1 //===- InstCombineVectorOps.cpp -------------------------------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file implements instcombine for ExtractElement, InsertElement and 11 // ShuffleVector. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #include "InstCombine.h" 16 #include "llvm/Support/PatternMatch.h" 17 using namespace llvm; 18 using namespace PatternMatch; 19 20 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it 21 /// is to leave as a vector operation. isConstant indicates whether we're 22 /// extracting one known element. If false we're extracting a variable index. 23 static bool CheapToScalarize(Value *V, bool isConstant) { 24 if (Constant *C = dyn_cast<Constant>(V)) { 25 if (isConstant) return true; 26 27 // If all elts are the same, we can extract it and use any of the values. 28 Constant *Op0 = C->getAggregateElement(0U); 29 for (unsigned i = 1, e = V->getType()->getVectorNumElements(); i != e; ++i) 30 if (C->getAggregateElement(i) != Op0) 31 return false; 32 return true; 33 } 34 Instruction *I = dyn_cast<Instruction>(V); 35 if (!I) return false; 36 37 // Insert element gets simplified to the inserted element or is deleted if 38 // this is constant idx extract element and its a constant idx insertelt. 39 if (I->getOpcode() == Instruction::InsertElement && isConstant && 40 isa<ConstantInt>(I->getOperand(2))) 41 return true; 42 if (I->getOpcode() == Instruction::Load && I->hasOneUse()) 43 return true; 44 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) 45 if (BO->hasOneUse() && 46 (CheapToScalarize(BO->getOperand(0), isConstant) || 47 CheapToScalarize(BO->getOperand(1), isConstant))) 48 return true; 49 if (CmpInst *CI = dyn_cast<CmpInst>(I)) 50 if (CI->hasOneUse() && 51 (CheapToScalarize(CI->getOperand(0), isConstant) || 52 CheapToScalarize(CI->getOperand(1), isConstant))) 53 return true; 54 55 return false; 56 } 57 58 /// FindScalarElement - Given a vector and an element number, see if the scalar 59 /// value is already around as a register, for example if it were inserted then 60 /// extracted from the vector. 61 static Value *FindScalarElement(Value *V, unsigned EltNo) { 62 assert(V->getType()->isVectorTy() && "Not looking at a vector?"); 63 VectorType *VTy = cast<VectorType>(V->getType()); 64 unsigned Width = VTy->getNumElements(); 65 if (EltNo >= Width) // Out of range access. 66 return UndefValue::get(VTy->getElementType()); 67 68 if (Constant *C = dyn_cast<Constant>(V)) 69 return C->getAggregateElement(EltNo); 70 71 if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) { 72 // If this is an insert to a variable element, we don't know what it is. 73 if (!isa<ConstantInt>(III->getOperand(2))) 74 return 0; 75 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue(); 76 77 // If this is an insert to the element we are looking for, return the 78 // inserted value. 79 if (EltNo == IIElt) 80 return III->getOperand(1); 81 82 // Otherwise, the insertelement doesn't modify the value, recurse on its 83 // vector input. 84 return FindScalarElement(III->getOperand(0), EltNo); 85 } 86 87 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) { 88 unsigned LHSWidth = SVI->getOperand(0)->getType()->getVectorNumElements(); 89 int InEl = SVI->getMaskValue(EltNo); 90 if (InEl < 0) 91 return UndefValue::get(VTy->getElementType()); 92 if (InEl < (int)LHSWidth) 93 return FindScalarElement(SVI->getOperand(0), InEl); 94 return FindScalarElement(SVI->getOperand(1), InEl - LHSWidth); 95 } 96 97 // Extract a value from a vector add operation with a constant zero. 98 Value *Val = 0; Constant *Con = 0; 99 if (match(V, m_Add(m_Value(Val), m_Constant(Con)))) { 100 if (Con->getAggregateElement(EltNo)->isNullValue()) 101 return FindScalarElement(Val, EltNo); 102 } 103 104 // Otherwise, we don't know. 105 return 0; 106 } 107 108 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) { 109 // If vector val is constant with all elements the same, replace EI with 110 // that element. We handle a known element # below. 111 if (Constant *C = dyn_cast<Constant>(EI.getOperand(0))) 112 if (CheapToScalarize(C, false)) 113 return ReplaceInstUsesWith(EI, C->getAggregateElement(0U)); 114 115 // If extracting a specified index from the vector, see if we can recursively 116 // find a previously computed scalar that was inserted into the vector. 117 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) { 118 unsigned IndexVal = IdxC->getZExtValue(); 119 unsigned VectorWidth = EI.getVectorOperandType()->getNumElements(); 120 121 // If this is extracting an invalid index, turn this into undef, to avoid 122 // crashing the code below. 123 if (IndexVal >= VectorWidth) 124 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType())); 125 126 // This instruction only demands the single element from the input vector. 127 // If the input vector has a single use, simplify it based on this use 128 // property. 129 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) { 130 APInt UndefElts(VectorWidth, 0); 131 APInt DemandedMask(VectorWidth, 0); 132 DemandedMask.setBit(IndexVal); 133 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0), 134 DemandedMask, UndefElts)) { 135 EI.setOperand(0, V); 136 return &EI; 137 } 138 } 139 140 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal)) 141 return ReplaceInstUsesWith(EI, Elt); 142 143 // If the this extractelement is directly using a bitcast from a vector of 144 // the same number of elements, see if we can find the source element from 145 // it. In this case, we will end up needing to bitcast the scalars. 146 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) { 147 if (VectorType *VT = dyn_cast<VectorType>(BCI->getOperand(0)->getType())) 148 if (VT->getNumElements() == VectorWidth) 149 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal)) 150 return new BitCastInst(Elt, EI.getType()); 151 } 152 } 153 154 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) { 155 // Push extractelement into predecessor operation if legal and 156 // profitable to do so 157 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) { 158 if (I->hasOneUse() && 159 CheapToScalarize(BO, isa<ConstantInt>(EI.getOperand(1)))) { 160 Value *newEI0 = 161 Builder->CreateExtractElement(BO->getOperand(0), EI.getOperand(1), 162 EI.getName()+".lhs"); 163 Value *newEI1 = 164 Builder->CreateExtractElement(BO->getOperand(1), EI.getOperand(1), 165 EI.getName()+".rhs"); 166 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1); 167 } 168 } else if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) { 169 // Extracting the inserted element? 170 if (IE->getOperand(2) == EI.getOperand(1)) 171 return ReplaceInstUsesWith(EI, IE->getOperand(1)); 172 // If the inserted and extracted elements are constants, they must not 173 // be the same value, extract from the pre-inserted value instead. 174 if (isa<Constant>(IE->getOperand(2)) && isa<Constant>(EI.getOperand(1))) { 175 Worklist.AddValue(EI.getOperand(0)); 176 EI.setOperand(0, IE->getOperand(0)); 177 return &EI; 178 } 179 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) { 180 // If this is extracting an element from a shufflevector, figure out where 181 // it came from and extract from the appropriate input element instead. 182 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) { 183 int SrcIdx = SVI->getMaskValue(Elt->getZExtValue()); 184 Value *Src; 185 unsigned LHSWidth = 186 SVI->getOperand(0)->getType()->getVectorNumElements(); 187 188 if (SrcIdx < 0) 189 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType())); 190 if (SrcIdx < (int)LHSWidth) 191 Src = SVI->getOperand(0); 192 else { 193 SrcIdx -= LHSWidth; 194 Src = SVI->getOperand(1); 195 } 196 Type *Int32Ty = Type::getInt32Ty(EI.getContext()); 197 return ExtractElementInst::Create(Src, 198 ConstantInt::get(Int32Ty, 199 SrcIdx, false)); 200 } 201 } else if (CastInst *CI = dyn_cast<CastInst>(I)) { 202 // Canonicalize extractelement(cast) -> cast(extractelement) 203 // bitcasts can change the number of vector elements and they cost nothing 204 if (CI->hasOneUse() && EI.hasOneUse() && 205 (CI->getOpcode() != Instruction::BitCast)) { 206 Value *EE = Builder->CreateExtractElement(CI->getOperand(0), 207 EI.getIndexOperand()); 208 return CastInst::Create(CI->getOpcode(), EE, EI.getType()); 209 } 210 } 211 } 212 return 0; 213 } 214 215 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns 216 /// elements from either LHS or RHS, return the shuffle mask and true. 217 /// Otherwise, return false. 218 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS, 219 SmallVectorImpl<Constant*> &Mask) { 220 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() && 221 "Invalid CollectSingleShuffleElements"); 222 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements(); 223 224 if (isa<UndefValue>(V)) { 225 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(V->getContext()))); 226 return true; 227 } 228 229 if (V == LHS) { 230 for (unsigned i = 0; i != NumElts; ++i) 231 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()), i)); 232 return true; 233 } 234 235 if (V == RHS) { 236 for (unsigned i = 0; i != NumElts; ++i) 237 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()), 238 i+NumElts)); 239 return true; 240 } 241 242 if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) { 243 // If this is an insert of an extract from some other vector, include it. 244 Value *VecOp = IEI->getOperand(0); 245 Value *ScalarOp = IEI->getOperand(1); 246 Value *IdxOp = IEI->getOperand(2); 247 248 if (!isa<ConstantInt>(IdxOp)) 249 return false; 250 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue(); 251 252 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector. 253 // Okay, we can handle this if the vector we are insertinting into is 254 // transitively ok. 255 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) { 256 // If so, update the mask to reflect the inserted undef. 257 Mask[InsertedIdx] = UndefValue::get(Type::getInt32Ty(V->getContext())); 258 return true; 259 } 260 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){ 261 if (isa<ConstantInt>(EI->getOperand(1)) && 262 EI->getOperand(0)->getType() == V->getType()) { 263 unsigned ExtractedIdx = 264 cast<ConstantInt>(EI->getOperand(1))->getZExtValue(); 265 266 // This must be extracting from either LHS or RHS. 267 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) { 268 // Okay, we can handle this if the vector we are insertinting into is 269 // transitively ok. 270 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) { 271 // If so, update the mask to reflect the inserted value. 272 if (EI->getOperand(0) == LHS) { 273 Mask[InsertedIdx % NumElts] = 274 ConstantInt::get(Type::getInt32Ty(V->getContext()), 275 ExtractedIdx); 276 } else { 277 assert(EI->getOperand(0) == RHS); 278 Mask[InsertedIdx % NumElts] = 279 ConstantInt::get(Type::getInt32Ty(V->getContext()), 280 ExtractedIdx+NumElts); 281 } 282 return true; 283 } 284 } 285 } 286 } 287 } 288 // TODO: Handle shufflevector here! 289 290 return false; 291 } 292 293 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the 294 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask 295 /// that computes V and the LHS value of the shuffle. 296 static Value *CollectShuffleElements(Value *V, SmallVectorImpl<Constant*> &Mask, 297 Value *&RHS) { 298 assert(V->getType()->isVectorTy() && 299 (RHS == 0 || V->getType() == RHS->getType()) && 300 "Invalid shuffle!"); 301 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements(); 302 303 if (isa<UndefValue>(V)) { 304 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(V->getContext()))); 305 return V; 306 } 307 308 if (isa<ConstantAggregateZero>(V)) { 309 Mask.assign(NumElts, ConstantInt::get(Type::getInt32Ty(V->getContext()),0)); 310 return V; 311 } 312 313 if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) { 314 // If this is an insert of an extract from some other vector, include it. 315 Value *VecOp = IEI->getOperand(0); 316 Value *ScalarOp = IEI->getOperand(1); 317 Value *IdxOp = IEI->getOperand(2); 318 319 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) { 320 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) && 321 EI->getOperand(0)->getType() == V->getType()) { 322 unsigned ExtractedIdx = 323 cast<ConstantInt>(EI->getOperand(1))->getZExtValue(); 324 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue(); 325 326 // Either the extracted from or inserted into vector must be RHSVec, 327 // otherwise we'd end up with a shuffle of three inputs. 328 if (EI->getOperand(0) == RHS || RHS == 0) { 329 RHS = EI->getOperand(0); 330 Value *V = CollectShuffleElements(VecOp, Mask, RHS); 331 Mask[InsertedIdx % NumElts] = 332 ConstantInt::get(Type::getInt32Ty(V->getContext()), 333 NumElts+ExtractedIdx); 334 return V; 335 } 336 337 if (VecOp == RHS) { 338 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS); 339 // Everything but the extracted element is replaced with the RHS. 340 for (unsigned i = 0; i != NumElts; ++i) { 341 if (i != InsertedIdx) 342 Mask[i] = ConstantInt::get(Type::getInt32Ty(V->getContext()), 343 NumElts+i); 344 } 345 return V; 346 } 347 348 // If this insertelement is a chain that comes from exactly these two 349 // vectors, return the vector and the effective shuffle. 350 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask)) 351 return EI->getOperand(0); 352 } 353 } 354 } 355 // TODO: Handle shufflevector here! 356 357 // Otherwise, can't do anything fancy. Return an identity vector. 358 for (unsigned i = 0; i != NumElts; ++i) 359 Mask.push_back(ConstantInt::get(Type::getInt32Ty(V->getContext()), i)); 360 return V; 361 } 362 363 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) { 364 Value *VecOp = IE.getOperand(0); 365 Value *ScalarOp = IE.getOperand(1); 366 Value *IdxOp = IE.getOperand(2); 367 368 // Inserting an undef or into an undefined place, remove this. 369 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp)) 370 ReplaceInstUsesWith(IE, VecOp); 371 372 // If the inserted element was extracted from some other vector, and if the 373 // indexes are constant, try to turn this into a shufflevector operation. 374 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) { 375 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) && 376 EI->getOperand(0)->getType() == IE.getType()) { 377 unsigned NumVectorElts = IE.getType()->getNumElements(); 378 unsigned ExtractedIdx = 379 cast<ConstantInt>(EI->getOperand(1))->getZExtValue(); 380 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue(); 381 382 if (ExtractedIdx >= NumVectorElts) // Out of range extract. 383 return ReplaceInstUsesWith(IE, VecOp); 384 385 if (InsertedIdx >= NumVectorElts) // Out of range insert. 386 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType())); 387 388 // If we are extracting a value from a vector, then inserting it right 389 // back into the same place, just use the input vector. 390 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx) 391 return ReplaceInstUsesWith(IE, VecOp); 392 393 // If this insertelement isn't used by some other insertelement, turn it 394 // (and any insertelements it points to), into one big shuffle. 395 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) { 396 SmallVector<Constant*, 16> Mask; 397 Value *RHS = 0; 398 Value *LHS = CollectShuffleElements(&IE, Mask, RHS); 399 if (RHS == 0) RHS = UndefValue::get(LHS->getType()); 400 // We now have a shuffle of LHS, RHS, Mask. 401 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask)); 402 } 403 } 404 } 405 406 unsigned VWidth = cast<VectorType>(VecOp->getType())->getNumElements(); 407 APInt UndefElts(VWidth, 0); 408 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth)); 409 if (Value *V = SimplifyDemandedVectorElts(&IE, AllOnesEltMask, UndefElts)) { 410 if (V != &IE) 411 return ReplaceInstUsesWith(IE, V); 412 return &IE; 413 } 414 415 return 0; 416 } 417 418 419 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) { 420 Value *LHS = SVI.getOperand(0); 421 Value *RHS = SVI.getOperand(1); 422 SmallVector<int, 16> Mask = SVI.getShuffleMask(); 423 424 bool MadeChange = false; 425 426 // Undefined shuffle mask -> undefined value. 427 if (isa<UndefValue>(SVI.getOperand(2))) 428 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType())); 429 430 unsigned VWidth = cast<VectorType>(SVI.getType())->getNumElements(); 431 432 APInt UndefElts(VWidth, 0); 433 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth)); 434 if (Value *V = SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) { 435 if (V != &SVI) 436 return ReplaceInstUsesWith(SVI, V); 437 LHS = SVI.getOperand(0); 438 RHS = SVI.getOperand(1); 439 MadeChange = true; 440 } 441 442 unsigned LHSWidth = cast<VectorType>(LHS->getType())->getNumElements(); 443 444 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask') 445 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask'). 446 if (LHS == RHS || isa<UndefValue>(LHS)) { 447 if (isa<UndefValue>(LHS) && LHS == RHS) { 448 // shuffle(undef,undef,mask) -> undef. 449 Value* result = (VWidth == LHSWidth) 450 ? LHS : UndefValue::get(SVI.getType()); 451 return ReplaceInstUsesWith(SVI, result); 452 } 453 454 // Remap any references to RHS to use LHS. 455 SmallVector<Constant*, 16> Elts; 456 for (unsigned i = 0, e = LHSWidth; i != VWidth; ++i) { 457 if (Mask[i] < 0) { 458 Elts.push_back(UndefValue::get(Type::getInt32Ty(SVI.getContext()))); 459 continue; 460 } 461 462 if ((Mask[i] >= (int)e && isa<UndefValue>(RHS)) || 463 (Mask[i] < (int)e && isa<UndefValue>(LHS))) { 464 Mask[i] = -1; // Turn into undef. 465 Elts.push_back(UndefValue::get(Type::getInt32Ty(SVI.getContext()))); 466 } else { 467 Mask[i] = Mask[i] % e; // Force to LHS. 468 Elts.push_back(ConstantInt::get(Type::getInt32Ty(SVI.getContext()), 469 Mask[i])); 470 } 471 } 472 SVI.setOperand(0, SVI.getOperand(1)); 473 SVI.setOperand(1, UndefValue::get(RHS->getType())); 474 SVI.setOperand(2, ConstantVector::get(Elts)); 475 LHS = SVI.getOperand(0); 476 RHS = SVI.getOperand(1); 477 MadeChange = true; 478 } 479 480 if (VWidth == LHSWidth) { 481 // Analyze the shuffle, are the LHS or RHS and identity shuffles? 482 bool isLHSID = true, isRHSID = true; 483 484 for (unsigned i = 0, e = Mask.size(); i != e; ++i) { 485 if (Mask[i] < 0) continue; // Ignore undef values. 486 // Is this an identity shuffle of the LHS value? 487 isLHSID &= (Mask[i] == (int)i); 488 489 // Is this an identity shuffle of the RHS value? 490 isRHSID &= (Mask[i]-e == i); 491 } 492 493 // Eliminate identity shuffles. 494 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS); 495 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS); 496 } 497 498 // If the LHS is a shufflevector itself, see if we can combine it with this 499 // one without producing an unusual shuffle. 500 // Cases that might be simplified: 501 // 1. 502 // x1=shuffle(v1,v2,mask1) 503 // x=shuffle(x1,undef,mask) 504 // ==> 505 // x=shuffle(v1,undef,newMask) 506 // newMask[i] = (mask[i] < x1.size()) ? mask1[mask[i]] : -1 507 // 2. 508 // x1=shuffle(v1,undef,mask1) 509 // x=shuffle(x1,x2,mask) 510 // where v1.size() == mask1.size() 511 // ==> 512 // x=shuffle(v1,x2,newMask) 513 // newMask[i] = (mask[i] < x1.size()) ? mask1[mask[i]] : mask[i] 514 // 3. 515 // x2=shuffle(v2,undef,mask2) 516 // x=shuffle(x1,x2,mask) 517 // where v2.size() == mask2.size() 518 // ==> 519 // x=shuffle(x1,v2,newMask) 520 // newMask[i] = (mask[i] < x1.size()) 521 // ? mask[i] : mask2[mask[i]-x1.size()]+x1.size() 522 // 4. 523 // x1=shuffle(v1,undef,mask1) 524 // x2=shuffle(v2,undef,mask2) 525 // x=shuffle(x1,x2,mask) 526 // where v1.size() == v2.size() 527 // ==> 528 // x=shuffle(v1,v2,newMask) 529 // newMask[i] = (mask[i] < x1.size()) 530 // ? mask1[mask[i]] : mask2[mask[i]-x1.size()]+v1.size() 531 // 532 // Here we are really conservative: 533 // we are absolutely afraid of producing a shuffle mask not in the input 534 // program, because the code gen may not be smart enough to turn a merged 535 // shuffle into two specific shuffles: it may produce worse code. As such, 536 // we only merge two shuffles if the result is either a splat or one of the 537 // input shuffle masks. In this case, merging the shuffles just removes 538 // one instruction, which we know is safe. This is good for things like 539 // turning: (splat(splat)) -> splat, or 540 // merge(V[0..n], V[n+1..2n]) -> V[0..2n] 541 ShuffleVectorInst* LHSShuffle = dyn_cast<ShuffleVectorInst>(LHS); 542 ShuffleVectorInst* RHSShuffle = dyn_cast<ShuffleVectorInst>(RHS); 543 if (LHSShuffle) 544 if (!isa<UndefValue>(LHSShuffle->getOperand(1)) && !isa<UndefValue>(RHS)) 545 LHSShuffle = NULL; 546 if (RHSShuffle) 547 if (!isa<UndefValue>(RHSShuffle->getOperand(1))) 548 RHSShuffle = NULL; 549 if (!LHSShuffle && !RHSShuffle) 550 return MadeChange ? &SVI : 0; 551 552 Value* LHSOp0 = NULL; 553 Value* LHSOp1 = NULL; 554 Value* RHSOp0 = NULL; 555 unsigned LHSOp0Width = 0; 556 unsigned RHSOp0Width = 0; 557 if (LHSShuffle) { 558 LHSOp0 = LHSShuffle->getOperand(0); 559 LHSOp1 = LHSShuffle->getOperand(1); 560 LHSOp0Width = cast<VectorType>(LHSOp0->getType())->getNumElements(); 561 } 562 if (RHSShuffle) { 563 RHSOp0 = RHSShuffle->getOperand(0); 564 RHSOp0Width = cast<VectorType>(RHSOp0->getType())->getNumElements(); 565 } 566 Value* newLHS = LHS; 567 Value* newRHS = RHS; 568 if (LHSShuffle) { 569 // case 1 570 if (isa<UndefValue>(RHS)) { 571 newLHS = LHSOp0; 572 newRHS = LHSOp1; 573 } 574 // case 2 or 4 575 else if (LHSOp0Width == LHSWidth) { 576 newLHS = LHSOp0; 577 } 578 } 579 // case 3 or 4 580 if (RHSShuffle && RHSOp0Width == LHSWidth) { 581 newRHS = RHSOp0; 582 } 583 // case 4 584 if (LHSOp0 == RHSOp0) { 585 newLHS = LHSOp0; 586 newRHS = NULL; 587 } 588 589 if (newLHS == LHS && newRHS == RHS) 590 return MadeChange ? &SVI : 0; 591 592 SmallVector<int, 16> LHSMask; 593 SmallVector<int, 16> RHSMask; 594 if (newLHS != LHS) 595 LHSMask = LHSShuffle->getShuffleMask(); 596 if (RHSShuffle && newRHS != RHS) 597 RHSMask = RHSShuffle->getShuffleMask(); 598 599 unsigned newLHSWidth = (newLHS != LHS) ? LHSOp0Width : LHSWidth; 600 SmallVector<int, 16> newMask; 601 bool isSplat = true; 602 int SplatElt = -1; 603 // Create a new mask for the new ShuffleVectorInst so that the new 604 // ShuffleVectorInst is equivalent to the original one. 605 for (unsigned i = 0; i < VWidth; ++i) { 606 int eltMask; 607 if (Mask[i] < 0) { 608 // This element is an undef value. 609 eltMask = -1; 610 } else if (Mask[i] < (int)LHSWidth) { 611 // This element is from left hand side vector operand. 612 // 613 // If LHS is going to be replaced (case 1, 2, or 4), calculate the 614 // new mask value for the element. 615 if (newLHS != LHS) { 616 eltMask = LHSMask[Mask[i]]; 617 // If the value selected is an undef value, explicitly specify it 618 // with a -1 mask value. 619 if (eltMask >= (int)LHSOp0Width && isa<UndefValue>(LHSOp1)) 620 eltMask = -1; 621 } else 622 eltMask = Mask[i]; 623 } else { 624 // This element is from right hand side vector operand 625 // 626 // If the value selected is an undef value, explicitly specify it 627 // with a -1 mask value. (case 1) 628 if (isa<UndefValue>(RHS)) 629 eltMask = -1; 630 // If RHS is going to be replaced (case 3 or 4), calculate the 631 // new mask value for the element. 632 else if (newRHS != RHS) { 633 eltMask = RHSMask[Mask[i]-LHSWidth]; 634 // If the value selected is an undef value, explicitly specify it 635 // with a -1 mask value. 636 if (eltMask >= (int)RHSOp0Width) { 637 assert(isa<UndefValue>(RHSShuffle->getOperand(1)) 638 && "should have been check above"); 639 eltMask = -1; 640 } 641 } else 642 eltMask = Mask[i]-LHSWidth; 643 644 // If LHS's width is changed, shift the mask value accordingly. 645 // If newRHS == NULL, i.e. LHSOp0 == RHSOp0, we want to remap any 646 // references from RHSOp0 to LHSOp0, so we don't need to shift the mask. 647 // If newRHS == newLHS, we want to remap any references from newRHS to 648 // newLHS so that we can properly identify splats that may occur due to 649 // obfuscation accross the two vectors. 650 if (eltMask >= 0 && newRHS != NULL && newLHS != newRHS) 651 eltMask += newLHSWidth; 652 } 653 654 // Check if this could still be a splat. 655 if (eltMask >= 0) { 656 if (SplatElt >= 0 && SplatElt != eltMask) 657 isSplat = false; 658 SplatElt = eltMask; 659 } 660 661 newMask.push_back(eltMask); 662 } 663 664 // If the result mask is equal to one of the original shuffle masks, 665 // or is a splat, do the replacement. 666 if (isSplat || newMask == LHSMask || newMask == RHSMask || newMask == Mask) { 667 SmallVector<Constant*, 16> Elts; 668 Type *Int32Ty = Type::getInt32Ty(SVI.getContext()); 669 for (unsigned i = 0, e = newMask.size(); i != e; ++i) { 670 if (newMask[i] < 0) { 671 Elts.push_back(UndefValue::get(Int32Ty)); 672 } else { 673 Elts.push_back(ConstantInt::get(Int32Ty, newMask[i])); 674 } 675 } 676 if (newRHS == NULL) 677 newRHS = UndefValue::get(newLHS->getType()); 678 return new ShuffleVectorInst(newLHS, newRHS, ConstantVector::get(Elts)); 679 } 680 681 return MadeChange ? &SVI : 0; 682 } 683