1 //===- InstCombineSimplifyDemanded.cpp ------------------------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file contains logic for simplifying instructions based on information 10 // about how they are used. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "InstCombineInternal.h" 15 #include "llvm/Analysis/ValueTracking.h" 16 #include "llvm/IR/IntrinsicInst.h" 17 #include "llvm/IR/IntrinsicsAMDGPU.h" 18 #include "llvm/IR/IntrinsicsX86.h" 19 #include "llvm/IR/PatternMatch.h" 20 #include "llvm/Support/KnownBits.h" 21 22 using namespace llvm; 23 using namespace llvm::PatternMatch; 24 25 #define DEBUG_TYPE "instcombine" 26 27 namespace { 28 29 struct AMDGPUImageDMaskIntrinsic { 30 unsigned Intr; 31 }; 32 33 #define GET_AMDGPUImageDMaskIntrinsicTable_IMPL 34 #include "InstCombineTables.inc" 35 36 } // end anonymous namespace 37 38 /// Check to see if the specified operand of the specified instruction is a 39 /// constant integer. If so, check to see if there are any bits set in the 40 /// constant that are not demanded. If so, shrink the constant and return true. 41 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo, 42 const APInt &Demanded) { 43 assert(I && "No instruction?"); 44 assert(OpNo < I->getNumOperands() && "Operand index too large"); 45 46 // The operand must be a constant integer or splat integer. 47 Value *Op = I->getOperand(OpNo); 48 const APInt *C; 49 if (!match(Op, m_APInt(C))) 50 return false; 51 52 // If there are no bits set that aren't demanded, nothing to do. 53 if (C->isSubsetOf(Demanded)) 54 return false; 55 56 // This instruction is producing bits that are not demanded. Shrink the RHS. 57 I->setOperand(OpNo, ConstantInt::get(Op->getType(), *C & Demanded)); 58 59 return true; 60 } 61 62 63 64 /// Inst is an integer instruction that SimplifyDemandedBits knows about. See if 65 /// the instruction has any properties that allow us to simplify its operands. 66 bool InstCombiner::SimplifyDemandedInstructionBits(Instruction &Inst) { 67 unsigned BitWidth = Inst.getType()->getScalarSizeInBits(); 68 KnownBits Known(BitWidth); 69 APInt DemandedMask(APInt::getAllOnesValue(BitWidth)); 70 71 Value *V = SimplifyDemandedUseBits(&Inst, DemandedMask, Known, 72 0, &Inst); 73 if (!V) return false; 74 if (V == &Inst) return true; 75 replaceInstUsesWith(Inst, V); 76 return true; 77 } 78 79 /// This form of SimplifyDemandedBits simplifies the specified instruction 80 /// operand if possible, updating it in place. It returns true if it made any 81 /// change and false otherwise. 82 bool InstCombiner::SimplifyDemandedBits(Instruction *I, unsigned OpNo, 83 const APInt &DemandedMask, 84 KnownBits &Known, 85 unsigned Depth) { 86 Use &U = I->getOperandUse(OpNo); 87 Value *NewVal = SimplifyDemandedUseBits(U.get(), DemandedMask, Known, 88 Depth, I); 89 if (!NewVal) return false; 90 if (Instruction* OpInst = dyn_cast<Instruction>(U)) 91 salvageDebugInfo(*OpInst); 92 93 replaceUse(U, NewVal); 94 return true; 95 } 96 97 98 /// This function attempts to replace V with a simpler value based on the 99 /// demanded bits. When this function is called, it is known that only the bits 100 /// set in DemandedMask of the result of V are ever used downstream. 101 /// Consequently, depending on the mask and V, it may be possible to replace V 102 /// with a constant or one of its operands. In such cases, this function does 103 /// the replacement and returns true. In all other cases, it returns false after 104 /// analyzing the expression and setting KnownOne and known to be one in the 105 /// expression. Known.Zero contains all the bits that are known to be zero in 106 /// the expression. These are provided to potentially allow the caller (which 107 /// might recursively be SimplifyDemandedBits itself) to simplify the 108 /// expression. 109 /// Known.One and Known.Zero always follow the invariant that: 110 /// Known.One & Known.Zero == 0. 111 /// That is, a bit can't be both 1 and 0. Note that the bits in Known.One and 112 /// Known.Zero may only be accurate for those bits set in DemandedMask. Note 113 /// also that the bitwidth of V, DemandedMask, Known.Zero and Known.One must all 114 /// be the same. 115 /// 116 /// This returns null if it did not change anything and it permits no 117 /// simplification. This returns V itself if it did some simplification of V's 118 /// operands based on the information about what bits are demanded. This returns 119 /// some other non-null value if it found out that V is equal to another value 120 /// in the context where the specified bits are demanded, but not for all users. 121 Value *InstCombiner::SimplifyDemandedUseBits(Value *V, APInt DemandedMask, 122 KnownBits &Known, unsigned Depth, 123 Instruction *CxtI) { 124 assert(V != nullptr && "Null pointer of Value???"); 125 assert(Depth <= 6 && "Limit Search Depth"); 126 uint32_t BitWidth = DemandedMask.getBitWidth(); 127 Type *VTy = V->getType(); 128 assert( 129 (!VTy->isIntOrIntVectorTy() || VTy->getScalarSizeInBits() == BitWidth) && 130 Known.getBitWidth() == BitWidth && 131 "Value *V, DemandedMask and Known must have same BitWidth"); 132 133 if (isa<Constant>(V)) { 134 computeKnownBits(V, Known, Depth, CxtI); 135 return nullptr; 136 } 137 138 Known.resetAll(); 139 if (DemandedMask.isNullValue()) // Not demanding any bits from V. 140 return UndefValue::get(VTy); 141 142 if (Depth == 6) // Limit search depth. 143 return nullptr; 144 145 Instruction *I = dyn_cast<Instruction>(V); 146 if (!I) { 147 computeKnownBits(V, Known, Depth, CxtI); 148 return nullptr; // Only analyze instructions. 149 } 150 151 // If there are multiple uses of this value and we aren't at the root, then 152 // we can't do any simplifications of the operands, because DemandedMask 153 // only reflects the bits demanded by *one* of the users. 154 if (Depth != 0 && !I->hasOneUse()) 155 return SimplifyMultipleUseDemandedBits(I, DemandedMask, Known, Depth, CxtI); 156 157 KnownBits LHSKnown(BitWidth), RHSKnown(BitWidth); 158 159 // If this is the root being simplified, allow it to have multiple uses, 160 // just set the DemandedMask to all bits so that we can try to simplify the 161 // operands. This allows visitTruncInst (for example) to simplify the 162 // operand of a trunc without duplicating all the logic below. 163 if (Depth == 0 && !V->hasOneUse()) 164 DemandedMask.setAllBits(); 165 166 switch (I->getOpcode()) { 167 default: 168 computeKnownBits(I, Known, Depth, CxtI); 169 break; 170 case Instruction::And: { 171 // If either the LHS or the RHS are Zero, the result is zero. 172 if (SimplifyDemandedBits(I, 1, DemandedMask, RHSKnown, Depth + 1) || 173 SimplifyDemandedBits(I, 0, DemandedMask & ~RHSKnown.Zero, LHSKnown, 174 Depth + 1)) 175 return I; 176 assert(!RHSKnown.hasConflict() && "Bits known to be one AND zero?"); 177 assert(!LHSKnown.hasConflict() && "Bits known to be one AND zero?"); 178 179 Known = LHSKnown & RHSKnown; 180 181 // If the client is only demanding bits that we know, return the known 182 // constant. 183 if (DemandedMask.isSubsetOf(Known.Zero | Known.One)) 184 return Constant::getIntegerValue(VTy, Known.One); 185 186 // If all of the demanded bits are known 1 on one side, return the other. 187 // These bits cannot contribute to the result of the 'and'. 188 if (DemandedMask.isSubsetOf(LHSKnown.Zero | RHSKnown.One)) 189 return I->getOperand(0); 190 if (DemandedMask.isSubsetOf(RHSKnown.Zero | LHSKnown.One)) 191 return I->getOperand(1); 192 193 // If the RHS is a constant, see if we can simplify it. 194 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnown.Zero)) 195 return I; 196 197 break; 198 } 199 case Instruction::Or: { 200 // If either the LHS or the RHS are One, the result is One. 201 if (SimplifyDemandedBits(I, 1, DemandedMask, RHSKnown, Depth + 1) || 202 SimplifyDemandedBits(I, 0, DemandedMask & ~RHSKnown.One, LHSKnown, 203 Depth + 1)) 204 return I; 205 assert(!RHSKnown.hasConflict() && "Bits known to be one AND zero?"); 206 assert(!LHSKnown.hasConflict() && "Bits known to be one AND zero?"); 207 208 Known = LHSKnown | RHSKnown; 209 210 // If the client is only demanding bits that we know, return the known 211 // constant. 212 if (DemandedMask.isSubsetOf(Known.Zero | Known.One)) 213 return Constant::getIntegerValue(VTy, Known.One); 214 215 // If all of the demanded bits are known zero on one side, return the other. 216 // These bits cannot contribute to the result of the 'or'. 217 if (DemandedMask.isSubsetOf(LHSKnown.One | RHSKnown.Zero)) 218 return I->getOperand(0); 219 if (DemandedMask.isSubsetOf(RHSKnown.One | LHSKnown.Zero)) 220 return I->getOperand(1); 221 222 // If the RHS is a constant, see if we can simplify it. 223 if (ShrinkDemandedConstant(I, 1, DemandedMask)) 224 return I; 225 226 break; 227 } 228 case Instruction::Xor: { 229 if (SimplifyDemandedBits(I, 1, DemandedMask, RHSKnown, Depth + 1) || 230 SimplifyDemandedBits(I, 0, DemandedMask, LHSKnown, Depth + 1)) 231 return I; 232 assert(!RHSKnown.hasConflict() && "Bits known to be one AND zero?"); 233 assert(!LHSKnown.hasConflict() && "Bits known to be one AND zero?"); 234 235 Known = LHSKnown ^ RHSKnown; 236 237 // If the client is only demanding bits that we know, return the known 238 // constant. 239 if (DemandedMask.isSubsetOf(Known.Zero | Known.One)) 240 return Constant::getIntegerValue(VTy, Known.One); 241 242 // If all of the demanded bits are known zero on one side, return the other. 243 // These bits cannot contribute to the result of the 'xor'. 244 if (DemandedMask.isSubsetOf(RHSKnown.Zero)) 245 return I->getOperand(0); 246 if (DemandedMask.isSubsetOf(LHSKnown.Zero)) 247 return I->getOperand(1); 248 249 // If all of the demanded bits are known to be zero on one side or the 250 // other, turn this into an *inclusive* or. 251 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0 252 if (DemandedMask.isSubsetOf(RHSKnown.Zero | LHSKnown.Zero)) { 253 Instruction *Or = 254 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1), 255 I->getName()); 256 return InsertNewInstWith(Or, *I); 257 } 258 259 // If all of the demanded bits on one side are known, and all of the set 260 // bits on that side are also known to be set on the other side, turn this 261 // into an AND, as we know the bits will be cleared. 262 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2 263 if (DemandedMask.isSubsetOf(RHSKnown.Zero|RHSKnown.One) && 264 RHSKnown.One.isSubsetOf(LHSKnown.One)) { 265 Constant *AndC = Constant::getIntegerValue(VTy, 266 ~RHSKnown.One & DemandedMask); 267 Instruction *And = BinaryOperator::CreateAnd(I->getOperand(0), AndC); 268 return InsertNewInstWith(And, *I); 269 } 270 271 // If the RHS is a constant, see if we can simplify it. 272 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1. 273 if (ShrinkDemandedConstant(I, 1, DemandedMask)) 274 return I; 275 276 // If our LHS is an 'and' and if it has one use, and if any of the bits we 277 // are flipping are known to be set, then the xor is just resetting those 278 // bits to zero. We can just knock out bits from the 'and' and the 'xor', 279 // simplifying both of them. 280 if (Instruction *LHSInst = dyn_cast<Instruction>(I->getOperand(0))) 281 if (LHSInst->getOpcode() == Instruction::And && LHSInst->hasOneUse() && 282 isa<ConstantInt>(I->getOperand(1)) && 283 isa<ConstantInt>(LHSInst->getOperand(1)) && 284 (LHSKnown.One & RHSKnown.One & DemandedMask) != 0) { 285 ConstantInt *AndRHS = cast<ConstantInt>(LHSInst->getOperand(1)); 286 ConstantInt *XorRHS = cast<ConstantInt>(I->getOperand(1)); 287 APInt NewMask = ~(LHSKnown.One & RHSKnown.One & DemandedMask); 288 289 Constant *AndC = 290 ConstantInt::get(I->getType(), NewMask & AndRHS->getValue()); 291 Instruction *NewAnd = BinaryOperator::CreateAnd(I->getOperand(0), AndC); 292 InsertNewInstWith(NewAnd, *I); 293 294 Constant *XorC = 295 ConstantInt::get(I->getType(), NewMask & XorRHS->getValue()); 296 Instruction *NewXor = BinaryOperator::CreateXor(NewAnd, XorC); 297 return InsertNewInstWith(NewXor, *I); 298 } 299 300 break; 301 } 302 case Instruction::Select: { 303 Value *LHS, *RHS; 304 SelectPatternFlavor SPF = matchSelectPattern(I, LHS, RHS).Flavor; 305 if (SPF == SPF_UMAX) { 306 // UMax(A, C) == A if ... 307 // The lowest non-zero bit of DemandMask is higher than the highest 308 // non-zero bit of C. 309 const APInt *C; 310 unsigned CTZ = DemandedMask.countTrailingZeros(); 311 if (match(RHS, m_APInt(C)) && CTZ >= C->getActiveBits()) 312 return LHS; 313 } else if (SPF == SPF_UMIN) { 314 // UMin(A, C) == A if ... 315 // The lowest non-zero bit of DemandMask is higher than the highest 316 // non-one bit of C. 317 // This comes from using DeMorgans on the above umax example. 318 const APInt *C; 319 unsigned CTZ = DemandedMask.countTrailingZeros(); 320 if (match(RHS, m_APInt(C)) && 321 CTZ >= C->getBitWidth() - C->countLeadingOnes()) 322 return LHS; 323 } 324 325 // If this is a select as part of any other min/max pattern, don't simplify 326 // any further in case we break the structure. 327 if (SPF != SPF_UNKNOWN) 328 return nullptr; 329 330 if (SimplifyDemandedBits(I, 2, DemandedMask, RHSKnown, Depth + 1) || 331 SimplifyDemandedBits(I, 1, DemandedMask, LHSKnown, Depth + 1)) 332 return I; 333 assert(!RHSKnown.hasConflict() && "Bits known to be one AND zero?"); 334 assert(!LHSKnown.hasConflict() && "Bits known to be one AND zero?"); 335 336 // If the operands are constants, see if we can simplify them. 337 // This is similar to ShrinkDemandedConstant, but for a select we want to 338 // try to keep the selected constants the same as icmp value constants, if 339 // we can. This helps not break apart (or helps put back together) 340 // canonical patterns like min and max. 341 auto CanonicalizeSelectConstant = [](Instruction *I, unsigned OpNo, 342 APInt DemandedMask) { 343 const APInt *SelC; 344 if (!match(I->getOperand(OpNo), m_APInt(SelC))) 345 return false; 346 347 // Get the constant out of the ICmp, if there is one. 348 const APInt *CmpC; 349 ICmpInst::Predicate Pred; 350 if (!match(I->getOperand(0), m_c_ICmp(Pred, m_APInt(CmpC), m_Value())) || 351 CmpC->getBitWidth() != SelC->getBitWidth()) 352 return ShrinkDemandedConstant(I, OpNo, DemandedMask); 353 354 // If the constant is already the same as the ICmp, leave it as-is. 355 if (*CmpC == *SelC) 356 return false; 357 // If the constants are not already the same, but can be with the demand 358 // mask, use the constant value from the ICmp. 359 if ((*CmpC & DemandedMask) == (*SelC & DemandedMask)) { 360 I->setOperand(OpNo, ConstantInt::get(I->getType(), *CmpC)); 361 return true; 362 } 363 return ShrinkDemandedConstant(I, OpNo, DemandedMask); 364 }; 365 if (CanonicalizeSelectConstant(I, 1, DemandedMask) || 366 CanonicalizeSelectConstant(I, 2, DemandedMask)) 367 return I; 368 369 // Only known if known in both the LHS and RHS. 370 Known.One = RHSKnown.One & LHSKnown.One; 371 Known.Zero = RHSKnown.Zero & LHSKnown.Zero; 372 break; 373 } 374 case Instruction::ZExt: 375 case Instruction::Trunc: { 376 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits(); 377 378 APInt InputDemandedMask = DemandedMask.zextOrTrunc(SrcBitWidth); 379 KnownBits InputKnown(SrcBitWidth); 380 if (SimplifyDemandedBits(I, 0, InputDemandedMask, InputKnown, Depth + 1)) 381 return I; 382 assert(InputKnown.getBitWidth() == SrcBitWidth && "Src width changed?"); 383 Known = InputKnown.zextOrTrunc(BitWidth); 384 assert(!Known.hasConflict() && "Bits known to be one AND zero?"); 385 break; 386 } 387 case Instruction::BitCast: 388 if (!I->getOperand(0)->getType()->isIntOrIntVectorTy()) 389 return nullptr; // vector->int or fp->int? 390 391 if (VectorType *DstVTy = dyn_cast<VectorType>(I->getType())) { 392 if (VectorType *SrcVTy = 393 dyn_cast<VectorType>(I->getOperand(0)->getType())) { 394 if (DstVTy->getNumElements() != SrcVTy->getNumElements()) 395 // Don't touch a bitcast between vectors of different element counts. 396 return nullptr; 397 } else 398 // Don't touch a scalar-to-vector bitcast. 399 return nullptr; 400 } else if (I->getOperand(0)->getType()->isVectorTy()) 401 // Don't touch a vector-to-scalar bitcast. 402 return nullptr; 403 404 if (SimplifyDemandedBits(I, 0, DemandedMask, Known, Depth + 1)) 405 return I; 406 assert(!Known.hasConflict() && "Bits known to be one AND zero?"); 407 break; 408 case Instruction::SExt: { 409 // Compute the bits in the result that are not present in the input. 410 unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits(); 411 412 APInt InputDemandedBits = DemandedMask.trunc(SrcBitWidth); 413 414 // If any of the sign extended bits are demanded, we know that the sign 415 // bit is demanded. 416 if (DemandedMask.getActiveBits() > SrcBitWidth) 417 InputDemandedBits.setBit(SrcBitWidth-1); 418 419 KnownBits InputKnown(SrcBitWidth); 420 if (SimplifyDemandedBits(I, 0, InputDemandedBits, InputKnown, Depth + 1)) 421 return I; 422 423 // If the input sign bit is known zero, or if the NewBits are not demanded 424 // convert this into a zero extension. 425 if (InputKnown.isNonNegative() || 426 DemandedMask.getActiveBits() <= SrcBitWidth) { 427 // Convert to ZExt cast. 428 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName()); 429 return InsertNewInstWith(NewCast, *I); 430 } 431 432 // If the sign bit of the input is known set or clear, then we know the 433 // top bits of the result. 434 Known = InputKnown.sext(BitWidth); 435 assert(!Known.hasConflict() && "Bits known to be one AND zero?"); 436 break; 437 } 438 case Instruction::Add: 439 if ((DemandedMask & 1) == 0) { 440 // If we do not need the low bit, try to convert bool math to logic: 441 // add iN (zext i1 X), (sext i1 Y) --> sext (~X & Y) to iN 442 Value *X, *Y; 443 if (match(I, m_c_Add(m_OneUse(m_ZExt(m_Value(X))), 444 m_OneUse(m_SExt(m_Value(Y))))) && 445 X->getType()->isIntOrIntVectorTy(1) && X->getType() == Y->getType()) { 446 // Truth table for inputs and output signbits: 447 // X:0 | X:1 448 // ---------- 449 // Y:0 | 0 | 0 | 450 // Y:1 | -1 | 0 | 451 // ---------- 452 IRBuilderBase::InsertPointGuard Guard(Builder); 453 Builder.SetInsertPoint(I); 454 Value *AndNot = Builder.CreateAnd(Builder.CreateNot(X), Y); 455 return Builder.CreateSExt(AndNot, VTy); 456 } 457 458 // add iN (sext i1 X), (sext i1 Y) --> sext (X | Y) to iN 459 // TODO: Relax the one-use checks because we are removing an instruction? 460 if (match(I, m_Add(m_OneUse(m_SExt(m_Value(X))), 461 m_OneUse(m_SExt(m_Value(Y))))) && 462 X->getType()->isIntOrIntVectorTy(1) && X->getType() == Y->getType()) { 463 // Truth table for inputs and output signbits: 464 // X:0 | X:1 465 // ----------- 466 // Y:0 | -1 | -1 | 467 // Y:1 | -1 | 0 | 468 // ----------- 469 IRBuilderBase::InsertPointGuard Guard(Builder); 470 Builder.SetInsertPoint(I); 471 Value *Or = Builder.CreateOr(X, Y); 472 return Builder.CreateSExt(Or, VTy); 473 } 474 } 475 LLVM_FALLTHROUGH; 476 case Instruction::Sub: { 477 /// If the high-bits of an ADD/SUB are not demanded, then we do not care 478 /// about the high bits of the operands. 479 unsigned NLZ = DemandedMask.countLeadingZeros(); 480 // Right fill the mask of bits for this ADD/SUB to demand the most 481 // significant bit and all those below it. 482 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ)); 483 if (ShrinkDemandedConstant(I, 0, DemandedFromOps) || 484 SimplifyDemandedBits(I, 0, DemandedFromOps, LHSKnown, Depth + 1) || 485 ShrinkDemandedConstant(I, 1, DemandedFromOps) || 486 SimplifyDemandedBits(I, 1, DemandedFromOps, RHSKnown, Depth + 1)) { 487 if (NLZ > 0) { 488 // Disable the nsw and nuw flags here: We can no longer guarantee that 489 // we won't wrap after simplification. Removing the nsw/nuw flags is 490 // legal here because the top bit is not demanded. 491 BinaryOperator &BinOP = *cast<BinaryOperator>(I); 492 BinOP.setHasNoSignedWrap(false); 493 BinOP.setHasNoUnsignedWrap(false); 494 } 495 return I; 496 } 497 498 // If we are known to be adding/subtracting zeros to every bit below 499 // the highest demanded bit, we just return the other side. 500 if (DemandedFromOps.isSubsetOf(RHSKnown.Zero)) 501 return I->getOperand(0); 502 // We can't do this with the LHS for subtraction, unless we are only 503 // demanding the LSB. 504 if ((I->getOpcode() == Instruction::Add || 505 DemandedFromOps.isOneValue()) && 506 DemandedFromOps.isSubsetOf(LHSKnown.Zero)) 507 return I->getOperand(1); 508 509 // Otherwise just compute the known bits of the result. 510 bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap(); 511 Known = KnownBits::computeForAddSub(I->getOpcode() == Instruction::Add, 512 NSW, LHSKnown, RHSKnown); 513 break; 514 } 515 case Instruction::Shl: { 516 const APInt *SA; 517 if (match(I->getOperand(1), m_APInt(SA))) { 518 const APInt *ShrAmt; 519 if (match(I->getOperand(0), m_Shr(m_Value(), m_APInt(ShrAmt)))) 520 if (Instruction *Shr = dyn_cast<Instruction>(I->getOperand(0))) 521 if (Value *R = simplifyShrShlDemandedBits(Shr, *ShrAmt, I, *SA, 522 DemandedMask, Known)) 523 return R; 524 525 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth-1); 526 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt)); 527 528 // If the shift is NUW/NSW, then it does demand the high bits. 529 ShlOperator *IOp = cast<ShlOperator>(I); 530 if (IOp->hasNoSignedWrap()) 531 DemandedMaskIn.setHighBits(ShiftAmt+1); 532 else if (IOp->hasNoUnsignedWrap()) 533 DemandedMaskIn.setHighBits(ShiftAmt); 534 535 if (SimplifyDemandedBits(I, 0, DemandedMaskIn, Known, Depth + 1)) 536 return I; 537 assert(!Known.hasConflict() && "Bits known to be one AND zero?"); 538 539 bool SignBitZero = Known.Zero.isSignBitSet(); 540 bool SignBitOne = Known.One.isSignBitSet(); 541 Known.Zero <<= ShiftAmt; 542 Known.One <<= ShiftAmt; 543 // low bits known zero. 544 if (ShiftAmt) 545 Known.Zero.setLowBits(ShiftAmt); 546 547 // If this shift has "nsw" keyword, then the result is either a poison 548 // value or has the same sign bit as the first operand. 549 if (IOp->hasNoSignedWrap()) { 550 if (SignBitZero) 551 Known.Zero.setSignBit(); 552 else if (SignBitOne) 553 Known.One.setSignBit(); 554 if (Known.hasConflict()) 555 return UndefValue::get(I->getType()); 556 } 557 } else { 558 computeKnownBits(I, Known, Depth, CxtI); 559 } 560 break; 561 } 562 case Instruction::LShr: { 563 const APInt *SA; 564 if (match(I->getOperand(1), m_APInt(SA))) { 565 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth-1); 566 567 // Unsigned shift right. 568 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt)); 569 570 // If the shift is exact, then it does demand the low bits (and knows that 571 // they are zero). 572 if (cast<LShrOperator>(I)->isExact()) 573 DemandedMaskIn.setLowBits(ShiftAmt); 574 575 if (SimplifyDemandedBits(I, 0, DemandedMaskIn, Known, Depth + 1)) 576 return I; 577 assert(!Known.hasConflict() && "Bits known to be one AND zero?"); 578 Known.Zero.lshrInPlace(ShiftAmt); 579 Known.One.lshrInPlace(ShiftAmt); 580 if (ShiftAmt) 581 Known.Zero.setHighBits(ShiftAmt); // high bits known zero. 582 } else { 583 computeKnownBits(I, Known, Depth, CxtI); 584 } 585 break; 586 } 587 case Instruction::AShr: { 588 // If this is an arithmetic shift right and only the low-bit is set, we can 589 // always convert this into a logical shr, even if the shift amount is 590 // variable. The low bit of the shift cannot be an input sign bit unless 591 // the shift amount is >= the size of the datatype, which is undefined. 592 if (DemandedMask.isOneValue()) { 593 // Perform the logical shift right. 594 Instruction *NewVal = BinaryOperator::CreateLShr( 595 I->getOperand(0), I->getOperand(1), I->getName()); 596 return InsertNewInstWith(NewVal, *I); 597 } 598 599 // If the sign bit is the only bit demanded by this ashr, then there is no 600 // need to do it, the shift doesn't change the high bit. 601 if (DemandedMask.isSignMask()) 602 return I->getOperand(0); 603 604 const APInt *SA; 605 if (match(I->getOperand(1), m_APInt(SA))) { 606 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth-1); 607 608 // Signed shift right. 609 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt)); 610 // If any of the high bits are demanded, we should set the sign bit as 611 // demanded. 612 if (DemandedMask.countLeadingZeros() <= ShiftAmt) 613 DemandedMaskIn.setSignBit(); 614 615 // If the shift is exact, then it does demand the low bits (and knows that 616 // they are zero). 617 if (cast<AShrOperator>(I)->isExact()) 618 DemandedMaskIn.setLowBits(ShiftAmt); 619 620 if (SimplifyDemandedBits(I, 0, DemandedMaskIn, Known, Depth + 1)) 621 return I; 622 623 unsigned SignBits = ComputeNumSignBits(I->getOperand(0), Depth + 1, CxtI); 624 625 assert(!Known.hasConflict() && "Bits known to be one AND zero?"); 626 // Compute the new bits that are at the top now plus sign bits. 627 APInt HighBits(APInt::getHighBitsSet( 628 BitWidth, std::min(SignBits + ShiftAmt - 1, BitWidth))); 629 Known.Zero.lshrInPlace(ShiftAmt); 630 Known.One.lshrInPlace(ShiftAmt); 631 632 // If the input sign bit is known to be zero, or if none of the top bits 633 // are demanded, turn this into an unsigned shift right. 634 assert(BitWidth > ShiftAmt && "Shift amount not saturated?"); 635 if (Known.Zero[BitWidth-ShiftAmt-1] || 636 !DemandedMask.intersects(HighBits)) { 637 BinaryOperator *LShr = BinaryOperator::CreateLShr(I->getOperand(0), 638 I->getOperand(1)); 639 LShr->setIsExact(cast<BinaryOperator>(I)->isExact()); 640 return InsertNewInstWith(LShr, *I); 641 } else if (Known.One[BitWidth-ShiftAmt-1]) { // New bits are known one. 642 Known.One |= HighBits; 643 } 644 } else { 645 computeKnownBits(I, Known, Depth, CxtI); 646 } 647 break; 648 } 649 case Instruction::UDiv: { 650 // UDiv doesn't demand low bits that are zero in the divisor. 651 const APInt *SA; 652 if (match(I->getOperand(1), m_APInt(SA))) { 653 // If the shift is exact, then it does demand the low bits. 654 if (cast<UDivOperator>(I)->isExact()) 655 break; 656 657 // FIXME: Take the demanded mask of the result into account. 658 unsigned RHSTrailingZeros = SA->countTrailingZeros(); 659 APInt DemandedMaskIn = 660 APInt::getHighBitsSet(BitWidth, BitWidth - RHSTrailingZeros); 661 if (SimplifyDemandedBits(I, 0, DemandedMaskIn, LHSKnown, Depth + 1)) 662 return I; 663 664 // Propagate zero bits from the input. 665 Known.Zero.setHighBits(std::min( 666 BitWidth, LHSKnown.Zero.countLeadingOnes() + RHSTrailingZeros)); 667 } else { 668 computeKnownBits(I, Known, Depth, CxtI); 669 } 670 break; 671 } 672 case Instruction::SRem: 673 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) { 674 // X % -1 demands all the bits because we don't want to introduce 675 // INT_MIN % -1 (== undef) by accident. 676 if (Rem->isMinusOne()) 677 break; 678 APInt RA = Rem->getValue().abs(); 679 if (RA.isPowerOf2()) { 680 if (DemandedMask.ult(RA)) // srem won't affect demanded bits 681 return I->getOperand(0); 682 683 APInt LowBits = RA - 1; 684 APInt Mask2 = LowBits | APInt::getSignMask(BitWidth); 685 if (SimplifyDemandedBits(I, 0, Mask2, LHSKnown, Depth + 1)) 686 return I; 687 688 // The low bits of LHS are unchanged by the srem. 689 Known.Zero = LHSKnown.Zero & LowBits; 690 Known.One = LHSKnown.One & LowBits; 691 692 // If LHS is non-negative or has all low bits zero, then the upper bits 693 // are all zero. 694 if (LHSKnown.isNonNegative() || LowBits.isSubsetOf(LHSKnown.Zero)) 695 Known.Zero |= ~LowBits; 696 697 // If LHS is negative and not all low bits are zero, then the upper bits 698 // are all one. 699 if (LHSKnown.isNegative() && LowBits.intersects(LHSKnown.One)) 700 Known.One |= ~LowBits; 701 702 assert(!Known.hasConflict() && "Bits known to be one AND zero?"); 703 break; 704 } 705 } 706 707 // The sign bit is the LHS's sign bit, except when the result of the 708 // remainder is zero. 709 if (DemandedMask.isSignBitSet()) { 710 computeKnownBits(I->getOperand(0), LHSKnown, Depth + 1, CxtI); 711 // If it's known zero, our sign bit is also zero. 712 if (LHSKnown.isNonNegative()) 713 Known.makeNonNegative(); 714 } 715 break; 716 case Instruction::URem: { 717 KnownBits Known2(BitWidth); 718 APInt AllOnes = APInt::getAllOnesValue(BitWidth); 719 if (SimplifyDemandedBits(I, 0, AllOnes, Known2, Depth + 1) || 720 SimplifyDemandedBits(I, 1, AllOnes, Known2, Depth + 1)) 721 return I; 722 723 unsigned Leaders = Known2.countMinLeadingZeros(); 724 Known.Zero = APInt::getHighBitsSet(BitWidth, Leaders) & DemandedMask; 725 break; 726 } 727 case Instruction::Call: { 728 bool KnownBitsComputed = false; 729 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 730 switch (II->getIntrinsicID()) { 731 default: break; 732 case Intrinsic::bswap: { 733 // If the only bits demanded come from one byte of the bswap result, 734 // just shift the input byte into position to eliminate the bswap. 735 unsigned NLZ = DemandedMask.countLeadingZeros(); 736 unsigned NTZ = DemandedMask.countTrailingZeros(); 737 738 // Round NTZ down to the next byte. If we have 11 trailing zeros, then 739 // we need all the bits down to bit 8. Likewise, round NLZ. If we 740 // have 14 leading zeros, round to 8. 741 NLZ &= ~7; 742 NTZ &= ~7; 743 // If we need exactly one byte, we can do this transformation. 744 if (BitWidth-NLZ-NTZ == 8) { 745 unsigned ResultBit = NTZ; 746 unsigned InputBit = BitWidth-NTZ-8; 747 748 // Replace this with either a left or right shift to get the byte into 749 // the right place. 750 Instruction *NewVal; 751 if (InputBit > ResultBit) 752 NewVal = BinaryOperator::CreateLShr(II->getArgOperand(0), 753 ConstantInt::get(I->getType(), InputBit-ResultBit)); 754 else 755 NewVal = BinaryOperator::CreateShl(II->getArgOperand(0), 756 ConstantInt::get(I->getType(), ResultBit-InputBit)); 757 NewVal->takeName(I); 758 return InsertNewInstWith(NewVal, *I); 759 } 760 break; 761 } 762 case Intrinsic::fshr: 763 case Intrinsic::fshl: { 764 const APInt *SA; 765 if (!match(I->getOperand(2), m_APInt(SA))) 766 break; 767 768 // Normalize to funnel shift left. APInt shifts of BitWidth are well- 769 // defined, so no need to special-case zero shifts here. 770 uint64_t ShiftAmt = SA->urem(BitWidth); 771 if (II->getIntrinsicID() == Intrinsic::fshr) 772 ShiftAmt = BitWidth - ShiftAmt; 773 774 APInt DemandedMaskLHS(DemandedMask.lshr(ShiftAmt)); 775 APInt DemandedMaskRHS(DemandedMask.shl(BitWidth - ShiftAmt)); 776 if (SimplifyDemandedBits(I, 0, DemandedMaskLHS, LHSKnown, Depth + 1) || 777 SimplifyDemandedBits(I, 1, DemandedMaskRHS, RHSKnown, Depth + 1)) 778 return I; 779 780 Known.Zero = LHSKnown.Zero.shl(ShiftAmt) | 781 RHSKnown.Zero.lshr(BitWidth - ShiftAmt); 782 Known.One = LHSKnown.One.shl(ShiftAmt) | 783 RHSKnown.One.lshr(BitWidth - ShiftAmt); 784 KnownBitsComputed = true; 785 break; 786 } 787 case Intrinsic::x86_mmx_pmovmskb: 788 case Intrinsic::x86_sse_movmsk_ps: 789 case Intrinsic::x86_sse2_movmsk_pd: 790 case Intrinsic::x86_sse2_pmovmskb_128: 791 case Intrinsic::x86_avx_movmsk_ps_256: 792 case Intrinsic::x86_avx_movmsk_pd_256: 793 case Intrinsic::x86_avx2_pmovmskb: { 794 // MOVMSK copies the vector elements' sign bits to the low bits 795 // and zeros the high bits. 796 unsigned ArgWidth; 797 if (II->getIntrinsicID() == Intrinsic::x86_mmx_pmovmskb) { 798 ArgWidth = 8; // Arg is x86_mmx, but treated as <8 x i8>. 799 } else { 800 auto Arg = II->getArgOperand(0); 801 auto ArgType = cast<VectorType>(Arg->getType()); 802 ArgWidth = ArgType->getNumElements(); 803 } 804 805 // If we don't need any of low bits then return zero, 806 // we know that DemandedMask is non-zero already. 807 APInt DemandedElts = DemandedMask.zextOrTrunc(ArgWidth); 808 if (DemandedElts.isNullValue()) 809 return ConstantInt::getNullValue(VTy); 810 811 // We know that the upper bits are set to zero. 812 Known.Zero.setBitsFrom(ArgWidth); 813 KnownBitsComputed = true; 814 break; 815 } 816 case Intrinsic::x86_sse42_crc32_64_64: 817 Known.Zero.setBitsFrom(32); 818 KnownBitsComputed = true; 819 break; 820 } 821 } 822 823 if (!KnownBitsComputed) 824 computeKnownBits(V, Known, Depth, CxtI); 825 break; 826 } 827 } 828 829 // If the client is only demanding bits that we know, return the known 830 // constant. 831 if (DemandedMask.isSubsetOf(Known.Zero|Known.One)) 832 return Constant::getIntegerValue(VTy, Known.One); 833 return nullptr; 834 } 835 836 /// Helper routine of SimplifyDemandedUseBits. It computes Known 837 /// bits. It also tries to handle simplifications that can be done based on 838 /// DemandedMask, but without modifying the Instruction. 839 Value *InstCombiner::SimplifyMultipleUseDemandedBits(Instruction *I, 840 const APInt &DemandedMask, 841 KnownBits &Known, 842 unsigned Depth, 843 Instruction *CxtI) { 844 unsigned BitWidth = DemandedMask.getBitWidth(); 845 Type *ITy = I->getType(); 846 847 KnownBits LHSKnown(BitWidth); 848 KnownBits RHSKnown(BitWidth); 849 850 // Despite the fact that we can't simplify this instruction in all User's 851 // context, we can at least compute the known bits, and we can 852 // do simplifications that apply to *just* the one user if we know that 853 // this instruction has a simpler value in that context. 854 switch (I->getOpcode()) { 855 case Instruction::And: { 856 // If either the LHS or the RHS are Zero, the result is zero. 857 computeKnownBits(I->getOperand(1), RHSKnown, Depth + 1, CxtI); 858 computeKnownBits(I->getOperand(0), LHSKnown, Depth + 1, 859 CxtI); 860 861 Known = LHSKnown & RHSKnown; 862 863 // If the client is only demanding bits that we know, return the known 864 // constant. 865 if (DemandedMask.isSubsetOf(Known.Zero | Known.One)) 866 return Constant::getIntegerValue(ITy, Known.One); 867 868 // If all of the demanded bits are known 1 on one side, return the other. 869 // These bits cannot contribute to the result of the 'and' in this 870 // context. 871 if (DemandedMask.isSubsetOf(LHSKnown.Zero | RHSKnown.One)) 872 return I->getOperand(0); 873 if (DemandedMask.isSubsetOf(RHSKnown.Zero | LHSKnown.One)) 874 return I->getOperand(1); 875 876 break; 877 } 878 case Instruction::Or: { 879 // We can simplify (X|Y) -> X or Y in the user's context if we know that 880 // only bits from X or Y are demanded. 881 882 // If either the LHS or the RHS are One, the result is One. 883 computeKnownBits(I->getOperand(1), RHSKnown, Depth + 1, CxtI); 884 computeKnownBits(I->getOperand(0), LHSKnown, Depth + 1, 885 CxtI); 886 887 Known = LHSKnown | RHSKnown; 888 889 // If the client is only demanding bits that we know, return the known 890 // constant. 891 if (DemandedMask.isSubsetOf(Known.Zero | Known.One)) 892 return Constant::getIntegerValue(ITy, Known.One); 893 894 // If all of the demanded bits are known zero on one side, return the 895 // other. These bits cannot contribute to the result of the 'or' in this 896 // context. 897 if (DemandedMask.isSubsetOf(LHSKnown.One | RHSKnown.Zero)) 898 return I->getOperand(0); 899 if (DemandedMask.isSubsetOf(RHSKnown.One | LHSKnown.Zero)) 900 return I->getOperand(1); 901 902 break; 903 } 904 case Instruction::Xor: { 905 // We can simplify (X^Y) -> X or Y in the user's context if we know that 906 // only bits from X or Y are demanded. 907 908 computeKnownBits(I->getOperand(1), RHSKnown, Depth + 1, CxtI); 909 computeKnownBits(I->getOperand(0), LHSKnown, Depth + 1, 910 CxtI); 911 912 Known = LHSKnown ^ RHSKnown; 913 914 // If the client is only demanding bits that we know, return the known 915 // constant. 916 if (DemandedMask.isSubsetOf(Known.Zero | Known.One)) 917 return Constant::getIntegerValue(ITy, Known.One); 918 919 // If all of the demanded bits are known zero on one side, return the 920 // other. 921 if (DemandedMask.isSubsetOf(RHSKnown.Zero)) 922 return I->getOperand(0); 923 if (DemandedMask.isSubsetOf(LHSKnown.Zero)) 924 return I->getOperand(1); 925 926 break; 927 } 928 default: 929 // Compute the Known bits to simplify things downstream. 930 computeKnownBits(I, Known, Depth, CxtI); 931 932 // If this user is only demanding bits that we know, return the known 933 // constant. 934 if (DemandedMask.isSubsetOf(Known.Zero|Known.One)) 935 return Constant::getIntegerValue(ITy, Known.One); 936 937 break; 938 } 939 940 return nullptr; 941 } 942 943 944 /// Helper routine of SimplifyDemandedUseBits. It tries to simplify 945 /// "E1 = (X lsr C1) << C2", where the C1 and C2 are constant, into 946 /// "E2 = X << (C2 - C1)" or "E2 = X >> (C1 - C2)", depending on the sign 947 /// of "C2-C1". 948 /// 949 /// Suppose E1 and E2 are generally different in bits S={bm, bm+1, 950 /// ..., bn}, without considering the specific value X is holding. 951 /// This transformation is legal iff one of following conditions is hold: 952 /// 1) All the bit in S are 0, in this case E1 == E2. 953 /// 2) We don't care those bits in S, per the input DemandedMask. 954 /// 3) Combination of 1) and 2). Some bits in S are 0, and we don't care the 955 /// rest bits. 956 /// 957 /// Currently we only test condition 2). 958 /// 959 /// As with SimplifyDemandedUseBits, it returns NULL if the simplification was 960 /// not successful. 961 Value * 962 InstCombiner::simplifyShrShlDemandedBits(Instruction *Shr, const APInt &ShrOp1, 963 Instruction *Shl, const APInt &ShlOp1, 964 const APInt &DemandedMask, 965 KnownBits &Known) { 966 if (!ShlOp1 || !ShrOp1) 967 return nullptr; // No-op. 968 969 Value *VarX = Shr->getOperand(0); 970 Type *Ty = VarX->getType(); 971 unsigned BitWidth = Ty->getScalarSizeInBits(); 972 if (ShlOp1.uge(BitWidth) || ShrOp1.uge(BitWidth)) 973 return nullptr; // Undef. 974 975 unsigned ShlAmt = ShlOp1.getZExtValue(); 976 unsigned ShrAmt = ShrOp1.getZExtValue(); 977 978 Known.One.clearAllBits(); 979 Known.Zero.setLowBits(ShlAmt - 1); 980 Known.Zero &= DemandedMask; 981 982 APInt BitMask1(APInt::getAllOnesValue(BitWidth)); 983 APInt BitMask2(APInt::getAllOnesValue(BitWidth)); 984 985 bool isLshr = (Shr->getOpcode() == Instruction::LShr); 986 BitMask1 = isLshr ? (BitMask1.lshr(ShrAmt) << ShlAmt) : 987 (BitMask1.ashr(ShrAmt) << ShlAmt); 988 989 if (ShrAmt <= ShlAmt) { 990 BitMask2 <<= (ShlAmt - ShrAmt); 991 } else { 992 BitMask2 = isLshr ? BitMask2.lshr(ShrAmt - ShlAmt): 993 BitMask2.ashr(ShrAmt - ShlAmt); 994 } 995 996 // Check if condition-2 (see the comment to this function) is satified. 997 if ((BitMask1 & DemandedMask) == (BitMask2 & DemandedMask)) { 998 if (ShrAmt == ShlAmt) 999 return VarX; 1000 1001 if (!Shr->hasOneUse()) 1002 return nullptr; 1003 1004 BinaryOperator *New; 1005 if (ShrAmt < ShlAmt) { 1006 Constant *Amt = ConstantInt::get(VarX->getType(), ShlAmt - ShrAmt); 1007 New = BinaryOperator::CreateShl(VarX, Amt); 1008 BinaryOperator *Orig = cast<BinaryOperator>(Shl); 1009 New->setHasNoSignedWrap(Orig->hasNoSignedWrap()); 1010 New->setHasNoUnsignedWrap(Orig->hasNoUnsignedWrap()); 1011 } else { 1012 Constant *Amt = ConstantInt::get(VarX->getType(), ShrAmt - ShlAmt); 1013 New = isLshr ? BinaryOperator::CreateLShr(VarX, Amt) : 1014 BinaryOperator::CreateAShr(VarX, Amt); 1015 if (cast<BinaryOperator>(Shr)->isExact()) 1016 New->setIsExact(true); 1017 } 1018 1019 return InsertNewInstWith(New, *Shl); 1020 } 1021 1022 return nullptr; 1023 } 1024 1025 /// Implement SimplifyDemandedVectorElts for amdgcn buffer and image intrinsics. 1026 /// 1027 /// Note: This only supports non-TFE/LWE image intrinsic calls; those have 1028 /// struct returns. 1029 Value *InstCombiner::simplifyAMDGCNMemoryIntrinsicDemanded(IntrinsicInst *II, 1030 APInt DemandedElts, 1031 int DMaskIdx) { 1032 1033 auto *IIVTy = cast<VectorType>(II->getType()); 1034 unsigned VWidth = IIVTy->getNumElements(); 1035 if (VWidth == 1) 1036 return nullptr; 1037 1038 IRBuilderBase::InsertPointGuard Guard(Builder); 1039 Builder.SetInsertPoint(II); 1040 1041 // Assume the arguments are unchanged and later override them, if needed. 1042 SmallVector<Value *, 16> Args(II->arg_begin(), II->arg_end()); 1043 1044 if (DMaskIdx < 0) { 1045 // Buffer case. 1046 1047 const unsigned ActiveBits = DemandedElts.getActiveBits(); 1048 const unsigned UnusedComponentsAtFront = DemandedElts.countTrailingZeros(); 1049 1050 // Start assuming the prefix of elements is demanded, but possibly clear 1051 // some other bits if there are trailing zeros (unused components at front) 1052 // and update offset. 1053 DemandedElts = (1 << ActiveBits) - 1; 1054 1055 if (UnusedComponentsAtFront > 0) { 1056 static const unsigned InvalidOffsetIdx = 0xf; 1057 1058 unsigned OffsetIdx; 1059 switch (II->getIntrinsicID()) { 1060 case Intrinsic::amdgcn_raw_buffer_load: 1061 OffsetIdx = 1; 1062 break; 1063 case Intrinsic::amdgcn_s_buffer_load: 1064 // If resulting type is vec3, there is no point in trimming the 1065 // load with updated offset, as the vec3 would most likely be widened to 1066 // vec4 anyway during lowering. 1067 if (ActiveBits == 4 && UnusedComponentsAtFront == 1) 1068 OffsetIdx = InvalidOffsetIdx; 1069 else 1070 OffsetIdx = 1; 1071 break; 1072 case Intrinsic::amdgcn_struct_buffer_load: 1073 OffsetIdx = 2; 1074 break; 1075 default: 1076 // TODO: handle tbuffer* intrinsics. 1077 OffsetIdx = InvalidOffsetIdx; 1078 break; 1079 } 1080 1081 if (OffsetIdx != InvalidOffsetIdx) { 1082 // Clear demanded bits and update the offset. 1083 DemandedElts &= ~((1 << UnusedComponentsAtFront) - 1); 1084 auto *Offset = II->getArgOperand(OffsetIdx); 1085 unsigned SingleComponentSizeInBits = 1086 getDataLayout().getTypeSizeInBits(II->getType()->getScalarType()); 1087 unsigned OffsetAdd = 1088 UnusedComponentsAtFront * SingleComponentSizeInBits / 8; 1089 auto *OffsetAddVal = ConstantInt::get(Offset->getType(), OffsetAdd); 1090 Args[OffsetIdx] = Builder.CreateAdd(Offset, OffsetAddVal); 1091 } 1092 } 1093 } else { 1094 // Image case. 1095 1096 ConstantInt *DMask = cast<ConstantInt>(II->getArgOperand(DMaskIdx)); 1097 unsigned DMaskVal = DMask->getZExtValue() & 0xf; 1098 1099 // Mask off values that are undefined because the dmask doesn't cover them 1100 DemandedElts &= (1 << countPopulation(DMaskVal)) - 1; 1101 1102 unsigned NewDMaskVal = 0; 1103 unsigned OrigLoadIdx = 0; 1104 for (unsigned SrcIdx = 0; SrcIdx < 4; ++SrcIdx) { 1105 const unsigned Bit = 1 << SrcIdx; 1106 if (!!(DMaskVal & Bit)) { 1107 if (!!DemandedElts[OrigLoadIdx]) 1108 NewDMaskVal |= Bit; 1109 OrigLoadIdx++; 1110 } 1111 } 1112 1113 if (DMaskVal != NewDMaskVal) 1114 Args[DMaskIdx] = ConstantInt::get(DMask->getType(), NewDMaskVal); 1115 } 1116 1117 unsigned NewNumElts = DemandedElts.countPopulation(); 1118 if (!NewNumElts) 1119 return UndefValue::get(II->getType()); 1120 1121 // FIXME: Allow v3i16/v3f16 in buffer and image intrinsics when the types are 1122 // fully supported. 1123 if (II->getType()->getScalarSizeInBits() == 16 && NewNumElts == 3) 1124 return nullptr; 1125 1126 if (NewNumElts >= VWidth && DemandedElts.isMask()) { 1127 if (DMaskIdx >= 0) 1128 II->setArgOperand(DMaskIdx, Args[DMaskIdx]); 1129 return nullptr; 1130 } 1131 1132 // Validate function argument and return types, extracting overloaded types 1133 // along the way. 1134 SmallVector<Type *, 6> OverloadTys; 1135 if (!Intrinsic::getIntrinsicSignature(II->getCalledFunction(), OverloadTys)) 1136 return nullptr; 1137 1138 Module *M = II->getParent()->getParent()->getParent(); 1139 Type *EltTy = IIVTy->getElementType(); 1140 Type *NewTy = 1141 (NewNumElts == 1) ? EltTy : FixedVectorType::get(EltTy, NewNumElts); 1142 1143 OverloadTys[0] = NewTy; 1144 Function *NewIntrin = 1145 Intrinsic::getDeclaration(M, II->getIntrinsicID(), OverloadTys); 1146 1147 CallInst *NewCall = Builder.CreateCall(NewIntrin, Args); 1148 NewCall->takeName(II); 1149 NewCall->copyMetadata(*II); 1150 1151 if (NewNumElts == 1) { 1152 return Builder.CreateInsertElement(UndefValue::get(II->getType()), NewCall, 1153 DemandedElts.countTrailingZeros()); 1154 } 1155 1156 SmallVector<int, 8> EltMask; 1157 unsigned NewLoadIdx = 0; 1158 for (unsigned OrigLoadIdx = 0; OrigLoadIdx < VWidth; ++OrigLoadIdx) { 1159 if (!!DemandedElts[OrigLoadIdx]) 1160 EltMask.push_back(NewLoadIdx++); 1161 else 1162 EltMask.push_back(NewNumElts); 1163 } 1164 1165 Value *Shuffle = 1166 Builder.CreateShuffleVector(NewCall, UndefValue::get(NewTy), EltMask); 1167 1168 return Shuffle; 1169 } 1170 1171 /// The specified value produces a vector with any number of elements. 1172 /// This method analyzes which elements of the operand are undef and returns 1173 /// that information in UndefElts. 1174 /// 1175 /// DemandedElts contains the set of elements that are actually used by the 1176 /// caller, and by default (AllowMultipleUsers equals false) the value is 1177 /// simplified only if it has a single caller. If AllowMultipleUsers is set 1178 /// to true, DemandedElts refers to the union of sets of elements that are 1179 /// used by all callers. 1180 /// 1181 /// If the information about demanded elements can be used to simplify the 1182 /// operation, the operation is simplified, then the resultant value is 1183 /// returned. This returns null if no change was made. 1184 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, APInt DemandedElts, 1185 APInt &UndefElts, 1186 unsigned Depth, 1187 bool AllowMultipleUsers) { 1188 // Cannot analyze scalable type. The number of vector elements is not a 1189 // compile-time constant. 1190 if (isa<ScalableVectorType>(V->getType())) 1191 return nullptr; 1192 1193 unsigned VWidth = cast<FixedVectorType>(V->getType())->getNumElements(); 1194 APInt EltMask(APInt::getAllOnesValue(VWidth)); 1195 assert((DemandedElts & ~EltMask) == 0 && "Invalid DemandedElts!"); 1196 1197 if (isa<UndefValue>(V)) { 1198 // If the entire vector is undefined, just return this info. 1199 UndefElts = EltMask; 1200 return nullptr; 1201 } 1202 1203 if (DemandedElts.isNullValue()) { // If nothing is demanded, provide undef. 1204 UndefElts = EltMask; 1205 return UndefValue::get(V->getType()); 1206 } 1207 1208 UndefElts = 0; 1209 1210 if (auto *C = dyn_cast<Constant>(V)) { 1211 // Check if this is identity. If so, return 0 since we are not simplifying 1212 // anything. 1213 if (DemandedElts.isAllOnesValue()) 1214 return nullptr; 1215 1216 Type *EltTy = cast<VectorType>(V->getType())->getElementType(); 1217 Constant *Undef = UndefValue::get(EltTy); 1218 SmallVector<Constant*, 16> Elts; 1219 for (unsigned i = 0; i != VWidth; ++i) { 1220 if (!DemandedElts[i]) { // If not demanded, set to undef. 1221 Elts.push_back(Undef); 1222 UndefElts.setBit(i); 1223 continue; 1224 } 1225 1226 Constant *Elt = C->getAggregateElement(i); 1227 if (!Elt) return nullptr; 1228 1229 if (isa<UndefValue>(Elt)) { // Already undef. 1230 Elts.push_back(Undef); 1231 UndefElts.setBit(i); 1232 } else { // Otherwise, defined. 1233 Elts.push_back(Elt); 1234 } 1235 } 1236 1237 // If we changed the constant, return it. 1238 Constant *NewCV = ConstantVector::get(Elts); 1239 return NewCV != C ? NewCV : nullptr; 1240 } 1241 1242 // Limit search depth. 1243 if (Depth == 10) 1244 return nullptr; 1245 1246 if (!AllowMultipleUsers) { 1247 // If multiple users are using the root value, proceed with 1248 // simplification conservatively assuming that all elements 1249 // are needed. 1250 if (!V->hasOneUse()) { 1251 // Quit if we find multiple users of a non-root value though. 1252 // They'll be handled when it's their turn to be visited by 1253 // the main instcombine process. 1254 if (Depth != 0) 1255 // TODO: Just compute the UndefElts information recursively. 1256 return nullptr; 1257 1258 // Conservatively assume that all elements are needed. 1259 DemandedElts = EltMask; 1260 } 1261 } 1262 1263 Instruction *I = dyn_cast<Instruction>(V); 1264 if (!I) return nullptr; // Only analyze instructions. 1265 1266 bool MadeChange = false; 1267 auto simplifyAndSetOp = [&](Instruction *Inst, unsigned OpNum, 1268 APInt Demanded, APInt &Undef) { 1269 auto *II = dyn_cast<IntrinsicInst>(Inst); 1270 Value *Op = II ? II->getArgOperand(OpNum) : Inst->getOperand(OpNum); 1271 if (Value *V = SimplifyDemandedVectorElts(Op, Demanded, Undef, Depth + 1)) { 1272 replaceOperand(*Inst, OpNum, V); 1273 MadeChange = true; 1274 } 1275 }; 1276 1277 APInt UndefElts2(VWidth, 0); 1278 APInt UndefElts3(VWidth, 0); 1279 switch (I->getOpcode()) { 1280 default: break; 1281 1282 case Instruction::GetElementPtr: { 1283 // The LangRef requires that struct geps have all constant indices. As 1284 // such, we can't convert any operand to partial undef. 1285 auto mayIndexStructType = [](GetElementPtrInst &GEP) { 1286 for (auto I = gep_type_begin(GEP), E = gep_type_end(GEP); 1287 I != E; I++) 1288 if (I.isStruct()) 1289 return true;; 1290 return false; 1291 }; 1292 if (mayIndexStructType(cast<GetElementPtrInst>(*I))) 1293 break; 1294 1295 // Conservatively track the demanded elements back through any vector 1296 // operands we may have. We know there must be at least one, or we 1297 // wouldn't have a vector result to get here. Note that we intentionally 1298 // merge the undef bits here since gepping with either an undef base or 1299 // index results in undef. 1300 for (unsigned i = 0; i < I->getNumOperands(); i++) { 1301 if (isa<UndefValue>(I->getOperand(i))) { 1302 // If the entire vector is undefined, just return this info. 1303 UndefElts = EltMask; 1304 return nullptr; 1305 } 1306 if (I->getOperand(i)->getType()->isVectorTy()) { 1307 APInt UndefEltsOp(VWidth, 0); 1308 simplifyAndSetOp(I, i, DemandedElts, UndefEltsOp); 1309 UndefElts |= UndefEltsOp; 1310 } 1311 } 1312 1313 break; 1314 } 1315 case Instruction::InsertElement: { 1316 // If this is a variable index, we don't know which element it overwrites. 1317 // demand exactly the same input as we produce. 1318 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2)); 1319 if (!Idx) { 1320 // Note that we can't propagate undef elt info, because we don't know 1321 // which elt is getting updated. 1322 simplifyAndSetOp(I, 0, DemandedElts, UndefElts2); 1323 break; 1324 } 1325 1326 // The element inserted overwrites whatever was there, so the input demanded 1327 // set is simpler than the output set. 1328 unsigned IdxNo = Idx->getZExtValue(); 1329 APInt PreInsertDemandedElts = DemandedElts; 1330 if (IdxNo < VWidth) 1331 PreInsertDemandedElts.clearBit(IdxNo); 1332 1333 simplifyAndSetOp(I, 0, PreInsertDemandedElts, UndefElts); 1334 1335 // If this is inserting an element that isn't demanded, remove this 1336 // insertelement. 1337 if (IdxNo >= VWidth || !DemandedElts[IdxNo]) { 1338 Worklist.push(I); 1339 return I->getOperand(0); 1340 } 1341 1342 // The inserted element is defined. 1343 UndefElts.clearBit(IdxNo); 1344 break; 1345 } 1346 case Instruction::ShuffleVector: { 1347 auto *Shuffle = cast<ShuffleVectorInst>(I); 1348 assert(Shuffle->getOperand(0)->getType() == 1349 Shuffle->getOperand(1)->getType() && 1350 "Expected shuffle operands to have same type"); 1351 unsigned OpWidth = 1352 cast<VectorType>(Shuffle->getOperand(0)->getType())->getNumElements(); 1353 // Handle trivial case of a splat. Only check the first element of LHS 1354 // operand. 1355 if (all_of(Shuffle->getShuffleMask(), [](int Elt) { return Elt == 0; }) && 1356 DemandedElts.isAllOnesValue()) { 1357 if (!isa<UndefValue>(I->getOperand(1))) { 1358 I->setOperand(1, UndefValue::get(I->getOperand(1)->getType())); 1359 MadeChange = true; 1360 } 1361 APInt LeftDemanded(OpWidth, 1); 1362 APInt LHSUndefElts(OpWidth, 0); 1363 simplifyAndSetOp(I, 0, LeftDemanded, LHSUndefElts); 1364 if (LHSUndefElts[0]) 1365 UndefElts = EltMask; 1366 else 1367 UndefElts.clearAllBits(); 1368 break; 1369 } 1370 1371 APInt LeftDemanded(OpWidth, 0), RightDemanded(OpWidth, 0); 1372 for (unsigned i = 0; i < VWidth; i++) { 1373 if (DemandedElts[i]) { 1374 unsigned MaskVal = Shuffle->getMaskValue(i); 1375 if (MaskVal != -1u) { 1376 assert(MaskVal < OpWidth * 2 && 1377 "shufflevector mask index out of range!"); 1378 if (MaskVal < OpWidth) 1379 LeftDemanded.setBit(MaskVal); 1380 else 1381 RightDemanded.setBit(MaskVal - OpWidth); 1382 } 1383 } 1384 } 1385 1386 APInt LHSUndefElts(OpWidth, 0); 1387 simplifyAndSetOp(I, 0, LeftDemanded, LHSUndefElts); 1388 1389 APInt RHSUndefElts(OpWidth, 0); 1390 simplifyAndSetOp(I, 1, RightDemanded, RHSUndefElts); 1391 1392 // If this shuffle does not change the vector length and the elements 1393 // demanded by this shuffle are an identity mask, then this shuffle is 1394 // unnecessary. 1395 // 1396 // We are assuming canonical form for the mask, so the source vector is 1397 // operand 0 and operand 1 is not used. 1398 // 1399 // Note that if an element is demanded and this shuffle mask is undefined 1400 // for that element, then the shuffle is not considered an identity 1401 // operation. The shuffle prevents poison from the operand vector from 1402 // leaking to the result by replacing poison with an undefined value. 1403 if (VWidth == OpWidth) { 1404 bool IsIdentityShuffle = true; 1405 for (unsigned i = 0; i < VWidth; i++) { 1406 unsigned MaskVal = Shuffle->getMaskValue(i); 1407 if (DemandedElts[i] && i != MaskVal) { 1408 IsIdentityShuffle = false; 1409 break; 1410 } 1411 } 1412 if (IsIdentityShuffle) 1413 return Shuffle->getOperand(0); 1414 } 1415 1416 bool NewUndefElts = false; 1417 unsigned LHSIdx = -1u, LHSValIdx = -1u; 1418 unsigned RHSIdx = -1u, RHSValIdx = -1u; 1419 bool LHSUniform = true; 1420 bool RHSUniform = true; 1421 for (unsigned i = 0; i < VWidth; i++) { 1422 unsigned MaskVal = Shuffle->getMaskValue(i); 1423 if (MaskVal == -1u) { 1424 UndefElts.setBit(i); 1425 } else if (!DemandedElts[i]) { 1426 NewUndefElts = true; 1427 UndefElts.setBit(i); 1428 } else if (MaskVal < OpWidth) { 1429 if (LHSUndefElts[MaskVal]) { 1430 NewUndefElts = true; 1431 UndefElts.setBit(i); 1432 } else { 1433 LHSIdx = LHSIdx == -1u ? i : OpWidth; 1434 LHSValIdx = LHSValIdx == -1u ? MaskVal : OpWidth; 1435 LHSUniform = LHSUniform && (MaskVal == i); 1436 } 1437 } else { 1438 if (RHSUndefElts[MaskVal - OpWidth]) { 1439 NewUndefElts = true; 1440 UndefElts.setBit(i); 1441 } else { 1442 RHSIdx = RHSIdx == -1u ? i : OpWidth; 1443 RHSValIdx = RHSValIdx == -1u ? MaskVal - OpWidth : OpWidth; 1444 RHSUniform = RHSUniform && (MaskVal - OpWidth == i); 1445 } 1446 } 1447 } 1448 1449 // Try to transform shuffle with constant vector and single element from 1450 // this constant vector to single insertelement instruction. 1451 // shufflevector V, C, <v1, v2, .., ci, .., vm> -> 1452 // insertelement V, C[ci], ci-n 1453 if (OpWidth == Shuffle->getType()->getNumElements()) { 1454 Value *Op = nullptr; 1455 Constant *Value = nullptr; 1456 unsigned Idx = -1u; 1457 1458 // Find constant vector with the single element in shuffle (LHS or RHS). 1459 if (LHSIdx < OpWidth && RHSUniform) { 1460 if (auto *CV = dyn_cast<ConstantVector>(Shuffle->getOperand(0))) { 1461 Op = Shuffle->getOperand(1); 1462 Value = CV->getOperand(LHSValIdx); 1463 Idx = LHSIdx; 1464 } 1465 } 1466 if (RHSIdx < OpWidth && LHSUniform) { 1467 if (auto *CV = dyn_cast<ConstantVector>(Shuffle->getOperand(1))) { 1468 Op = Shuffle->getOperand(0); 1469 Value = CV->getOperand(RHSValIdx); 1470 Idx = RHSIdx; 1471 } 1472 } 1473 // Found constant vector with single element - convert to insertelement. 1474 if (Op && Value) { 1475 Instruction *New = InsertElementInst::Create( 1476 Op, Value, ConstantInt::get(Type::getInt32Ty(I->getContext()), Idx), 1477 Shuffle->getName()); 1478 InsertNewInstWith(New, *Shuffle); 1479 return New; 1480 } 1481 } 1482 if (NewUndefElts) { 1483 // Add additional discovered undefs. 1484 SmallVector<int, 16> Elts; 1485 for (unsigned i = 0; i < VWidth; ++i) { 1486 if (UndefElts[i]) 1487 Elts.push_back(UndefMaskElem); 1488 else 1489 Elts.push_back(Shuffle->getMaskValue(i)); 1490 } 1491 Shuffle->setShuffleMask(Elts); 1492 MadeChange = true; 1493 } 1494 break; 1495 } 1496 case Instruction::Select: { 1497 // If this is a vector select, try to transform the select condition based 1498 // on the current demanded elements. 1499 SelectInst *Sel = cast<SelectInst>(I); 1500 if (Sel->getCondition()->getType()->isVectorTy()) { 1501 // TODO: We are not doing anything with UndefElts based on this call. 1502 // It is overwritten below based on the other select operands. If an 1503 // element of the select condition is known undef, then we are free to 1504 // choose the output value from either arm of the select. If we know that 1505 // one of those values is undef, then the output can be undef. 1506 simplifyAndSetOp(I, 0, DemandedElts, UndefElts); 1507 } 1508 1509 // Next, see if we can transform the arms of the select. 1510 APInt DemandedLHS(DemandedElts), DemandedRHS(DemandedElts); 1511 if (auto *CV = dyn_cast<ConstantVector>(Sel->getCondition())) { 1512 for (unsigned i = 0; i < VWidth; i++) { 1513 // isNullValue() always returns false when called on a ConstantExpr. 1514 // Skip constant expressions to avoid propagating incorrect information. 1515 Constant *CElt = CV->getAggregateElement(i); 1516 if (isa<ConstantExpr>(CElt)) 1517 continue; 1518 // TODO: If a select condition element is undef, we can demand from 1519 // either side. If one side is known undef, choosing that side would 1520 // propagate undef. 1521 if (CElt->isNullValue()) 1522 DemandedLHS.clearBit(i); 1523 else 1524 DemandedRHS.clearBit(i); 1525 } 1526 } 1527 1528 simplifyAndSetOp(I, 1, DemandedLHS, UndefElts2); 1529 simplifyAndSetOp(I, 2, DemandedRHS, UndefElts3); 1530 1531 // Output elements are undefined if the element from each arm is undefined. 1532 // TODO: This can be improved. See comment in select condition handling. 1533 UndefElts = UndefElts2 & UndefElts3; 1534 break; 1535 } 1536 case Instruction::BitCast: { 1537 // Vector->vector casts only. 1538 VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType()); 1539 if (!VTy) break; 1540 unsigned InVWidth = VTy->getNumElements(); 1541 APInt InputDemandedElts(InVWidth, 0); 1542 UndefElts2 = APInt(InVWidth, 0); 1543 unsigned Ratio; 1544 1545 if (VWidth == InVWidth) { 1546 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same 1547 // elements as are demanded of us. 1548 Ratio = 1; 1549 InputDemandedElts = DemandedElts; 1550 } else if ((VWidth % InVWidth) == 0) { 1551 // If the number of elements in the output is a multiple of the number of 1552 // elements in the input then an input element is live if any of the 1553 // corresponding output elements are live. 1554 Ratio = VWidth / InVWidth; 1555 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) 1556 if (DemandedElts[OutIdx]) 1557 InputDemandedElts.setBit(OutIdx / Ratio); 1558 } else if ((InVWidth % VWidth) == 0) { 1559 // If the number of elements in the input is a multiple of the number of 1560 // elements in the output then an input element is live if the 1561 // corresponding output element is live. 1562 Ratio = InVWidth / VWidth; 1563 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx) 1564 if (DemandedElts[InIdx / Ratio]) 1565 InputDemandedElts.setBit(InIdx); 1566 } else { 1567 // Unsupported so far. 1568 break; 1569 } 1570 1571 simplifyAndSetOp(I, 0, InputDemandedElts, UndefElts2); 1572 1573 if (VWidth == InVWidth) { 1574 UndefElts = UndefElts2; 1575 } else if ((VWidth % InVWidth) == 0) { 1576 // If the number of elements in the output is a multiple of the number of 1577 // elements in the input then an output element is undef if the 1578 // corresponding input element is undef. 1579 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) 1580 if (UndefElts2[OutIdx / Ratio]) 1581 UndefElts.setBit(OutIdx); 1582 } else if ((InVWidth % VWidth) == 0) { 1583 // If the number of elements in the input is a multiple of the number of 1584 // elements in the output then an output element is undef if all of the 1585 // corresponding input elements are undef. 1586 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) { 1587 APInt SubUndef = UndefElts2.lshr(OutIdx * Ratio).zextOrTrunc(Ratio); 1588 if (SubUndef.countPopulation() == Ratio) 1589 UndefElts.setBit(OutIdx); 1590 } 1591 } else { 1592 llvm_unreachable("Unimp"); 1593 } 1594 break; 1595 } 1596 case Instruction::FPTrunc: 1597 case Instruction::FPExt: 1598 simplifyAndSetOp(I, 0, DemandedElts, UndefElts); 1599 break; 1600 1601 case Instruction::Call: { 1602 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I); 1603 if (!II) break; 1604 switch (II->getIntrinsicID()) { 1605 case Intrinsic::masked_gather: // fallthrough 1606 case Intrinsic::masked_load: { 1607 // Subtlety: If we load from a pointer, the pointer must be valid 1608 // regardless of whether the element is demanded. Doing otherwise risks 1609 // segfaults which didn't exist in the original program. 1610 APInt DemandedPtrs(APInt::getAllOnesValue(VWidth)), 1611 DemandedPassThrough(DemandedElts); 1612 if (auto *CV = dyn_cast<ConstantVector>(II->getOperand(2))) 1613 for (unsigned i = 0; i < VWidth; i++) { 1614 Constant *CElt = CV->getAggregateElement(i); 1615 if (CElt->isNullValue()) 1616 DemandedPtrs.clearBit(i); 1617 else if (CElt->isAllOnesValue()) 1618 DemandedPassThrough.clearBit(i); 1619 } 1620 if (II->getIntrinsicID() == Intrinsic::masked_gather) 1621 simplifyAndSetOp(II, 0, DemandedPtrs, UndefElts2); 1622 simplifyAndSetOp(II, 3, DemandedPassThrough, UndefElts3); 1623 1624 // Output elements are undefined if the element from both sources are. 1625 // TODO: can strengthen via mask as well. 1626 UndefElts = UndefElts2 & UndefElts3; 1627 break; 1628 } 1629 case Intrinsic::x86_xop_vfrcz_ss: 1630 case Intrinsic::x86_xop_vfrcz_sd: 1631 // The instructions for these intrinsics are speced to zero upper bits not 1632 // pass them through like other scalar intrinsics. So we shouldn't just 1633 // use Arg0 if DemandedElts[0] is clear like we do for other intrinsics. 1634 // Instead we should return a zero vector. 1635 if (!DemandedElts[0]) { 1636 Worklist.push(II); 1637 return ConstantAggregateZero::get(II->getType()); 1638 } 1639 1640 // Only the lower element is used. 1641 DemandedElts = 1; 1642 simplifyAndSetOp(II, 0, DemandedElts, UndefElts); 1643 1644 // Only the lower element is undefined. The high elements are zero. 1645 UndefElts = UndefElts[0]; 1646 break; 1647 1648 // Unary scalar-as-vector operations that work column-wise. 1649 case Intrinsic::x86_sse_rcp_ss: 1650 case Intrinsic::x86_sse_rsqrt_ss: 1651 simplifyAndSetOp(II, 0, DemandedElts, UndefElts); 1652 1653 // If lowest element of a scalar op isn't used then use Arg0. 1654 if (!DemandedElts[0]) { 1655 Worklist.push(II); 1656 return II->getArgOperand(0); 1657 } 1658 // TODO: If only low elt lower SQRT to FSQRT (with rounding/exceptions 1659 // checks). 1660 break; 1661 1662 // Binary scalar-as-vector operations that work column-wise. The high 1663 // elements come from operand 0. The low element is a function of both 1664 // operands. 1665 case Intrinsic::x86_sse_min_ss: 1666 case Intrinsic::x86_sse_max_ss: 1667 case Intrinsic::x86_sse_cmp_ss: 1668 case Intrinsic::x86_sse2_min_sd: 1669 case Intrinsic::x86_sse2_max_sd: 1670 case Intrinsic::x86_sse2_cmp_sd: { 1671 simplifyAndSetOp(II, 0, DemandedElts, UndefElts); 1672 1673 // If lowest element of a scalar op isn't used then use Arg0. 1674 if (!DemandedElts[0]) { 1675 Worklist.push(II); 1676 return II->getArgOperand(0); 1677 } 1678 1679 // Only lower element is used for operand 1. 1680 DemandedElts = 1; 1681 simplifyAndSetOp(II, 1, DemandedElts, UndefElts2); 1682 1683 // Lower element is undefined if both lower elements are undefined. 1684 // Consider things like undef&0. The result is known zero, not undef. 1685 if (!UndefElts2[0]) 1686 UndefElts.clearBit(0); 1687 1688 break; 1689 } 1690 1691 // Binary scalar-as-vector operations that work column-wise. The high 1692 // elements come from operand 0 and the low element comes from operand 1. 1693 case Intrinsic::x86_sse41_round_ss: 1694 case Intrinsic::x86_sse41_round_sd: { 1695 // Don't use the low element of operand 0. 1696 APInt DemandedElts2 = DemandedElts; 1697 DemandedElts2.clearBit(0); 1698 simplifyAndSetOp(II, 0, DemandedElts2, UndefElts); 1699 1700 // If lowest element of a scalar op isn't used then use Arg0. 1701 if (!DemandedElts[0]) { 1702 Worklist.push(II); 1703 return II->getArgOperand(0); 1704 } 1705 1706 // Only lower element is used for operand 1. 1707 DemandedElts = 1; 1708 simplifyAndSetOp(II, 1, DemandedElts, UndefElts2); 1709 1710 // Take the high undef elements from operand 0 and take the lower element 1711 // from operand 1. 1712 UndefElts.clearBit(0); 1713 UndefElts |= UndefElts2[0]; 1714 break; 1715 } 1716 1717 // Three input scalar-as-vector operations that work column-wise. The high 1718 // elements come from operand 0 and the low element is a function of all 1719 // three inputs. 1720 case Intrinsic::x86_avx512_mask_add_ss_round: 1721 case Intrinsic::x86_avx512_mask_div_ss_round: 1722 case Intrinsic::x86_avx512_mask_mul_ss_round: 1723 case Intrinsic::x86_avx512_mask_sub_ss_round: 1724 case Intrinsic::x86_avx512_mask_max_ss_round: 1725 case Intrinsic::x86_avx512_mask_min_ss_round: 1726 case Intrinsic::x86_avx512_mask_add_sd_round: 1727 case Intrinsic::x86_avx512_mask_div_sd_round: 1728 case Intrinsic::x86_avx512_mask_mul_sd_round: 1729 case Intrinsic::x86_avx512_mask_sub_sd_round: 1730 case Intrinsic::x86_avx512_mask_max_sd_round: 1731 case Intrinsic::x86_avx512_mask_min_sd_round: 1732 simplifyAndSetOp(II, 0, DemandedElts, UndefElts); 1733 1734 // If lowest element of a scalar op isn't used then use Arg0. 1735 if (!DemandedElts[0]) { 1736 Worklist.push(II); 1737 return II->getArgOperand(0); 1738 } 1739 1740 // Only lower element is used for operand 1 and 2. 1741 DemandedElts = 1; 1742 simplifyAndSetOp(II, 1, DemandedElts, UndefElts2); 1743 simplifyAndSetOp(II, 2, DemandedElts, UndefElts3); 1744 1745 // Lower element is undefined if all three lower elements are undefined. 1746 // Consider things like undef&0. The result is known zero, not undef. 1747 if (!UndefElts2[0] || !UndefElts3[0]) 1748 UndefElts.clearBit(0); 1749 1750 break; 1751 1752 case Intrinsic::x86_sse2_packssdw_128: 1753 case Intrinsic::x86_sse2_packsswb_128: 1754 case Intrinsic::x86_sse2_packuswb_128: 1755 case Intrinsic::x86_sse41_packusdw: 1756 case Intrinsic::x86_avx2_packssdw: 1757 case Intrinsic::x86_avx2_packsswb: 1758 case Intrinsic::x86_avx2_packusdw: 1759 case Intrinsic::x86_avx2_packuswb: 1760 case Intrinsic::x86_avx512_packssdw_512: 1761 case Intrinsic::x86_avx512_packsswb_512: 1762 case Intrinsic::x86_avx512_packusdw_512: 1763 case Intrinsic::x86_avx512_packuswb_512: { 1764 auto *Ty0 = II->getArgOperand(0)->getType(); 1765 unsigned InnerVWidth = cast<VectorType>(Ty0)->getNumElements(); 1766 assert(VWidth == (InnerVWidth * 2) && "Unexpected input size"); 1767 1768 unsigned NumLanes = Ty0->getPrimitiveSizeInBits() / 128; 1769 unsigned VWidthPerLane = VWidth / NumLanes; 1770 unsigned InnerVWidthPerLane = InnerVWidth / NumLanes; 1771 1772 // Per lane, pack the elements of the first input and then the second. 1773 // e.g. 1774 // v8i16 PACK(v4i32 X, v4i32 Y) - (X[0..3],Y[0..3]) 1775 // v32i8 PACK(v16i16 X, v16i16 Y) - (X[0..7],Y[0..7]),(X[8..15],Y[8..15]) 1776 for (int OpNum = 0; OpNum != 2; ++OpNum) { 1777 APInt OpDemandedElts(InnerVWidth, 0); 1778 for (unsigned Lane = 0; Lane != NumLanes; ++Lane) { 1779 unsigned LaneIdx = Lane * VWidthPerLane; 1780 for (unsigned Elt = 0; Elt != InnerVWidthPerLane; ++Elt) { 1781 unsigned Idx = LaneIdx + Elt + InnerVWidthPerLane * OpNum; 1782 if (DemandedElts[Idx]) 1783 OpDemandedElts.setBit((Lane * InnerVWidthPerLane) + Elt); 1784 } 1785 } 1786 1787 // Demand elements from the operand. 1788 APInt OpUndefElts(InnerVWidth, 0); 1789 simplifyAndSetOp(II, OpNum, OpDemandedElts, OpUndefElts); 1790 1791 // Pack the operand's UNDEF elements, one lane at a time. 1792 OpUndefElts = OpUndefElts.zext(VWidth); 1793 for (unsigned Lane = 0; Lane != NumLanes; ++Lane) { 1794 APInt LaneElts = OpUndefElts.lshr(InnerVWidthPerLane * Lane); 1795 LaneElts = LaneElts.getLoBits(InnerVWidthPerLane); 1796 LaneElts <<= InnerVWidthPerLane * (2 * Lane + OpNum); 1797 UndefElts |= LaneElts; 1798 } 1799 } 1800 break; 1801 } 1802 1803 // PSHUFB 1804 case Intrinsic::x86_ssse3_pshuf_b_128: 1805 case Intrinsic::x86_avx2_pshuf_b: 1806 case Intrinsic::x86_avx512_pshuf_b_512: 1807 // PERMILVAR 1808 case Intrinsic::x86_avx_vpermilvar_ps: 1809 case Intrinsic::x86_avx_vpermilvar_ps_256: 1810 case Intrinsic::x86_avx512_vpermilvar_ps_512: 1811 case Intrinsic::x86_avx_vpermilvar_pd: 1812 case Intrinsic::x86_avx_vpermilvar_pd_256: 1813 case Intrinsic::x86_avx512_vpermilvar_pd_512: 1814 // PERMV 1815 case Intrinsic::x86_avx2_permd: 1816 case Intrinsic::x86_avx2_permps: { 1817 simplifyAndSetOp(II, 1, DemandedElts, UndefElts); 1818 break; 1819 } 1820 1821 // SSE4A instructions leave the upper 64-bits of the 128-bit result 1822 // in an undefined state. 1823 case Intrinsic::x86_sse4a_extrq: 1824 case Intrinsic::x86_sse4a_extrqi: 1825 case Intrinsic::x86_sse4a_insertq: 1826 case Intrinsic::x86_sse4a_insertqi: 1827 UndefElts.setHighBits(VWidth / 2); 1828 break; 1829 case Intrinsic::amdgcn_buffer_load: 1830 case Intrinsic::amdgcn_buffer_load_format: 1831 case Intrinsic::amdgcn_raw_buffer_load: 1832 case Intrinsic::amdgcn_raw_buffer_load_format: 1833 case Intrinsic::amdgcn_raw_tbuffer_load: 1834 case Intrinsic::amdgcn_s_buffer_load: 1835 case Intrinsic::amdgcn_struct_buffer_load: 1836 case Intrinsic::amdgcn_struct_buffer_load_format: 1837 case Intrinsic::amdgcn_struct_tbuffer_load: 1838 case Intrinsic::amdgcn_tbuffer_load: 1839 return simplifyAMDGCNMemoryIntrinsicDemanded(II, DemandedElts); 1840 default: { 1841 if (getAMDGPUImageDMaskIntrinsic(II->getIntrinsicID())) 1842 return simplifyAMDGCNMemoryIntrinsicDemanded(II, DemandedElts, 0); 1843 1844 break; 1845 } 1846 } // switch on IntrinsicID 1847 break; 1848 } // case Call 1849 } // switch on Opcode 1850 1851 // TODO: We bail completely on integer div/rem and shifts because they have 1852 // UB/poison potential, but that should be refined. 1853 BinaryOperator *BO; 1854 if (match(I, m_BinOp(BO)) && !BO->isIntDivRem() && !BO->isShift()) { 1855 simplifyAndSetOp(I, 0, DemandedElts, UndefElts); 1856 simplifyAndSetOp(I, 1, DemandedElts, UndefElts2); 1857 1858 // Any change to an instruction with potential poison must clear those flags 1859 // because we can not guarantee those constraints now. Other analysis may 1860 // determine that it is safe to re-apply the flags. 1861 if (MadeChange) 1862 BO->dropPoisonGeneratingFlags(); 1863 1864 // Output elements are undefined if both are undefined. Consider things 1865 // like undef & 0. The result is known zero, not undef. 1866 UndefElts &= UndefElts2; 1867 } 1868 1869 // If we've proven all of the lanes undef, return an undef value. 1870 // TODO: Intersect w/demanded lanes 1871 if (UndefElts.isAllOnesValue()) 1872 return UndefValue::get(I->getType());; 1873 1874 return MadeChange ? I : nullptr; 1875 } 1876