DAGCombiner.cpp revision da95a84a11e684072ea25fbb9830e769039fd985
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: Should add a corresponding version of fold AND with 20// ZERO_EXTEND/SIGN_EXTEND by converting them to an ANY_EXTEND node which 21// we don't have yet. 22// 23// FIXME: select C, pow2, pow2 -> something smart 24// FIXME: trunc(select X, Y, Z) -> select X, trunc(Y), trunc(Z) 25// FIXME: Dead stores -> nuke 26// FIXME: shr X, (and Y,31) -> shr X, Y (TRICKY!) 27// FIXME: mul (x, const) -> shifts + adds 28// FIXME: undef values 29// FIXME: make truncate see through SIGN_EXTEND and AND 30// FIXME: (sra (sra x, c1), c2) -> (sra x, c1+c2) 31// FIXME: verify that getNode can't return extends with an operand whose type 32// is >= to that of the extend. 33// FIXME: divide by zero is currently left unfolded. do we want to turn this 34// into an undef? 35// FIXME: select ne (select cc, 1, 0), 0, true, false -> select cc, true, false 36// FIXME: reassociate (X+C)+Y into (X+Y)+C if the inner expression has one use 37// 38//===----------------------------------------------------------------------===// 39 40#define DEBUG_TYPE "dagcombine" 41#include "llvm/ADT/Statistic.h" 42#include "llvm/CodeGen/SelectionDAG.h" 43#include "llvm/Support/Debug.h" 44#include "llvm/Support/MathExtras.h" 45#include "llvm/Target/TargetLowering.h" 46#include <algorithm> 47#include <cmath> 48using namespace llvm; 49 50namespace { 51 Statistic<> NodesCombined ("dagcombiner", "Number of dag nodes combined"); 52 53 class DAGCombiner { 54 SelectionDAG &DAG; 55 TargetLowering &TLI; 56 bool AfterLegalize; 57 58 // Worklist of all of the nodes that need to be simplified. 59 std::vector<SDNode*> WorkList; 60 61 /// AddUsersToWorkList - When an instruction is simplified, add all users of 62 /// the instruction to the work lists because they might get more simplified 63 /// now. 64 /// 65 void AddUsersToWorkList(SDNode *N) { 66 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end(); 67 UI != UE; ++UI) 68 WorkList.push_back(*UI); 69 } 70 71 /// removeFromWorkList - remove all instances of N from the worklist. 72 void removeFromWorkList(SDNode *N) { 73 WorkList.erase(std::remove(WorkList.begin(), WorkList.end(), N), 74 WorkList.end()); 75 } 76 77 SDOperand CombineTo(SDNode *N, const std::vector<SDOperand> &To) { 78 ++NodesCombined; 79 DEBUG(std::cerr << "\nReplacing "; N->dump(); 80 std::cerr << "\nWith: "; To[0].Val->dump(); 81 std::cerr << " and " << To.size()-1 << " other values\n"); 82 std::vector<SDNode*> NowDead; 83 DAG.ReplaceAllUsesWith(N, To, &NowDead); 84 85 // Push the new nodes and any users onto the worklist 86 for (unsigned i = 0, e = To.size(); i != e; ++i) { 87 WorkList.push_back(To[i].Val); 88 AddUsersToWorkList(To[i].Val); 89 } 90 91 // Nodes can end up on the worklist more than once. Make sure we do 92 // not process a node that has been replaced. 93 removeFromWorkList(N); 94 for (unsigned i = 0, e = NowDead.size(); i != e; ++i) 95 removeFromWorkList(NowDead[i]); 96 97 // Finally, since the node is now dead, remove it from the graph. 98 DAG.DeleteNode(N); 99 return SDOperand(N, 0); 100 } 101 102 SDOperand CombineTo(SDNode *N, SDOperand Res) { 103 std::vector<SDOperand> To; 104 To.push_back(Res); 105 return CombineTo(N, To); 106 } 107 108 SDOperand CombineTo(SDNode *N, SDOperand Res0, SDOperand Res1) { 109 std::vector<SDOperand> To; 110 To.push_back(Res0); 111 To.push_back(Res1); 112 return CombineTo(N, To); 113 } 114 115 /// visit - call the node-specific routine that knows how to fold each 116 /// particular type of node. 117 SDOperand visit(SDNode *N); 118 119 // Visitation implementation - Implement dag node combining for different 120 // node types. The semantics are as follows: 121 // Return Value: 122 // SDOperand.Val == 0 - No change was made 123 // SDOperand.Val == N - N was replaced, is dead, and is already handled. 124 // otherwise - N should be replaced by the returned Operand. 125 // 126 SDOperand visitTokenFactor(SDNode *N); 127 SDOperand visitADD(SDNode *N); 128 SDOperand visitSUB(SDNode *N); 129 SDOperand visitMUL(SDNode *N); 130 SDOperand visitSDIV(SDNode *N); 131 SDOperand visitUDIV(SDNode *N); 132 SDOperand visitSREM(SDNode *N); 133 SDOperand visitUREM(SDNode *N); 134 SDOperand visitMULHU(SDNode *N); 135 SDOperand visitMULHS(SDNode *N); 136 SDOperand visitAND(SDNode *N); 137 SDOperand visitOR(SDNode *N); 138 SDOperand visitXOR(SDNode *N); 139 SDOperand visitSHL(SDNode *N); 140 SDOperand visitSRA(SDNode *N); 141 SDOperand visitSRL(SDNode *N); 142 SDOperand visitCTLZ(SDNode *N); 143 SDOperand visitCTTZ(SDNode *N); 144 SDOperand visitCTPOP(SDNode *N); 145 SDOperand visitSELECT(SDNode *N); 146 SDOperand visitSELECT_CC(SDNode *N); 147 SDOperand visitSETCC(SDNode *N); 148 SDOperand visitADD_PARTS(SDNode *N); 149 SDOperand visitSUB_PARTS(SDNode *N); 150 SDOperand visitSIGN_EXTEND(SDNode *N); 151 SDOperand visitZERO_EXTEND(SDNode *N); 152 SDOperand visitSIGN_EXTEND_INREG(SDNode *N); 153 SDOperand visitTRUNCATE(SDNode *N); 154 SDOperand visitBIT_CONVERT(SDNode *N); 155 156 SDOperand visitFADD(SDNode *N); 157 SDOperand visitFSUB(SDNode *N); 158 SDOperand visitFMUL(SDNode *N); 159 SDOperand visitFDIV(SDNode *N); 160 SDOperand visitFREM(SDNode *N); 161 SDOperand visitSINT_TO_FP(SDNode *N); 162 SDOperand visitUINT_TO_FP(SDNode *N); 163 SDOperand visitFP_TO_SINT(SDNode *N); 164 SDOperand visitFP_TO_UINT(SDNode *N); 165 SDOperand visitFP_ROUND(SDNode *N); 166 SDOperand visitFP_ROUND_INREG(SDNode *N); 167 SDOperand visitFP_EXTEND(SDNode *N); 168 SDOperand visitFNEG(SDNode *N); 169 SDOperand visitFABS(SDNode *N); 170 SDOperand visitBRCOND(SDNode *N); 171 SDOperand visitBRCONDTWOWAY(SDNode *N); 172 SDOperand visitBR_CC(SDNode *N); 173 SDOperand visitBRTWOWAY_CC(SDNode *N); 174 175 SDOperand visitLOAD(SDNode *N); 176 SDOperand visitSTORE(SDNode *N); 177 178 SDOperand visitLOCATION(SDNode *N); 179 SDOperand visitDEBUGLOC(SDNode *N); 180 181 bool SimplifySelectOps(SDNode *SELECT, SDOperand LHS, SDOperand RHS); 182 SDOperand SimplifySelect(SDOperand N0, SDOperand N1, SDOperand N2); 183 SDOperand SimplifySelectCC(SDOperand N0, SDOperand N1, SDOperand N2, 184 SDOperand N3, ISD::CondCode CC); 185 SDOperand SimplifySetCC(MVT::ValueType VT, SDOperand N0, SDOperand N1, 186 ISD::CondCode Cond, bool foldBooleans = true); 187 188 SDOperand BuildSDIV(SDNode *N); 189 SDOperand BuildUDIV(SDNode *N); 190public: 191 DAGCombiner(SelectionDAG &D) 192 : DAG(D), TLI(D.getTargetLoweringInfo()), AfterLegalize(false) {} 193 194 /// Run - runs the dag combiner on all nodes in the work list 195 void Run(bool RunningAfterLegalize); 196 }; 197} 198 199struct ms { 200 int64_t m; // magic number 201 int64_t s; // shift amount 202}; 203 204struct mu { 205 uint64_t m; // magic number 206 int64_t a; // add indicator 207 int64_t s; // shift amount 208}; 209 210/// magic - calculate the magic numbers required to codegen an integer sdiv as 211/// a sequence of multiply and shifts. Requires that the divisor not be 0, 1, 212/// or -1. 213static ms magic32(int32_t d) { 214 int32_t p; 215 uint32_t ad, anc, delta, q1, r1, q2, r2, t; 216 const uint32_t two31 = 0x80000000U; 217 struct ms mag; 218 219 ad = abs(d); 220 t = two31 + ((uint32_t)d >> 31); 221 anc = t - 1 - t%ad; // absolute value of nc 222 p = 31; // initialize p 223 q1 = two31/anc; // initialize q1 = 2p/abs(nc) 224 r1 = two31 - q1*anc; // initialize r1 = rem(2p,abs(nc)) 225 q2 = two31/ad; // initialize q2 = 2p/abs(d) 226 r2 = two31 - q2*ad; // initialize r2 = rem(2p,abs(d)) 227 do { 228 p = p + 1; 229 q1 = 2*q1; // update q1 = 2p/abs(nc) 230 r1 = 2*r1; // update r1 = rem(2p/abs(nc)) 231 if (r1 >= anc) { // must be unsigned comparison 232 q1 = q1 + 1; 233 r1 = r1 - anc; 234 } 235 q2 = 2*q2; // update q2 = 2p/abs(d) 236 r2 = 2*r2; // update r2 = rem(2p/abs(d)) 237 if (r2 >= ad) { // must be unsigned comparison 238 q2 = q2 + 1; 239 r2 = r2 - ad; 240 } 241 delta = ad - r2; 242 } while (q1 < delta || (q1 == delta && r1 == 0)); 243 244 mag.m = (int32_t)(q2 + 1); // make sure to sign extend 245 if (d < 0) mag.m = -mag.m; // resulting magic number 246 mag.s = p - 32; // resulting shift 247 return mag; 248} 249 250/// magicu - calculate the magic numbers required to codegen an integer udiv as 251/// a sequence of multiply, add and shifts. Requires that the divisor not be 0. 252static mu magicu32(uint32_t d) { 253 int32_t p; 254 uint32_t nc, delta, q1, r1, q2, r2; 255 struct mu magu; 256 magu.a = 0; // initialize "add" indicator 257 nc = - 1 - (-d)%d; 258 p = 31; // initialize p 259 q1 = 0x80000000/nc; // initialize q1 = 2p/nc 260 r1 = 0x80000000 - q1*nc; // initialize r1 = rem(2p,nc) 261 q2 = 0x7FFFFFFF/d; // initialize q2 = (2p-1)/d 262 r2 = 0x7FFFFFFF - q2*d; // initialize r2 = rem((2p-1),d) 263 do { 264 p = p + 1; 265 if (r1 >= nc - r1 ) { 266 q1 = 2*q1 + 1; // update q1 267 r1 = 2*r1 - nc; // update r1 268 } 269 else { 270 q1 = 2*q1; // update q1 271 r1 = 2*r1; // update r1 272 } 273 if (r2 + 1 >= d - r2) { 274 if (q2 >= 0x7FFFFFFF) magu.a = 1; 275 q2 = 2*q2 + 1; // update q2 276 r2 = 2*r2 + 1 - d; // update r2 277 } 278 else { 279 if (q2 >= 0x80000000) magu.a = 1; 280 q2 = 2*q2; // update q2 281 r2 = 2*r2 + 1; // update r2 282 } 283 delta = d - 1 - r2; 284 } while (p < 64 && (q1 < delta || (q1 == delta && r1 == 0))); 285 magu.m = q2 + 1; // resulting magic number 286 magu.s = p - 32; // resulting shift 287 return magu; 288} 289 290/// magic - calculate the magic numbers required to codegen an integer sdiv as 291/// a sequence of multiply and shifts. Requires that the divisor not be 0, 1, 292/// or -1. 293static ms magic64(int64_t d) { 294 int64_t p; 295 uint64_t ad, anc, delta, q1, r1, q2, r2, t; 296 const uint64_t two63 = 9223372036854775808ULL; // 2^63 297 struct ms mag; 298 299 ad = d >= 0 ? d : -d; 300 t = two63 + ((uint64_t)d >> 63); 301 anc = t - 1 - t%ad; // absolute value of nc 302 p = 63; // initialize p 303 q1 = two63/anc; // initialize q1 = 2p/abs(nc) 304 r1 = two63 - q1*anc; // initialize r1 = rem(2p,abs(nc)) 305 q2 = two63/ad; // initialize q2 = 2p/abs(d) 306 r2 = two63 - q2*ad; // initialize r2 = rem(2p,abs(d)) 307 do { 308 p = p + 1; 309 q1 = 2*q1; // update q1 = 2p/abs(nc) 310 r1 = 2*r1; // update r1 = rem(2p/abs(nc)) 311 if (r1 >= anc) { // must be unsigned comparison 312 q1 = q1 + 1; 313 r1 = r1 - anc; 314 } 315 q2 = 2*q2; // update q2 = 2p/abs(d) 316 r2 = 2*r2; // update r2 = rem(2p/abs(d)) 317 if (r2 >= ad) { // must be unsigned comparison 318 q2 = q2 + 1; 319 r2 = r2 - ad; 320 } 321 delta = ad - r2; 322 } while (q1 < delta || (q1 == delta && r1 == 0)); 323 324 mag.m = q2 + 1; 325 if (d < 0) mag.m = -mag.m; // resulting magic number 326 mag.s = p - 64; // resulting shift 327 return mag; 328} 329 330/// magicu - calculate the magic numbers required to codegen an integer udiv as 331/// a sequence of multiply, add and shifts. Requires that the divisor not be 0. 332static mu magicu64(uint64_t d) 333{ 334 int64_t p; 335 uint64_t nc, delta, q1, r1, q2, r2; 336 struct mu magu; 337 magu.a = 0; // initialize "add" indicator 338 nc = - 1 - (-d)%d; 339 p = 63; // initialize p 340 q1 = 0x8000000000000000ull/nc; // initialize q1 = 2p/nc 341 r1 = 0x8000000000000000ull - q1*nc; // initialize r1 = rem(2p,nc) 342 q2 = 0x7FFFFFFFFFFFFFFFull/d; // initialize q2 = (2p-1)/d 343 r2 = 0x7FFFFFFFFFFFFFFFull - q2*d; // initialize r2 = rem((2p-1),d) 344 do { 345 p = p + 1; 346 if (r1 >= nc - r1 ) { 347 q1 = 2*q1 + 1; // update q1 348 r1 = 2*r1 - nc; // update r1 349 } 350 else { 351 q1 = 2*q1; // update q1 352 r1 = 2*r1; // update r1 353 } 354 if (r2 + 1 >= d - r2) { 355 if (q2 >= 0x7FFFFFFFFFFFFFFFull) magu.a = 1; 356 q2 = 2*q2 + 1; // update q2 357 r2 = 2*r2 + 1 - d; // update r2 358 } 359 else { 360 if (q2 >= 0x8000000000000000ull) magu.a = 1; 361 q2 = 2*q2; // update q2 362 r2 = 2*r2 + 1; // update r2 363 } 364 delta = d - 1 - r2; 365 } while (p < 64 && (q1 < delta || (q1 == delta && r1 == 0))); 366 magu.m = q2 + 1; // resulting magic number 367 magu.s = p - 64; // resulting shift 368 return magu; 369} 370 371/// MaskedValueIsZero - Return true if 'Op & Mask' is known to be zero. We use 372/// this predicate to simplify operations downstream. Op and Mask are known to 373/// be the same type. 374static bool MaskedValueIsZero(const SDOperand &Op, uint64_t Mask, 375 const TargetLowering &TLI) { 376 unsigned SrcBits; 377 if (Mask == 0) return true; 378 379 // If we know the result of a setcc has the top bits zero, use this info. 380 switch (Op.getOpcode()) { 381 case ISD::Constant: 382 return (cast<ConstantSDNode>(Op)->getValue() & Mask) == 0; 383 case ISD::SETCC: 384 return ((Mask & 1) == 0) && 385 TLI.getSetCCResultContents() == TargetLowering::ZeroOrOneSetCCResult; 386 case ISD::ZEXTLOAD: 387 SrcBits = MVT::getSizeInBits(cast<VTSDNode>(Op.getOperand(3))->getVT()); 388 return (Mask & ((1ULL << SrcBits)-1)) == 0; // Returning only the zext bits. 389 case ISD::ZERO_EXTEND: 390 SrcBits = MVT::getSizeInBits(Op.getOperand(0).getValueType()); 391 return MaskedValueIsZero(Op.getOperand(0),Mask & (~0ULL >> (64-SrcBits)),TLI); 392 case ISD::AssertZext: 393 SrcBits = MVT::getSizeInBits(cast<VTSDNode>(Op.getOperand(1))->getVT()); 394 return (Mask & ((1ULL << SrcBits)-1)) == 0; // Returning only the zext bits. 395 case ISD::AND: 396 // If either of the operands has zero bits, the result will too. 397 if (MaskedValueIsZero(Op.getOperand(1), Mask, TLI) || 398 MaskedValueIsZero(Op.getOperand(0), Mask, TLI)) 399 return true; 400 // (X & C1) & C2 == 0 iff C1 & C2 == 0. 401 if (ConstantSDNode *AndRHS = dyn_cast<ConstantSDNode>(Op.getOperand(1))) 402 return MaskedValueIsZero(Op.getOperand(0),AndRHS->getValue() & Mask, TLI); 403 return false; 404 case ISD::OR: 405 case ISD::XOR: 406 return MaskedValueIsZero(Op.getOperand(0), Mask, TLI) && 407 MaskedValueIsZero(Op.getOperand(1), Mask, TLI); 408 case ISD::SELECT: 409 return MaskedValueIsZero(Op.getOperand(1), Mask, TLI) && 410 MaskedValueIsZero(Op.getOperand(2), Mask, TLI); 411 case ISD::SELECT_CC: 412 return MaskedValueIsZero(Op.getOperand(2), Mask, TLI) && 413 MaskedValueIsZero(Op.getOperand(3), Mask, TLI); 414 case ISD::SRL: 415 // (ushr X, C1) & C2 == 0 iff X & (C2 << C1) == 0 416 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 417 uint64_t NewVal = Mask << ShAmt->getValue(); 418 SrcBits = MVT::getSizeInBits(Op.getValueType()); 419 if (SrcBits != 64) NewVal &= (1ULL << SrcBits)-1; 420 return MaskedValueIsZero(Op.getOperand(0), NewVal, TLI); 421 } 422 return false; 423 case ISD::SHL: 424 // (ushl X, C1) & C2 == 0 iff X & (C2 >> C1) == 0 425 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 426 uint64_t NewVal = Mask >> ShAmt->getValue(); 427 return MaskedValueIsZero(Op.getOperand(0), NewVal, TLI); 428 } 429 return false; 430 case ISD::ADD: 431 // (add X, Y) & C == 0 iff (X&C)|(Y&C) == 0 and all bits are low bits. 432 if ((Mask&(Mask+1)) == 0) { // All low bits 433 if (MaskedValueIsZero(Op.getOperand(0), Mask, TLI) && 434 MaskedValueIsZero(Op.getOperand(1), Mask, TLI)) 435 return true; 436 } 437 break; 438 case ISD::SUB: 439 if (ConstantSDNode *CLHS = dyn_cast<ConstantSDNode>(Op.getOperand(0))) { 440 // We know that the top bits of C-X are clear if X contains less bits 441 // than C (i.e. no wrap-around can happen). For example, 20-X is 442 // positive if we can prove that X is >= 0 and < 16. 443 unsigned Bits = MVT::getSizeInBits(CLHS->getValueType(0)); 444 if ((CLHS->getValue() & (1 << (Bits-1))) == 0) { // sign bit clear 445 unsigned NLZ = CountLeadingZeros_64(CLHS->getValue()+1); 446 uint64_t MaskV = (1ULL << (63-NLZ))-1; 447 if (MaskedValueIsZero(Op.getOperand(1), ~MaskV, TLI)) { 448 // High bits are clear this value is known to be >= C. 449 unsigned NLZ2 = CountLeadingZeros_64(CLHS->getValue()); 450 if ((Mask & ((1ULL << (64-NLZ2))-1)) == 0) 451 return true; 452 } 453 } 454 } 455 break; 456 case ISD::CTTZ: 457 case ISD::CTLZ: 458 case ISD::CTPOP: 459 // Bit counting instructions can not set the high bits of the result 460 // register. The max number of bits sets depends on the input. 461 return (Mask & (MVT::getSizeInBits(Op.getValueType())*2-1)) == 0; 462 default: 463 if (Op.getOpcode() >= ISD::BUILTIN_OP_END) 464 return TLI.isMaskedValueZeroForTargetNode(Op, Mask); 465 break; 466 } 467 return false; 468} 469 470// isSetCCEquivalent - Return true if this node is a setcc, or is a select_cc 471// that selects between the values 1 and 0, making it equivalent to a setcc. 472// Also, set the incoming LHS, RHS, and CC references to the appropriate 473// nodes based on the type of node we are checking. This simplifies life a 474// bit for the callers. 475static bool isSetCCEquivalent(SDOperand N, SDOperand &LHS, SDOperand &RHS, 476 SDOperand &CC) { 477 if (N.getOpcode() == ISD::SETCC) { 478 LHS = N.getOperand(0); 479 RHS = N.getOperand(1); 480 CC = N.getOperand(2); 481 return true; 482 } 483 if (N.getOpcode() == ISD::SELECT_CC && 484 N.getOperand(2).getOpcode() == ISD::Constant && 485 N.getOperand(3).getOpcode() == ISD::Constant && 486 cast<ConstantSDNode>(N.getOperand(2))->getValue() == 1 && 487 cast<ConstantSDNode>(N.getOperand(3))->isNullValue()) { 488 LHS = N.getOperand(0); 489 RHS = N.getOperand(1); 490 CC = N.getOperand(4); 491 return true; 492 } 493 return false; 494} 495 496// isOneUseSetCC - Return true if this is a SetCC-equivalent operation with only 497// one use. If this is true, it allows the users to invert the operation for 498// free when it is profitable to do so. 499static bool isOneUseSetCC(SDOperand N) { 500 SDOperand N0, N1, N2; 501 if (isSetCCEquivalent(N, N0, N1, N2) && N.Val->hasOneUse()) 502 return true; 503 return false; 504} 505 506// FIXME: This should probably go in the ISD class rather than being duplicated 507// in several files. 508static bool isCommutativeBinOp(unsigned Opcode) { 509 switch (Opcode) { 510 case ISD::ADD: 511 case ISD::MUL: 512 case ISD::AND: 513 case ISD::OR: 514 case ISD::XOR: return true; 515 default: return false; // FIXME: Need commutative info for user ops! 516 } 517} 518 519void DAGCombiner::Run(bool RunningAfterLegalize) { 520 // set the instance variable, so that the various visit routines may use it. 521 AfterLegalize = RunningAfterLegalize; 522 523 // Add all the dag nodes to the worklist. 524 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(), 525 E = DAG.allnodes_end(); I != E; ++I) 526 WorkList.push_back(I); 527 528 // Create a dummy node (which is not added to allnodes), that adds a reference 529 // to the root node, preventing it from being deleted, and tracking any 530 // changes of the root. 531 HandleSDNode Dummy(DAG.getRoot()); 532 533 // while the worklist isn't empty, inspect the node on the end of it and 534 // try and combine it. 535 while (!WorkList.empty()) { 536 SDNode *N = WorkList.back(); 537 WorkList.pop_back(); 538 539 // If N has no uses, it is dead. Make sure to revisit all N's operands once 540 // N is deleted from the DAG, since they too may now be dead or may have a 541 // reduced number of uses, allowing other xforms. 542 if (N->use_empty() && N != &Dummy) { 543 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) 544 WorkList.push_back(N->getOperand(i).Val); 545 546 removeFromWorkList(N); 547 DAG.DeleteNode(N); 548 continue; 549 } 550 551 SDOperand RV = visit(N); 552 if (RV.Val) { 553 ++NodesCombined; 554 // If we get back the same node we passed in, rather than a new node or 555 // zero, we know that the node must have defined multiple values and 556 // CombineTo was used. Since CombineTo takes care of the worklist 557 // mechanics for us, we have no work to do in this case. 558 if (RV.Val != N) { 559 DEBUG(std::cerr << "\nReplacing "; N->dump(); 560 std::cerr << "\nWith: "; RV.Val->dump(); 561 std::cerr << '\n'); 562 std::vector<SDNode*> NowDead; 563 DAG.ReplaceAllUsesWith(N, std::vector<SDOperand>(1, RV), &NowDead); 564 565 // Push the new node and any users onto the worklist 566 WorkList.push_back(RV.Val); 567 AddUsersToWorkList(RV.Val); 568 569 // Nodes can end up on the worklist more than once. Make sure we do 570 // not process a node that has been replaced. 571 removeFromWorkList(N); 572 for (unsigned i = 0, e = NowDead.size(); i != e; ++i) 573 removeFromWorkList(NowDead[i]); 574 575 // Finally, since the node is now dead, remove it from the graph. 576 DAG.DeleteNode(N); 577 } 578 } 579 } 580 581 // If the root changed (e.g. it was a dead load, update the root). 582 DAG.setRoot(Dummy.getValue()); 583} 584 585SDOperand DAGCombiner::visit(SDNode *N) { 586 switch(N->getOpcode()) { 587 default: break; 588 case ISD::TokenFactor: return visitTokenFactor(N); 589 case ISD::ADD: return visitADD(N); 590 case ISD::SUB: return visitSUB(N); 591 case ISD::MUL: return visitMUL(N); 592 case ISD::SDIV: return visitSDIV(N); 593 case ISD::UDIV: return visitUDIV(N); 594 case ISD::SREM: return visitSREM(N); 595 case ISD::UREM: return visitUREM(N); 596 case ISD::MULHU: return visitMULHU(N); 597 case ISD::MULHS: return visitMULHS(N); 598 case ISD::AND: return visitAND(N); 599 case ISD::OR: return visitOR(N); 600 case ISD::XOR: return visitXOR(N); 601 case ISD::SHL: return visitSHL(N); 602 case ISD::SRA: return visitSRA(N); 603 case ISD::SRL: return visitSRL(N); 604 case ISD::CTLZ: return visitCTLZ(N); 605 case ISD::CTTZ: return visitCTTZ(N); 606 case ISD::CTPOP: return visitCTPOP(N); 607 case ISD::SELECT: return visitSELECT(N); 608 case ISD::SELECT_CC: return visitSELECT_CC(N); 609 case ISD::SETCC: return visitSETCC(N); 610 case ISD::ADD_PARTS: return visitADD_PARTS(N); 611 case ISD::SUB_PARTS: return visitSUB_PARTS(N); 612 case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N); 613 case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N); 614 case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N); 615 case ISD::TRUNCATE: return visitTRUNCATE(N); 616 case ISD::BIT_CONVERT: return visitBIT_CONVERT(N); 617 case ISD::FADD: return visitFADD(N); 618 case ISD::FSUB: return visitFSUB(N); 619 case ISD::FMUL: return visitFMUL(N); 620 case ISD::FDIV: return visitFDIV(N); 621 case ISD::FREM: return visitFREM(N); 622 case ISD::SINT_TO_FP: return visitSINT_TO_FP(N); 623 case ISD::UINT_TO_FP: return visitUINT_TO_FP(N); 624 case ISD::FP_TO_SINT: return visitFP_TO_SINT(N); 625 case ISD::FP_TO_UINT: return visitFP_TO_UINT(N); 626 case ISD::FP_ROUND: return visitFP_ROUND(N); 627 case ISD::FP_ROUND_INREG: return visitFP_ROUND_INREG(N); 628 case ISD::FP_EXTEND: return visitFP_EXTEND(N); 629 case ISD::FNEG: return visitFNEG(N); 630 case ISD::FABS: return visitFABS(N); 631 case ISD::BRCOND: return visitBRCOND(N); 632 case ISD::BRCONDTWOWAY: return visitBRCONDTWOWAY(N); 633 case ISD::BR_CC: return visitBR_CC(N); 634 case ISD::BRTWOWAY_CC: return visitBRTWOWAY_CC(N); 635 case ISD::LOAD: return visitLOAD(N); 636 case ISD::STORE: return visitSTORE(N); 637 case ISD::LOCATION: return visitLOCATION(N); 638 case ISD::DEBUG_LOC: return visitDEBUGLOC(N); 639 } 640 return SDOperand(); 641} 642 643SDOperand DAGCombiner::visitTokenFactor(SDNode *N) { 644 std::vector<SDOperand> Ops; 645 bool Changed = false; 646 647 // If the token factor has two operands and one is the entry token, replace 648 // the token factor with the other operand. 649 if (N->getNumOperands() == 2) { 650 if (N->getOperand(0).getOpcode() == ISD::EntryToken) 651 return N->getOperand(1); 652 if (N->getOperand(1).getOpcode() == ISD::EntryToken) 653 return N->getOperand(0); 654 } 655 656 // fold (tokenfactor (tokenfactor)) -> tokenfactor 657 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) { 658 SDOperand Op = N->getOperand(i); 659 if (Op.getOpcode() == ISD::TokenFactor && Op.hasOneUse()) { 660 Changed = true; 661 for (unsigned j = 0, e = Op.getNumOperands(); j != e; ++j) 662 Ops.push_back(Op.getOperand(j)); 663 } else { 664 Ops.push_back(Op); 665 } 666 } 667 if (Changed) 668 return DAG.getNode(ISD::TokenFactor, MVT::Other, Ops); 669 return SDOperand(); 670} 671 672SDOperand DAGCombiner::visitADD(SDNode *N) { 673 SDOperand N0 = N->getOperand(0); 674 SDOperand N1 = N->getOperand(1); 675 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 676 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 677 MVT::ValueType VT = N0.getValueType(); 678 679 // fold (add c1, c2) -> c1+c2 680 if (N0C && N1C) 681 return DAG.getConstant(N0C->getValue() + N1C->getValue(), VT); 682 // canonicalize constant to RHS 683 if (N0C && !N1C) 684 return DAG.getNode(ISD::ADD, VT, N1, N0); 685 // fold (add x, 0) -> x 686 if (N1C && N1C->isNullValue()) 687 return N0; 688 // fold (add (add x, c1), c2) -> (add x, c1+c2) 689 if (N1C && N0.getOpcode() == ISD::ADD) { 690 ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0)); 691 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1)); 692 if (N00C) 693 return DAG.getNode(ISD::ADD, VT, N0.getOperand(1), 694 DAG.getConstant(N1C->getValue()+N00C->getValue(), VT)); 695 if (N01C) 696 return DAG.getNode(ISD::ADD, VT, N0.getOperand(0), 697 DAG.getConstant(N1C->getValue()+N01C->getValue(), VT)); 698 } 699 // fold ((0-A) + B) -> B-A 700 if (N0.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N0.getOperand(0)) && 701 cast<ConstantSDNode>(N0.getOperand(0))->isNullValue()) 702 return DAG.getNode(ISD::SUB, VT, N1, N0.getOperand(1)); 703 // fold (A + (0-B)) -> A-B 704 if (N1.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N1.getOperand(0)) && 705 cast<ConstantSDNode>(N1.getOperand(0))->isNullValue()) 706 return DAG.getNode(ISD::SUB, VT, N0, N1.getOperand(1)); 707 // fold (A+(B-A)) -> B 708 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1)) 709 return N1.getOperand(0); 710 return SDOperand(); 711} 712 713SDOperand DAGCombiner::visitSUB(SDNode *N) { 714 SDOperand N0 = N->getOperand(0); 715 SDOperand N1 = N->getOperand(1); 716 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val); 717 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val); 718 719 // fold (sub x, x) -> 0 720 if (N0 == N1) 721 return DAG.getConstant(0, N->getValueType(0)); 722 723 // fold (sub c1, c2) -> c1-c2 724 if (N0C && N1C) 725 return DAG.getConstant(N0C->getValue() - N1C->getValue(), 726 N->getValueType(0)); 727 // fold (sub x, c) -> (add x, -c) 728 if (N1C) 729 return DAG.getNode(ISD::ADD, N0.getValueType(), N0, 730 DAG.getConstant(-N1C->getValue(), N0.getValueType())); 731 732 // fold (A+B)-A -> B 733 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1) 734 return N0.getOperand(1); 735 // fold (A+B)-B -> A 736 if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1) 737 return N0.getOperand(0); 738 return SDOperand(); 739} 740 741SDOperand DAGCombiner::visitMUL(SDNode *N) { 742 SDOperand N0 = N->getOperand(0); 743 SDOperand N1 = N->getOperand(1); 744 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 745 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 746 MVT::ValueType VT = N0.getValueType(); 747 748 // fold (mul c1, c2) -> c1*c2 749 if (N0C && N1C) 750 return DAG.getConstant(N0C->getValue() * N1C->getValue(), VT); 751 // canonicalize constant to RHS 752 if (N0C && !N1C) 753 return DAG.getNode(ISD::MUL, VT, N1, N0); 754 // fold (mul x, 0) -> 0 755 if (N1C && N1C->isNullValue()) 756 return N1; 757 // fold (mul x, -1) -> 0-x 758 if (N1C && N1C->isAllOnesValue()) 759 return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT), N0); 760 // fold (mul x, (1 << c)) -> x << c 761 if (N1C && isPowerOf2_64(N1C->getValue())) 762 return DAG.getNode(ISD::SHL, VT, N0, 763 DAG.getConstant(Log2_64(N1C->getValue()), 764 TLI.getShiftAmountTy())); 765 // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c 766 if (N1C && isPowerOf2_64(-N1C->getSignExtended())) { 767 // FIXME: If the input is something that is easily negated (e.g. a 768 // single-use add), we should put the negate there. 769 return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT), 770 DAG.getNode(ISD::SHL, VT, N0, 771 DAG.getConstant(Log2_64(-N1C->getSignExtended()), 772 TLI.getShiftAmountTy()))); 773 } 774 775 776 // fold (mul (mul x, c1), c2) -> (mul x, c1*c2) 777 if (N1C && N0.getOpcode() == ISD::MUL) { 778 ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0)); 779 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1)); 780 if (N00C) 781 return DAG.getNode(ISD::MUL, VT, N0.getOperand(1), 782 DAG.getConstant(N1C->getValue()*N00C->getValue(), VT)); 783 if (N01C) 784 return DAG.getNode(ISD::MUL, VT, N0.getOperand(0), 785 DAG.getConstant(N1C->getValue()*N01C->getValue(), VT)); 786 } 787 return SDOperand(); 788} 789 790SDOperand DAGCombiner::visitSDIV(SDNode *N) { 791 SDOperand N0 = N->getOperand(0); 792 SDOperand N1 = N->getOperand(1); 793 MVT::ValueType VT = N->getValueType(0); 794 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val); 795 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val); 796 797 // fold (sdiv c1, c2) -> c1/c2 798 if (N0C && N1C && !N1C->isNullValue()) 799 return DAG.getConstant(N0C->getSignExtended() / N1C->getSignExtended(), 800 N->getValueType(0)); 801 // fold (sdiv X, 1) -> X 802 if (N1C && N1C->getSignExtended() == 1LL) 803 return N0; 804 // fold (sdiv X, -1) -> 0-X 805 if (N1C && N1C->isAllOnesValue()) 806 return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT), N0); 807 // If we know the sign bits of both operands are zero, strength reduce to a 808 // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2 809 uint64_t SignBit = 1ULL << (MVT::getSizeInBits(VT)-1); 810 if (MaskedValueIsZero(N1, SignBit, TLI) && 811 MaskedValueIsZero(N0, SignBit, TLI)) 812 return DAG.getNode(ISD::UDIV, N1.getValueType(), N0, N1); 813 // fold (sdiv X, pow2) -> (add (sra X, log(pow2)), (srl X, sizeof(X)-1)) 814 if (N1C && N1C->getValue() && !TLI.isIntDivCheap() && 815 (isPowerOf2_64(N1C->getSignExtended()) || 816 isPowerOf2_64(-N1C->getSignExtended()))) { 817 // If dividing by powers of two is cheap, then don't perform the following 818 // fold. 819 if (TLI.isPow2DivCheap()) 820 return SDOperand(); 821 int64_t pow2 = N1C->getSignExtended(); 822 int64_t abs2 = pow2 > 0 ? pow2 : -pow2; 823 SDOperand SRL = DAG.getNode(ISD::SRL, VT, N0, 824 DAG.getConstant(MVT::getSizeInBits(VT)-1, 825 TLI.getShiftAmountTy())); 826 WorkList.push_back(SRL.Val); 827 SDOperand SGN = DAG.getNode(ISD::ADD, VT, N0, SRL); 828 WorkList.push_back(SGN.Val); 829 SDOperand SRA = DAG.getNode(ISD::SRA, VT, SGN, 830 DAG.getConstant(Log2_64(abs2), 831 TLI.getShiftAmountTy())); 832 // If we're dividing by a positive value, we're done. Otherwise, we must 833 // negate the result. 834 if (pow2 > 0) 835 return SRA; 836 WorkList.push_back(SRA.Val); 837 return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT), SRA); 838 } 839 // if integer divide is expensive and we satisfy the requirements, emit an 840 // alternate sequence. 841 if (N1C && (N1C->getSignExtended() < -1 || N1C->getSignExtended() > 1) && 842 !TLI.isIntDivCheap()) { 843 SDOperand Op = BuildSDIV(N); 844 if (Op.Val) return Op; 845 } 846 return SDOperand(); 847} 848 849SDOperand DAGCombiner::visitUDIV(SDNode *N) { 850 SDOperand N0 = N->getOperand(0); 851 SDOperand N1 = N->getOperand(1); 852 MVT::ValueType VT = N->getValueType(0); 853 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val); 854 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val); 855 856 // fold (udiv c1, c2) -> c1/c2 857 if (N0C && N1C && !N1C->isNullValue()) 858 return DAG.getConstant(N0C->getValue() / N1C->getValue(), 859 N->getValueType(0)); 860 // fold (udiv x, (1 << c)) -> x >>u c 861 if (N1C && isPowerOf2_64(N1C->getValue())) 862 return DAG.getNode(ISD::SRL, N->getValueType(0), N0, 863 DAG.getConstant(Log2_64(N1C->getValue()), 864 TLI.getShiftAmountTy())); 865 // fold (udiv x, c) -> alternate 866 if (N1C && N1C->getValue() && !TLI.isIntDivCheap()) { 867 SDOperand Op = BuildUDIV(N); 868 if (Op.Val) return Op; 869 } 870 871 return SDOperand(); 872} 873 874SDOperand DAGCombiner::visitSREM(SDNode *N) { 875 SDOperand N0 = N->getOperand(0); 876 SDOperand N1 = N->getOperand(1); 877 MVT::ValueType VT = N->getValueType(0); 878 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 879 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 880 881 // fold (srem c1, c2) -> c1%c2 882 if (N0C && N1C && !N1C->isNullValue()) 883 return DAG.getConstant(N0C->getSignExtended() % N1C->getSignExtended(), 884 N->getValueType(0)); 885 // If we know the sign bits of both operands are zero, strength reduce to a 886 // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15 887 uint64_t SignBit = 1ULL << (MVT::getSizeInBits(VT)-1); 888 if (MaskedValueIsZero(N1, SignBit, TLI) && 889 MaskedValueIsZero(N0, SignBit, TLI)) 890 return DAG.getNode(ISD::UREM, N1.getValueType(), N0, N1); 891 return SDOperand(); 892} 893 894SDOperand DAGCombiner::visitUREM(SDNode *N) { 895 SDOperand N0 = N->getOperand(0); 896 SDOperand N1 = N->getOperand(1); 897 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 898 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 899 900 // fold (urem c1, c2) -> c1%c2 901 if (N0C && N1C && !N1C->isNullValue()) 902 return DAG.getConstant(N0C->getValue() % N1C->getValue(), 903 N->getValueType(0)); 904 // fold (urem x, pow2) -> (and x, pow2-1) 905 if (N1C && !N1C->isNullValue() && isPowerOf2_64(N1C->getValue())) 906 return DAG.getNode(ISD::AND, N0.getValueType(), N0, 907 DAG.getConstant(N1C->getValue()-1, N1.getValueType())); 908 return SDOperand(); 909} 910 911SDOperand DAGCombiner::visitMULHS(SDNode *N) { 912 SDOperand N0 = N->getOperand(0); 913 SDOperand N1 = N->getOperand(1); 914 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 915 916 // fold (mulhs x, 0) -> 0 917 if (N1C && N1C->isNullValue()) 918 return N1; 919 // fold (mulhs x, 1) -> (sra x, size(x)-1) 920 if (N1C && N1C->getValue() == 1) 921 return DAG.getNode(ISD::SRA, N0.getValueType(), N0, 922 DAG.getConstant(MVT::getSizeInBits(N0.getValueType())-1, 923 TLI.getShiftAmountTy())); 924 return SDOperand(); 925} 926 927SDOperand DAGCombiner::visitMULHU(SDNode *N) { 928 SDOperand N0 = N->getOperand(0); 929 SDOperand N1 = N->getOperand(1); 930 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 931 932 // fold (mulhu x, 0) -> 0 933 if (N1C && N1C->isNullValue()) 934 return N1; 935 // fold (mulhu x, 1) -> 0 936 if (N1C && N1C->getValue() == 1) 937 return DAG.getConstant(0, N0.getValueType()); 938 return SDOperand(); 939} 940 941SDOperand DAGCombiner::visitAND(SDNode *N) { 942 SDOperand N0 = N->getOperand(0); 943 SDOperand N1 = N->getOperand(1); 944 SDOperand LL, LR, RL, RR, CC0, CC1; 945 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 946 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 947 MVT::ValueType VT = N1.getValueType(); 948 unsigned OpSizeInBits = MVT::getSizeInBits(VT); 949 950 // fold (and c1, c2) -> c1&c2 951 if (N0C && N1C) 952 return DAG.getConstant(N0C->getValue() & N1C->getValue(), VT); 953 // canonicalize constant to RHS 954 if (N0C && !N1C) 955 return DAG.getNode(ISD::AND, VT, N1, N0); 956 // fold (and x, -1) -> x 957 if (N1C && N1C->isAllOnesValue()) 958 return N0; 959 // if (and x, c) is known to be zero, return 0 960 if (N1C && MaskedValueIsZero(SDOperand(N, 0), ~0ULL >> (64-OpSizeInBits),TLI)) 961 return DAG.getConstant(0, VT); 962 // fold (and x, c) -> x iff (x & ~c) == 0 963 if (N1C && MaskedValueIsZero(N0,~N1C->getValue() & (~0ULL>>(64-OpSizeInBits)), 964 TLI)) 965 return N0; 966 // fold (and (and x, c1), c2) -> (and x, c1^c2) 967 if (N1C && N0.getOpcode() == ISD::AND) { 968 ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0)); 969 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1)); 970 if (N00C) 971 return DAG.getNode(ISD::AND, VT, N0.getOperand(1), 972 DAG.getConstant(N1C->getValue()&N00C->getValue(), VT)); 973 if (N01C) 974 return DAG.getNode(ISD::AND, VT, N0.getOperand(0), 975 DAG.getConstant(N1C->getValue()&N01C->getValue(), VT)); 976 } 977 // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1) 978 if (N1C && N0.getOpcode() == ISD::SIGN_EXTEND_INREG) { 979 unsigned ExtendBits = 980 MVT::getSizeInBits(cast<VTSDNode>(N0.getOperand(1))->getVT()); 981 if (ExtendBits == 64 || ((N1C->getValue() & (~0ULL << ExtendBits)) == 0)) 982 return DAG.getNode(ISD::AND, VT, N0.getOperand(0), N1); 983 } 984 // fold (and (or x, 0xFFFF), 0xFF) -> 0xFF 985 if (N1C && N0.getOpcode() == ISD::OR) 986 if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1))) 987 if ((ORI->getValue() & N1C->getValue()) == N1C->getValue()) 988 return N1; 989 // fold (and (setcc x), (setcc y)) -> (setcc (and x, y)) 990 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){ 991 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get(); 992 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get(); 993 994 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 && 995 MVT::isInteger(LL.getValueType())) { 996 // fold (X == 0) & (Y == 0) -> (X|Y == 0) 997 if (cast<ConstantSDNode>(LR)->getValue() == 0 && Op1 == ISD::SETEQ) { 998 SDOperand ORNode = DAG.getNode(ISD::OR, LR.getValueType(), LL, RL); 999 WorkList.push_back(ORNode.Val); 1000 return DAG.getSetCC(VT, ORNode, LR, Op1); 1001 } 1002 // fold (X == -1) & (Y == -1) -> (X&Y == -1) 1003 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETEQ) { 1004 SDOperand ANDNode = DAG.getNode(ISD::AND, LR.getValueType(), LL, RL); 1005 WorkList.push_back(ANDNode.Val); 1006 return DAG.getSetCC(VT, ANDNode, LR, Op1); 1007 } 1008 // fold (X > -1) & (Y > -1) -> (X|Y > -1) 1009 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETGT) { 1010 SDOperand ORNode = DAG.getNode(ISD::OR, LR.getValueType(), LL, RL); 1011 WorkList.push_back(ORNode.Val); 1012 return DAG.getSetCC(VT, ORNode, LR, Op1); 1013 } 1014 } 1015 // canonicalize equivalent to ll == rl 1016 if (LL == RR && LR == RL) { 1017 Op1 = ISD::getSetCCSwappedOperands(Op1); 1018 std::swap(RL, RR); 1019 } 1020 if (LL == RL && LR == RR) { 1021 bool isInteger = MVT::isInteger(LL.getValueType()); 1022 ISD::CondCode Result = ISD::getSetCCAndOperation(Op0, Op1, isInteger); 1023 if (Result != ISD::SETCC_INVALID) 1024 return DAG.getSetCC(N0.getValueType(), LL, LR, Result); 1025 } 1026 } 1027 // fold (and (zext x), (zext y)) -> (zext (and x, y)) 1028 if (N0.getOpcode() == ISD::ZERO_EXTEND && 1029 N1.getOpcode() == ISD::ZERO_EXTEND && 1030 N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()) { 1031 SDOperand ANDNode = DAG.getNode(ISD::AND, N0.getOperand(0).getValueType(), 1032 N0.getOperand(0), N1.getOperand(0)); 1033 WorkList.push_back(ANDNode.Val); 1034 return DAG.getNode(ISD::ZERO_EXTEND, VT, ANDNode); 1035 } 1036 // fold (and (shl/srl x), (shl/srl y)) -> (shl/srl (and x, y)) 1037 if (((N0.getOpcode() == ISD::SHL && N1.getOpcode() == ISD::SHL) || 1038 (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SRL)) && 1039 N0.getOperand(1) == N1.getOperand(1)) { 1040 SDOperand ANDNode = DAG.getNode(ISD::AND, N0.getOperand(0).getValueType(), 1041 N0.getOperand(0), N1.getOperand(0)); 1042 WorkList.push_back(ANDNode.Val); 1043 return DAG.getNode(N0.getOpcode(), VT, ANDNode, N0.getOperand(1)); 1044 } 1045 // fold (and (sra)) -> (and (srl)) when possible. 1046 if (N0.getOpcode() == ISD::SRA && N0.Val->hasOneUse()) { 1047 if (ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1))) { 1048 // If the RHS of the AND has zeros where the sign bits of the SRA will 1049 // land, turn the SRA into an SRL. 1050 if (MaskedValueIsZero(N1, (~0ULL << (OpSizeInBits-N01C->getValue())) & 1051 (~0ULL>>(64-OpSizeInBits)), TLI)) { 1052 WorkList.push_back(N); 1053 CombineTo(N0.Val, DAG.getNode(ISD::SRL, VT, N0.getOperand(0), 1054 N0.getOperand(1))); 1055 return SDOperand(); 1056 } 1057 } 1058 } 1059 // fold (zext_inreg (extload x)) -> (zextload x) 1060 if (N0.getOpcode() == ISD::EXTLOAD) { 1061 MVT::ValueType EVT = cast<VTSDNode>(N0.getOperand(3))->getVT(); 1062 // If we zero all the possible extended bits, then we can turn this into 1063 // a zextload if we are running before legalize or the operation is legal. 1064 if (MaskedValueIsZero(N1, ~0ULL << MVT::getSizeInBits(EVT), TLI) && 1065 (!AfterLegalize || TLI.isOperationLegal(ISD::ZEXTLOAD, EVT))) { 1066 SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, N0.getOperand(0), 1067 N0.getOperand(1), N0.getOperand(2), 1068 EVT); 1069 WorkList.push_back(N); 1070 CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1)); 1071 return SDOperand(); 1072 } 1073 } 1074 // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use 1075 if (N0.getOpcode() == ISD::SEXTLOAD && N0.hasOneUse()) { 1076 MVT::ValueType EVT = cast<VTSDNode>(N0.getOperand(3))->getVT(); 1077 // If we zero all the possible extended bits, then we can turn this into 1078 // a zextload if we are running before legalize or the operation is legal. 1079 if (MaskedValueIsZero(N1, ~0ULL << MVT::getSizeInBits(EVT), TLI) && 1080 (!AfterLegalize || TLI.isOperationLegal(ISD::ZEXTLOAD, EVT))) { 1081 SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, N0.getOperand(0), 1082 N0.getOperand(1), N0.getOperand(2), 1083 EVT); 1084 WorkList.push_back(N); 1085 CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1)); 1086 return SDOperand(); 1087 } 1088 } 1089 return SDOperand(); 1090} 1091 1092SDOperand DAGCombiner::visitOR(SDNode *N) { 1093 SDOperand N0 = N->getOperand(0); 1094 SDOperand N1 = N->getOperand(1); 1095 SDOperand LL, LR, RL, RR, CC0, CC1; 1096 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1097 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 1098 MVT::ValueType VT = N1.getValueType(); 1099 unsigned OpSizeInBits = MVT::getSizeInBits(VT); 1100 1101 // fold (or c1, c2) -> c1|c2 1102 if (N0C && N1C) 1103 return DAG.getConstant(N0C->getValue() | N1C->getValue(), 1104 N->getValueType(0)); 1105 // canonicalize constant to RHS 1106 if (N0C && !N1C) 1107 return DAG.getNode(ISD::OR, VT, N1, N0); 1108 // fold (or x, 0) -> x 1109 if (N1C && N1C->isNullValue()) 1110 return N0; 1111 // fold (or x, -1) -> -1 1112 if (N1C && N1C->isAllOnesValue()) 1113 return N1; 1114 // fold (or x, c) -> c iff (x & ~c) == 0 1115 if (N1C && MaskedValueIsZero(N0,~N1C->getValue() & (~0ULL>>(64-OpSizeInBits)), 1116 TLI)) 1117 return N1; 1118 // fold (or (or x, c1), c2) -> (or x, c1|c2) 1119 if (N1C && N0.getOpcode() == ISD::OR) { 1120 ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0)); 1121 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1)); 1122 if (N00C) 1123 return DAG.getNode(ISD::OR, VT, N0.getOperand(1), 1124 DAG.getConstant(N1C->getValue()|N00C->getValue(), VT)); 1125 if (N01C) 1126 return DAG.getNode(ISD::OR, VT, N0.getOperand(0), 1127 DAG.getConstant(N1C->getValue()|N01C->getValue(), VT)); 1128 } else if (N1C && N0.getOpcode() == ISD::AND && N0.Val->hasOneUse() && 1129 isa<ConstantSDNode>(N0.getOperand(1))) { 1130 // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2) 1131 ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1)); 1132 return DAG.getNode(ISD::AND, VT, DAG.getNode(ISD::OR, VT, N0.getOperand(0), 1133 N1), 1134 DAG.getConstant(N1C->getValue() | C1->getValue(), VT)); 1135 } 1136 1137 1138 // fold (or (setcc x), (setcc y)) -> (setcc (or x, y)) 1139 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){ 1140 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get(); 1141 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get(); 1142 1143 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 && 1144 MVT::isInteger(LL.getValueType())) { 1145 // fold (X != 0) | (Y != 0) -> (X|Y != 0) 1146 // fold (X < 0) | (Y < 0) -> (X|Y < 0) 1147 if (cast<ConstantSDNode>(LR)->getValue() == 0 && 1148 (Op1 == ISD::SETNE || Op1 == ISD::SETLT)) { 1149 SDOperand ORNode = DAG.getNode(ISD::OR, LR.getValueType(), LL, RL); 1150 WorkList.push_back(ORNode.Val); 1151 return DAG.getSetCC(VT, ORNode, LR, Op1); 1152 } 1153 // fold (X != -1) | (Y != -1) -> (X&Y != -1) 1154 // fold (X > -1) | (Y > -1) -> (X&Y > -1) 1155 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && 1156 (Op1 == ISD::SETNE || Op1 == ISD::SETGT)) { 1157 SDOperand ANDNode = DAG.getNode(ISD::AND, LR.getValueType(), LL, RL); 1158 WorkList.push_back(ANDNode.Val); 1159 return DAG.getSetCC(VT, ANDNode, LR, Op1); 1160 } 1161 } 1162 // canonicalize equivalent to ll == rl 1163 if (LL == RR && LR == RL) { 1164 Op1 = ISD::getSetCCSwappedOperands(Op1); 1165 std::swap(RL, RR); 1166 } 1167 if (LL == RL && LR == RR) { 1168 bool isInteger = MVT::isInteger(LL.getValueType()); 1169 ISD::CondCode Result = ISD::getSetCCOrOperation(Op0, Op1, isInteger); 1170 if (Result != ISD::SETCC_INVALID) 1171 return DAG.getSetCC(N0.getValueType(), LL, LR, Result); 1172 } 1173 } 1174 // fold (or (zext x), (zext y)) -> (zext (or x, y)) 1175 if (N0.getOpcode() == ISD::ZERO_EXTEND && 1176 N1.getOpcode() == ISD::ZERO_EXTEND && 1177 N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()) { 1178 SDOperand ORNode = DAG.getNode(ISD::OR, N0.getOperand(0).getValueType(), 1179 N0.getOperand(0), N1.getOperand(0)); 1180 WorkList.push_back(ORNode.Val); 1181 return DAG.getNode(ISD::ZERO_EXTEND, VT, ORNode); 1182 } 1183 return SDOperand(); 1184} 1185 1186SDOperand DAGCombiner::visitXOR(SDNode *N) { 1187 SDOperand N0 = N->getOperand(0); 1188 SDOperand N1 = N->getOperand(1); 1189 SDOperand LHS, RHS, CC; 1190 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1191 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 1192 MVT::ValueType VT = N0.getValueType(); 1193 1194 // fold (xor c1, c2) -> c1^c2 1195 if (N0C && N1C) 1196 return DAG.getConstant(N0C->getValue() ^ N1C->getValue(), VT); 1197 // canonicalize constant to RHS 1198 if (N0C && !N1C) 1199 return DAG.getNode(ISD::XOR, VT, N1, N0); 1200 // fold (xor x, 0) -> x 1201 if (N1C && N1C->isNullValue()) 1202 return N0; 1203 // fold !(x cc y) -> (x !cc y) 1204 if (N1C && N1C->getValue() == 1 && isSetCCEquivalent(N0, LHS, RHS, CC)) { 1205 bool isInt = MVT::isInteger(LHS.getValueType()); 1206 ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(), 1207 isInt); 1208 if (N0.getOpcode() == ISD::SETCC) 1209 return DAG.getSetCC(VT, LHS, RHS, NotCC); 1210 if (N0.getOpcode() == ISD::SELECT_CC) 1211 return DAG.getSelectCC(LHS, RHS, N0.getOperand(2),N0.getOperand(3),NotCC); 1212 assert(0 && "Unhandled SetCC Equivalent!"); 1213 abort(); 1214 } 1215 // fold !(x or y) -> (!x and !y) iff x or y are setcc 1216 if (N1C && N1C->getValue() == 1 && 1217 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) { 1218 SDOperand LHS = N0.getOperand(0), RHS = N0.getOperand(1); 1219 if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) { 1220 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND; 1221 LHS = DAG.getNode(ISD::XOR, VT, LHS, N1); // RHS = ~LHS 1222 RHS = DAG.getNode(ISD::XOR, VT, RHS, N1); // RHS = ~RHS 1223 WorkList.push_back(LHS.Val); WorkList.push_back(RHS.Val); 1224 return DAG.getNode(NewOpcode, VT, LHS, RHS); 1225 } 1226 } 1227 // fold !(x or y) -> (!x and !y) iff x or y are constants 1228 if (N1C && N1C->isAllOnesValue() && 1229 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) { 1230 SDOperand LHS = N0.getOperand(0), RHS = N0.getOperand(1); 1231 if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) { 1232 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND; 1233 LHS = DAG.getNode(ISD::XOR, VT, LHS, N1); // RHS = ~LHS 1234 RHS = DAG.getNode(ISD::XOR, VT, RHS, N1); // RHS = ~RHS 1235 WorkList.push_back(LHS.Val); WorkList.push_back(RHS.Val); 1236 return DAG.getNode(NewOpcode, VT, LHS, RHS); 1237 } 1238 } 1239 // fold (xor (xor x, c1), c2) -> (xor x, c1^c2) 1240 if (N1C && N0.getOpcode() == ISD::XOR) { 1241 ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0)); 1242 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1)); 1243 if (N00C) 1244 return DAG.getNode(ISD::XOR, VT, N0.getOperand(1), 1245 DAG.getConstant(N1C->getValue()^N00C->getValue(), VT)); 1246 if (N01C) 1247 return DAG.getNode(ISD::XOR, VT, N0.getOperand(0), 1248 DAG.getConstant(N1C->getValue()^N01C->getValue(), VT)); 1249 } 1250 // fold (xor x, x) -> 0 1251 if (N0 == N1) 1252 return DAG.getConstant(0, VT); 1253 // fold (xor (zext x), (zext y)) -> (zext (xor x, y)) 1254 if (N0.getOpcode() == ISD::ZERO_EXTEND && 1255 N1.getOpcode() == ISD::ZERO_EXTEND && 1256 N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()) { 1257 SDOperand XORNode = DAG.getNode(ISD::XOR, N0.getOperand(0).getValueType(), 1258 N0.getOperand(0), N1.getOperand(0)); 1259 WorkList.push_back(XORNode.Val); 1260 return DAG.getNode(ISD::ZERO_EXTEND, VT, XORNode); 1261 } 1262 return SDOperand(); 1263} 1264 1265SDOperand DAGCombiner::visitSHL(SDNode *N) { 1266 SDOperand N0 = N->getOperand(0); 1267 SDOperand N1 = N->getOperand(1); 1268 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1269 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 1270 MVT::ValueType VT = N0.getValueType(); 1271 unsigned OpSizeInBits = MVT::getSizeInBits(VT); 1272 1273 // fold (shl c1, c2) -> c1<<c2 1274 if (N0C && N1C) 1275 return DAG.getConstant(N0C->getValue() << N1C->getValue(), VT); 1276 // fold (shl 0, x) -> 0 1277 if (N0C && N0C->isNullValue()) 1278 return N0; 1279 // fold (shl x, c >= size(x)) -> undef 1280 if (N1C && N1C->getValue() >= OpSizeInBits) 1281 return DAG.getNode(ISD::UNDEF, VT); 1282 // fold (shl x, 0) -> x 1283 if (N1C && N1C->isNullValue()) 1284 return N0; 1285 // fold (shl x, 1) -> (add x, x) 1286 if (N1C && N1C->getValue() == 1) 1287 return DAG.getNode(ISD::ADD, VT, N0, N0); 1288 // if (shl x, c) is known to be zero, return 0 1289 if (N1C && MaskedValueIsZero(SDOperand(N, 0), ~0ULL >> (64-OpSizeInBits),TLI)) 1290 return DAG.getConstant(0, VT); 1291 // fold (shl (shl x, c1), c2) -> 0 or (shl x, c1+c2) 1292 if (N1C && N0.getOpcode() == ISD::SHL && 1293 N0.getOperand(1).getOpcode() == ISD::Constant) { 1294 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getValue(); 1295 uint64_t c2 = N1C->getValue(); 1296 if (c1 + c2 > OpSizeInBits) 1297 return DAG.getConstant(0, VT); 1298 return DAG.getNode(ISD::SHL, VT, N0.getOperand(0), 1299 DAG.getConstant(c1 + c2, N1.getValueType())); 1300 } 1301 // fold (shl (srl x, c1), c2) -> (shl (and x, -1 << c1), c2-c1) or 1302 // (srl (and x, -1 << c1), c1-c2) 1303 if (N1C && N0.getOpcode() == ISD::SRL && 1304 N0.getOperand(1).getOpcode() == ISD::Constant) { 1305 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getValue(); 1306 uint64_t c2 = N1C->getValue(); 1307 SDOperand Mask = DAG.getNode(ISD::AND, VT, N0.getOperand(0), 1308 DAG.getConstant(~0ULL << c1, VT)); 1309 if (c2 > c1) 1310 return DAG.getNode(ISD::SHL, VT, Mask, 1311 DAG.getConstant(c2-c1, N1.getValueType())); 1312 else 1313 return DAG.getNode(ISD::SRL, VT, Mask, 1314 DAG.getConstant(c1-c2, N1.getValueType())); 1315 } 1316 // fold (shl (sra x, c1), c1) -> (and x, -1 << c1) 1317 if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1)) 1318 return DAG.getNode(ISD::AND, VT, N0.getOperand(0), 1319 DAG.getConstant(~0ULL << N1C->getValue(), VT)); 1320 return SDOperand(); 1321} 1322 1323SDOperand DAGCombiner::visitSRA(SDNode *N) { 1324 SDOperand N0 = N->getOperand(0); 1325 SDOperand N1 = N->getOperand(1); 1326 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1327 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 1328 MVT::ValueType VT = N0.getValueType(); 1329 unsigned OpSizeInBits = MVT::getSizeInBits(VT); 1330 1331 // fold (sra c1, c2) -> c1>>c2 1332 if (N0C && N1C) 1333 return DAG.getConstant(N0C->getSignExtended() >> N1C->getValue(), VT); 1334 // fold (sra 0, x) -> 0 1335 if (N0C && N0C->isNullValue()) 1336 return N0; 1337 // fold (sra -1, x) -> -1 1338 if (N0C && N0C->isAllOnesValue()) 1339 return N0; 1340 // fold (sra x, c >= size(x)) -> undef 1341 if (N1C && N1C->getValue() >= OpSizeInBits) 1342 return DAG.getNode(ISD::UNDEF, VT); 1343 // fold (sra x, 0) -> x 1344 if (N1C && N1C->isNullValue()) 1345 return N0; 1346 // If the sign bit is known to be zero, switch this to a SRL. 1347 if (MaskedValueIsZero(N0, (1ULL << (OpSizeInBits-1)), TLI)) 1348 return DAG.getNode(ISD::SRL, VT, N0, N1); 1349 return SDOperand(); 1350} 1351 1352SDOperand DAGCombiner::visitSRL(SDNode *N) { 1353 SDOperand N0 = N->getOperand(0); 1354 SDOperand N1 = N->getOperand(1); 1355 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1356 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 1357 MVT::ValueType VT = N0.getValueType(); 1358 unsigned OpSizeInBits = MVT::getSizeInBits(VT); 1359 1360 // fold (srl c1, c2) -> c1 >>u c2 1361 if (N0C && N1C) 1362 return DAG.getConstant(N0C->getValue() >> N1C->getValue(), VT); 1363 // fold (srl 0, x) -> 0 1364 if (N0C && N0C->isNullValue()) 1365 return N0; 1366 // fold (srl x, c >= size(x)) -> undef 1367 if (N1C && N1C->getValue() >= OpSizeInBits) 1368 return DAG.getNode(ISD::UNDEF, VT); 1369 // fold (srl x, 0) -> x 1370 if (N1C && N1C->isNullValue()) 1371 return N0; 1372 // if (srl x, c) is known to be zero, return 0 1373 if (N1C && MaskedValueIsZero(SDOperand(N, 0), ~0ULL >> (64-OpSizeInBits),TLI)) 1374 return DAG.getConstant(0, VT); 1375 // fold (srl (srl x, c1), c2) -> 0 or (srl x, c1+c2) 1376 if (N1C && N0.getOpcode() == ISD::SRL && 1377 N0.getOperand(1).getOpcode() == ISD::Constant) { 1378 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getValue(); 1379 uint64_t c2 = N1C->getValue(); 1380 if (c1 + c2 > OpSizeInBits) 1381 return DAG.getConstant(0, VT); 1382 return DAG.getNode(ISD::SRL, VT, N0.getOperand(0), 1383 DAG.getConstant(c1 + c2, N1.getValueType())); 1384 } 1385 return SDOperand(); 1386} 1387 1388SDOperand DAGCombiner::visitCTLZ(SDNode *N) { 1389 SDOperand N0 = N->getOperand(0); 1390 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1391 1392 // fold (ctlz c1) -> c2 1393 if (N0C) 1394 return DAG.getConstant(CountLeadingZeros_64(N0C->getValue()), 1395 N0.getValueType()); 1396 return SDOperand(); 1397} 1398 1399SDOperand DAGCombiner::visitCTTZ(SDNode *N) { 1400 SDOperand N0 = N->getOperand(0); 1401 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1402 1403 // fold (cttz c1) -> c2 1404 if (N0C) 1405 return DAG.getConstant(CountTrailingZeros_64(N0C->getValue()), 1406 N0.getValueType()); 1407 return SDOperand(); 1408} 1409 1410SDOperand DAGCombiner::visitCTPOP(SDNode *N) { 1411 SDOperand N0 = N->getOperand(0); 1412 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1413 1414 // fold (ctpop c1) -> c2 1415 if (N0C) 1416 return DAG.getConstant(CountPopulation_64(N0C->getValue()), 1417 N0.getValueType()); 1418 return SDOperand(); 1419} 1420 1421SDOperand DAGCombiner::visitSELECT(SDNode *N) { 1422 SDOperand N0 = N->getOperand(0); 1423 SDOperand N1 = N->getOperand(1); 1424 SDOperand N2 = N->getOperand(2); 1425 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1426 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 1427 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2); 1428 MVT::ValueType VT = N->getValueType(0); 1429 1430 // fold select C, X, X -> X 1431 if (N1 == N2) 1432 return N1; 1433 // fold select true, X, Y -> X 1434 if (N0C && !N0C->isNullValue()) 1435 return N1; 1436 // fold select false, X, Y -> Y 1437 if (N0C && N0C->isNullValue()) 1438 return N2; 1439 // fold select C, 1, X -> C | X 1440 if (MVT::i1 == VT && N1C && N1C->getValue() == 1) 1441 return DAG.getNode(ISD::OR, VT, N0, N2); 1442 // fold select C, 0, X -> ~C & X 1443 // FIXME: this should check for C type == X type, not i1? 1444 if (MVT::i1 == VT && N1C && N1C->isNullValue()) { 1445 SDOperand XORNode = DAG.getNode(ISD::XOR, VT, N0, DAG.getConstant(1, VT)); 1446 WorkList.push_back(XORNode.Val); 1447 return DAG.getNode(ISD::AND, VT, XORNode, N2); 1448 } 1449 // fold select C, X, 1 -> ~C | X 1450 if (MVT::i1 == VT && N2C && N2C->getValue() == 1) { 1451 SDOperand XORNode = DAG.getNode(ISD::XOR, VT, N0, DAG.getConstant(1, VT)); 1452 WorkList.push_back(XORNode.Val); 1453 return DAG.getNode(ISD::OR, VT, XORNode, N1); 1454 } 1455 // fold select C, X, 0 -> C & X 1456 // FIXME: this should check for C type == X type, not i1? 1457 if (MVT::i1 == VT && N2C && N2C->isNullValue()) 1458 return DAG.getNode(ISD::AND, VT, N0, N1); 1459 // fold X ? X : Y --> X ? 1 : Y --> X | Y 1460 if (MVT::i1 == VT && N0 == N1) 1461 return DAG.getNode(ISD::OR, VT, N0, N2); 1462 // fold X ? Y : X --> X ? Y : 0 --> X & Y 1463 if (MVT::i1 == VT && N0 == N2) 1464 return DAG.getNode(ISD::AND, VT, N0, N1); 1465 1466 // If we can fold this based on the true/false value, do so. 1467 if (SimplifySelectOps(N, N1, N2)) 1468 return SDOperand(); 1469 1470 // fold selects based on a setcc into other things, such as min/max/abs 1471 if (N0.getOpcode() == ISD::SETCC) 1472 return SimplifySelect(N0, N1, N2); 1473 return SDOperand(); 1474} 1475 1476SDOperand DAGCombiner::visitSELECT_CC(SDNode *N) { 1477 SDOperand N0 = N->getOperand(0); 1478 SDOperand N1 = N->getOperand(1); 1479 SDOperand N2 = N->getOperand(2); 1480 SDOperand N3 = N->getOperand(3); 1481 SDOperand N4 = N->getOperand(4); 1482 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1483 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 1484 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2); 1485 ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get(); 1486 1487 // Determine if the condition we're dealing with is constant 1488 SDOperand SCC = SimplifySetCC(TLI.getSetCCResultTy(), N0, N1, CC, false); 1489 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.Val); 1490 1491 // fold select_cc lhs, rhs, x, x, cc -> x 1492 if (N2 == N3) 1493 return N2; 1494 1495 // If we can fold this based on the true/false value, do so. 1496 if (SimplifySelectOps(N, N2, N3)) 1497 return SDOperand(); 1498 1499 // fold select_cc into other things, such as min/max/abs 1500 return SimplifySelectCC(N0, N1, N2, N3, CC); 1501} 1502 1503SDOperand DAGCombiner::visitSETCC(SDNode *N) { 1504 return SimplifySetCC(N->getValueType(0), N->getOperand(0), N->getOperand(1), 1505 cast<CondCodeSDNode>(N->getOperand(2))->get()); 1506} 1507 1508SDOperand DAGCombiner::visitADD_PARTS(SDNode *N) { 1509 SDOperand LHSLo = N->getOperand(0); 1510 SDOperand RHSLo = N->getOperand(2); 1511 MVT::ValueType VT = LHSLo.getValueType(); 1512 1513 // fold (a_Hi, 0) + (b_Hi, b_Lo) -> (b_Hi + a_Hi, b_Lo) 1514 if (MaskedValueIsZero(LHSLo, (1ULL << MVT::getSizeInBits(VT))-1, TLI)) { 1515 SDOperand Hi = DAG.getNode(ISD::ADD, VT, N->getOperand(1), 1516 N->getOperand(3)); 1517 WorkList.push_back(Hi.Val); 1518 CombineTo(N, RHSLo, Hi); 1519 return SDOperand(); 1520 } 1521 // fold (a_Hi, a_Lo) + (b_Hi, 0) -> (a_Hi + b_Hi, a_Lo) 1522 if (MaskedValueIsZero(RHSLo, (1ULL << MVT::getSizeInBits(VT))-1, TLI)) { 1523 SDOperand Hi = DAG.getNode(ISD::ADD, VT, N->getOperand(1), 1524 N->getOperand(3)); 1525 WorkList.push_back(Hi.Val); 1526 CombineTo(N, LHSLo, Hi); 1527 return SDOperand(); 1528 } 1529 return SDOperand(); 1530} 1531 1532SDOperand DAGCombiner::visitSUB_PARTS(SDNode *N) { 1533 SDOperand LHSLo = N->getOperand(0); 1534 SDOperand RHSLo = N->getOperand(2); 1535 MVT::ValueType VT = LHSLo.getValueType(); 1536 1537 // fold (a_Hi, a_Lo) - (b_Hi, 0) -> (a_Hi - b_Hi, a_Lo) 1538 if (MaskedValueIsZero(RHSLo, (1ULL << MVT::getSizeInBits(VT))-1, TLI)) { 1539 SDOperand Hi = DAG.getNode(ISD::SUB, VT, N->getOperand(1), 1540 N->getOperand(3)); 1541 WorkList.push_back(Hi.Val); 1542 CombineTo(N, LHSLo, Hi); 1543 return SDOperand(); 1544 } 1545 return SDOperand(); 1546} 1547 1548SDOperand DAGCombiner::visitSIGN_EXTEND(SDNode *N) { 1549 SDOperand N0 = N->getOperand(0); 1550 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1551 MVT::ValueType VT = N->getValueType(0); 1552 1553 // fold (sext c1) -> c1 1554 if (N0C) 1555 return DAG.getConstant(N0C->getSignExtended(), VT); 1556 // fold (sext (sext x)) -> (sext x) 1557 if (N0.getOpcode() == ISD::SIGN_EXTEND) 1558 return DAG.getNode(ISD::SIGN_EXTEND, VT, N0.getOperand(0)); 1559 // fold (sext (truncate x)) -> (sextinreg x) iff x size == sext size. 1560 if (N0.getOpcode() == ISD::TRUNCATE && N0.getOperand(0).getValueType() == VT&& 1561 (!AfterLegalize || 1562 TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, N0.getValueType()))) 1563 return DAG.getNode(ISD::SIGN_EXTEND_INREG, VT, N0.getOperand(0), 1564 DAG.getValueType(N0.getValueType())); 1565 // fold (sext (load x)) -> (sext (truncate (sextload x))) 1566 if (N0.getOpcode() == ISD::LOAD && N0.hasOneUse() && 1567 (!AfterLegalize||TLI.isOperationLegal(ISD::SEXTLOAD, N0.getValueType()))){ 1568 SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, N0.getOperand(0), 1569 N0.getOperand(1), N0.getOperand(2), 1570 N0.getValueType()); 1571 CombineTo(N, ExtLoad); 1572 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad), 1573 ExtLoad.getValue(1)); 1574 return SDOperand(); 1575 } 1576 1577 // fold (sext (sextload x)) -> (sext (truncate (sextload x))) 1578 // fold (sext ( extload x)) -> (sext (truncate (sextload x))) 1579 if ((N0.getOpcode() == ISD::SEXTLOAD || N0.getOpcode() == ISD::EXTLOAD) && 1580 N0.hasOneUse()) { 1581 SDOperand ExtLoad = DAG.getNode(ISD::SEXTLOAD, VT, N0.getOperand(0), 1582 N0.getOperand(1), N0.getOperand(2), 1583 N0.getOperand(3)); 1584 CombineTo(N, ExtLoad); 1585 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad), 1586 ExtLoad.getValue(1)); 1587 return SDOperand(); 1588 } 1589 1590 return SDOperand(); 1591} 1592 1593SDOperand DAGCombiner::visitZERO_EXTEND(SDNode *N) { 1594 SDOperand N0 = N->getOperand(0); 1595 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1596 MVT::ValueType VT = N->getValueType(0); 1597 1598 // fold (zext c1) -> c1 1599 if (N0C) 1600 return DAG.getConstant(N0C->getValue(), VT); 1601 // fold (zext (zext x)) -> (zext x) 1602 if (N0.getOpcode() == ISD::ZERO_EXTEND) 1603 return DAG.getNode(ISD::ZERO_EXTEND, VT, N0.getOperand(0)); 1604 // fold (zext (truncate x)) -> (zextinreg x) iff x size == zext size. 1605 if (N0.getOpcode() == ISD::TRUNCATE && N0.getOperand(0).getValueType() == VT&& 1606 (!AfterLegalize || TLI.isOperationLegal(ISD::AND, N0.getValueType()))) 1607 return DAG.getZeroExtendInReg(N0.getOperand(0), N0.getValueType()); 1608 // fold (zext (load x)) -> (zext (truncate (zextload x))) 1609 if (N0.getOpcode() == ISD::LOAD && N0.hasOneUse() && 1610 (!AfterLegalize||TLI.isOperationLegal(ISD::ZEXTLOAD, N0.getValueType()))){ 1611 SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, N0.getOperand(0), 1612 N0.getOperand(1), N0.getOperand(2), 1613 N0.getValueType()); 1614 CombineTo(N, ExtLoad); 1615 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad), 1616 ExtLoad.getValue(1)); 1617 return SDOperand(); 1618 } 1619 1620 // fold (zext (zextload x)) -> (zext (truncate (zextload x))) 1621 // fold (zext ( extload x)) -> (zext (truncate (zextload x))) 1622 if ((N0.getOpcode() == ISD::ZEXTLOAD || N0.getOpcode() == ISD::EXTLOAD) && 1623 N0.hasOneUse()) { 1624 SDOperand ExtLoad = DAG.getNode(ISD::ZEXTLOAD, VT, N0.getOperand(0), 1625 N0.getOperand(1), N0.getOperand(2), 1626 N0.getOperand(3)); 1627 CombineTo(N, ExtLoad); 1628 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad), 1629 ExtLoad.getValue(1)); 1630 return SDOperand(); 1631 } 1632 return SDOperand(); 1633} 1634 1635SDOperand DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) { 1636 SDOperand N0 = N->getOperand(0); 1637 SDOperand N1 = N->getOperand(1); 1638 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1639 MVT::ValueType VT = N->getValueType(0); 1640 MVT::ValueType EVT = cast<VTSDNode>(N1)->getVT(); 1641 unsigned EVTBits = MVT::getSizeInBits(EVT); 1642 1643 // fold (sext_in_reg c1) -> c1 1644 if (N0C) { 1645 SDOperand Truncate = DAG.getConstant(N0C->getValue(), EVT); 1646 return DAG.getNode(ISD::SIGN_EXTEND, VT, Truncate); 1647 } 1648 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt1 1649 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG && 1650 cast<VTSDNode>(N0.getOperand(1))->getVT() <= EVT) { 1651 return N0; 1652 } 1653 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2 1654 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG && 1655 EVT < cast<VTSDNode>(N0.getOperand(1))->getVT()) { 1656 return DAG.getNode(ISD::SIGN_EXTEND_INREG, VT, N0.getOperand(0), N1); 1657 } 1658 // fold (sext_in_reg (assert_sext x)) -> (assert_sext x) 1659 if (N0.getOpcode() == ISD::AssertSext && 1660 cast<VTSDNode>(N0.getOperand(1))->getVT() <= EVT) { 1661 return N0; 1662 } 1663 // fold (sext_in_reg (sextload x)) -> (sextload x) 1664 if (N0.getOpcode() == ISD::SEXTLOAD && 1665 cast<VTSDNode>(N0.getOperand(3))->getVT() <= EVT) { 1666 return N0; 1667 } 1668 // fold (sext_in_reg (setcc x)) -> setcc x iff (setcc x) == 0 or -1 1669 if (N0.getOpcode() == ISD::SETCC && 1670 TLI.getSetCCResultContents() == 1671 TargetLowering::ZeroOrNegativeOneSetCCResult) 1672 return N0; 1673 // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is zero 1674 if (MaskedValueIsZero(N0, 1ULL << (EVTBits-1), TLI)) 1675 return DAG.getNode(ISD::AND, N0.getValueType(), N0, 1676 DAG.getConstant(~0ULL >> (64-EVTBits), VT)); 1677 // fold (sext_in_reg (srl x)) -> sra x 1678 if (N0.getOpcode() == ISD::SRL && 1679 N0.getOperand(1).getOpcode() == ISD::Constant && 1680 cast<ConstantSDNode>(N0.getOperand(1))->getValue() == EVTBits) { 1681 return DAG.getNode(ISD::SRA, N0.getValueType(), N0.getOperand(0), 1682 N0.getOperand(1)); 1683 } 1684 // fold (sext_inreg (extload x)) -> (sextload x) 1685 if (N0.getOpcode() == ISD::EXTLOAD && 1686 EVT == cast<VTSDNode>(N0.getOperand(3))->getVT() && 1687 (!AfterLegalize || TLI.isOperationLegal(ISD::SEXTLOAD, EVT))) { 1688 SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, N0.getOperand(0), 1689 N0.getOperand(1), N0.getOperand(2), 1690 EVT); 1691 CombineTo(N, ExtLoad); 1692 CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1)); 1693 return SDOperand(); 1694 } 1695 // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use 1696 if (N0.getOpcode() == ISD::ZEXTLOAD && N0.hasOneUse() && 1697 EVT == cast<VTSDNode>(N0.getOperand(3))->getVT() && 1698 (!AfterLegalize || TLI.isOperationLegal(ISD::SEXTLOAD, EVT))) { 1699 SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, N0.getOperand(0), 1700 N0.getOperand(1), N0.getOperand(2), 1701 EVT); 1702 CombineTo(N, ExtLoad); 1703 CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1)); 1704 return SDOperand(); 1705 } 1706 return SDOperand(); 1707} 1708 1709SDOperand DAGCombiner::visitTRUNCATE(SDNode *N) { 1710 SDOperand N0 = N->getOperand(0); 1711 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1712 MVT::ValueType VT = N->getValueType(0); 1713 1714 // noop truncate 1715 if (N0.getValueType() == N->getValueType(0)) 1716 return N0; 1717 // fold (truncate c1) -> c1 1718 if (N0C) 1719 return DAG.getConstant(N0C->getValue(), VT); 1720 // fold (truncate (truncate x)) -> (truncate x) 1721 if (N0.getOpcode() == ISD::TRUNCATE) 1722 return DAG.getNode(ISD::TRUNCATE, VT, N0.getOperand(0)); 1723 // fold (truncate (ext x)) -> (ext x) or (truncate x) or x 1724 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::SIGN_EXTEND){ 1725 if (N0.getValueType() < VT) 1726 // if the source is smaller than the dest, we still need an extend 1727 return DAG.getNode(N0.getOpcode(), VT, N0.getOperand(0)); 1728 else if (N0.getValueType() > VT) 1729 // if the source is larger than the dest, than we just need the truncate 1730 return DAG.getNode(ISD::TRUNCATE, VT, N0.getOperand(0)); 1731 else 1732 // if the source and dest are the same type, we can drop both the extend 1733 // and the truncate 1734 return N0.getOperand(0); 1735 } 1736 // fold (truncate (load x)) -> (smaller load x) 1737 if (N0.getOpcode() == ISD::LOAD && N0.hasOneUse()) { 1738 assert(MVT::getSizeInBits(N0.getValueType()) > MVT::getSizeInBits(VT) && 1739 "Cannot truncate to larger type!"); 1740 MVT::ValueType PtrType = N0.getOperand(1).getValueType(); 1741 // For big endian targets, we need to add an offset to the pointer to load 1742 // the correct bytes. For little endian systems, we merely need to read 1743 // fewer bytes from the same pointer. 1744 uint64_t PtrOff = 1745 (MVT::getSizeInBits(N0.getValueType()) - MVT::getSizeInBits(VT)) / 8; 1746 SDOperand NewPtr = TLI.isLittleEndian() ? N0.getOperand(1) : 1747 DAG.getNode(ISD::ADD, PtrType, N0.getOperand(1), 1748 DAG.getConstant(PtrOff, PtrType)); 1749 WorkList.push_back(NewPtr.Val); 1750 SDOperand Load = DAG.getLoad(VT, N0.getOperand(0), NewPtr,N0.getOperand(2)); 1751 WorkList.push_back(N); 1752 CombineTo(N0.Val, Load, Load.getValue(1)); 1753 return SDOperand(); 1754 } 1755 return SDOperand(); 1756} 1757 1758SDOperand DAGCombiner::visitBIT_CONVERT(SDNode *N) { 1759 SDOperand N0 = N->getOperand(0); 1760 MVT::ValueType VT = N->getValueType(0); 1761 1762 // If the input is a constant, let getNode() fold it. 1763 if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) { 1764 SDOperand Res = DAG.getNode(ISD::BIT_CONVERT, VT, N0); 1765 if (Res.Val != N) return Res; 1766 } 1767 1768 if (N0.getOpcode() == ISD::BIT_CONVERT) // conv(conv(x,t1),t2) -> conv(x,t2) 1769 return DAG.getNode(ISD::BIT_CONVERT, VT, N0.getOperand(0)); 1770 1771 // fold (conv (load x)) -> (load (conv*)x) 1772 if (N0.getOpcode() == ISD::LOAD && N0.hasOneUse()) { 1773 SDOperand Load = DAG.getLoad(VT, N0.getOperand(0), N0.getOperand(1), 1774 N0.getOperand(2)); 1775 WorkList.push_back(N); 1776 CombineTo(N0.Val, DAG.getNode(ISD::BIT_CONVERT, N0.getValueType(), Load), 1777 Load.getValue(1)); 1778 return Load; 1779 } 1780 1781 return SDOperand(); 1782} 1783 1784SDOperand DAGCombiner::visitFADD(SDNode *N) { 1785 SDOperand N0 = N->getOperand(0); 1786 SDOperand N1 = N->getOperand(1); 1787 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 1788 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1); 1789 MVT::ValueType VT = N->getValueType(0); 1790 1791 // fold (fadd c1, c2) -> c1+c2 1792 if (N0CFP && N1CFP) 1793 return DAG.getConstantFP(N0CFP->getValue() + N1CFP->getValue(), VT); 1794 // canonicalize constant to RHS 1795 if (N0CFP && !N1CFP) 1796 return DAG.getNode(ISD::FADD, VT, N1, N0); 1797 // fold (A + (-B)) -> A-B 1798 if (N1.getOpcode() == ISD::FNEG) 1799 return DAG.getNode(ISD::FSUB, VT, N0, N1.getOperand(0)); 1800 // fold ((-A) + B) -> B-A 1801 if (N0.getOpcode() == ISD::FNEG) 1802 return DAG.getNode(ISD::FSUB, VT, N1, N0.getOperand(0)); 1803 return SDOperand(); 1804} 1805 1806SDOperand DAGCombiner::visitFSUB(SDNode *N) { 1807 SDOperand N0 = N->getOperand(0); 1808 SDOperand N1 = N->getOperand(1); 1809 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 1810 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1); 1811 MVT::ValueType VT = N->getValueType(0); 1812 1813 // fold (fsub c1, c2) -> c1-c2 1814 if (N0CFP && N1CFP) 1815 return DAG.getConstantFP(N0CFP->getValue() - N1CFP->getValue(), VT); 1816 // fold (A-(-B)) -> A+B 1817 if (N1.getOpcode() == ISD::FNEG) 1818 return DAG.getNode(ISD::FADD, N0.getValueType(), N0, N1.getOperand(0)); 1819 return SDOperand(); 1820} 1821 1822SDOperand DAGCombiner::visitFMUL(SDNode *N) { 1823 SDOperand N0 = N->getOperand(0); 1824 SDOperand N1 = N->getOperand(1); 1825 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 1826 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1); 1827 MVT::ValueType VT = N->getValueType(0); 1828 1829 // fold (fmul c1, c2) -> c1*c2 1830 if (N0CFP && N1CFP) 1831 return DAG.getConstantFP(N0CFP->getValue() * N1CFP->getValue(), VT); 1832 // canonicalize constant to RHS 1833 if (N0CFP && !N1CFP) 1834 return DAG.getNode(ISD::FMUL, VT, N1, N0); 1835 // fold (fmul X, 2.0) -> (fadd X, X) 1836 if (N1CFP && N1CFP->isExactlyValue(+2.0)) 1837 return DAG.getNode(ISD::FADD, VT, N0, N0); 1838 return SDOperand(); 1839} 1840 1841SDOperand DAGCombiner::visitFDIV(SDNode *N) { 1842 SDOperand N0 = N->getOperand(0); 1843 SDOperand N1 = N->getOperand(1); 1844 MVT::ValueType VT = N->getValueType(0); 1845 1846 if (ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0)) 1847 if (ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1)) { 1848 // fold floating point (fdiv c1, c2) 1849 return DAG.getConstantFP(N0CFP->getValue() / N1CFP->getValue(), VT); 1850 } 1851 return SDOperand(); 1852} 1853 1854SDOperand DAGCombiner::visitFREM(SDNode *N) { 1855 SDOperand N0 = N->getOperand(0); 1856 SDOperand N1 = N->getOperand(1); 1857 MVT::ValueType VT = N->getValueType(0); 1858 1859 if (ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0)) 1860 if (ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1)) { 1861 // fold floating point (frem c1, c2) -> fmod(c1, c2) 1862 return DAG.getConstantFP(fmod(N0CFP->getValue(),N1CFP->getValue()), VT); 1863 } 1864 return SDOperand(); 1865} 1866 1867 1868SDOperand DAGCombiner::visitSINT_TO_FP(SDNode *N) { 1869 SDOperand N0 = N->getOperand(0); 1870 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1871 1872 // fold (sint_to_fp c1) -> c1fp 1873 if (N0C) 1874 return DAG.getConstantFP(N0C->getSignExtended(), N->getValueType(0)); 1875 return SDOperand(); 1876} 1877 1878SDOperand DAGCombiner::visitUINT_TO_FP(SDNode *N) { 1879 SDOperand N0 = N->getOperand(0); 1880 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1881 1882 // fold (uint_to_fp c1) -> c1fp 1883 if (N0C) 1884 return DAG.getConstantFP(N0C->getValue(), N->getValueType(0)); 1885 return SDOperand(); 1886} 1887 1888SDOperand DAGCombiner::visitFP_TO_SINT(SDNode *N) { 1889 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N->getOperand(0)); 1890 1891 // fold (fp_to_sint c1fp) -> c1 1892 if (N0CFP) 1893 return DAG.getConstant((int64_t)N0CFP->getValue(), N->getValueType(0)); 1894 return SDOperand(); 1895} 1896 1897SDOperand DAGCombiner::visitFP_TO_UINT(SDNode *N) { 1898 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N->getOperand(0)); 1899 1900 // fold (fp_to_uint c1fp) -> c1 1901 if (N0CFP) 1902 return DAG.getConstant((uint64_t)N0CFP->getValue(), N->getValueType(0)); 1903 return SDOperand(); 1904} 1905 1906SDOperand DAGCombiner::visitFP_ROUND(SDNode *N) { 1907 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N->getOperand(0)); 1908 1909 // fold (fp_round c1fp) -> c1fp 1910 if (N0CFP) 1911 return DAG.getConstantFP(N0CFP->getValue(), N->getValueType(0)); 1912 return SDOperand(); 1913} 1914 1915SDOperand DAGCombiner::visitFP_ROUND_INREG(SDNode *N) { 1916 SDOperand N0 = N->getOperand(0); 1917 MVT::ValueType VT = N->getValueType(0); 1918 MVT::ValueType EVT = cast<VTSDNode>(N->getOperand(1))->getVT(); 1919 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 1920 1921 // fold (fp_round_inreg c1fp) -> c1fp 1922 if (N0CFP) { 1923 SDOperand Round = DAG.getConstantFP(N0CFP->getValue(), EVT); 1924 return DAG.getNode(ISD::FP_EXTEND, VT, Round); 1925 } 1926 return SDOperand(); 1927} 1928 1929SDOperand DAGCombiner::visitFP_EXTEND(SDNode *N) { 1930 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N->getOperand(0)); 1931 1932 // fold (fp_extend c1fp) -> c1fp 1933 if (N0CFP) 1934 return DAG.getConstantFP(N0CFP->getValue(), N->getValueType(0)); 1935 return SDOperand(); 1936} 1937 1938SDOperand DAGCombiner::visitFNEG(SDNode *N) { 1939 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N->getOperand(0)); 1940 // fold (neg c1) -> -c1 1941 if (N0CFP) 1942 return DAG.getConstantFP(-N0CFP->getValue(), N->getValueType(0)); 1943 // fold (neg (sub x, y)) -> (sub y, x) 1944 if (N->getOperand(0).getOpcode() == ISD::SUB) 1945 return DAG.getNode(ISD::SUB, N->getValueType(0), N->getOperand(1), 1946 N->getOperand(0)); 1947 // fold (neg (neg x)) -> x 1948 if (N->getOperand(0).getOpcode() == ISD::FNEG) 1949 return N->getOperand(0).getOperand(0); 1950 return SDOperand(); 1951} 1952 1953SDOperand DAGCombiner::visitFABS(SDNode *N) { 1954 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N->getOperand(0)); 1955 // fold (fabs c1) -> fabs(c1) 1956 if (N0CFP) 1957 return DAG.getConstantFP(fabs(N0CFP->getValue()), N->getValueType(0)); 1958 // fold (fabs (fabs x)) -> (fabs x) 1959 if (N->getOperand(0).getOpcode() == ISD::FABS) 1960 return N->getOperand(0); 1961 // fold (fabs (fneg x)) -> (fabs x) 1962 if (N->getOperand(0).getOpcode() == ISD::FNEG) 1963 return DAG.getNode(ISD::FABS, N->getValueType(0), 1964 N->getOperand(0).getOperand(0)); 1965 return SDOperand(); 1966} 1967 1968SDOperand DAGCombiner::visitBRCOND(SDNode *N) { 1969 SDOperand Chain = N->getOperand(0); 1970 SDOperand N1 = N->getOperand(1); 1971 SDOperand N2 = N->getOperand(2); 1972 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 1973 1974 // never taken branch, fold to chain 1975 if (N1C && N1C->isNullValue()) 1976 return Chain; 1977 // unconditional branch 1978 if (N1C && N1C->getValue() == 1) 1979 return DAG.getNode(ISD::BR, MVT::Other, Chain, N2); 1980 return SDOperand(); 1981} 1982 1983SDOperand DAGCombiner::visitBRCONDTWOWAY(SDNode *N) { 1984 SDOperand Chain = N->getOperand(0); 1985 SDOperand N1 = N->getOperand(1); 1986 SDOperand N2 = N->getOperand(2); 1987 SDOperand N3 = N->getOperand(3); 1988 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 1989 1990 // unconditional branch to true mbb 1991 if (N1C && N1C->getValue() == 1) 1992 return DAG.getNode(ISD::BR, MVT::Other, Chain, N2); 1993 // unconditional branch to false mbb 1994 if (N1C && N1C->isNullValue()) 1995 return DAG.getNode(ISD::BR, MVT::Other, Chain, N3); 1996 return SDOperand(); 1997} 1998 1999// Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB. 2000// 2001SDOperand DAGCombiner::visitBR_CC(SDNode *N) { 2002 CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1)); 2003 SDOperand CondLHS = N->getOperand(2), CondRHS = N->getOperand(3); 2004 2005 // Use SimplifySetCC to simplify SETCC's. 2006 SDOperand Simp = SimplifySetCC(MVT::i1, CondLHS, CondRHS, CC->get(), false); 2007 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(Simp.Val); 2008 2009 // fold br_cc true, dest -> br dest (unconditional branch) 2010 if (SCCC && SCCC->getValue()) 2011 return DAG.getNode(ISD::BR, MVT::Other, N->getOperand(0), 2012 N->getOperand(4)); 2013 // fold br_cc false, dest -> unconditional fall through 2014 if (SCCC && SCCC->isNullValue()) 2015 return N->getOperand(0); 2016 // fold to a simpler setcc 2017 if (Simp.Val && Simp.getOpcode() == ISD::SETCC) 2018 return DAG.getNode(ISD::BR_CC, MVT::Other, N->getOperand(0), 2019 Simp.getOperand(2), Simp.getOperand(0), 2020 Simp.getOperand(1), N->getOperand(4)); 2021 return SDOperand(); 2022} 2023 2024SDOperand DAGCombiner::visitBRTWOWAY_CC(SDNode *N) { 2025 SDOperand Chain = N->getOperand(0); 2026 SDOperand CCN = N->getOperand(1); 2027 SDOperand LHS = N->getOperand(2); 2028 SDOperand RHS = N->getOperand(3); 2029 SDOperand N4 = N->getOperand(4); 2030 SDOperand N5 = N->getOperand(5); 2031 2032 SDOperand SCC = SimplifySetCC(TLI.getSetCCResultTy(), LHS, RHS, 2033 cast<CondCodeSDNode>(CCN)->get(), false); 2034 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.Val); 2035 2036 // fold select_cc lhs, rhs, x, x, cc -> x 2037 if (N4 == N5) 2038 return DAG.getNode(ISD::BR, MVT::Other, Chain, N4); 2039 // fold select_cc true, x, y -> x 2040 if (SCCC && SCCC->getValue()) 2041 return DAG.getNode(ISD::BR, MVT::Other, Chain, N4); 2042 // fold select_cc false, x, y -> y 2043 if (SCCC && SCCC->isNullValue()) 2044 return DAG.getNode(ISD::BR, MVT::Other, Chain, N5); 2045 // fold to a simpler setcc 2046 if (SCC.Val && SCC.getOpcode() == ISD::SETCC) 2047 return DAG.getBR2Way_CC(Chain, SCC.getOperand(2), SCC.getOperand(0), 2048 SCC.getOperand(1), N4, N5); 2049 return SDOperand(); 2050} 2051 2052SDOperand DAGCombiner::visitLOAD(SDNode *N) { 2053 SDOperand Chain = N->getOperand(0); 2054 SDOperand Ptr = N->getOperand(1); 2055 SDOperand SrcValue = N->getOperand(2); 2056 2057 // If this load is directly stored, replace the load value with the stored 2058 // value. 2059 // TODO: Handle store large -> read small portion. 2060 // TODO: Handle TRUNCSTORE/EXTLOAD 2061 if (Chain.getOpcode() == ISD::STORE && Chain.getOperand(2) == Ptr && 2062 Chain.getOperand(1).getValueType() == N->getValueType(0)) 2063 return CombineTo(N, Chain.getOperand(1), Chain); 2064 2065 return SDOperand(); 2066} 2067 2068SDOperand DAGCombiner::visitSTORE(SDNode *N) { 2069 SDOperand Chain = N->getOperand(0); 2070 SDOperand Value = N->getOperand(1); 2071 SDOperand Ptr = N->getOperand(2); 2072 SDOperand SrcValue = N->getOperand(3); 2073 2074 // If this is a store that kills a previous store, remove the previous store. 2075 if (Chain.getOpcode() == ISD::STORE && Chain.getOperand(2) == Ptr && 2076 Chain.Val->hasOneUse() /* Avoid introducing DAG cycles */ && 2077 // Make sure that these stores are the same value type: 2078 // FIXME: we really care that the second store is >= size of the first. 2079 Value.getValueType() == Chain.getOperand(1).getValueType()) { 2080 // Create a new store of Value that replaces both stores. 2081 SDNode *PrevStore = Chain.Val; 2082 if (PrevStore->getOperand(1) == Value) // Same value multiply stored. 2083 return Chain; 2084 SDOperand NewStore = DAG.getNode(ISD::STORE, MVT::Other, 2085 PrevStore->getOperand(0), Value, Ptr, 2086 SrcValue); 2087 CombineTo(N, NewStore); // Nuke this store. 2088 CombineTo(PrevStore, NewStore); // Nuke the previous store. 2089 return SDOperand(N, 0); 2090 } 2091 2092 // If this is a store of a bit convert, store the input value. 2093 if (Value.getOpcode() == ISD::BIT_CONVERT) 2094 return DAG.getNode(ISD::STORE, MVT::Other, Chain, Value.getOperand(0), 2095 Ptr, SrcValue); 2096 2097 return SDOperand(); 2098} 2099 2100SDOperand DAGCombiner::visitLOCATION(SDNode *N) { 2101 SDOperand Chain = N->getOperand(0); 2102 2103 // Remove redundant locations (last one holds) 2104 if (Chain.getOpcode() == ISD::LOCATION && Chain.hasOneUse()) { 2105 return DAG.getNode(ISD::LOCATION, MVT::Other, Chain.getOperand(0), 2106 N->getOperand(1), 2107 N->getOperand(2), 2108 N->getOperand(3), 2109 N->getOperand(4)); 2110 } 2111 2112 return SDOperand(); 2113} 2114 2115SDOperand DAGCombiner::visitDEBUGLOC(SDNode *N) { 2116 SDOperand Chain = N->getOperand(0); 2117 2118 // Remove redundant debug locations (last one holds) 2119 if (Chain.getOpcode() == ISD::DEBUG_LOC && Chain.hasOneUse()) { 2120 return DAG.getNode(ISD::DEBUG_LOC, MVT::Other, Chain.getOperand(0), 2121 N->getOperand(1), 2122 N->getOperand(2), 2123 N->getOperand(3)); 2124 } 2125 2126 return SDOperand(); 2127} 2128 2129SDOperand DAGCombiner::SimplifySelect(SDOperand N0, SDOperand N1, SDOperand N2){ 2130 assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!"); 2131 2132 SDOperand SCC = SimplifySelectCC(N0.getOperand(0), N0.getOperand(1), N1, N2, 2133 cast<CondCodeSDNode>(N0.getOperand(2))->get()); 2134 // If we got a simplified select_cc node back from SimplifySelectCC, then 2135 // break it down into a new SETCC node, and a new SELECT node, and then return 2136 // the SELECT node, since we were called with a SELECT node. 2137 if (SCC.Val) { 2138 // Check to see if we got a select_cc back (to turn into setcc/select). 2139 // Otherwise, just return whatever node we got back, like fabs. 2140 if (SCC.getOpcode() == ISD::SELECT_CC) { 2141 SDOperand SETCC = DAG.getNode(ISD::SETCC, N0.getValueType(), 2142 SCC.getOperand(0), SCC.getOperand(1), 2143 SCC.getOperand(4)); 2144 WorkList.push_back(SETCC.Val); 2145 return DAG.getNode(ISD::SELECT, SCC.getValueType(), SCC.getOperand(2), 2146 SCC.getOperand(3), SETCC); 2147 } 2148 return SCC; 2149 } 2150 return SDOperand(); 2151} 2152 2153/// SimplifySelectOps - Given a SELECT or a SELECT_CC node, where LHS and RHS 2154/// are the two values being selected between, see if we can simplify the 2155/// select. 2156/// 2157bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDOperand LHS, 2158 SDOperand RHS) { 2159 2160 // If this is a select from two identical things, try to pull the operation 2161 // through the select. 2162 if (LHS.getOpcode() == RHS.getOpcode() && LHS.hasOneUse() && RHS.hasOneUse()){ 2163#if 0 2164 std::cerr << "SELECT: ["; LHS.Val->dump(); 2165 std::cerr << "] ["; RHS.Val->dump(); 2166 std::cerr << "]\n"; 2167#endif 2168 2169 // If this is a load and the token chain is identical, replace the select 2170 // of two loads with a load through a select of the address to load from. 2171 // This triggers in things like "select bool X, 10.0, 123.0" after the FP 2172 // constants have been dropped into the constant pool. 2173 if ((LHS.getOpcode() == ISD::LOAD || 2174 LHS.getOpcode() == ISD::EXTLOAD || 2175 LHS.getOpcode() == ISD::ZEXTLOAD || 2176 LHS.getOpcode() == ISD::SEXTLOAD) && 2177 // Token chains must be identical. 2178 LHS.getOperand(0) == RHS.getOperand(0) && 2179 // If this is an EXTLOAD, the VT's must match. 2180 (LHS.getOpcode() == ISD::LOAD || 2181 LHS.getOperand(3) == RHS.getOperand(3))) { 2182 // FIXME: this conflates two src values, discarding one. This is not 2183 // the right thing to do, but nothing uses srcvalues now. When they do, 2184 // turn SrcValue into a list of locations. 2185 SDOperand Addr; 2186 if (TheSelect->getOpcode() == ISD::SELECT) 2187 Addr = DAG.getNode(ISD::SELECT, LHS.getOperand(1).getValueType(), 2188 TheSelect->getOperand(0), LHS.getOperand(1), 2189 RHS.getOperand(1)); 2190 else 2191 Addr = DAG.getNode(ISD::SELECT_CC, LHS.getOperand(1).getValueType(), 2192 TheSelect->getOperand(0), 2193 TheSelect->getOperand(1), 2194 LHS.getOperand(1), RHS.getOperand(1), 2195 TheSelect->getOperand(4)); 2196 2197 SDOperand Load; 2198 if (LHS.getOpcode() == ISD::LOAD) 2199 Load = DAG.getLoad(TheSelect->getValueType(0), LHS.getOperand(0), 2200 Addr, LHS.getOperand(2)); 2201 else 2202 Load = DAG.getExtLoad(LHS.getOpcode(), TheSelect->getValueType(0), 2203 LHS.getOperand(0), Addr, LHS.getOperand(2), 2204 cast<VTSDNode>(LHS.getOperand(3))->getVT()); 2205 // Users of the select now use the result of the load. 2206 CombineTo(TheSelect, Load); 2207 2208 // Users of the old loads now use the new load's chain. We know the 2209 // old-load value is dead now. 2210 CombineTo(LHS.Val, Load.getValue(0), Load.getValue(1)); 2211 CombineTo(RHS.Val, Load.getValue(0), Load.getValue(1)); 2212 return true; 2213 } 2214 } 2215 2216 return false; 2217} 2218 2219SDOperand DAGCombiner::SimplifySelectCC(SDOperand N0, SDOperand N1, 2220 SDOperand N2, SDOperand N3, 2221 ISD::CondCode CC) { 2222 2223 MVT::ValueType VT = N2.getValueType(); 2224 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val); 2225 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val); 2226 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.Val); 2227 ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N3.Val); 2228 2229 // Determine if the condition we're dealing with is constant 2230 SDOperand SCC = SimplifySetCC(TLI.getSetCCResultTy(), N0, N1, CC, false); 2231 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.Val); 2232 2233 // fold select_cc true, x, y -> x 2234 if (SCCC && SCCC->getValue()) 2235 return N2; 2236 // fold select_cc false, x, y -> y 2237 if (SCCC && SCCC->getValue() == 0) 2238 return N3; 2239 2240 // Check to see if we can simplify the select into an fabs node 2241 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) { 2242 // Allow either -0.0 or 0.0 2243 if (CFP->getValue() == 0.0) { 2244 // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs 2245 if ((CC == ISD::SETGE || CC == ISD::SETGT) && 2246 N0 == N2 && N3.getOpcode() == ISD::FNEG && 2247 N2 == N3.getOperand(0)) 2248 return DAG.getNode(ISD::FABS, VT, N0); 2249 2250 // select (setl[te] X, +/-0.0), fneg(X), X -> fabs 2251 if ((CC == ISD::SETLT || CC == ISD::SETLE) && 2252 N0 == N3 && N2.getOpcode() == ISD::FNEG && 2253 N2.getOperand(0) == N3) 2254 return DAG.getNode(ISD::FABS, VT, N3); 2255 } 2256 } 2257 2258 // Check to see if we can perform the "gzip trick", transforming 2259 // select_cc setlt X, 0, A, 0 -> and (sra X, size(X)-1), A 2260 if (N1C && N1C->isNullValue() && N3C && N3C->isNullValue() && 2261 MVT::isInteger(N0.getValueType()) && 2262 MVT::isInteger(N2.getValueType()) && CC == ISD::SETLT) { 2263 MVT::ValueType XType = N0.getValueType(); 2264 MVT::ValueType AType = N2.getValueType(); 2265 if (XType >= AType) { 2266 // and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a 2267 // single-bit constant. 2268 if (N2C && ((N2C->getValue() & (N2C->getValue()-1)) == 0)) { 2269 unsigned ShCtV = Log2_64(N2C->getValue()); 2270 ShCtV = MVT::getSizeInBits(XType)-ShCtV-1; 2271 SDOperand ShCt = DAG.getConstant(ShCtV, TLI.getShiftAmountTy()); 2272 SDOperand Shift = DAG.getNode(ISD::SRL, XType, N0, ShCt); 2273 WorkList.push_back(Shift.Val); 2274 if (XType > AType) { 2275 Shift = DAG.getNode(ISD::TRUNCATE, AType, Shift); 2276 WorkList.push_back(Shift.Val); 2277 } 2278 return DAG.getNode(ISD::AND, AType, Shift, N2); 2279 } 2280 SDOperand Shift = DAG.getNode(ISD::SRA, XType, N0, 2281 DAG.getConstant(MVT::getSizeInBits(XType)-1, 2282 TLI.getShiftAmountTy())); 2283 WorkList.push_back(Shift.Val); 2284 if (XType > AType) { 2285 Shift = DAG.getNode(ISD::TRUNCATE, AType, Shift); 2286 WorkList.push_back(Shift.Val); 2287 } 2288 return DAG.getNode(ISD::AND, AType, Shift, N2); 2289 } 2290 } 2291 2292 // fold select C, 16, 0 -> shl C, 4 2293 if (N2C && N3C && N3C->isNullValue() && isPowerOf2_64(N2C->getValue()) && 2294 TLI.getSetCCResultContents() == TargetLowering::ZeroOrOneSetCCResult) { 2295 // Get a SetCC of the condition 2296 // FIXME: Should probably make sure that setcc is legal if we ever have a 2297 // target where it isn't. 2298 SDOperand Temp, SCC = DAG.getSetCC(TLI.getSetCCResultTy(), N0, N1, CC); 2299 WorkList.push_back(SCC.Val); 2300 // cast from setcc result type to select result type 2301 if (AfterLegalize) 2302 Temp = DAG.getZeroExtendInReg(SCC, N2.getValueType()); 2303 else 2304 Temp = DAG.getNode(ISD::ZERO_EXTEND, N2.getValueType(), SCC); 2305 WorkList.push_back(Temp.Val); 2306 // shl setcc result by log2 n2c 2307 return DAG.getNode(ISD::SHL, N2.getValueType(), Temp, 2308 DAG.getConstant(Log2_64(N2C->getValue()), 2309 TLI.getShiftAmountTy())); 2310 } 2311 2312 // Check to see if this is the equivalent of setcc 2313 // FIXME: Turn all of these into setcc if setcc if setcc is legal 2314 // otherwise, go ahead with the folds. 2315 if (0 && N3C && N3C->isNullValue() && N2C && (N2C->getValue() == 1ULL)) { 2316 MVT::ValueType XType = N0.getValueType(); 2317 if (TLI.isOperationLegal(ISD::SETCC, TLI.getSetCCResultTy())) { 2318 SDOperand Res = DAG.getSetCC(TLI.getSetCCResultTy(), N0, N1, CC); 2319 if (Res.getValueType() != VT) 2320 Res = DAG.getNode(ISD::ZERO_EXTEND, VT, Res); 2321 return Res; 2322 } 2323 2324 // seteq X, 0 -> srl (ctlz X, log2(size(X))) 2325 if (N1C && N1C->isNullValue() && CC == ISD::SETEQ && 2326 TLI.isOperationLegal(ISD::CTLZ, XType)) { 2327 SDOperand Ctlz = DAG.getNode(ISD::CTLZ, XType, N0); 2328 return DAG.getNode(ISD::SRL, XType, Ctlz, 2329 DAG.getConstant(Log2_32(MVT::getSizeInBits(XType)), 2330 TLI.getShiftAmountTy())); 2331 } 2332 // setgt X, 0 -> srl (and (-X, ~X), size(X)-1) 2333 if (N1C && N1C->isNullValue() && CC == ISD::SETGT) { 2334 SDOperand NegN0 = DAG.getNode(ISD::SUB, XType, DAG.getConstant(0, XType), 2335 N0); 2336 SDOperand NotN0 = DAG.getNode(ISD::XOR, XType, N0, 2337 DAG.getConstant(~0ULL, XType)); 2338 return DAG.getNode(ISD::SRL, XType, 2339 DAG.getNode(ISD::AND, XType, NegN0, NotN0), 2340 DAG.getConstant(MVT::getSizeInBits(XType)-1, 2341 TLI.getShiftAmountTy())); 2342 } 2343 // setgt X, -1 -> xor (srl (X, size(X)-1), 1) 2344 if (N1C && N1C->isAllOnesValue() && CC == ISD::SETGT) { 2345 SDOperand Sign = DAG.getNode(ISD::SRL, XType, N0, 2346 DAG.getConstant(MVT::getSizeInBits(XType)-1, 2347 TLI.getShiftAmountTy())); 2348 return DAG.getNode(ISD::XOR, XType, Sign, DAG.getConstant(1, XType)); 2349 } 2350 } 2351 2352 // Check to see if this is an integer abs. select_cc setl[te] X, 0, -X, X -> 2353 // Y = sra (X, size(X)-1); xor (add (X, Y), Y) 2354 if (N1C && N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE) && 2355 N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1)) { 2356 if (ConstantSDNode *SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0))) { 2357 MVT::ValueType XType = N0.getValueType(); 2358 if (SubC->isNullValue() && MVT::isInteger(XType)) { 2359 SDOperand Shift = DAG.getNode(ISD::SRA, XType, N0, 2360 DAG.getConstant(MVT::getSizeInBits(XType)-1, 2361 TLI.getShiftAmountTy())); 2362 SDOperand Add = DAG.getNode(ISD::ADD, XType, N0, Shift); 2363 WorkList.push_back(Shift.Val); 2364 WorkList.push_back(Add.Val); 2365 return DAG.getNode(ISD::XOR, XType, Add, Shift); 2366 } 2367 } 2368 } 2369 2370 return SDOperand(); 2371} 2372 2373SDOperand DAGCombiner::SimplifySetCC(MVT::ValueType VT, SDOperand N0, 2374 SDOperand N1, ISD::CondCode Cond, 2375 bool foldBooleans) { 2376 // These setcc operations always fold. 2377 switch (Cond) { 2378 default: break; 2379 case ISD::SETFALSE: 2380 case ISD::SETFALSE2: return DAG.getConstant(0, VT); 2381 case ISD::SETTRUE: 2382 case ISD::SETTRUE2: return DAG.getConstant(1, VT); 2383 } 2384 2385 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val)) { 2386 uint64_t C1 = N1C->getValue(); 2387 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val)) { 2388 uint64_t C0 = N0C->getValue(); 2389 2390 // Sign extend the operands if required 2391 if (ISD::isSignedIntSetCC(Cond)) { 2392 C0 = N0C->getSignExtended(); 2393 C1 = N1C->getSignExtended(); 2394 } 2395 2396 switch (Cond) { 2397 default: assert(0 && "Unknown integer setcc!"); 2398 case ISD::SETEQ: return DAG.getConstant(C0 == C1, VT); 2399 case ISD::SETNE: return DAG.getConstant(C0 != C1, VT); 2400 case ISD::SETULT: return DAG.getConstant(C0 < C1, VT); 2401 case ISD::SETUGT: return DAG.getConstant(C0 > C1, VT); 2402 case ISD::SETULE: return DAG.getConstant(C0 <= C1, VT); 2403 case ISD::SETUGE: return DAG.getConstant(C0 >= C1, VT); 2404 case ISD::SETLT: return DAG.getConstant((int64_t)C0 < (int64_t)C1, VT); 2405 case ISD::SETGT: return DAG.getConstant((int64_t)C0 > (int64_t)C1, VT); 2406 case ISD::SETLE: return DAG.getConstant((int64_t)C0 <= (int64_t)C1, VT); 2407 case ISD::SETGE: return DAG.getConstant((int64_t)C0 >= (int64_t)C1, VT); 2408 } 2409 } else { 2410 // If the LHS is a ZERO_EXTEND, perform the comparison on the input. 2411 if (N0.getOpcode() == ISD::ZERO_EXTEND) { 2412 unsigned InSize = MVT::getSizeInBits(N0.getOperand(0).getValueType()); 2413 2414 // If the comparison constant has bits in the upper part, the 2415 // zero-extended value could never match. 2416 if (C1 & (~0ULL << InSize)) { 2417 unsigned VSize = MVT::getSizeInBits(N0.getValueType()); 2418 switch (Cond) { 2419 case ISD::SETUGT: 2420 case ISD::SETUGE: 2421 case ISD::SETEQ: return DAG.getConstant(0, VT); 2422 case ISD::SETULT: 2423 case ISD::SETULE: 2424 case ISD::SETNE: return DAG.getConstant(1, VT); 2425 case ISD::SETGT: 2426 case ISD::SETGE: 2427 // True if the sign bit of C1 is set. 2428 return DAG.getConstant((C1 & (1ULL << VSize)) != 0, VT); 2429 case ISD::SETLT: 2430 case ISD::SETLE: 2431 // True if the sign bit of C1 isn't set. 2432 return DAG.getConstant((C1 & (1ULL << VSize)) == 0, VT); 2433 default: 2434 break; 2435 } 2436 } 2437 2438 // Otherwise, we can perform the comparison with the low bits. 2439 switch (Cond) { 2440 case ISD::SETEQ: 2441 case ISD::SETNE: 2442 case ISD::SETUGT: 2443 case ISD::SETUGE: 2444 case ISD::SETULT: 2445 case ISD::SETULE: 2446 return DAG.getSetCC(VT, N0.getOperand(0), 2447 DAG.getConstant(C1, N0.getOperand(0).getValueType()), 2448 Cond); 2449 default: 2450 break; // todo, be more careful with signed comparisons 2451 } 2452 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG && 2453 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) { 2454 MVT::ValueType ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT(); 2455 unsigned ExtSrcTyBits = MVT::getSizeInBits(ExtSrcTy); 2456 MVT::ValueType ExtDstTy = N0.getValueType(); 2457 unsigned ExtDstTyBits = MVT::getSizeInBits(ExtDstTy); 2458 2459 // If the extended part has any inconsistent bits, it cannot ever 2460 // compare equal. In other words, they have to be all ones or all 2461 // zeros. 2462 uint64_t ExtBits = 2463 (~0ULL >> (64-ExtSrcTyBits)) & (~0ULL << (ExtDstTyBits-1)); 2464 if ((C1 & ExtBits) != 0 && (C1 & ExtBits) != ExtBits) 2465 return DAG.getConstant(Cond == ISD::SETNE, VT); 2466 2467 SDOperand ZextOp; 2468 MVT::ValueType Op0Ty = N0.getOperand(0).getValueType(); 2469 if (Op0Ty == ExtSrcTy) { 2470 ZextOp = N0.getOperand(0); 2471 } else { 2472 int64_t Imm = ~0ULL >> (64-ExtSrcTyBits); 2473 ZextOp = DAG.getNode(ISD::AND, Op0Ty, N0.getOperand(0), 2474 DAG.getConstant(Imm, Op0Ty)); 2475 } 2476 WorkList.push_back(ZextOp.Val); 2477 // Otherwise, make this a use of a zext. 2478 return DAG.getSetCC(VT, ZextOp, 2479 DAG.getConstant(C1 & (~0ULL>>(64-ExtSrcTyBits)), 2480 ExtDstTy), 2481 Cond); 2482 } 2483 2484 uint64_t MinVal, MaxVal; 2485 unsigned OperandBitSize = MVT::getSizeInBits(N1C->getValueType(0)); 2486 if (ISD::isSignedIntSetCC(Cond)) { 2487 MinVal = 1ULL << (OperandBitSize-1); 2488 if (OperandBitSize != 1) // Avoid X >> 64, which is undefined. 2489 MaxVal = ~0ULL >> (65-OperandBitSize); 2490 else 2491 MaxVal = 0; 2492 } else { 2493 MinVal = 0; 2494 MaxVal = ~0ULL >> (64-OperandBitSize); 2495 } 2496 2497 // Canonicalize GE/LE comparisons to use GT/LT comparisons. 2498 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) { 2499 if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true 2500 --C1; // X >= C0 --> X > (C0-1) 2501 return DAG.getSetCC(VT, N0, DAG.getConstant(C1, N1.getValueType()), 2502 (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT); 2503 } 2504 2505 if (Cond == ISD::SETLE || Cond == ISD::SETULE) { 2506 if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true 2507 ++C1; // X <= C0 --> X < (C0+1) 2508 return DAG.getSetCC(VT, N0, DAG.getConstant(C1, N1.getValueType()), 2509 (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT); 2510 } 2511 2512 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal) 2513 return DAG.getConstant(0, VT); // X < MIN --> false 2514 2515 // Canonicalize setgt X, Min --> setne X, Min 2516 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal) 2517 return DAG.getSetCC(VT, N0, N1, ISD::SETNE); 2518 // Canonicalize setlt X, Max --> setne X, Max 2519 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal) 2520 return DAG.getSetCC(VT, N0, N1, ISD::SETNE); 2521 2522 // If we have setult X, 1, turn it into seteq X, 0 2523 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1) 2524 return DAG.getSetCC(VT, N0, DAG.getConstant(MinVal, N0.getValueType()), 2525 ISD::SETEQ); 2526 // If we have setugt X, Max-1, turn it into seteq X, Max 2527 else if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1) 2528 return DAG.getSetCC(VT, N0, DAG.getConstant(MaxVal, N0.getValueType()), 2529 ISD::SETEQ); 2530 2531 // If we have "setcc X, C0", check to see if we can shrink the immediate 2532 // by changing cc. 2533 2534 // SETUGT X, SINTMAX -> SETLT X, 0 2535 if (Cond == ISD::SETUGT && OperandBitSize != 1 && 2536 C1 == (~0ULL >> (65-OperandBitSize))) 2537 return DAG.getSetCC(VT, N0, DAG.getConstant(0, N1.getValueType()), 2538 ISD::SETLT); 2539 2540 // FIXME: Implement the rest of these. 2541 2542 // Fold bit comparisons when we can. 2543 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && 2544 VT == N0.getValueType() && N0.getOpcode() == ISD::AND) 2545 if (ConstantSDNode *AndRHS = 2546 dyn_cast<ConstantSDNode>(N0.getOperand(1))) { 2547 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3 2548 // Perform the xform if the AND RHS is a single bit. 2549 if ((AndRHS->getValue() & (AndRHS->getValue()-1)) == 0) { 2550 return DAG.getNode(ISD::SRL, VT, N0, 2551 DAG.getConstant(Log2_64(AndRHS->getValue()), 2552 TLI.getShiftAmountTy())); 2553 } 2554 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getValue()) { 2555 // (X & 8) == 8 --> (X & 8) >> 3 2556 // Perform the xform if C1 is a single bit. 2557 if ((C1 & (C1-1)) == 0) { 2558 return DAG.getNode(ISD::SRL, VT, N0, 2559 DAG.getConstant(Log2_64(C1),TLI.getShiftAmountTy())); 2560 } 2561 } 2562 } 2563 } 2564 } else if (isa<ConstantSDNode>(N0.Val)) { 2565 // Ensure that the constant occurs on the RHS. 2566 return DAG.getSetCC(VT, N1, N0, ISD::getSetCCSwappedOperands(Cond)); 2567 } 2568 2569 if (ConstantFPSDNode *N0C = dyn_cast<ConstantFPSDNode>(N0.Val)) 2570 if (ConstantFPSDNode *N1C = dyn_cast<ConstantFPSDNode>(N1.Val)) { 2571 double C0 = N0C->getValue(), C1 = N1C->getValue(); 2572 2573 switch (Cond) { 2574 default: break; // FIXME: Implement the rest of these! 2575 case ISD::SETEQ: return DAG.getConstant(C0 == C1, VT); 2576 case ISD::SETNE: return DAG.getConstant(C0 != C1, VT); 2577 case ISD::SETLT: return DAG.getConstant(C0 < C1, VT); 2578 case ISD::SETGT: return DAG.getConstant(C0 > C1, VT); 2579 case ISD::SETLE: return DAG.getConstant(C0 <= C1, VT); 2580 case ISD::SETGE: return DAG.getConstant(C0 >= C1, VT); 2581 } 2582 } else { 2583 // Ensure that the constant occurs on the RHS. 2584 return DAG.getSetCC(VT, N1, N0, ISD::getSetCCSwappedOperands(Cond)); 2585 } 2586 2587 if (N0 == N1) { 2588 // We can always fold X == Y for integer setcc's. 2589 if (MVT::isInteger(N0.getValueType())) 2590 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT); 2591 unsigned UOF = ISD::getUnorderedFlavor(Cond); 2592 if (UOF == 2) // FP operators that are undefined on NaNs. 2593 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT); 2594 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond))) 2595 return DAG.getConstant(UOF, VT); 2596 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO 2597 // if it is not already. 2598 ISD::CondCode NewCond = UOF == 0 ? ISD::SETUO : ISD::SETO; 2599 if (NewCond != Cond) 2600 return DAG.getSetCC(VT, N0, N1, NewCond); 2601 } 2602 2603 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && 2604 MVT::isInteger(N0.getValueType())) { 2605 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB || 2606 N0.getOpcode() == ISD::XOR) { 2607 // Simplify (X+Y) == (X+Z) --> Y == Z 2608 if (N0.getOpcode() == N1.getOpcode()) { 2609 if (N0.getOperand(0) == N1.getOperand(0)) 2610 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(1), Cond); 2611 if (N0.getOperand(1) == N1.getOperand(1)) 2612 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(0), Cond); 2613 if (isCommutativeBinOp(N0.getOpcode())) { 2614 // If X op Y == Y op X, try other combinations. 2615 if (N0.getOperand(0) == N1.getOperand(1)) 2616 return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(0), Cond); 2617 if (N0.getOperand(1) == N1.getOperand(0)) 2618 return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(1), Cond); 2619 } 2620 } 2621 2622 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0. Common for condcodes. 2623 if (N0.getOpcode() == ISD::XOR) 2624 if (ConstantSDNode *XORC = dyn_cast<ConstantSDNode>(N0.getOperand(1))) 2625 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) { 2626 // If we know that all of the inverted bits are zero, don't bother 2627 // performing the inversion. 2628 if (MaskedValueIsZero(N0.getOperand(0), ~XORC->getValue(), TLI)) 2629 return DAG.getSetCC(VT, N0.getOperand(0), 2630 DAG.getConstant(XORC->getValue()^RHSC->getValue(), 2631 N0.getValueType()), Cond); 2632 } 2633 2634 // Simplify (X+Z) == X --> Z == 0 2635 if (N0.getOperand(0) == N1) 2636 return DAG.getSetCC(VT, N0.getOperand(1), 2637 DAG.getConstant(0, N0.getValueType()), Cond); 2638 if (N0.getOperand(1) == N1) { 2639 if (isCommutativeBinOp(N0.getOpcode())) 2640 return DAG.getSetCC(VT, N0.getOperand(0), 2641 DAG.getConstant(0, N0.getValueType()), Cond); 2642 else { 2643 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!"); 2644 // (Z-X) == X --> Z == X<<1 2645 SDOperand SH = DAG.getNode(ISD::SHL, N1.getValueType(), 2646 N1, 2647 DAG.getConstant(1,TLI.getShiftAmountTy())); 2648 WorkList.push_back(SH.Val); 2649 return DAG.getSetCC(VT, N0.getOperand(0), SH, Cond); 2650 } 2651 } 2652 } 2653 2654 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB || 2655 N1.getOpcode() == ISD::XOR) { 2656 // Simplify X == (X+Z) --> Z == 0 2657 if (N1.getOperand(0) == N0) { 2658 return DAG.getSetCC(VT, N1.getOperand(1), 2659 DAG.getConstant(0, N1.getValueType()), Cond); 2660 } else if (N1.getOperand(1) == N0) { 2661 if (isCommutativeBinOp(N1.getOpcode())) { 2662 return DAG.getSetCC(VT, N1.getOperand(0), 2663 DAG.getConstant(0, N1.getValueType()), Cond); 2664 } else { 2665 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!"); 2666 // X == (Z-X) --> X<<1 == Z 2667 SDOperand SH = DAG.getNode(ISD::SHL, N1.getValueType(), N0, 2668 DAG.getConstant(1,TLI.getShiftAmountTy())); 2669 WorkList.push_back(SH.Val); 2670 return DAG.getSetCC(VT, SH, N1.getOperand(0), Cond); 2671 } 2672 } 2673 } 2674 } 2675 2676 // Fold away ALL boolean setcc's. 2677 SDOperand Temp; 2678 if (N0.getValueType() == MVT::i1 && foldBooleans) { 2679 switch (Cond) { 2680 default: assert(0 && "Unknown integer setcc!"); 2681 case ISD::SETEQ: // X == Y -> (X^Y)^1 2682 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, N1); 2683 N0 = DAG.getNode(ISD::XOR, MVT::i1, Temp, DAG.getConstant(1, MVT::i1)); 2684 WorkList.push_back(Temp.Val); 2685 break; 2686 case ISD::SETNE: // X != Y --> (X^Y) 2687 N0 = DAG.getNode(ISD::XOR, MVT::i1, N0, N1); 2688 break; 2689 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> X^1 & Y 2690 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> X^1 & Y 2691 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1)); 2692 N0 = DAG.getNode(ISD::AND, MVT::i1, N1, Temp); 2693 WorkList.push_back(Temp.Val); 2694 break; 2695 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> Y^1 & X 2696 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> Y^1 & X 2697 Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1)); 2698 N0 = DAG.getNode(ISD::AND, MVT::i1, N0, Temp); 2699 WorkList.push_back(Temp.Val); 2700 break; 2701 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> X^1 | Y 2702 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> X^1 | Y 2703 Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1)); 2704 N0 = DAG.getNode(ISD::OR, MVT::i1, N1, Temp); 2705 WorkList.push_back(Temp.Val); 2706 break; 2707 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> Y^1 | X 2708 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> Y^1 | X 2709 Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1)); 2710 N0 = DAG.getNode(ISD::OR, MVT::i1, N0, Temp); 2711 break; 2712 } 2713 if (VT != MVT::i1) { 2714 WorkList.push_back(N0.Val); 2715 // FIXME: If running after legalize, we probably can't do this. 2716 N0 = DAG.getNode(ISD::ZERO_EXTEND, VT, N0); 2717 } 2718 return N0; 2719 } 2720 2721 // Could not fold it. 2722 return SDOperand(); 2723} 2724 2725/// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant, 2726/// return a DAG expression to select that will generate the same value by 2727/// multiplying by a magic number. See: 2728/// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html> 2729SDOperand DAGCombiner::BuildSDIV(SDNode *N) { 2730 MVT::ValueType VT = N->getValueType(0); 2731 2732 // Check to see if we can do this. 2733 if (!TLI.isTypeLegal(VT) || (VT != MVT::i32 && VT != MVT::i64)) 2734 return SDOperand(); // BuildSDIV only operates on i32 or i64 2735 if (!TLI.isOperationLegal(ISD::MULHS, VT)) 2736 return SDOperand(); // Make sure the target supports MULHS. 2737 2738 int64_t d = cast<ConstantSDNode>(N->getOperand(1))->getSignExtended(); 2739 ms magics = (VT == MVT::i32) ? magic32(d) : magic64(d); 2740 2741 // Multiply the numerator (operand 0) by the magic value 2742 SDOperand Q = DAG.getNode(ISD::MULHS, VT, N->getOperand(0), 2743 DAG.getConstant(magics.m, VT)); 2744 // If d > 0 and m < 0, add the numerator 2745 if (d > 0 && magics.m < 0) { 2746 Q = DAG.getNode(ISD::ADD, VT, Q, N->getOperand(0)); 2747 WorkList.push_back(Q.Val); 2748 } 2749 // If d < 0 and m > 0, subtract the numerator. 2750 if (d < 0 && magics.m > 0) { 2751 Q = DAG.getNode(ISD::SUB, VT, Q, N->getOperand(0)); 2752 WorkList.push_back(Q.Val); 2753 } 2754 // Shift right algebraic if shift value is nonzero 2755 if (magics.s > 0) { 2756 Q = DAG.getNode(ISD::SRA, VT, Q, 2757 DAG.getConstant(magics.s, TLI.getShiftAmountTy())); 2758 WorkList.push_back(Q.Val); 2759 } 2760 // Extract the sign bit and add it to the quotient 2761 SDOperand T = 2762 DAG.getNode(ISD::SRL, VT, Q, DAG.getConstant(MVT::getSizeInBits(VT)-1, 2763 TLI.getShiftAmountTy())); 2764 WorkList.push_back(T.Val); 2765 return DAG.getNode(ISD::ADD, VT, Q, T); 2766} 2767 2768/// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant, 2769/// return a DAG expression to select that will generate the same value by 2770/// multiplying by a magic number. See: 2771/// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html> 2772SDOperand DAGCombiner::BuildUDIV(SDNode *N) { 2773 MVT::ValueType VT = N->getValueType(0); 2774 2775 // Check to see if we can do this. 2776 if (!TLI.isTypeLegal(VT) || (VT != MVT::i32 && VT != MVT::i64)) 2777 return SDOperand(); // BuildUDIV only operates on i32 or i64 2778 if (!TLI.isOperationLegal(ISD::MULHU, VT)) 2779 return SDOperand(); // Make sure the target supports MULHU. 2780 2781 uint64_t d = cast<ConstantSDNode>(N->getOperand(1))->getValue(); 2782 mu magics = (VT == MVT::i32) ? magicu32(d) : magicu64(d); 2783 2784 // Multiply the numerator (operand 0) by the magic value 2785 SDOperand Q = DAG.getNode(ISD::MULHU, VT, N->getOperand(0), 2786 DAG.getConstant(magics.m, VT)); 2787 WorkList.push_back(Q.Val); 2788 2789 if (magics.a == 0) { 2790 return DAG.getNode(ISD::SRL, VT, Q, 2791 DAG.getConstant(magics.s, TLI.getShiftAmountTy())); 2792 } else { 2793 SDOperand NPQ = DAG.getNode(ISD::SUB, VT, N->getOperand(0), Q); 2794 WorkList.push_back(NPQ.Val); 2795 NPQ = DAG.getNode(ISD::SRL, VT, NPQ, 2796 DAG.getConstant(1, TLI.getShiftAmountTy())); 2797 WorkList.push_back(NPQ.Val); 2798 NPQ = DAG.getNode(ISD::ADD, VT, NPQ, Q); 2799 WorkList.push_back(NPQ.Val); 2800 return DAG.getNode(ISD::SRL, VT, NPQ, 2801 DAG.getConstant(magics.s-1, TLI.getShiftAmountTy())); 2802 } 2803} 2804 2805// SelectionDAG::Combine - This is the entry point for the file. 2806// 2807void SelectionDAG::Combine(bool RunningAfterLegalize) { 2808 /// run - This is the main entry point to this class. 2809 /// 2810 DAGCombiner(*this).Run(RunningAfterLegalize); 2811} 2812