TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode()))
564 RV = TLI.PerformDAGCombine(N, DagCombineInfo);
569 // If we get back the same node we passed in, rather than a new node or
570 // zero, we know that the node must have defined multiple values and
571 // CombineTo was used. Since CombineTo takes care of the worklist
572 // mechanics for us, we have no work to do in this case.
574 DEBUG(std::cerr << "\nReplacing "; N->dump();
575 std::cerr << "\nWith: "; RV.Val->dump();
577 std::vector<SDNode*> NowDead;
578 DAG.ReplaceAllUsesWith(N, std::vector<SDOperand>(1, RV), &NowDead);
580 // Push the new node and any users onto the worklist
581 WorkList.push_back(RV.Val);
582 AddUsersToWorkList(RV.Val);
584 // Nodes can end up on the worklist more than once. Make sure we do
585 // not process a node that has been replaced.
586 removeFromWorkList(N);
587 for (unsigned i = 0, e = NowDead.size(); i != e; ++i)
588 removeFromWorkList(NowDead[i]);
590 // Finally, since the node is now dead, remove it from the graph.
596 // If the root changed (e.g. it was a dead load, update the root).
597 DAG.setRoot(Dummy.getValue());
600 SDOperand DAGCombiner::visit(SDNode *N) {
601 switch(N->getOpcode()) {
603 case ISD::TokenFactor: return visitTokenFactor(N);
604 case ISD::ADD: return visitADD(N);
605 case ISD::SUB: return visitSUB(N);
606 case ISD::MUL: return visitMUL(N);
607 case ISD::SDIV: return visitSDIV(N);
608 case ISD::UDIV: return visitUDIV(N);
609 case ISD::SREM: return visitSREM(N);
610 case ISD::UREM: return visitUREM(N);
611 case ISD::MULHU: return visitMULHU(N);
612 case ISD::MULHS: return visitMULHS(N);
613 case ISD::AND: return visitAND(N);
614 case ISD::OR: return visitOR(N);
615 case ISD::XOR: return visitXOR(N);
616 case ISD::SHL: return visitSHL(N);
617 case ISD::SRA: return visitSRA(N);
618 case ISD::SRL: return visitSRL(N);
619 case ISD::CTLZ: return visitCTLZ(N);
620 case ISD::CTTZ: return visitCTTZ(N);
621 case ISD::CTPOP: return visitCTPOP(N);
622 case ISD::SELECT: return visitSELECT(N);
623 case ISD::SELECT_CC: return visitSELECT_CC(N);
624 case ISD::SETCC: return visitSETCC(N);
625 case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N);
626 case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N);
627 case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N);
628 case ISD::TRUNCATE: return visitTRUNCATE(N);
629 case ISD::BIT_CONVERT: return visitBIT_CONVERT(N);
630 case ISD::FADD: return visitFADD(N);
631 case ISD::FSUB: return visitFSUB(N);
632 case ISD::FMUL: return visitFMUL(N);
633 case ISD::FDIV: return visitFDIV(N);
634 case ISD::FREM: return visitFREM(N);
635 case ISD::FCOPYSIGN: return visitFCOPYSIGN(N);
636 case ISD::SINT_TO_FP: return visitSINT_TO_FP(N);
637 case ISD::UINT_TO_FP: return visitUINT_TO_FP(N);
638 case ISD::FP_TO_SINT: return visitFP_TO_SINT(N);
639 case ISD::FP_TO_UINT: return visitFP_TO_UINT(N);
640 case ISD::FP_ROUND: return visitFP_ROUND(N);
641 case ISD::FP_ROUND_INREG: return visitFP_ROUND_INREG(N);
642 case ISD::FP_EXTEND: return visitFP_EXTEND(N);
643 case ISD::FNEG: return visitFNEG(N);
644 case ISD::FABS: return visitFABS(N);
645 case ISD::BRCOND: return visitBRCOND(N);
646 case ISD::BR_CC: return visitBR_CC(N);
647 case ISD::LOAD: return visitLOAD(N);
650 case ISD::ZEXTLOAD: return visitXEXTLOAD(N);
651 case ISD::STORE: return visitSTORE(N);
652 case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N);
653 case ISD::VINSERT_VECTOR_ELT: return visitVINSERT_VECTOR_ELT(N);
654 case ISD::VBUILD_VECTOR: return visitVBUILD_VECTOR(N);
655 case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N);
656 case ISD::VVECTOR_SHUFFLE: return visitVVECTOR_SHUFFLE(N);
661 SDOperand DAGCombiner::visitTokenFactor(SDNode *N) {
662 std::vector<SDOperand> Ops;
663 bool Changed = false;
665 // If the token factor has two operands and one is the entry token, replace
666 // the token factor with the other operand.
667 if (N->getNumOperands() == 2) {
668 if (N->getOperand(0).getOpcode() == ISD::EntryToken)
669 return N->getOperand(1);
670 if (N->getOperand(1).getOpcode() == ISD::EntryToken)
671 return N->getOperand(0);
674 // fold (tokenfactor (tokenfactor)) -> tokenfactor
675 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
676 SDOperand Op = N->getOperand(i);
677 if (Op.getOpcode() == ISD::TokenFactor && Op.hasOneUse()) {
678 AddToWorkList(Op.Val); // Remove dead node.
680 for (unsigned j = 0, e = Op.getNumOperands(); j != e; ++j)
681 Ops.push_back(Op.getOperand(j));
687 return DAG.getNode(ISD::TokenFactor, MVT::Other, Ops);
691 SDOperand DAGCombiner::visitADD(SDNode *N) {
692 SDOperand N0 = N->getOperand(0);
693 SDOperand N1 = N->getOperand(1);
694 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
695 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
696 MVT::ValueType VT = N0.getValueType();
698 // fold (add c1, c2) -> c1+c2
700 return DAG.getNode(ISD::ADD, VT, N0, N1);
701 // canonicalize constant to RHS
703 return DAG.getNode(ISD::ADD, VT, N1, N0);
704 // fold (add x, 0) -> x
705 if (N1C && N1C->isNullValue())
707 // fold ((c1-A)+c2) -> (c1+c2)-A
708 if (N1C && N0.getOpcode() == ISD::SUB)
709 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getOperand(0)))
710 return DAG.getNode(ISD::SUB, VT,
711 DAG.getConstant(N1C->getValue()+N0C->getValue(), VT),
714 SDOperand RADD = ReassociateOps(ISD::ADD, N0, N1);
717 // fold ((0-A) + B) -> B-A
718 if (N0.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N0.getOperand(0)) &&
719 cast<ConstantSDNode>(N0.getOperand(0))->isNullValue())
720 return DAG.getNode(ISD::SUB, VT, N1, N0.getOperand(1));
721 // fold (A + (0-B)) -> A-B
722 if (N1.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N1.getOperand(0)) &&
723 cast<ConstantSDNode>(N1.getOperand(0))->isNullValue())
724 return DAG.getNode(ISD::SUB, VT, N0, N1.getOperand(1));
725 // fold (A+(B-A)) -> B
726 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1))
727 return N1.getOperand(0);
729 if (!MVT::isVector(VT) && SimplifyDemandedBits(SDOperand(N, 0)))
732 // fold (a+b) -> (a|b) iff a and b share no bits.
733 if (MVT::isInteger(VT) && !MVT::isVector(VT)) {
734 uint64_t LHSZero, LHSOne;
735 uint64_t RHSZero, RHSOne;
736 uint64_t Mask = MVT::getIntVTBitMask(VT);
737 TLI.ComputeMaskedBits(N0, Mask, LHSZero, LHSOne);
739 TLI.ComputeMaskedBits(N1, Mask, RHSZero, RHSOne);
741 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
742 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
743 if ((RHSZero & (~LHSZero & Mask)) == (~LHSZero & Mask) ||
744 (LHSZero & (~RHSZero & Mask)) == (~RHSZero & Mask))
745 return DAG.getNode(ISD::OR, VT, N0, N1);
752 SDOperand DAGCombiner::visitSUB(SDNode *N) {
753 SDOperand N0 = N->getOperand(0);
754 SDOperand N1 = N->getOperand(1);
755 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val);
756 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val);
757 MVT::ValueType VT = N0.getValueType();
759 // fold (sub x, x) -> 0
761 return DAG.getConstant(0, N->getValueType(0));
762 // fold (sub c1, c2) -> c1-c2
764 return DAG.getNode(ISD::SUB, VT, N0, N1);
765 // fold (sub x, c) -> (add x, -c)
767 return DAG.getNode(ISD::ADD, VT, N0, DAG.getConstant(-N1C->getValue(), VT));
769 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1)
770 return N0.getOperand(1);
772 if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1)
773 return N0.getOperand(0);
777 SDOperand DAGCombiner::visitMUL(SDNode *N) {
778 SDOperand N0 = N->getOperand(0);
779 SDOperand N1 = N->getOperand(1);
780 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
781 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
782 MVT::ValueType VT = N0.getValueType();
784 // fold (mul c1, c2) -> c1*c2
786 return DAG.getNode(ISD::MUL, VT, N0, N1);
787 // canonicalize constant to RHS
789 return DAG.getNode(ISD::MUL, VT, N1, N0);
790 // fold (mul x, 0) -> 0
791 if (N1C && N1C->isNullValue())
793 // fold (mul x, -1) -> 0-x
794 if (N1C && N1C->isAllOnesValue())
795 return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT), N0);
796 // fold (mul x, (1 << c)) -> x << c
797 if (N1C && isPowerOf2_64(N1C->getValue()))
798 return DAG.getNode(ISD::SHL, VT, N0,
799 DAG.getConstant(Log2_64(N1C->getValue()),
800 TLI.getShiftAmountTy()));
801 // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c
802 if (N1C && isPowerOf2_64(-N1C->getSignExtended())) {
803 // FIXME: If the input is something that is easily negated (e.g. a
804 // single-use add), we should put the negate there.
805 return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT),
806 DAG.getNode(ISD::SHL, VT, N0,
807 DAG.getConstant(Log2_64(-N1C->getSignExtended()),
808 TLI.getShiftAmountTy())));
811 // (mul (shl X, c1), c2) -> (mul X, c2 << c1)
812 if (N1C && N0.getOpcode() == ISD::SHL &&
813 isa<ConstantSDNode>(N0.getOperand(1))) {
814 SDOperand C3 = DAG.getNode(ISD::SHL, VT, N1, N0.getOperand(1));
815 AddToWorkList(C3.Val);
816 return DAG.getNode(ISD::MUL, VT, N0.getOperand(0), C3);
819 // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one
822 SDOperand Sh(0,0), Y(0,0);
823 // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)).
824 if (N0.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N0.getOperand(1)) &&
825 N0.Val->hasOneUse()) {
827 } else if (N1.getOpcode() == ISD::SHL &&
828 isa<ConstantSDNode>(N1.getOperand(1)) && N1.Val->hasOneUse()) {
832 SDOperand Mul = DAG.getNode(ISD::MUL, VT, Sh.getOperand(0), Y);
833 return DAG.getNode(ISD::SHL, VT, Mul, Sh.getOperand(1));
836 // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2)
837 if (N1C && N0.getOpcode() == ISD::ADD && N0.Val->hasOneUse() &&
838 isa<ConstantSDNode>(N0.getOperand(1))) {
839 return DAG.getNode(ISD::ADD, VT,
840 DAG.getNode(ISD::MUL, VT, N0.getOperand(0), N1),
841 DAG.getNode(ISD::MUL, VT, N0.getOperand(1), N1));
845 SDOperand RMUL = ReassociateOps(ISD::MUL, N0, N1);
851 SDOperand DAGCombiner::visitSDIV(SDNode *N) {
852 SDOperand N0 = N->getOperand(0);
853 SDOperand N1 = N->getOperand(1);
854 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val);
855 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val);
856 MVT::ValueType VT = N->getValueType(0);
858 // fold (sdiv c1, c2) -> c1/c2
859 if (N0C && N1C && !N1C->isNullValue())
860 return DAG.getNode(ISD::SDIV, VT, N0, N1);
861 // fold (sdiv X, 1) -> X
862 if (N1C && N1C->getSignExtended() == 1LL)
864 // fold (sdiv X, -1) -> 0-X
865 if (N1C && N1C->isAllOnesValue())
866 return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT), N0);
867 // If we know the sign bits of both operands are zero, strength reduce to a
868 // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2
869 uint64_t SignBit = 1ULL << (MVT::getSizeInBits(VT)-1);
870 if (TLI.MaskedValueIsZero(N1, SignBit) &&
871 TLI.MaskedValueIsZero(N0, SignBit))
872 return DAG.getNode(ISD::UDIV, N1.getValueType(), N0, N1);
873 // fold (sdiv X, pow2) -> simple ops after legalize
874 if (N1C && N1C->getValue() && !TLI.isIntDivCheap() &&
875 (isPowerOf2_64(N1C->getSignExtended()) ||
876 isPowerOf2_64(-N1C->getSignExtended()))) {
877 // If dividing by powers of two is cheap, then don't perform the following
879 if (TLI.isPow2DivCheap())
881 int64_t pow2 = N1C->getSignExtended();
882 int64_t abs2 = pow2 > 0 ? pow2 : -pow2;
883 unsigned lg2 = Log2_64(abs2);
884 // Splat the sign bit into the register
885 SDOperand SGN = DAG.getNode(ISD::SRA, VT, N0,
886 DAG.getConstant(MVT::getSizeInBits(VT)-1,
887 TLI.getShiftAmountTy()));
888 AddToWorkList(SGN.Val);
889 // Add (N0 < 0) ? abs2 - 1 : 0;
890 SDOperand SRL = DAG.getNode(ISD::SRL, VT, SGN,
891 DAG.getConstant(MVT::getSizeInBits(VT)-lg2,
892 TLI.getShiftAmountTy()));
893 SDOperand ADD = DAG.getNode(ISD::ADD, VT, N0, SRL);
894 AddToWorkList(SRL.Val);
895 AddToWorkList(ADD.Val); // Divide by pow2
896 SDOperand SRA = DAG.getNode(ISD::SRA, VT, ADD,
897 DAG.getConstant(lg2, TLI.getShiftAmountTy()));
898 // If we're dividing by a positive value, we're done. Otherwise, we must
899 // negate the result.
902 AddToWorkList(SRA.Val);
903 return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT), SRA);
905 // if integer divide is expensive and we satisfy the requirements, emit an
906 // alternate sequence.
907 if (N1C && (N1C->getSignExtended() < -1 || N1C->getSignExtended() > 1) &&
908 !TLI.isIntDivCheap()) {
909 SDOperand Op = BuildSDIV(N);
910 if (Op.Val) return Op;
915 SDOperand DAGCombiner::visitUDIV(SDNode *N) {
916 SDOperand N0 = N->getOperand(0);
917 SDOperand N1 = N->getOperand(1);
918 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val);
919 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val);
920 MVT::ValueType VT = N->getValueType(0);
922 // fold (udiv c1, c2) -> c1/c2
923 if (N0C && N1C && !N1C->isNullValue())
924 return DAG.getNode(ISD::UDIV, VT, N0, N1);
925 // fold (udiv x, (1 << c)) -> x >>u c
926 if (N1C && isPowerOf2_64(N1C->getValue()))
927 return DAG.getNode(ISD::SRL, VT, N0,
928 DAG.getConstant(Log2_64(N1C->getValue()),
929 TLI.getShiftAmountTy()));
930 // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2
931 if (N1.getOpcode() == ISD::SHL) {
932 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
933 if (isPowerOf2_64(SHC->getValue())) {
934 MVT::ValueType ADDVT = N1.getOperand(1).getValueType();
935 SDOperand Add = DAG.getNode(ISD::ADD, ADDVT, N1.getOperand(1),
936 DAG.getConstant(Log2_64(SHC->getValue()),
938 AddToWorkList(Add.Val);
939 return DAG.getNode(ISD::SRL, VT, N0, Add);
943 // fold (udiv x, c) -> alternate
944 if (N1C && N1C->getValue() && !TLI.isIntDivCheap()) {
945 SDOperand Op = BuildUDIV(N);
946 if (Op.Val) return Op;
951 SDOperand DAGCombiner::visitSREM(SDNode *N) {
952 SDOperand N0 = N->getOperand(0);
953 SDOperand N1 = N->getOperand(1);
954 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
955 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
956 MVT::ValueType VT = N->getValueType(0);
958 // fold (srem c1, c2) -> c1%c2
959 if (N0C && N1C && !N1C->isNullValue())
960 return DAG.getNode(ISD::SREM, VT, N0, N1);
961 // If we know the sign bits of both operands are zero, strength reduce to a
962 // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15
963 uint64_t SignBit = 1ULL << (MVT::getSizeInBits(VT)-1);
964 if (TLI.MaskedValueIsZero(N1, SignBit) &&
965 TLI.MaskedValueIsZero(N0, SignBit))
966 return DAG.getNode(ISD::UREM, VT, N0, N1);
970 SDOperand DAGCombiner::visitUREM(SDNode *N) {
971 SDOperand N0 = N->getOperand(0);
972 SDOperand N1 = N->getOperand(1);
973 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
974 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
975 MVT::ValueType VT = N->getValueType(0);
977 // fold (urem c1, c2) -> c1%c2
978 if (N0C && N1C && !N1C->isNullValue())
979 return DAG.getNode(ISD::UREM, VT, N0, N1);
980 // fold (urem x, pow2) -> (and x, pow2-1)
981 if (N1C && !N1C->isNullValue() && isPowerOf2_64(N1C->getValue()))
982 return DAG.getNode(ISD::AND, VT, N0, DAG.getConstant(N1C->getValue()-1,VT));
983 // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1))
984 if (N1.getOpcode() == ISD::SHL) {
985 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
986 if (isPowerOf2_64(SHC->getValue())) {
987 SDOperand Add = DAG.getNode(ISD::ADD, VT, N1,DAG.getConstant(~0ULL,VT));
988 AddToWorkList(Add.Val);
989 return DAG.getNode(ISD::AND, VT, N0, Add);
996 SDOperand DAGCombiner::visitMULHS(SDNode *N) {
997 SDOperand N0 = N->getOperand(0);
998 SDOperand N1 = N->getOperand(1);
999 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1001 // fold (mulhs x, 0) -> 0
1002 if (N1C && N1C->isNullValue())
1004 // fold (mulhs x, 1) -> (sra x, size(x)-1)
1005 if (N1C && N1C->getValue() == 1)
1006 return DAG.getNode(ISD::SRA, N0.getValueType(), N0,
1007 DAG.getConstant(MVT::getSizeInBits(N0.getValueType())-1,
1008 TLI.getShiftAmountTy()));
1012 SDOperand DAGCombiner::visitMULHU(SDNode *N) {
1013 SDOperand N0 = N->getOperand(0);
1014 SDOperand N1 = N->getOperand(1);
1015 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1017 // fold (mulhu x, 0) -> 0
1018 if (N1C && N1C->isNullValue())
1020 // fold (mulhu x, 1) -> 0
1021 if (N1C && N1C->getValue() == 1)
1022 return DAG.getConstant(0, N0.getValueType());
1026 SDOperand DAGCombiner::visitAND(SDNode *N) {
1027 SDOperand N0 = N->getOperand(0);
1028 SDOperand N1 = N->getOperand(1);
1029 SDOperand LL, LR, RL, RR, CC0, CC1;
1030 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1031 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1032 MVT::ValueType VT = N1.getValueType();
1033 unsigned OpSizeInBits = MVT::getSizeInBits(VT);
1035 // fold (and c1, c2) -> c1&c2
1037 return DAG.getNode(ISD::AND, VT, N0, N1);
1038 // canonicalize constant to RHS
1040 return DAG.getNode(ISD::AND, VT, N1, N0);
1041 // fold (and x, -1) -> x
1042 if (N1C && N1C->isAllOnesValue())
1044 // if (and x, c) is known to be zero, return 0
1045 if (N1C && TLI.MaskedValueIsZero(SDOperand(N, 0), MVT::getIntVTBitMask(VT)))
1046 return DAG.getConstant(0, VT);
1048 SDOperand RAND = ReassociateOps(ISD::AND, N0, N1);
1051 // fold (and (or x, 0xFFFF), 0xFF) -> 0xFF
1052 if (N1C && N0.getOpcode() == ISD::OR)
1053 if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
1054 if ((ORI->getValue() & N1C->getValue()) == N1C->getValue())
1056 // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits.
1057 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
1058 unsigned InMask = MVT::getIntVTBitMask(N0.getOperand(0).getValueType());
1059 if (TLI.MaskedValueIsZero(N0.getOperand(0),
1060 ~N1C->getValue() & InMask)) {
1061 SDOperand Zext = DAG.getNode(ISD::ZERO_EXTEND, N0.getValueType(),
1064 // Replace uses of the AND with uses of the Zero extend node.
1067 // We actually want to replace all uses of the any_extend with the
1068 // zero_extend, to avoid duplicating things. This will later cause this
1069 // AND to be folded.
1070 CombineTo(N0.Val, Zext);
1074 // fold (and (setcc x), (setcc y)) -> (setcc (and x, y))
1075 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
1076 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
1077 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
1079 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
1080 MVT::isInteger(LL.getValueType())) {
1081 // fold (X == 0) & (Y == 0) -> (X|Y == 0)
1082 if (cast<ConstantSDNode>(LR)->getValue() == 0 && Op1 == ISD::SETEQ) {
1083 SDOperand ORNode = DAG.getNode(ISD::OR, LR.getValueType(), LL, RL);
1084 AddToWorkList(ORNode.Val);
1085 return DAG.getSetCC(VT, ORNode, LR, Op1);
1087 // fold (X == -1) & (Y == -1) -> (X&Y == -1)
1088 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETEQ) {
1089 SDOperand ANDNode = DAG.getNode(ISD::AND, LR.getValueType(), LL, RL);
1090 AddToWorkList(ANDNode.Val);
1091 return DAG.getSetCC(VT, ANDNode, LR, Op1);
1093 // fold (X > -1) & (Y > -1) -> (X|Y > -1)
1094 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETGT) {
1095 SDOperand ORNode = DAG.getNode(ISD::OR, LR.getValueType(), LL, RL);
1096 AddToWorkList(ORNode.Val);
1097 return DAG.getSetCC(VT, ORNode, LR, Op1);
1100 // canonicalize equivalent to ll == rl
1101 if (LL == RR && LR == RL) {
1102 Op1 = ISD::getSetCCSwappedOperands(Op1);
1105 if (LL == RL && LR == RR) {
1106 bool isInteger = MVT::isInteger(LL.getValueType());
1107 ISD::CondCode Result = ISD::getSetCCAndOperation(Op0, Op1, isInteger);
1108 if (Result != ISD::SETCC_INVALID)
1109 return DAG.getSetCC(N0.getValueType(), LL, LR, Result);
1112 // fold (and (zext x), (zext y)) -> (zext (and x, y))
1113 if (N0.getOpcode() == ISD::ZERO_EXTEND &&
1114 N1.getOpcode() == ISD::ZERO_EXTEND &&
1115 N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()) {
1116 SDOperand ANDNode = DAG.getNode(ISD::AND, N0.getOperand(0).getValueType(),
1117 N0.getOperand(0), N1.getOperand(0));
1118 AddToWorkList(ANDNode.Val);
1119 return DAG.getNode(ISD::ZERO_EXTEND, VT, ANDNode);
1121 // fold (and (shl/srl/sra x), (shl/srl/sra y)) -> (shl/srl/sra (and x, y))
1122 if (((N0.getOpcode() == ISD::SHL && N1.getOpcode() == ISD::SHL) ||
1123 (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SRL) ||
1124 (N0.getOpcode() == ISD::SRA && N1.getOpcode() == ISD::SRA)) &&
1125 N0.getOperand(1) == N1.getOperand(1)) {
1126 SDOperand ANDNode = DAG.getNode(ISD::AND, N0.getOperand(0).getValueType(),
1127 N0.getOperand(0), N1.getOperand(0));
1128 AddToWorkList(ANDNode.Val);
1129 return DAG.getNode(N0.getOpcode(), VT, ANDNode, N0.getOperand(1));
1131 // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1)
1132 // fold (and (sra)) -> (and (srl)) when possible.
1133 if (!MVT::isVector(VT) &&
1134 SimplifyDemandedBits(SDOperand(N, 0)))
1136 // fold (zext_inreg (extload x)) -> (zextload x)
1137 if (N0.getOpcode() == ISD::EXTLOAD) {
1138 MVT::ValueType EVT = cast<VTSDNode>(N0.getOperand(3))->getVT();
1139 // If we zero all the possible extended bits, then we can turn this into
1140 // a zextload if we are running before legalize or the operation is legal.
1141 if (TLI.MaskedValueIsZero(N1, ~0ULL << MVT::getSizeInBits(EVT)) &&
1142 (!AfterLegalize || TLI.isOperationLegal(ISD::ZEXTLOAD, EVT))) {
1143 SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, N0.getOperand(0),
1144 N0.getOperand(1), N0.getOperand(2),
1147 CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1));
1151 // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use
1152 if (N0.getOpcode() == ISD::SEXTLOAD && N0.hasOneUse()) {
1153 MVT::ValueType EVT = cast<VTSDNode>(N0.getOperand(3))->getVT();
1154 // If we zero all the possible extended bits, then we can turn this into
1155 // a zextload if we are running before legalize or the operation is legal.
1156 if (TLI.MaskedValueIsZero(N1, ~0ULL << MVT::getSizeInBits(EVT)) &&
1157 (!AfterLegalize || TLI.isOperationLegal(ISD::ZEXTLOAD, EVT))) {
1158 SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, N0.getOperand(0),
1159 N0.getOperand(1), N0.getOperand(2),
1162 CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1));
1167 // fold (and (load x), 255) -> (zextload x, i8)
1168 // fold (and (extload x, i16), 255) -> (zextload x, i8)
1170 (N0.getOpcode() == ISD::LOAD || N0.getOpcode() == ISD::EXTLOAD ||
1171 N0.getOpcode() == ISD::ZEXTLOAD) &&
1173 MVT::ValueType EVT, LoadedVT;
1174 if (N1C->getValue() == 255)
1176 else if (N1C->getValue() == 65535)
1178 else if (N1C->getValue() == ~0U)
1183 LoadedVT = N0.getOpcode() == ISD::LOAD ? VT :
1184 cast<VTSDNode>(N0.getOperand(3))->getVT();
1185 if (EVT != MVT::Other && LoadedVT > EVT) {
1186 MVT::ValueType PtrType = N0.getOperand(1).getValueType();
1187 // For big endian targets, we need to add an offset to the pointer to load
1188 // the correct bytes. For little endian systems, we merely need to read
1189 // fewer bytes from the same pointer.
1191 (MVT::getSizeInBits(LoadedVT) - MVT::getSizeInBits(EVT)) / 8;
1192 SDOperand NewPtr = N0.getOperand(1);
1193 if (!TLI.isLittleEndian())
1194 NewPtr = DAG.getNode(ISD::ADD, PtrType, NewPtr,
1195 DAG.getConstant(PtrOff, PtrType));
1196 AddToWorkList(NewPtr.Val);
1198 DAG.getExtLoad(ISD::ZEXTLOAD, VT, N0.getOperand(0), NewPtr,
1199 N0.getOperand(2), EVT);
1201 CombineTo(N0.Val, Load, Load.getValue(1));
1209 SDOperand DAGCombiner::visitOR(SDNode *N) {
1210 SDOperand N0 = N->getOperand(0);
1211 SDOperand N1 = N->getOperand(1);
1212 SDOperand LL, LR, RL, RR, CC0, CC1;
1213 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1214 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1215 MVT::ValueType VT = N1.getValueType();
1216 unsigned OpSizeInBits = MVT::getSizeInBits(VT);
1218 // fold (or c1, c2) -> c1|c2
1220 return DAG.getNode(ISD::OR, VT, N0, N1);
1221 // canonicalize constant to RHS
1223 return DAG.getNode(ISD::OR, VT, N1, N0);
1224 // fold (or x, 0) -> x
1225 if (N1C && N1C->isNullValue())
1227 // fold (or x, -1) -> -1
1228 if (N1C && N1C->isAllOnesValue())
1230 // fold (or x, c) -> c iff (x & ~c) == 0
1232 TLI.MaskedValueIsZero(N0,~N1C->getValue() & (~0ULL>>(64-OpSizeInBits))))
1235 SDOperand ROR = ReassociateOps(ISD::OR, N0, N1);
1238 // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2)
1239 if (N1C && N0.getOpcode() == ISD::AND && N0.Val->hasOneUse() &&
1240 isa<ConstantSDNode>(N0.getOperand(1))) {
1241 ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1));
1242 return DAG.getNode(ISD::AND, VT, DAG.getNode(ISD::OR, VT, N0.getOperand(0),
1244 DAG.getConstant(N1C->getValue() | C1->getValue(), VT));
1246 // fold (or (setcc x), (setcc y)) -> (setcc (or x, y))
1247 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
1248 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
1249 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
1251 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
1252 MVT::isInteger(LL.getValueType())) {
1253 // fold (X != 0) | (Y != 0) -> (X|Y != 0)
1254 // fold (X < 0) | (Y < 0) -> (X|Y < 0)
1255 if (cast<ConstantSDNode>(LR)->getValue() == 0 &&
1256 (Op1 == ISD::SETNE || Op1 == ISD::SETLT)) {
1257 SDOperand ORNode = DAG.getNode(ISD::OR, LR.getValueType(), LL, RL);
1258 AddToWorkList(ORNode.Val);
1259 return DAG.getSetCC(VT, ORNode, LR, Op1);
1261 // fold (X != -1) | (Y != -1) -> (X&Y != -1)
1262 // fold (X > -1) | (Y > -1) -> (X&Y > -1)
1263 if (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
1264 (Op1 == ISD::SETNE || Op1 == ISD::SETGT)) {
1265 SDOperand ANDNode = DAG.getNode(ISD::AND, LR.getValueType(), LL, RL);
1266 AddToWorkList(ANDNode.Val);
1267 return DAG.getSetCC(VT, ANDNode, LR, Op1);
1270 // canonicalize equivalent to ll == rl
1271 if (LL == RR && LR == RL) {
1272 Op1 = ISD::getSetCCSwappedOperands(Op1);
1275 if (LL == RL && LR == RR) {
1276 bool isInteger = MVT::isInteger(LL.getValueType());
1277 ISD::CondCode Result = ISD::getSetCCOrOperation(Op0, Op1, isInteger);
1278 if (Result != ISD::SETCC_INVALID)
1279 return DAG.getSetCC(N0.getValueType(), LL, LR, Result);
1282 // fold (or (zext x), (zext y)) -> (zext (or x, y))
1283 if (N0.getOpcode() == ISD::ZERO_EXTEND &&
1284 N1.getOpcode() == ISD::ZERO_EXTEND &&
1285 N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()) {
1286 SDOperand ORNode = DAG.getNode(ISD::OR, N0.getOperand(0).getValueType(),
1287 N0.getOperand(0), N1.getOperand(0));
1288 AddToWorkList(ORNode.Val);
1289 return DAG.getNode(ISD::ZERO_EXTEND, VT, ORNode);
1291 // fold (or (shl/srl/sra x), (shl/srl/sra y)) -> (shl/srl/sra (or x, y))
1292 if (((N0.getOpcode() == ISD::SHL && N1.getOpcode() == ISD::SHL) ||
1293 (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SRL) ||
1294 (N0.getOpcode() == ISD::SRA && N1.getOpcode() == ISD::SRA)) &&
1295 N0.getOperand(1) == N1.getOperand(1)) {
1296 SDOperand ORNode = DAG.getNode(ISD::OR, N0.getOperand(0).getValueType(),
1297 N0.getOperand(0), N1.getOperand(0));
1298 AddToWorkList(ORNode.Val);
1299 return DAG.getNode(N0.getOpcode(), VT, ORNode, N0.getOperand(1));
1301 // canonicalize shl to left side in a shl/srl pair, to match rotate
1302 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
1304 // check for rotl, rotr
1305 if (N0.getOpcode() == ISD::SHL && N1.getOpcode() == ISD::SRL &&
1306 N0.getOperand(0) == N1.getOperand(0) &&
1307 TLI.isOperationLegal(ISD::ROTL, VT) && TLI.isTypeLegal(VT)) {
1308 // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1)
1309 if (N0.getOperand(1).getOpcode() == ISD::Constant &&
1310 N1.getOperand(1).getOpcode() == ISD::Constant) {
1311 uint64_t c1val = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
1312 uint64_t c2val = cast<ConstantSDNode>(N1.getOperand(1))->getValue();
1313 if ((c1val + c2val) == OpSizeInBits)
1314 return DAG.getNode(ISD::ROTL, VT, N0.getOperand(0), N0.getOperand(1));
1316 // fold (or (shl x, y), (srl x, (sub 32, y))) -> (rotl x, y)
1317 if (N1.getOperand(1).getOpcode() == ISD::SUB &&
1318 N0.getOperand(1) == N1.getOperand(1).getOperand(1))
1319 if (ConstantSDNode *SUBC =
1320 dyn_cast<ConstantSDNode>(N1.getOperand(1).getOperand(0)))
1321 if (SUBC->getValue() == OpSizeInBits)
1322 return DAG.getNode(ISD::ROTL, VT, N0.getOperand(0), N0.getOperand(1));
1323 // fold (or (shl x, (sub 32, y)), (srl x, r)) -> (rotr x, y)
1324 if (N0.getOperand(1).getOpcode() == ISD::SUB &&
1325 N1.getOperand(1) == N0.getOperand(1).getOperand(1))
1326 if (ConstantSDNode *SUBC =
1327 dyn_cast<ConstantSDNode>(N0.getOperand(1).getOperand(0)))
1328 if (SUBC->getValue() == OpSizeInBits) {
1329 if (TLI.isOperationLegal(ISD::ROTR, VT) && TLI.isTypeLegal(VT))
1330 return DAG.getNode(ISD::ROTR, VT, N0.getOperand(0),
1333 return DAG.getNode(ISD::ROTL, VT, N0.getOperand(0),
1340 SDOperand DAGCombiner::visitXOR(SDNode *N) {
1341 SDOperand N0 = N->getOperand(0);
1342 SDOperand N1 = N->getOperand(1);
1343 SDOperand LHS, RHS, CC;
1344 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1345 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1346 MVT::ValueType VT = N0.getValueType();
1348 // fold (xor c1, c2) -> c1^c2
1350 return DAG.getNode(ISD::XOR, VT, N0, N1);
1351 // canonicalize constant to RHS
1353 return DAG.getNode(ISD::XOR, VT, N1, N0);
1354 // fold (xor x, 0) -> x
1355 if (N1C && N1C->isNullValue())
1358 SDOperand RXOR = ReassociateOps(ISD::XOR, N0, N1);
1361 // fold !(x cc y) -> (x !cc y)
1362 if (N1C && N1C->getValue() == 1 && isSetCCEquivalent(N0, LHS, RHS, CC)) {
1363 bool isInt = MVT::isInteger(LHS.getValueType());
1364 ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
1366 if (N0.getOpcode() == ISD::SETCC)
1367 return DAG.getSetCC(VT, LHS, RHS, NotCC);
1368 if (N0.getOpcode() == ISD::SELECT_CC)
1369 return DAG.getSelectCC(LHS, RHS, N0.getOperand(2),N0.getOperand(3),NotCC);
1370 assert(0 && "Unhandled SetCC Equivalent!");
1373 // fold !(x or y) -> (!x and !y) iff x or y are setcc
1374 if (N1C && N1C->getValue() == 1 &&
1375 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
1376 SDOperand LHS = N0.getOperand(0), RHS = N0.getOperand(1);
1377 if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) {
1378 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
1379 LHS = DAG.getNode(ISD::XOR, VT, LHS, N1); // RHS = ~LHS
1380 RHS = DAG.getNode(ISD::XOR, VT, RHS, N1); // RHS = ~RHS
1381 AddToWorkList(LHS.Val); AddToWorkList(RHS.Val);
1382 return DAG.getNode(NewOpcode, VT, LHS, RHS);
1385 // fold !(x or y) -> (!x and !y) iff x or y are constants
1386 if (N1C && N1C->isAllOnesValue() &&
1387 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
1388 SDOperand LHS = N0.getOperand(0), RHS = N0.getOperand(1);
1389 if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) {
1390 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
1391 LHS = DAG.getNode(ISD::XOR, VT, LHS, N1); // RHS = ~LHS
1392 RHS = DAG.getNode(ISD::XOR, VT, RHS, N1); // RHS = ~RHS
1393 AddToWorkList(LHS.Val); AddToWorkList(RHS.Val);
1394 return DAG.getNode(NewOpcode, VT, LHS, RHS);
1397 // fold (xor (xor x, c1), c2) -> (xor x, c1^c2)
1398 if (N1C && N0.getOpcode() == ISD::XOR) {
1399 ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0));
1400 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
1402 return DAG.getNode(ISD::XOR, VT, N0.getOperand(1),
1403 DAG.getConstant(N1C->getValue()^N00C->getValue(), VT));
1405 return DAG.getNode(ISD::XOR, VT, N0.getOperand(0),
1406 DAG.getConstant(N1C->getValue()^N01C->getValue(), VT));
1408 // fold (xor x, x) -> 0
1410 if (!MVT::isVector(VT)) {
1411 return DAG.getConstant(0, VT);
1412 } else if (!AfterLegalize || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT)) {
1413 // Produce a vector of zeros.
1414 SDOperand El = DAG.getConstant(0, MVT::getVectorBaseType(VT));
1415 std::vector<SDOperand> Ops(MVT::getVectorNumElements(VT), El);
1416 return DAG.getNode(ISD::BUILD_VECTOR, VT, Ops);
1419 // fold (xor (zext x), (zext y)) -> (zext (xor x, y))
1420 if (N0.getOpcode() == ISD::ZERO_EXTEND &&
1421 N1.getOpcode() == ISD::ZERO_EXTEND &&
1422 N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()) {
1423 SDOperand XORNode = DAG.getNode(ISD::XOR, N0.getOperand(0).getValueType(),
1424 N0.getOperand(0), N1.getOperand(0));
1425 AddToWorkList(XORNode.Val);
1426 return DAG.getNode(ISD::ZERO_EXTEND, VT, XORNode);
1428 // fold (xor (shl/srl/sra x), (shl/srl/sra y)) -> (shl/srl/sra (xor x, y))
1429 if (((N0.getOpcode() == ISD::SHL && N1.getOpcode() == ISD::SHL) ||
1430 (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SRL) ||
1431 (N0.getOpcode() == ISD::SRA && N1.getOpcode() == ISD::SRA)) &&
1432 N0.getOperand(1) == N1.getOperand(1)) {
1433 SDOperand XORNode = DAG.getNode(ISD::XOR, N0.getOperand(0).getValueType(),
1434 N0.getOperand(0), N1.getOperand(0));
1435 AddToWorkList(XORNode.Val);
1436 return DAG.getNode(N0.getOpcode(), VT, XORNode, N0.getOperand(1));
1441 SDOperand DAGCombiner::visitSHL(SDNode *N) {
1442 SDOperand N0 = N->getOperand(0);
1443 SDOperand N1 = N->getOperand(1);
1444 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1445 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1446 MVT::ValueType VT = N0.getValueType();
1447 unsigned OpSizeInBits = MVT::getSizeInBits(VT);
1449 // fold (shl c1, c2) -> c1<<c2
1451 return DAG.getNode(ISD::SHL, VT, N0, N1);
1452 // fold (shl 0, x) -> 0
1453 if (N0C && N0C->isNullValue())
1455 // fold (shl x, c >= size(x)) -> undef
1456 if (N1C && N1C->getValue() >= OpSizeInBits)
1457 return DAG.getNode(ISD::UNDEF, VT);
1458 // fold (shl x, 0) -> x
1459 if (N1C && N1C->isNullValue())
1461 // if (shl x, c) is known to be zero, return 0
1462 if (TLI.MaskedValueIsZero(SDOperand(N, 0), MVT::getIntVTBitMask(VT)))
1463 return DAG.getConstant(0, VT);
1464 if (SimplifyDemandedBits(SDOperand(N, 0)))
1466 // fold (shl (shl x, c1), c2) -> 0 or (shl x, c1+c2)
1467 if (N1C && N0.getOpcode() == ISD::SHL &&
1468 N0.getOperand(1).getOpcode() == ISD::Constant) {
1469 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
1470 uint64_t c2 = N1C->getValue();
1471 if (c1 + c2 > OpSizeInBits)
1472 return DAG.getConstant(0, VT);
1473 return DAG.getNode(ISD::SHL, VT, N0.getOperand(0),
1474 DAG.getConstant(c1 + c2, N1.getValueType()));
1476 // fold (shl (srl x, c1), c2) -> (shl (and x, -1 << c1), c2-c1) or
1477 // (srl (and x, -1 << c1), c1-c2)
1478 if (N1C && N0.getOpcode() == ISD::SRL &&
1479 N0.getOperand(1).getOpcode() == ISD::Constant) {
1480 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
1481 uint64_t c2 = N1C->getValue();
1482 SDOperand Mask = DAG.getNode(ISD::AND, VT, N0.getOperand(0),
1483 DAG.getConstant(~0ULL << c1, VT));
1485 return DAG.getNode(ISD::SHL, VT, Mask,
1486 DAG.getConstant(c2-c1, N1.getValueType()));
1488 return DAG.getNode(ISD::SRL, VT, Mask,
1489 DAG.getConstant(c1-c2, N1.getValueType()));
1491 // fold (shl (sra x, c1), c1) -> (and x, -1 << c1)
1492 if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1))
1493 return DAG.getNode(ISD::AND, VT, N0.getOperand(0),
1494 DAG.getConstant(~0ULL << N1C->getValue(), VT));
1495 // fold (shl (add x, c1), c2) -> (add (shl x, c2), c1<<c2)
1496 if (N1C && N0.getOpcode() == ISD::ADD && N0.Val->hasOneUse() &&
1497 isa<ConstantSDNode>(N0.getOperand(1))) {
1498 return DAG.getNode(ISD::ADD, VT,
1499 DAG.getNode(ISD::SHL, VT, N0.getOperand(0), N1),
1500 DAG.getNode(ISD::SHL, VT, N0.getOperand(1), N1));
1505 SDOperand DAGCombiner::visitSRA(SDNode *N) {
1506 SDOperand N0 = N->getOperand(0);
1507 SDOperand N1 = N->getOperand(1);
1508 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1509 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1510 MVT::ValueType VT = N0.getValueType();
1512 // fold (sra c1, c2) -> c1>>c2
1514 return DAG.getNode(ISD::SRA, VT, N0, N1);
1515 // fold (sra 0, x) -> 0
1516 if (N0C && N0C->isNullValue())
1518 // fold (sra -1, x) -> -1
1519 if (N0C && N0C->isAllOnesValue())
1521 // fold (sra x, c >= size(x)) -> undef
1522 if (N1C && N1C->getValue() >= MVT::getSizeInBits(VT))
1523 return DAG.getNode(ISD::UNDEF, VT);
1524 // fold (sra x, 0) -> x
1525 if (N1C && N1C->isNullValue())
1527 // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports
1529 if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) {
1530 unsigned LowBits = MVT::getSizeInBits(VT) - (unsigned)N1C->getValue();
1533 default: EVT = MVT::Other; break;
1534 case 1: EVT = MVT::i1; break;
1535 case 8: EVT = MVT::i8; break;
1536 case 16: EVT = MVT::i16; break;
1537 case 32: EVT = MVT::i32; break;
1539 if (EVT > MVT::Other && TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, EVT))
1540 return DAG.getNode(ISD::SIGN_EXTEND_INREG, VT, N0.getOperand(0),
1541 DAG.getValueType(EVT));
1544 // fold (sra (sra x, c1), c2) -> (sra x, c1+c2)
1545 if (N1C && N0.getOpcode() == ISD::SRA) {
1546 if (ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1547 unsigned Sum = N1C->getValue() + C1->getValue();
1548 if (Sum >= MVT::getSizeInBits(VT)) Sum = MVT::getSizeInBits(VT)-1;
1549 return DAG.getNode(ISD::SRA, VT, N0.getOperand(0),
1550 DAG.getConstant(Sum, N1C->getValueType(0)));
1554 // If the sign bit is known to be zero, switch this to a SRL.
1555 if (TLI.MaskedValueIsZero(N0, MVT::getIntVTSignBit(VT)))
1556 return DAG.getNode(ISD::SRL, VT, N0, N1);
1560 SDOperand DAGCombiner::visitSRL(SDNode *N) {
1561 SDOperand N0 = N->getOperand(0);
1562 SDOperand N1 = N->getOperand(1);
1563 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1564 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1565 MVT::ValueType VT = N0.getValueType();
1566 unsigned OpSizeInBits = MVT::getSizeInBits(VT);
1568 // fold (srl c1, c2) -> c1 >>u c2
1570 return DAG.getNode(ISD::SRL, VT, N0, N1);
1571 // fold (srl 0, x) -> 0
1572 if (N0C && N0C->isNullValue())
1574 // fold (srl x, c >= size(x)) -> undef
1575 if (N1C && N1C->getValue() >= OpSizeInBits)
1576 return DAG.getNode(ISD::UNDEF, VT);
1577 // fold (srl x, 0) -> x
1578 if (N1C && N1C->isNullValue())
1580 // if (srl x, c) is known to be zero, return 0
1581 if (N1C && TLI.MaskedValueIsZero(SDOperand(N, 0), ~0ULL >> (64-OpSizeInBits)))
1582 return DAG.getConstant(0, VT);
1583 // fold (srl (srl x, c1), c2) -> 0 or (srl x, c1+c2)
1584 if (N1C && N0.getOpcode() == ISD::SRL &&
1585 N0.getOperand(1).getOpcode() == ISD::Constant) {
1586 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
1587 uint64_t c2 = N1C->getValue();
1588 if (c1 + c2 > OpSizeInBits)
1589 return DAG.getConstant(0, VT);
1590 return DAG.getNode(ISD::SRL, VT, N0.getOperand(0),
1591 DAG.getConstant(c1 + c2, N1.getValueType()));
1596 SDOperand DAGCombiner::visitCTLZ(SDNode *N) {
1597 SDOperand N0 = N->getOperand(0);
1598 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1599 MVT::ValueType VT = N->getValueType(0);
1601 // fold (ctlz c1) -> c2
1603 return DAG.getNode(ISD::CTLZ, VT, N0);
1607 SDOperand DAGCombiner::visitCTTZ(SDNode *N) {
1608 SDOperand N0 = N->getOperand(0);
1609 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1610 MVT::ValueType VT = N->getValueType(0);
1612 // fold (cttz c1) -> c2
1614 return DAG.getNode(ISD::CTTZ, VT, N0);
1618 SDOperand DAGCombiner::visitCTPOP(SDNode *N) {
1619 SDOperand N0 = N->getOperand(0);
1620 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1621 MVT::ValueType VT = N->getValueType(0);
1623 // fold (ctpop c1) -> c2
1625 return DAG.getNode(ISD::CTPOP, VT, N0);
1629 SDOperand DAGCombiner::visitSELECT(SDNode *N) {
1630 SDOperand N0 = N->getOperand(0);
1631 SDOperand N1 = N->getOperand(1);
1632 SDOperand N2 = N->getOperand(2);
1633 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1634 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1635 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
1636 MVT::ValueType VT = N->getValueType(0);
1638 // fold select C, X, X -> X
1641 // fold select true, X, Y -> X
1642 if (N0C && !N0C->isNullValue())
1644 // fold select false, X, Y -> Y
1645 if (N0C && N0C->isNullValue())
1647 // fold select C, 1, X -> C | X
1648 if (MVT::i1 == VT && N1C && N1C->getValue() == 1)
1649 return DAG.getNode(ISD::OR, VT, N0, N2);
1650 // fold select C, 0, X -> ~C & X
1651 // FIXME: this should check for C type == X type, not i1?
1652 if (MVT::i1 == VT && N1C && N1C->isNullValue()) {
1653 SDOperand XORNode = DAG.getNode(ISD::XOR, VT, N0, DAG.getConstant(1, VT));
1654 AddToWorkList(XORNode.Val);
1655 return DAG.getNode(ISD::AND, VT, XORNode, N2);
1657 // fold select C, X, 1 -> ~C | X
1658 if (MVT::i1 == VT && N2C && N2C->getValue() == 1) {
1659 SDOperand XORNode = DAG.getNode(ISD::XOR, VT, N0, DAG.getConstant(1, VT));
1660 AddToWorkList(XORNode.Val);
1661 return DAG.getNode(ISD::OR, VT, XORNode, N1);
1663 // fold select C, X, 0 -> C & X
1664 // FIXME: this should check for C type == X type, not i1?
1665 if (MVT::i1 == VT && N2C && N2C->isNullValue())
1666 return DAG.getNode(ISD::AND, VT, N0, N1);
1667 // fold X ? X : Y --> X ? 1 : Y --> X | Y
1668 if (MVT::i1 == VT && N0 == N1)
1669 return DAG.getNode(ISD::OR, VT, N0, N2);
1670 // fold X ? Y : X --> X ? Y : 0 --> X & Y
1671 if (MVT::i1 == VT && N0 == N2)
1672 return DAG.getNode(ISD::AND, VT, N0, N1);
1673 // If we can fold this based on the true/false value, do so.
1674 if (SimplifySelectOps(N, N1, N2))
1676 // fold selects based on a setcc into other things, such as min/max/abs
1677 if (N0.getOpcode() == ISD::SETCC)
1679 // Check against MVT::Other for SELECT_CC, which is a workaround for targets
1680 // having to say they don't support SELECT_CC on every type the DAG knows
1681 // about, since there is no way to mark an opcode illegal at all value types
1682 if (TLI.isOperationLegal(ISD::SELECT_CC, MVT::Other))
1683 return DAG.getNode(ISD::SELECT_CC, VT, N0.getOperand(0), N0.getOperand(1),
1684 N1, N2, N0.getOperand(2));
1686 return SimplifySelect(N0, N1, N2);
1690 SDOperand DAGCombiner::visitSELECT_CC(SDNode *N) {
1691 SDOperand N0 = N->getOperand(0);
1692 SDOperand N1 = N->getOperand(1);
1693 SDOperand N2 = N->getOperand(2);
1694 SDOperand N3 = N->getOperand(3);
1695 SDOperand N4 = N->getOperand(4);
1696 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1697 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1698 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
1699 ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get();
1701 // Determine if the condition we're dealing with is constant
1702 SDOperand SCC = SimplifySetCC(TLI.getSetCCResultTy(), N0, N1, CC, false);
1703 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.Val);
1705 // fold select_cc lhs, rhs, x, x, cc -> x
1709 // If we can fold this based on the true/false value, do so.
1710 if (SimplifySelectOps(N, N2, N3))
1713 // fold select_cc into other things, such as min/max/abs
1714 return SimplifySelectCC(N0, N1, N2, N3, CC);
1717 SDOperand DAGCombiner::visitSETCC(SDNode *N) {
1718 return SimplifySetCC(N->getValueType(0), N->getOperand(0), N->getOperand(1),
1719 cast<CondCodeSDNode>(N->getOperand(2))->get());
1722 SDOperand DAGCombiner::visitSIGN_EXTEND(SDNode *N) {
1723 SDOperand N0 = N->getOperand(0);
1724 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1725 MVT::ValueType VT = N->getValueType(0);
1727 // fold (sext c1) -> c1
1729 return DAG.getNode(ISD::SIGN_EXTEND, VT, N0);
1730 // fold (sext (sext x)) -> (sext x)
1731 if (N0.getOpcode() == ISD::SIGN_EXTEND)
1732 return DAG.getNode(ISD::SIGN_EXTEND, VT, N0.getOperand(0));
1733 // fold (sext (truncate x)) -> (sextinreg x) iff x size == sext size.
1734 if (N0.getOpcode() == ISD::TRUNCATE && N0.getOperand(0).getValueType() == VT&&
1736 TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, N0.getValueType())))
1737 return DAG.getNode(ISD::SIGN_EXTEND_INREG, VT, N0.getOperand(0),
1738 DAG.getValueType(N0.getValueType()));
1739 // fold (sext (load x)) -> (sext (truncate (sextload x)))
1740 if (N0.getOpcode() == ISD::LOAD && N0.hasOneUse() &&
1741 (!AfterLegalize||TLI.isOperationLegal(ISD::SEXTLOAD, N0.getValueType()))){
1742 SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, N0.getOperand(0),
1743 N0.getOperand(1), N0.getOperand(2),
1745 CombineTo(N, ExtLoad);
1746 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
1747 ExtLoad.getValue(1));
1751 // fold (sext (sextload x)) -> (sext (truncate (sextload x)))
1752 // fold (sext ( extload x)) -> (sext (truncate (sextload x)))
1753 if ((N0.getOpcode() == ISD::SEXTLOAD || N0.getOpcode() == ISD::EXTLOAD) &&
1755 SDOperand ExtLoad = DAG.getNode(ISD::SEXTLOAD, VT, N0.getOperand(0),
1756 N0.getOperand(1), N0.getOperand(2),
1758 CombineTo(N, ExtLoad);
1759 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
1760 ExtLoad.getValue(1));
1767 SDOperand DAGCombiner::visitZERO_EXTEND(SDNode *N) {
1768 SDOperand N0 = N->getOperand(0);
1769 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1770 MVT::ValueType VT = N->getValueType(0);
1772 // fold (zext c1) -> c1
1774 return DAG.getNode(ISD::ZERO_EXTEND, VT, N0);
1775 // fold (zext (zext x)) -> (zext x)
1776 if (N0.getOpcode() == ISD::ZERO_EXTEND)
1777 return DAG.getNode(ISD::ZERO_EXTEND, VT, N0.getOperand(0));
1778 // fold (zext (truncate x)) -> (zextinreg x) iff x size == zext size.
1779 if (N0.getOpcode() == ISD::TRUNCATE && N0.getOperand(0).getValueType() == VT&&
1780 (!AfterLegalize || TLI.isOperationLegal(ISD::AND, N0.getValueType())))
1781 return DAG.getZeroExtendInReg(N0.getOperand(0), N0.getValueType());
1782 // fold (zext (load x)) -> (zext (truncate (zextload x)))
1783 if (N0.getOpcode() == ISD::LOAD && N0.hasOneUse() &&
1784 (!AfterLegalize||TLI.isOperationLegal(ISD::ZEXTLOAD, N0.getValueType()))){
1785 SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, N0.getOperand(0),
1786 N0.getOperand(1), N0.getOperand(2),
1788 CombineTo(N, ExtLoad);
1789 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
1790 ExtLoad.getValue(1));
1794 // fold (zext (zextload x)) -> (zext (truncate (zextload x)))
1795 // fold (zext ( extload x)) -> (zext (truncate (zextload x)))
1796 if ((N0.getOpcode() == ISD::ZEXTLOAD || N0.getOpcode() == ISD::EXTLOAD) &&
1798 SDOperand ExtLoad = DAG.getNode(ISD::ZEXTLOAD, VT, N0.getOperand(0),
1799 N0.getOperand(1), N0.getOperand(2),
1801 CombineTo(N, ExtLoad);
1802 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
1803 ExtLoad.getValue(1));
1809 SDOperand DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
1810 SDOperand N0 = N->getOperand(0);
1811 SDOperand N1 = N->getOperand(1);
1812 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1813 MVT::ValueType VT = N->getValueType(0);
1814 MVT::ValueType EVT = cast<VTSDNode>(N1)->getVT();
1815 unsigned EVTBits = MVT::getSizeInBits(EVT);
1817 // fold (sext_in_reg c1) -> c1
1819 SDOperand Truncate = DAG.getConstant(N0C->getValue(), EVT);
1820 return DAG.getNode(ISD::SIGN_EXTEND, VT, Truncate);
1822 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt1
1823 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
1824 cast<VTSDNode>(N0.getOperand(1))->getVT() <= EVT) {
1827 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2
1828 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
1829 EVT < cast<VTSDNode>(N0.getOperand(1))->getVT()) {
1830 return DAG.getNode(ISD::SIGN_EXTEND_INREG, VT, N0.getOperand(0), N1);
1832 // fold (sext_in_reg (assert_sext x)) -> (assert_sext x)
1833 if (N0.getOpcode() == ISD::AssertSext &&
1834 cast<VTSDNode>(N0.getOperand(1))->getVT() <= EVT) {
1837 // fold (sext_in_reg (sextload x)) -> (sextload x)
1838 if (N0.getOpcode() == ISD::SEXTLOAD &&
1839 cast<VTSDNode>(N0.getOperand(3))->getVT() <= EVT) {
1842 // fold (sext_in_reg (setcc x)) -> setcc x iff (setcc x) == 0 or -1
1843 if (N0.getOpcode() == ISD::SETCC &&
1844 TLI.getSetCCResultContents() ==
1845 TargetLowering::ZeroOrNegativeOneSetCCResult)
1847 // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is zero
1848 if (TLI.MaskedValueIsZero(N0, 1ULL << (EVTBits-1)))
1849 return DAG.getZeroExtendInReg(N0, EVT);
1850 // fold (sext_in_reg (srl x)) -> sra x
1851 if (N0.getOpcode() == ISD::SRL &&
1852 N0.getOperand(1).getOpcode() == ISD::Constant &&
1853 cast<ConstantSDNode>(N0.getOperand(1))->getValue() == EVTBits) {
1854 return DAG.getNode(ISD::SRA, N0.getValueType(), N0.getOperand(0),
1857 // fold (sext_inreg (extload x)) -> (sextload x)
1858 if (N0.getOpcode() == ISD::EXTLOAD &&
1859 EVT == cast<VTSDNode>(N0.getOperand(3))->getVT() &&
1860 (!AfterLegalize || TLI.isOperationLegal(ISD::SEXTLOAD, EVT))) {
1861 SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, N0.getOperand(0),
1862 N0.getOperand(1), N0.getOperand(2),
1864 CombineTo(N, ExtLoad);
1865 CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1));
1868 // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use
1869 if (N0.getOpcode() == ISD::ZEXTLOAD && N0.hasOneUse() &&
1870 EVT == cast<VTSDNode>(N0.getOperand(3))->getVT() &&
1871 (!AfterLegalize || TLI.isOperationLegal(ISD::SEXTLOAD, EVT))) {
1872 SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, N0.getOperand(0),
1873 N0.getOperand(1), N0.getOperand(2),
1875 CombineTo(N, ExtLoad);
1876 CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1));
1882 SDOperand DAGCombiner::visitTRUNCATE(SDNode *N) {
1883 SDOperand N0 = N->getOperand(0);
1884 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1885 MVT::ValueType VT = N->getValueType(0);
1888 if (N0.getValueType() == N->getValueType(0))
1890 // fold (truncate c1) -> c1
1892 return DAG.getNode(ISD::TRUNCATE, VT, N0);
1893 // fold (truncate (truncate x)) -> (truncate x)
1894 if (N0.getOpcode() == ISD::TRUNCATE)
1895 return DAG.getNode(ISD::TRUNCATE, VT, N0.getOperand(0));
1896 // fold (truncate (ext x)) -> (ext x) or (truncate x) or x
1897 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::SIGN_EXTEND){
1898 if (N0.getValueType() < VT)
1899 // if the source is smaller than the dest, we still need an extend
1900 return DAG.getNode(N0.getOpcode(), VT, N0.getOperand(0));
1901 else if (N0.getValueType() > VT)
1902 // if the source is larger than the dest, than we just need the truncate
1903 return DAG.getNode(ISD::TRUNCATE, VT, N0.getOperand(0));
1905 // if the source and dest are the same type, we can drop both the extend
1907 return N0.getOperand(0);
1909 // fold (truncate (load x)) -> (smaller load x)
1910 if (N0.getOpcode() == ISD::LOAD && N0.hasOneUse()) {
1911 assert(MVT::getSizeInBits(N0.getValueType()) > MVT::getSizeInBits(VT) &&
1912 "Cannot truncate to larger type!");
1913 MVT::ValueType PtrType = N0.getOperand(1).getValueType();
1914 // For big endian targets, we need to add an offset to the pointer to load
1915 // the correct bytes. For little endian systems, we merely need to read
1916 // fewer bytes from the same pointer.
1918 (MVT::getSizeInBits(N0.getValueType()) - MVT::getSizeInBits(VT)) / 8;
1919 SDOperand NewPtr = TLI.isLittleEndian() ? N0.getOperand(1) :
1920 DAG.getNode(ISD::ADD, PtrType, N0.getOperand(1),
1921 DAG.getConstant(PtrOff, PtrType));
1922 AddToWorkList(NewPtr.Val);
1923 SDOperand Load = DAG.getLoad(VT, N0.getOperand(0), NewPtr,N0.getOperand(2));
1925 CombineTo(N0.Val, Load, Load.getValue(1));
1931 SDOperand DAGCombiner::visitBIT_CONVERT(SDNode *N) {
1932 SDOperand N0 = N->getOperand(0);
1933 MVT::ValueType VT = N->getValueType(0);
1935 // If the input is a constant, let getNode() fold it.
1936 if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) {
1937 SDOperand Res = DAG.getNode(ISD::BIT_CONVERT, VT, N0);
1938 if (Res.Val != N) return Res;
1941 if (N0.getOpcode() == ISD::BIT_CONVERT) // conv(conv(x,t1),t2) -> conv(x,t2)
1942 return DAG.getNode(ISD::BIT_CONVERT, VT, N0.getOperand(0));
1944 // fold (conv (load x)) -> (load (conv*)x)
1945 // FIXME: These xforms need to know that the resultant load doesn't need a
1946 // higher alignment than the original!
1947 if (0 && N0.getOpcode() == ISD::LOAD && N0.hasOneUse()) {
1948 SDOperand Load = DAG.getLoad(VT, N0.getOperand(0), N0.getOperand(1),
1951 CombineTo(N0.Val, DAG.getNode(ISD::BIT_CONVERT, N0.getValueType(), Load),
1959 SDOperand DAGCombiner::visitFADD(SDNode *N) {
1960 SDOperand N0 = N->getOperand(0);
1961 SDOperand N1 = N->getOperand(1);
1962 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
1963 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
1964 MVT::ValueType VT = N->getValueType(0);
1966 // fold (fadd c1, c2) -> c1+c2
1968 return DAG.getNode(ISD::FADD, VT, N0, N1);
1969 // canonicalize constant to RHS
1970 if (N0CFP && !N1CFP)
1971 return DAG.getNode(ISD::FADD, VT, N1, N0);
1972 // fold (A + (-B)) -> A-B
1973 if (N1.getOpcode() == ISD::FNEG)
1974 return DAG.getNode(ISD::FSUB, VT, N0, N1.getOperand(0));
1975 // fold ((-A) + B) -> B-A
1976 if (N0.getOpcode() == ISD::FNEG)
1977 return DAG.getNode(ISD::FSUB, VT, N1, N0.getOperand(0));
1981 SDOperand DAGCombiner::visitFSUB(SDNode *N) {
1982 SDOperand N0 = N->getOperand(0);
1983 SDOperand N1 = N->getOperand(1);
1984 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
1985 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
1986 MVT::ValueType VT = N->getValueType(0);
1988 // fold (fsub c1, c2) -> c1-c2
1990 return DAG.getNode(ISD::FSUB, VT, N0, N1);
1991 // fold (A-(-B)) -> A+B
1992 if (N1.getOpcode() == ISD::FNEG)
1993 return DAG.getNode(ISD::FADD, VT, N0, N1.getOperand(0));
1997 SDOperand DAGCombiner::visitFMUL(SDNode *N) {
1998 SDOperand N0 = N->getOperand(0);
1999 SDOperand N1 = N->getOperand(1);
2000 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2001 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
2002 MVT::ValueType VT = N->getValueType(0);
2004 // fold (fmul c1, c2) -> c1*c2
2006 return DAG.getNode(ISD::FMUL, VT, N0, N1);
2007 // canonicalize constant to RHS
2008 if (N0CFP && !N1CFP)
2009 return DAG.getNode(ISD::FMUL, VT, N1, N0);
2010 // fold (fmul X, 2.0) -> (fadd X, X)
2011 if (N1CFP && N1CFP->isExactlyValue(+2.0))
2012 return DAG.getNode(ISD::FADD, VT, N0, N0);
2016 SDOperand DAGCombiner::visitFDIV(SDNode *N) {
2017 SDOperand N0 = N->getOperand(0);
2018 SDOperand N1 = N->getOperand(1);
2019 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2020 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
2021 MVT::ValueType VT = N->getValueType(0);
2023 // fold (fdiv c1, c2) -> c1/c2
2025 return DAG.getNode(ISD::FDIV, VT, N0, N1);
2029 SDOperand DAGCombiner::visitFREM(SDNode *N) {
2030 SDOperand N0 = N->getOperand(0);
2031 SDOperand N1 = N->getOperand(1);
2032 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2033 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
2034 MVT::ValueType VT = N->getValueType(0);
2036 // fold (frem c1, c2) -> fmod(c1,c2)
2038 return DAG.getNode(ISD::FREM, VT, N0, N1);
2042 SDOperand DAGCombiner::visitFCOPYSIGN(SDNode *N) {
2043 SDOperand N0 = N->getOperand(0);
2044 SDOperand N1 = N->getOperand(1);
2045 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2046 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
2047 MVT::ValueType VT = N->getValueType(0);
2049 if (N0CFP && N1CFP) // Constant fold
2050 return DAG.getNode(ISD::FCOPYSIGN, VT, N0, N1);
2053 // copysign(x, c1) -> fabs(x) iff ispos(c1)
2054 // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1)
2059 u.d = N1CFP->getValue();
2061 return DAG.getNode(ISD::FABS, VT, N0);
2063 return DAG.getNode(ISD::FNEG, VT, DAG.getNode(ISD::FABS, VT, N0));
2066 // copysign(fabs(x), y) -> copysign(x, y)
2067 // copysign(fneg(x), y) -> copysign(x, y)
2068 // copysign(copysign(x,z), y) -> copysign(x, y)
2069 if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG ||
2070 N0.getOpcode() == ISD::FCOPYSIGN)
2071 return DAG.getNode(ISD::FCOPYSIGN, VT, N0.getOperand(0), N1);
2073 // copysign(x, abs(y)) -> abs(x)
2074 if (N1.getOpcode() == ISD::FABS)
2075 return DAG.getNode(ISD::FABS, VT, N0);
2077 // copysign(x, copysign(y,z)) -> copysign(x, z)
2078 if (N1.getOpcode() == ISD::FCOPYSIGN)
2079 return DAG.getNode(ISD::FCOPYSIGN, VT, N0, N1.getOperand(1));
2081 // copysign(x, fp_extend(y)) -> copysign(x, y)
2082 // copysign(x, fp_round(y)) -> copysign(x, y)
2083 if (N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND)
2084 return DAG.getNode(ISD::FCOPYSIGN, VT, N0, N1.getOperand(0));
2091 SDOperand DAGCombiner::visitSINT_TO_FP(SDNode *N) {
2092 SDOperand N0 = N->getOperand(0);
2093 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2094 MVT::ValueType VT = N->getValueType(0);
2096 // fold (sint_to_fp c1) -> c1fp
2098 return DAG.getNode(ISD::SINT_TO_FP, VT, N0);
2102 SDOperand DAGCombiner::visitUINT_TO_FP(SDNode *N) {
2103 SDOperand N0 = N->getOperand(0);
2104 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2105 MVT::ValueType VT = N->getValueType(0);
2107 // fold (uint_to_fp c1) -> c1fp
2109 return DAG.getNode(ISD::UINT_TO_FP, VT, N0);
2113 SDOperand DAGCombiner::visitFP_TO_SINT(SDNode *N) {
2114 SDOperand N0 = N->getOperand(0);
2115 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2116 MVT::ValueType VT = N->getValueType(0);
2118 // fold (fp_to_sint c1fp) -> c1
2120 return DAG.getNode(ISD::FP_TO_SINT, VT, N0);
2124 SDOperand DAGCombiner::visitFP_TO_UINT(SDNode *N) {
2125 SDOperand N0 = N->getOperand(0);
2126 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2127 MVT::ValueType VT = N->getValueType(0);
2129 // fold (fp_to_uint c1fp) -> c1
2131 return DAG.getNode(ISD::FP_TO_UINT, VT, N0);
2135 SDOperand DAGCombiner::visitFP_ROUND(SDNode *N) {
2136 SDOperand N0 = N->getOperand(0);
2137 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2138 MVT::ValueType VT = N->getValueType(0);
2140 // fold (fp_round c1fp) -> c1fp
2142 return DAG.getNode(ISD::FP_ROUND, VT, N0);
2144 // fold (fp_round (fp_extend x)) -> x
2145 if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType())
2146 return N0.getOperand(0);
2148 // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y)
2149 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.Val->hasOneUse()) {
2150 SDOperand Tmp = DAG.getNode(ISD::FP_ROUND, VT, N0.getOperand(0));
2151 AddToWorkList(Tmp.Val);
2152 return DAG.getNode(ISD::FCOPYSIGN, VT, Tmp, N0.getOperand(1));
2158 SDOperand DAGCombiner::visitFP_ROUND_INREG(SDNode *N) {
2159 SDOperand N0 = N->getOperand(0);
2160 MVT::ValueType VT = N->getValueType(0);
2161 MVT::ValueType EVT = cast<VTSDNode>(N->getOperand(1))->getVT();
2162 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2164 // fold (fp_round_inreg c1fp) -> c1fp
2166 SDOperand Round = DAG.getConstantFP(N0CFP->getValue(), EVT);
2167 return DAG.getNode(ISD::FP_EXTEND, VT, Round);
2172 SDOperand DAGCombiner::visitFP_EXTEND(SDNode *N) {
2173 SDOperand N0 = N->getOperand(0);
2174 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2175 MVT::ValueType VT = N->getValueType(0);
2177 // fold (fp_extend c1fp) -> c1fp
2179 return DAG.getNode(ISD::FP_EXTEND, VT, N0);
2183 SDOperand DAGCombiner::visitFNEG(SDNode *N) {
2184 SDOperand N0 = N->getOperand(0);
2185 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2186 MVT::ValueType VT = N->getValueType(0);
2188 // fold (fneg c1) -> -c1
2190 return DAG.getNode(ISD::FNEG, VT, N0);
2191 // fold (fneg (sub x, y)) -> (sub y, x)
2192 if (N0.getOpcode() == ISD::SUB)
2193 return DAG.getNode(ISD::SUB, VT, N0.getOperand(1), N0.getOperand(0));
2194 // fold (fneg (fneg x)) -> x
2195 if (N0.getOpcode() == ISD::FNEG)
2196 return N0.getOperand(0);
2200 SDOperand DAGCombiner::visitFABS(SDNode *N) {
2201 SDOperand N0 = N->getOperand(0);
2202 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2203 MVT::ValueType VT = N->getValueType(0);
2205 // fold (fabs c1) -> fabs(c1)
2207 return DAG.getNode(ISD::FABS, VT, N0);
2208 // fold (fabs (fabs x)) -> (fabs x)
2209 if (N0.getOpcode() == ISD::FABS)
2210 return N->getOperand(0);
2211 // fold (fabs (fneg x)) -> (fabs x)
2212 // fold (fabs (fcopysign x, y)) -> (fabs x)
2213 if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN)
2214 return DAG.getNode(ISD::FABS, VT, N0.getOperand(0));
2219 SDOperand DAGCombiner::visitBRCOND(SDNode *N) {
2220 SDOperand Chain = N->getOperand(0);
2221 SDOperand N1 = N->getOperand(1);
2222 SDOperand N2 = N->getOperand(2);
2223 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2225 // never taken branch, fold to chain
2226 if (N1C && N1C->isNullValue())
2228 // unconditional branch
2229 if (N1C && N1C->getValue() == 1)
2230 return DAG.getNode(ISD::BR, MVT::Other, Chain, N2);
2231 // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal
2233 if (N1.getOpcode() == ISD::SETCC &&
2234 TLI.isOperationLegal(ISD::BR_CC, MVT::Other)) {
2235 return DAG.getNode(ISD::BR_CC, MVT::Other, Chain, N1.getOperand(2),
2236 N1.getOperand(0), N1.getOperand(1), N2);
2241 // Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB.
2243 SDOperand DAGCombiner::visitBR_CC(SDNode *N) {
2244 CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1));
2245 SDOperand CondLHS = N->getOperand(2), CondRHS = N->getOperand(3);
2247 // Use SimplifySetCC to simplify SETCC's.
2248 SDOperand Simp = SimplifySetCC(MVT::i1, CondLHS, CondRHS, CC->get(), false);
2249 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(Simp.Val);
2251 // fold br_cc true, dest -> br dest (unconditional branch)
2252 if (SCCC && SCCC->getValue())
2253 return DAG.getNode(ISD::BR, MVT::Other, N->getOperand(0),
2255 // fold br_cc false, dest -> unconditional fall through
2256 if (SCCC && SCCC->isNullValue())
2257 return N->getOperand(0);
2258 // fold to a simpler setcc
2259 if (Simp.Val && Simp.getOpcode() == ISD::SETCC)
2260 return DAG.getNode(ISD::BR_CC, MVT::Other, N->getOperand(0),
2261 Simp.getOperand(2), Simp.getOperand(0),
2262 Simp.getOperand(1), N->getOperand(4));
2266 SDOperand DAGCombiner::visitLOAD(SDNode *N) {
2267 SDOperand Chain = N->getOperand(0);
2268 SDOperand Ptr = N->getOperand(1);
2269 SDOperand SrcValue = N->getOperand(2);
2271 // If there are no uses of the loaded value, change uses of the chain value
2272 // into uses of the chain input (i.e. delete the dead load).
2273 if (N->hasNUsesOfValue(0, 0))
2274 return CombineTo(N, DAG.getNode(ISD::UNDEF, N->getValueType(0)), Chain);
2276 // If this load is directly stored, replace the load value with the stored
2278 // TODO: Handle store large -> read small portion.
2279 // TODO: Handle TRUNCSTORE/EXTLOAD
2280 if (Chain.getOpcode() == ISD::STORE && Chain.getOperand(2) == Ptr &&
2281 Chain.getOperand(1).getValueType() == N->getValueType(0))
2282 return CombineTo(N, Chain.getOperand(1), Chain);
2287 /// visitXEXTLOAD - Handle EXTLOAD/ZEXTLOAD/SEXTLOAD.
2288 SDOperand DAGCombiner::visitXEXTLOAD(SDNode *N) {
2289 SDOperand Chain = N->getOperand(0);
2290 SDOperand Ptr = N->getOperand(1);
2291 SDOperand SrcValue = N->getOperand(2);
2292 SDOperand EVT = N->getOperand(3);
2294 // If there are no uses of the loaded value, change uses of the chain value
2295 // into uses of the chain input (i.e. delete the dead load).
2296 if (N->hasNUsesOfValue(0, 0))
2297 return CombineTo(N, DAG.getNode(ISD::UNDEF, N->getValueType(0)), Chain);
2302 SDOperand DAGCombiner::visitSTORE(SDNode *N) {
2303 SDOperand Chain = N->getOperand(0);
2304 SDOperand Value = N->getOperand(1);
2305 SDOperand Ptr = N->getOperand(2);
2306 SDOperand SrcValue = N->getOperand(3);
2308 // If this is a store that kills a previous store, remove the previous store.
2309 if (Chain.getOpcode() == ISD::STORE && Chain.getOperand(2) == Ptr &&
2310 Chain.Val->hasOneUse() /* Avoid introducing DAG cycles */ &&
2311 // Make sure that these stores are the same value type:
2312 // FIXME: we really care that the second store is >= size of the first.
2313 Value.getValueType() == Chain.getOperand(1).getValueType()) {
2314 // Create a new store of Value that replaces both stores.
2315 SDNode *PrevStore = Chain.Val;
2316 if (PrevStore->getOperand(1) == Value) // Same value multiply stored.
2318 SDOperand NewStore = DAG.getNode(ISD::STORE, MVT::Other,
2319 PrevStore->getOperand(0), Value, Ptr,
2321 CombineTo(N, NewStore); // Nuke this store.
2322 CombineTo(PrevStore, NewStore); // Nuke the previous store.
2323 return SDOperand(N, 0);
2326 // If this is a store of a bit convert, store the input value.
2327 // FIXME: This needs to know that the resultant store does not need a
2328 // higher alignment than the original.
2329 if (0 && Value.getOpcode() == ISD::BIT_CONVERT)
2330 return DAG.getNode(ISD::STORE, MVT::Other, Chain, Value.getOperand(0),
2336 SDOperand DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) {
2337 SDOperand InVec = N->getOperand(0);
2338 SDOperand InVal = N->getOperand(1);
2339 SDOperand EltNo = N->getOperand(2);
2341 // If the invec is a BUILD_VECTOR and if EltNo is a constant, build a new
2342 // vector with the inserted element.
2343 if (InVec.getOpcode() == ISD::BUILD_VECTOR && isa<ConstantSDNode>(EltNo)) {
2344 unsigned Elt = cast<ConstantSDNode>(EltNo)->getValue();
2345 std::vector<SDOperand> Ops(InVec.Val->op_begin(), InVec.Val->op_end());
2346 if (Elt < Ops.size())
2348 return DAG.getNode(ISD::BUILD_VECTOR, InVec.getValueType(), Ops);
2354 SDOperand DAGCombiner::visitVINSERT_VECTOR_ELT(SDNode *N) {
2355 SDOperand InVec = N->getOperand(0);
2356 SDOperand InVal = N->getOperand(1);
2357 SDOperand EltNo = N->getOperand(2);
2358 SDOperand NumElts = N->getOperand(3);
2359 SDOperand EltType = N->getOperand(4);
2361 // If the invec is a VBUILD_VECTOR and if EltNo is a constant, build a new
2362 // vector with the inserted element.
2363 if (InVec.getOpcode() == ISD::VBUILD_VECTOR && isa<ConstantSDNode>(EltNo)) {
2364 unsigned Elt = cast<ConstantSDNode>(EltNo)->getValue();
2365 std::vector<SDOperand> Ops(InVec.Val->op_begin(), InVec.Val->op_end());
2366 if (Elt < Ops.size()-2)
2368 return DAG.getNode(ISD::VBUILD_VECTOR, InVec.getValueType(), Ops);
2374 SDOperand DAGCombiner::visitVBUILD_VECTOR(SDNode *N) {
2375 unsigned NumInScalars = N->getNumOperands()-2;
2376 SDOperand NumElts = N->getOperand(NumInScalars);
2377 SDOperand EltType = N->getOperand(NumInScalars+1);
2379 // Check to see if this is a VBUILD_VECTOR of a bunch of VEXTRACT_VECTOR_ELT
2380 // operations. If so, and if the EXTRACT_ELT vector inputs come from at most
2381 // two distinct vectors, turn this into a shuffle node.
2382 SDOperand VecIn1, VecIn2;
2383 for (unsigned i = 0; i != NumInScalars; ++i) {
2384 // Ignore undef inputs.
2385 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
2387 // If this input is something other than a VEXTRACT_VECTOR_ELT with a
2388 // constant index, bail out.
2389 if (N->getOperand(i).getOpcode() != ISD::VEXTRACT_VECTOR_ELT ||
2390 !isa<ConstantSDNode>(N->getOperand(i).getOperand(1))) {
2391 VecIn1 = VecIn2 = SDOperand(0, 0);
2395 // If the input vector type disagrees with the result of the vbuild_vector,
2396 // we can't make a shuffle.
2397 SDOperand ExtractedFromVec = N->getOperand(i).getOperand(0);
2398 if (*(ExtractedFromVec.Val->op_end()-2) != NumElts ||
2399 *(ExtractedFromVec.Val->op_end()-1) != EltType) {
2400 VecIn1 = VecIn2 = SDOperand(0, 0);
2404 // Otherwise, remember this. We allow up to two distinct input vectors.
2405 if (ExtractedFromVec == VecIn1 || ExtractedFromVec == VecIn2)
2408 if (VecIn1.Val == 0) {
2409 VecIn1 = ExtractedFromVec;
2410 } else if (VecIn2.Val == 0) {
2411 VecIn2 = ExtractedFromVec;
2414 VecIn1 = VecIn2 = SDOperand(0, 0);
2419 // If everything is good, we can make a shuffle operation.
2421 std::vector<SDOperand> BuildVecIndices;
2422 for (unsigned i = 0; i != NumInScalars; ++i) {
2423 if (N->getOperand(i).getOpcode() == ISD::UNDEF) {
2424 BuildVecIndices.push_back(DAG.getNode(ISD::UNDEF, MVT::i32));
2428 SDOperand Extract = N->getOperand(i);
2430 // If extracting from the first vector, just use the index directly.
2431 if (Extract.getOperand(0) == VecIn1) {
2432 BuildVecIndices.push_back(Extract.getOperand(1));
2436 // Otherwise, use InIdx + VecSize
2437 unsigned Idx = cast<ConstantSDNode>(Extract.getOperand(1))->getValue();
2438 BuildVecIndices.push_back(DAG.getConstant(Idx+NumInScalars, MVT::i32));
2441 // Add count and size info.
2442 BuildVecIndices.push_back(NumElts);
2443 BuildVecIndices.push_back(DAG.getValueType(MVT::i32));
2445 // Return the new VVECTOR_SHUFFLE node.
2446 std::vector<SDOperand> Ops;
2447 Ops.push_back(VecIn1);
2449 Ops.push_back(VecIn2);
2451 // Use an undef vbuild_vector as input for the second operand.
2452 std::vector<SDOperand> UnOps(NumInScalars,
2453 DAG.getNode(ISD::UNDEF,
2454 cast<VTSDNode>(EltType)->getVT()));
2455 UnOps.push_back(NumElts);
2456 UnOps.push_back(EltType);
2457 Ops.push_back(DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, UnOps));
2459 Ops.push_back(DAG.getNode(ISD::VBUILD_VECTOR,MVT::Vector, BuildVecIndices));
2460 Ops.push_back(NumElts);
2461 Ops.push_back(EltType);
2462 return DAG.getNode(ISD::VVECTOR_SHUFFLE, MVT::Vector, Ops);
2468 SDOperand DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
2469 SDOperand ShufMask = N->getOperand(2);
2470 unsigned NumElts = ShufMask.getNumOperands();
2472 // If the shuffle mask is an identity operation on the LHS, return the LHS.
2473 bool isIdentity = true;
2474 for (unsigned i = 0; i != NumElts; ++i) {
2475 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF &&
2476 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() != i) {
2481 if (isIdentity) return N->getOperand(0);
2483 // If the shuffle mask is an identity operation on the RHS, return the RHS.
2485 for (unsigned i = 0; i != NumElts; ++i) {
2486 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF &&
2487 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() != i+NumElts) {
2492 if (isIdentity) return N->getOperand(1);
2494 // If the LHS and the RHS are the same node, turn the RHS into an undef.
2495 if (N->getOperand(0) == N->getOperand(1)) {
2496 // Check the SHUFFLE mask, mapping any inputs from the 2nd operand into the
2498 std::vector<SDOperand> MappedOps;
2499 for (unsigned i = 0, e = ShufMask.getNumOperands(); i != e; ++i) {
2500 if (cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() >= NumElts) {
2502 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() - NumElts;
2503 MappedOps.push_back(DAG.getConstant(NewIdx, MVT::i32));
2505 MappedOps.push_back(ShufMask.getOperand(i));
2508 ShufMask = DAG.getNode(ISD::BUILD_VECTOR, ShufMask.getValueType(),
2510 return DAG.getNode(ISD::VECTOR_SHUFFLE, N->getValueType(0),
2512 DAG.getNode(ISD::UNDEF, N->getValueType(0)),
2519 SDOperand DAGCombiner::visitVVECTOR_SHUFFLE(SDNode *N) {
2520 SDOperand ShufMask = N->getOperand(2);
2521 unsigned NumElts = ShufMask.getNumOperands()-2;
2523 // If the shuffle mask is an identity operation on the LHS, return the LHS.
2524 bool isIdentity = true;
2525 for (unsigned i = 0; i != NumElts; ++i) {
2526 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF &&
2527 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() != i) {
2532 if (isIdentity) return N->getOperand(0);
2534 // If the shuffle mask is an identity operation on the RHS, return the RHS.
2536 for (unsigned i = 0; i != NumElts; ++i) {
2537 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF &&
2538 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() != i+NumElts) {
2543 if (isIdentity) return N->getOperand(1);
2548 SDOperand DAGCombiner::SimplifySelect(SDOperand N0, SDOperand N1, SDOperand N2){
2549 assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!");
2551 SDOperand SCC = SimplifySelectCC(N0.getOperand(0), N0.getOperand(1), N1, N2,
2552 cast<CondCodeSDNode>(N0.getOperand(2))->get());
2553 // If we got a simplified select_cc node back from SimplifySelectCC, then
2554 // break it down into a new SETCC node, and a new SELECT node, and then return
2555 // the SELECT node, since we were called with a SELECT node.
2557 // Check to see if we got a select_cc back (to turn into setcc/select).
2558 // Otherwise, just return whatever node we got back, like fabs.
2559 if (SCC.getOpcode() == ISD::SELECT_CC) {
2560 SDOperand SETCC = DAG.getNode(ISD::SETCC, N0.getValueType(),
2561 SCC.getOperand(0), SCC.getOperand(1),
2563 AddToWorkList(SETCC.Val);
2564 return DAG.getNode(ISD::SELECT, SCC.getValueType(), SCC.getOperand(2),
2565 SCC.getOperand(3), SETCC);
2572 /// SimplifySelectOps - Given a SELECT or a SELECT_CC node, where LHS and RHS
2573 /// are the two values being selected between, see if we can simplify the
2576 bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDOperand LHS,
2579 // If this is a select from two identical things, try to pull the operation
2580 // through the select.
2581 if (LHS.getOpcode() == RHS.getOpcode() && LHS.hasOneUse() && RHS.hasOneUse()){
2583 std::cerr << "SELECT: ["; LHS.Val->dump();
2584 std::cerr << "] ["; RHS.Val->dump();
2588 // If this is a load and the token chain is identical, replace the select
2589 // of two loads with a load through a select of the address to load from.
2590 // This triggers in things like "select bool X, 10.0, 123.0" after the FP
2591 // constants have been dropped into the constant pool.
2592 if ((LHS.getOpcode() == ISD::LOAD ||
2593 LHS.getOpcode() == ISD::EXTLOAD ||
2594 LHS.getOpcode() == ISD::ZEXTLOAD ||
2595 LHS.getOpcode() == ISD::SEXTLOAD) &&
2596 // Token chains must be identical.
2597 LHS.getOperand(0) == RHS.getOperand(0) &&
2598 // If this is an EXTLOAD, the VT's must match.
2599 (LHS.getOpcode() == ISD::LOAD ||
2600 LHS.getOperand(3) == RHS.getOperand(3))) {
2601 // FIXME: this conflates two src values, discarding one. This is not
2602 // the right thing to do, but nothing uses srcvalues now. When they do,
2603 // turn SrcValue into a list of locations.
2605 if (TheSelect->getOpcode() == ISD::SELECT)
2606 Addr = DAG.getNode(ISD::SELECT, LHS.getOperand(1).getValueType(),
2607 TheSelect->getOperand(0), LHS.getOperand(1),
2610 Addr = DAG.getNode(ISD::SELECT_CC, LHS.getOperand(1).getValueType(),
2611 TheSelect->getOperand(0),
2612 TheSelect->getOperand(1),
2613 LHS.getOperand(1), RHS.getOperand(1),
2614 TheSelect->getOperand(4));
2617 if (LHS.getOpcode() == ISD::LOAD)
2618 Load = DAG.getLoad(TheSelect->getValueType(0), LHS.getOperand(0),
2619 Addr, LHS.getOperand(2));
2621 Load = DAG.getExtLoad(LHS.getOpcode(), TheSelect->getValueType(0),
2622 LHS.getOperand(0), Addr, LHS.getOperand(2),
2623 cast<VTSDNode>(LHS.getOperand(3))->getVT());
2624 // Users of the select now use the result of the load.
2625 CombineTo(TheSelect, Load);
2627 // Users of the old loads now use the new load's chain. We know the
2628 // old-load value is dead now.
2629 CombineTo(LHS.Val, Load.getValue(0), Load.getValue(1));
2630 CombineTo(RHS.Val, Load.getValue(0), Load.getValue(1));
2638 SDOperand DAGCombiner::SimplifySelectCC(SDOperand N0, SDOperand N1,
2639 SDOperand N2, SDOperand N3,
2642 MVT::ValueType VT = N2.getValueType();
2643 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val);
2644 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val);
2645 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.Val);
2646 ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N3.Val);
2648 // Determine if the condition we're dealing with is constant
2649 SDOperand SCC = SimplifySetCC(TLI.getSetCCResultTy(), N0, N1, CC, false);
2650 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.Val);
2652 // fold select_cc true, x, y -> x
2653 if (SCCC && SCCC->getValue())
2655 // fold select_cc false, x, y -> y
2656 if (SCCC && SCCC->getValue() == 0)
2659 // Check to see if we can simplify the select into an fabs node
2660 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) {
2661 // Allow either -0.0 or 0.0
2662 if (CFP->getValue() == 0.0) {
2663 // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs
2664 if ((CC == ISD::SETGE || CC == ISD::SETGT) &&
2665 N0 == N2 && N3.getOpcode() == ISD::FNEG &&
2666 N2 == N3.getOperand(0))
2667 return DAG.getNode(ISD::FABS, VT, N0);
2669 // select (setl[te] X, +/-0.0), fneg(X), X -> fabs
2670 if ((CC == ISD::SETLT || CC == ISD::SETLE) &&
2671 N0 == N3 && N2.getOpcode() == ISD::FNEG &&
2672 N2.getOperand(0) == N3)
2673 return DAG.getNode(ISD::FABS, VT, N3);
2677 // Check to see if we can perform the "gzip trick", transforming
2678 // select_cc setlt X, 0, A, 0 -> and (sra X, size(X)-1), A
2679 if (N1C && N1C->isNullValue() && N3C && N3C->isNullValue() &&
2680 MVT::isInteger(N0.getValueType()) &&
2681 MVT::isInteger(N2.getValueType()) && CC == ISD::SETLT) {
2682 MVT::ValueType XType = N0.getValueType();
2683 MVT::ValueType AType = N2.getValueType();
2684 if (XType >= AType) {
2685 // and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a
2686 // single-bit constant.
2687 if (N2C && ((N2C->getValue() & (N2C->getValue()-1)) == 0)) {
2688 unsigned ShCtV = Log2_64(N2C->getValue());
2689 ShCtV = MVT::getSizeInBits(XType)-ShCtV-1;
2690 SDOperand ShCt = DAG.getConstant(ShCtV, TLI.getShiftAmountTy());
2691 SDOperand Shift = DAG.getNode(ISD::SRL, XType, N0, ShCt);
2692 AddToWorkList(Shift.Val);
2693 if (XType > AType) {
2694 Shift = DAG.getNode(ISD::TRUNCATE, AType, Shift);
2695 AddToWorkList(Shift.Val);
2697 return DAG.getNode(ISD::AND, AType, Shift, N2);
2699 SDOperand Shift = DAG.getNode(ISD::SRA, XType, N0,
2700 DAG.getConstant(MVT::getSizeInBits(XType)-1,
2701 TLI.getShiftAmountTy()));
2702 AddToWorkList(Shift.Val);
2703 if (XType > AType) {
2704 Shift = DAG.getNode(ISD::TRUNCATE, AType, Shift);
2705 AddToWorkList(Shift.Val);
2707 return DAG.getNode(ISD::AND, AType, Shift, N2);
2711 // fold select C, 16, 0 -> shl C, 4
2712 if (N2C && N3C && N3C->isNullValue() && isPowerOf2_64(N2C->getValue()) &&
2713 TLI.getSetCCResultContents() == TargetLowering::ZeroOrOneSetCCResult) {
2714 // Get a SetCC of the condition
2715 // FIXME: Should probably make sure that setcc is legal if we ever have a
2716 // target where it isn't.
2717 SDOperand Temp, SCC;
2718 // cast from setcc result type to select result type
2719 if (AfterLegalize) {
2720 SCC = DAG.getSetCC(TLI.getSetCCResultTy(), N0, N1, CC);
2721 Temp = DAG.getZeroExtendInReg(SCC, N2.getValueType());
2723 SCC = DAG.getSetCC(MVT::i1, N0, N1, CC);
2724 Temp = DAG.getNode(ISD::ZERO_EXTEND, N2.getValueType(), SCC);
2726 AddToWorkList(SCC.Val);
2727 AddToWorkList(Temp.Val);
2728 // shl setcc result by log2 n2c
2729 return DAG.getNode(ISD::SHL, N2.getValueType(), Temp,
2730 DAG.getConstant(Log2_64(N2C->getValue()),
2731 TLI.getShiftAmountTy()));
2734 // Check to see if this is the equivalent of setcc
2735 // FIXME: Turn all of these into setcc if setcc if setcc is legal
2736 // otherwise, go ahead with the folds.
2737 if (0 && N3C && N3C->isNullValue() && N2C && (N2C->getValue() == 1ULL)) {
2738 MVT::ValueType XType = N0.getValueType();
2739 if (TLI.isOperationLegal(ISD::SETCC, TLI.getSetCCResultTy())) {
2740 SDOperand Res = DAG.getSetCC(TLI.getSetCCResultTy(), N0, N1, CC);
2741 if (Res.getValueType() != VT)
2742 Res = DAG.getNode(ISD::ZERO_EXTEND, VT, Res);
2746 // seteq X, 0 -> srl (ctlz X, log2(size(X)))
2747 if (N1C && N1C->isNullValue() && CC == ISD::SETEQ &&
2748 TLI.isOperationLegal(ISD::CTLZ, XType)) {
2749 SDOperand Ctlz = DAG.getNode(ISD::CTLZ, XType, N0);
2750 return DAG.getNode(ISD::SRL, XType, Ctlz,
2751 DAG.getConstant(Log2_32(MVT::getSizeInBits(XType)),
2752 TLI.getShiftAmountTy()));
2754 // setgt X, 0 -> srl (and (-X, ~X), size(X)-1)
2755 if (N1C && N1C->isNullValue() && CC == ISD::SETGT) {
2756 SDOperand NegN0 = DAG.getNode(ISD::SUB, XType, DAG.getConstant(0, XType),
2758 SDOperand NotN0 = DAG.getNode(ISD::XOR, XType, N0,
2759 DAG.getConstant(~0ULL, XType));
2760 return DAG.getNode(ISD::SRL, XType,
2761 DAG.getNode(ISD::AND, XType, NegN0, NotN0),
2762 DAG.getConstant(MVT::getSizeInBits(XType)-1,
2763 TLI.getShiftAmountTy()));
2765 // setgt X, -1 -> xor (srl (X, size(X)-1), 1)
2766 if (N1C && N1C->isAllOnesValue() && CC == ISD::SETGT) {
2767 SDOperand Sign = DAG.getNode(ISD::SRL, XType, N0,
2768 DAG.getConstant(MVT::getSizeInBits(XType)-1,
2769 TLI.getShiftAmountTy()));
2770 return DAG.getNode(ISD::XOR, XType, Sign, DAG.getConstant(1, XType));
2774 // Check to see if this is an integer abs. select_cc setl[te] X, 0, -X, X ->
2775 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
2776 if (N1C && N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE) &&
2777 N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1)) {
2778 if (ConstantSDNode *SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0))) {
2779 MVT::ValueType XType = N0.getValueType();
2780 if (SubC->isNullValue() && MVT::isInteger(XType)) {
2781 SDOperand Shift = DAG.getNode(ISD::SRA, XType, N0,
2782 DAG.getConstant(MVT::getSizeInBits(XType)-1,
2783 TLI.getShiftAmountTy()));
2784 SDOperand Add = DAG.getNode(ISD::ADD, XType, N0, Shift);
2785 AddToWorkList(Shift.Val);
2786 AddToWorkList(Add.Val);
2787 return DAG.getNode(ISD::XOR, XType, Add, Shift);
2795 SDOperand DAGCombiner::SimplifySetCC(MVT::ValueType VT, SDOperand N0,
2796 SDOperand N1, ISD::CondCode Cond,
2797 bool foldBooleans) {
2798 // These setcc operations always fold.
2802 case ISD::SETFALSE2: return DAG.getConstant(0, VT);
2804 case ISD::SETTRUE2: return DAG.getConstant(1, VT);
2807 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val)) {
2808 uint64_t C1 = N1C->getValue();
2809 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val)) {
2810 uint64_t C0 = N0C->getValue();
2812 // Sign extend the operands if required
2813 if (ISD::isSignedIntSetCC(Cond)) {
2814 C0 = N0C->getSignExtended();
2815 C1 = N1C->getSignExtended();
2819 default: assert(0 && "Unknown integer setcc!");
2820 case ISD::SETEQ: return DAG.getConstant(C0 == C1, VT);
2821 case ISD::SETNE: return DAG.getConstant(C0 != C1, VT);
2822 case ISD::SETULT: return DAG.getConstant(C0 < C1, VT);
2823 case ISD::SETUGT: return DAG.getConstant(C0 > C1, VT);
2824 case ISD::SETULE: return DAG.getConstant(C0 <= C1, VT);
2825 case ISD::SETUGE: return DAG.getConstant(C0 >= C1, VT);
2826 case ISD::SETLT: return DAG.getConstant((int64_t)C0 < (int64_t)C1, VT);
2827 case ISD::SETGT: return DAG.getConstant((int64_t)C0 > (int64_t)C1, VT);
2828 case ISD::SETLE: return DAG.getConstant((int64_t)C0 <= (int64_t)C1, VT);
2829 case ISD::SETGE: return DAG.getConstant((int64_t)C0 >= (int64_t)C1, VT);
2832 // If the LHS is a ZERO_EXTEND, perform the comparison on the input.
2833 if (N0.getOpcode() == ISD::ZERO_EXTEND) {
2834 unsigned InSize = MVT::getSizeInBits(N0.getOperand(0).getValueType());
2836 // If the comparison constant has bits in the upper part, the
2837 // zero-extended value could never match.
2838 if (C1 & (~0ULL << InSize)) {
2839 unsigned VSize = MVT::getSizeInBits(N0.getValueType());
2843 case ISD::SETEQ: return DAG.getConstant(0, VT);
2846 case ISD::SETNE: return DAG.getConstant(1, VT);
2849 // True if the sign bit of C1 is set.
2850 return DAG.getConstant((C1 & (1ULL << VSize)) != 0, VT);
2853 // True if the sign bit of C1 isn't set.
2854 return DAG.getConstant((C1 & (1ULL << VSize)) == 0, VT);
2860 // Otherwise, we can perform the comparison with the low bits.
2868 return DAG.getSetCC(VT, N0.getOperand(0),
2869 DAG.getConstant(C1, N0.getOperand(0).getValueType()),
2872 break; // todo, be more careful with signed comparisons
2874 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
2875 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
2876 MVT::ValueType ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT();
2877 unsigned ExtSrcTyBits = MVT::getSizeInBits(ExtSrcTy);
2878 MVT::ValueType ExtDstTy = N0.getValueType();
2879 unsigned ExtDstTyBits = MVT::getSizeInBits(ExtDstTy);
2881 // If the extended part has any inconsistent bits, it cannot ever
2882 // compare equal. In other words, they have to be all ones or all
2885 (~0ULL >> (64-ExtSrcTyBits)) & (~0ULL << (ExtDstTyBits-1));
2886 if ((C1 & ExtBits) != 0 && (C1 & ExtBits) != ExtBits)
2887 return DAG.getConstant(Cond == ISD::SETNE, VT);
2890 MVT::ValueType Op0Ty = N0.getOperand(0).getValueType();
2891 if (Op0Ty == ExtSrcTy) {
2892 ZextOp = N0.getOperand(0);
2894 int64_t Imm = ~0ULL >> (64-ExtSrcTyBits);
2895 ZextOp = DAG.getNode(ISD::AND, Op0Ty, N0.getOperand(0),
2896 DAG.getConstant(Imm, Op0Ty));
2898 AddToWorkList(ZextOp.Val);
2899 // Otherwise, make this a use of a zext.
2900 return DAG.getSetCC(VT, ZextOp,
2901 DAG.getConstant(C1 & (~0ULL>>(64-ExtSrcTyBits)),
2904 } else if ((N1C->getValue() == 0 || N1C->getValue() == 1) &&
2905 (Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
2906 (N0.getOpcode() == ISD::XOR ||
2907 (N0.getOpcode() == ISD::AND &&
2908 N0.getOperand(0).getOpcode() == ISD::XOR &&
2909 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) &&
2910 isa<ConstantSDNode>(N0.getOperand(1)) &&
2911 cast<ConstantSDNode>(N0.getOperand(1))->getValue() == 1) {
2912 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We can
2913 // only do this if the top bits are known zero.
2914 if (TLI.MaskedValueIsZero(N1,
2915 MVT::getIntVTBitMask(N0.getValueType())-1)) {
2916 // Okay, get the un-inverted input value.
2918 if (N0.getOpcode() == ISD::XOR)
2919 Val = N0.getOperand(0);
2921 assert(N0.getOpcode() == ISD::AND &&
2922 N0.getOperand(0).getOpcode() == ISD::XOR);
2923 // ((X^1)&1)^1 -> X & 1
2924 Val = DAG.getNode(ISD::AND, N0.getValueType(),
2925 N0.getOperand(0).getOperand(0), N0.getOperand(1));
2927 return DAG.getSetCC(VT, Val, N1,
2928 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
2932 uint64_t MinVal, MaxVal;
2933 unsigned OperandBitSize = MVT::getSizeInBits(N1C->getValueType(0));
2934 if (ISD::isSignedIntSetCC(Cond)) {
2935 MinVal = 1ULL << (OperandBitSize-1);
2936 if (OperandBitSize != 1) // Avoid X >> 64, which is undefined.
2937 MaxVal = ~0ULL >> (65-OperandBitSize);
2942 MaxVal = ~0ULL >> (64-OperandBitSize);
2945 // Canonicalize GE/LE comparisons to use GT/LT comparisons.
2946 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) {
2947 if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true
2948 --C1; // X >= C0 --> X > (C0-1)
2949 return DAG.getSetCC(VT, N0, DAG.getConstant(C1, N1.getValueType()),
2950 (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT);
2953 if (Cond == ISD::SETLE || Cond == ISD::SETULE) {
2954 if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true
2955 ++C1; // X <= C0 --> X < (C0+1)
2956 return DAG.getSetCC(VT, N0, DAG.getConstant(C1, N1.getValueType()),
2957 (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT);
2960 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal)
2961 return DAG.getConstant(0, VT); // X < MIN --> false
2963 // Canonicalize setgt X, Min --> setne X, Min
2964 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal)
2965 return DAG.getSetCC(VT, N0, N1, ISD::SETNE);
2966 // Canonicalize setlt X, Max --> setne X, Max
2967 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal)
2968 return DAG.getSetCC(VT, N0, N1, ISD::SETNE);
2970 // If we have setult X, 1, turn it into seteq X, 0
2971 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1)
2972 return DAG.getSetCC(VT, N0, DAG.getConstant(MinVal, N0.getValueType()),
2974 // If we have setugt X, Max-1, turn it into seteq X, Max
2975 else if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1)
2976 return DAG.getSetCC(VT, N0, DAG.getConstant(MaxVal, N0.getValueType()),
2979 // If we have "setcc X, C0", check to see if we can shrink the immediate
2982 // SETUGT X, SINTMAX -> SETLT X, 0
2983 if (Cond == ISD::SETUGT && OperandBitSize != 1 &&
2984 C1 == (~0ULL >> (65-OperandBitSize)))
2985 return DAG.getSetCC(VT, N0, DAG.getConstant(0, N1.getValueType()),
2988 // FIXME: Implement the rest of these.
2990 // Fold bit comparisons when we can.
2991 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
2992 VT == N0.getValueType() && N0.getOpcode() == ISD::AND)
2993 if (ConstantSDNode *AndRHS =
2994 dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
2995 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3
2996 // Perform the xform if the AND RHS is a single bit.
2997 if ((AndRHS->getValue() & (AndRHS->getValue()-1)) == 0) {
2998 return DAG.getNode(ISD::SRL, VT, N0,
2999 DAG.getConstant(Log2_64(AndRHS->getValue()),
3000 TLI.getShiftAmountTy()));
3002 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getValue()) {
3003 // (X & 8) == 8 --> (X & 8) >> 3
3004 // Perform the xform if C1 is a single bit.
3005 if ((C1 & (C1-1)) == 0) {
3006 return DAG.getNode(ISD::SRL, VT, N0,
3007 DAG.getConstant(Log2_64(C1),TLI.getShiftAmountTy()));
3012 } else if (isa<ConstantSDNode>(N0.Val)) {
3013 // Ensure that the constant occurs on the RHS.
3014 return DAG.getSetCC(VT, N1, N0, ISD::getSetCCSwappedOperands(Cond));
3017 if (ConstantFPSDNode *N0C = dyn_cast<ConstantFPSDNode>(N0.Val))
3018 if (ConstantFPSDNode *N1C = dyn_cast<ConstantFPSDNode>(N1.Val)) {
3019 double C0 = N0C->getValue(), C1 = N1C->getValue();
3022 default: break; // FIXME: Implement the rest of these!
3023 case ISD::SETEQ: return DAG.getConstant(C0 == C1, VT);
3024 case ISD::SETNE: return DAG.getConstant(C0 != C1, VT);
3025 case ISD::SETLT: return DAG.getConstant(C0 < C1, VT);
3026 case ISD::SETGT: return DAG.getConstant(C0 > C1, VT);
3027 case ISD::SETLE: return DAG.getConstant(C0 <= C1, VT);
3028 case ISD::SETGE: return DAG.getConstant(C0 >= C1, VT);
3031 // Ensure that the constant occurs on the RHS.
3032 return DAG.getSetCC(VT, N1, N0, ISD::getSetCCSwappedOperands(Cond));
3036 // We can always fold X == Y for integer setcc's.
3037 if (MVT::isInteger(N0.getValueType()))
3038 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
3039 unsigned UOF = ISD::getUnorderedFlavor(Cond);
3040 if (UOF == 2) // FP operators that are undefined on NaNs.
3041 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
3042 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond)))
3043 return DAG.getConstant(UOF, VT);
3044 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO
3045 // if it is not already.
3046 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO;
3047 if (NewCond != Cond)
3048 return DAG.getSetCC(VT, N0, N1, NewCond);
3051 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
3052 MVT::isInteger(N0.getValueType())) {
3053 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB ||
3054 N0.getOpcode() == ISD::XOR) {
3055 // Simplify (X+Y) == (X+Z) --> Y == Z
3056 if (N0.getOpcode() == N1.getOpcode()) {
3057 if (N0.getOperand(0) == N1.getOperand(0))
3058 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(1), Cond);
3059 if (N0.getOperand(1) == N1.getOperand(1))
3060 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(0), Cond);
3061 if (isCommutativeBinOp(N0.getOpcode())) {
3062 // If X op Y == Y op X, try other combinations.
3063 if (N0.getOperand(0) == N1.getOperand(1))
3064 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(0), Cond);
3065 if (N0.getOperand(1) == N1.getOperand(0))
3066 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(1), Cond);
3070 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) {
3071 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
3072 // Turn (X+C1) == C2 --> X == C2-C1
3073 if (N0.getOpcode() == ISD::ADD && N0.Val->hasOneUse()) {
3074 return DAG.getSetCC(VT, N0.getOperand(0),
3075 DAG.getConstant(RHSC->getValue()-LHSR->getValue(),
3076 N0.getValueType()), Cond);
3079 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0.
3080 if (N0.getOpcode() == ISD::XOR)
3081 // If we know that all of the inverted bits are zero, don't bother
3082 // performing the inversion.
3083 if (TLI.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getValue()))
3084 return DAG.getSetCC(VT, N0.getOperand(0),
3085 DAG.getConstant(LHSR->getValue()^RHSC->getValue(),
3086 N0.getValueType()), Cond);
3089 // Turn (C1-X) == C2 --> X == C1-C2
3090 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) {
3091 if (N0.getOpcode() == ISD::SUB && N0.Val->hasOneUse()) {
3092 return DAG.getSetCC(VT, N0.getOperand(1),
3093 DAG.getConstant(SUBC->getValue()-RHSC->getValue(),
3094 N0.getValueType()), Cond);
3099 // Simplify (X+Z) == X --> Z == 0
3100 if (N0.getOperand(0) == N1)
3101 return DAG.getSetCC(VT, N0.getOperand(1),
3102 DAG.getConstant(0, N0.getValueType()), Cond);
3103 if (N0.getOperand(1) == N1) {
3104 if (isCommutativeBinOp(N0.getOpcode()))
3105 return DAG.getSetCC(VT, N0.getOperand(0),
3106 DAG.getConstant(0, N0.getValueType()), Cond);
3108 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!");
3109 // (Z-X) == X --> Z == X<<1
3110 SDOperand SH = DAG.getNode(ISD::SHL, N1.getValueType(),
3112 DAG.getConstant(1,TLI.getShiftAmountTy()));
3113 AddToWorkList(SH.Val);
3114 return DAG.getSetCC(VT, N0.getOperand(0), SH, Cond);
3119 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB ||
3120 N1.getOpcode() == ISD::XOR) {
3121 // Simplify X == (X+Z) --> Z == 0
3122 if (N1.getOperand(0) == N0) {
3123 return DAG.getSetCC(VT, N1.getOperand(1),
3124 DAG.getConstant(0, N1.getValueType()), Cond);
3125 } else if (N1.getOperand(1) == N0) {
3126 if (isCommutativeBinOp(N1.getOpcode())) {
3127 return DAG.getSetCC(VT, N1.getOperand(0),
3128 DAG.getConstant(0, N1.getValueType()), Cond);
3130 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!");
3131 // X == (Z-X) --> X<<1 == Z
3132 SDOperand SH = DAG.getNode(ISD::SHL, N1.getValueType(), N0,
3133 DAG.getConstant(1,TLI.getShiftAmountTy()));
3134 AddToWorkList(SH.Val);
3135 return DAG.getSetCC(VT, SH, N1.getOperand(0), Cond);
3141 // Fold away ALL boolean setcc's.
3143 if (N0.getValueType() == MVT::i1 && foldBooleans) {
3145 default: assert(0 && "Unknown integer setcc!");
3146 case ISD::SETEQ: // X == Y -> (X^Y)^1
3147 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, N1);
3148 N0 = DAG.getNode(ISD::XOR, MVT::i1, Temp, DAG.getConstant(1, MVT::i1));
3149 AddToWorkList(Temp.Val);
3151 case ISD::SETNE: // X != Y --> (X^Y)
3152 N0 = DAG.getNode(ISD::XOR, MVT::i1, N0, N1);
3154 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> X^1 & Y
3155 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> X^1 & Y
3156 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1));
3157 N0 = DAG.getNode(ISD::AND, MVT::i1, N1, Temp);
3158 AddToWorkList(Temp.Val);
3160 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> Y^1 & X
3161 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> Y^1 & X
3162 Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1));
3163 N0 = DAG.getNode(ISD::AND, MVT::i1, N0, Temp);
3164 AddToWorkList(Temp.Val);
3166 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> X^1 | Y
3167 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> X^1 | Y
3168 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1));
3169 N0 = DAG.getNode(ISD::OR, MVT::i1, N1, Temp);
3170 AddToWorkList(Temp.Val);
3172 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> Y^1 | X
3173 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> Y^1 | X
3174 Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1));
3175 N0 = DAG.getNode(ISD::OR, MVT::i1, N0, Temp);
3178 if (VT != MVT::i1) {
3179 AddToWorkList(N0.Val);
3180 // FIXME: If running after legalize, we probably can't do this.
3181 N0 = DAG.getNode(ISD::ZERO_EXTEND, VT, N0);
3186 // Could not fold it.
3190 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
3191 /// return a DAG expression to select that will generate the same value by
3192 /// multiplying by a magic number. See:
3193 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
3194 SDOperand DAGCombiner::BuildSDIV(SDNode *N) {
3195 MVT::ValueType VT = N->getValueType(0);
3197 // Check to see if we can do this.
3198 if (!TLI.isTypeLegal(VT) || (VT != MVT::i32 && VT != MVT::i64))
3199 return SDOperand(); // BuildSDIV only operates on i32 or i64
3200 if (!TLI.isOperationLegal(ISD::MULHS, VT))
3201 return SDOperand(); // Make sure the target supports MULHS.
3203 int64_t d = cast<ConstantSDNode>(N->getOperand(1))->getSignExtended();
3204 ms magics = (VT == MVT::i32) ? magic32(d) : magic64(d);
3206 // Multiply the numerator (operand 0) by the magic value
3207 SDOperand Q = DAG.getNode(ISD::MULHS, VT, N->getOperand(0),
3208 DAG.getConstant(magics.m, VT));
3209 // If d > 0 and m < 0, add the numerator
3210 if (d > 0 && magics.m < 0) {
3211 Q = DAG.getNode(ISD::ADD, VT, Q, N->getOperand(0));
3212 AddToWorkList(Q.Val);
3214 // If d < 0 and m > 0, subtract the numerator.
3215 if (d < 0 && magics.m > 0) {
3216 Q = DAG.getNode(ISD::SUB, VT, Q, N->getOperand(0));
3217 AddToWorkList(Q.Val);
3219 // Shift right algebraic if shift value is nonzero
3221 Q = DAG.getNode(ISD::SRA, VT, Q,
3222 DAG.getConstant(magics.s, TLI.getShiftAmountTy()));
3223 AddToWorkList(Q.Val);
3225 // Extract the sign bit and add it to the quotient
3227 DAG.getNode(ISD::SRL, VT, Q, DAG.getConstant(MVT::getSizeInBits(VT)-1,
3228 TLI.getShiftAmountTy()));
3229 AddToWorkList(T.Val);
3230 return DAG.getNode(ISD::ADD, VT, Q, T);
3233 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
3234 /// return a DAG expression to select that will generate the same value by
3235 /// multiplying by a magic number. See:
3236 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
3237 SDOperand DAGCombiner::BuildUDIV(SDNode *N) {
3238 MVT::ValueType VT = N->getValueType(0);
3240 // Check to see if we can do this.
3241 if (!TLI.isTypeLegal(VT) || (VT != MVT::i32 && VT != MVT::i64))
3242 return SDOperand(); // BuildUDIV only operates on i32 or i64
3243 if (!TLI.isOperationLegal(ISD::MULHU, VT))
3244 return SDOperand(); // Make sure the target supports MULHU.
3246 uint64_t d = cast<ConstantSDNode>(N->getOperand(1))->getValue();
3247 mu magics = (VT == MVT::i32) ? magicu32(d) : magicu64(d);
3249 // Multiply the numerator (operand 0) by the magic value
3250 SDOperand Q = DAG.getNode(ISD::MULHU, VT, N->getOperand(0),
3251 DAG.getConstant(magics.m, VT));
3252 AddToWorkList(Q.Val);
3254 if (magics.a == 0) {
3255 return DAG.getNode(ISD::SRL, VT, Q,
3256 DAG.getConstant(magics.s, TLI.getShiftAmountTy()));
3258 SDOperand NPQ = DAG.getNode(ISD::SUB, VT, N->getOperand(0), Q);
3259 AddToWorkList(NPQ.Val);
3260 NPQ = DAG.getNode(ISD::SRL, VT, NPQ,
3261 DAG.getConstant(1, TLI.getShiftAmountTy()));
3262 AddToWorkList(NPQ.Val);
3263 NPQ = DAG.getNode(ISD::ADD, VT, NPQ, Q);
3264 AddToWorkList(NPQ.Val);
3265 return DAG.getNode(ISD::SRL, VT, NPQ,
3266 DAG.getConstant(magics.s-1, TLI.getShiftAmountTy()));
3270 // SelectionDAG::Combine - This is the entry point for the file.
3272 void SelectionDAG::Combine(bool RunningAfterLegalize) {
3273 /// run - This is the main entry point to this class.
3275 DAGCombiner(*this).Run(RunningAfterLegalize);