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 visitSTORE(SDNode *N);
212 SDOperand visitINSERT_VECTOR_ELT(SDNode *N);
213 SDOperand visitVINSERT_VECTOR_ELT(SDNode *N);
214 SDOperand visitVBUILD_VECTOR(SDNode *N);
215 SDOperand visitVECTOR_SHUFFLE(SDNode *N);
217 SDOperand ReassociateOps(unsigned Opc, SDOperand LHS, SDOperand RHS);
219 bool SimplifySelectOps(SDNode *SELECT, SDOperand LHS, SDOperand RHS);
220 SDOperand SimplifySelect(SDOperand N0, SDOperand N1, SDOperand N2);
221 SDOperand SimplifySelectCC(SDOperand N0, SDOperand N1, SDOperand N2,
222 SDOperand N3, ISD::CondCode CC);
223 SDOperand SimplifySetCC(MVT::ValueType VT, SDOperand N0, SDOperand N1,
224 ISD::CondCode Cond, bool foldBooleans = true);
226 SDOperand BuildSDIV(SDNode *N);
227 SDOperand BuildUDIV(SDNode *N);
229 DAGCombiner(SelectionDAG &D)
230 : DAG(D), TLI(D.getTargetLoweringInfo()), AfterLegalize(false) {}
232 /// Run - runs the dag combiner on all nodes in the work list
233 void Run(bool RunningAfterLegalize);
237 //===----------------------------------------------------------------------===//
238 // TargetLowering::DAGCombinerInfo implementation
239 //===----------------------------------------------------------------------===//
241 void TargetLowering::DAGCombinerInfo::AddToWorklist(SDNode *N) {
242 ((DAGCombiner*)DC)->AddToWorkList(N);
245 SDOperand TargetLowering::DAGCombinerInfo::
246 CombineTo(SDNode *N, const std::vector<SDOperand> &To) {
247 return ((DAGCombiner*)DC)->CombineTo(N, To);
250 SDOperand TargetLowering::DAGCombinerInfo::
251 CombineTo(SDNode *N, SDOperand Res) {
252 return ((DAGCombiner*)DC)->CombineTo(N, Res);
256 SDOperand TargetLowering::DAGCombinerInfo::
257 CombineTo(SDNode *N, SDOperand Res0, SDOperand Res1) {
258 return ((DAGCombiner*)DC)->CombineTo(N, Res0, Res1);
264 //===----------------------------------------------------------------------===//
268 int64_t m; // magic number
269 int64_t s; // shift amount
273 uint64_t m; // magic number
274 int64_t a; // add indicator
275 int64_t s; // shift amount
278 /// magic - calculate the magic numbers required to codegen an integer sdiv as
279 /// a sequence of multiply and shifts. Requires that the divisor not be 0, 1,
281 static ms magic32(int32_t d) {
283 uint32_t ad, anc, delta, q1, r1, q2, r2, t;
284 const uint32_t two31 = 0x80000000U;
288 t = two31 + ((uint32_t)d >> 31);
289 anc = t - 1 - t%ad; // absolute value of nc
290 p = 31; // initialize p
291 q1 = two31/anc; // initialize q1 = 2p/abs(nc)
292 r1 = two31 - q1*anc; // initialize r1 = rem(2p,abs(nc))
293 q2 = two31/ad; // initialize q2 = 2p/abs(d)
294 r2 = two31 - q2*ad; // initialize r2 = rem(2p,abs(d))
297 q1 = 2*q1; // update q1 = 2p/abs(nc)
298 r1 = 2*r1; // update r1 = rem(2p/abs(nc))
299 if (r1 >= anc) { // must be unsigned comparison
303 q2 = 2*q2; // update q2 = 2p/abs(d)
304 r2 = 2*r2; // update r2 = rem(2p/abs(d))
305 if (r2 >= ad) { // must be unsigned comparison
310 } while (q1 < delta || (q1 == delta && r1 == 0));
312 mag.m = (int32_t)(q2 + 1); // make sure to sign extend
313 if (d < 0) mag.m = -mag.m; // resulting magic number
314 mag.s = p - 32; // resulting shift
318 /// magicu - calculate the magic numbers required to codegen an integer udiv as
319 /// a sequence of multiply, add and shifts. Requires that the divisor not be 0.
320 static mu magicu32(uint32_t d) {
322 uint32_t nc, delta, q1, r1, q2, r2;
324 magu.a = 0; // initialize "add" indicator
326 p = 31; // initialize p
327 q1 = 0x80000000/nc; // initialize q1 = 2p/nc
328 r1 = 0x80000000 - q1*nc; // initialize r1 = rem(2p,nc)
329 q2 = 0x7FFFFFFF/d; // initialize q2 = (2p-1)/d
330 r2 = 0x7FFFFFFF - q2*d; // initialize r2 = rem((2p-1),d)
333 if (r1 >= nc - r1 ) {
334 q1 = 2*q1 + 1; // update q1
335 r1 = 2*r1 - nc; // update r1
338 q1 = 2*q1; // update q1
339 r1 = 2*r1; // update r1
341 if (r2 + 1 >= d - r2) {
342 if (q2 >= 0x7FFFFFFF) magu.a = 1;
343 q2 = 2*q2 + 1; // update q2
344 r2 = 2*r2 + 1 - d; // update r2
347 if (q2 >= 0x80000000) magu.a = 1;
348 q2 = 2*q2; // update q2
349 r2 = 2*r2 + 1; // update r2
352 } while (p < 64 && (q1 < delta || (q1 == delta && r1 == 0)));
353 magu.m = q2 + 1; // resulting magic number
354 magu.s = p - 32; // resulting shift
358 /// magic - calculate the magic numbers required to codegen an integer sdiv as
359 /// a sequence of multiply and shifts. Requires that the divisor not be 0, 1,
361 static ms magic64(int64_t d) {
363 uint64_t ad, anc, delta, q1, r1, q2, r2, t;
364 const uint64_t two63 = 9223372036854775808ULL; // 2^63
367 ad = d >= 0 ? d : -d;
368 t = two63 + ((uint64_t)d >> 63);
369 anc = t - 1 - t%ad; // absolute value of nc
370 p = 63; // initialize p
371 q1 = two63/anc; // initialize q1 = 2p/abs(nc)
372 r1 = two63 - q1*anc; // initialize r1 = rem(2p,abs(nc))
373 q2 = two63/ad; // initialize q2 = 2p/abs(d)
374 r2 = two63 - q2*ad; // initialize r2 = rem(2p,abs(d))
377 q1 = 2*q1; // update q1 = 2p/abs(nc)
378 r1 = 2*r1; // update r1 = rem(2p/abs(nc))
379 if (r1 >= anc) { // must be unsigned comparison
383 q2 = 2*q2; // update q2 = 2p/abs(d)
384 r2 = 2*r2; // update r2 = rem(2p/abs(d))
385 if (r2 >= ad) { // must be unsigned comparison
390 } while (q1 < delta || (q1 == delta && r1 == 0));
393 if (d < 0) mag.m = -mag.m; // resulting magic number
394 mag.s = p - 64; // resulting shift
398 /// magicu - calculate the magic numbers required to codegen an integer udiv as
399 /// a sequence of multiply, add and shifts. Requires that the divisor not be 0.
400 static mu magicu64(uint64_t d)
403 uint64_t nc, delta, q1, r1, q2, r2;
405 magu.a = 0; // initialize "add" indicator
407 p = 63; // initialize p
408 q1 = 0x8000000000000000ull/nc; // initialize q1 = 2p/nc
409 r1 = 0x8000000000000000ull - q1*nc; // initialize r1 = rem(2p,nc)
410 q2 = 0x7FFFFFFFFFFFFFFFull/d; // initialize q2 = (2p-1)/d
411 r2 = 0x7FFFFFFFFFFFFFFFull - q2*d; // initialize r2 = rem((2p-1),d)
414 if (r1 >= nc - r1 ) {
415 q1 = 2*q1 + 1; // update q1
416 r1 = 2*r1 - nc; // update r1
419 q1 = 2*q1; // update q1
420 r1 = 2*r1; // update r1
422 if (r2 + 1 >= d - r2) {
423 if (q2 >= 0x7FFFFFFFFFFFFFFFull) magu.a = 1;
424 q2 = 2*q2 + 1; // update q2
425 r2 = 2*r2 + 1 - d; // update r2
428 if (q2 >= 0x8000000000000000ull) magu.a = 1;
429 q2 = 2*q2; // update q2
430 r2 = 2*r2 + 1; // update r2
433 } while (p < 64 && (q1 < delta || (q1 == delta && r1 == 0)));
434 magu.m = q2 + 1; // resulting magic number
435 magu.s = p - 64; // resulting shift
439 // isSetCCEquivalent - Return true if this node is a setcc, or is a select_cc
440 // that selects between the values 1 and 0, making it equivalent to a setcc.
441 // Also, set the incoming LHS, RHS, and CC references to the appropriate
442 // nodes based on the type of node we are checking. This simplifies life a
443 // bit for the callers.
444 static bool isSetCCEquivalent(SDOperand N, SDOperand &LHS, SDOperand &RHS,
446 if (N.getOpcode() == ISD::SETCC) {
447 LHS = N.getOperand(0);
448 RHS = N.getOperand(1);
449 CC = N.getOperand(2);
452 if (N.getOpcode() == ISD::SELECT_CC &&
453 N.getOperand(2).getOpcode() == ISD::Constant &&
454 N.getOperand(3).getOpcode() == ISD::Constant &&
455 cast<ConstantSDNode>(N.getOperand(2))->getValue() == 1 &&
456 cast<ConstantSDNode>(N.getOperand(3))->isNullValue()) {
457 LHS = N.getOperand(0);
458 RHS = N.getOperand(1);
459 CC = N.getOperand(4);
465 // isOneUseSetCC - Return true if this is a SetCC-equivalent operation with only
466 // one use. If this is true, it allows the users to invert the operation for
467 // free when it is profitable to do so.
468 static bool isOneUseSetCC(SDOperand N) {
469 SDOperand N0, N1, N2;
470 if (isSetCCEquivalent(N, N0, N1, N2) && N.Val->hasOneUse())
475 // FIXME: This should probably go in the ISD class rather than being duplicated
477 static bool isCommutativeBinOp(unsigned Opcode) {
483 case ISD::XOR: return true;
484 default: return false; // FIXME: Need commutative info for user ops!
488 SDOperand DAGCombiner::ReassociateOps(unsigned Opc, SDOperand N0, SDOperand N1){
489 MVT::ValueType VT = N0.getValueType();
490 // reassoc. (op (op x, c1), y) -> (op (op x, y), c1) iff x+c1 has one use
491 // reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2))
492 if (N0.getOpcode() == Opc && isa<ConstantSDNode>(N0.getOperand(1))) {
493 if (isa<ConstantSDNode>(N1)) {
494 SDOperand OpNode = DAG.getNode(Opc, VT, N0.getOperand(1), N1);
495 AddToWorkList(OpNode.Val);
496 return DAG.getNode(Opc, VT, OpNode, N0.getOperand(0));
497 } else if (N0.hasOneUse()) {
498 SDOperand OpNode = DAG.getNode(Opc, VT, N0.getOperand(0), N1);
499 AddToWorkList(OpNode.Val);
500 return DAG.getNode(Opc, VT, OpNode, N0.getOperand(1));
503 // reassoc. (op y, (op x, c1)) -> (op (op x, y), c1) iff x+c1 has one use
504 // reassoc. (op c2, (op x, c1)) -> (op x, (op c1, c2))
505 if (N1.getOpcode() == Opc && isa<ConstantSDNode>(N1.getOperand(1))) {
506 if (isa<ConstantSDNode>(N0)) {
507 SDOperand OpNode = DAG.getNode(Opc, VT, N1.getOperand(1), N0);
508 AddToWorkList(OpNode.Val);
509 return DAG.getNode(Opc, VT, OpNode, N1.getOperand(0));
510 } else if (N1.hasOneUse()) {
511 SDOperand OpNode = DAG.getNode(Opc, VT, N1.getOperand(0), N0);
512 AddToWorkList(OpNode.Val);
513 return DAG.getNode(Opc, VT, OpNode, N1.getOperand(1));
519 void DAGCombiner::Run(bool RunningAfterLegalize) {
520 // set the instance variable, so that the various visit routines may use it.
521 AfterLegalize = RunningAfterLegalize;
523 // Add all the dag nodes to the worklist.
524 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
525 E = DAG.allnodes_end(); I != E; ++I)
526 WorkList.push_back(I);
528 // Create a dummy node (which is not added to allnodes), that adds a reference
529 // to the root node, preventing it from being deleted, and tracking any
530 // changes of the root.
531 HandleSDNode Dummy(DAG.getRoot());
534 /// DagCombineInfo - Expose the DAG combiner to the target combiner impls.
535 TargetLowering::DAGCombinerInfo
536 DagCombineInfo(DAG, !RunningAfterLegalize, this);
538 // while the worklist isn't empty, inspect the node on the end of it and
539 // try and combine it.
540 while (!WorkList.empty()) {
541 SDNode *N = WorkList.back();
544 // If N has no uses, it is dead. Make sure to revisit all N's operands once
545 // N is deleted from the DAG, since they too may now be dead or may have a
546 // reduced number of uses, allowing other xforms.
547 if (N->use_empty() && N != &Dummy) {
548 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
549 WorkList.push_back(N->getOperand(i).Val);
551 removeFromWorkList(N);
556 SDOperand RV = visit(N);
558 // If nothing happened, try a target-specific DAG combine.
560 if (N->getOpcode() >= ISD::BUILTIN_OP_END ||
561 TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode()))
562 RV = TLI.PerformDAGCombine(N, DagCombineInfo);
567 // If we get back the same node we passed in, rather than a new node or
568 // zero, we know that the node must have defined multiple values and
569 // CombineTo was used. Since CombineTo takes care of the worklist
570 // mechanics for us, we have no work to do in this case.
572 DEBUG(std::cerr << "\nReplacing "; N->dump();
573 std::cerr << "\nWith: "; RV.Val->dump();
575 std::vector<SDNode*> NowDead;
576 DAG.ReplaceAllUsesWith(N, std::vector<SDOperand>(1, RV), &NowDead);
578 // Push the new node and any users onto the worklist
579 WorkList.push_back(RV.Val);
580 AddUsersToWorkList(RV.Val);
582 // Nodes can end up on the worklist more than once. Make sure we do
583 // not process a node that has been replaced.
584 removeFromWorkList(N);
585 for (unsigned i = 0, e = NowDead.size(); i != e; ++i)
586 removeFromWorkList(NowDead[i]);
588 // Finally, since the node is now dead, remove it from the graph.
594 // If the root changed (e.g. it was a dead load, update the root).
595 DAG.setRoot(Dummy.getValue());
598 SDOperand DAGCombiner::visit(SDNode *N) {
599 switch(N->getOpcode()) {
601 case ISD::TokenFactor: return visitTokenFactor(N);
602 case ISD::ADD: return visitADD(N);
603 case ISD::SUB: return visitSUB(N);
604 case ISD::MUL: return visitMUL(N);
605 case ISD::SDIV: return visitSDIV(N);
606 case ISD::UDIV: return visitUDIV(N);
607 case ISD::SREM: return visitSREM(N);
608 case ISD::UREM: return visitUREM(N);
609 case ISD::MULHU: return visitMULHU(N);
610 case ISD::MULHS: return visitMULHS(N);
611 case ISD::AND: return visitAND(N);
612 case ISD::OR: return visitOR(N);
613 case ISD::XOR: return visitXOR(N);
614 case ISD::SHL: return visitSHL(N);
615 case ISD::SRA: return visitSRA(N);
616 case ISD::SRL: return visitSRL(N);
617 case ISD::CTLZ: return visitCTLZ(N);
618 case ISD::CTTZ: return visitCTTZ(N);
619 case ISD::CTPOP: return visitCTPOP(N);
620 case ISD::SELECT: return visitSELECT(N);
621 case ISD::SELECT_CC: return visitSELECT_CC(N);
622 case ISD::SETCC: return visitSETCC(N);
623 case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N);
624 case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N);
625 case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N);
626 case ISD::TRUNCATE: return visitTRUNCATE(N);
627 case ISD::BIT_CONVERT: return visitBIT_CONVERT(N);
628 case ISD::FADD: return visitFADD(N);
629 case ISD::FSUB: return visitFSUB(N);
630 case ISD::FMUL: return visitFMUL(N);
631 case ISD::FDIV: return visitFDIV(N);
632 case ISD::FREM: return visitFREM(N);
633 case ISD::FCOPYSIGN: return visitFCOPYSIGN(N);
634 case ISD::SINT_TO_FP: return visitSINT_TO_FP(N);
635 case ISD::UINT_TO_FP: return visitUINT_TO_FP(N);
636 case ISD::FP_TO_SINT: return visitFP_TO_SINT(N);
637 case ISD::FP_TO_UINT: return visitFP_TO_UINT(N);
638 case ISD::FP_ROUND: return visitFP_ROUND(N);
639 case ISD::FP_ROUND_INREG: return visitFP_ROUND_INREG(N);
640 case ISD::FP_EXTEND: return visitFP_EXTEND(N);
641 case ISD::FNEG: return visitFNEG(N);
642 case ISD::FABS: return visitFABS(N);
643 case ISD::BRCOND: return visitBRCOND(N);
644 case ISD::BR_CC: return visitBR_CC(N);
645 case ISD::LOAD: return visitLOAD(N);
646 case ISD::STORE: return visitSTORE(N);
647 case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N);
648 case ISD::VINSERT_VECTOR_ELT: return visitVINSERT_VECTOR_ELT(N);
649 case ISD::VBUILD_VECTOR: return visitVBUILD_VECTOR(N);
650 case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N);
655 SDOperand DAGCombiner::visitTokenFactor(SDNode *N) {
656 std::vector<SDOperand> Ops;
657 bool Changed = false;
659 // If the token factor has two operands and one is the entry token, replace
660 // the token factor with the other operand.
661 if (N->getNumOperands() == 2) {
662 if (N->getOperand(0).getOpcode() == ISD::EntryToken)
663 return N->getOperand(1);
664 if (N->getOperand(1).getOpcode() == ISD::EntryToken)
665 return N->getOperand(0);
668 // fold (tokenfactor (tokenfactor)) -> tokenfactor
669 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
670 SDOperand Op = N->getOperand(i);
671 if (Op.getOpcode() == ISD::TokenFactor && Op.hasOneUse()) {
672 AddToWorkList(Op.Val); // Remove dead node.
674 for (unsigned j = 0, e = Op.getNumOperands(); j != e; ++j)
675 Ops.push_back(Op.getOperand(j));
681 return DAG.getNode(ISD::TokenFactor, MVT::Other, Ops);
685 SDOperand DAGCombiner::visitADD(SDNode *N) {
686 SDOperand N0 = N->getOperand(0);
687 SDOperand N1 = N->getOperand(1);
688 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
689 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
690 MVT::ValueType VT = N0.getValueType();
692 // fold (add c1, c2) -> c1+c2
694 return DAG.getNode(ISD::ADD, VT, N0, N1);
695 // canonicalize constant to RHS
697 return DAG.getNode(ISD::ADD, VT, N1, N0);
698 // fold (add x, 0) -> x
699 if (N1C && N1C->isNullValue())
701 // fold ((c1-A)+c2) -> (c1+c2)-A
702 if (N1C && N0.getOpcode() == ISD::SUB)
703 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getOperand(0)))
704 return DAG.getNode(ISD::SUB, VT,
705 DAG.getConstant(N1C->getValue()+N0C->getValue(), VT),
708 SDOperand RADD = ReassociateOps(ISD::ADD, N0, N1);
711 // fold ((0-A) + B) -> B-A
712 if (N0.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N0.getOperand(0)) &&
713 cast<ConstantSDNode>(N0.getOperand(0))->isNullValue())
714 return DAG.getNode(ISD::SUB, VT, N1, N0.getOperand(1));
715 // fold (A + (0-B)) -> A-B
716 if (N1.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N1.getOperand(0)) &&
717 cast<ConstantSDNode>(N1.getOperand(0))->isNullValue())
718 return DAG.getNode(ISD::SUB, VT, N0, N1.getOperand(1));
719 // fold (A+(B-A)) -> B
720 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1))
721 return N1.getOperand(0);
723 if (!MVT::isVector(VT) && SimplifyDemandedBits(SDOperand(N, 0)))
726 // fold (a+b) -> (a|b) iff a and b share no bits.
727 if (MVT::isInteger(VT) && !MVT::isVector(VT)) {
728 uint64_t LHSZero, LHSOne;
729 uint64_t RHSZero, RHSOne;
730 uint64_t Mask = MVT::getIntVTBitMask(VT);
731 TLI.ComputeMaskedBits(N0, Mask, LHSZero, LHSOne);
733 TLI.ComputeMaskedBits(N1, Mask, RHSZero, RHSOne);
735 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
736 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
737 if ((RHSZero & (~LHSZero & Mask)) == (~LHSZero & Mask) ||
738 (LHSZero & (~RHSZero & Mask)) == (~RHSZero & Mask))
739 return DAG.getNode(ISD::OR, VT, N0, N1);
746 SDOperand DAGCombiner::visitSUB(SDNode *N) {
747 SDOperand N0 = N->getOperand(0);
748 SDOperand N1 = N->getOperand(1);
749 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val);
750 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val);
751 MVT::ValueType VT = N0.getValueType();
753 // fold (sub x, x) -> 0
755 return DAG.getConstant(0, N->getValueType(0));
756 // fold (sub c1, c2) -> c1-c2
758 return DAG.getNode(ISD::SUB, VT, N0, N1);
759 // fold (sub x, c) -> (add x, -c)
761 return DAG.getNode(ISD::ADD, VT, N0, DAG.getConstant(-N1C->getValue(), VT));
763 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1)
764 return N0.getOperand(1);
766 if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1)
767 return N0.getOperand(0);
771 SDOperand DAGCombiner::visitMUL(SDNode *N) {
772 SDOperand N0 = N->getOperand(0);
773 SDOperand N1 = N->getOperand(1);
774 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
775 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
776 MVT::ValueType VT = N0.getValueType();
778 // fold (mul c1, c2) -> c1*c2
780 return DAG.getNode(ISD::MUL, VT, N0, N1);
781 // canonicalize constant to RHS
783 return DAG.getNode(ISD::MUL, VT, N1, N0);
784 // fold (mul x, 0) -> 0
785 if (N1C && N1C->isNullValue())
787 // fold (mul x, -1) -> 0-x
788 if (N1C && N1C->isAllOnesValue())
789 return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT), N0);
790 // fold (mul x, (1 << c)) -> x << c
791 if (N1C && isPowerOf2_64(N1C->getValue()))
792 return DAG.getNode(ISD::SHL, VT, N0,
793 DAG.getConstant(Log2_64(N1C->getValue()),
794 TLI.getShiftAmountTy()));
795 // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c
796 if (N1C && isPowerOf2_64(-N1C->getSignExtended())) {
797 // FIXME: If the input is something that is easily negated (e.g. a
798 // single-use add), we should put the negate there.
799 return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT),
800 DAG.getNode(ISD::SHL, VT, N0,
801 DAG.getConstant(Log2_64(-N1C->getSignExtended()),
802 TLI.getShiftAmountTy())));
805 // (mul (shl X, c1), c2) -> (mul X, c2 << c1)
806 if (N1C && N0.getOpcode() == ISD::SHL &&
807 isa<ConstantSDNode>(N0.getOperand(1))) {
808 SDOperand C3 = DAG.getNode(ISD::SHL, VT, N1, N0.getOperand(1));
809 AddToWorkList(C3.Val);
810 return DAG.getNode(ISD::MUL, VT, N0.getOperand(0), C3);
813 // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one
816 SDOperand Sh(0,0), Y(0,0);
817 // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)).
818 if (N0.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N0.getOperand(1)) &&
819 N0.Val->hasOneUse()) {
821 } else if (N1.getOpcode() == ISD::SHL &&
822 isa<ConstantSDNode>(N1.getOperand(1)) && N1.Val->hasOneUse()) {
826 SDOperand Mul = DAG.getNode(ISD::MUL, VT, Sh.getOperand(0), Y);
827 return DAG.getNode(ISD::SHL, VT, Mul, Sh.getOperand(1));
830 // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2)
831 if (N1C && N0.getOpcode() == ISD::ADD && N0.Val->hasOneUse() &&
832 isa<ConstantSDNode>(N0.getOperand(1))) {
833 return DAG.getNode(ISD::ADD, VT,
834 DAG.getNode(ISD::MUL, VT, N0.getOperand(0), N1),
835 DAG.getNode(ISD::MUL, VT, N0.getOperand(1), N1));
839 SDOperand RMUL = ReassociateOps(ISD::MUL, N0, N1);
845 SDOperand DAGCombiner::visitSDIV(SDNode *N) {
846 SDOperand N0 = N->getOperand(0);
847 SDOperand N1 = N->getOperand(1);
848 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val);
849 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val);
850 MVT::ValueType VT = N->getValueType(0);
852 // fold (sdiv c1, c2) -> c1/c2
853 if (N0C && N1C && !N1C->isNullValue())
854 return DAG.getNode(ISD::SDIV, VT, N0, N1);
855 // fold (sdiv X, 1) -> X
856 if (N1C && N1C->getSignExtended() == 1LL)
858 // fold (sdiv X, -1) -> 0-X
859 if (N1C && N1C->isAllOnesValue())
860 return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT), N0);
861 // If we know the sign bits of both operands are zero, strength reduce to a
862 // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2
863 uint64_t SignBit = 1ULL << (MVT::getSizeInBits(VT)-1);
864 if (TLI.MaskedValueIsZero(N1, SignBit) &&
865 TLI.MaskedValueIsZero(N0, SignBit))
866 return DAG.getNode(ISD::UDIV, N1.getValueType(), N0, N1);
867 // fold (sdiv X, pow2) -> simple ops after legalize
868 if (N1C && N1C->getValue() && !TLI.isIntDivCheap() &&
869 (isPowerOf2_64(N1C->getSignExtended()) ||
870 isPowerOf2_64(-N1C->getSignExtended()))) {
871 // If dividing by powers of two is cheap, then don't perform the following
873 if (TLI.isPow2DivCheap())
875 int64_t pow2 = N1C->getSignExtended();
876 int64_t abs2 = pow2 > 0 ? pow2 : -pow2;
877 unsigned lg2 = Log2_64(abs2);
878 // Splat the sign bit into the register
879 SDOperand SGN = DAG.getNode(ISD::SRA, VT, N0,
880 DAG.getConstant(MVT::getSizeInBits(VT)-1,
881 TLI.getShiftAmountTy()));
882 AddToWorkList(SGN.Val);
883 // Add (N0 < 0) ? abs2 - 1 : 0;
884 SDOperand SRL = DAG.getNode(ISD::SRL, VT, SGN,
885 DAG.getConstant(MVT::getSizeInBits(VT)-lg2,
886 TLI.getShiftAmountTy()));
887 SDOperand ADD = DAG.getNode(ISD::ADD, VT, N0, SRL);
888 AddToWorkList(SRL.Val);
889 AddToWorkList(ADD.Val); // Divide by pow2
890 SDOperand SRA = DAG.getNode(ISD::SRA, VT, ADD,
891 DAG.getConstant(lg2, TLI.getShiftAmountTy()));
892 // If we're dividing by a positive value, we're done. Otherwise, we must
893 // negate the result.
896 AddToWorkList(SRA.Val);
897 return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT), SRA);
899 // if integer divide is expensive and we satisfy the requirements, emit an
900 // alternate sequence.
901 if (N1C && (N1C->getSignExtended() < -1 || N1C->getSignExtended() > 1) &&
902 !TLI.isIntDivCheap()) {
903 SDOperand Op = BuildSDIV(N);
904 if (Op.Val) return Op;
909 SDOperand DAGCombiner::visitUDIV(SDNode *N) {
910 SDOperand N0 = N->getOperand(0);
911 SDOperand N1 = N->getOperand(1);
912 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val);
913 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val);
914 MVT::ValueType VT = N->getValueType(0);
916 // fold (udiv c1, c2) -> c1/c2
917 if (N0C && N1C && !N1C->isNullValue())
918 return DAG.getNode(ISD::UDIV, VT, N0, N1);
919 // fold (udiv x, (1 << c)) -> x >>u c
920 if (N1C && isPowerOf2_64(N1C->getValue()))
921 return DAG.getNode(ISD::SRL, VT, N0,
922 DAG.getConstant(Log2_64(N1C->getValue()),
923 TLI.getShiftAmountTy()));
924 // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2
925 if (N1.getOpcode() == ISD::SHL) {
926 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
927 if (isPowerOf2_64(SHC->getValue())) {
928 MVT::ValueType ADDVT = N1.getOperand(1).getValueType();
929 SDOperand Add = DAG.getNode(ISD::ADD, ADDVT, N1.getOperand(1),
930 DAG.getConstant(Log2_64(SHC->getValue()),
932 AddToWorkList(Add.Val);
933 return DAG.getNode(ISD::SRL, VT, N0, Add);
937 // fold (udiv x, c) -> alternate
938 if (N1C && N1C->getValue() && !TLI.isIntDivCheap()) {
939 SDOperand Op = BuildUDIV(N);
940 if (Op.Val) return Op;
945 SDOperand DAGCombiner::visitSREM(SDNode *N) {
946 SDOperand N0 = N->getOperand(0);
947 SDOperand N1 = N->getOperand(1);
948 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
949 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
950 MVT::ValueType VT = N->getValueType(0);
952 // fold (srem c1, c2) -> c1%c2
953 if (N0C && N1C && !N1C->isNullValue())
954 return DAG.getNode(ISD::SREM, VT, N0, N1);
955 // If we know the sign bits of both operands are zero, strength reduce to a
956 // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15
957 uint64_t SignBit = 1ULL << (MVT::getSizeInBits(VT)-1);
958 if (TLI.MaskedValueIsZero(N1, SignBit) &&
959 TLI.MaskedValueIsZero(N0, SignBit))
960 return DAG.getNode(ISD::UREM, VT, N0, N1);
964 SDOperand DAGCombiner::visitUREM(SDNode *N) {
965 SDOperand N0 = N->getOperand(0);
966 SDOperand N1 = N->getOperand(1);
967 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
968 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
969 MVT::ValueType VT = N->getValueType(0);
971 // fold (urem c1, c2) -> c1%c2
972 if (N0C && N1C && !N1C->isNullValue())
973 return DAG.getNode(ISD::UREM, VT, N0, N1);
974 // fold (urem x, pow2) -> (and x, pow2-1)
975 if (N1C && !N1C->isNullValue() && isPowerOf2_64(N1C->getValue()))
976 return DAG.getNode(ISD::AND, VT, N0, DAG.getConstant(N1C->getValue()-1,VT));
977 // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1))
978 if (N1.getOpcode() == ISD::SHL) {
979 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
980 if (isPowerOf2_64(SHC->getValue())) {
981 SDOperand Add = DAG.getNode(ISD::ADD, VT, N1,DAG.getConstant(~0ULL,VT));
982 AddToWorkList(Add.Val);
983 return DAG.getNode(ISD::AND, VT, N0, Add);
990 SDOperand DAGCombiner::visitMULHS(SDNode *N) {
991 SDOperand N0 = N->getOperand(0);
992 SDOperand N1 = N->getOperand(1);
993 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
995 // fold (mulhs x, 0) -> 0
996 if (N1C && N1C->isNullValue())
998 // fold (mulhs x, 1) -> (sra x, size(x)-1)
999 if (N1C && N1C->getValue() == 1)
1000 return DAG.getNode(ISD::SRA, N0.getValueType(), N0,
1001 DAG.getConstant(MVT::getSizeInBits(N0.getValueType())-1,
1002 TLI.getShiftAmountTy()));
1006 SDOperand DAGCombiner::visitMULHU(SDNode *N) {
1007 SDOperand N0 = N->getOperand(0);
1008 SDOperand N1 = N->getOperand(1);
1009 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1011 // fold (mulhu x, 0) -> 0
1012 if (N1C && N1C->isNullValue())
1014 // fold (mulhu x, 1) -> 0
1015 if (N1C && N1C->getValue() == 1)
1016 return DAG.getConstant(0, N0.getValueType());
1020 SDOperand DAGCombiner::visitAND(SDNode *N) {
1021 SDOperand N0 = N->getOperand(0);
1022 SDOperand N1 = N->getOperand(1);
1023 SDOperand LL, LR, RL, RR, CC0, CC1;
1024 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1025 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1026 MVT::ValueType VT = N1.getValueType();
1027 unsigned OpSizeInBits = MVT::getSizeInBits(VT);
1029 // fold (and c1, c2) -> c1&c2
1031 return DAG.getNode(ISD::AND, VT, N0, N1);
1032 // canonicalize constant to RHS
1034 return DAG.getNode(ISD::AND, VT, N1, N0);
1035 // fold (and x, -1) -> x
1036 if (N1C && N1C->isAllOnesValue())
1038 // if (and x, c) is known to be zero, return 0
1039 if (N1C && TLI.MaskedValueIsZero(SDOperand(N, 0), MVT::getIntVTBitMask(VT)))
1040 return DAG.getConstant(0, VT);
1042 SDOperand RAND = ReassociateOps(ISD::AND, N0, N1);
1045 // fold (and (or x, 0xFFFF), 0xFF) -> 0xFF
1046 if (N1C && N0.getOpcode() == ISD::OR)
1047 if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
1048 if ((ORI->getValue() & N1C->getValue()) == N1C->getValue())
1050 // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits.
1051 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
1052 unsigned InMask = MVT::getIntVTBitMask(N0.getOperand(0).getValueType());
1053 if (TLI.MaskedValueIsZero(N0.getOperand(0),
1054 ~N1C->getValue() & InMask)) {
1055 SDOperand Zext = DAG.getNode(ISD::ZERO_EXTEND, N0.getValueType(),
1058 // Replace uses of the AND with uses of the Zero extend node.
1061 // We actually want to replace all uses of the any_extend with the
1062 // zero_extend, to avoid duplicating things. This will later cause this
1063 // AND to be folded.
1064 CombineTo(N0.Val, Zext);
1068 // fold (and (setcc x), (setcc y)) -> (setcc (and x, y))
1069 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
1070 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
1071 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
1073 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
1074 MVT::isInteger(LL.getValueType())) {
1075 // fold (X == 0) & (Y == 0) -> (X|Y == 0)
1076 if (cast<ConstantSDNode>(LR)->getValue() == 0 && Op1 == ISD::SETEQ) {
1077 SDOperand ORNode = DAG.getNode(ISD::OR, LR.getValueType(), LL, RL);
1078 AddToWorkList(ORNode.Val);
1079 return DAG.getSetCC(VT, ORNode, LR, Op1);
1081 // fold (X == -1) & (Y == -1) -> (X&Y == -1)
1082 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETEQ) {
1083 SDOperand ANDNode = DAG.getNode(ISD::AND, LR.getValueType(), LL, RL);
1084 AddToWorkList(ANDNode.Val);
1085 return DAG.getSetCC(VT, ANDNode, LR, Op1);
1087 // fold (X > -1) & (Y > -1) -> (X|Y > -1)
1088 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETGT) {
1089 SDOperand ORNode = DAG.getNode(ISD::OR, LR.getValueType(), LL, RL);
1090 AddToWorkList(ORNode.Val);
1091 return DAG.getSetCC(VT, ORNode, LR, Op1);
1094 // canonicalize equivalent to ll == rl
1095 if (LL == RR && LR == RL) {
1096 Op1 = ISD::getSetCCSwappedOperands(Op1);
1099 if (LL == RL && LR == RR) {
1100 bool isInteger = MVT::isInteger(LL.getValueType());
1101 ISD::CondCode Result = ISD::getSetCCAndOperation(Op0, Op1, isInteger);
1102 if (Result != ISD::SETCC_INVALID)
1103 return DAG.getSetCC(N0.getValueType(), LL, LR, Result);
1106 // fold (and (zext x), (zext y)) -> (zext (and x, y))
1107 if (N0.getOpcode() == ISD::ZERO_EXTEND &&
1108 N1.getOpcode() == ISD::ZERO_EXTEND &&
1109 N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()) {
1110 SDOperand ANDNode = DAG.getNode(ISD::AND, N0.getOperand(0).getValueType(),
1111 N0.getOperand(0), N1.getOperand(0));
1112 AddToWorkList(ANDNode.Val);
1113 return DAG.getNode(ISD::ZERO_EXTEND, VT, ANDNode);
1115 // fold (and (shl/srl/sra x), (shl/srl/sra y)) -> (shl/srl/sra (and x, y))
1116 if (((N0.getOpcode() == ISD::SHL && N1.getOpcode() == ISD::SHL) ||
1117 (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SRL) ||
1118 (N0.getOpcode() == ISD::SRA && N1.getOpcode() == ISD::SRA)) &&
1119 N0.getOperand(1) == N1.getOperand(1)) {
1120 SDOperand ANDNode = DAG.getNode(ISD::AND, N0.getOperand(0).getValueType(),
1121 N0.getOperand(0), N1.getOperand(0));
1122 AddToWorkList(ANDNode.Val);
1123 return DAG.getNode(N0.getOpcode(), VT, ANDNode, N0.getOperand(1));
1125 // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1)
1126 // fold (and (sra)) -> (and (srl)) when possible.
1127 if (!MVT::isVector(VT) &&
1128 SimplifyDemandedBits(SDOperand(N, 0)))
1130 // fold (zext_inreg (extload x)) -> (zextload x)
1131 if (N0.getOpcode() == ISD::EXTLOAD) {
1132 MVT::ValueType EVT = cast<VTSDNode>(N0.getOperand(3))->getVT();
1133 // If we zero all the possible extended bits, then we can turn this into
1134 // a zextload if we are running before legalize or the operation is legal.
1135 if (TLI.MaskedValueIsZero(N1, ~0ULL << MVT::getSizeInBits(EVT)) &&
1136 (!AfterLegalize || TLI.isOperationLegal(ISD::ZEXTLOAD, EVT))) {
1137 SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, N0.getOperand(0),
1138 N0.getOperand(1), N0.getOperand(2),
1141 CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1));
1145 // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use
1146 if (N0.getOpcode() == ISD::SEXTLOAD && N0.hasOneUse()) {
1147 MVT::ValueType EVT = cast<VTSDNode>(N0.getOperand(3))->getVT();
1148 // If we zero all the possible extended bits, then we can turn this into
1149 // a zextload if we are running before legalize or the operation is legal.
1150 if (TLI.MaskedValueIsZero(N1, ~0ULL << MVT::getSizeInBits(EVT)) &&
1151 (!AfterLegalize || TLI.isOperationLegal(ISD::ZEXTLOAD, EVT))) {
1152 SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, N0.getOperand(0),
1153 N0.getOperand(1), N0.getOperand(2),
1156 CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1));
1161 // fold (and (load x), 255) -> (zextload x, i8)
1162 // fold (and (extload x, i16), 255) -> (zextload x, i8)
1164 (N0.getOpcode() == ISD::LOAD || N0.getOpcode() == ISD::EXTLOAD ||
1165 N0.getOpcode() == ISD::ZEXTLOAD) &&
1167 MVT::ValueType EVT, LoadedVT;
1168 if (N1C->getValue() == 255)
1170 else if (N1C->getValue() == 65535)
1172 else if (N1C->getValue() == ~0U)
1177 LoadedVT = N0.getOpcode() == ISD::LOAD ? VT :
1178 cast<VTSDNode>(N0.getOperand(3))->getVT();
1179 if (EVT != MVT::Other && LoadedVT > EVT) {
1180 MVT::ValueType PtrType = N0.getOperand(1).getValueType();
1181 // For big endian targets, we need to add an offset to the pointer to load
1182 // the correct bytes. For little endian systems, we merely need to read
1183 // fewer bytes from the same pointer.
1185 (MVT::getSizeInBits(LoadedVT) - MVT::getSizeInBits(EVT)) / 8;
1186 SDOperand NewPtr = N0.getOperand(1);
1187 if (!TLI.isLittleEndian())
1188 NewPtr = DAG.getNode(ISD::ADD, PtrType, NewPtr,
1189 DAG.getConstant(PtrOff, PtrType));
1190 AddToWorkList(NewPtr.Val);
1192 DAG.getExtLoad(ISD::ZEXTLOAD, VT, N0.getOperand(0), NewPtr,
1193 N0.getOperand(2), EVT);
1195 CombineTo(N0.Val, Load, Load.getValue(1));
1203 SDOperand DAGCombiner::visitOR(SDNode *N) {
1204 SDOperand N0 = N->getOperand(0);
1205 SDOperand N1 = N->getOperand(1);
1206 SDOperand LL, LR, RL, RR, CC0, CC1;
1207 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1208 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1209 MVT::ValueType VT = N1.getValueType();
1210 unsigned OpSizeInBits = MVT::getSizeInBits(VT);
1212 // fold (or c1, c2) -> c1|c2
1214 return DAG.getNode(ISD::OR, VT, N0, N1);
1215 // canonicalize constant to RHS
1217 return DAG.getNode(ISD::OR, VT, N1, N0);
1218 // fold (or x, 0) -> x
1219 if (N1C && N1C->isNullValue())
1221 // fold (or x, -1) -> -1
1222 if (N1C && N1C->isAllOnesValue())
1224 // fold (or x, c) -> c iff (x & ~c) == 0
1226 TLI.MaskedValueIsZero(N0,~N1C->getValue() & (~0ULL>>(64-OpSizeInBits))))
1229 SDOperand ROR = ReassociateOps(ISD::OR, N0, N1);
1232 // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2)
1233 if (N1C && N0.getOpcode() == ISD::AND && N0.Val->hasOneUse() &&
1234 isa<ConstantSDNode>(N0.getOperand(1))) {
1235 ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1));
1236 return DAG.getNode(ISD::AND, VT, DAG.getNode(ISD::OR, VT, N0.getOperand(0),
1238 DAG.getConstant(N1C->getValue() | C1->getValue(), VT));
1240 // fold (or (setcc x), (setcc y)) -> (setcc (or x, y))
1241 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
1242 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
1243 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
1245 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
1246 MVT::isInteger(LL.getValueType())) {
1247 // fold (X != 0) | (Y != 0) -> (X|Y != 0)
1248 // fold (X < 0) | (Y < 0) -> (X|Y < 0)
1249 if (cast<ConstantSDNode>(LR)->getValue() == 0 &&
1250 (Op1 == ISD::SETNE || Op1 == ISD::SETLT)) {
1251 SDOperand ORNode = DAG.getNode(ISD::OR, LR.getValueType(), LL, RL);
1252 AddToWorkList(ORNode.Val);
1253 return DAG.getSetCC(VT, ORNode, LR, Op1);
1255 // fold (X != -1) | (Y != -1) -> (X&Y != -1)
1256 // fold (X > -1) | (Y > -1) -> (X&Y > -1)
1257 if (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
1258 (Op1 == ISD::SETNE || Op1 == ISD::SETGT)) {
1259 SDOperand ANDNode = DAG.getNode(ISD::AND, LR.getValueType(), LL, RL);
1260 AddToWorkList(ANDNode.Val);
1261 return DAG.getSetCC(VT, ANDNode, LR, Op1);
1264 // canonicalize equivalent to ll == rl
1265 if (LL == RR && LR == RL) {
1266 Op1 = ISD::getSetCCSwappedOperands(Op1);
1269 if (LL == RL && LR == RR) {
1270 bool isInteger = MVT::isInteger(LL.getValueType());
1271 ISD::CondCode Result = ISD::getSetCCOrOperation(Op0, Op1, isInteger);
1272 if (Result != ISD::SETCC_INVALID)
1273 return DAG.getSetCC(N0.getValueType(), LL, LR, Result);
1276 // fold (or (zext x), (zext y)) -> (zext (or x, y))
1277 if (N0.getOpcode() == ISD::ZERO_EXTEND &&
1278 N1.getOpcode() == ISD::ZERO_EXTEND &&
1279 N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()) {
1280 SDOperand ORNode = DAG.getNode(ISD::OR, N0.getOperand(0).getValueType(),
1281 N0.getOperand(0), N1.getOperand(0));
1282 AddToWorkList(ORNode.Val);
1283 return DAG.getNode(ISD::ZERO_EXTEND, VT, ORNode);
1285 // fold (or (shl/srl/sra x), (shl/srl/sra y)) -> (shl/srl/sra (or x, y))
1286 if (((N0.getOpcode() == ISD::SHL && N1.getOpcode() == ISD::SHL) ||
1287 (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SRL) ||
1288 (N0.getOpcode() == ISD::SRA && N1.getOpcode() == ISD::SRA)) &&
1289 N0.getOperand(1) == N1.getOperand(1)) {
1290 SDOperand ORNode = DAG.getNode(ISD::OR, N0.getOperand(0).getValueType(),
1291 N0.getOperand(0), N1.getOperand(0));
1292 AddToWorkList(ORNode.Val);
1293 return DAG.getNode(N0.getOpcode(), VT, ORNode, N0.getOperand(1));
1295 // canonicalize shl to left side in a shl/srl pair, to match rotate
1296 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
1298 // check for rotl, rotr
1299 if (N0.getOpcode() == ISD::SHL && N1.getOpcode() == ISD::SRL &&
1300 N0.getOperand(0) == N1.getOperand(0) &&
1301 TLI.isOperationLegal(ISD::ROTL, VT) && TLI.isTypeLegal(VT)) {
1302 // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1)
1303 if (N0.getOperand(1).getOpcode() == ISD::Constant &&
1304 N1.getOperand(1).getOpcode() == ISD::Constant) {
1305 uint64_t c1val = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
1306 uint64_t c2val = cast<ConstantSDNode>(N1.getOperand(1))->getValue();
1307 if ((c1val + c2val) == OpSizeInBits)
1308 return DAG.getNode(ISD::ROTL, VT, N0.getOperand(0), N0.getOperand(1));
1310 // fold (or (shl x, y), (srl x, (sub 32, y))) -> (rotl x, y)
1311 if (N1.getOperand(1).getOpcode() == ISD::SUB &&
1312 N0.getOperand(1) == N1.getOperand(1).getOperand(1))
1313 if (ConstantSDNode *SUBC =
1314 dyn_cast<ConstantSDNode>(N1.getOperand(1).getOperand(0)))
1315 if (SUBC->getValue() == OpSizeInBits)
1316 return DAG.getNode(ISD::ROTL, VT, N0.getOperand(0), N0.getOperand(1));
1317 // fold (or (shl x, (sub 32, y)), (srl x, r)) -> (rotr x, y)
1318 if (N0.getOperand(1).getOpcode() == ISD::SUB &&
1319 N1.getOperand(1) == N0.getOperand(1).getOperand(1))
1320 if (ConstantSDNode *SUBC =
1321 dyn_cast<ConstantSDNode>(N0.getOperand(1).getOperand(0)))
1322 if (SUBC->getValue() == OpSizeInBits) {
1323 if (TLI.isOperationLegal(ISD::ROTR, VT) && TLI.isTypeLegal(VT))
1324 return DAG.getNode(ISD::ROTR, VT, N0.getOperand(0),
1327 return DAG.getNode(ISD::ROTL, VT, N0.getOperand(0),
1334 SDOperand DAGCombiner::visitXOR(SDNode *N) {
1335 SDOperand N0 = N->getOperand(0);
1336 SDOperand N1 = N->getOperand(1);
1337 SDOperand LHS, RHS, CC;
1338 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1339 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1340 MVT::ValueType VT = N0.getValueType();
1342 // fold (xor c1, c2) -> c1^c2
1344 return DAG.getNode(ISD::XOR, VT, N0, N1);
1345 // canonicalize constant to RHS
1347 return DAG.getNode(ISD::XOR, VT, N1, N0);
1348 // fold (xor x, 0) -> x
1349 if (N1C && N1C->isNullValue())
1352 SDOperand RXOR = ReassociateOps(ISD::XOR, N0, N1);
1355 // fold !(x cc y) -> (x !cc y)
1356 if (N1C && N1C->getValue() == 1 && isSetCCEquivalent(N0, LHS, RHS, CC)) {
1357 bool isInt = MVT::isInteger(LHS.getValueType());
1358 ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
1360 if (N0.getOpcode() == ISD::SETCC)
1361 return DAG.getSetCC(VT, LHS, RHS, NotCC);
1362 if (N0.getOpcode() == ISD::SELECT_CC)
1363 return DAG.getSelectCC(LHS, RHS, N0.getOperand(2),N0.getOperand(3),NotCC);
1364 assert(0 && "Unhandled SetCC Equivalent!");
1367 // fold !(x or y) -> (!x and !y) iff x or y are setcc
1368 if (N1C && N1C->getValue() == 1 &&
1369 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
1370 SDOperand LHS = N0.getOperand(0), RHS = N0.getOperand(1);
1371 if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) {
1372 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
1373 LHS = DAG.getNode(ISD::XOR, VT, LHS, N1); // RHS = ~LHS
1374 RHS = DAG.getNode(ISD::XOR, VT, RHS, N1); // RHS = ~RHS
1375 AddToWorkList(LHS.Val); AddToWorkList(RHS.Val);
1376 return DAG.getNode(NewOpcode, VT, LHS, RHS);
1379 // fold !(x or y) -> (!x and !y) iff x or y are constants
1380 if (N1C && N1C->isAllOnesValue() &&
1381 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
1382 SDOperand LHS = N0.getOperand(0), RHS = N0.getOperand(1);
1383 if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) {
1384 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
1385 LHS = DAG.getNode(ISD::XOR, VT, LHS, N1); // RHS = ~LHS
1386 RHS = DAG.getNode(ISD::XOR, VT, RHS, N1); // RHS = ~RHS
1387 AddToWorkList(LHS.Val); AddToWorkList(RHS.Val);
1388 return DAG.getNode(NewOpcode, VT, LHS, RHS);
1391 // fold (xor (xor x, c1), c2) -> (xor x, c1^c2)
1392 if (N1C && N0.getOpcode() == ISD::XOR) {
1393 ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0));
1394 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
1396 return DAG.getNode(ISD::XOR, VT, N0.getOperand(1),
1397 DAG.getConstant(N1C->getValue()^N00C->getValue(), VT));
1399 return DAG.getNode(ISD::XOR, VT, N0.getOperand(0),
1400 DAG.getConstant(N1C->getValue()^N01C->getValue(), VT));
1402 // fold (xor x, x) -> 0
1404 if (!MVT::isVector(VT)) {
1405 return DAG.getConstant(0, VT);
1406 } else if (!AfterLegalize || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT)) {
1407 // Produce a vector of zeros.
1408 SDOperand El = DAG.getConstant(0, MVT::getVectorBaseType(VT));
1409 std::vector<SDOperand> Ops(MVT::getVectorNumElements(VT), El);
1410 return DAG.getNode(ISD::BUILD_VECTOR, VT, Ops);
1413 // fold (xor (zext x), (zext y)) -> (zext (xor x, y))
1414 if (N0.getOpcode() == ISD::ZERO_EXTEND &&
1415 N1.getOpcode() == ISD::ZERO_EXTEND &&
1416 N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()) {
1417 SDOperand XORNode = DAG.getNode(ISD::XOR, N0.getOperand(0).getValueType(),
1418 N0.getOperand(0), N1.getOperand(0));
1419 AddToWorkList(XORNode.Val);
1420 return DAG.getNode(ISD::ZERO_EXTEND, VT, XORNode);
1422 // fold (xor (shl/srl/sra x), (shl/srl/sra y)) -> (shl/srl/sra (xor x, y))
1423 if (((N0.getOpcode() == ISD::SHL && N1.getOpcode() == ISD::SHL) ||
1424 (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SRL) ||
1425 (N0.getOpcode() == ISD::SRA && N1.getOpcode() == ISD::SRA)) &&
1426 N0.getOperand(1) == N1.getOperand(1)) {
1427 SDOperand XORNode = DAG.getNode(ISD::XOR, N0.getOperand(0).getValueType(),
1428 N0.getOperand(0), N1.getOperand(0));
1429 AddToWorkList(XORNode.Val);
1430 return DAG.getNode(N0.getOpcode(), VT, XORNode, N0.getOperand(1));
1435 SDOperand DAGCombiner::visitSHL(SDNode *N) {
1436 SDOperand N0 = N->getOperand(0);
1437 SDOperand N1 = N->getOperand(1);
1438 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1439 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1440 MVT::ValueType VT = N0.getValueType();
1441 unsigned OpSizeInBits = MVT::getSizeInBits(VT);
1443 // fold (shl c1, c2) -> c1<<c2
1445 return DAG.getNode(ISD::SHL, VT, N0, N1);
1446 // fold (shl 0, x) -> 0
1447 if (N0C && N0C->isNullValue())
1449 // fold (shl x, c >= size(x)) -> undef
1450 if (N1C && N1C->getValue() >= OpSizeInBits)
1451 return DAG.getNode(ISD::UNDEF, VT);
1452 // fold (shl x, 0) -> x
1453 if (N1C && N1C->isNullValue())
1455 // if (shl x, c) is known to be zero, return 0
1456 if (TLI.MaskedValueIsZero(SDOperand(N, 0), MVT::getIntVTBitMask(VT)))
1457 return DAG.getConstant(0, VT);
1458 if (SimplifyDemandedBits(SDOperand(N, 0)))
1460 // fold (shl (shl x, c1), c2) -> 0 or (shl x, c1+c2)
1461 if (N1C && N0.getOpcode() == ISD::SHL &&
1462 N0.getOperand(1).getOpcode() == ISD::Constant) {
1463 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
1464 uint64_t c2 = N1C->getValue();
1465 if (c1 + c2 > OpSizeInBits)
1466 return DAG.getConstant(0, VT);
1467 return DAG.getNode(ISD::SHL, VT, N0.getOperand(0),
1468 DAG.getConstant(c1 + c2, N1.getValueType()));
1470 // fold (shl (srl x, c1), c2) -> (shl (and x, -1 << c1), c2-c1) or
1471 // (srl (and x, -1 << c1), c1-c2)
1472 if (N1C && N0.getOpcode() == ISD::SRL &&
1473 N0.getOperand(1).getOpcode() == ISD::Constant) {
1474 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
1475 uint64_t c2 = N1C->getValue();
1476 SDOperand Mask = DAG.getNode(ISD::AND, VT, N0.getOperand(0),
1477 DAG.getConstant(~0ULL << c1, VT));
1479 return DAG.getNode(ISD::SHL, VT, Mask,
1480 DAG.getConstant(c2-c1, N1.getValueType()));
1482 return DAG.getNode(ISD::SRL, VT, Mask,
1483 DAG.getConstant(c1-c2, N1.getValueType()));
1485 // fold (shl (sra x, c1), c1) -> (and x, -1 << c1)
1486 if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1))
1487 return DAG.getNode(ISD::AND, VT, N0.getOperand(0),
1488 DAG.getConstant(~0ULL << N1C->getValue(), VT));
1489 // fold (shl (add x, c1), c2) -> (add (shl x, c2), c1<<c2)
1490 if (N1C && N0.getOpcode() == ISD::ADD && N0.Val->hasOneUse() &&
1491 isa<ConstantSDNode>(N0.getOperand(1))) {
1492 return DAG.getNode(ISD::ADD, VT,
1493 DAG.getNode(ISD::SHL, VT, N0.getOperand(0), N1),
1494 DAG.getNode(ISD::SHL, VT, N0.getOperand(1), N1));
1499 SDOperand DAGCombiner::visitSRA(SDNode *N) {
1500 SDOperand N0 = N->getOperand(0);
1501 SDOperand N1 = N->getOperand(1);
1502 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1503 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1504 MVT::ValueType VT = N0.getValueType();
1506 // fold (sra c1, c2) -> c1>>c2
1508 return DAG.getNode(ISD::SRA, VT, N0, N1);
1509 // fold (sra 0, x) -> 0
1510 if (N0C && N0C->isNullValue())
1512 // fold (sra -1, x) -> -1
1513 if (N0C && N0C->isAllOnesValue())
1515 // fold (sra x, c >= size(x)) -> undef
1516 if (N1C && N1C->getValue() >= MVT::getSizeInBits(VT))
1517 return DAG.getNode(ISD::UNDEF, VT);
1518 // fold (sra x, 0) -> x
1519 if (N1C && N1C->isNullValue())
1521 // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports
1523 if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) {
1524 unsigned LowBits = MVT::getSizeInBits(VT) - (unsigned)N1C->getValue();
1527 default: EVT = MVT::Other; break;
1528 case 1: EVT = MVT::i1; break;
1529 case 8: EVT = MVT::i8; break;
1530 case 16: EVT = MVT::i16; break;
1531 case 32: EVT = MVT::i32; break;
1533 if (EVT > MVT::Other && TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, EVT))
1534 return DAG.getNode(ISD::SIGN_EXTEND_INREG, VT, N0.getOperand(0),
1535 DAG.getValueType(EVT));
1538 // fold (sra (sra x, c1), c2) -> (sra x, c1+c2)
1539 if (N1C && N0.getOpcode() == ISD::SRA) {
1540 if (ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
1541 unsigned Sum = N1C->getValue() + C1->getValue();
1542 if (Sum >= MVT::getSizeInBits(VT)) Sum = MVT::getSizeInBits(VT)-1;
1543 return DAG.getNode(ISD::SRA, VT, N0.getOperand(0),
1544 DAG.getConstant(Sum, N1C->getValueType(0)));
1548 // If the sign bit is known to be zero, switch this to a SRL.
1549 if (TLI.MaskedValueIsZero(N0, MVT::getIntVTSignBit(VT)))
1550 return DAG.getNode(ISD::SRL, VT, N0, N1);
1554 SDOperand DAGCombiner::visitSRL(SDNode *N) {
1555 SDOperand N0 = N->getOperand(0);
1556 SDOperand N1 = N->getOperand(1);
1557 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1558 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1559 MVT::ValueType VT = N0.getValueType();
1560 unsigned OpSizeInBits = MVT::getSizeInBits(VT);
1562 // fold (srl c1, c2) -> c1 >>u c2
1564 return DAG.getNode(ISD::SRL, VT, N0, N1);
1565 // fold (srl 0, x) -> 0
1566 if (N0C && N0C->isNullValue())
1568 // fold (srl x, c >= size(x)) -> undef
1569 if (N1C && N1C->getValue() >= OpSizeInBits)
1570 return DAG.getNode(ISD::UNDEF, VT);
1571 // fold (srl x, 0) -> x
1572 if (N1C && N1C->isNullValue())
1574 // if (srl x, c) is known to be zero, return 0
1575 if (N1C && TLI.MaskedValueIsZero(SDOperand(N, 0), ~0ULL >> (64-OpSizeInBits)))
1576 return DAG.getConstant(0, VT);
1577 // fold (srl (srl x, c1), c2) -> 0 or (srl x, c1+c2)
1578 if (N1C && N0.getOpcode() == ISD::SRL &&
1579 N0.getOperand(1).getOpcode() == ISD::Constant) {
1580 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getValue();
1581 uint64_t c2 = N1C->getValue();
1582 if (c1 + c2 > OpSizeInBits)
1583 return DAG.getConstant(0, VT);
1584 return DAG.getNode(ISD::SRL, VT, N0.getOperand(0),
1585 DAG.getConstant(c1 + c2, N1.getValueType()));
1590 SDOperand DAGCombiner::visitCTLZ(SDNode *N) {
1591 SDOperand N0 = N->getOperand(0);
1592 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1593 MVT::ValueType VT = N->getValueType(0);
1595 // fold (ctlz c1) -> c2
1597 return DAG.getNode(ISD::CTLZ, VT, N0);
1601 SDOperand DAGCombiner::visitCTTZ(SDNode *N) {
1602 SDOperand N0 = N->getOperand(0);
1603 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1604 MVT::ValueType VT = N->getValueType(0);
1606 // fold (cttz c1) -> c2
1608 return DAG.getNode(ISD::CTTZ, VT, N0);
1612 SDOperand DAGCombiner::visitCTPOP(SDNode *N) {
1613 SDOperand N0 = N->getOperand(0);
1614 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1615 MVT::ValueType VT = N->getValueType(0);
1617 // fold (ctpop c1) -> c2
1619 return DAG.getNode(ISD::CTPOP, VT, N0);
1623 SDOperand DAGCombiner::visitSELECT(SDNode *N) {
1624 SDOperand N0 = N->getOperand(0);
1625 SDOperand N1 = N->getOperand(1);
1626 SDOperand N2 = N->getOperand(2);
1627 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1628 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1629 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
1630 MVT::ValueType VT = N->getValueType(0);
1632 // fold select C, X, X -> X
1635 // fold select true, X, Y -> X
1636 if (N0C && !N0C->isNullValue())
1638 // fold select false, X, Y -> Y
1639 if (N0C && N0C->isNullValue())
1641 // fold select C, 1, X -> C | X
1642 if (MVT::i1 == VT && N1C && N1C->getValue() == 1)
1643 return DAG.getNode(ISD::OR, VT, N0, N2);
1644 // fold select C, 0, X -> ~C & X
1645 // FIXME: this should check for C type == X type, not i1?
1646 if (MVT::i1 == VT && N1C && N1C->isNullValue()) {
1647 SDOperand XORNode = DAG.getNode(ISD::XOR, VT, N0, DAG.getConstant(1, VT));
1648 AddToWorkList(XORNode.Val);
1649 return DAG.getNode(ISD::AND, VT, XORNode, N2);
1651 // fold select C, X, 1 -> ~C | X
1652 if (MVT::i1 == VT && N2C && N2C->getValue() == 1) {
1653 SDOperand XORNode = DAG.getNode(ISD::XOR, VT, N0, DAG.getConstant(1, VT));
1654 AddToWorkList(XORNode.Val);
1655 return DAG.getNode(ISD::OR, VT, XORNode, N1);
1657 // fold select C, X, 0 -> C & X
1658 // FIXME: this should check for C type == X type, not i1?
1659 if (MVT::i1 == VT && N2C && N2C->isNullValue())
1660 return DAG.getNode(ISD::AND, VT, N0, N1);
1661 // fold X ? X : Y --> X ? 1 : Y --> X | Y
1662 if (MVT::i1 == VT && N0 == N1)
1663 return DAG.getNode(ISD::OR, VT, N0, N2);
1664 // fold X ? Y : X --> X ? Y : 0 --> X & Y
1665 if (MVT::i1 == VT && N0 == N2)
1666 return DAG.getNode(ISD::AND, VT, N0, N1);
1667 // If we can fold this based on the true/false value, do so.
1668 if (SimplifySelectOps(N, N1, N2))
1670 // fold selects based on a setcc into other things, such as min/max/abs
1671 if (N0.getOpcode() == ISD::SETCC)
1673 // Check against MVT::Other for SELECT_CC, which is a workaround for targets
1674 // having to say they don't support SELECT_CC on every type the DAG knows
1675 // about, since there is no way to mark an opcode illegal at all value types
1676 if (TLI.isOperationLegal(ISD::SELECT_CC, MVT::Other))
1677 return DAG.getNode(ISD::SELECT_CC, VT, N0.getOperand(0), N0.getOperand(1),
1678 N1, N2, N0.getOperand(2));
1680 return SimplifySelect(N0, N1, N2);
1684 SDOperand DAGCombiner::visitSELECT_CC(SDNode *N) {
1685 SDOperand N0 = N->getOperand(0);
1686 SDOperand N1 = N->getOperand(1);
1687 SDOperand N2 = N->getOperand(2);
1688 SDOperand N3 = N->getOperand(3);
1689 SDOperand N4 = N->getOperand(4);
1690 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1691 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1692 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
1693 ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get();
1695 // Determine if the condition we're dealing with is constant
1696 SDOperand SCC = SimplifySetCC(TLI.getSetCCResultTy(), N0, N1, CC, false);
1697 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.Val);
1699 // fold select_cc lhs, rhs, x, x, cc -> x
1703 // If we can fold this based on the true/false value, do so.
1704 if (SimplifySelectOps(N, N2, N3))
1707 // fold select_cc into other things, such as min/max/abs
1708 return SimplifySelectCC(N0, N1, N2, N3, CC);
1711 SDOperand DAGCombiner::visitSETCC(SDNode *N) {
1712 return SimplifySetCC(N->getValueType(0), N->getOperand(0), N->getOperand(1),
1713 cast<CondCodeSDNode>(N->getOperand(2))->get());
1716 SDOperand DAGCombiner::visitSIGN_EXTEND(SDNode *N) {
1717 SDOperand N0 = N->getOperand(0);
1718 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1719 MVT::ValueType VT = N->getValueType(0);
1721 // fold (sext c1) -> c1
1723 return DAG.getNode(ISD::SIGN_EXTEND, VT, N0);
1724 // fold (sext (sext x)) -> (sext x)
1725 if (N0.getOpcode() == ISD::SIGN_EXTEND)
1726 return DAG.getNode(ISD::SIGN_EXTEND, VT, N0.getOperand(0));
1727 // fold (sext (truncate x)) -> (sextinreg x) iff x size == sext size.
1728 if (N0.getOpcode() == ISD::TRUNCATE && N0.getOperand(0).getValueType() == VT&&
1730 TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, N0.getValueType())))
1731 return DAG.getNode(ISD::SIGN_EXTEND_INREG, VT, N0.getOperand(0),
1732 DAG.getValueType(N0.getValueType()));
1733 // fold (sext (load x)) -> (sext (truncate (sextload x)))
1734 if (N0.getOpcode() == ISD::LOAD && N0.hasOneUse() &&
1735 (!AfterLegalize||TLI.isOperationLegal(ISD::SEXTLOAD, N0.getValueType()))){
1736 SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, N0.getOperand(0),
1737 N0.getOperand(1), N0.getOperand(2),
1739 CombineTo(N, ExtLoad);
1740 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
1741 ExtLoad.getValue(1));
1745 // fold (sext (sextload x)) -> (sext (truncate (sextload x)))
1746 // fold (sext ( extload x)) -> (sext (truncate (sextload x)))
1747 if ((N0.getOpcode() == ISD::SEXTLOAD || N0.getOpcode() == ISD::EXTLOAD) &&
1749 SDOperand ExtLoad = DAG.getNode(ISD::SEXTLOAD, VT, N0.getOperand(0),
1750 N0.getOperand(1), N0.getOperand(2),
1752 CombineTo(N, ExtLoad);
1753 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
1754 ExtLoad.getValue(1));
1761 SDOperand DAGCombiner::visitZERO_EXTEND(SDNode *N) {
1762 SDOperand N0 = N->getOperand(0);
1763 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1764 MVT::ValueType VT = N->getValueType(0);
1766 // fold (zext c1) -> c1
1768 return DAG.getNode(ISD::ZERO_EXTEND, VT, N0);
1769 // fold (zext (zext x)) -> (zext x)
1770 if (N0.getOpcode() == ISD::ZERO_EXTEND)
1771 return DAG.getNode(ISD::ZERO_EXTEND, VT, N0.getOperand(0));
1772 // fold (zext (truncate x)) -> (zextinreg x) iff x size == zext size.
1773 if (N0.getOpcode() == ISD::TRUNCATE && N0.getOperand(0).getValueType() == VT&&
1774 (!AfterLegalize || TLI.isOperationLegal(ISD::AND, N0.getValueType())))
1775 return DAG.getZeroExtendInReg(N0.getOperand(0), N0.getValueType());
1776 // fold (zext (load x)) -> (zext (truncate (zextload x)))
1777 if (N0.getOpcode() == ISD::LOAD && N0.hasOneUse() &&
1778 (!AfterLegalize||TLI.isOperationLegal(ISD::ZEXTLOAD, N0.getValueType()))){
1779 SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, N0.getOperand(0),
1780 N0.getOperand(1), N0.getOperand(2),
1782 CombineTo(N, ExtLoad);
1783 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
1784 ExtLoad.getValue(1));
1788 // fold (zext (zextload x)) -> (zext (truncate (zextload x)))
1789 // fold (zext ( extload x)) -> (zext (truncate (zextload x)))
1790 if ((N0.getOpcode() == ISD::ZEXTLOAD || N0.getOpcode() == ISD::EXTLOAD) &&
1792 SDOperand ExtLoad = DAG.getNode(ISD::ZEXTLOAD, VT, N0.getOperand(0),
1793 N0.getOperand(1), N0.getOperand(2),
1795 CombineTo(N, ExtLoad);
1796 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
1797 ExtLoad.getValue(1));
1803 SDOperand DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
1804 SDOperand N0 = N->getOperand(0);
1805 SDOperand N1 = N->getOperand(1);
1806 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1807 MVT::ValueType VT = N->getValueType(0);
1808 MVT::ValueType EVT = cast<VTSDNode>(N1)->getVT();
1809 unsigned EVTBits = MVT::getSizeInBits(EVT);
1811 // fold (sext_in_reg c1) -> c1
1813 SDOperand Truncate = DAG.getConstant(N0C->getValue(), EVT);
1814 return DAG.getNode(ISD::SIGN_EXTEND, VT, Truncate);
1816 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt1
1817 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
1818 cast<VTSDNode>(N0.getOperand(1))->getVT() <= EVT) {
1821 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2
1822 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
1823 EVT < cast<VTSDNode>(N0.getOperand(1))->getVT()) {
1824 return DAG.getNode(ISD::SIGN_EXTEND_INREG, VT, N0.getOperand(0), N1);
1826 // fold (sext_in_reg (assert_sext x)) -> (assert_sext x)
1827 if (N0.getOpcode() == ISD::AssertSext &&
1828 cast<VTSDNode>(N0.getOperand(1))->getVT() <= EVT) {
1831 // fold (sext_in_reg (sextload x)) -> (sextload x)
1832 if (N0.getOpcode() == ISD::SEXTLOAD &&
1833 cast<VTSDNode>(N0.getOperand(3))->getVT() <= EVT) {
1836 // fold (sext_in_reg (setcc x)) -> setcc x iff (setcc x) == 0 or -1
1837 if (N0.getOpcode() == ISD::SETCC &&
1838 TLI.getSetCCResultContents() ==
1839 TargetLowering::ZeroOrNegativeOneSetCCResult)
1841 // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is zero
1842 if (TLI.MaskedValueIsZero(N0, 1ULL << (EVTBits-1)))
1843 return DAG.getZeroExtendInReg(N0, EVT);
1844 // fold (sext_in_reg (srl x)) -> sra x
1845 if (N0.getOpcode() == ISD::SRL &&
1846 N0.getOperand(1).getOpcode() == ISD::Constant &&
1847 cast<ConstantSDNode>(N0.getOperand(1))->getValue() == EVTBits) {
1848 return DAG.getNode(ISD::SRA, N0.getValueType(), N0.getOperand(0),
1851 // fold (sext_inreg (extload x)) -> (sextload x)
1852 if (N0.getOpcode() == ISD::EXTLOAD &&
1853 EVT == cast<VTSDNode>(N0.getOperand(3))->getVT() &&
1854 (!AfterLegalize || TLI.isOperationLegal(ISD::SEXTLOAD, EVT))) {
1855 SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, N0.getOperand(0),
1856 N0.getOperand(1), N0.getOperand(2),
1858 CombineTo(N, ExtLoad);
1859 CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1));
1862 // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use
1863 if (N0.getOpcode() == ISD::ZEXTLOAD && N0.hasOneUse() &&
1864 EVT == cast<VTSDNode>(N0.getOperand(3))->getVT() &&
1865 (!AfterLegalize || TLI.isOperationLegal(ISD::SEXTLOAD, EVT))) {
1866 SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, N0.getOperand(0),
1867 N0.getOperand(1), N0.getOperand(2),
1869 CombineTo(N, ExtLoad);
1870 CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1));
1876 SDOperand DAGCombiner::visitTRUNCATE(SDNode *N) {
1877 SDOperand N0 = N->getOperand(0);
1878 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1879 MVT::ValueType VT = N->getValueType(0);
1882 if (N0.getValueType() == N->getValueType(0))
1884 // fold (truncate c1) -> c1
1886 return DAG.getNode(ISD::TRUNCATE, VT, N0);
1887 // fold (truncate (truncate x)) -> (truncate x)
1888 if (N0.getOpcode() == ISD::TRUNCATE)
1889 return DAG.getNode(ISD::TRUNCATE, VT, N0.getOperand(0));
1890 // fold (truncate (ext x)) -> (ext x) or (truncate x) or x
1891 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::SIGN_EXTEND){
1892 if (N0.getValueType() < VT)
1893 // if the source is smaller than the dest, we still need an extend
1894 return DAG.getNode(N0.getOpcode(), VT, N0.getOperand(0));
1895 else if (N0.getValueType() > VT)
1896 // if the source is larger than the dest, than we just need the truncate
1897 return DAG.getNode(ISD::TRUNCATE, VT, N0.getOperand(0));
1899 // if the source and dest are the same type, we can drop both the extend
1901 return N0.getOperand(0);
1903 // fold (truncate (load x)) -> (smaller load x)
1904 if (N0.getOpcode() == ISD::LOAD && N0.hasOneUse()) {
1905 assert(MVT::getSizeInBits(N0.getValueType()) > MVT::getSizeInBits(VT) &&
1906 "Cannot truncate to larger type!");
1907 MVT::ValueType PtrType = N0.getOperand(1).getValueType();
1908 // For big endian targets, we need to add an offset to the pointer to load
1909 // the correct bytes. For little endian systems, we merely need to read
1910 // fewer bytes from the same pointer.
1912 (MVT::getSizeInBits(N0.getValueType()) - MVT::getSizeInBits(VT)) / 8;
1913 SDOperand NewPtr = TLI.isLittleEndian() ? N0.getOperand(1) :
1914 DAG.getNode(ISD::ADD, PtrType, N0.getOperand(1),
1915 DAG.getConstant(PtrOff, PtrType));
1916 AddToWorkList(NewPtr.Val);
1917 SDOperand Load = DAG.getLoad(VT, N0.getOperand(0), NewPtr,N0.getOperand(2));
1919 CombineTo(N0.Val, Load, Load.getValue(1));
1925 SDOperand DAGCombiner::visitBIT_CONVERT(SDNode *N) {
1926 SDOperand N0 = N->getOperand(0);
1927 MVT::ValueType VT = N->getValueType(0);
1929 // If the input is a constant, let getNode() fold it.
1930 if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) {
1931 SDOperand Res = DAG.getNode(ISD::BIT_CONVERT, VT, N0);
1932 if (Res.Val != N) return Res;
1935 if (N0.getOpcode() == ISD::BIT_CONVERT) // conv(conv(x,t1),t2) -> conv(x,t2)
1936 return DAG.getNode(ISD::BIT_CONVERT, VT, N0.getOperand(0));
1938 // fold (conv (load x)) -> (load (conv*)x)
1939 // FIXME: These xforms need to know that the resultant load doesn't need a
1940 // higher alignment than the original!
1941 if (0 && N0.getOpcode() == ISD::LOAD && N0.hasOneUse()) {
1942 SDOperand Load = DAG.getLoad(VT, N0.getOperand(0), N0.getOperand(1),
1945 CombineTo(N0.Val, DAG.getNode(ISD::BIT_CONVERT, N0.getValueType(), Load),
1953 SDOperand DAGCombiner::visitFADD(SDNode *N) {
1954 SDOperand N0 = N->getOperand(0);
1955 SDOperand N1 = N->getOperand(1);
1956 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
1957 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
1958 MVT::ValueType VT = N->getValueType(0);
1960 // fold (fadd c1, c2) -> c1+c2
1962 return DAG.getNode(ISD::FADD, VT, N0, N1);
1963 // canonicalize constant to RHS
1964 if (N0CFP && !N1CFP)
1965 return DAG.getNode(ISD::FADD, VT, N1, N0);
1966 // fold (A + (-B)) -> A-B
1967 if (N1.getOpcode() == ISD::FNEG)
1968 return DAG.getNode(ISD::FSUB, VT, N0, N1.getOperand(0));
1969 // fold ((-A) + B) -> B-A
1970 if (N0.getOpcode() == ISD::FNEG)
1971 return DAG.getNode(ISD::FSUB, VT, N1, N0.getOperand(0));
1975 SDOperand DAGCombiner::visitFSUB(SDNode *N) {
1976 SDOperand N0 = N->getOperand(0);
1977 SDOperand N1 = N->getOperand(1);
1978 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
1979 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
1980 MVT::ValueType VT = N->getValueType(0);
1982 // fold (fsub c1, c2) -> c1-c2
1984 return DAG.getNode(ISD::FSUB, VT, N0, N1);
1985 // fold (A-(-B)) -> A+B
1986 if (N1.getOpcode() == ISD::FNEG)
1987 return DAG.getNode(ISD::FADD, VT, N0, N1.getOperand(0));
1991 SDOperand DAGCombiner::visitFMUL(SDNode *N) {
1992 SDOperand N0 = N->getOperand(0);
1993 SDOperand N1 = N->getOperand(1);
1994 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
1995 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
1996 MVT::ValueType VT = N->getValueType(0);
1998 // fold (fmul c1, c2) -> c1*c2
2000 return DAG.getNode(ISD::FMUL, VT, N0, N1);
2001 // canonicalize constant to RHS
2002 if (N0CFP && !N1CFP)
2003 return DAG.getNode(ISD::FMUL, VT, N1, N0);
2004 // fold (fmul X, 2.0) -> (fadd X, X)
2005 if (N1CFP && N1CFP->isExactlyValue(+2.0))
2006 return DAG.getNode(ISD::FADD, VT, N0, N0);
2010 SDOperand DAGCombiner::visitFDIV(SDNode *N) {
2011 SDOperand N0 = N->getOperand(0);
2012 SDOperand N1 = N->getOperand(1);
2013 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2014 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
2015 MVT::ValueType VT = N->getValueType(0);
2017 // fold (fdiv c1, c2) -> c1/c2
2019 return DAG.getNode(ISD::FDIV, VT, N0, N1);
2023 SDOperand DAGCombiner::visitFREM(SDNode *N) {
2024 SDOperand N0 = N->getOperand(0);
2025 SDOperand N1 = N->getOperand(1);
2026 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2027 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
2028 MVT::ValueType VT = N->getValueType(0);
2030 // fold (frem c1, c2) -> fmod(c1,c2)
2032 return DAG.getNode(ISD::FREM, VT, N0, N1);
2036 SDOperand DAGCombiner::visitFCOPYSIGN(SDNode *N) {
2037 SDOperand N0 = N->getOperand(0);
2038 SDOperand N1 = N->getOperand(1);
2039 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2040 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
2041 MVT::ValueType VT = N->getValueType(0);
2043 if (N0CFP && N1CFP) // Constant fold
2044 return DAG.getNode(ISD::FCOPYSIGN, VT, N0, N1);
2047 // copysign(x, c1) -> fabs(x) iff ispos(c1)
2048 // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1)
2053 u.d = N1CFP->getValue();
2055 return DAG.getNode(ISD::FABS, VT, N0);
2057 return DAG.getNode(ISD::FNEG, VT, DAG.getNode(ISD::FABS, VT, N0));
2060 // copysign(fabs(x), y) -> copysign(x, y)
2061 // copysign(fneg(x), y) -> copysign(x, y)
2062 // copysign(copysign(x,z), y) -> copysign(x, y)
2063 if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG ||
2064 N0.getOpcode() == ISD::FCOPYSIGN)
2065 return DAG.getNode(ISD::FCOPYSIGN, VT, N0.getOperand(0), N1);
2067 // copysign(x, abs(y)) -> abs(x)
2068 if (N1.getOpcode() == ISD::FABS)
2069 return DAG.getNode(ISD::FABS, VT, N0);
2071 // copysign(x, copysign(y,z)) -> copysign(x, z)
2072 if (N1.getOpcode() == ISD::FCOPYSIGN)
2073 return DAG.getNode(ISD::FCOPYSIGN, VT, N0, N1.getOperand(1));
2075 // copysign(x, fp_extend(y)) -> copysign(x, y)
2076 // copysign(x, fp_round(y)) -> copysign(x, y)
2077 if (N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND)
2078 return DAG.getNode(ISD::FCOPYSIGN, VT, N0, N1.getOperand(0));
2085 SDOperand DAGCombiner::visitSINT_TO_FP(SDNode *N) {
2086 SDOperand N0 = N->getOperand(0);
2087 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2088 MVT::ValueType VT = N->getValueType(0);
2090 // fold (sint_to_fp c1) -> c1fp
2092 return DAG.getNode(ISD::SINT_TO_FP, VT, N0);
2096 SDOperand DAGCombiner::visitUINT_TO_FP(SDNode *N) {
2097 SDOperand N0 = N->getOperand(0);
2098 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2099 MVT::ValueType VT = N->getValueType(0);
2101 // fold (uint_to_fp c1) -> c1fp
2103 return DAG.getNode(ISD::UINT_TO_FP, VT, N0);
2107 SDOperand DAGCombiner::visitFP_TO_SINT(SDNode *N) {
2108 SDOperand N0 = N->getOperand(0);
2109 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2110 MVT::ValueType VT = N->getValueType(0);
2112 // fold (fp_to_sint c1fp) -> c1
2114 return DAG.getNode(ISD::FP_TO_SINT, VT, N0);
2118 SDOperand DAGCombiner::visitFP_TO_UINT(SDNode *N) {
2119 SDOperand N0 = N->getOperand(0);
2120 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2121 MVT::ValueType VT = N->getValueType(0);
2123 // fold (fp_to_uint c1fp) -> c1
2125 return DAG.getNode(ISD::FP_TO_UINT, VT, N0);
2129 SDOperand DAGCombiner::visitFP_ROUND(SDNode *N) {
2130 SDOperand N0 = N->getOperand(0);
2131 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2132 MVT::ValueType VT = N->getValueType(0);
2134 // fold (fp_round c1fp) -> c1fp
2136 return DAG.getNode(ISD::FP_ROUND, VT, N0);
2138 // fold (fp_round (fp_extend x)) -> x
2139 if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType())
2140 return N0.getOperand(0);
2142 // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y)
2143 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.Val->hasOneUse()) {
2144 SDOperand Tmp = DAG.getNode(ISD::FP_ROUND, VT, N0.getOperand(0));
2145 AddToWorkList(Tmp.Val);
2146 return DAG.getNode(ISD::FCOPYSIGN, VT, Tmp, N0.getOperand(1));
2152 SDOperand DAGCombiner::visitFP_ROUND_INREG(SDNode *N) {
2153 SDOperand N0 = N->getOperand(0);
2154 MVT::ValueType VT = N->getValueType(0);
2155 MVT::ValueType EVT = cast<VTSDNode>(N->getOperand(1))->getVT();
2156 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2158 // fold (fp_round_inreg c1fp) -> c1fp
2160 SDOperand Round = DAG.getConstantFP(N0CFP->getValue(), EVT);
2161 return DAG.getNode(ISD::FP_EXTEND, VT, Round);
2166 SDOperand DAGCombiner::visitFP_EXTEND(SDNode *N) {
2167 SDOperand N0 = N->getOperand(0);
2168 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2169 MVT::ValueType VT = N->getValueType(0);
2171 // fold (fp_extend c1fp) -> c1fp
2173 return DAG.getNode(ISD::FP_EXTEND, VT, N0);
2177 SDOperand DAGCombiner::visitFNEG(SDNode *N) {
2178 SDOperand N0 = N->getOperand(0);
2179 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2180 MVT::ValueType VT = N->getValueType(0);
2182 // fold (fneg c1) -> -c1
2184 return DAG.getNode(ISD::FNEG, VT, N0);
2185 // fold (fneg (sub x, y)) -> (sub y, x)
2186 if (N0.getOpcode() == ISD::SUB)
2187 return DAG.getNode(ISD::SUB, VT, N0.getOperand(1), N0.getOperand(0));
2188 // fold (fneg (fneg x)) -> x
2189 if (N0.getOpcode() == ISD::FNEG)
2190 return N0.getOperand(0);
2194 SDOperand DAGCombiner::visitFABS(SDNode *N) {
2195 SDOperand N0 = N->getOperand(0);
2196 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
2197 MVT::ValueType VT = N->getValueType(0);
2199 // fold (fabs c1) -> fabs(c1)
2201 return DAG.getNode(ISD::FABS, VT, N0);
2202 // fold (fabs (fabs x)) -> (fabs x)
2203 if (N0.getOpcode() == ISD::FABS)
2204 return N->getOperand(0);
2205 // fold (fabs (fneg x)) -> (fabs x)
2206 // fold (fabs (fcopysign x, y)) -> (fabs x)
2207 if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN)
2208 return DAG.getNode(ISD::FABS, VT, N0.getOperand(0));
2213 SDOperand DAGCombiner::visitBRCOND(SDNode *N) {
2214 SDOperand Chain = N->getOperand(0);
2215 SDOperand N1 = N->getOperand(1);
2216 SDOperand N2 = N->getOperand(2);
2217 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2219 // never taken branch, fold to chain
2220 if (N1C && N1C->isNullValue())
2222 // unconditional branch
2223 if (N1C && N1C->getValue() == 1)
2224 return DAG.getNode(ISD::BR, MVT::Other, Chain, N2);
2225 // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal
2227 if (N1.getOpcode() == ISD::SETCC &&
2228 TLI.isOperationLegal(ISD::BR_CC, MVT::Other)) {
2229 return DAG.getNode(ISD::BR_CC, MVT::Other, Chain, N1.getOperand(2),
2230 N1.getOperand(0), N1.getOperand(1), N2);
2235 // Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB.
2237 SDOperand DAGCombiner::visitBR_CC(SDNode *N) {
2238 CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1));
2239 SDOperand CondLHS = N->getOperand(2), CondRHS = N->getOperand(3);
2241 // Use SimplifySetCC to simplify SETCC's.
2242 SDOperand Simp = SimplifySetCC(MVT::i1, CondLHS, CondRHS, CC->get(), false);
2243 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(Simp.Val);
2245 // fold br_cc true, dest -> br dest (unconditional branch)
2246 if (SCCC && SCCC->getValue())
2247 return DAG.getNode(ISD::BR, MVT::Other, N->getOperand(0),
2249 // fold br_cc false, dest -> unconditional fall through
2250 if (SCCC && SCCC->isNullValue())
2251 return N->getOperand(0);
2252 // fold to a simpler setcc
2253 if (Simp.Val && Simp.getOpcode() == ISD::SETCC)
2254 return DAG.getNode(ISD::BR_CC, MVT::Other, N->getOperand(0),
2255 Simp.getOperand(2), Simp.getOperand(0),
2256 Simp.getOperand(1), N->getOperand(4));
2260 SDOperand DAGCombiner::visitLOAD(SDNode *N) {
2261 SDOperand Chain = N->getOperand(0);
2262 SDOperand Ptr = N->getOperand(1);
2263 SDOperand SrcValue = N->getOperand(2);
2265 // If this load is directly stored, replace the load value with the stored
2267 // TODO: Handle store large -> read small portion.
2268 // TODO: Handle TRUNCSTORE/EXTLOAD
2269 if (Chain.getOpcode() == ISD::STORE && Chain.getOperand(2) == Ptr &&
2270 Chain.getOperand(1).getValueType() == N->getValueType(0))
2271 return CombineTo(N, Chain.getOperand(1), Chain);
2276 SDOperand DAGCombiner::visitSTORE(SDNode *N) {
2277 SDOperand Chain = N->getOperand(0);
2278 SDOperand Value = N->getOperand(1);
2279 SDOperand Ptr = N->getOperand(2);
2280 SDOperand SrcValue = N->getOperand(3);
2282 // If this is a store that kills a previous store, remove the previous store.
2283 if (Chain.getOpcode() == ISD::STORE && Chain.getOperand(2) == Ptr &&
2284 Chain.Val->hasOneUse() /* Avoid introducing DAG cycles */ &&
2285 // Make sure that these stores are the same value type:
2286 // FIXME: we really care that the second store is >= size of the first.
2287 Value.getValueType() == Chain.getOperand(1).getValueType()) {
2288 // Create a new store of Value that replaces both stores.
2289 SDNode *PrevStore = Chain.Val;
2290 if (PrevStore->getOperand(1) == Value) // Same value multiply stored.
2292 SDOperand NewStore = DAG.getNode(ISD::STORE, MVT::Other,
2293 PrevStore->getOperand(0), Value, Ptr,
2295 CombineTo(N, NewStore); // Nuke this store.
2296 CombineTo(PrevStore, NewStore); // Nuke the previous store.
2297 return SDOperand(N, 0);
2300 // If this is a store of a bit convert, store the input value.
2301 // FIXME: This needs to know that the resultant store does not need a
2302 // higher alignment than the original.
2303 if (0 && Value.getOpcode() == ISD::BIT_CONVERT)
2304 return DAG.getNode(ISD::STORE, MVT::Other, Chain, Value.getOperand(0),
2310 SDOperand DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) {
2311 SDOperand InVec = N->getOperand(0);
2312 SDOperand InVal = N->getOperand(1);
2313 SDOperand EltNo = N->getOperand(2);
2315 // If the invec is a BUILD_VECTOR and if EltNo is a constant, build a new
2316 // vector with the inserted element.
2317 if (InVec.getOpcode() == ISD::BUILD_VECTOR && isa<ConstantSDNode>(EltNo)) {
2318 unsigned Elt = cast<ConstantSDNode>(EltNo)->getValue();
2319 std::vector<SDOperand> Ops(InVec.Val->op_begin(), InVec.Val->op_end());
2320 if (Elt < Ops.size())
2322 return DAG.getNode(ISD::BUILD_VECTOR, InVec.getValueType(), Ops);
2328 SDOperand DAGCombiner::visitVINSERT_VECTOR_ELT(SDNode *N) {
2329 SDOperand InVec = N->getOperand(0);
2330 SDOperand InVal = N->getOperand(1);
2331 SDOperand EltNo = N->getOperand(2);
2332 SDOperand NumElts = N->getOperand(3);
2333 SDOperand EltType = N->getOperand(4);
2335 // If the invec is a VBUILD_VECTOR and if EltNo is a constant, build a new
2336 // vector with the inserted element.
2337 if (InVec.getOpcode() == ISD::VBUILD_VECTOR && isa<ConstantSDNode>(EltNo)) {
2338 unsigned Elt = cast<ConstantSDNode>(EltNo)->getValue();
2339 std::vector<SDOperand> Ops(InVec.Val->op_begin(), InVec.Val->op_end());
2340 if (Elt < Ops.size()-2)
2342 return DAG.getNode(ISD::VBUILD_VECTOR, InVec.getValueType(), Ops);
2348 SDOperand DAGCombiner::visitVBUILD_VECTOR(SDNode *N) {
2349 unsigned NumInScalars = N->getNumOperands()-2;
2350 SDOperand NumElts = N->getOperand(NumInScalars);
2351 SDOperand EltType = N->getOperand(NumInScalars+1);
2353 // Check to see if this is a VBUILD_VECTOR of a bunch of VEXTRACT_VECTOR_ELT
2354 // operations. If so, and if the EXTRACT_ELT vector inputs come from at most
2355 // two distinct vectors, turn this into a shuffle node.
2356 SDOperand VecIn1, VecIn2;
2357 for (unsigned i = 0; i != NumInScalars; ++i) {
2358 // Ignore undef inputs.
2359 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
2361 // If this input is something other than a VEXTRACT_VECTOR_ELT with a
2362 // constant index, bail out.
2363 if (N->getOperand(i).getOpcode() != ISD::VEXTRACT_VECTOR_ELT ||
2364 !isa<ConstantSDNode>(N->getOperand(i).getOperand(1))) {
2365 VecIn1 = VecIn2 = SDOperand(0, 0);
2369 // If the input vector type disagrees with the result of the vbuild_vector,
2370 // we can't make a shuffle.
2371 SDOperand ExtractedFromVec = N->getOperand(i).getOperand(0);
2372 if (*(ExtractedFromVec.Val->op_end()-2) != NumElts ||
2373 *(ExtractedFromVec.Val->op_end()-1) != EltType) {
2374 VecIn1 = VecIn2 = SDOperand(0, 0);
2378 // Otherwise, remember this. We allow up to two distinct input vectors.
2379 if (ExtractedFromVec == VecIn1 || ExtractedFromVec == VecIn2)
2382 if (VecIn1.Val == 0) {
2383 VecIn1 = ExtractedFromVec;
2384 } else if (VecIn2.Val == 0) {
2385 VecIn2 = ExtractedFromVec;
2388 VecIn1 = VecIn2 = SDOperand(0, 0);
2393 // If everything is good, we can make a shuffle operation.
2395 std::vector<SDOperand> BuildVecIndices;
2396 for (unsigned i = 0; i != NumInScalars; ++i) {
2397 if (N->getOperand(i).getOpcode() == ISD::UNDEF) {
2398 BuildVecIndices.push_back(DAG.getNode(ISD::UNDEF, MVT::i32));
2402 SDOperand Extract = N->getOperand(i);
2404 // If extracting from the first vector, just use the index directly.
2405 if (Extract.getOperand(0) == VecIn1) {
2406 BuildVecIndices.push_back(Extract.getOperand(1));
2410 // Otherwise, use InIdx + VecSize
2411 unsigned Idx = cast<ConstantSDNode>(Extract.getOperand(1))->getValue();
2412 BuildVecIndices.push_back(DAG.getConstant(Idx+NumInScalars, MVT::i32));
2415 // Add count and size info.
2416 BuildVecIndices.push_back(NumElts);
2417 BuildVecIndices.push_back(DAG.getValueType(MVT::i32));
2419 // Return the new VVECTOR_SHUFFLE node.
2420 std::vector<SDOperand> Ops;
2421 Ops.push_back(VecIn1);
2423 Ops.push_back(VecIn2);
2425 // Use an undef vbuild_vector as input for the second operand.
2426 std::vector<SDOperand> UnOps(NumInScalars,
2427 DAG.getNode(ISD::UNDEF,
2428 cast<VTSDNode>(EltType)->getVT()));
2429 UnOps.push_back(NumElts);
2430 UnOps.push_back(EltType);
2431 Ops.push_back(DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, UnOps));
2433 Ops.push_back(DAG.getNode(ISD::VBUILD_VECTOR,MVT::Vector, BuildVecIndices));
2434 Ops.push_back(NumElts);
2435 Ops.push_back(EltType);
2436 return DAG.getNode(ISD::VVECTOR_SHUFFLE, MVT::Vector, Ops);
2442 SDOperand DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
2443 // If the LHS and the RHS are the same node, turn the RHS into an undef.
2444 if (N->getOperand(0) == N->getOperand(1)) {
2445 // Check the SHUFFLE mask, mapping any inputs from the 2nd operand into the
2447 std::vector<SDOperand> MappedOps;
2448 SDOperand ShufMask = N->getOperand(2);
2449 unsigned NumElts = ShufMask.getNumOperands();
2450 for (unsigned i = 0, e = ShufMask.getNumOperands(); i != e; ++i) {
2451 if (cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() >= NumElts) {
2453 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() - NumElts;
2454 MappedOps.push_back(DAG.getConstant(NewIdx, MVT::i32));
2456 MappedOps.push_back(ShufMask.getOperand(i));
2459 ShufMask = DAG.getNode(ISD::BUILD_VECTOR, ShufMask.getValueType(),
2461 return DAG.getNode(ISD::VECTOR_SHUFFLE, N->getValueType(0),
2463 DAG.getNode(ISD::UNDEF, N->getValueType(0)),
2470 SDOperand DAGCombiner::SimplifySelect(SDOperand N0, SDOperand N1, SDOperand N2){
2471 assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!");
2473 SDOperand SCC = SimplifySelectCC(N0.getOperand(0), N0.getOperand(1), N1, N2,
2474 cast<CondCodeSDNode>(N0.getOperand(2))->get());
2475 // If we got a simplified select_cc node back from SimplifySelectCC, then
2476 // break it down into a new SETCC node, and a new SELECT node, and then return
2477 // the SELECT node, since we were called with a SELECT node.
2479 // Check to see if we got a select_cc back (to turn into setcc/select).
2480 // Otherwise, just return whatever node we got back, like fabs.
2481 if (SCC.getOpcode() == ISD::SELECT_CC) {
2482 SDOperand SETCC = DAG.getNode(ISD::SETCC, N0.getValueType(),
2483 SCC.getOperand(0), SCC.getOperand(1),
2485 AddToWorkList(SETCC.Val);
2486 return DAG.getNode(ISD::SELECT, SCC.getValueType(), SCC.getOperand(2),
2487 SCC.getOperand(3), SETCC);
2494 /// SimplifySelectOps - Given a SELECT or a SELECT_CC node, where LHS and RHS
2495 /// are the two values being selected between, see if we can simplify the
2498 bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDOperand LHS,
2501 // If this is a select from two identical things, try to pull the operation
2502 // through the select.
2503 if (LHS.getOpcode() == RHS.getOpcode() && LHS.hasOneUse() && RHS.hasOneUse()){
2505 std::cerr << "SELECT: ["; LHS.Val->dump();
2506 std::cerr << "] ["; RHS.Val->dump();
2510 // If this is a load and the token chain is identical, replace the select
2511 // of two loads with a load through a select of the address to load from.
2512 // This triggers in things like "select bool X, 10.0, 123.0" after the FP
2513 // constants have been dropped into the constant pool.
2514 if ((LHS.getOpcode() == ISD::LOAD ||
2515 LHS.getOpcode() == ISD::EXTLOAD ||
2516 LHS.getOpcode() == ISD::ZEXTLOAD ||
2517 LHS.getOpcode() == ISD::SEXTLOAD) &&
2518 // Token chains must be identical.
2519 LHS.getOperand(0) == RHS.getOperand(0) &&
2520 // If this is an EXTLOAD, the VT's must match.
2521 (LHS.getOpcode() == ISD::LOAD ||
2522 LHS.getOperand(3) == RHS.getOperand(3))) {
2523 // FIXME: this conflates two src values, discarding one. This is not
2524 // the right thing to do, but nothing uses srcvalues now. When they do,
2525 // turn SrcValue into a list of locations.
2527 if (TheSelect->getOpcode() == ISD::SELECT)
2528 Addr = DAG.getNode(ISD::SELECT, LHS.getOperand(1).getValueType(),
2529 TheSelect->getOperand(0), LHS.getOperand(1),
2532 Addr = DAG.getNode(ISD::SELECT_CC, LHS.getOperand(1).getValueType(),
2533 TheSelect->getOperand(0),
2534 TheSelect->getOperand(1),
2535 LHS.getOperand(1), RHS.getOperand(1),
2536 TheSelect->getOperand(4));
2539 if (LHS.getOpcode() == ISD::LOAD)
2540 Load = DAG.getLoad(TheSelect->getValueType(0), LHS.getOperand(0),
2541 Addr, LHS.getOperand(2));
2543 Load = DAG.getExtLoad(LHS.getOpcode(), TheSelect->getValueType(0),
2544 LHS.getOperand(0), Addr, LHS.getOperand(2),
2545 cast<VTSDNode>(LHS.getOperand(3))->getVT());
2546 // Users of the select now use the result of the load.
2547 CombineTo(TheSelect, Load);
2549 // Users of the old loads now use the new load's chain. We know the
2550 // old-load value is dead now.
2551 CombineTo(LHS.Val, Load.getValue(0), Load.getValue(1));
2552 CombineTo(RHS.Val, Load.getValue(0), Load.getValue(1));
2560 SDOperand DAGCombiner::SimplifySelectCC(SDOperand N0, SDOperand N1,
2561 SDOperand N2, SDOperand N3,
2564 MVT::ValueType VT = N2.getValueType();
2565 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val);
2566 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val);
2567 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.Val);
2568 ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N3.Val);
2570 // Determine if the condition we're dealing with is constant
2571 SDOperand SCC = SimplifySetCC(TLI.getSetCCResultTy(), N0, N1, CC, false);
2572 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.Val);
2574 // fold select_cc true, x, y -> x
2575 if (SCCC && SCCC->getValue())
2577 // fold select_cc false, x, y -> y
2578 if (SCCC && SCCC->getValue() == 0)
2581 // Check to see if we can simplify the select into an fabs node
2582 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) {
2583 // Allow either -0.0 or 0.0
2584 if (CFP->getValue() == 0.0) {
2585 // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs
2586 if ((CC == ISD::SETGE || CC == ISD::SETGT) &&
2587 N0 == N2 && N3.getOpcode() == ISD::FNEG &&
2588 N2 == N3.getOperand(0))
2589 return DAG.getNode(ISD::FABS, VT, N0);
2591 // select (setl[te] X, +/-0.0), fneg(X), X -> fabs
2592 if ((CC == ISD::SETLT || CC == ISD::SETLE) &&
2593 N0 == N3 && N2.getOpcode() == ISD::FNEG &&
2594 N2.getOperand(0) == N3)
2595 return DAG.getNode(ISD::FABS, VT, N3);
2599 // Check to see if we can perform the "gzip trick", transforming
2600 // select_cc setlt X, 0, A, 0 -> and (sra X, size(X)-1), A
2601 if (N1C && N1C->isNullValue() && N3C && N3C->isNullValue() &&
2602 MVT::isInteger(N0.getValueType()) &&
2603 MVT::isInteger(N2.getValueType()) && CC == ISD::SETLT) {
2604 MVT::ValueType XType = N0.getValueType();
2605 MVT::ValueType AType = N2.getValueType();
2606 if (XType >= AType) {
2607 // and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a
2608 // single-bit constant.
2609 if (N2C && ((N2C->getValue() & (N2C->getValue()-1)) == 0)) {
2610 unsigned ShCtV = Log2_64(N2C->getValue());
2611 ShCtV = MVT::getSizeInBits(XType)-ShCtV-1;
2612 SDOperand ShCt = DAG.getConstant(ShCtV, TLI.getShiftAmountTy());
2613 SDOperand Shift = DAG.getNode(ISD::SRL, XType, N0, ShCt);
2614 AddToWorkList(Shift.Val);
2615 if (XType > AType) {
2616 Shift = DAG.getNode(ISD::TRUNCATE, AType, Shift);
2617 AddToWorkList(Shift.Val);
2619 return DAG.getNode(ISD::AND, AType, Shift, N2);
2621 SDOperand Shift = DAG.getNode(ISD::SRA, XType, N0,
2622 DAG.getConstant(MVT::getSizeInBits(XType)-1,
2623 TLI.getShiftAmountTy()));
2624 AddToWorkList(Shift.Val);
2625 if (XType > AType) {
2626 Shift = DAG.getNode(ISD::TRUNCATE, AType, Shift);
2627 AddToWorkList(Shift.Val);
2629 return DAG.getNode(ISD::AND, AType, Shift, N2);
2633 // fold select C, 16, 0 -> shl C, 4
2634 if (N2C && N3C && N3C->isNullValue() && isPowerOf2_64(N2C->getValue()) &&
2635 TLI.getSetCCResultContents() == TargetLowering::ZeroOrOneSetCCResult) {
2636 // Get a SetCC of the condition
2637 // FIXME: Should probably make sure that setcc is legal if we ever have a
2638 // target where it isn't.
2639 SDOperand Temp, SCC;
2640 // cast from setcc result type to select result type
2641 if (AfterLegalize) {
2642 SCC = DAG.getSetCC(TLI.getSetCCResultTy(), N0, N1, CC);
2643 Temp = DAG.getZeroExtendInReg(SCC, N2.getValueType());
2645 SCC = DAG.getSetCC(MVT::i1, N0, N1, CC);
2646 Temp = DAG.getNode(ISD::ZERO_EXTEND, N2.getValueType(), SCC);
2648 AddToWorkList(SCC.Val);
2649 AddToWorkList(Temp.Val);
2650 // shl setcc result by log2 n2c
2651 return DAG.getNode(ISD::SHL, N2.getValueType(), Temp,
2652 DAG.getConstant(Log2_64(N2C->getValue()),
2653 TLI.getShiftAmountTy()));
2656 // Check to see if this is the equivalent of setcc
2657 // FIXME: Turn all of these into setcc if setcc if setcc is legal
2658 // otherwise, go ahead with the folds.
2659 if (0 && N3C && N3C->isNullValue() && N2C && (N2C->getValue() == 1ULL)) {
2660 MVT::ValueType XType = N0.getValueType();
2661 if (TLI.isOperationLegal(ISD::SETCC, TLI.getSetCCResultTy())) {
2662 SDOperand Res = DAG.getSetCC(TLI.getSetCCResultTy(), N0, N1, CC);
2663 if (Res.getValueType() != VT)
2664 Res = DAG.getNode(ISD::ZERO_EXTEND, VT, Res);
2668 // seteq X, 0 -> srl (ctlz X, log2(size(X)))
2669 if (N1C && N1C->isNullValue() && CC == ISD::SETEQ &&
2670 TLI.isOperationLegal(ISD::CTLZ, XType)) {
2671 SDOperand Ctlz = DAG.getNode(ISD::CTLZ, XType, N0);
2672 return DAG.getNode(ISD::SRL, XType, Ctlz,
2673 DAG.getConstant(Log2_32(MVT::getSizeInBits(XType)),
2674 TLI.getShiftAmountTy()));
2676 // setgt X, 0 -> srl (and (-X, ~X), size(X)-1)
2677 if (N1C && N1C->isNullValue() && CC == ISD::SETGT) {
2678 SDOperand NegN0 = DAG.getNode(ISD::SUB, XType, DAG.getConstant(0, XType),
2680 SDOperand NotN0 = DAG.getNode(ISD::XOR, XType, N0,
2681 DAG.getConstant(~0ULL, XType));
2682 return DAG.getNode(ISD::SRL, XType,
2683 DAG.getNode(ISD::AND, XType, NegN0, NotN0),
2684 DAG.getConstant(MVT::getSizeInBits(XType)-1,
2685 TLI.getShiftAmountTy()));
2687 // setgt X, -1 -> xor (srl (X, size(X)-1), 1)
2688 if (N1C && N1C->isAllOnesValue() && CC == ISD::SETGT) {
2689 SDOperand Sign = DAG.getNode(ISD::SRL, XType, N0,
2690 DAG.getConstant(MVT::getSizeInBits(XType)-1,
2691 TLI.getShiftAmountTy()));
2692 return DAG.getNode(ISD::XOR, XType, Sign, DAG.getConstant(1, XType));
2696 // Check to see if this is an integer abs. select_cc setl[te] X, 0, -X, X ->
2697 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
2698 if (N1C && N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE) &&
2699 N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1)) {
2700 if (ConstantSDNode *SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0))) {
2701 MVT::ValueType XType = N0.getValueType();
2702 if (SubC->isNullValue() && MVT::isInteger(XType)) {
2703 SDOperand Shift = DAG.getNode(ISD::SRA, XType, N0,
2704 DAG.getConstant(MVT::getSizeInBits(XType)-1,
2705 TLI.getShiftAmountTy()));
2706 SDOperand Add = DAG.getNode(ISD::ADD, XType, N0, Shift);
2707 AddToWorkList(Shift.Val);
2708 AddToWorkList(Add.Val);
2709 return DAG.getNode(ISD::XOR, XType, Add, Shift);
2717 SDOperand DAGCombiner::SimplifySetCC(MVT::ValueType VT, SDOperand N0,
2718 SDOperand N1, ISD::CondCode Cond,
2719 bool foldBooleans) {
2720 // These setcc operations always fold.
2724 case ISD::SETFALSE2: return DAG.getConstant(0, VT);
2726 case ISD::SETTRUE2: return DAG.getConstant(1, VT);
2729 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val)) {
2730 uint64_t C1 = N1C->getValue();
2731 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val)) {
2732 uint64_t C0 = N0C->getValue();
2734 // Sign extend the operands if required
2735 if (ISD::isSignedIntSetCC(Cond)) {
2736 C0 = N0C->getSignExtended();
2737 C1 = N1C->getSignExtended();
2741 default: assert(0 && "Unknown integer setcc!");
2742 case ISD::SETEQ: return DAG.getConstant(C0 == C1, VT);
2743 case ISD::SETNE: return DAG.getConstant(C0 != C1, VT);
2744 case ISD::SETULT: return DAG.getConstant(C0 < C1, VT);
2745 case ISD::SETUGT: return DAG.getConstant(C0 > C1, VT);
2746 case ISD::SETULE: return DAG.getConstant(C0 <= C1, VT);
2747 case ISD::SETUGE: return DAG.getConstant(C0 >= C1, VT);
2748 case ISD::SETLT: return DAG.getConstant((int64_t)C0 < (int64_t)C1, VT);
2749 case ISD::SETGT: return DAG.getConstant((int64_t)C0 > (int64_t)C1, VT);
2750 case ISD::SETLE: return DAG.getConstant((int64_t)C0 <= (int64_t)C1, VT);
2751 case ISD::SETGE: return DAG.getConstant((int64_t)C0 >= (int64_t)C1, VT);
2754 // If the LHS is a ZERO_EXTEND, perform the comparison on the input.
2755 if (N0.getOpcode() == ISD::ZERO_EXTEND) {
2756 unsigned InSize = MVT::getSizeInBits(N0.getOperand(0).getValueType());
2758 // If the comparison constant has bits in the upper part, the
2759 // zero-extended value could never match.
2760 if (C1 & (~0ULL << InSize)) {
2761 unsigned VSize = MVT::getSizeInBits(N0.getValueType());
2765 case ISD::SETEQ: return DAG.getConstant(0, VT);
2768 case ISD::SETNE: return DAG.getConstant(1, VT);
2771 // True if the sign bit of C1 is set.
2772 return DAG.getConstant((C1 & (1ULL << VSize)) != 0, VT);
2775 // True if the sign bit of C1 isn't set.
2776 return DAG.getConstant((C1 & (1ULL << VSize)) == 0, VT);
2782 // Otherwise, we can perform the comparison with the low bits.
2790 return DAG.getSetCC(VT, N0.getOperand(0),
2791 DAG.getConstant(C1, N0.getOperand(0).getValueType()),
2794 break; // todo, be more careful with signed comparisons
2796 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
2797 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) {
2798 MVT::ValueType ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT();
2799 unsigned ExtSrcTyBits = MVT::getSizeInBits(ExtSrcTy);
2800 MVT::ValueType ExtDstTy = N0.getValueType();
2801 unsigned ExtDstTyBits = MVT::getSizeInBits(ExtDstTy);
2803 // If the extended part has any inconsistent bits, it cannot ever
2804 // compare equal. In other words, they have to be all ones or all
2807 (~0ULL >> (64-ExtSrcTyBits)) & (~0ULL << (ExtDstTyBits-1));
2808 if ((C1 & ExtBits) != 0 && (C1 & ExtBits) != ExtBits)
2809 return DAG.getConstant(Cond == ISD::SETNE, VT);
2812 MVT::ValueType Op0Ty = N0.getOperand(0).getValueType();
2813 if (Op0Ty == ExtSrcTy) {
2814 ZextOp = N0.getOperand(0);
2816 int64_t Imm = ~0ULL >> (64-ExtSrcTyBits);
2817 ZextOp = DAG.getNode(ISD::AND, Op0Ty, N0.getOperand(0),
2818 DAG.getConstant(Imm, Op0Ty));
2820 AddToWorkList(ZextOp.Val);
2821 // Otherwise, make this a use of a zext.
2822 return DAG.getSetCC(VT, ZextOp,
2823 DAG.getConstant(C1 & (~0ULL>>(64-ExtSrcTyBits)),
2826 } else if ((N1C->getValue() == 0 || N1C->getValue() == 1) &&
2827 (Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
2828 (N0.getOpcode() == ISD::XOR ||
2829 (N0.getOpcode() == ISD::AND &&
2830 N0.getOperand(0).getOpcode() == ISD::XOR &&
2831 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) &&
2832 isa<ConstantSDNode>(N0.getOperand(1)) &&
2833 cast<ConstantSDNode>(N0.getOperand(1))->getValue() == 1) {
2834 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We can
2835 // only do this if the top bits are known zero.
2836 if (TLI.MaskedValueIsZero(N1,
2837 MVT::getIntVTBitMask(N0.getValueType())-1)) {
2838 // Okay, get the un-inverted input value.
2840 if (N0.getOpcode() == ISD::XOR)
2841 Val = N0.getOperand(0);
2843 assert(N0.getOpcode() == ISD::AND &&
2844 N0.getOperand(0).getOpcode() == ISD::XOR);
2845 // ((X^1)&1)^1 -> X & 1
2846 Val = DAG.getNode(ISD::AND, N0.getValueType(),
2847 N0.getOperand(0).getOperand(0), N0.getOperand(1));
2849 return DAG.getSetCC(VT, Val, N1,
2850 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
2854 uint64_t MinVal, MaxVal;
2855 unsigned OperandBitSize = MVT::getSizeInBits(N1C->getValueType(0));
2856 if (ISD::isSignedIntSetCC(Cond)) {
2857 MinVal = 1ULL << (OperandBitSize-1);
2858 if (OperandBitSize != 1) // Avoid X >> 64, which is undefined.
2859 MaxVal = ~0ULL >> (65-OperandBitSize);
2864 MaxVal = ~0ULL >> (64-OperandBitSize);
2867 // Canonicalize GE/LE comparisons to use GT/LT comparisons.
2868 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) {
2869 if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true
2870 --C1; // X >= C0 --> X > (C0-1)
2871 return DAG.getSetCC(VT, N0, DAG.getConstant(C1, N1.getValueType()),
2872 (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT);
2875 if (Cond == ISD::SETLE || Cond == ISD::SETULE) {
2876 if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true
2877 ++C1; // X <= C0 --> X < (C0+1)
2878 return DAG.getSetCC(VT, N0, DAG.getConstant(C1, N1.getValueType()),
2879 (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT);
2882 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal)
2883 return DAG.getConstant(0, VT); // X < MIN --> false
2885 // Canonicalize setgt X, Min --> setne X, Min
2886 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal)
2887 return DAG.getSetCC(VT, N0, N1, ISD::SETNE);
2888 // Canonicalize setlt X, Max --> setne X, Max
2889 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal)
2890 return DAG.getSetCC(VT, N0, N1, ISD::SETNE);
2892 // If we have setult X, 1, turn it into seteq X, 0
2893 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1)
2894 return DAG.getSetCC(VT, N0, DAG.getConstant(MinVal, N0.getValueType()),
2896 // If we have setugt X, Max-1, turn it into seteq X, Max
2897 else if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1)
2898 return DAG.getSetCC(VT, N0, DAG.getConstant(MaxVal, N0.getValueType()),
2901 // If we have "setcc X, C0", check to see if we can shrink the immediate
2904 // SETUGT X, SINTMAX -> SETLT X, 0
2905 if (Cond == ISD::SETUGT && OperandBitSize != 1 &&
2906 C1 == (~0ULL >> (65-OperandBitSize)))
2907 return DAG.getSetCC(VT, N0, DAG.getConstant(0, N1.getValueType()),
2910 // FIXME: Implement the rest of these.
2912 // Fold bit comparisons when we can.
2913 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
2914 VT == N0.getValueType() && N0.getOpcode() == ISD::AND)
2915 if (ConstantSDNode *AndRHS =
2916 dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
2917 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3
2918 // Perform the xform if the AND RHS is a single bit.
2919 if ((AndRHS->getValue() & (AndRHS->getValue()-1)) == 0) {
2920 return DAG.getNode(ISD::SRL, VT, N0,
2921 DAG.getConstant(Log2_64(AndRHS->getValue()),
2922 TLI.getShiftAmountTy()));
2924 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getValue()) {
2925 // (X & 8) == 8 --> (X & 8) >> 3
2926 // Perform the xform if C1 is a single bit.
2927 if ((C1 & (C1-1)) == 0) {
2928 return DAG.getNode(ISD::SRL, VT, N0,
2929 DAG.getConstant(Log2_64(C1),TLI.getShiftAmountTy()));
2934 } else if (isa<ConstantSDNode>(N0.Val)) {
2935 // Ensure that the constant occurs on the RHS.
2936 return DAG.getSetCC(VT, N1, N0, ISD::getSetCCSwappedOperands(Cond));
2939 if (ConstantFPSDNode *N0C = dyn_cast<ConstantFPSDNode>(N0.Val))
2940 if (ConstantFPSDNode *N1C = dyn_cast<ConstantFPSDNode>(N1.Val)) {
2941 double C0 = N0C->getValue(), C1 = N1C->getValue();
2944 default: break; // FIXME: Implement the rest of these!
2945 case ISD::SETEQ: return DAG.getConstant(C0 == C1, VT);
2946 case ISD::SETNE: return DAG.getConstant(C0 != C1, VT);
2947 case ISD::SETLT: return DAG.getConstant(C0 < C1, VT);
2948 case ISD::SETGT: return DAG.getConstant(C0 > C1, VT);
2949 case ISD::SETLE: return DAG.getConstant(C0 <= C1, VT);
2950 case ISD::SETGE: return DAG.getConstant(C0 >= C1, VT);
2953 // Ensure that the constant occurs on the RHS.
2954 return DAG.getSetCC(VT, N1, N0, ISD::getSetCCSwappedOperands(Cond));
2958 // We can always fold X == Y for integer setcc's.
2959 if (MVT::isInteger(N0.getValueType()))
2960 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
2961 unsigned UOF = ISD::getUnorderedFlavor(Cond);
2962 if (UOF == 2) // FP operators that are undefined on NaNs.
2963 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT);
2964 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond)))
2965 return DAG.getConstant(UOF, VT);
2966 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO
2967 // if it is not already.
2968 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO;
2969 if (NewCond != Cond)
2970 return DAG.getSetCC(VT, N0, N1, NewCond);
2973 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
2974 MVT::isInteger(N0.getValueType())) {
2975 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB ||
2976 N0.getOpcode() == ISD::XOR) {
2977 // Simplify (X+Y) == (X+Z) --> Y == Z
2978 if (N0.getOpcode() == N1.getOpcode()) {
2979 if (N0.getOperand(0) == N1.getOperand(0))
2980 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(1), Cond);
2981 if (N0.getOperand(1) == N1.getOperand(1))
2982 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(0), Cond);
2983 if (isCommutativeBinOp(N0.getOpcode())) {
2984 // If X op Y == Y op X, try other combinations.
2985 if (N0.getOperand(0) == N1.getOperand(1))
2986 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(0), Cond);
2987 if (N0.getOperand(1) == N1.getOperand(0))
2988 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(1), Cond);
2992 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) {
2993 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
2994 // Turn (X+C1) == C2 --> X == C2-C1
2995 if (N0.getOpcode() == ISD::ADD && N0.Val->hasOneUse()) {
2996 return DAG.getSetCC(VT, N0.getOperand(0),
2997 DAG.getConstant(RHSC->getValue()-LHSR->getValue(),
2998 N0.getValueType()), Cond);
3001 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0.
3002 if (N0.getOpcode() == ISD::XOR)
3003 // If we know that all of the inverted bits are zero, don't bother
3004 // performing the inversion.
3005 if (TLI.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getValue()))
3006 return DAG.getSetCC(VT, N0.getOperand(0),
3007 DAG.getConstant(LHSR->getValue()^RHSC->getValue(),
3008 N0.getValueType()), Cond);
3011 // Turn (C1-X) == C2 --> X == C1-C2
3012 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) {
3013 if (N0.getOpcode() == ISD::SUB && N0.Val->hasOneUse()) {
3014 return DAG.getSetCC(VT, N0.getOperand(1),
3015 DAG.getConstant(SUBC->getValue()-RHSC->getValue(),
3016 N0.getValueType()), Cond);
3021 // Simplify (X+Z) == X --> Z == 0
3022 if (N0.getOperand(0) == N1)
3023 return DAG.getSetCC(VT, N0.getOperand(1),
3024 DAG.getConstant(0, N0.getValueType()), Cond);
3025 if (N0.getOperand(1) == N1) {
3026 if (isCommutativeBinOp(N0.getOpcode()))
3027 return DAG.getSetCC(VT, N0.getOperand(0),
3028 DAG.getConstant(0, N0.getValueType()), Cond);
3030 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!");
3031 // (Z-X) == X --> Z == X<<1
3032 SDOperand SH = DAG.getNode(ISD::SHL, N1.getValueType(),
3034 DAG.getConstant(1,TLI.getShiftAmountTy()));
3035 AddToWorkList(SH.Val);
3036 return DAG.getSetCC(VT, N0.getOperand(0), SH, Cond);
3041 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB ||
3042 N1.getOpcode() == ISD::XOR) {
3043 // Simplify X == (X+Z) --> Z == 0
3044 if (N1.getOperand(0) == N0) {
3045 return DAG.getSetCC(VT, N1.getOperand(1),
3046 DAG.getConstant(0, N1.getValueType()), Cond);
3047 } else if (N1.getOperand(1) == N0) {
3048 if (isCommutativeBinOp(N1.getOpcode())) {
3049 return DAG.getSetCC(VT, N1.getOperand(0),
3050 DAG.getConstant(0, N1.getValueType()), Cond);
3052 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!");
3053 // X == (Z-X) --> X<<1 == Z
3054 SDOperand SH = DAG.getNode(ISD::SHL, N1.getValueType(), N0,
3055 DAG.getConstant(1,TLI.getShiftAmountTy()));
3056 AddToWorkList(SH.Val);
3057 return DAG.getSetCC(VT, SH, N1.getOperand(0), Cond);
3063 // Fold away ALL boolean setcc's.
3065 if (N0.getValueType() == MVT::i1 && foldBooleans) {
3067 default: assert(0 && "Unknown integer setcc!");
3068 case ISD::SETEQ: // X == Y -> (X^Y)^1
3069 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, N1);
3070 N0 = DAG.getNode(ISD::XOR, MVT::i1, Temp, DAG.getConstant(1, MVT::i1));
3071 AddToWorkList(Temp.Val);
3073 case ISD::SETNE: // X != Y --> (X^Y)
3074 N0 = DAG.getNode(ISD::XOR, MVT::i1, N0, N1);
3076 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> X^1 & Y
3077 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> X^1 & Y
3078 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1));
3079 N0 = DAG.getNode(ISD::AND, MVT::i1, N1, Temp);
3080 AddToWorkList(Temp.Val);
3082 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> Y^1 & X
3083 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> Y^1 & X
3084 Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1));
3085 N0 = DAG.getNode(ISD::AND, MVT::i1, N0, Temp);
3086 AddToWorkList(Temp.Val);
3088 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> X^1 | Y
3089 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> X^1 | Y
3090 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1));
3091 N0 = DAG.getNode(ISD::OR, MVT::i1, N1, Temp);
3092 AddToWorkList(Temp.Val);
3094 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> Y^1 | X
3095 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> Y^1 | X
3096 Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1));
3097 N0 = DAG.getNode(ISD::OR, MVT::i1, N0, Temp);
3100 if (VT != MVT::i1) {
3101 AddToWorkList(N0.Val);
3102 // FIXME: If running after legalize, we probably can't do this.
3103 N0 = DAG.getNode(ISD::ZERO_EXTEND, VT, N0);
3108 // Could not fold it.
3112 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
3113 /// return a DAG expression to select that will generate the same value by
3114 /// multiplying by a magic number. See:
3115 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
3116 SDOperand DAGCombiner::BuildSDIV(SDNode *N) {
3117 MVT::ValueType VT = N->getValueType(0);
3119 // Check to see if we can do this.
3120 if (!TLI.isTypeLegal(VT) || (VT != MVT::i32 && VT != MVT::i64))
3121 return SDOperand(); // BuildSDIV only operates on i32 or i64
3122 if (!TLI.isOperationLegal(ISD::MULHS, VT))
3123 return SDOperand(); // Make sure the target supports MULHS.
3125 int64_t d = cast<ConstantSDNode>(N->getOperand(1))->getSignExtended();
3126 ms magics = (VT == MVT::i32) ? magic32(d) : magic64(d);
3128 // Multiply the numerator (operand 0) by the magic value
3129 SDOperand Q = DAG.getNode(ISD::MULHS, VT, N->getOperand(0),
3130 DAG.getConstant(magics.m, VT));
3131 // If d > 0 and m < 0, add the numerator
3132 if (d > 0 && magics.m < 0) {
3133 Q = DAG.getNode(ISD::ADD, VT, Q, N->getOperand(0));
3134 AddToWorkList(Q.Val);
3136 // If d < 0 and m > 0, subtract the numerator.
3137 if (d < 0 && magics.m > 0) {
3138 Q = DAG.getNode(ISD::SUB, VT, Q, N->getOperand(0));
3139 AddToWorkList(Q.Val);
3141 // Shift right algebraic if shift value is nonzero
3143 Q = DAG.getNode(ISD::SRA, VT, Q,
3144 DAG.getConstant(magics.s, TLI.getShiftAmountTy()));
3145 AddToWorkList(Q.Val);
3147 // Extract the sign bit and add it to the quotient
3149 DAG.getNode(ISD::SRL, VT, Q, DAG.getConstant(MVT::getSizeInBits(VT)-1,
3150 TLI.getShiftAmountTy()));
3151 AddToWorkList(T.Val);
3152 return DAG.getNode(ISD::ADD, VT, Q, T);
3155 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
3156 /// return a DAG expression to select that will generate the same value by
3157 /// multiplying by a magic number. See:
3158 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
3159 SDOperand DAGCombiner::BuildUDIV(SDNode *N) {
3160 MVT::ValueType VT = N->getValueType(0);
3162 // Check to see if we can do this.
3163 if (!TLI.isTypeLegal(VT) || (VT != MVT::i32 && VT != MVT::i64))
3164 return SDOperand(); // BuildUDIV only operates on i32 or i64
3165 if (!TLI.isOperationLegal(ISD::MULHU, VT))
3166 return SDOperand(); // Make sure the target supports MULHU.
3168 uint64_t d = cast<ConstantSDNode>(N->getOperand(1))->getValue();
3169 mu magics = (VT == MVT::i32) ? magicu32(d) : magicu64(d);
3171 // Multiply the numerator (operand 0) by the magic value
3172 SDOperand Q = DAG.getNode(ISD::MULHU, VT, N->getOperand(0),
3173 DAG.getConstant(magics.m, VT));
3174 AddToWorkList(Q.Val);
3176 if (magics.a == 0) {
3177 return DAG.getNode(ISD::SRL, VT, Q,
3178 DAG.getConstant(magics.s, TLI.getShiftAmountTy()));
3180 SDOperand NPQ = DAG.getNode(ISD::SUB, VT, N->getOperand(0), Q);
3181 AddToWorkList(NPQ.Val);
3182 NPQ = DAG.getNode(ISD::SRL, VT, NPQ,
3183 DAG.getConstant(1, TLI.getShiftAmountTy()));
3184 AddToWorkList(NPQ.Val);
3185 NPQ = DAG.getNode(ISD::ADD, VT, NPQ, Q);
3186 AddToWorkList(NPQ.Val);
3187 return DAG.getNode(ISD::SRL, VT, NPQ,
3188 DAG.getConstant(magics.s-1, TLI.getShiftAmountTy()));
3192 // SelectionDAG::Combine - This is the entry point for the file.
3194 void SelectionDAG::Combine(bool RunningAfterLegalize) {
3195 /// run - This is the main entry point to this class.
3197 DAGCombiner(*this).Run(RunningAfterLegalize);