DAGCombiner.cpp revision 60e8c71c9f63959890ecabb70c6e2cde2f947224
1//===-- DAGCombiner.cpp - Implement a DAG node combiner -------------------===// 2// 3// The LLVM Compiler Infrastructure 4// 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. 7// 8//===----------------------------------------------------------------------===// 9// 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. 12// 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. 18// 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 27// into an undef? 28// FIXME: select ne (select cc, 1, 0), 0, true, false -> select cc, true, false 29// 30//===----------------------------------------------------------------------===// 31 32#define DEBUG_TYPE "dagcombine" 33#include "llvm/ADT/Statistic.h" 34#include "llvm/CodeGen/SelectionDAG.h" 35#include "llvm/Support/Debug.h" 36#include "llvm/Support/MathExtras.h" 37#include "llvm/Target/TargetLowering.h" 38#include <algorithm> 39#include <cmath> 40#include <iostream> 41using namespace llvm; 42 43namespace { 44 Statistic<> NodesCombined ("dagcombiner", "Number of dag nodes combined"); 45 46 class DAGCombiner { 47 SelectionDAG &DAG; 48 TargetLowering &TLI; 49 bool AfterLegalize; 50 51 // Worklist of all of the nodes that need to be simplified. 52 std::vector<SDNode*> WorkList; 53 54 /// AddUsersToWorkList - When an instruction is simplified, add all users of 55 /// the instruction to the work lists because they might get more simplified 56 /// now. 57 /// 58 void AddUsersToWorkList(SDNode *N) { 59 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end(); 60 UI != UE; ++UI) 61 WorkList.push_back(*UI); 62 } 63 64 /// removeFromWorkList - remove all instances of N from the worklist. 65 /// 66 void removeFromWorkList(SDNode *N) { 67 WorkList.erase(std::remove(WorkList.begin(), WorkList.end(), N), 68 WorkList.end()); 69 } 70 71 public: 72 void AddToWorkList(SDNode *N) { 73 WorkList.push_back(N); 74 } 75 76 SDOperand CombineTo(SDNode *N, const std::vector<SDOperand> &To) { 77 ++NodesCombined; 78 DEBUG(std::cerr << "\nReplacing "; N->dump(); 79 std::cerr << "\nWith: "; To[0].Val->dump(&DAG); 80 std::cerr << " and " << To.size()-1 << " other values\n"); 81 std::vector<SDNode*> NowDead; 82 DAG.ReplaceAllUsesWith(N, To, &NowDead); 83 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); 88 } 89 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]); 95 96 // Finally, since the node is now dead, remove it from the graph. 97 DAG.DeleteNode(N); 98 return SDOperand(N, 0); 99 } 100 101 SDOperand CombineTo(SDNode *N, SDOperand Res) { 102 std::vector<SDOperand> To; 103 To.push_back(Res); 104 return CombineTo(N, To); 105 } 106 107 SDOperand CombineTo(SDNode *N, SDOperand Res0, SDOperand Res1) { 108 std::vector<SDOperand> To; 109 To.push_back(Res0); 110 To.push_back(Res1); 111 return CombineTo(N, To); 112 } 113 private: 114 115 /// 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)) 123 return false; 124 125 // Revisit the node. 126 WorkList.push_back(Op.Val); 127 128 // Replace the old value with the new one. 129 ++NodesCombined; 130 DEBUG(std::cerr << "\nReplacing "; TLO.Old.Val->dump(); 131 std::cerr << "\nWith: "; TLO.New.Val->dump(&DAG)); 132 133 std::vector<SDNode*> NowDead; 134 DAG.ReplaceAllUsesOfValueWith(TLO.Old, TLO.New, NowDead); 135 136 // Push the new node and any (possibly new) users onto the worklist. 137 WorkList.push_back(TLO.New.Val); 138 AddUsersToWorkList(TLO.New.Val); 139 140 // 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]); 144 145 // 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); 151 } 152 return true; 153 } 154 155 /// visit - call the node-specific routine that knows how to fold each 156 /// particular type of node. 157 SDOperand visit(SDNode *N); 158 159 // Visitation implementation - Implement dag node combining for different 160 // node types. The semantics are as follows: 161 // Return Value: 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. 165 // 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 visitVBinOp(SDNode *N, ISD::NodeType IntOp, ISD::NodeType FPOp); 180 SDOperand visitSHL(SDNode *N); 181 SDOperand visitSRA(SDNode *N); 182 SDOperand visitSRL(SDNode *N); 183 SDOperand visitCTLZ(SDNode *N); 184 SDOperand visitCTTZ(SDNode *N); 185 SDOperand visitCTPOP(SDNode *N); 186 SDOperand visitSELECT(SDNode *N); 187 SDOperand visitSELECT_CC(SDNode *N); 188 SDOperand visitSETCC(SDNode *N); 189 SDOperand visitSIGN_EXTEND(SDNode *N); 190 SDOperand visitZERO_EXTEND(SDNode *N); 191 SDOperand visitANY_EXTEND(SDNode *N); 192 SDOperand visitSIGN_EXTEND_INREG(SDNode *N); 193 SDOperand visitTRUNCATE(SDNode *N); 194 SDOperand visitBIT_CONVERT(SDNode *N); 195 SDOperand visitVBIT_CONVERT(SDNode *N); 196 SDOperand visitFADD(SDNode *N); 197 SDOperand visitFSUB(SDNode *N); 198 SDOperand visitFMUL(SDNode *N); 199 SDOperand visitFDIV(SDNode *N); 200 SDOperand visitFREM(SDNode *N); 201 SDOperand visitFCOPYSIGN(SDNode *N); 202 SDOperand visitSINT_TO_FP(SDNode *N); 203 SDOperand visitUINT_TO_FP(SDNode *N); 204 SDOperand visitFP_TO_SINT(SDNode *N); 205 SDOperand visitFP_TO_UINT(SDNode *N); 206 SDOperand visitFP_ROUND(SDNode *N); 207 SDOperand visitFP_ROUND_INREG(SDNode *N); 208 SDOperand visitFP_EXTEND(SDNode *N); 209 SDOperand visitFNEG(SDNode *N); 210 SDOperand visitFABS(SDNode *N); 211 SDOperand visitBRCOND(SDNode *N); 212 SDOperand visitBR_CC(SDNode *N); 213 SDOperand visitLOAD(SDNode *N); 214 SDOperand visitXEXTLOAD(SDNode *N); 215 SDOperand visitSTORE(SDNode *N); 216 SDOperand visitINSERT_VECTOR_ELT(SDNode *N); 217 SDOperand visitVINSERT_VECTOR_ELT(SDNode *N); 218 SDOperand visitVBUILD_VECTOR(SDNode *N); 219 SDOperand visitVECTOR_SHUFFLE(SDNode *N); 220 SDOperand visitVVECTOR_SHUFFLE(SDNode *N); 221 222 SDOperand XformToShuffleWithZero(SDNode *N); 223 SDOperand ReassociateOps(unsigned Opc, SDOperand LHS, SDOperand RHS); 224 225 bool SimplifySelectOps(SDNode *SELECT, SDOperand LHS, SDOperand RHS); 226 SDOperand SimplifyBinOpWithSameOpcodeHands(SDNode *N); 227 SDOperand SimplifySelect(SDOperand N0, SDOperand N1, SDOperand N2); 228 SDOperand SimplifySelectCC(SDOperand N0, SDOperand N1, SDOperand N2, 229 SDOperand N3, ISD::CondCode CC); 230 SDOperand SimplifySetCC(MVT::ValueType VT, SDOperand N0, SDOperand N1, 231 ISD::CondCode Cond, bool foldBooleans = true); 232 SDOperand ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(SDNode *, MVT::ValueType); 233 SDOperand BuildSDIV(SDNode *N); 234 SDOperand BuildUDIV(SDNode *N); 235public: 236 DAGCombiner(SelectionDAG &D) 237 : DAG(D), TLI(D.getTargetLoweringInfo()), AfterLegalize(false) {} 238 239 /// Run - runs the dag combiner on all nodes in the work list 240 void Run(bool RunningAfterLegalize); 241 }; 242} 243 244//===----------------------------------------------------------------------===// 245// TargetLowering::DAGCombinerInfo implementation 246//===----------------------------------------------------------------------===// 247 248void TargetLowering::DAGCombinerInfo::AddToWorklist(SDNode *N) { 249 ((DAGCombiner*)DC)->AddToWorkList(N); 250} 251 252SDOperand TargetLowering::DAGCombinerInfo:: 253CombineTo(SDNode *N, const std::vector<SDOperand> &To) { 254 return ((DAGCombiner*)DC)->CombineTo(N, To); 255} 256 257SDOperand TargetLowering::DAGCombinerInfo:: 258CombineTo(SDNode *N, SDOperand Res) { 259 return ((DAGCombiner*)DC)->CombineTo(N, Res); 260} 261 262 263SDOperand TargetLowering::DAGCombinerInfo:: 264CombineTo(SDNode *N, SDOperand Res0, SDOperand Res1) { 265 return ((DAGCombiner*)DC)->CombineTo(N, Res0, Res1); 266} 267 268 269 270 271//===----------------------------------------------------------------------===// 272 273 274struct ms { 275 int64_t m; // magic number 276 int64_t s; // shift amount 277}; 278 279struct mu { 280 uint64_t m; // magic number 281 int64_t a; // add indicator 282 int64_t s; // shift amount 283}; 284 285/// magic - calculate the magic numbers required to codegen an integer sdiv as 286/// a sequence of multiply and shifts. Requires that the divisor not be 0, 1, 287/// or -1. 288static ms magic32(int32_t d) { 289 int32_t p; 290 uint32_t ad, anc, delta, q1, r1, q2, r2, t; 291 const uint32_t two31 = 0x80000000U; 292 struct ms mag; 293 294 ad = abs(d); 295 t = two31 + ((uint32_t)d >> 31); 296 anc = t - 1 - t%ad; // absolute value of nc 297 p = 31; // initialize p 298 q1 = two31/anc; // initialize q1 = 2p/abs(nc) 299 r1 = two31 - q1*anc; // initialize r1 = rem(2p,abs(nc)) 300 q2 = two31/ad; // initialize q2 = 2p/abs(d) 301 r2 = two31 - q2*ad; // initialize r2 = rem(2p,abs(d)) 302 do { 303 p = p + 1; 304 q1 = 2*q1; // update q1 = 2p/abs(nc) 305 r1 = 2*r1; // update r1 = rem(2p/abs(nc)) 306 if (r1 >= anc) { // must be unsigned comparison 307 q1 = q1 + 1; 308 r1 = r1 - anc; 309 } 310 q2 = 2*q2; // update q2 = 2p/abs(d) 311 r2 = 2*r2; // update r2 = rem(2p/abs(d)) 312 if (r2 >= ad) { // must be unsigned comparison 313 q2 = q2 + 1; 314 r2 = r2 - ad; 315 } 316 delta = ad - r2; 317 } while (q1 < delta || (q1 == delta && r1 == 0)); 318 319 mag.m = (int32_t)(q2 + 1); // make sure to sign extend 320 if (d < 0) mag.m = -mag.m; // resulting magic number 321 mag.s = p - 32; // resulting shift 322 return mag; 323} 324 325/// magicu - calculate the magic numbers required to codegen an integer udiv as 326/// a sequence of multiply, add and shifts. Requires that the divisor not be 0. 327static mu magicu32(uint32_t d) { 328 int32_t p; 329 uint32_t nc, delta, q1, r1, q2, r2; 330 struct mu magu; 331 magu.a = 0; // initialize "add" indicator 332 nc = - 1 - (-d)%d; 333 p = 31; // initialize p 334 q1 = 0x80000000/nc; // initialize q1 = 2p/nc 335 r1 = 0x80000000 - q1*nc; // initialize r1 = rem(2p,nc) 336 q2 = 0x7FFFFFFF/d; // initialize q2 = (2p-1)/d 337 r2 = 0x7FFFFFFF - q2*d; // initialize r2 = rem((2p-1),d) 338 do { 339 p = p + 1; 340 if (r1 >= nc - r1 ) { 341 q1 = 2*q1 + 1; // update q1 342 r1 = 2*r1 - nc; // update r1 343 } 344 else { 345 q1 = 2*q1; // update q1 346 r1 = 2*r1; // update r1 347 } 348 if (r2 + 1 >= d - r2) { 349 if (q2 >= 0x7FFFFFFF) magu.a = 1; 350 q2 = 2*q2 + 1; // update q2 351 r2 = 2*r2 + 1 - d; // update r2 352 } 353 else { 354 if (q2 >= 0x80000000) magu.a = 1; 355 q2 = 2*q2; // update q2 356 r2 = 2*r2 + 1; // update r2 357 } 358 delta = d - 1 - r2; 359 } while (p < 64 && (q1 < delta || (q1 == delta && r1 == 0))); 360 magu.m = q2 + 1; // resulting magic number 361 magu.s = p - 32; // resulting shift 362 return magu; 363} 364 365/// magic - calculate the magic numbers required to codegen an integer sdiv as 366/// a sequence of multiply and shifts. Requires that the divisor not be 0, 1, 367/// or -1. 368static ms magic64(int64_t d) { 369 int64_t p; 370 uint64_t ad, anc, delta, q1, r1, q2, r2, t; 371 const uint64_t two63 = 9223372036854775808ULL; // 2^63 372 struct ms mag; 373 374 ad = d >= 0 ? d : -d; 375 t = two63 + ((uint64_t)d >> 63); 376 anc = t - 1 - t%ad; // absolute value of nc 377 p = 63; // initialize p 378 q1 = two63/anc; // initialize q1 = 2p/abs(nc) 379 r1 = two63 - q1*anc; // initialize r1 = rem(2p,abs(nc)) 380 q2 = two63/ad; // initialize q2 = 2p/abs(d) 381 r2 = two63 - q2*ad; // initialize r2 = rem(2p,abs(d)) 382 do { 383 p = p + 1; 384 q1 = 2*q1; // update q1 = 2p/abs(nc) 385 r1 = 2*r1; // update r1 = rem(2p/abs(nc)) 386 if (r1 >= anc) { // must be unsigned comparison 387 q1 = q1 + 1; 388 r1 = r1 - anc; 389 } 390 q2 = 2*q2; // update q2 = 2p/abs(d) 391 r2 = 2*r2; // update r2 = rem(2p/abs(d)) 392 if (r2 >= ad) { // must be unsigned comparison 393 q2 = q2 + 1; 394 r2 = r2 - ad; 395 } 396 delta = ad - r2; 397 } while (q1 < delta || (q1 == delta && r1 == 0)); 398 399 mag.m = q2 + 1; 400 if (d < 0) mag.m = -mag.m; // resulting magic number 401 mag.s = p - 64; // resulting shift 402 return mag; 403} 404 405/// magicu - calculate the magic numbers required to codegen an integer udiv as 406/// a sequence of multiply, add and shifts. Requires that the divisor not be 0. 407static mu magicu64(uint64_t d) 408{ 409 int64_t p; 410 uint64_t nc, delta, q1, r1, q2, r2; 411 struct mu magu; 412 magu.a = 0; // initialize "add" indicator 413 nc = - 1 - (-d)%d; 414 p = 63; // initialize p 415 q1 = 0x8000000000000000ull/nc; // initialize q1 = 2p/nc 416 r1 = 0x8000000000000000ull - q1*nc; // initialize r1 = rem(2p,nc) 417 q2 = 0x7FFFFFFFFFFFFFFFull/d; // initialize q2 = (2p-1)/d 418 r2 = 0x7FFFFFFFFFFFFFFFull - q2*d; // initialize r2 = rem((2p-1),d) 419 do { 420 p = p + 1; 421 if (r1 >= nc - r1 ) { 422 q1 = 2*q1 + 1; // update q1 423 r1 = 2*r1 - nc; // update r1 424 } 425 else { 426 q1 = 2*q1; // update q1 427 r1 = 2*r1; // update r1 428 } 429 if (r2 + 1 >= d - r2) { 430 if (q2 >= 0x7FFFFFFFFFFFFFFFull) magu.a = 1; 431 q2 = 2*q2 + 1; // update q2 432 r2 = 2*r2 + 1 - d; // update r2 433 } 434 else { 435 if (q2 >= 0x8000000000000000ull) magu.a = 1; 436 q2 = 2*q2; // update q2 437 r2 = 2*r2 + 1; // update r2 438 } 439 delta = d - 1 - r2; 440 } while (p < 64 && (q1 < delta || (q1 == delta && r1 == 0))); 441 magu.m = q2 + 1; // resulting magic number 442 magu.s = p - 64; // resulting shift 443 return magu; 444} 445 446// isSetCCEquivalent - Return true if this node is a setcc, or is a select_cc 447// that selects between the values 1 and 0, making it equivalent to a setcc. 448// Also, set the incoming LHS, RHS, and CC references to the appropriate 449// nodes based on the type of node we are checking. This simplifies life a 450// bit for the callers. 451static bool isSetCCEquivalent(SDOperand N, SDOperand &LHS, SDOperand &RHS, 452 SDOperand &CC) { 453 if (N.getOpcode() == ISD::SETCC) { 454 LHS = N.getOperand(0); 455 RHS = N.getOperand(1); 456 CC = N.getOperand(2); 457 return true; 458 } 459 if (N.getOpcode() == ISD::SELECT_CC && 460 N.getOperand(2).getOpcode() == ISD::Constant && 461 N.getOperand(3).getOpcode() == ISD::Constant && 462 cast<ConstantSDNode>(N.getOperand(2))->getValue() == 1 && 463 cast<ConstantSDNode>(N.getOperand(3))->isNullValue()) { 464 LHS = N.getOperand(0); 465 RHS = N.getOperand(1); 466 CC = N.getOperand(4); 467 return true; 468 } 469 return false; 470} 471 472// isOneUseSetCC - Return true if this is a SetCC-equivalent operation with only 473// one use. If this is true, it allows the users to invert the operation for 474// free when it is profitable to do so. 475static bool isOneUseSetCC(SDOperand N) { 476 SDOperand N0, N1, N2; 477 if (isSetCCEquivalent(N, N0, N1, N2) && N.Val->hasOneUse()) 478 return true; 479 return false; 480} 481 482// FIXME: This should probably go in the ISD class rather than being duplicated 483// in several files. 484static bool isCommutativeBinOp(unsigned Opcode) { 485 switch (Opcode) { 486 case ISD::ADD: 487 case ISD::MUL: 488 case ISD::AND: 489 case ISD::OR: 490 case ISD::XOR: return true; 491 default: return false; // FIXME: Need commutative info for user ops! 492 } 493} 494 495SDOperand DAGCombiner::ReassociateOps(unsigned Opc, SDOperand N0, SDOperand N1){ 496 MVT::ValueType VT = N0.getValueType(); 497 // reassoc. (op (op x, c1), y) -> (op (op x, y), c1) iff x+c1 has one use 498 // reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2)) 499 if (N0.getOpcode() == Opc && isa<ConstantSDNode>(N0.getOperand(1))) { 500 if (isa<ConstantSDNode>(N1)) { 501 SDOperand OpNode = DAG.getNode(Opc, VT, N0.getOperand(1), N1); 502 AddToWorkList(OpNode.Val); 503 return DAG.getNode(Opc, VT, OpNode, N0.getOperand(0)); 504 } else if (N0.hasOneUse()) { 505 SDOperand OpNode = DAG.getNode(Opc, VT, N0.getOperand(0), N1); 506 AddToWorkList(OpNode.Val); 507 return DAG.getNode(Opc, VT, OpNode, N0.getOperand(1)); 508 } 509 } 510 // reassoc. (op y, (op x, c1)) -> (op (op x, y), c1) iff x+c1 has one use 511 // reassoc. (op c2, (op x, c1)) -> (op x, (op c1, c2)) 512 if (N1.getOpcode() == Opc && isa<ConstantSDNode>(N1.getOperand(1))) { 513 if (isa<ConstantSDNode>(N0)) { 514 SDOperand OpNode = DAG.getNode(Opc, VT, N1.getOperand(1), N0); 515 AddToWorkList(OpNode.Val); 516 return DAG.getNode(Opc, VT, OpNode, N1.getOperand(0)); 517 } else if (N1.hasOneUse()) { 518 SDOperand OpNode = DAG.getNode(Opc, VT, N1.getOperand(0), N0); 519 AddToWorkList(OpNode.Val); 520 return DAG.getNode(Opc, VT, OpNode, N1.getOperand(1)); 521 } 522 } 523 return SDOperand(); 524} 525 526void DAGCombiner::Run(bool RunningAfterLegalize) { 527 // set the instance variable, so that the various visit routines may use it. 528 AfterLegalize = RunningAfterLegalize; 529 530 // Add all the dag nodes to the worklist. 531 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(), 532 E = DAG.allnodes_end(); I != E; ++I) 533 WorkList.push_back(I); 534 535 // Create a dummy node (which is not added to allnodes), that adds a reference 536 // to the root node, preventing it from being deleted, and tracking any 537 // changes of the root. 538 HandleSDNode Dummy(DAG.getRoot()); 539 540 541 /// DagCombineInfo - Expose the DAG combiner to the target combiner impls. 542 TargetLowering::DAGCombinerInfo 543 DagCombineInfo(DAG, !RunningAfterLegalize, this); 544 545 // while the worklist isn't empty, inspect the node on the end of it and 546 // try and combine it. 547 while (!WorkList.empty()) { 548 SDNode *N = WorkList.back(); 549 WorkList.pop_back(); 550 551 // If N has no uses, it is dead. Make sure to revisit all N's operands once 552 // N is deleted from the DAG, since they too may now be dead or may have a 553 // reduced number of uses, allowing other xforms. 554 if (N->use_empty() && N != &Dummy) { 555 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) 556 WorkList.push_back(N->getOperand(i).Val); 557 558 removeFromWorkList(N); 559 DAG.DeleteNode(N); 560 continue; 561 } 562 563 SDOperand RV = visit(N); 564 565 // If nothing happened, try a target-specific DAG combine. 566 if (RV.Val == 0) { 567 if (N->getOpcode() >= ISD::BUILTIN_OP_END || 568 TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode())) 569 RV = TLI.PerformDAGCombine(N, DagCombineInfo); 570 } 571 572 if (RV.Val) { 573 ++NodesCombined; 574 // If we get back the same node we passed in, rather than a new node or 575 // zero, we know that the node must have defined multiple values and 576 // CombineTo was used. Since CombineTo takes care of the worklist 577 // mechanics for us, we have no work to do in this case. 578 if (RV.Val != N) { 579 DEBUG(std::cerr << "\nReplacing "; N->dump(); 580 std::cerr << "\nWith: "; RV.Val->dump(&DAG); 581 std::cerr << '\n'); 582 std::vector<SDNode*> NowDead; 583 DAG.ReplaceAllUsesWith(N, std::vector<SDOperand>(1, RV), &NowDead); 584 585 // Push the new node and any users onto the worklist 586 WorkList.push_back(RV.Val); 587 AddUsersToWorkList(RV.Val); 588 589 // Nodes can end up on the worklist more than once. Make sure we do 590 // not process a node that has been replaced. 591 removeFromWorkList(N); 592 for (unsigned i = 0, e = NowDead.size(); i != e; ++i) 593 removeFromWorkList(NowDead[i]); 594 595 // Finally, since the node is now dead, remove it from the graph. 596 DAG.DeleteNode(N); 597 } 598 } 599 } 600 601 // If the root changed (e.g. it was a dead load, update the root). 602 DAG.setRoot(Dummy.getValue()); 603} 604 605SDOperand DAGCombiner::visit(SDNode *N) { 606 switch(N->getOpcode()) { 607 default: break; 608 case ISD::TokenFactor: return visitTokenFactor(N); 609 case ISD::ADD: return visitADD(N); 610 case ISD::SUB: return visitSUB(N); 611 case ISD::MUL: return visitMUL(N); 612 case ISD::SDIV: return visitSDIV(N); 613 case ISD::UDIV: return visitUDIV(N); 614 case ISD::SREM: return visitSREM(N); 615 case ISD::UREM: return visitUREM(N); 616 case ISD::MULHU: return visitMULHU(N); 617 case ISD::MULHS: return visitMULHS(N); 618 case ISD::AND: return visitAND(N); 619 case ISD::OR: return visitOR(N); 620 case ISD::XOR: return visitXOR(N); 621 case ISD::SHL: return visitSHL(N); 622 case ISD::SRA: return visitSRA(N); 623 case ISD::SRL: return visitSRL(N); 624 case ISD::CTLZ: return visitCTLZ(N); 625 case ISD::CTTZ: return visitCTTZ(N); 626 case ISD::CTPOP: return visitCTPOP(N); 627 case ISD::SELECT: return visitSELECT(N); 628 case ISD::SELECT_CC: return visitSELECT_CC(N); 629 case ISD::SETCC: return visitSETCC(N); 630 case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N); 631 case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N); 632 case ISD::ANY_EXTEND: return visitANY_EXTEND(N); 633 case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N); 634 case ISD::TRUNCATE: return visitTRUNCATE(N); 635 case ISD::BIT_CONVERT: return visitBIT_CONVERT(N); 636 case ISD::VBIT_CONVERT: return visitVBIT_CONVERT(N); 637 case ISD::FADD: return visitFADD(N); 638 case ISD::FSUB: return visitFSUB(N); 639 case ISD::FMUL: return visitFMUL(N); 640 case ISD::FDIV: return visitFDIV(N); 641 case ISD::FREM: return visitFREM(N); 642 case ISD::FCOPYSIGN: return visitFCOPYSIGN(N); 643 case ISD::SINT_TO_FP: return visitSINT_TO_FP(N); 644 case ISD::UINT_TO_FP: return visitUINT_TO_FP(N); 645 case ISD::FP_TO_SINT: return visitFP_TO_SINT(N); 646 case ISD::FP_TO_UINT: return visitFP_TO_UINT(N); 647 case ISD::FP_ROUND: return visitFP_ROUND(N); 648 case ISD::FP_ROUND_INREG: return visitFP_ROUND_INREG(N); 649 case ISD::FP_EXTEND: return visitFP_EXTEND(N); 650 case ISD::FNEG: return visitFNEG(N); 651 case ISD::FABS: return visitFABS(N); 652 case ISD::BRCOND: return visitBRCOND(N); 653 case ISD::BR_CC: return visitBR_CC(N); 654 case ISD::LOAD: return visitLOAD(N); 655 case ISD::EXTLOAD: 656 case ISD::SEXTLOAD: 657 case ISD::ZEXTLOAD: return visitXEXTLOAD(N); 658 case ISD::STORE: return visitSTORE(N); 659 case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N); 660 case ISD::VINSERT_VECTOR_ELT: return visitVINSERT_VECTOR_ELT(N); 661 case ISD::VBUILD_VECTOR: return visitVBUILD_VECTOR(N); 662 case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N); 663 case ISD::VVECTOR_SHUFFLE: return visitVVECTOR_SHUFFLE(N); 664 case ISD::VADD: return visitVBinOp(N, ISD::ADD , ISD::FADD); 665 case ISD::VSUB: return visitVBinOp(N, ISD::SUB , ISD::FSUB); 666 case ISD::VMUL: return visitVBinOp(N, ISD::MUL , ISD::FMUL); 667 case ISD::VSDIV: return visitVBinOp(N, ISD::SDIV, ISD::FDIV); 668 case ISD::VUDIV: return visitVBinOp(N, ISD::UDIV, ISD::UDIV); 669 case ISD::VAND: return visitVBinOp(N, ISD::AND , ISD::AND); 670 case ISD::VOR: return visitVBinOp(N, ISD::OR , ISD::OR); 671 case ISD::VXOR: return visitVBinOp(N, ISD::XOR , ISD::XOR); 672 } 673 return SDOperand(); 674} 675 676SDOperand DAGCombiner::visitTokenFactor(SDNode *N) { 677 std::vector<SDOperand> Ops; 678 bool Changed = false; 679 680 // If the token factor has two operands and one is the entry token, replace 681 // the token factor with the other operand. 682 if (N->getNumOperands() == 2) { 683 if (N->getOperand(0).getOpcode() == ISD::EntryToken) 684 return N->getOperand(1); 685 if (N->getOperand(1).getOpcode() == ISD::EntryToken) 686 return N->getOperand(0); 687 } 688 689 // fold (tokenfactor (tokenfactor)) -> tokenfactor 690 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) { 691 SDOperand Op = N->getOperand(i); 692 if (Op.getOpcode() == ISD::TokenFactor && Op.hasOneUse()) { 693 AddToWorkList(Op.Val); // Remove dead node. 694 Changed = true; 695 for (unsigned j = 0, e = Op.getNumOperands(); j != e; ++j) 696 Ops.push_back(Op.getOperand(j)); 697 } else { 698 Ops.push_back(Op); 699 } 700 } 701 if (Changed) 702 return DAG.getNode(ISD::TokenFactor, MVT::Other, Ops); 703 return SDOperand(); 704} 705 706SDOperand DAGCombiner::visitADD(SDNode *N) { 707 SDOperand N0 = N->getOperand(0); 708 SDOperand N1 = N->getOperand(1); 709 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 710 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 711 MVT::ValueType VT = N0.getValueType(); 712 713 // fold (add c1, c2) -> c1+c2 714 if (N0C && N1C) 715 return DAG.getNode(ISD::ADD, VT, N0, N1); 716 // canonicalize constant to RHS 717 if (N0C && !N1C) 718 return DAG.getNode(ISD::ADD, VT, N1, N0); 719 // fold (add x, 0) -> x 720 if (N1C && N1C->isNullValue()) 721 return N0; 722 // fold ((c1-A)+c2) -> (c1+c2)-A 723 if (N1C && N0.getOpcode() == ISD::SUB) 724 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getOperand(0))) 725 return DAG.getNode(ISD::SUB, VT, 726 DAG.getConstant(N1C->getValue()+N0C->getValue(), VT), 727 N0.getOperand(1)); 728 // reassociate add 729 SDOperand RADD = ReassociateOps(ISD::ADD, N0, N1); 730 if (RADD.Val != 0) 731 return RADD; 732 // fold ((0-A) + B) -> B-A 733 if (N0.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N0.getOperand(0)) && 734 cast<ConstantSDNode>(N0.getOperand(0))->isNullValue()) 735 return DAG.getNode(ISD::SUB, VT, N1, N0.getOperand(1)); 736 // fold (A + (0-B)) -> A-B 737 if (N1.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N1.getOperand(0)) && 738 cast<ConstantSDNode>(N1.getOperand(0))->isNullValue()) 739 return DAG.getNode(ISD::SUB, VT, N0, N1.getOperand(1)); 740 // fold (A+(B-A)) -> B 741 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1)) 742 return N1.getOperand(0); 743 744 if (!MVT::isVector(VT) && SimplifyDemandedBits(SDOperand(N, 0))) 745 return SDOperand(N, 0); 746 747 // fold (a+b) -> (a|b) iff a and b share no bits. 748 if (MVT::isInteger(VT) && !MVT::isVector(VT)) { 749 uint64_t LHSZero, LHSOne; 750 uint64_t RHSZero, RHSOne; 751 uint64_t Mask = MVT::getIntVTBitMask(VT); 752 TLI.ComputeMaskedBits(N0, Mask, LHSZero, LHSOne); 753 if (LHSZero) { 754 TLI.ComputeMaskedBits(N1, Mask, RHSZero, RHSOne); 755 756 // If all possibly-set bits on the LHS are clear on the RHS, return an OR. 757 // If all possibly-set bits on the RHS are clear on the LHS, return an OR. 758 if ((RHSZero & (~LHSZero & Mask)) == (~LHSZero & Mask) || 759 (LHSZero & (~RHSZero & Mask)) == (~RHSZero & Mask)) 760 return DAG.getNode(ISD::OR, VT, N0, N1); 761 } 762 } 763 764 return SDOperand(); 765} 766 767SDOperand DAGCombiner::visitSUB(SDNode *N) { 768 SDOperand N0 = N->getOperand(0); 769 SDOperand N1 = N->getOperand(1); 770 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val); 771 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val); 772 MVT::ValueType VT = N0.getValueType(); 773 774 // fold (sub x, x) -> 0 775 if (N0 == N1) 776 return DAG.getConstant(0, N->getValueType(0)); 777 // fold (sub c1, c2) -> c1-c2 778 if (N0C && N1C) 779 return DAG.getNode(ISD::SUB, VT, N0, N1); 780 // fold (sub x, c) -> (add x, -c) 781 if (N1C) 782 return DAG.getNode(ISD::ADD, VT, N0, DAG.getConstant(-N1C->getValue(), VT)); 783 // fold (A+B)-A -> B 784 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1) 785 return N0.getOperand(1); 786 // fold (A+B)-B -> A 787 if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1) 788 return N0.getOperand(0); 789 return SDOperand(); 790} 791 792SDOperand DAGCombiner::visitMUL(SDNode *N) { 793 SDOperand N0 = N->getOperand(0); 794 SDOperand N1 = N->getOperand(1); 795 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 796 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 797 MVT::ValueType VT = N0.getValueType(); 798 799 // fold (mul c1, c2) -> c1*c2 800 if (N0C && N1C) 801 return DAG.getNode(ISD::MUL, VT, N0, N1); 802 // canonicalize constant to RHS 803 if (N0C && !N1C) 804 return DAG.getNode(ISD::MUL, VT, N1, N0); 805 // fold (mul x, 0) -> 0 806 if (N1C && N1C->isNullValue()) 807 return N1; 808 // fold (mul x, -1) -> 0-x 809 if (N1C && N1C->isAllOnesValue()) 810 return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT), N0); 811 // fold (mul x, (1 << c)) -> x << c 812 if (N1C && isPowerOf2_64(N1C->getValue())) 813 return DAG.getNode(ISD::SHL, VT, N0, 814 DAG.getConstant(Log2_64(N1C->getValue()), 815 TLI.getShiftAmountTy())); 816 // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c 817 if (N1C && isPowerOf2_64(-N1C->getSignExtended())) { 818 // FIXME: If the input is something that is easily negated (e.g. a 819 // single-use add), we should put the negate there. 820 return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT), 821 DAG.getNode(ISD::SHL, VT, N0, 822 DAG.getConstant(Log2_64(-N1C->getSignExtended()), 823 TLI.getShiftAmountTy()))); 824 } 825 826 // (mul (shl X, c1), c2) -> (mul X, c2 << c1) 827 if (N1C && N0.getOpcode() == ISD::SHL && 828 isa<ConstantSDNode>(N0.getOperand(1))) { 829 SDOperand C3 = DAG.getNode(ISD::SHL, VT, N1, N0.getOperand(1)); 830 AddToWorkList(C3.Val); 831 return DAG.getNode(ISD::MUL, VT, N0.getOperand(0), C3); 832 } 833 834 // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one 835 // use. 836 { 837 SDOperand Sh(0,0), Y(0,0); 838 // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)). 839 if (N0.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N0.getOperand(1)) && 840 N0.Val->hasOneUse()) { 841 Sh = N0; Y = N1; 842 } else if (N1.getOpcode() == ISD::SHL && 843 isa<ConstantSDNode>(N1.getOperand(1)) && N1.Val->hasOneUse()) { 844 Sh = N1; Y = N0; 845 } 846 if (Sh.Val) { 847 SDOperand Mul = DAG.getNode(ISD::MUL, VT, Sh.getOperand(0), Y); 848 return DAG.getNode(ISD::SHL, VT, Mul, Sh.getOperand(1)); 849 } 850 } 851 // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2) 852 if (N1C && N0.getOpcode() == ISD::ADD && N0.Val->hasOneUse() && 853 isa<ConstantSDNode>(N0.getOperand(1))) { 854 return DAG.getNode(ISD::ADD, VT, 855 DAG.getNode(ISD::MUL, VT, N0.getOperand(0), N1), 856 DAG.getNode(ISD::MUL, VT, N0.getOperand(1), N1)); 857 } 858 859 // reassociate mul 860 SDOperand RMUL = ReassociateOps(ISD::MUL, N0, N1); 861 if (RMUL.Val != 0) 862 return RMUL; 863 return SDOperand(); 864} 865 866SDOperand DAGCombiner::visitSDIV(SDNode *N) { 867 SDOperand N0 = N->getOperand(0); 868 SDOperand N1 = N->getOperand(1); 869 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val); 870 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val); 871 MVT::ValueType VT = N->getValueType(0); 872 873 // fold (sdiv c1, c2) -> c1/c2 874 if (N0C && N1C && !N1C->isNullValue()) 875 return DAG.getNode(ISD::SDIV, VT, N0, N1); 876 // fold (sdiv X, 1) -> X 877 if (N1C && N1C->getSignExtended() == 1LL) 878 return N0; 879 // fold (sdiv X, -1) -> 0-X 880 if (N1C && N1C->isAllOnesValue()) 881 return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT), N0); 882 // If we know the sign bits of both operands are zero, strength reduce to a 883 // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2 884 uint64_t SignBit = 1ULL << (MVT::getSizeInBits(VT)-1); 885 if (TLI.MaskedValueIsZero(N1, SignBit) && 886 TLI.MaskedValueIsZero(N0, SignBit)) 887 return DAG.getNode(ISD::UDIV, N1.getValueType(), N0, N1); 888 // fold (sdiv X, pow2) -> simple ops after legalize 889 if (N1C && N1C->getValue() && !TLI.isIntDivCheap() && 890 (isPowerOf2_64(N1C->getSignExtended()) || 891 isPowerOf2_64(-N1C->getSignExtended()))) { 892 // If dividing by powers of two is cheap, then don't perform the following 893 // fold. 894 if (TLI.isPow2DivCheap()) 895 return SDOperand(); 896 int64_t pow2 = N1C->getSignExtended(); 897 int64_t abs2 = pow2 > 0 ? pow2 : -pow2; 898 unsigned lg2 = Log2_64(abs2); 899 // Splat the sign bit into the register 900 SDOperand SGN = DAG.getNode(ISD::SRA, VT, N0, 901 DAG.getConstant(MVT::getSizeInBits(VT)-1, 902 TLI.getShiftAmountTy())); 903 AddToWorkList(SGN.Val); 904 // Add (N0 < 0) ? abs2 - 1 : 0; 905 SDOperand SRL = DAG.getNode(ISD::SRL, VT, SGN, 906 DAG.getConstant(MVT::getSizeInBits(VT)-lg2, 907 TLI.getShiftAmountTy())); 908 SDOperand ADD = DAG.getNode(ISD::ADD, VT, N0, SRL); 909 AddToWorkList(SRL.Val); 910 AddToWorkList(ADD.Val); // Divide by pow2 911 SDOperand SRA = DAG.getNode(ISD::SRA, VT, ADD, 912 DAG.getConstant(lg2, TLI.getShiftAmountTy())); 913 // If we're dividing by a positive value, we're done. Otherwise, we must 914 // negate the result. 915 if (pow2 > 0) 916 return SRA; 917 AddToWorkList(SRA.Val); 918 return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT), SRA); 919 } 920 // if integer divide is expensive and we satisfy the requirements, emit an 921 // alternate sequence. 922 if (N1C && (N1C->getSignExtended() < -1 || N1C->getSignExtended() > 1) && 923 !TLI.isIntDivCheap()) { 924 SDOperand Op = BuildSDIV(N); 925 if (Op.Val) return Op; 926 } 927 return SDOperand(); 928} 929 930SDOperand DAGCombiner::visitUDIV(SDNode *N) { 931 SDOperand N0 = N->getOperand(0); 932 SDOperand N1 = N->getOperand(1); 933 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val); 934 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val); 935 MVT::ValueType VT = N->getValueType(0); 936 937 // fold (udiv c1, c2) -> c1/c2 938 if (N0C && N1C && !N1C->isNullValue()) 939 return DAG.getNode(ISD::UDIV, VT, N0, N1); 940 // fold (udiv x, (1 << c)) -> x >>u c 941 if (N1C && isPowerOf2_64(N1C->getValue())) 942 return DAG.getNode(ISD::SRL, VT, N0, 943 DAG.getConstant(Log2_64(N1C->getValue()), 944 TLI.getShiftAmountTy())); 945 // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2 946 if (N1.getOpcode() == ISD::SHL) { 947 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) { 948 if (isPowerOf2_64(SHC->getValue())) { 949 MVT::ValueType ADDVT = N1.getOperand(1).getValueType(); 950 SDOperand Add = DAG.getNode(ISD::ADD, ADDVT, N1.getOperand(1), 951 DAG.getConstant(Log2_64(SHC->getValue()), 952 ADDVT)); 953 AddToWorkList(Add.Val); 954 return DAG.getNode(ISD::SRL, VT, N0, Add); 955 } 956 } 957 } 958 // fold (udiv x, c) -> alternate 959 if (N1C && N1C->getValue() && !TLI.isIntDivCheap()) { 960 SDOperand Op = BuildUDIV(N); 961 if (Op.Val) return Op; 962 } 963 return SDOperand(); 964} 965 966SDOperand DAGCombiner::visitSREM(SDNode *N) { 967 SDOperand N0 = N->getOperand(0); 968 SDOperand N1 = N->getOperand(1); 969 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 970 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 971 MVT::ValueType VT = N->getValueType(0); 972 973 // fold (srem c1, c2) -> c1%c2 974 if (N0C && N1C && !N1C->isNullValue()) 975 return DAG.getNode(ISD::SREM, VT, N0, N1); 976 // If we know the sign bits of both operands are zero, strength reduce to a 977 // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15 978 uint64_t SignBit = 1ULL << (MVT::getSizeInBits(VT)-1); 979 if (TLI.MaskedValueIsZero(N1, SignBit) && 980 TLI.MaskedValueIsZero(N0, SignBit)) 981 return DAG.getNode(ISD::UREM, VT, N0, N1); 982 return SDOperand(); 983} 984 985SDOperand DAGCombiner::visitUREM(SDNode *N) { 986 SDOperand N0 = N->getOperand(0); 987 SDOperand N1 = N->getOperand(1); 988 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 989 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 990 MVT::ValueType VT = N->getValueType(0); 991 992 // fold (urem c1, c2) -> c1%c2 993 if (N0C && N1C && !N1C->isNullValue()) 994 return DAG.getNode(ISD::UREM, VT, N0, N1); 995 // fold (urem x, pow2) -> (and x, pow2-1) 996 if (N1C && !N1C->isNullValue() && isPowerOf2_64(N1C->getValue())) 997 return DAG.getNode(ISD::AND, VT, N0, DAG.getConstant(N1C->getValue()-1,VT)); 998 // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1)) 999 if (N1.getOpcode() == ISD::SHL) { 1000 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) { 1001 if (isPowerOf2_64(SHC->getValue())) { 1002 SDOperand Add = DAG.getNode(ISD::ADD, VT, N1,DAG.getConstant(~0ULL,VT)); 1003 AddToWorkList(Add.Val); 1004 return DAG.getNode(ISD::AND, VT, N0, Add); 1005 } 1006 } 1007 } 1008 return SDOperand(); 1009} 1010 1011SDOperand DAGCombiner::visitMULHS(SDNode *N) { 1012 SDOperand N0 = N->getOperand(0); 1013 SDOperand N1 = N->getOperand(1); 1014 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 1015 1016 // fold (mulhs x, 0) -> 0 1017 if (N1C && N1C->isNullValue()) 1018 return N1; 1019 // fold (mulhs x, 1) -> (sra x, size(x)-1) 1020 if (N1C && N1C->getValue() == 1) 1021 return DAG.getNode(ISD::SRA, N0.getValueType(), N0, 1022 DAG.getConstant(MVT::getSizeInBits(N0.getValueType())-1, 1023 TLI.getShiftAmountTy())); 1024 return SDOperand(); 1025} 1026 1027SDOperand DAGCombiner::visitMULHU(SDNode *N) { 1028 SDOperand N0 = N->getOperand(0); 1029 SDOperand N1 = N->getOperand(1); 1030 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 1031 1032 // fold (mulhu x, 0) -> 0 1033 if (N1C && N1C->isNullValue()) 1034 return N1; 1035 // fold (mulhu x, 1) -> 0 1036 if (N1C && N1C->getValue() == 1) 1037 return DAG.getConstant(0, N0.getValueType()); 1038 return SDOperand(); 1039} 1040 1041/// SimplifyBinOpWithSameOpcodeHands - If this is a binary operator with 1042/// two operands of the same opcode, try to simplify it. 1043SDOperand DAGCombiner::SimplifyBinOpWithSameOpcodeHands(SDNode *N) { 1044 SDOperand N0 = N->getOperand(0), N1 = N->getOperand(1); 1045 MVT::ValueType VT = N0.getValueType(); 1046 assert(N0.getOpcode() == N1.getOpcode() && "Bad input!"); 1047 1048 // For each of OP in AND/OR/XOR: 1049 // fold (OP (zext x), (zext y)) -> (zext (OP x, y)) 1050 // fold (OP (sext x), (sext y)) -> (sext (OP x, y)) 1051 // fold (OP (aext x), (aext y)) -> (aext (OP x, y)) 1052 // fold (OP (trunc x), (trunc y)) -> (trunc (OP x, y)) 1053 if ((N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND|| 1054 N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::TRUNCATE) && 1055 N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()) { 1056 SDOperand ORNode = DAG.getNode(N->getOpcode(), 1057 N0.getOperand(0).getValueType(), 1058 N0.getOperand(0), N1.getOperand(0)); 1059 AddToWorkList(ORNode.Val); 1060 return DAG.getNode(N0.getOpcode(), VT, ORNode); 1061 } 1062 1063 // For each of OP in SHL/SRL/SRA/AND... 1064 // fold (and (OP x, z), (OP y, z)) -> (OP (and x, y), z) 1065 // fold (or (OP x, z), (OP y, z)) -> (OP (or x, y), z) 1066 // fold (xor (OP x, z), (OP y, z)) -> (OP (xor x, y), z) 1067 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL || 1068 N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::AND) && 1069 N0.getOperand(1) == N1.getOperand(1)) { 1070 SDOperand ORNode = DAG.getNode(N->getOpcode(), 1071 N0.getOperand(0).getValueType(), 1072 N0.getOperand(0), N1.getOperand(0)); 1073 AddToWorkList(ORNode.Val); 1074 return DAG.getNode(N0.getOpcode(), VT, ORNode, N0.getOperand(1)); 1075 } 1076 1077 return SDOperand(); 1078} 1079 1080SDOperand DAGCombiner::visitAND(SDNode *N) { 1081 SDOperand N0 = N->getOperand(0); 1082 SDOperand N1 = N->getOperand(1); 1083 SDOperand LL, LR, RL, RR, CC0, CC1; 1084 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1085 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 1086 MVT::ValueType VT = N1.getValueType(); 1087 unsigned OpSizeInBits = MVT::getSizeInBits(VT); 1088 1089 // fold (and c1, c2) -> c1&c2 1090 if (N0C && N1C) 1091 return DAG.getNode(ISD::AND, VT, N0, N1); 1092 // canonicalize constant to RHS 1093 if (N0C && !N1C) 1094 return DAG.getNode(ISD::AND, VT, N1, N0); 1095 // fold (and x, -1) -> x 1096 if (N1C && N1C->isAllOnesValue()) 1097 return N0; 1098 // if (and x, c) is known to be zero, return 0 1099 if (N1C && TLI.MaskedValueIsZero(SDOperand(N, 0), MVT::getIntVTBitMask(VT))) 1100 return DAG.getConstant(0, VT); 1101 // reassociate and 1102 SDOperand RAND = ReassociateOps(ISD::AND, N0, N1); 1103 if (RAND.Val != 0) 1104 return RAND; 1105 // fold (and (or x, 0xFFFF), 0xFF) -> 0xFF 1106 if (N1C && N0.getOpcode() == ISD::OR) 1107 if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1))) 1108 if ((ORI->getValue() & N1C->getValue()) == N1C->getValue()) 1109 return N1; 1110 // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits. 1111 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) { 1112 unsigned InMask = MVT::getIntVTBitMask(N0.getOperand(0).getValueType()); 1113 if (TLI.MaskedValueIsZero(N0.getOperand(0), 1114 ~N1C->getValue() & InMask)) { 1115 SDOperand Zext = DAG.getNode(ISD::ZERO_EXTEND, N0.getValueType(), 1116 N0.getOperand(0)); 1117 1118 // Replace uses of the AND with uses of the Zero extend node. 1119 CombineTo(N, Zext); 1120 1121 // We actually want to replace all uses of the any_extend with the 1122 // zero_extend, to avoid duplicating things. This will later cause this 1123 // AND to be folded. 1124 CombineTo(N0.Val, Zext); 1125 return SDOperand(N, 0); // Return N so it doesn't get rechecked! 1126 } 1127 } 1128 // fold (and (setcc x), (setcc y)) -> (setcc (and x, y)) 1129 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){ 1130 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get(); 1131 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get(); 1132 1133 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 && 1134 MVT::isInteger(LL.getValueType())) { 1135 // fold (X == 0) & (Y == 0) -> (X|Y == 0) 1136 if (cast<ConstantSDNode>(LR)->getValue() == 0 && Op1 == ISD::SETEQ) { 1137 SDOperand ORNode = DAG.getNode(ISD::OR, LR.getValueType(), LL, RL); 1138 AddToWorkList(ORNode.Val); 1139 return DAG.getSetCC(VT, ORNode, LR, Op1); 1140 } 1141 // fold (X == -1) & (Y == -1) -> (X&Y == -1) 1142 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETEQ) { 1143 SDOperand ANDNode = DAG.getNode(ISD::AND, LR.getValueType(), LL, RL); 1144 AddToWorkList(ANDNode.Val); 1145 return DAG.getSetCC(VT, ANDNode, LR, Op1); 1146 } 1147 // fold (X > -1) & (Y > -1) -> (X|Y > -1) 1148 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETGT) { 1149 SDOperand ORNode = DAG.getNode(ISD::OR, LR.getValueType(), LL, RL); 1150 AddToWorkList(ORNode.Val); 1151 return DAG.getSetCC(VT, ORNode, LR, Op1); 1152 } 1153 } 1154 // canonicalize equivalent to ll == rl 1155 if (LL == RR && LR == RL) { 1156 Op1 = ISD::getSetCCSwappedOperands(Op1); 1157 std::swap(RL, RR); 1158 } 1159 if (LL == RL && LR == RR) { 1160 bool isInteger = MVT::isInteger(LL.getValueType()); 1161 ISD::CondCode Result = ISD::getSetCCAndOperation(Op0, Op1, isInteger); 1162 if (Result != ISD::SETCC_INVALID) 1163 return DAG.getSetCC(N0.getValueType(), LL, LR, Result); 1164 } 1165 } 1166 1167 // Simplify: and (op x...), (op y...) -> (op (and x, y)) 1168 if (N0.getOpcode() == N1.getOpcode()) { 1169 SDOperand Tmp = SimplifyBinOpWithSameOpcodeHands(N); 1170 if (Tmp.Val) return Tmp; 1171 } 1172 1173 // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1) 1174 // fold (and (sra)) -> (and (srl)) when possible. 1175 if (!MVT::isVector(VT) && 1176 SimplifyDemandedBits(SDOperand(N, 0))) 1177 return SDOperand(N, 0); 1178 // fold (zext_inreg (extload x)) -> (zextload x) 1179 if (N0.getOpcode() == ISD::EXTLOAD) { 1180 MVT::ValueType EVT = cast<VTSDNode>(N0.getOperand(3))->getVT(); 1181 // If we zero all the possible extended bits, then we can turn this into 1182 // a zextload if we are running before legalize or the operation is legal. 1183 if (TLI.MaskedValueIsZero(N1, ~0ULL << MVT::getSizeInBits(EVT)) && 1184 (!AfterLegalize || TLI.isOperationLegal(ISD::ZEXTLOAD, EVT))) { 1185 SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, N0.getOperand(0), 1186 N0.getOperand(1), N0.getOperand(2), 1187 EVT); 1188 AddToWorkList(N); 1189 CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1)); 1190 return SDOperand(N, 0); // Return N so it doesn't get rechecked! 1191 } 1192 } 1193 // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use 1194 if (N0.getOpcode() == ISD::SEXTLOAD && N0.hasOneUse()) { 1195 MVT::ValueType EVT = cast<VTSDNode>(N0.getOperand(3))->getVT(); 1196 // If we zero all the possible extended bits, then we can turn this into 1197 // a zextload if we are running before legalize or the operation is legal. 1198 if (TLI.MaskedValueIsZero(N1, ~0ULL << MVT::getSizeInBits(EVT)) && 1199 (!AfterLegalize || TLI.isOperationLegal(ISD::ZEXTLOAD, EVT))) { 1200 SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, N0.getOperand(0), 1201 N0.getOperand(1), N0.getOperand(2), 1202 EVT); 1203 AddToWorkList(N); 1204 CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1)); 1205 return SDOperand(N, 0); // Return N so it doesn't get rechecked! 1206 } 1207 } 1208 1209 // fold (and (load x), 255) -> (zextload x, i8) 1210 // fold (and (extload x, i16), 255) -> (zextload x, i8) 1211 if (N1C && 1212 (N0.getOpcode() == ISD::LOAD || N0.getOpcode() == ISD::EXTLOAD || 1213 N0.getOpcode() == ISD::ZEXTLOAD) && 1214 N0.hasOneUse()) { 1215 MVT::ValueType EVT, LoadedVT; 1216 if (N1C->getValue() == 255) 1217 EVT = MVT::i8; 1218 else if (N1C->getValue() == 65535) 1219 EVT = MVT::i16; 1220 else if (N1C->getValue() == ~0U) 1221 EVT = MVT::i32; 1222 else 1223 EVT = MVT::Other; 1224 1225 LoadedVT = N0.getOpcode() == ISD::LOAD ? VT : 1226 cast<VTSDNode>(N0.getOperand(3))->getVT(); 1227 if (EVT != MVT::Other && LoadedVT > EVT && 1228 (!AfterLegalize || TLI.isOperationLegal(ISD::ZEXTLOAD, EVT))) { 1229 MVT::ValueType PtrType = N0.getOperand(1).getValueType(); 1230 // For big endian targets, we need to add an offset to the pointer to load 1231 // the correct bytes. For little endian systems, we merely need to read 1232 // fewer bytes from the same pointer. 1233 unsigned PtrOff = 1234 (MVT::getSizeInBits(LoadedVT) - MVT::getSizeInBits(EVT)) / 8; 1235 SDOperand NewPtr = N0.getOperand(1); 1236 if (!TLI.isLittleEndian()) 1237 NewPtr = DAG.getNode(ISD::ADD, PtrType, NewPtr, 1238 DAG.getConstant(PtrOff, PtrType)); 1239 AddToWorkList(NewPtr.Val); 1240 SDOperand Load = 1241 DAG.getExtLoad(ISD::ZEXTLOAD, VT, N0.getOperand(0), NewPtr, 1242 N0.getOperand(2), EVT); 1243 AddToWorkList(N); 1244 CombineTo(N0.Val, Load, Load.getValue(1)); 1245 return SDOperand(N, 0); // Return N so it doesn't get rechecked! 1246 } 1247 } 1248 1249 return SDOperand(); 1250} 1251 1252SDOperand DAGCombiner::visitOR(SDNode *N) { 1253 SDOperand N0 = N->getOperand(0); 1254 SDOperand N1 = N->getOperand(1); 1255 SDOperand LL, LR, RL, RR, CC0, CC1; 1256 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1257 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 1258 MVT::ValueType VT = N1.getValueType(); 1259 unsigned OpSizeInBits = MVT::getSizeInBits(VT); 1260 1261 // fold (or c1, c2) -> c1|c2 1262 if (N0C && N1C) 1263 return DAG.getNode(ISD::OR, VT, N0, N1); 1264 // canonicalize constant to RHS 1265 if (N0C && !N1C) 1266 return DAG.getNode(ISD::OR, VT, N1, N0); 1267 // fold (or x, 0) -> x 1268 if (N1C && N1C->isNullValue()) 1269 return N0; 1270 // fold (or x, -1) -> -1 1271 if (N1C && N1C->isAllOnesValue()) 1272 return N1; 1273 // fold (or x, c) -> c iff (x & ~c) == 0 1274 if (N1C && 1275 TLI.MaskedValueIsZero(N0,~N1C->getValue() & (~0ULL>>(64-OpSizeInBits)))) 1276 return N1; 1277 // reassociate or 1278 SDOperand ROR = ReassociateOps(ISD::OR, N0, N1); 1279 if (ROR.Val != 0) 1280 return ROR; 1281 // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2) 1282 if (N1C && N0.getOpcode() == ISD::AND && N0.Val->hasOneUse() && 1283 isa<ConstantSDNode>(N0.getOperand(1))) { 1284 ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1)); 1285 return DAG.getNode(ISD::AND, VT, DAG.getNode(ISD::OR, VT, N0.getOperand(0), 1286 N1), 1287 DAG.getConstant(N1C->getValue() | C1->getValue(), VT)); 1288 } 1289 // fold (or (setcc x), (setcc y)) -> (setcc (or x, y)) 1290 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){ 1291 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get(); 1292 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get(); 1293 1294 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 && 1295 MVT::isInteger(LL.getValueType())) { 1296 // fold (X != 0) | (Y != 0) -> (X|Y != 0) 1297 // fold (X < 0) | (Y < 0) -> (X|Y < 0) 1298 if (cast<ConstantSDNode>(LR)->getValue() == 0 && 1299 (Op1 == ISD::SETNE || Op1 == ISD::SETLT)) { 1300 SDOperand ORNode = DAG.getNode(ISD::OR, LR.getValueType(), LL, RL); 1301 AddToWorkList(ORNode.Val); 1302 return DAG.getSetCC(VT, ORNode, LR, Op1); 1303 } 1304 // fold (X != -1) | (Y != -1) -> (X&Y != -1) 1305 // fold (X > -1) | (Y > -1) -> (X&Y > -1) 1306 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && 1307 (Op1 == ISD::SETNE || Op1 == ISD::SETGT)) { 1308 SDOperand ANDNode = DAG.getNode(ISD::AND, LR.getValueType(), LL, RL); 1309 AddToWorkList(ANDNode.Val); 1310 return DAG.getSetCC(VT, ANDNode, LR, Op1); 1311 } 1312 } 1313 // canonicalize equivalent to ll == rl 1314 if (LL == RR && LR == RL) { 1315 Op1 = ISD::getSetCCSwappedOperands(Op1); 1316 std::swap(RL, RR); 1317 } 1318 if (LL == RL && LR == RR) { 1319 bool isInteger = MVT::isInteger(LL.getValueType()); 1320 ISD::CondCode Result = ISD::getSetCCOrOperation(Op0, Op1, isInteger); 1321 if (Result != ISD::SETCC_INVALID) 1322 return DAG.getSetCC(N0.getValueType(), LL, LR, Result); 1323 } 1324 } 1325 1326 // Simplify: or (op x...), (op y...) -> (op (or x, y)) 1327 if (N0.getOpcode() == N1.getOpcode()) { 1328 SDOperand Tmp = SimplifyBinOpWithSameOpcodeHands(N); 1329 if (Tmp.Val) return Tmp; 1330 } 1331 1332 // canonicalize shl to left side in a shl/srl pair, to match rotate 1333 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL) 1334 std::swap(N0, N1); 1335 // check for rotl, rotr 1336 if (N0.getOpcode() == ISD::SHL && N1.getOpcode() == ISD::SRL && 1337 N0.getOperand(0) == N1.getOperand(0) && 1338 TLI.isOperationLegal(ISD::ROTL, VT) && TLI.isTypeLegal(VT)) { 1339 // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1) 1340 if (N0.getOperand(1).getOpcode() == ISD::Constant && 1341 N1.getOperand(1).getOpcode() == ISD::Constant) { 1342 uint64_t c1val = cast<ConstantSDNode>(N0.getOperand(1))->getValue(); 1343 uint64_t c2val = cast<ConstantSDNode>(N1.getOperand(1))->getValue(); 1344 if ((c1val + c2val) == OpSizeInBits) 1345 return DAG.getNode(ISD::ROTL, VT, N0.getOperand(0), N0.getOperand(1)); 1346 } 1347 // fold (or (shl x, y), (srl x, (sub 32, y))) -> (rotl x, y) 1348 if (N1.getOperand(1).getOpcode() == ISD::SUB && 1349 N0.getOperand(1) == N1.getOperand(1).getOperand(1)) 1350 if (ConstantSDNode *SUBC = 1351 dyn_cast<ConstantSDNode>(N1.getOperand(1).getOperand(0))) 1352 if (SUBC->getValue() == OpSizeInBits) 1353 return DAG.getNode(ISD::ROTL, VT, N0.getOperand(0), N0.getOperand(1)); 1354 // fold (or (shl x, (sub 32, y)), (srl x, r)) -> (rotr x, y) 1355 if (N0.getOperand(1).getOpcode() == ISD::SUB && 1356 N1.getOperand(1) == N0.getOperand(1).getOperand(1)) 1357 if (ConstantSDNode *SUBC = 1358 dyn_cast<ConstantSDNode>(N0.getOperand(1).getOperand(0))) 1359 if (SUBC->getValue() == OpSizeInBits) { 1360 if (TLI.isOperationLegal(ISD::ROTR, VT) && TLI.isTypeLegal(VT)) 1361 return DAG.getNode(ISD::ROTR, VT, N0.getOperand(0), 1362 N1.getOperand(1)); 1363 else 1364 return DAG.getNode(ISD::ROTL, VT, N0.getOperand(0), 1365 N0.getOperand(1)); 1366 } 1367 } 1368 return SDOperand(); 1369} 1370 1371SDOperand DAGCombiner::visitXOR(SDNode *N) { 1372 SDOperand N0 = N->getOperand(0); 1373 SDOperand N1 = N->getOperand(1); 1374 SDOperand LHS, RHS, CC; 1375 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1376 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 1377 MVT::ValueType VT = N0.getValueType(); 1378 1379 // fold (xor c1, c2) -> c1^c2 1380 if (N0C && N1C) 1381 return DAG.getNode(ISD::XOR, VT, N0, N1); 1382 // canonicalize constant to RHS 1383 if (N0C && !N1C) 1384 return DAG.getNode(ISD::XOR, VT, N1, N0); 1385 // fold (xor x, 0) -> x 1386 if (N1C && N1C->isNullValue()) 1387 return N0; 1388 // reassociate xor 1389 SDOperand RXOR = ReassociateOps(ISD::XOR, N0, N1); 1390 if (RXOR.Val != 0) 1391 return RXOR; 1392 // fold !(x cc y) -> (x !cc y) 1393 if (N1C && N1C->getValue() == 1 && isSetCCEquivalent(N0, LHS, RHS, CC)) { 1394 bool isInt = MVT::isInteger(LHS.getValueType()); 1395 ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(), 1396 isInt); 1397 if (N0.getOpcode() == ISD::SETCC) 1398 return DAG.getSetCC(VT, LHS, RHS, NotCC); 1399 if (N0.getOpcode() == ISD::SELECT_CC) 1400 return DAG.getSelectCC(LHS, RHS, N0.getOperand(2),N0.getOperand(3),NotCC); 1401 assert(0 && "Unhandled SetCC Equivalent!"); 1402 abort(); 1403 } 1404 // fold !(x or y) -> (!x and !y) iff x or y are setcc 1405 if (N1C && N1C->getValue() == 1 && 1406 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) { 1407 SDOperand LHS = N0.getOperand(0), RHS = N0.getOperand(1); 1408 if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) { 1409 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND; 1410 LHS = DAG.getNode(ISD::XOR, VT, LHS, N1); // RHS = ~LHS 1411 RHS = DAG.getNode(ISD::XOR, VT, RHS, N1); // RHS = ~RHS 1412 AddToWorkList(LHS.Val); AddToWorkList(RHS.Val); 1413 return DAG.getNode(NewOpcode, VT, LHS, RHS); 1414 } 1415 } 1416 // fold !(x or y) -> (!x and !y) iff x or y are constants 1417 if (N1C && N1C->isAllOnesValue() && 1418 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) { 1419 SDOperand LHS = N0.getOperand(0), RHS = N0.getOperand(1); 1420 if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) { 1421 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND; 1422 LHS = DAG.getNode(ISD::XOR, VT, LHS, N1); // RHS = ~LHS 1423 RHS = DAG.getNode(ISD::XOR, VT, RHS, N1); // RHS = ~RHS 1424 AddToWorkList(LHS.Val); AddToWorkList(RHS.Val); 1425 return DAG.getNode(NewOpcode, VT, LHS, RHS); 1426 } 1427 } 1428 // fold (xor (xor x, c1), c2) -> (xor x, c1^c2) 1429 if (N1C && N0.getOpcode() == ISD::XOR) { 1430 ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0)); 1431 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1)); 1432 if (N00C) 1433 return DAG.getNode(ISD::XOR, VT, N0.getOperand(1), 1434 DAG.getConstant(N1C->getValue()^N00C->getValue(), VT)); 1435 if (N01C) 1436 return DAG.getNode(ISD::XOR, VT, N0.getOperand(0), 1437 DAG.getConstant(N1C->getValue()^N01C->getValue(), VT)); 1438 } 1439 // fold (xor x, x) -> 0 1440 if (N0 == N1) { 1441 if (!MVT::isVector(VT)) { 1442 return DAG.getConstant(0, VT); 1443 } else if (!AfterLegalize || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT)) { 1444 // Produce a vector of zeros. 1445 SDOperand El = DAG.getConstant(0, MVT::getVectorBaseType(VT)); 1446 std::vector<SDOperand> Ops(MVT::getVectorNumElements(VT), El); 1447 return DAG.getNode(ISD::BUILD_VECTOR, VT, Ops); 1448 } 1449 } 1450 1451 // Simplify: xor (op x...), (op y...) -> (op (xor x, y)) 1452 if (N0.getOpcode() == N1.getOpcode()) { 1453 SDOperand Tmp = SimplifyBinOpWithSameOpcodeHands(N); 1454 if (Tmp.Val) return Tmp; 1455 } 1456 1457 // Simplify the expression using non-local knowledge. 1458 if (!MVT::isVector(VT) && 1459 SimplifyDemandedBits(SDOperand(N, 0))) 1460 return SDOperand(N, 0); 1461 1462 return SDOperand(); 1463} 1464 1465SDOperand DAGCombiner::visitSHL(SDNode *N) { 1466 SDOperand N0 = N->getOperand(0); 1467 SDOperand N1 = N->getOperand(1); 1468 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1469 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 1470 MVT::ValueType VT = N0.getValueType(); 1471 unsigned OpSizeInBits = MVT::getSizeInBits(VT); 1472 1473 // fold (shl c1, c2) -> c1<<c2 1474 if (N0C && N1C) 1475 return DAG.getNode(ISD::SHL, VT, N0, N1); 1476 // fold (shl 0, x) -> 0 1477 if (N0C && N0C->isNullValue()) 1478 return N0; 1479 // fold (shl x, c >= size(x)) -> undef 1480 if (N1C && N1C->getValue() >= OpSizeInBits) 1481 return DAG.getNode(ISD::UNDEF, VT); 1482 // fold (shl x, 0) -> x 1483 if (N1C && N1C->isNullValue()) 1484 return N0; 1485 // if (shl x, c) is known to be zero, return 0 1486 if (TLI.MaskedValueIsZero(SDOperand(N, 0), MVT::getIntVTBitMask(VT))) 1487 return DAG.getConstant(0, VT); 1488 if (SimplifyDemandedBits(SDOperand(N, 0))) 1489 return SDOperand(N, 0); 1490 // fold (shl (shl x, c1), c2) -> 0 or (shl x, c1+c2) 1491 if (N1C && N0.getOpcode() == ISD::SHL && 1492 N0.getOperand(1).getOpcode() == ISD::Constant) { 1493 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getValue(); 1494 uint64_t c2 = N1C->getValue(); 1495 if (c1 + c2 > OpSizeInBits) 1496 return DAG.getConstant(0, VT); 1497 return DAG.getNode(ISD::SHL, VT, N0.getOperand(0), 1498 DAG.getConstant(c1 + c2, N1.getValueType())); 1499 } 1500 // fold (shl (srl x, c1), c2) -> (shl (and x, -1 << c1), c2-c1) or 1501 // (srl (and x, -1 << c1), c1-c2) 1502 if (N1C && N0.getOpcode() == ISD::SRL && 1503 N0.getOperand(1).getOpcode() == ISD::Constant) { 1504 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getValue(); 1505 uint64_t c2 = N1C->getValue(); 1506 SDOperand Mask = DAG.getNode(ISD::AND, VT, N0.getOperand(0), 1507 DAG.getConstant(~0ULL << c1, VT)); 1508 if (c2 > c1) 1509 return DAG.getNode(ISD::SHL, VT, Mask, 1510 DAG.getConstant(c2-c1, N1.getValueType())); 1511 else 1512 return DAG.getNode(ISD::SRL, VT, Mask, 1513 DAG.getConstant(c1-c2, N1.getValueType())); 1514 } 1515 // fold (shl (sra x, c1), c1) -> (and x, -1 << c1) 1516 if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1)) 1517 return DAG.getNode(ISD::AND, VT, N0.getOperand(0), 1518 DAG.getConstant(~0ULL << N1C->getValue(), VT)); 1519 // fold (shl (add x, c1), c2) -> (add (shl x, c2), c1<<c2) 1520 if (N1C && N0.getOpcode() == ISD::ADD && N0.Val->hasOneUse() && 1521 isa<ConstantSDNode>(N0.getOperand(1))) { 1522 return DAG.getNode(ISD::ADD, VT, 1523 DAG.getNode(ISD::SHL, VT, N0.getOperand(0), N1), 1524 DAG.getNode(ISD::SHL, VT, N0.getOperand(1), N1)); 1525 } 1526 return SDOperand(); 1527} 1528 1529SDOperand DAGCombiner::visitSRA(SDNode *N) { 1530 SDOperand N0 = N->getOperand(0); 1531 SDOperand N1 = N->getOperand(1); 1532 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1533 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 1534 MVT::ValueType VT = N0.getValueType(); 1535 1536 // fold (sra c1, c2) -> c1>>c2 1537 if (N0C && N1C) 1538 return DAG.getNode(ISD::SRA, VT, N0, N1); 1539 // fold (sra 0, x) -> 0 1540 if (N0C && N0C->isNullValue()) 1541 return N0; 1542 // fold (sra -1, x) -> -1 1543 if (N0C && N0C->isAllOnesValue()) 1544 return N0; 1545 // fold (sra x, c >= size(x)) -> undef 1546 if (N1C && N1C->getValue() >= MVT::getSizeInBits(VT)) 1547 return DAG.getNode(ISD::UNDEF, VT); 1548 // fold (sra x, 0) -> x 1549 if (N1C && N1C->isNullValue()) 1550 return N0; 1551 // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports 1552 // sext_inreg. 1553 if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) { 1554 unsigned LowBits = MVT::getSizeInBits(VT) - (unsigned)N1C->getValue(); 1555 MVT::ValueType EVT; 1556 switch (LowBits) { 1557 default: EVT = MVT::Other; break; 1558 case 1: EVT = MVT::i1; break; 1559 case 8: EVT = MVT::i8; break; 1560 case 16: EVT = MVT::i16; break; 1561 case 32: EVT = MVT::i32; break; 1562 } 1563 if (EVT > MVT::Other && TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, EVT)) 1564 return DAG.getNode(ISD::SIGN_EXTEND_INREG, VT, N0.getOperand(0), 1565 DAG.getValueType(EVT)); 1566 } 1567 1568 // fold (sra (sra x, c1), c2) -> (sra x, c1+c2) 1569 if (N1C && N0.getOpcode() == ISD::SRA) { 1570 if (ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) { 1571 unsigned Sum = N1C->getValue() + C1->getValue(); 1572 if (Sum >= MVT::getSizeInBits(VT)) Sum = MVT::getSizeInBits(VT)-1; 1573 return DAG.getNode(ISD::SRA, VT, N0.getOperand(0), 1574 DAG.getConstant(Sum, N1C->getValueType(0))); 1575 } 1576 } 1577 1578 // Simplify, based on bits shifted out of the LHS. 1579 if (N1C && SimplifyDemandedBits(SDOperand(N, 0))) 1580 return SDOperand(N, 0); 1581 1582 1583 // If the sign bit is known to be zero, switch this to a SRL. 1584 if (TLI.MaskedValueIsZero(N0, MVT::getIntVTSignBit(VT))) 1585 return DAG.getNode(ISD::SRL, VT, N0, N1); 1586 return SDOperand(); 1587} 1588 1589SDOperand DAGCombiner::visitSRL(SDNode *N) { 1590 SDOperand N0 = N->getOperand(0); 1591 SDOperand N1 = N->getOperand(1); 1592 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1593 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 1594 MVT::ValueType VT = N0.getValueType(); 1595 unsigned OpSizeInBits = MVT::getSizeInBits(VT); 1596 1597 // fold (srl c1, c2) -> c1 >>u c2 1598 if (N0C && N1C) 1599 return DAG.getNode(ISD::SRL, VT, N0, N1); 1600 // fold (srl 0, x) -> 0 1601 if (N0C && N0C->isNullValue()) 1602 return N0; 1603 // fold (srl x, c >= size(x)) -> undef 1604 if (N1C && N1C->getValue() >= OpSizeInBits) 1605 return DAG.getNode(ISD::UNDEF, VT); 1606 // fold (srl x, 0) -> x 1607 if (N1C && N1C->isNullValue()) 1608 return N0; 1609 // if (srl x, c) is known to be zero, return 0 1610 if (N1C && TLI.MaskedValueIsZero(SDOperand(N, 0), ~0ULL >> (64-OpSizeInBits))) 1611 return DAG.getConstant(0, VT); 1612 // fold (srl (srl x, c1), c2) -> 0 or (srl x, c1+c2) 1613 if (N1C && N0.getOpcode() == ISD::SRL && 1614 N0.getOperand(1).getOpcode() == ISD::Constant) { 1615 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getValue(); 1616 uint64_t c2 = N1C->getValue(); 1617 if (c1 + c2 > OpSizeInBits) 1618 return DAG.getConstant(0, VT); 1619 return DAG.getNode(ISD::SRL, VT, N0.getOperand(0), 1620 DAG.getConstant(c1 + c2, N1.getValueType())); 1621 } 1622 1623 // fold (srl (anyextend x), c) -> (anyextend (srl x, c)) 1624 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) { 1625 // Shifting in all undef bits? 1626 MVT::ValueType SmallVT = N0.getOperand(0).getValueType(); 1627 if (N1C->getValue() >= MVT::getSizeInBits(SmallVT)) 1628 return DAG.getNode(ISD::UNDEF, VT); 1629 1630 SDOperand SmallShift = DAG.getNode(ISD::SRL, SmallVT, N0.getOperand(0), N1); 1631 AddToWorkList(SmallShift.Val); 1632 return DAG.getNode(ISD::ANY_EXTEND, VT, SmallShift); 1633 } 1634 1635 // fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit). 1636 if (N1C && N0.getOpcode() == ISD::CTLZ && 1637 N1C->getValue() == Log2_32(MVT::getSizeInBits(VT))) { 1638 uint64_t KnownZero, KnownOne, Mask = MVT::getIntVTBitMask(VT); 1639 TLI.ComputeMaskedBits(N0.getOperand(0), Mask, KnownZero, KnownOne); 1640 1641 // If any of the input bits are KnownOne, then the input couldn't be all 1642 // zeros, thus the result of the srl will always be zero. 1643 if (KnownOne) return DAG.getConstant(0, VT); 1644 1645 // If all of the bits input the to ctlz node are known to be zero, then 1646 // the result of the ctlz is "32" and the result of the shift is one. 1647 uint64_t UnknownBits = ~KnownZero & Mask; 1648 if (UnknownBits == 0) return DAG.getConstant(1, VT); 1649 1650 // Otherwise, check to see if there is exactly one bit input to the ctlz. 1651 if ((UnknownBits & (UnknownBits-1)) == 0) { 1652 // Okay, we know that only that the single bit specified by UnknownBits 1653 // could be set on input to the CTLZ node. If this bit is set, the SRL 1654 // will return 0, if it is clear, it returns 1. Change the CTLZ/SRL pair 1655 // to an SRL,XOR pair, which is likely to simplify more. 1656 unsigned ShAmt = CountTrailingZeros_64(UnknownBits); 1657 SDOperand Op = N0.getOperand(0); 1658 if (ShAmt) { 1659 Op = DAG.getNode(ISD::SRL, VT, Op, 1660 DAG.getConstant(ShAmt, TLI.getShiftAmountTy())); 1661 AddToWorkList(Op.Val); 1662 } 1663 return DAG.getNode(ISD::XOR, VT, Op, DAG.getConstant(1, VT)); 1664 } 1665 } 1666 1667 return SDOperand(); 1668} 1669 1670SDOperand DAGCombiner::visitCTLZ(SDNode *N) { 1671 SDOperand N0 = N->getOperand(0); 1672 MVT::ValueType VT = N->getValueType(0); 1673 1674 // fold (ctlz c1) -> c2 1675 if (isa<ConstantSDNode>(N0)) 1676 return DAG.getNode(ISD::CTLZ, VT, N0); 1677 return SDOperand(); 1678} 1679 1680SDOperand DAGCombiner::visitCTTZ(SDNode *N) { 1681 SDOperand N0 = N->getOperand(0); 1682 MVT::ValueType VT = N->getValueType(0); 1683 1684 // fold (cttz c1) -> c2 1685 if (isa<ConstantSDNode>(N0)) 1686 return DAG.getNode(ISD::CTTZ, VT, N0); 1687 return SDOperand(); 1688} 1689 1690SDOperand DAGCombiner::visitCTPOP(SDNode *N) { 1691 SDOperand N0 = N->getOperand(0); 1692 MVT::ValueType VT = N->getValueType(0); 1693 1694 // fold (ctpop c1) -> c2 1695 if (isa<ConstantSDNode>(N0)) 1696 return DAG.getNode(ISD::CTPOP, VT, N0); 1697 return SDOperand(); 1698} 1699 1700SDOperand DAGCombiner::visitSELECT(SDNode *N) { 1701 SDOperand N0 = N->getOperand(0); 1702 SDOperand N1 = N->getOperand(1); 1703 SDOperand N2 = N->getOperand(2); 1704 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1705 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 1706 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2); 1707 MVT::ValueType VT = N->getValueType(0); 1708 1709 // fold select C, X, X -> X 1710 if (N1 == N2) 1711 return N1; 1712 // fold select true, X, Y -> X 1713 if (N0C && !N0C->isNullValue()) 1714 return N1; 1715 // fold select false, X, Y -> Y 1716 if (N0C && N0C->isNullValue()) 1717 return N2; 1718 // fold select C, 1, X -> C | X 1719 if (MVT::i1 == VT && N1C && N1C->getValue() == 1) 1720 return DAG.getNode(ISD::OR, VT, N0, N2); 1721 // fold select C, 0, X -> ~C & X 1722 // FIXME: this should check for C type == X type, not i1? 1723 if (MVT::i1 == VT && N1C && N1C->isNullValue()) { 1724 SDOperand XORNode = DAG.getNode(ISD::XOR, VT, N0, DAG.getConstant(1, VT)); 1725 AddToWorkList(XORNode.Val); 1726 return DAG.getNode(ISD::AND, VT, XORNode, N2); 1727 } 1728 // fold select C, X, 1 -> ~C | X 1729 if (MVT::i1 == VT && N2C && N2C->getValue() == 1) { 1730 SDOperand XORNode = DAG.getNode(ISD::XOR, VT, N0, DAG.getConstant(1, VT)); 1731 AddToWorkList(XORNode.Val); 1732 return DAG.getNode(ISD::OR, VT, XORNode, N1); 1733 } 1734 // fold select C, X, 0 -> C & X 1735 // FIXME: this should check for C type == X type, not i1? 1736 if (MVT::i1 == VT && N2C && N2C->isNullValue()) 1737 return DAG.getNode(ISD::AND, VT, N0, N1); 1738 // fold X ? X : Y --> X ? 1 : Y --> X | Y 1739 if (MVT::i1 == VT && N0 == N1) 1740 return DAG.getNode(ISD::OR, VT, N0, N2); 1741 // fold X ? Y : X --> X ? Y : 0 --> X & Y 1742 if (MVT::i1 == VT && N0 == N2) 1743 return DAG.getNode(ISD::AND, VT, N0, N1); 1744 // If we can fold this based on the true/false value, do so. 1745 if (SimplifySelectOps(N, N1, N2)) 1746 return SDOperand(); 1747 // fold selects based on a setcc into other things, such as min/max/abs 1748 if (N0.getOpcode() == ISD::SETCC) 1749 // FIXME: 1750 // Check against MVT::Other for SELECT_CC, which is a workaround for targets 1751 // having to say they don't support SELECT_CC on every type the DAG knows 1752 // about, since there is no way to mark an opcode illegal at all value types 1753 if (TLI.isOperationLegal(ISD::SELECT_CC, MVT::Other)) 1754 return DAG.getNode(ISD::SELECT_CC, VT, N0.getOperand(0), N0.getOperand(1), 1755 N1, N2, N0.getOperand(2)); 1756 else 1757 return SimplifySelect(N0, N1, N2); 1758 return SDOperand(); 1759} 1760 1761SDOperand DAGCombiner::visitSELECT_CC(SDNode *N) { 1762 SDOperand N0 = N->getOperand(0); 1763 SDOperand N1 = N->getOperand(1); 1764 SDOperand N2 = N->getOperand(2); 1765 SDOperand N3 = N->getOperand(3); 1766 SDOperand N4 = N->getOperand(4); 1767 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1768 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 1769 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2); 1770 ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get(); 1771 1772 // Determine if the condition we're dealing with is constant 1773 SDOperand SCC = SimplifySetCC(TLI.getSetCCResultTy(), N0, N1, CC, false); 1774 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.Val); 1775 1776 // fold select_cc lhs, rhs, x, x, cc -> x 1777 if (N2 == N3) 1778 return N2; 1779 1780 // If we can fold this based on the true/false value, do so. 1781 if (SimplifySelectOps(N, N2, N3)) 1782 return SDOperand(); 1783 1784 // fold select_cc into other things, such as min/max/abs 1785 return SimplifySelectCC(N0, N1, N2, N3, CC); 1786} 1787 1788SDOperand DAGCombiner::visitSETCC(SDNode *N) { 1789 return SimplifySetCC(N->getValueType(0), N->getOperand(0), N->getOperand(1), 1790 cast<CondCodeSDNode>(N->getOperand(2))->get()); 1791} 1792 1793SDOperand DAGCombiner::visitSIGN_EXTEND(SDNode *N) { 1794 SDOperand N0 = N->getOperand(0); 1795 MVT::ValueType VT = N->getValueType(0); 1796 1797 // fold (sext c1) -> c1 1798 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0)) 1799 return DAG.getNode(ISD::SIGN_EXTEND, VT, N0); 1800 1801 // fold (sext (sext x)) -> (sext x) 1802 // fold (sext (aext x)) -> (sext x) 1803 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) 1804 return DAG.getNode(ISD::SIGN_EXTEND, VT, N0.getOperand(0)); 1805 1806 // fold (sext (truncate x)) -> (sextinreg x) iff x size == sext size. 1807 if (N0.getOpcode() == ISD::TRUNCATE && N0.getOperand(0).getValueType() == VT&& 1808 (!AfterLegalize || 1809 TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, N0.getValueType()))) 1810 return DAG.getNode(ISD::SIGN_EXTEND_INREG, VT, N0.getOperand(0), 1811 DAG.getValueType(N0.getValueType())); 1812 1813 // fold (sext (load x)) -> (sext (truncate (sextload x))) 1814 if (N0.getOpcode() == ISD::LOAD && N0.hasOneUse() && 1815 (!AfterLegalize||TLI.isOperationLegal(ISD::SEXTLOAD, N0.getValueType()))){ 1816 SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, N0.getOperand(0), 1817 N0.getOperand(1), N0.getOperand(2), 1818 N0.getValueType()); 1819 CombineTo(N, ExtLoad); 1820 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad), 1821 ExtLoad.getValue(1)); 1822 return SDOperand(N, 0); // Return N so it doesn't get rechecked! 1823 } 1824 1825 // fold (sext (sextload x)) -> (sext (truncate (sextload x))) 1826 // fold (sext ( extload x)) -> (sext (truncate (sextload x))) 1827 if ((N0.getOpcode() == ISD::SEXTLOAD || N0.getOpcode() == ISD::EXTLOAD) && 1828 N0.hasOneUse()) { 1829 MVT::ValueType EVT = cast<VTSDNode>(N0.getOperand(3))->getVT(); 1830 SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, N0.getOperand(0), 1831 N0.getOperand(1), N0.getOperand(2), EVT); 1832 CombineTo(N, ExtLoad); 1833 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad), 1834 ExtLoad.getValue(1)); 1835 return SDOperand(N, 0); // Return N so it doesn't get rechecked! 1836 } 1837 1838 return SDOperand(); 1839} 1840 1841SDOperand DAGCombiner::visitZERO_EXTEND(SDNode *N) { 1842 SDOperand N0 = N->getOperand(0); 1843 MVT::ValueType VT = N->getValueType(0); 1844 1845 // fold (zext c1) -> c1 1846 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0)) 1847 return DAG.getNode(ISD::ZERO_EXTEND, VT, N0); 1848 // fold (zext (zext x)) -> (zext x) 1849 // fold (zext (aext x)) -> (zext x) 1850 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) 1851 return DAG.getNode(ISD::ZERO_EXTEND, VT, N0.getOperand(0)); 1852 // fold (zext (truncate x)) -> (zextinreg x) iff x size == zext size. 1853 if (N0.getOpcode() == ISD::TRUNCATE && N0.getOperand(0).getValueType() == VT&& 1854 (!AfterLegalize || TLI.isOperationLegal(ISD::AND, N0.getValueType()))) 1855 return DAG.getZeroExtendInReg(N0.getOperand(0), N0.getValueType()); 1856 // fold (zext (load x)) -> (zext (truncate (zextload x))) 1857 if (N0.getOpcode() == ISD::LOAD && N0.hasOneUse() && 1858 (!AfterLegalize||TLI.isOperationLegal(ISD::ZEXTLOAD, N0.getValueType()))){ 1859 SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, N0.getOperand(0), 1860 N0.getOperand(1), N0.getOperand(2), 1861 N0.getValueType()); 1862 CombineTo(N, ExtLoad); 1863 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad), 1864 ExtLoad.getValue(1)); 1865 return SDOperand(N, 0); // Return N so it doesn't get rechecked! 1866 } 1867 1868 // fold (zext (zextload x)) -> (zext (truncate (zextload x))) 1869 // fold (zext ( extload x)) -> (zext (truncate (zextload x))) 1870 if ((N0.getOpcode() == ISD::ZEXTLOAD || N0.getOpcode() == ISD::EXTLOAD) && 1871 N0.hasOneUse()) { 1872 MVT::ValueType EVT = cast<VTSDNode>(N0.getOperand(3))->getVT(); 1873 SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, N0.getOperand(0), 1874 N0.getOperand(1), N0.getOperand(2), EVT); 1875 CombineTo(N, ExtLoad); 1876 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad), 1877 ExtLoad.getValue(1)); 1878 return SDOperand(N, 0); // Return N so it doesn't get rechecked! 1879 } 1880 return SDOperand(); 1881} 1882 1883SDOperand DAGCombiner::visitANY_EXTEND(SDNode *N) { 1884 SDOperand N0 = N->getOperand(0); 1885 MVT::ValueType VT = N->getValueType(0); 1886 1887 // fold (aext c1) -> c1 1888 if (isa<ConstantSDNode>(N0)) 1889 return DAG.getNode(ISD::ANY_EXTEND, VT, N0); 1890 // fold (aext (aext x)) -> (aext x) 1891 // fold (aext (zext x)) -> (zext x) 1892 // fold (aext (sext x)) -> (sext x) 1893 if (N0.getOpcode() == ISD::ANY_EXTEND || 1894 N0.getOpcode() == ISD::ZERO_EXTEND || 1895 N0.getOpcode() == ISD::SIGN_EXTEND) 1896 return DAG.getNode(N0.getOpcode(), VT, N0.getOperand(0)); 1897 1898 // fold (aext (truncate x)) -> x iff x size == zext size. 1899 if (N0.getOpcode() == ISD::TRUNCATE && N0.getOperand(0).getValueType() == VT) 1900 return N0.getOperand(0); 1901 // fold (aext (load x)) -> (aext (truncate (extload x))) 1902 if (N0.getOpcode() == ISD::LOAD && N0.hasOneUse() && 1903 (!AfterLegalize||TLI.isOperationLegal(ISD::EXTLOAD, N0.getValueType()))) { 1904 SDOperand ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, VT, N0.getOperand(0), 1905 N0.getOperand(1), N0.getOperand(2), 1906 N0.getValueType()); 1907 CombineTo(N, ExtLoad); 1908 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad), 1909 ExtLoad.getValue(1)); 1910 return SDOperand(N, 0); // Return N so it doesn't get rechecked! 1911 } 1912 1913 // fold (aext (zextload x)) -> (aext (truncate (zextload x))) 1914 // fold (aext (sextload x)) -> (aext (truncate (sextload x))) 1915 // fold (aext ( extload x)) -> (aext (truncate (extload x))) 1916 if ((N0.getOpcode() == ISD::ZEXTLOAD || N0.getOpcode() == ISD::EXTLOAD || 1917 N0.getOpcode() == ISD::SEXTLOAD) && 1918 N0.hasOneUse()) { 1919 MVT::ValueType EVT = cast<VTSDNode>(N0.getOperand(3))->getVT(); 1920 SDOperand ExtLoad = DAG.getExtLoad(N0.getOpcode(), VT, N0.getOperand(0), 1921 N0.getOperand(1), N0.getOperand(2), EVT); 1922 CombineTo(N, ExtLoad); 1923 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad), 1924 ExtLoad.getValue(1)); 1925 return SDOperand(N, 0); // Return N so it doesn't get rechecked! 1926 } 1927 return SDOperand(); 1928} 1929 1930 1931SDOperand DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) { 1932 SDOperand N0 = N->getOperand(0); 1933 SDOperand N1 = N->getOperand(1); 1934 MVT::ValueType VT = N->getValueType(0); 1935 MVT::ValueType EVT = cast<VTSDNode>(N1)->getVT(); 1936 unsigned EVTBits = MVT::getSizeInBits(EVT); 1937 1938 // fold (sext_in_reg c1) -> c1 1939 if (isa<ConstantSDNode>(N0) || N0.getOpcode() == ISD::UNDEF) 1940 return DAG.getNode(ISD::SIGN_EXTEND_INREG, VT, N0, N1); 1941 1942 // If the input is already sign extended, just drop the extension. 1943 if (TLI.ComputeNumSignBits(N0) >= MVT::getSizeInBits(VT)-EVTBits+1) 1944 return N0; 1945 1946 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2 1947 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG && 1948 EVT < cast<VTSDNode>(N0.getOperand(1))->getVT()) { 1949 return DAG.getNode(ISD::SIGN_EXTEND_INREG, VT, N0.getOperand(0), N1); 1950 } 1951 1952 // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is zero 1953 if (TLI.MaskedValueIsZero(N0, 1ULL << (EVTBits-1))) 1954 return DAG.getZeroExtendInReg(N0, EVT); 1955 1956 // fold (sext_in_reg (srl X, 24), i8) -> sra X, 24 1957 // fold (sext_in_reg (srl X, 23), i8) -> sra X, 23 iff possible. 1958 // We already fold "(sext_in_reg (srl X, 25), i8) -> srl X, 25" above. 1959 if (N0.getOpcode() == ISD::SRL) { 1960 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1))) 1961 if (ShAmt->getValue()+EVTBits <= MVT::getSizeInBits(VT)) { 1962 // We can turn this into an SRA iff the input to the SRL is already sign 1963 // extended enough. 1964 unsigned InSignBits = TLI.ComputeNumSignBits(N0.getOperand(0)); 1965 if (MVT::getSizeInBits(VT)-(ShAmt->getValue()+EVTBits) < InSignBits) 1966 return DAG.getNode(ISD::SRA, VT, N0.getOperand(0), N0.getOperand(1)); 1967 } 1968 } 1969 1970 // fold (sext_inreg (extload x)) -> (sextload x) 1971 if (N0.getOpcode() == ISD::EXTLOAD && 1972 EVT == cast<VTSDNode>(N0.getOperand(3))->getVT() && 1973 (!AfterLegalize || TLI.isOperationLegal(ISD::SEXTLOAD, EVT))) { 1974 SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, N0.getOperand(0), 1975 N0.getOperand(1), N0.getOperand(2), 1976 EVT); 1977 CombineTo(N, ExtLoad); 1978 CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1)); 1979 return SDOperand(N, 0); // Return N so it doesn't get rechecked! 1980 } 1981 // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use 1982 if (N0.getOpcode() == ISD::ZEXTLOAD && N0.hasOneUse() && 1983 EVT == cast<VTSDNode>(N0.getOperand(3))->getVT() && 1984 (!AfterLegalize || TLI.isOperationLegal(ISD::SEXTLOAD, EVT))) { 1985 SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, N0.getOperand(0), 1986 N0.getOperand(1), N0.getOperand(2), 1987 EVT); 1988 CombineTo(N, ExtLoad); 1989 CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1)); 1990 return SDOperand(N, 0); // Return N so it doesn't get rechecked! 1991 } 1992 return SDOperand(); 1993} 1994 1995SDOperand DAGCombiner::visitTRUNCATE(SDNode *N) { 1996 SDOperand N0 = N->getOperand(0); 1997 MVT::ValueType VT = N->getValueType(0); 1998 1999 // noop truncate 2000 if (N0.getValueType() == N->getValueType(0)) 2001 return N0; 2002 // fold (truncate c1) -> c1 2003 if (isa<ConstantSDNode>(N0)) 2004 return DAG.getNode(ISD::TRUNCATE, VT, N0); 2005 // fold (truncate (truncate x)) -> (truncate x) 2006 if (N0.getOpcode() == ISD::TRUNCATE) 2007 return DAG.getNode(ISD::TRUNCATE, VT, N0.getOperand(0)); 2008 // fold (truncate (ext x)) -> (ext x) or (truncate x) or x 2009 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::SIGN_EXTEND|| 2010 N0.getOpcode() == ISD::ANY_EXTEND) { 2011 if (N0.getValueType() < VT) 2012 // if the source is smaller than the dest, we still need an extend 2013 return DAG.getNode(N0.getOpcode(), VT, N0.getOperand(0)); 2014 else if (N0.getValueType() > VT) 2015 // if the source is larger than the dest, than we just need the truncate 2016 return DAG.getNode(ISD::TRUNCATE, VT, N0.getOperand(0)); 2017 else 2018 // if the source and dest are the same type, we can drop both the extend 2019 // and the truncate 2020 return N0.getOperand(0); 2021 } 2022 // fold (truncate (load x)) -> (smaller load x) 2023 if (N0.getOpcode() == ISD::LOAD && N0.hasOneUse()) { 2024 assert(MVT::getSizeInBits(N0.getValueType()) > MVT::getSizeInBits(VT) && 2025 "Cannot truncate to larger type!"); 2026 MVT::ValueType PtrType = N0.getOperand(1).getValueType(); 2027 // For big endian targets, we need to add an offset to the pointer to load 2028 // the correct bytes. For little endian systems, we merely need to read 2029 // fewer bytes from the same pointer. 2030 uint64_t PtrOff = 2031 (MVT::getSizeInBits(N0.getValueType()) - MVT::getSizeInBits(VT)) / 8; 2032 SDOperand NewPtr = TLI.isLittleEndian() ? N0.getOperand(1) : 2033 DAG.getNode(ISD::ADD, PtrType, N0.getOperand(1), 2034 DAG.getConstant(PtrOff, PtrType)); 2035 AddToWorkList(NewPtr.Val); 2036 SDOperand Load = DAG.getLoad(VT, N0.getOperand(0), NewPtr,N0.getOperand(2)); 2037 AddToWorkList(N); 2038 CombineTo(N0.Val, Load, Load.getValue(1)); 2039 return SDOperand(N, 0); // Return N so it doesn't get rechecked! 2040 } 2041 return SDOperand(); 2042} 2043 2044SDOperand DAGCombiner::visitBIT_CONVERT(SDNode *N) { 2045 SDOperand N0 = N->getOperand(0); 2046 MVT::ValueType VT = N->getValueType(0); 2047 2048 // If the input is a constant, let getNode() fold it. 2049 if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) { 2050 SDOperand Res = DAG.getNode(ISD::BIT_CONVERT, VT, N0); 2051 if (Res.Val != N) return Res; 2052 } 2053 2054 if (N0.getOpcode() == ISD::BIT_CONVERT) // conv(conv(x,t1),t2) -> conv(x,t2) 2055 return DAG.getNode(ISD::BIT_CONVERT, VT, N0.getOperand(0)); 2056 2057 // fold (conv (load x)) -> (load (conv*)x) 2058 // FIXME: These xforms need to know that the resultant load doesn't need a 2059 // higher alignment than the original! 2060 if (0 && N0.getOpcode() == ISD::LOAD && N0.hasOneUse()) { 2061 SDOperand Load = DAG.getLoad(VT, N0.getOperand(0), N0.getOperand(1), 2062 N0.getOperand(2)); 2063 AddToWorkList(N); 2064 CombineTo(N0.Val, DAG.getNode(ISD::BIT_CONVERT, N0.getValueType(), Load), 2065 Load.getValue(1)); 2066 return Load; 2067 } 2068 2069 return SDOperand(); 2070} 2071 2072SDOperand DAGCombiner::visitVBIT_CONVERT(SDNode *N) { 2073 SDOperand N0 = N->getOperand(0); 2074 MVT::ValueType VT = N->getValueType(0); 2075 2076 // If the input is a VBUILD_VECTOR with all constant elements, fold this now. 2077 // First check to see if this is all constant. 2078 if (N0.getOpcode() == ISD::VBUILD_VECTOR && N0.Val->hasOneUse() && 2079 VT == MVT::Vector) { 2080 bool isSimple = true; 2081 for (unsigned i = 0, e = N0.getNumOperands()-2; i != e; ++i) 2082 if (N0.getOperand(i).getOpcode() != ISD::UNDEF && 2083 N0.getOperand(i).getOpcode() != ISD::Constant && 2084 N0.getOperand(i).getOpcode() != ISD::ConstantFP) { 2085 isSimple = false; 2086 break; 2087 } 2088 2089 MVT::ValueType DestEltVT = cast<VTSDNode>(N->getOperand(2))->getVT(); 2090 if (isSimple && !MVT::isVector(DestEltVT)) { 2091 return ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(N0.Val, DestEltVT); 2092 } 2093 } 2094 2095 return SDOperand(); 2096} 2097 2098/// ConstantFoldVBIT_CONVERTofVBUILD_VECTOR - We know that BV is a vbuild_vector 2099/// node with Constant, ConstantFP or Undef operands. DstEltVT indicates the 2100/// destination element value type. 2101SDOperand DAGCombiner:: 2102ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(SDNode *BV, MVT::ValueType DstEltVT) { 2103 MVT::ValueType SrcEltVT = BV->getOperand(0).getValueType(); 2104 2105 // If this is already the right type, we're done. 2106 if (SrcEltVT == DstEltVT) return SDOperand(BV, 0); 2107 2108 unsigned SrcBitSize = MVT::getSizeInBits(SrcEltVT); 2109 unsigned DstBitSize = MVT::getSizeInBits(DstEltVT); 2110 2111 // If this is a conversion of N elements of one type to N elements of another 2112 // type, convert each element. This handles FP<->INT cases. 2113 if (SrcBitSize == DstBitSize) { 2114 std::vector<SDOperand> Ops; 2115 for (unsigned i = 0, e = BV->getNumOperands()-2; i != e; ++i) { 2116 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, DstEltVT, BV->getOperand(i))); 2117 AddToWorkList(Ops.back().Val); 2118 } 2119 Ops.push_back(*(BV->op_end()-2)); // Add num elements. 2120 Ops.push_back(DAG.getValueType(DstEltVT)); 2121 return DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, Ops); 2122 } 2123 2124 // Otherwise, we're growing or shrinking the elements. To avoid having to 2125 // handle annoying details of growing/shrinking FP values, we convert them to 2126 // int first. 2127 if (MVT::isFloatingPoint(SrcEltVT)) { 2128 // Convert the input float vector to a int vector where the elements are the 2129 // same sizes. 2130 assert((SrcEltVT == MVT::f32 || SrcEltVT == MVT::f64) && "Unknown FP VT!"); 2131 MVT::ValueType IntVT = SrcEltVT == MVT::f32 ? MVT::i32 : MVT::i64; 2132 BV = ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(BV, IntVT).Val; 2133 SrcEltVT = IntVT; 2134 } 2135 2136 // Now we know the input is an integer vector. If the output is a FP type, 2137 // convert to integer first, then to FP of the right size. 2138 if (MVT::isFloatingPoint(DstEltVT)) { 2139 assert((DstEltVT == MVT::f32 || DstEltVT == MVT::f64) && "Unknown FP VT!"); 2140 MVT::ValueType TmpVT = DstEltVT == MVT::f32 ? MVT::i32 : MVT::i64; 2141 SDNode *Tmp = ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(BV, TmpVT).Val; 2142 2143 // Next, convert to FP elements of the same size. 2144 return ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(Tmp, DstEltVT); 2145 } 2146 2147 // Okay, we know the src/dst types are both integers of differing types. 2148 // Handling growing first. 2149 assert(MVT::isInteger(SrcEltVT) && MVT::isInteger(DstEltVT)); 2150 if (SrcBitSize < DstBitSize) { 2151 unsigned NumInputsPerOutput = DstBitSize/SrcBitSize; 2152 2153 std::vector<SDOperand> Ops; 2154 for (unsigned i = 0, e = BV->getNumOperands()-2; i != e; 2155 i += NumInputsPerOutput) { 2156 bool isLE = TLI.isLittleEndian(); 2157 uint64_t NewBits = 0; 2158 bool EltIsUndef = true; 2159 for (unsigned j = 0; j != NumInputsPerOutput; ++j) { 2160 // Shift the previously computed bits over. 2161 NewBits <<= SrcBitSize; 2162 SDOperand Op = BV->getOperand(i+ (isLE ? (NumInputsPerOutput-j-1) : j)); 2163 if (Op.getOpcode() == ISD::UNDEF) continue; 2164 EltIsUndef = false; 2165 2166 NewBits |= cast<ConstantSDNode>(Op)->getValue(); 2167 } 2168 2169 if (EltIsUndef) 2170 Ops.push_back(DAG.getNode(ISD::UNDEF, DstEltVT)); 2171 else 2172 Ops.push_back(DAG.getConstant(NewBits, DstEltVT)); 2173 } 2174 2175 Ops.push_back(DAG.getConstant(Ops.size(), MVT::i32)); // Add num elements. 2176 Ops.push_back(DAG.getValueType(DstEltVT)); // Add element size. 2177 return DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, Ops); 2178 } 2179 2180 // Finally, this must be the case where we are shrinking elements: each input 2181 // turns into multiple outputs. 2182 unsigned NumOutputsPerInput = SrcBitSize/DstBitSize; 2183 std::vector<SDOperand> Ops; 2184 for (unsigned i = 0, e = BV->getNumOperands()-2; i != e; ++i) { 2185 if (BV->getOperand(i).getOpcode() == ISD::UNDEF) { 2186 for (unsigned j = 0; j != NumOutputsPerInput; ++j) 2187 Ops.push_back(DAG.getNode(ISD::UNDEF, DstEltVT)); 2188 continue; 2189 } 2190 uint64_t OpVal = cast<ConstantSDNode>(BV->getOperand(i))->getValue(); 2191 2192 for (unsigned j = 0; j != NumOutputsPerInput; ++j) { 2193 unsigned ThisVal = OpVal & ((1ULL << DstBitSize)-1); 2194 OpVal >>= DstBitSize; 2195 Ops.push_back(DAG.getConstant(ThisVal, DstEltVT)); 2196 } 2197 2198 // For big endian targets, swap the order of the pieces of each element. 2199 if (!TLI.isLittleEndian()) 2200 std::reverse(Ops.end()-NumOutputsPerInput, Ops.end()); 2201 } 2202 Ops.push_back(DAG.getConstant(Ops.size(), MVT::i32)); // Add num elements. 2203 Ops.push_back(DAG.getValueType(DstEltVT)); // Add element size. 2204 return DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, Ops); 2205} 2206 2207 2208 2209SDOperand DAGCombiner::visitFADD(SDNode *N) { 2210 SDOperand N0 = N->getOperand(0); 2211 SDOperand N1 = N->getOperand(1); 2212 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 2213 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1); 2214 MVT::ValueType VT = N->getValueType(0); 2215 2216 // fold (fadd c1, c2) -> c1+c2 2217 if (N0CFP && N1CFP) 2218 return DAG.getNode(ISD::FADD, VT, N0, N1); 2219 // canonicalize constant to RHS 2220 if (N0CFP && !N1CFP) 2221 return DAG.getNode(ISD::FADD, VT, N1, N0); 2222 // fold (A + (-B)) -> A-B 2223 if (N1.getOpcode() == ISD::FNEG) 2224 return DAG.getNode(ISD::FSUB, VT, N0, N1.getOperand(0)); 2225 // fold ((-A) + B) -> B-A 2226 if (N0.getOpcode() == ISD::FNEG) 2227 return DAG.getNode(ISD::FSUB, VT, N1, N0.getOperand(0)); 2228 return SDOperand(); 2229} 2230 2231SDOperand DAGCombiner::visitFSUB(SDNode *N) { 2232 SDOperand N0 = N->getOperand(0); 2233 SDOperand N1 = N->getOperand(1); 2234 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 2235 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1); 2236 MVT::ValueType VT = N->getValueType(0); 2237 2238 // fold (fsub c1, c2) -> c1-c2 2239 if (N0CFP && N1CFP) 2240 return DAG.getNode(ISD::FSUB, VT, N0, N1); 2241 // fold (A-(-B)) -> A+B 2242 if (N1.getOpcode() == ISD::FNEG) 2243 return DAG.getNode(ISD::FADD, VT, N0, N1.getOperand(0)); 2244 return SDOperand(); 2245} 2246 2247SDOperand DAGCombiner::visitFMUL(SDNode *N) { 2248 SDOperand N0 = N->getOperand(0); 2249 SDOperand N1 = N->getOperand(1); 2250 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 2251 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1); 2252 MVT::ValueType VT = N->getValueType(0); 2253 2254 // fold (fmul c1, c2) -> c1*c2 2255 if (N0CFP && N1CFP) 2256 return DAG.getNode(ISD::FMUL, VT, N0, N1); 2257 // canonicalize constant to RHS 2258 if (N0CFP && !N1CFP) 2259 return DAG.getNode(ISD::FMUL, VT, N1, N0); 2260 // fold (fmul X, 2.0) -> (fadd X, X) 2261 if (N1CFP && N1CFP->isExactlyValue(+2.0)) 2262 return DAG.getNode(ISD::FADD, VT, N0, N0); 2263 return SDOperand(); 2264} 2265 2266SDOperand DAGCombiner::visitFDIV(SDNode *N) { 2267 SDOperand N0 = N->getOperand(0); 2268 SDOperand N1 = N->getOperand(1); 2269 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 2270 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1); 2271 MVT::ValueType VT = N->getValueType(0); 2272 2273 // fold (fdiv c1, c2) -> c1/c2 2274 if (N0CFP && N1CFP) 2275 return DAG.getNode(ISD::FDIV, VT, N0, N1); 2276 return SDOperand(); 2277} 2278 2279SDOperand DAGCombiner::visitFREM(SDNode *N) { 2280 SDOperand N0 = N->getOperand(0); 2281 SDOperand N1 = N->getOperand(1); 2282 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 2283 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1); 2284 MVT::ValueType VT = N->getValueType(0); 2285 2286 // fold (frem c1, c2) -> fmod(c1,c2) 2287 if (N0CFP && N1CFP) 2288 return DAG.getNode(ISD::FREM, VT, N0, N1); 2289 return SDOperand(); 2290} 2291 2292SDOperand DAGCombiner::visitFCOPYSIGN(SDNode *N) { 2293 SDOperand N0 = N->getOperand(0); 2294 SDOperand N1 = N->getOperand(1); 2295 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 2296 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1); 2297 MVT::ValueType VT = N->getValueType(0); 2298 2299 if (N0CFP && N1CFP) // Constant fold 2300 return DAG.getNode(ISD::FCOPYSIGN, VT, N0, N1); 2301 2302 if (N1CFP) { 2303 // copysign(x, c1) -> fabs(x) iff ispos(c1) 2304 // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1) 2305 union { 2306 double d; 2307 int64_t i; 2308 } u; 2309 u.d = N1CFP->getValue(); 2310 if (u.i >= 0) 2311 return DAG.getNode(ISD::FABS, VT, N0); 2312 else 2313 return DAG.getNode(ISD::FNEG, VT, DAG.getNode(ISD::FABS, VT, N0)); 2314 } 2315 2316 // copysign(fabs(x), y) -> copysign(x, y) 2317 // copysign(fneg(x), y) -> copysign(x, y) 2318 // copysign(copysign(x,z), y) -> copysign(x, y) 2319 if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG || 2320 N0.getOpcode() == ISD::FCOPYSIGN) 2321 return DAG.getNode(ISD::FCOPYSIGN, VT, N0.getOperand(0), N1); 2322 2323 // copysign(x, abs(y)) -> abs(x) 2324 if (N1.getOpcode() == ISD::FABS) 2325 return DAG.getNode(ISD::FABS, VT, N0); 2326 2327 // copysign(x, copysign(y,z)) -> copysign(x, z) 2328 if (N1.getOpcode() == ISD::FCOPYSIGN) 2329 return DAG.getNode(ISD::FCOPYSIGN, VT, N0, N1.getOperand(1)); 2330 2331 // copysign(x, fp_extend(y)) -> copysign(x, y) 2332 // copysign(x, fp_round(y)) -> copysign(x, y) 2333 if (N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND) 2334 return DAG.getNode(ISD::FCOPYSIGN, VT, N0, N1.getOperand(0)); 2335 2336 return SDOperand(); 2337} 2338 2339 2340 2341SDOperand DAGCombiner::visitSINT_TO_FP(SDNode *N) { 2342 SDOperand N0 = N->getOperand(0); 2343 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 2344 MVT::ValueType VT = N->getValueType(0); 2345 2346 // fold (sint_to_fp c1) -> c1fp 2347 if (N0C) 2348 return DAG.getNode(ISD::SINT_TO_FP, VT, N0); 2349 return SDOperand(); 2350} 2351 2352SDOperand DAGCombiner::visitUINT_TO_FP(SDNode *N) { 2353 SDOperand N0 = N->getOperand(0); 2354 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 2355 MVT::ValueType VT = N->getValueType(0); 2356 2357 // fold (uint_to_fp c1) -> c1fp 2358 if (N0C) 2359 return DAG.getNode(ISD::UINT_TO_FP, VT, N0); 2360 return SDOperand(); 2361} 2362 2363SDOperand DAGCombiner::visitFP_TO_SINT(SDNode *N) { 2364 SDOperand N0 = N->getOperand(0); 2365 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 2366 MVT::ValueType VT = N->getValueType(0); 2367 2368 // fold (fp_to_sint c1fp) -> c1 2369 if (N0CFP) 2370 return DAG.getNode(ISD::FP_TO_SINT, VT, N0); 2371 return SDOperand(); 2372} 2373 2374SDOperand DAGCombiner::visitFP_TO_UINT(SDNode *N) { 2375 SDOperand N0 = N->getOperand(0); 2376 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 2377 MVT::ValueType VT = N->getValueType(0); 2378 2379 // fold (fp_to_uint c1fp) -> c1 2380 if (N0CFP) 2381 return DAG.getNode(ISD::FP_TO_UINT, VT, N0); 2382 return SDOperand(); 2383} 2384 2385SDOperand DAGCombiner::visitFP_ROUND(SDNode *N) { 2386 SDOperand N0 = N->getOperand(0); 2387 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 2388 MVT::ValueType VT = N->getValueType(0); 2389 2390 // fold (fp_round c1fp) -> c1fp 2391 if (N0CFP) 2392 return DAG.getNode(ISD::FP_ROUND, VT, N0); 2393 2394 // fold (fp_round (fp_extend x)) -> x 2395 if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType()) 2396 return N0.getOperand(0); 2397 2398 // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y) 2399 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.Val->hasOneUse()) { 2400 SDOperand Tmp = DAG.getNode(ISD::FP_ROUND, VT, N0.getOperand(0)); 2401 AddToWorkList(Tmp.Val); 2402 return DAG.getNode(ISD::FCOPYSIGN, VT, Tmp, N0.getOperand(1)); 2403 } 2404 2405 return SDOperand(); 2406} 2407 2408SDOperand DAGCombiner::visitFP_ROUND_INREG(SDNode *N) { 2409 SDOperand N0 = N->getOperand(0); 2410 MVT::ValueType VT = N->getValueType(0); 2411 MVT::ValueType EVT = cast<VTSDNode>(N->getOperand(1))->getVT(); 2412 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 2413 2414 // fold (fp_round_inreg c1fp) -> c1fp 2415 if (N0CFP) { 2416 SDOperand Round = DAG.getConstantFP(N0CFP->getValue(), EVT); 2417 return DAG.getNode(ISD::FP_EXTEND, VT, Round); 2418 } 2419 return SDOperand(); 2420} 2421 2422SDOperand DAGCombiner::visitFP_EXTEND(SDNode *N) { 2423 SDOperand N0 = N->getOperand(0); 2424 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 2425 MVT::ValueType VT = N->getValueType(0); 2426 2427 // fold (fp_extend c1fp) -> c1fp 2428 if (N0CFP) 2429 return DAG.getNode(ISD::FP_EXTEND, VT, N0); 2430 2431 // fold (fpext (load x)) -> (fpext (fpround (extload x))) 2432 if (N0.getOpcode() == ISD::LOAD && N0.hasOneUse() && 2433 (!AfterLegalize||TLI.isOperationLegal(ISD::EXTLOAD, N0.getValueType()))) { 2434 SDOperand ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, VT, N0.getOperand(0), 2435 N0.getOperand(1), N0.getOperand(2), 2436 N0.getValueType()); 2437 CombineTo(N, ExtLoad); 2438 CombineTo(N0.Val, DAG.getNode(ISD::FP_ROUND, N0.getValueType(), ExtLoad), 2439 ExtLoad.getValue(1)); 2440 return SDOperand(N, 0); // Return N so it doesn't get rechecked! 2441 } 2442 2443 2444 return SDOperand(); 2445} 2446 2447SDOperand DAGCombiner::visitFNEG(SDNode *N) { 2448 SDOperand N0 = N->getOperand(0); 2449 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 2450 MVT::ValueType VT = N->getValueType(0); 2451 2452 // fold (fneg c1) -> -c1 2453 if (N0CFP) 2454 return DAG.getNode(ISD::FNEG, VT, N0); 2455 // fold (fneg (sub x, y)) -> (sub y, x) 2456 if (N0.getOpcode() == ISD::SUB) 2457 return DAG.getNode(ISD::SUB, VT, N0.getOperand(1), N0.getOperand(0)); 2458 // fold (fneg (fneg x)) -> x 2459 if (N0.getOpcode() == ISD::FNEG) 2460 return N0.getOperand(0); 2461 return SDOperand(); 2462} 2463 2464SDOperand DAGCombiner::visitFABS(SDNode *N) { 2465 SDOperand N0 = N->getOperand(0); 2466 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 2467 MVT::ValueType VT = N->getValueType(0); 2468 2469 // fold (fabs c1) -> fabs(c1) 2470 if (N0CFP) 2471 return DAG.getNode(ISD::FABS, VT, N0); 2472 // fold (fabs (fabs x)) -> (fabs x) 2473 if (N0.getOpcode() == ISD::FABS) 2474 return N->getOperand(0); 2475 // fold (fabs (fneg x)) -> (fabs x) 2476 // fold (fabs (fcopysign x, y)) -> (fabs x) 2477 if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN) 2478 return DAG.getNode(ISD::FABS, VT, N0.getOperand(0)); 2479 2480 return SDOperand(); 2481} 2482 2483SDOperand DAGCombiner::visitBRCOND(SDNode *N) { 2484 SDOperand Chain = N->getOperand(0); 2485 SDOperand N1 = N->getOperand(1); 2486 SDOperand N2 = N->getOperand(2); 2487 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 2488 2489 // never taken branch, fold to chain 2490 if (N1C && N1C->isNullValue()) 2491 return Chain; 2492 // unconditional branch 2493 if (N1C && N1C->getValue() == 1) 2494 return DAG.getNode(ISD::BR, MVT::Other, Chain, N2); 2495 // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal 2496 // on the target. 2497 if (N1.getOpcode() == ISD::SETCC && 2498 TLI.isOperationLegal(ISD::BR_CC, MVT::Other)) { 2499 return DAG.getNode(ISD::BR_CC, MVT::Other, Chain, N1.getOperand(2), 2500 N1.getOperand(0), N1.getOperand(1), N2); 2501 } 2502 return SDOperand(); 2503} 2504 2505// Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB. 2506// 2507SDOperand DAGCombiner::visitBR_CC(SDNode *N) { 2508 CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1)); 2509 SDOperand CondLHS = N->getOperand(2), CondRHS = N->getOperand(3); 2510 2511 // Use SimplifySetCC to simplify SETCC's. 2512 SDOperand Simp = SimplifySetCC(MVT::i1, CondLHS, CondRHS, CC->get(), false); 2513 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(Simp.Val); 2514 2515 // fold br_cc true, dest -> br dest (unconditional branch) 2516 if (SCCC && SCCC->getValue()) 2517 return DAG.getNode(ISD::BR, MVT::Other, N->getOperand(0), 2518 N->getOperand(4)); 2519 // fold br_cc false, dest -> unconditional fall through 2520 if (SCCC && SCCC->isNullValue()) 2521 return N->getOperand(0); 2522 // fold to a simpler setcc 2523 if (Simp.Val && Simp.getOpcode() == ISD::SETCC) 2524 return DAG.getNode(ISD::BR_CC, MVT::Other, N->getOperand(0), 2525 Simp.getOperand(2), Simp.getOperand(0), 2526 Simp.getOperand(1), N->getOperand(4)); 2527 return SDOperand(); 2528} 2529 2530SDOperand DAGCombiner::visitLOAD(SDNode *N) { 2531 SDOperand Chain = N->getOperand(0); 2532 SDOperand Ptr = N->getOperand(1); 2533 SDOperand SrcValue = N->getOperand(2); 2534 2535 // If there are no uses of the loaded value, change uses of the chain value 2536 // into uses of the chain input (i.e. delete the dead load). 2537 if (N->hasNUsesOfValue(0, 0)) 2538 return CombineTo(N, DAG.getNode(ISD::UNDEF, N->getValueType(0)), Chain); 2539 2540 // If this load is directly stored, replace the load value with the stored 2541 // value. 2542 // TODO: Handle store large -> read small portion. 2543 // TODO: Handle TRUNCSTORE/EXTLOAD 2544 if (Chain.getOpcode() == ISD::STORE && Chain.getOperand(2) == Ptr && 2545 Chain.getOperand(1).getValueType() == N->getValueType(0)) 2546 return CombineTo(N, Chain.getOperand(1), Chain); 2547 2548 return SDOperand(); 2549} 2550 2551/// visitXEXTLOAD - Handle EXTLOAD/ZEXTLOAD/SEXTLOAD. 2552SDOperand DAGCombiner::visitXEXTLOAD(SDNode *N) { 2553 SDOperand Chain = N->getOperand(0); 2554 SDOperand Ptr = N->getOperand(1); 2555 SDOperand SrcValue = N->getOperand(2); 2556 SDOperand EVT = N->getOperand(3); 2557 2558 // If there are no uses of the loaded value, change uses of the chain value 2559 // into uses of the chain input (i.e. delete the dead load). 2560 if (N->hasNUsesOfValue(0, 0)) 2561 return CombineTo(N, DAG.getNode(ISD::UNDEF, N->getValueType(0)), Chain); 2562 2563 return SDOperand(); 2564} 2565 2566SDOperand DAGCombiner::visitSTORE(SDNode *N) { 2567 SDOperand Chain = N->getOperand(0); 2568 SDOperand Value = N->getOperand(1); 2569 SDOperand Ptr = N->getOperand(2); 2570 SDOperand SrcValue = N->getOperand(3); 2571 2572 // If this is a store that kills a previous store, remove the previous store. 2573 if (Chain.getOpcode() == ISD::STORE && Chain.getOperand(2) == Ptr && 2574 Chain.Val->hasOneUse() /* Avoid introducing DAG cycles */ && 2575 // Make sure that these stores are the same value type: 2576 // FIXME: we really care that the second store is >= size of the first. 2577 Value.getValueType() == Chain.getOperand(1).getValueType()) { 2578 // Create a new store of Value that replaces both stores. 2579 SDNode *PrevStore = Chain.Val; 2580 if (PrevStore->getOperand(1) == Value) // Same value multiply stored. 2581 return Chain; 2582 SDOperand NewStore = DAG.getNode(ISD::STORE, MVT::Other, 2583 PrevStore->getOperand(0), Value, Ptr, 2584 SrcValue); 2585 CombineTo(N, NewStore); // Nuke this store. 2586 CombineTo(PrevStore, NewStore); // Nuke the previous store. 2587 return SDOperand(N, 0); 2588 } 2589 2590 // If this is a store of a bit convert, store the input value. 2591 // FIXME: This needs to know that the resultant store does not need a 2592 // higher alignment than the original. 2593 if (0 && Value.getOpcode() == ISD::BIT_CONVERT) 2594 return DAG.getNode(ISD::STORE, MVT::Other, Chain, Value.getOperand(0), 2595 Ptr, SrcValue); 2596 2597 return SDOperand(); 2598} 2599 2600SDOperand DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) { 2601 SDOperand InVec = N->getOperand(0); 2602 SDOperand InVal = N->getOperand(1); 2603 SDOperand EltNo = N->getOperand(2); 2604 2605 // If the invec is a BUILD_VECTOR and if EltNo is a constant, build a new 2606 // vector with the inserted element. 2607 if (InVec.getOpcode() == ISD::BUILD_VECTOR && isa<ConstantSDNode>(EltNo)) { 2608 unsigned Elt = cast<ConstantSDNode>(EltNo)->getValue(); 2609 std::vector<SDOperand> Ops(InVec.Val->op_begin(), InVec.Val->op_end()); 2610 if (Elt < Ops.size()) 2611 Ops[Elt] = InVal; 2612 return DAG.getNode(ISD::BUILD_VECTOR, InVec.getValueType(), Ops); 2613 } 2614 2615 return SDOperand(); 2616} 2617 2618SDOperand DAGCombiner::visitVINSERT_VECTOR_ELT(SDNode *N) { 2619 SDOperand InVec = N->getOperand(0); 2620 SDOperand InVal = N->getOperand(1); 2621 SDOperand EltNo = N->getOperand(2); 2622 SDOperand NumElts = N->getOperand(3); 2623 SDOperand EltType = N->getOperand(4); 2624 2625 // If the invec is a VBUILD_VECTOR and if EltNo is a constant, build a new 2626 // vector with the inserted element. 2627 if (InVec.getOpcode() == ISD::VBUILD_VECTOR && isa<ConstantSDNode>(EltNo)) { 2628 unsigned Elt = cast<ConstantSDNode>(EltNo)->getValue(); 2629 std::vector<SDOperand> Ops(InVec.Val->op_begin(), InVec.Val->op_end()); 2630 if (Elt < Ops.size()-2) 2631 Ops[Elt] = InVal; 2632 return DAG.getNode(ISD::VBUILD_VECTOR, InVec.getValueType(), Ops); 2633 } 2634 2635 return SDOperand(); 2636} 2637 2638SDOperand DAGCombiner::visitVBUILD_VECTOR(SDNode *N) { 2639 unsigned NumInScalars = N->getNumOperands()-2; 2640 SDOperand NumElts = N->getOperand(NumInScalars); 2641 SDOperand EltType = N->getOperand(NumInScalars+1); 2642 2643 // Check to see if this is a VBUILD_VECTOR of a bunch of VEXTRACT_VECTOR_ELT 2644 // operations. If so, and if the EXTRACT_ELT vector inputs come from at most 2645 // two distinct vectors, turn this into a shuffle node. 2646 SDOperand VecIn1, VecIn2; 2647 for (unsigned i = 0; i != NumInScalars; ++i) { 2648 // Ignore undef inputs. 2649 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue; 2650 2651 // If this input is something other than a VEXTRACT_VECTOR_ELT with a 2652 // constant index, bail out. 2653 if (N->getOperand(i).getOpcode() != ISD::VEXTRACT_VECTOR_ELT || 2654 !isa<ConstantSDNode>(N->getOperand(i).getOperand(1))) { 2655 VecIn1 = VecIn2 = SDOperand(0, 0); 2656 break; 2657 } 2658 2659 // If the input vector type disagrees with the result of the vbuild_vector, 2660 // we can't make a shuffle. 2661 SDOperand ExtractedFromVec = N->getOperand(i).getOperand(0); 2662 if (*(ExtractedFromVec.Val->op_end()-2) != NumElts || 2663 *(ExtractedFromVec.Val->op_end()-1) != EltType) { 2664 VecIn1 = VecIn2 = SDOperand(0, 0); 2665 break; 2666 } 2667 2668 // Otherwise, remember this. We allow up to two distinct input vectors. 2669 if (ExtractedFromVec == VecIn1 || ExtractedFromVec == VecIn2) 2670 continue; 2671 2672 if (VecIn1.Val == 0) { 2673 VecIn1 = ExtractedFromVec; 2674 } else if (VecIn2.Val == 0) { 2675 VecIn2 = ExtractedFromVec; 2676 } else { 2677 // Too many inputs. 2678 VecIn1 = VecIn2 = SDOperand(0, 0); 2679 break; 2680 } 2681 } 2682 2683 // If everything is good, we can make a shuffle operation. 2684 if (VecIn1.Val) { 2685 std::vector<SDOperand> BuildVecIndices; 2686 for (unsigned i = 0; i != NumInScalars; ++i) { 2687 if (N->getOperand(i).getOpcode() == ISD::UNDEF) { 2688 BuildVecIndices.push_back(DAG.getNode(ISD::UNDEF, MVT::i32)); 2689 continue; 2690 } 2691 2692 SDOperand Extract = N->getOperand(i); 2693 2694 // If extracting from the first vector, just use the index directly. 2695 if (Extract.getOperand(0) == VecIn1) { 2696 BuildVecIndices.push_back(Extract.getOperand(1)); 2697 continue; 2698 } 2699 2700 // Otherwise, use InIdx + VecSize 2701 unsigned Idx = cast<ConstantSDNode>(Extract.getOperand(1))->getValue(); 2702 BuildVecIndices.push_back(DAG.getConstant(Idx+NumInScalars, MVT::i32)); 2703 } 2704 2705 // Add count and size info. 2706 BuildVecIndices.push_back(NumElts); 2707 BuildVecIndices.push_back(DAG.getValueType(MVT::i32)); 2708 2709 // Return the new VVECTOR_SHUFFLE node. 2710 std::vector<SDOperand> Ops; 2711 Ops.push_back(VecIn1); 2712 if (VecIn2.Val) { 2713 Ops.push_back(VecIn2); 2714 } else { 2715 // Use an undef vbuild_vector as input for the second operand. 2716 std::vector<SDOperand> UnOps(NumInScalars, 2717 DAG.getNode(ISD::UNDEF, 2718 cast<VTSDNode>(EltType)->getVT())); 2719 UnOps.push_back(NumElts); 2720 UnOps.push_back(EltType); 2721 Ops.push_back(DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, UnOps)); 2722 AddToWorkList(Ops.back().Val); 2723 } 2724 Ops.push_back(DAG.getNode(ISD::VBUILD_VECTOR,MVT::Vector, BuildVecIndices)); 2725 Ops.push_back(NumElts); 2726 Ops.push_back(EltType); 2727 return DAG.getNode(ISD::VVECTOR_SHUFFLE, MVT::Vector, Ops); 2728 } 2729 2730 return SDOperand(); 2731} 2732 2733SDOperand DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) { 2734 SDOperand ShufMask = N->getOperand(2); 2735 unsigned NumElts = ShufMask.getNumOperands(); 2736 2737 // If the shuffle mask is an identity operation on the LHS, return the LHS. 2738 bool isIdentity = true; 2739 for (unsigned i = 0; i != NumElts; ++i) { 2740 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF && 2741 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() != i) { 2742 isIdentity = false; 2743 break; 2744 } 2745 } 2746 if (isIdentity) return N->getOperand(0); 2747 2748 // If the shuffle mask is an identity operation on the RHS, return the RHS. 2749 isIdentity = true; 2750 for (unsigned i = 0; i != NumElts; ++i) { 2751 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF && 2752 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() != i+NumElts) { 2753 isIdentity = false; 2754 break; 2755 } 2756 } 2757 if (isIdentity) return N->getOperand(1); 2758 2759 // If the LHS and the RHS are the same node, turn the RHS into an undef. 2760 if (N->getOperand(0) == N->getOperand(1)) { 2761 if (N->getOperand(0).getOpcode() == ISD::UNDEF) 2762 return DAG.getNode(ISD::UNDEF, N->getValueType(0)); 2763 // Check the SHUFFLE mask, mapping any inputs from the 2nd operand into the 2764 // first operand. 2765 std::vector<SDOperand> MappedOps; 2766 for (unsigned i = 0, e = ShufMask.getNumOperands(); i != e; ++i) { 2767 if (ShufMask.getOperand(i).getOpcode() == ISD::UNDEF || 2768 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() < NumElts) { 2769 MappedOps.push_back(ShufMask.getOperand(i)); 2770 } else { 2771 unsigned NewIdx = 2772 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() - NumElts; 2773 MappedOps.push_back(DAG.getConstant(NewIdx, MVT::i32)); 2774 } 2775 } 2776 ShufMask = DAG.getNode(ISD::BUILD_VECTOR, ShufMask.getValueType(), 2777 MappedOps); 2778 AddToWorkList(ShufMask.Val); 2779 return DAG.getNode(ISD::VECTOR_SHUFFLE, N->getValueType(0), 2780 N->getOperand(0), 2781 DAG.getNode(ISD::UNDEF, N->getValueType(0)), 2782 ShufMask); 2783 } 2784 2785 return SDOperand(); 2786} 2787 2788SDOperand DAGCombiner::visitVVECTOR_SHUFFLE(SDNode *N) { 2789 SDOperand ShufMask = N->getOperand(2); 2790 unsigned NumElts = ShufMask.getNumOperands()-2; 2791 2792 // If the shuffle mask is an identity operation on the LHS, return the LHS. 2793 bool isIdentity = true; 2794 for (unsigned i = 0; i != NumElts; ++i) { 2795 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF && 2796 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() != i) { 2797 isIdentity = false; 2798 break; 2799 } 2800 } 2801 if (isIdentity) return N->getOperand(0); 2802 2803 // If the shuffle mask is an identity operation on the RHS, return the RHS. 2804 isIdentity = true; 2805 for (unsigned i = 0; i != NumElts; ++i) { 2806 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF && 2807 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() != i+NumElts) { 2808 isIdentity = false; 2809 break; 2810 } 2811 } 2812 if (isIdentity) return N->getOperand(1); 2813 2814 // If the LHS and the RHS are the same node, turn the RHS into an undef. 2815 if (N->getOperand(0) == N->getOperand(1)) { 2816 // Check the SHUFFLE mask, mapping any inputs from the 2nd operand into the 2817 // first operand. 2818 std::vector<SDOperand> MappedOps; 2819 for (unsigned i = 0; i != NumElts; ++i) { 2820 if (ShufMask.getOperand(i).getOpcode() == ISD::UNDEF || 2821 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() < NumElts) { 2822 MappedOps.push_back(ShufMask.getOperand(i)); 2823 } else { 2824 unsigned NewIdx = 2825 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() - NumElts; 2826 MappedOps.push_back(DAG.getConstant(NewIdx, MVT::i32)); 2827 } 2828 } 2829 // Add the type/#elts values. 2830 MappedOps.push_back(ShufMask.getOperand(NumElts)); 2831 MappedOps.push_back(ShufMask.getOperand(NumElts+1)); 2832 2833 ShufMask = DAG.getNode(ISD::VBUILD_VECTOR, ShufMask.getValueType(), 2834 MappedOps); 2835 AddToWorkList(ShufMask.Val); 2836 2837 // Build the undef vector. 2838 SDOperand UDVal = DAG.getNode(ISD::UNDEF, MappedOps[0].getValueType()); 2839 for (unsigned i = 0; i != NumElts; ++i) 2840 MappedOps[i] = UDVal; 2841 MappedOps[NumElts ] = *(N->getOperand(0).Val->op_end()-2); 2842 MappedOps[NumElts+1] = *(N->getOperand(0).Val->op_end()-1); 2843 UDVal = DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, MappedOps); 2844 2845 return DAG.getNode(ISD::VVECTOR_SHUFFLE, MVT::Vector, 2846 N->getOperand(0), UDVal, ShufMask, 2847 MappedOps[NumElts], MappedOps[NumElts+1]); 2848 } 2849 2850 return SDOperand(); 2851} 2852 2853/// XformToShuffleWithZero - Returns a vector_shuffle if it able to transform 2854/// a VAND to a vector_shuffle with the destination vector and a zero vector. 2855/// e.g. VAND V, <0xffffffff, 0, 0xffffffff, 0>. ==> 2856/// vector_shuffle V, Zero, <0, 4, 2, 4> 2857SDOperand DAGCombiner::XformToShuffleWithZero(SDNode *N) { 2858 SDOperand LHS = N->getOperand(0); 2859 SDOperand RHS = N->getOperand(1); 2860 if (N->getOpcode() == ISD::VAND) { 2861 SDOperand DstVecSize = *(LHS.Val->op_end()-2); 2862 SDOperand DstVecEVT = *(LHS.Val->op_end()-1); 2863 if (RHS.getOpcode() == ISD::VBIT_CONVERT) 2864 RHS = RHS.getOperand(0); 2865 if (RHS.getOpcode() == ISD::VBUILD_VECTOR) { 2866 std::vector<SDOperand> IdxOps; 2867 unsigned NumOps = RHS.getNumOperands(); 2868 unsigned NumElts = NumOps-2; 2869 MVT::ValueType EVT = cast<VTSDNode>(RHS.getOperand(NumOps-1))->getVT(); 2870 for (unsigned i = 0; i != NumElts; ++i) { 2871 SDOperand Elt = RHS.getOperand(i); 2872 if (!isa<ConstantSDNode>(Elt)) 2873 return SDOperand(); 2874 else if (cast<ConstantSDNode>(Elt)->isAllOnesValue()) 2875 IdxOps.push_back(DAG.getConstant(i, EVT)); 2876 else if (cast<ConstantSDNode>(Elt)->isNullValue()) 2877 IdxOps.push_back(DAG.getConstant(NumElts, EVT)); 2878 else 2879 return SDOperand(); 2880 } 2881 2882 // Let's see if the target supports this vector_shuffle. 2883 if (!TLI.isVectorClearMaskLegal(IdxOps, EVT, DAG)) 2884 return SDOperand(); 2885 2886 // Return the new VVECTOR_SHUFFLE node. 2887 SDOperand NumEltsNode = DAG.getConstant(NumElts, MVT::i32); 2888 SDOperand EVTNode = DAG.getValueType(EVT); 2889 std::vector<SDOperand> Ops; 2890 LHS = DAG.getNode(ISD::VBIT_CONVERT, MVT::Vector, LHS, NumEltsNode, EVTNode); 2891 Ops.push_back(LHS); 2892 AddToWorkList(LHS.Val); 2893 std::vector<SDOperand> ZeroOps(NumElts, DAG.getConstant(0, EVT)); 2894 ZeroOps.push_back(NumEltsNode); 2895 ZeroOps.push_back(EVTNode); 2896 Ops.push_back(DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, ZeroOps)); 2897 IdxOps.push_back(NumEltsNode); 2898 IdxOps.push_back(EVTNode); 2899 Ops.push_back(DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, IdxOps)); 2900 Ops.push_back(NumEltsNode); 2901 Ops.push_back(EVTNode); 2902 SDOperand Result = DAG.getNode(ISD::VVECTOR_SHUFFLE, MVT::Vector, Ops); 2903 if (NumEltsNode != DstVecSize || EVTNode != DstVecEVT) { 2904 Result = DAG.getNode(ISD::VBIT_CONVERT, MVT::Vector, Result, 2905 DstVecSize, DstVecEVT); 2906 } 2907 return Result; 2908 } 2909 } 2910 return SDOperand(); 2911} 2912 2913/// visitVBinOp - Visit a binary vector operation, like VADD. IntOp indicates 2914/// the scalar operation of the vop if it is operating on an integer vector 2915/// (e.g. ADD) and FPOp indicates the FP version (e.g. FADD). 2916SDOperand DAGCombiner::visitVBinOp(SDNode *N, ISD::NodeType IntOp, 2917 ISD::NodeType FPOp) { 2918 MVT::ValueType EltType = cast<VTSDNode>(*(N->op_end()-1))->getVT(); 2919 ISD::NodeType ScalarOp = MVT::isInteger(EltType) ? IntOp : FPOp; 2920 SDOperand LHS = N->getOperand(0); 2921 SDOperand RHS = N->getOperand(1); 2922 SDOperand Shuffle = XformToShuffleWithZero(N); 2923 if (Shuffle.Val) return Shuffle; 2924 2925 // If the LHS and RHS are VBUILD_VECTOR nodes, see if we can constant fold 2926 // this operation. 2927 if (LHS.getOpcode() == ISD::VBUILD_VECTOR && 2928 RHS.getOpcode() == ISD::VBUILD_VECTOR) { 2929 std::vector<SDOperand> Ops; 2930 for (unsigned i = 0, e = LHS.getNumOperands()-2; i != e; ++i) { 2931 SDOperand LHSOp = LHS.getOperand(i); 2932 SDOperand RHSOp = RHS.getOperand(i); 2933 // If these two elements can't be folded, bail out. 2934 if ((LHSOp.getOpcode() != ISD::UNDEF && 2935 LHSOp.getOpcode() != ISD::Constant && 2936 LHSOp.getOpcode() != ISD::ConstantFP) || 2937 (RHSOp.getOpcode() != ISD::UNDEF && 2938 RHSOp.getOpcode() != ISD::Constant && 2939 RHSOp.getOpcode() != ISD::ConstantFP)) 2940 break; 2941 Ops.push_back(DAG.getNode(ScalarOp, EltType, LHSOp, RHSOp)); 2942 AddToWorkList(Ops.back().Val); 2943 assert((Ops.back().getOpcode() == ISD::UNDEF || 2944 Ops.back().getOpcode() == ISD::Constant || 2945 Ops.back().getOpcode() == ISD::ConstantFP) && 2946 "Scalar binop didn't fold!"); 2947 } 2948 2949 if (Ops.size() == LHS.getNumOperands()-2) { 2950 Ops.push_back(*(LHS.Val->op_end()-2)); 2951 Ops.push_back(*(LHS.Val->op_end()-1)); 2952 return DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, Ops); 2953 } 2954 } 2955 2956 return SDOperand(); 2957} 2958 2959SDOperand DAGCombiner::SimplifySelect(SDOperand N0, SDOperand N1, SDOperand N2){ 2960 assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!"); 2961 2962 SDOperand SCC = SimplifySelectCC(N0.getOperand(0), N0.getOperand(1), N1, N2, 2963 cast<CondCodeSDNode>(N0.getOperand(2))->get()); 2964 // If we got a simplified select_cc node back from SimplifySelectCC, then 2965 // break it down into a new SETCC node, and a new SELECT node, and then return 2966 // the SELECT node, since we were called with a SELECT node. 2967 if (SCC.Val) { 2968 // Check to see if we got a select_cc back (to turn into setcc/select). 2969 // Otherwise, just return whatever node we got back, like fabs. 2970 if (SCC.getOpcode() == ISD::SELECT_CC) { 2971 SDOperand SETCC = DAG.getNode(ISD::SETCC, N0.getValueType(), 2972 SCC.getOperand(0), SCC.getOperand(1), 2973 SCC.getOperand(4)); 2974 AddToWorkList(SETCC.Val); 2975 return DAG.getNode(ISD::SELECT, SCC.getValueType(), SCC.getOperand(2), 2976 SCC.getOperand(3), SETCC); 2977 } 2978 return SCC; 2979 } 2980 return SDOperand(); 2981} 2982 2983/// SimplifySelectOps - Given a SELECT or a SELECT_CC node, where LHS and RHS 2984/// are the two values being selected between, see if we can simplify the 2985/// select. 2986/// 2987bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDOperand LHS, 2988 SDOperand RHS) { 2989 2990 // If this is a select from two identical things, try to pull the operation 2991 // through the select. 2992 if (LHS.getOpcode() == RHS.getOpcode() && LHS.hasOneUse() && RHS.hasOneUse()){ 2993#if 0 2994 std::cerr << "SELECT: ["; LHS.Val->dump(); 2995 std::cerr << "] ["; RHS.Val->dump(); 2996 std::cerr << "]\n"; 2997#endif 2998 2999 // If this is a load and the token chain is identical, replace the select 3000 // of two loads with a load through a select of the address to load from. 3001 // This triggers in things like "select bool X, 10.0, 123.0" after the FP 3002 // constants have been dropped into the constant pool. 3003 if ((LHS.getOpcode() == ISD::LOAD || 3004 LHS.getOpcode() == ISD::EXTLOAD || 3005 LHS.getOpcode() == ISD::ZEXTLOAD || 3006 LHS.getOpcode() == ISD::SEXTLOAD) && 3007 // Token chains must be identical. 3008 LHS.getOperand(0) == RHS.getOperand(0) && 3009 // If this is an EXTLOAD, the VT's must match. 3010 (LHS.getOpcode() == ISD::LOAD || 3011 LHS.getOperand(3) == RHS.getOperand(3))) { 3012 // FIXME: this conflates two src values, discarding one. This is not 3013 // the right thing to do, but nothing uses srcvalues now. When they do, 3014 // turn SrcValue into a list of locations. 3015 SDOperand Addr; 3016 if (TheSelect->getOpcode() == ISD::SELECT) 3017 Addr = DAG.getNode(ISD::SELECT, LHS.getOperand(1).getValueType(), 3018 TheSelect->getOperand(0), LHS.getOperand(1), 3019 RHS.getOperand(1)); 3020 else 3021 Addr = DAG.getNode(ISD::SELECT_CC, LHS.getOperand(1).getValueType(), 3022 TheSelect->getOperand(0), 3023 TheSelect->getOperand(1), 3024 LHS.getOperand(1), RHS.getOperand(1), 3025 TheSelect->getOperand(4)); 3026 3027 SDOperand Load; 3028 if (LHS.getOpcode() == ISD::LOAD) 3029 Load = DAG.getLoad(TheSelect->getValueType(0), LHS.getOperand(0), 3030 Addr, LHS.getOperand(2)); 3031 else 3032 Load = DAG.getExtLoad(LHS.getOpcode(), TheSelect->getValueType(0), 3033 LHS.getOperand(0), Addr, LHS.getOperand(2), 3034 cast<VTSDNode>(LHS.getOperand(3))->getVT()); 3035 // Users of the select now use the result of the load. 3036 CombineTo(TheSelect, Load); 3037 3038 // Users of the old loads now use the new load's chain. We know the 3039 // old-load value is dead now. 3040 CombineTo(LHS.Val, Load.getValue(0), Load.getValue(1)); 3041 CombineTo(RHS.Val, Load.getValue(0), Load.getValue(1)); 3042 return true; 3043 } 3044 } 3045 3046 return false; 3047} 3048 3049SDOperand DAGCombiner::SimplifySelectCC(SDOperand N0, SDOperand N1, 3050 SDOperand N2, SDOperand N3, 3051 ISD::CondCode CC) { 3052 3053 MVT::ValueType VT = N2.getValueType(); 3054 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val); 3055 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val); 3056 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.Val); 3057 ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N3.Val); 3058 3059 // Determine if the condition we're dealing with is constant 3060 SDOperand SCC = SimplifySetCC(TLI.getSetCCResultTy(), N0, N1, CC, false); 3061 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.Val); 3062 3063 // fold select_cc true, x, y -> x 3064 if (SCCC && SCCC->getValue()) 3065 return N2; 3066 // fold select_cc false, x, y -> y 3067 if (SCCC && SCCC->getValue() == 0) 3068 return N3; 3069 3070 // Check to see if we can simplify the select into an fabs node 3071 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) { 3072 // Allow either -0.0 or 0.0 3073 if (CFP->getValue() == 0.0) { 3074 // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs 3075 if ((CC == ISD::SETGE || CC == ISD::SETGT) && 3076 N0 == N2 && N3.getOpcode() == ISD::FNEG && 3077 N2 == N3.getOperand(0)) 3078 return DAG.getNode(ISD::FABS, VT, N0); 3079 3080 // select (setl[te] X, +/-0.0), fneg(X), X -> fabs 3081 if ((CC == ISD::SETLT || CC == ISD::SETLE) && 3082 N0 == N3 && N2.getOpcode() == ISD::FNEG && 3083 N2.getOperand(0) == N3) 3084 return DAG.getNode(ISD::FABS, VT, N3); 3085 } 3086 } 3087 3088 // Check to see if we can perform the "gzip trick", transforming 3089 // select_cc setlt X, 0, A, 0 -> and (sra X, size(X)-1), A 3090 if (N1C && N1C->isNullValue() && N3C && N3C->isNullValue() && 3091 MVT::isInteger(N0.getValueType()) && 3092 MVT::isInteger(N2.getValueType()) && CC == ISD::SETLT) { 3093 MVT::ValueType XType = N0.getValueType(); 3094 MVT::ValueType AType = N2.getValueType(); 3095 if (XType >= AType) { 3096 // and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a 3097 // single-bit constant. 3098 if (N2C && ((N2C->getValue() & (N2C->getValue()-1)) == 0)) { 3099 unsigned ShCtV = Log2_64(N2C->getValue()); 3100 ShCtV = MVT::getSizeInBits(XType)-ShCtV-1; 3101 SDOperand ShCt = DAG.getConstant(ShCtV, TLI.getShiftAmountTy()); 3102 SDOperand Shift = DAG.getNode(ISD::SRL, XType, N0, ShCt); 3103 AddToWorkList(Shift.Val); 3104 if (XType > AType) { 3105 Shift = DAG.getNode(ISD::TRUNCATE, AType, Shift); 3106 AddToWorkList(Shift.Val); 3107 } 3108 return DAG.getNode(ISD::AND, AType, Shift, N2); 3109 } 3110 SDOperand Shift = DAG.getNode(ISD::SRA, XType, N0, 3111 DAG.getConstant(MVT::getSizeInBits(XType)-1, 3112 TLI.getShiftAmountTy())); 3113 AddToWorkList(Shift.Val); 3114 if (XType > AType) { 3115 Shift = DAG.getNode(ISD::TRUNCATE, AType, Shift); 3116 AddToWorkList(Shift.Val); 3117 } 3118 return DAG.getNode(ISD::AND, AType, Shift, N2); 3119 } 3120 } 3121 3122 // fold select C, 16, 0 -> shl C, 4 3123 if (N2C && N3C && N3C->isNullValue() && isPowerOf2_64(N2C->getValue()) && 3124 TLI.getSetCCResultContents() == TargetLowering::ZeroOrOneSetCCResult) { 3125 // Get a SetCC of the condition 3126 // FIXME: Should probably make sure that setcc is legal if we ever have a 3127 // target where it isn't. 3128 SDOperand Temp, SCC; 3129 // cast from setcc result type to select result type 3130 if (AfterLegalize) { 3131 SCC = DAG.getSetCC(TLI.getSetCCResultTy(), N0, N1, CC); 3132 Temp = DAG.getZeroExtendInReg(SCC, N2.getValueType()); 3133 } else { 3134 SCC = DAG.getSetCC(MVT::i1, N0, N1, CC); 3135 Temp = DAG.getNode(ISD::ZERO_EXTEND, N2.getValueType(), SCC); 3136 } 3137 AddToWorkList(SCC.Val); 3138 AddToWorkList(Temp.Val); 3139 // shl setcc result by log2 n2c 3140 return DAG.getNode(ISD::SHL, N2.getValueType(), Temp, 3141 DAG.getConstant(Log2_64(N2C->getValue()), 3142 TLI.getShiftAmountTy())); 3143 } 3144 3145 // Check to see if this is the equivalent of setcc 3146 // FIXME: Turn all of these into setcc if setcc if setcc is legal 3147 // otherwise, go ahead with the folds. 3148 if (0 && N3C && N3C->isNullValue() && N2C && (N2C->getValue() == 1ULL)) { 3149 MVT::ValueType XType = N0.getValueType(); 3150 if (TLI.isOperationLegal(ISD::SETCC, TLI.getSetCCResultTy())) { 3151 SDOperand Res = DAG.getSetCC(TLI.getSetCCResultTy(), N0, N1, CC); 3152 if (Res.getValueType() != VT) 3153 Res = DAG.getNode(ISD::ZERO_EXTEND, VT, Res); 3154 return Res; 3155 } 3156 3157 // seteq X, 0 -> srl (ctlz X, log2(size(X))) 3158 if (N1C && N1C->isNullValue() && CC == ISD::SETEQ && 3159 TLI.isOperationLegal(ISD::CTLZ, XType)) { 3160 SDOperand Ctlz = DAG.getNode(ISD::CTLZ, XType, N0); 3161 return DAG.getNode(ISD::SRL, XType, Ctlz, 3162 DAG.getConstant(Log2_32(MVT::getSizeInBits(XType)), 3163 TLI.getShiftAmountTy())); 3164 } 3165 // setgt X, 0 -> srl (and (-X, ~X), size(X)-1) 3166 if (N1C && N1C->isNullValue() && CC == ISD::SETGT) { 3167 SDOperand NegN0 = DAG.getNode(ISD::SUB, XType, DAG.getConstant(0, XType), 3168 N0); 3169 SDOperand NotN0 = DAG.getNode(ISD::XOR, XType, N0, 3170 DAG.getConstant(~0ULL, XType)); 3171 return DAG.getNode(ISD::SRL, XType, 3172 DAG.getNode(ISD::AND, XType, NegN0, NotN0), 3173 DAG.getConstant(MVT::getSizeInBits(XType)-1, 3174 TLI.getShiftAmountTy())); 3175 } 3176 // setgt X, -1 -> xor (srl (X, size(X)-1), 1) 3177 if (N1C && N1C->isAllOnesValue() && CC == ISD::SETGT) { 3178 SDOperand Sign = DAG.getNode(ISD::SRL, XType, N0, 3179 DAG.getConstant(MVT::getSizeInBits(XType)-1, 3180 TLI.getShiftAmountTy())); 3181 return DAG.getNode(ISD::XOR, XType, Sign, DAG.getConstant(1, XType)); 3182 } 3183 } 3184 3185 // Check to see if this is an integer abs. select_cc setl[te] X, 0, -X, X -> 3186 // Y = sra (X, size(X)-1); xor (add (X, Y), Y) 3187 if (N1C && N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE) && 3188 N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1)) { 3189 if (ConstantSDNode *SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0))) { 3190 MVT::ValueType XType = N0.getValueType(); 3191 if (SubC->isNullValue() && MVT::isInteger(XType)) { 3192 SDOperand Shift = DAG.getNode(ISD::SRA, XType, N0, 3193 DAG.getConstant(MVT::getSizeInBits(XType)-1, 3194 TLI.getShiftAmountTy())); 3195 SDOperand Add = DAG.getNode(ISD::ADD, XType, N0, Shift); 3196 AddToWorkList(Shift.Val); 3197 AddToWorkList(Add.Val); 3198 return DAG.getNode(ISD::XOR, XType, Add, Shift); 3199 } 3200 } 3201 } 3202 3203 return SDOperand(); 3204} 3205 3206SDOperand DAGCombiner::SimplifySetCC(MVT::ValueType VT, SDOperand N0, 3207 SDOperand N1, ISD::CondCode Cond, 3208 bool foldBooleans) { 3209 // These setcc operations always fold. 3210 switch (Cond) { 3211 default: break; 3212 case ISD::SETFALSE: 3213 case ISD::SETFALSE2: return DAG.getConstant(0, VT); 3214 case ISD::SETTRUE: 3215 case ISD::SETTRUE2: return DAG.getConstant(1, VT); 3216 } 3217 3218 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val)) { 3219 uint64_t C1 = N1C->getValue(); 3220 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val)) { 3221 uint64_t C0 = N0C->getValue(); 3222 3223 // Sign extend the operands if required 3224 if (ISD::isSignedIntSetCC(Cond)) { 3225 C0 = N0C->getSignExtended(); 3226 C1 = N1C->getSignExtended(); 3227 } 3228 3229 switch (Cond) { 3230 default: assert(0 && "Unknown integer setcc!"); 3231 case ISD::SETEQ: return DAG.getConstant(C0 == C1, VT); 3232 case ISD::SETNE: return DAG.getConstant(C0 != C1, VT); 3233 case ISD::SETULT: return DAG.getConstant(C0 < C1, VT); 3234 case ISD::SETUGT: return DAG.getConstant(C0 > C1, VT); 3235 case ISD::SETULE: return DAG.getConstant(C0 <= C1, VT); 3236 case ISD::SETUGE: return DAG.getConstant(C0 >= C1, VT); 3237 case ISD::SETLT: return DAG.getConstant((int64_t)C0 < (int64_t)C1, VT); 3238 case ISD::SETGT: return DAG.getConstant((int64_t)C0 > (int64_t)C1, VT); 3239 case ISD::SETLE: return DAG.getConstant((int64_t)C0 <= (int64_t)C1, VT); 3240 case ISD::SETGE: return DAG.getConstant((int64_t)C0 >= (int64_t)C1, VT); 3241 } 3242 } else { 3243 // If the LHS is a ZERO_EXTEND, perform the comparison on the input. 3244 if (N0.getOpcode() == ISD::ZERO_EXTEND) { 3245 unsigned InSize = MVT::getSizeInBits(N0.getOperand(0).getValueType()); 3246 3247 // If the comparison constant has bits in the upper part, the 3248 // zero-extended value could never match. 3249 if (C1 & (~0ULL << InSize)) { 3250 unsigned VSize = MVT::getSizeInBits(N0.getValueType()); 3251 switch (Cond) { 3252 case ISD::SETUGT: 3253 case ISD::SETUGE: 3254 case ISD::SETEQ: return DAG.getConstant(0, VT); 3255 case ISD::SETULT: 3256 case ISD::SETULE: 3257 case ISD::SETNE: return DAG.getConstant(1, VT); 3258 case ISD::SETGT: 3259 case ISD::SETGE: 3260 // True if the sign bit of C1 is set. 3261 return DAG.getConstant((C1 & (1ULL << VSize)) != 0, VT); 3262 case ISD::SETLT: 3263 case ISD::SETLE: 3264 // True if the sign bit of C1 isn't set. 3265 return DAG.getConstant((C1 & (1ULL << VSize)) == 0, VT); 3266 default: 3267 break; 3268 } 3269 } 3270 3271 // Otherwise, we can perform the comparison with the low bits. 3272 switch (Cond) { 3273 case ISD::SETEQ: 3274 case ISD::SETNE: 3275 case ISD::SETUGT: 3276 case ISD::SETUGE: 3277 case ISD::SETULT: 3278 case ISD::SETULE: 3279 return DAG.getSetCC(VT, N0.getOperand(0), 3280 DAG.getConstant(C1, N0.getOperand(0).getValueType()), 3281 Cond); 3282 default: 3283 break; // todo, be more careful with signed comparisons 3284 } 3285 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG && 3286 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) { 3287 MVT::ValueType ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT(); 3288 unsigned ExtSrcTyBits = MVT::getSizeInBits(ExtSrcTy); 3289 MVT::ValueType ExtDstTy = N0.getValueType(); 3290 unsigned ExtDstTyBits = MVT::getSizeInBits(ExtDstTy); 3291 3292 // If the extended part has any inconsistent bits, it cannot ever 3293 // compare equal. In other words, they have to be all ones or all 3294 // zeros. 3295 uint64_t ExtBits = 3296 (~0ULL >> (64-ExtSrcTyBits)) & (~0ULL << (ExtDstTyBits-1)); 3297 if ((C1 & ExtBits) != 0 && (C1 & ExtBits) != ExtBits) 3298 return DAG.getConstant(Cond == ISD::SETNE, VT); 3299 3300 SDOperand ZextOp; 3301 MVT::ValueType Op0Ty = N0.getOperand(0).getValueType(); 3302 if (Op0Ty == ExtSrcTy) { 3303 ZextOp = N0.getOperand(0); 3304 } else { 3305 int64_t Imm = ~0ULL >> (64-ExtSrcTyBits); 3306 ZextOp = DAG.getNode(ISD::AND, Op0Ty, N0.getOperand(0), 3307 DAG.getConstant(Imm, Op0Ty)); 3308 } 3309 AddToWorkList(ZextOp.Val); 3310 // Otherwise, make this a use of a zext. 3311 return DAG.getSetCC(VT, ZextOp, 3312 DAG.getConstant(C1 & (~0ULL>>(64-ExtSrcTyBits)), 3313 ExtDstTy), 3314 Cond); 3315 } else if ((N1C->getValue() == 0 || N1C->getValue() == 1) && 3316 (Cond == ISD::SETEQ || Cond == ISD::SETNE) && 3317 (N0.getOpcode() == ISD::XOR || 3318 (N0.getOpcode() == ISD::AND && 3319 N0.getOperand(0).getOpcode() == ISD::XOR && 3320 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) && 3321 isa<ConstantSDNode>(N0.getOperand(1)) && 3322 cast<ConstantSDNode>(N0.getOperand(1))->getValue() == 1) { 3323 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We can 3324 // only do this if the top bits are known zero. 3325 if (TLI.MaskedValueIsZero(N1, 3326 MVT::getIntVTBitMask(N0.getValueType())-1)) { 3327 // Okay, get the un-inverted input value. 3328 SDOperand Val; 3329 if (N0.getOpcode() == ISD::XOR) 3330 Val = N0.getOperand(0); 3331 else { 3332 assert(N0.getOpcode() == ISD::AND && 3333 N0.getOperand(0).getOpcode() == ISD::XOR); 3334 // ((X^1)&1)^1 -> X & 1 3335 Val = DAG.getNode(ISD::AND, N0.getValueType(), 3336 N0.getOperand(0).getOperand(0), N0.getOperand(1)); 3337 } 3338 return DAG.getSetCC(VT, Val, N1, 3339 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ); 3340 } 3341 } 3342 3343 uint64_t MinVal, MaxVal; 3344 unsigned OperandBitSize = MVT::getSizeInBits(N1C->getValueType(0)); 3345 if (ISD::isSignedIntSetCC(Cond)) { 3346 MinVal = 1ULL << (OperandBitSize-1); 3347 if (OperandBitSize != 1) // Avoid X >> 64, which is undefined. 3348 MaxVal = ~0ULL >> (65-OperandBitSize); 3349 else 3350 MaxVal = 0; 3351 } else { 3352 MinVal = 0; 3353 MaxVal = ~0ULL >> (64-OperandBitSize); 3354 } 3355 3356 // Canonicalize GE/LE comparisons to use GT/LT comparisons. 3357 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) { 3358 if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true 3359 --C1; // X >= C0 --> X > (C0-1) 3360 return DAG.getSetCC(VT, N0, DAG.getConstant(C1, N1.getValueType()), 3361 (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT); 3362 } 3363 3364 if (Cond == ISD::SETLE || Cond == ISD::SETULE) { 3365 if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true 3366 ++C1; // X <= C0 --> X < (C0+1) 3367 return DAG.getSetCC(VT, N0, DAG.getConstant(C1, N1.getValueType()), 3368 (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT); 3369 } 3370 3371 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal) 3372 return DAG.getConstant(0, VT); // X < MIN --> false 3373 3374 // Canonicalize setgt X, Min --> setne X, Min 3375 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal) 3376 return DAG.getSetCC(VT, N0, N1, ISD::SETNE); 3377 // Canonicalize setlt X, Max --> setne X, Max 3378 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal) 3379 return DAG.getSetCC(VT, N0, N1, ISD::SETNE); 3380 3381 // If we have setult X, 1, turn it into seteq X, 0 3382 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1) 3383 return DAG.getSetCC(VT, N0, DAG.getConstant(MinVal, N0.getValueType()), 3384 ISD::SETEQ); 3385 // If we have setugt X, Max-1, turn it into seteq X, Max 3386 else if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1) 3387 return DAG.getSetCC(VT, N0, DAG.getConstant(MaxVal, N0.getValueType()), 3388 ISD::SETEQ); 3389 3390 // If we have "setcc X, C0", check to see if we can shrink the immediate 3391 // by changing cc. 3392 3393 // SETUGT X, SINTMAX -> SETLT X, 0 3394 if (Cond == ISD::SETUGT && OperandBitSize != 1 && 3395 C1 == (~0ULL >> (65-OperandBitSize))) 3396 return DAG.getSetCC(VT, N0, DAG.getConstant(0, N1.getValueType()), 3397 ISD::SETLT); 3398 3399 // FIXME: Implement the rest of these. 3400 3401 // Fold bit comparisons when we can. 3402 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && 3403 VT == N0.getValueType() && N0.getOpcode() == ISD::AND) 3404 if (ConstantSDNode *AndRHS = 3405 dyn_cast<ConstantSDNode>(N0.getOperand(1))) { 3406 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3 3407 // Perform the xform if the AND RHS is a single bit. 3408 if ((AndRHS->getValue() & (AndRHS->getValue()-1)) == 0) { 3409 return DAG.getNode(ISD::SRL, VT, N0, 3410 DAG.getConstant(Log2_64(AndRHS->getValue()), 3411 TLI.getShiftAmountTy())); 3412 } 3413 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getValue()) { 3414 // (X & 8) == 8 --> (X & 8) >> 3 3415 // Perform the xform if C1 is a single bit. 3416 if ((C1 & (C1-1)) == 0) { 3417 return DAG.getNode(ISD::SRL, VT, N0, 3418 DAG.getConstant(Log2_64(C1),TLI.getShiftAmountTy())); 3419 } 3420 } 3421 } 3422 } 3423 } else if (isa<ConstantSDNode>(N0.Val)) { 3424 // Ensure that the constant occurs on the RHS. 3425 return DAG.getSetCC(VT, N1, N0, ISD::getSetCCSwappedOperands(Cond)); 3426 } 3427 3428 if (ConstantFPSDNode *N0C = dyn_cast<ConstantFPSDNode>(N0.Val)) 3429 if (ConstantFPSDNode *N1C = dyn_cast<ConstantFPSDNode>(N1.Val)) { 3430 double C0 = N0C->getValue(), C1 = N1C->getValue(); 3431 3432 switch (Cond) { 3433 default: break; // FIXME: Implement the rest of these! 3434 case ISD::SETEQ: return DAG.getConstant(C0 == C1, VT); 3435 case ISD::SETNE: return DAG.getConstant(C0 != C1, VT); 3436 case ISD::SETLT: return DAG.getConstant(C0 < C1, VT); 3437 case ISD::SETGT: return DAG.getConstant(C0 > C1, VT); 3438 case ISD::SETLE: return DAG.getConstant(C0 <= C1, VT); 3439 case ISD::SETGE: return DAG.getConstant(C0 >= C1, VT); 3440 } 3441 } else { 3442 // Ensure that the constant occurs on the RHS. 3443 return DAG.getSetCC(VT, N1, N0, ISD::getSetCCSwappedOperands(Cond)); 3444 } 3445 3446 if (N0 == N1) { 3447 // We can always fold X == Y for integer setcc's. 3448 if (MVT::isInteger(N0.getValueType())) 3449 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT); 3450 unsigned UOF = ISD::getUnorderedFlavor(Cond); 3451 if (UOF == 2) // FP operators that are undefined on NaNs. 3452 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT); 3453 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond))) 3454 return DAG.getConstant(UOF, VT); 3455 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO 3456 // if it is not already. 3457 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO; 3458 if (NewCond != Cond) 3459 return DAG.getSetCC(VT, N0, N1, NewCond); 3460 } 3461 3462 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && 3463 MVT::isInteger(N0.getValueType())) { 3464 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB || 3465 N0.getOpcode() == ISD::XOR) { 3466 // Simplify (X+Y) == (X+Z) --> Y == Z 3467 if (N0.getOpcode() == N1.getOpcode()) { 3468 if (N0.getOperand(0) == N1.getOperand(0)) 3469 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(1), Cond); 3470 if (N0.getOperand(1) == N1.getOperand(1)) 3471 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(0), Cond); 3472 if (isCommutativeBinOp(N0.getOpcode())) { 3473 // If X op Y == Y op X, try other combinations. 3474 if (N0.getOperand(0) == N1.getOperand(1)) 3475 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(0), Cond); 3476 if (N0.getOperand(1) == N1.getOperand(0)) 3477 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(1), Cond); 3478 } 3479 } 3480 3481 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) { 3482 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) { 3483 // Turn (X+C1) == C2 --> X == C2-C1 3484 if (N0.getOpcode() == ISD::ADD && N0.Val->hasOneUse()) { 3485 return DAG.getSetCC(VT, N0.getOperand(0), 3486 DAG.getConstant(RHSC->getValue()-LHSR->getValue(), 3487 N0.getValueType()), Cond); 3488 } 3489 3490 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0. 3491 if (N0.getOpcode() == ISD::XOR) 3492 // If we know that all of the inverted bits are zero, don't bother 3493 // performing the inversion. 3494 if (TLI.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getValue())) 3495 return DAG.getSetCC(VT, N0.getOperand(0), 3496 DAG.getConstant(LHSR->getValue()^RHSC->getValue(), 3497 N0.getValueType()), Cond); 3498 } 3499 3500 // Turn (C1-X) == C2 --> X == C1-C2 3501 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) { 3502 if (N0.getOpcode() == ISD::SUB && N0.Val->hasOneUse()) { 3503 return DAG.getSetCC(VT, N0.getOperand(1), 3504 DAG.getConstant(SUBC->getValue()-RHSC->getValue(), 3505 N0.getValueType()), Cond); 3506 } 3507 } 3508 } 3509 3510 // Simplify (X+Z) == X --> Z == 0 3511 if (N0.getOperand(0) == N1) 3512 return DAG.getSetCC(VT, N0.getOperand(1), 3513 DAG.getConstant(0, N0.getValueType()), Cond); 3514 if (N0.getOperand(1) == N1) { 3515 if (isCommutativeBinOp(N0.getOpcode())) 3516 return DAG.getSetCC(VT, N0.getOperand(0), 3517 DAG.getConstant(0, N0.getValueType()), Cond); 3518 else { 3519 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!"); 3520 // (Z-X) == X --> Z == X<<1 3521 SDOperand SH = DAG.getNode(ISD::SHL, N1.getValueType(), 3522 N1, 3523 DAG.getConstant(1,TLI.getShiftAmountTy())); 3524 AddToWorkList(SH.Val); 3525 return DAG.getSetCC(VT, N0.getOperand(0), SH, Cond); 3526 } 3527 } 3528 } 3529 3530 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB || 3531 N1.getOpcode() == ISD::XOR) { 3532 // Simplify X == (X+Z) --> Z == 0 3533 if (N1.getOperand(0) == N0) { 3534 return DAG.getSetCC(VT, N1.getOperand(1), 3535 DAG.getConstant(0, N1.getValueType()), Cond); 3536 } else if (N1.getOperand(1) == N0) { 3537 if (isCommutativeBinOp(N1.getOpcode())) { 3538 return DAG.getSetCC(VT, N1.getOperand(0), 3539 DAG.getConstant(0, N1.getValueType()), Cond); 3540 } else { 3541 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!"); 3542 // X == (Z-X) --> X<<1 == Z 3543 SDOperand SH = DAG.getNode(ISD::SHL, N1.getValueType(), N0, 3544 DAG.getConstant(1,TLI.getShiftAmountTy())); 3545 AddToWorkList(SH.Val); 3546 return DAG.getSetCC(VT, SH, N1.getOperand(0), Cond); 3547 } 3548 } 3549 } 3550 } 3551 3552 // Fold away ALL boolean setcc's. 3553 SDOperand Temp; 3554 if (N0.getValueType() == MVT::i1 && foldBooleans) { 3555 switch (Cond) { 3556 default: assert(0 && "Unknown integer setcc!"); 3557 case ISD::SETEQ: // X == Y -> (X^Y)^1 3558 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, N1); 3559 N0 = DAG.getNode(ISD::XOR, MVT::i1, Temp, DAG.getConstant(1, MVT::i1)); 3560 AddToWorkList(Temp.Val); 3561 break; 3562 case ISD::SETNE: // X != Y --> (X^Y) 3563 N0 = DAG.getNode(ISD::XOR, MVT::i1, N0, N1); 3564 break; 3565 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> X^1 & Y 3566 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> X^1 & Y 3567 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1)); 3568 N0 = DAG.getNode(ISD::AND, MVT::i1, N1, Temp); 3569 AddToWorkList(Temp.Val); 3570 break; 3571 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> Y^1 & X 3572 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> Y^1 & X 3573 Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1)); 3574 N0 = DAG.getNode(ISD::AND, MVT::i1, N0, Temp); 3575 AddToWorkList(Temp.Val); 3576 break; 3577 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> X^1 | Y 3578 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> X^1 | Y 3579 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1)); 3580 N0 = DAG.getNode(ISD::OR, MVT::i1, N1, Temp); 3581 AddToWorkList(Temp.Val); 3582 break; 3583 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> Y^1 | X 3584 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> Y^1 | X 3585 Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1)); 3586 N0 = DAG.getNode(ISD::OR, MVT::i1, N0, Temp); 3587 break; 3588 } 3589 if (VT != MVT::i1) { 3590 AddToWorkList(N0.Val); 3591 // FIXME: If running after legalize, we probably can't do this. 3592 N0 = DAG.getNode(ISD::ZERO_EXTEND, VT, N0); 3593 } 3594 return N0; 3595 } 3596 3597 // Could not fold it. 3598 return SDOperand(); 3599} 3600 3601/// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant, 3602/// return a DAG expression to select that will generate the same value by 3603/// multiplying by a magic number. See: 3604/// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html> 3605SDOperand DAGCombiner::BuildSDIV(SDNode *N) { 3606 MVT::ValueType VT = N->getValueType(0); 3607 3608 // Check to see if we can do this. 3609 if (!TLI.isTypeLegal(VT) || (VT != MVT::i32 && VT != MVT::i64)) 3610 return SDOperand(); // BuildSDIV only operates on i32 or i64 3611 if (!TLI.isOperationLegal(ISD::MULHS, VT)) 3612 return SDOperand(); // Make sure the target supports MULHS. 3613 3614 int64_t d = cast<ConstantSDNode>(N->getOperand(1))->getSignExtended(); 3615 ms magics = (VT == MVT::i32) ? magic32(d) : magic64(d); 3616 3617 // Multiply the numerator (operand 0) by the magic value 3618 SDOperand Q = DAG.getNode(ISD::MULHS, VT, N->getOperand(0), 3619 DAG.getConstant(magics.m, VT)); 3620 // If d > 0 and m < 0, add the numerator 3621 if (d > 0 && magics.m < 0) { 3622 Q = DAG.getNode(ISD::ADD, VT, Q, N->getOperand(0)); 3623 AddToWorkList(Q.Val); 3624 } 3625 // If d < 0 and m > 0, subtract the numerator. 3626 if (d < 0 && magics.m > 0) { 3627 Q = DAG.getNode(ISD::SUB, VT, Q, N->getOperand(0)); 3628 AddToWorkList(Q.Val); 3629 } 3630 // Shift right algebraic if shift value is nonzero 3631 if (magics.s > 0) { 3632 Q = DAG.getNode(ISD::SRA, VT, Q, 3633 DAG.getConstant(magics.s, TLI.getShiftAmountTy())); 3634 AddToWorkList(Q.Val); 3635 } 3636 // Extract the sign bit and add it to the quotient 3637 SDOperand T = 3638 DAG.getNode(ISD::SRL, VT, Q, DAG.getConstant(MVT::getSizeInBits(VT)-1, 3639 TLI.getShiftAmountTy())); 3640 AddToWorkList(T.Val); 3641 return DAG.getNode(ISD::ADD, VT, Q, T); 3642} 3643 3644/// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant, 3645/// return a DAG expression to select that will generate the same value by 3646/// multiplying by a magic number. See: 3647/// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html> 3648SDOperand DAGCombiner::BuildUDIV(SDNode *N) { 3649 MVT::ValueType VT = N->getValueType(0); 3650 3651 // Check to see if we can do this. 3652 if (!TLI.isTypeLegal(VT) || (VT != MVT::i32 && VT != MVT::i64)) 3653 return SDOperand(); // BuildUDIV only operates on i32 or i64 3654 if (!TLI.isOperationLegal(ISD::MULHU, VT)) 3655 return SDOperand(); // Make sure the target supports MULHU. 3656 3657 uint64_t d = cast<ConstantSDNode>(N->getOperand(1))->getValue(); 3658 mu magics = (VT == MVT::i32) ? magicu32(d) : magicu64(d); 3659 3660 // Multiply the numerator (operand 0) by the magic value 3661 SDOperand Q = DAG.getNode(ISD::MULHU, VT, N->getOperand(0), 3662 DAG.getConstant(magics.m, VT)); 3663 AddToWorkList(Q.Val); 3664 3665 if (magics.a == 0) { 3666 return DAG.getNode(ISD::SRL, VT, Q, 3667 DAG.getConstant(magics.s, TLI.getShiftAmountTy())); 3668 } else { 3669 SDOperand NPQ = DAG.getNode(ISD::SUB, VT, N->getOperand(0), Q); 3670 AddToWorkList(NPQ.Val); 3671 NPQ = DAG.getNode(ISD::SRL, VT, NPQ, 3672 DAG.getConstant(1, TLI.getShiftAmountTy())); 3673 AddToWorkList(NPQ.Val); 3674 NPQ = DAG.getNode(ISD::ADD, VT, NPQ, Q); 3675 AddToWorkList(NPQ.Val); 3676 return DAG.getNode(ISD::SRL, VT, NPQ, 3677 DAG.getConstant(magics.s-1, TLI.getShiftAmountTy())); 3678 } 3679} 3680 3681// SelectionDAG::Combine - This is the entry point for the file. 3682// 3683void SelectionDAG::Combine(bool RunningAfterLegalize) { 3684 /// run - This is the main entry point to this class. 3685 /// 3686 DAGCombiner(*this).Run(RunningAfterLegalize); 3687} 3688