1 #include "llvm/Transforms/Utils/VNCoercion.h" 2 #include "llvm/Analysis/ConstantFolding.h" 3 #include "llvm/Analysis/ValueTracking.h" 4 #include "llvm/IR/IRBuilder.h" 5 #include "llvm/IR/IntrinsicInst.h" 6 #include "llvm/Support/Debug.h" 7 8 #define DEBUG_TYPE "vncoerce" 9 10 namespace llvm { 11 namespace VNCoercion { 12 13 static bool isFirstClassAggregateOrScalableType(Type *Ty) { 14 return Ty->isStructTy() || Ty->isArrayTy() || isa<ScalableVectorType>(Ty); 15 } 16 17 /// Return true if coerceAvailableValueToLoadType will succeed. 18 bool canCoerceMustAliasedValueToLoad(Value *StoredVal, Type *LoadTy, 19 const DataLayout &DL) { 20 Type *StoredTy = StoredVal->getType(); 21 22 if (StoredTy == LoadTy) 23 return true; 24 25 // If the loaded/stored value is a first class array/struct, or scalable type, 26 // don't try to transform them. We need to be able to bitcast to integer. 27 if (isFirstClassAggregateOrScalableType(LoadTy) || 28 isFirstClassAggregateOrScalableType(StoredTy)) 29 return false; 30 31 uint64_t StoreSize = DL.getTypeSizeInBits(StoredTy).getFixedValue(); 32 33 // The store size must be byte-aligned to support future type casts. 34 if (llvm::alignTo(StoreSize, 8) != StoreSize) 35 return false; 36 37 // The store has to be at least as big as the load. 38 if (StoreSize < DL.getTypeSizeInBits(LoadTy).getFixedValue()) 39 return false; 40 41 bool StoredNI = DL.isNonIntegralPointerType(StoredTy->getScalarType()); 42 bool LoadNI = DL.isNonIntegralPointerType(LoadTy->getScalarType()); 43 // Don't coerce non-integral pointers to integers or vice versa. 44 if (StoredNI != LoadNI) { 45 // As a special case, allow coercion of memset used to initialize 46 // an array w/null. Despite non-integral pointers not generally having a 47 // specific bit pattern, we do assume null is zero. 48 if (auto *CI = dyn_cast<Constant>(StoredVal)) 49 return CI->isNullValue(); 50 return false; 51 } else if (StoredNI && LoadNI && 52 StoredTy->getPointerAddressSpace() != 53 LoadTy->getPointerAddressSpace()) { 54 return false; 55 } 56 57 58 // The implementation below uses inttoptr for vectors of unequal size; we 59 // can't allow this for non integral pointers. We could teach it to extract 60 // exact subvectors if desired. 61 if (StoredNI && StoreSize != DL.getTypeSizeInBits(LoadTy).getFixedValue()) 62 return false; 63 64 if (StoredTy->isTargetExtTy() || LoadTy->isTargetExtTy()) 65 return false; 66 67 return true; 68 } 69 70 /// If we saw a store of a value to memory, and 71 /// then a load from a must-aliased pointer of a different type, try to coerce 72 /// the stored value. LoadedTy is the type of the load we want to replace. 73 /// IRB is IRBuilder used to insert new instructions. 74 /// 75 /// If we can't do it, return null. 76 Value *coerceAvailableValueToLoadType(Value *StoredVal, Type *LoadedTy, 77 IRBuilderBase &Helper, 78 const DataLayout &DL) { 79 assert(canCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, DL) && 80 "precondition violation - materialization can't fail"); 81 if (auto *C = dyn_cast<Constant>(StoredVal)) 82 StoredVal = ConstantFoldConstant(C, DL); 83 84 // If this is already the right type, just return it. 85 Type *StoredValTy = StoredVal->getType(); 86 87 uint64_t StoredValSize = DL.getTypeSizeInBits(StoredValTy).getFixedValue(); 88 uint64_t LoadedValSize = DL.getTypeSizeInBits(LoadedTy).getFixedValue(); 89 90 // If the store and reload are the same size, we can always reuse it. 91 if (StoredValSize == LoadedValSize) { 92 // Pointer to Pointer -> use bitcast. 93 if (StoredValTy->isPtrOrPtrVectorTy() && LoadedTy->isPtrOrPtrVectorTy()) { 94 StoredVal = Helper.CreateBitCast(StoredVal, LoadedTy); 95 } else { 96 // Convert source pointers to integers, which can be bitcast. 97 if (StoredValTy->isPtrOrPtrVectorTy()) { 98 StoredValTy = DL.getIntPtrType(StoredValTy); 99 StoredVal = Helper.CreatePtrToInt(StoredVal, StoredValTy); 100 } 101 102 Type *TypeToCastTo = LoadedTy; 103 if (TypeToCastTo->isPtrOrPtrVectorTy()) 104 TypeToCastTo = DL.getIntPtrType(TypeToCastTo); 105 106 if (StoredValTy != TypeToCastTo) 107 StoredVal = Helper.CreateBitCast(StoredVal, TypeToCastTo); 108 109 // Cast to pointer if the load needs a pointer type. 110 if (LoadedTy->isPtrOrPtrVectorTy()) 111 StoredVal = Helper.CreateIntToPtr(StoredVal, LoadedTy); 112 } 113 114 if (auto *C = dyn_cast<ConstantExpr>(StoredVal)) 115 StoredVal = ConstantFoldConstant(C, DL); 116 117 return StoredVal; 118 } 119 // If the loaded value is smaller than the available value, then we can 120 // extract out a piece from it. If the available value is too small, then we 121 // can't do anything. 122 assert(StoredValSize >= LoadedValSize && 123 "canCoerceMustAliasedValueToLoad fail"); 124 125 // Convert source pointers to integers, which can be manipulated. 126 if (StoredValTy->isPtrOrPtrVectorTy()) { 127 StoredValTy = DL.getIntPtrType(StoredValTy); 128 StoredVal = Helper.CreatePtrToInt(StoredVal, StoredValTy); 129 } 130 131 // Convert vectors and fp to integer, which can be manipulated. 132 if (!StoredValTy->isIntegerTy()) { 133 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoredValSize); 134 StoredVal = Helper.CreateBitCast(StoredVal, StoredValTy); 135 } 136 137 // If this is a big-endian system, we need to shift the value down to the low 138 // bits so that a truncate will work. 139 if (DL.isBigEndian()) { 140 uint64_t ShiftAmt = DL.getTypeStoreSizeInBits(StoredValTy).getFixedValue() - 141 DL.getTypeStoreSizeInBits(LoadedTy).getFixedValue(); 142 StoredVal = Helper.CreateLShr( 143 StoredVal, ConstantInt::get(StoredVal->getType(), ShiftAmt)); 144 } 145 146 // Truncate the integer to the right size now. 147 Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadedValSize); 148 StoredVal = Helper.CreateTruncOrBitCast(StoredVal, NewIntTy); 149 150 if (LoadedTy != NewIntTy) { 151 // If the result is a pointer, inttoptr. 152 if (LoadedTy->isPtrOrPtrVectorTy()) 153 StoredVal = Helper.CreateIntToPtr(StoredVal, LoadedTy); 154 else 155 // Otherwise, bitcast. 156 StoredVal = Helper.CreateBitCast(StoredVal, LoadedTy); 157 } 158 159 if (auto *C = dyn_cast<Constant>(StoredVal)) 160 StoredVal = ConstantFoldConstant(C, DL); 161 162 return StoredVal; 163 } 164 165 /// This function is called when we have a memdep query of a load that ends up 166 /// being a clobbering memory write (store, memset, memcpy, memmove). This 167 /// means that the write *may* provide bits used by the load but we can't be 168 /// sure because the pointers don't must-alias. 169 /// 170 /// Check this case to see if there is anything more we can do before we give 171 /// up. This returns -1 if we have to give up, or a byte number in the stored 172 /// value of the piece that feeds the load. 173 static int analyzeLoadFromClobberingWrite(Type *LoadTy, Value *LoadPtr, 174 Value *WritePtr, 175 uint64_t WriteSizeInBits, 176 const DataLayout &DL) { 177 // If the loaded/stored value is a first class array/struct, or scalable type, 178 // don't try to transform them. We need to be able to bitcast to integer. 179 if (isFirstClassAggregateOrScalableType(LoadTy)) 180 return -1; 181 182 int64_t StoreOffset = 0, LoadOffset = 0; 183 Value *StoreBase = 184 GetPointerBaseWithConstantOffset(WritePtr, StoreOffset, DL); 185 Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, DL); 186 if (StoreBase != LoadBase) 187 return -1; 188 189 uint64_t LoadSize = DL.getTypeSizeInBits(LoadTy).getFixedValue(); 190 191 if ((WriteSizeInBits & 7) | (LoadSize & 7)) 192 return -1; 193 uint64_t StoreSize = WriteSizeInBits / 8; // Convert to bytes. 194 LoadSize /= 8; 195 196 // If the Load isn't completely contained within the stored bits, we don't 197 // have all the bits to feed it. We could do something crazy in the future 198 // (issue a smaller load then merge the bits in) but this seems unlikely to be 199 // valuable. 200 if (StoreOffset > LoadOffset || 201 StoreOffset + int64_t(StoreSize) < LoadOffset + int64_t(LoadSize)) 202 return -1; 203 204 // Okay, we can do this transformation. Return the number of bytes into the 205 // store that the load is. 206 return LoadOffset - StoreOffset; 207 } 208 209 /// This function is called when we have a 210 /// memdep query of a load that ends up being a clobbering store. 211 int analyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr, 212 StoreInst *DepSI, const DataLayout &DL) { 213 auto *StoredVal = DepSI->getValueOperand(); 214 215 // Cannot handle reading from store of first-class aggregate or scalable type. 216 if (isFirstClassAggregateOrScalableType(StoredVal->getType())) 217 return -1; 218 219 if (!canCoerceMustAliasedValueToLoad(StoredVal, LoadTy, DL)) 220 return -1; 221 222 Value *StorePtr = DepSI->getPointerOperand(); 223 uint64_t StoreSize = 224 DL.getTypeSizeInBits(DepSI->getValueOperand()->getType()).getFixedValue(); 225 return analyzeLoadFromClobberingWrite(LoadTy, LoadPtr, StorePtr, StoreSize, 226 DL); 227 } 228 229 /// This function is called when we have a 230 /// memdep query of a load that ends up being clobbered by another load. See if 231 /// the other load can feed into the second load. 232 int analyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr, LoadInst *DepLI, 233 const DataLayout &DL) { 234 // Cannot handle reading from store of first-class aggregate yet. 235 if (DepLI->getType()->isStructTy() || DepLI->getType()->isArrayTy()) 236 return -1; 237 238 if (!canCoerceMustAliasedValueToLoad(DepLI, LoadTy, DL)) 239 return -1; 240 241 Value *DepPtr = DepLI->getPointerOperand(); 242 uint64_t DepSize = DL.getTypeSizeInBits(DepLI->getType()).getFixedValue(); 243 return analyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, DepSize, DL); 244 } 245 246 int analyzeLoadFromClobberingMemInst(Type *LoadTy, Value *LoadPtr, 247 MemIntrinsic *MI, const DataLayout &DL) { 248 // If the mem operation is a non-constant size, we can't handle it. 249 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength()); 250 if (!SizeCst) 251 return -1; 252 uint64_t MemSizeInBits = SizeCst->getZExtValue() * 8; 253 254 // If this is memset, we just need to see if the offset is valid in the size 255 // of the memset.. 256 if (const auto *memset_inst = dyn_cast<MemSetInst>(MI)) { 257 if (DL.isNonIntegralPointerType(LoadTy->getScalarType())) { 258 auto *CI = dyn_cast<ConstantInt>(memset_inst->getValue()); 259 if (!CI || !CI->isZero()) 260 return -1; 261 } 262 return analyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(), 263 MemSizeInBits, DL); 264 } 265 266 // If we have a memcpy/memmove, the only case we can handle is if this is a 267 // copy from constant memory. In that case, we can read directly from the 268 // constant memory. 269 MemTransferInst *MTI = cast<MemTransferInst>(MI); 270 271 Constant *Src = dyn_cast<Constant>(MTI->getSource()); 272 if (!Src) 273 return -1; 274 275 GlobalVariable *GV = dyn_cast<GlobalVariable>(getUnderlyingObject(Src)); 276 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer()) 277 return -1; 278 279 // See if the access is within the bounds of the transfer. 280 int Offset = analyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(), 281 MemSizeInBits, DL); 282 if (Offset == -1) 283 return Offset; 284 285 // Otherwise, see if we can constant fold a load from the constant with the 286 // offset applied as appropriate. 287 unsigned IndexSize = DL.getIndexTypeSizeInBits(Src->getType()); 288 if (ConstantFoldLoadFromConstPtr(Src, LoadTy, APInt(IndexSize, Offset), DL)) 289 return Offset; 290 return -1; 291 } 292 293 static Value *getStoreValueForLoadHelper(Value *SrcVal, unsigned Offset, 294 Type *LoadTy, IRBuilderBase &Builder, 295 const DataLayout &DL) { 296 LLVMContext &Ctx = SrcVal->getType()->getContext(); 297 298 // If two pointers are in the same address space, they have the same size, 299 // so we don't need to do any truncation, etc. This avoids introducing 300 // ptrtoint instructions for pointers that may be non-integral. 301 if (SrcVal->getType()->isPointerTy() && LoadTy->isPointerTy() && 302 cast<PointerType>(SrcVal->getType())->getAddressSpace() == 303 cast<PointerType>(LoadTy)->getAddressSpace()) { 304 return SrcVal; 305 } 306 307 uint64_t StoreSize = 308 (DL.getTypeSizeInBits(SrcVal->getType()).getFixedValue() + 7) / 8; 309 uint64_t LoadSize = (DL.getTypeSizeInBits(LoadTy).getFixedValue() + 7) / 8; 310 // Compute which bits of the stored value are being used by the load. Convert 311 // to an integer type to start with. 312 if (SrcVal->getType()->isPtrOrPtrVectorTy()) 313 SrcVal = 314 Builder.CreatePtrToInt(SrcVal, DL.getIntPtrType(SrcVal->getType())); 315 if (!SrcVal->getType()->isIntegerTy()) 316 SrcVal = 317 Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize * 8)); 318 319 // Shift the bits to the least significant depending on endianness. 320 unsigned ShiftAmt; 321 if (DL.isLittleEndian()) 322 ShiftAmt = Offset * 8; 323 else 324 ShiftAmt = (StoreSize - LoadSize - Offset) * 8; 325 if (ShiftAmt) 326 SrcVal = Builder.CreateLShr(SrcVal, 327 ConstantInt::get(SrcVal->getType(), ShiftAmt)); 328 329 if (LoadSize != StoreSize) 330 SrcVal = Builder.CreateTruncOrBitCast(SrcVal, 331 IntegerType::get(Ctx, LoadSize * 8)); 332 return SrcVal; 333 } 334 335 Value *getValueForLoad(Value *SrcVal, unsigned Offset, Type *LoadTy, 336 Instruction *InsertPt, const DataLayout &DL) { 337 338 #ifndef NDEBUG 339 unsigned SrcValSize = DL.getTypeStoreSize(SrcVal->getType()).getFixedValue(); 340 unsigned LoadSize = DL.getTypeStoreSize(LoadTy).getFixedValue(); 341 assert(Offset + LoadSize <= SrcValSize); 342 #endif 343 IRBuilder<> Builder(InsertPt); 344 SrcVal = getStoreValueForLoadHelper(SrcVal, Offset, LoadTy, Builder, DL); 345 return coerceAvailableValueToLoadType(SrcVal, LoadTy, Builder, DL); 346 } 347 348 Constant *getConstantValueForLoad(Constant *SrcVal, unsigned Offset, 349 Type *LoadTy, const DataLayout &DL) { 350 #ifndef NDEBUG 351 unsigned SrcValSize = DL.getTypeStoreSize(SrcVal->getType()).getFixedValue(); 352 unsigned LoadSize = DL.getTypeStoreSize(LoadTy).getFixedValue(); 353 assert(Offset + LoadSize <= SrcValSize); 354 #endif 355 return ConstantFoldLoadFromConst(SrcVal, LoadTy, APInt(32, Offset), DL); 356 } 357 358 /// This function is called when we have a 359 /// memdep query of a load that ends up being a clobbering mem intrinsic. 360 Value *getMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset, 361 Type *LoadTy, Instruction *InsertPt, 362 const DataLayout &DL) { 363 LLVMContext &Ctx = LoadTy->getContext(); 364 uint64_t LoadSize = DL.getTypeSizeInBits(LoadTy).getFixedValue() / 8; 365 IRBuilder<> Builder(InsertPt); 366 367 // We know that this method is only called when the mem transfer fully 368 // provides the bits for the load. 369 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) { 370 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and 371 // independently of what the offset is. 372 Value *Val = MSI->getValue(); 373 if (LoadSize != 1) 374 Val = 375 Builder.CreateZExtOrBitCast(Val, IntegerType::get(Ctx, LoadSize * 8)); 376 Value *OneElt = Val; 377 378 // Splat the value out to the right number of bits. 379 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize;) { 380 // If we can double the number of bytes set, do it. 381 if (NumBytesSet * 2 <= LoadSize) { 382 Value *ShVal = Builder.CreateShl( 383 Val, ConstantInt::get(Val->getType(), NumBytesSet * 8)); 384 Val = Builder.CreateOr(Val, ShVal); 385 NumBytesSet <<= 1; 386 continue; 387 } 388 389 // Otherwise insert one byte at a time. 390 Value *ShVal = 391 Builder.CreateShl(Val, ConstantInt::get(Val->getType(), 1 * 8)); 392 Val = Builder.CreateOr(OneElt, ShVal); 393 ++NumBytesSet; 394 } 395 396 return coerceAvailableValueToLoadType(Val, LoadTy, Builder, DL); 397 } 398 399 // Otherwise, this is a memcpy/memmove from a constant global. 400 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst); 401 Constant *Src = cast<Constant>(MTI->getSource()); 402 unsigned IndexSize = DL.getIndexTypeSizeInBits(Src->getType()); 403 return ConstantFoldLoadFromConstPtr(Src, LoadTy, APInt(IndexSize, Offset), 404 DL); 405 } 406 407 Constant *getConstantMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset, 408 Type *LoadTy, const DataLayout &DL) { 409 LLVMContext &Ctx = LoadTy->getContext(); 410 uint64_t LoadSize = DL.getTypeSizeInBits(LoadTy).getFixedValue() / 8; 411 412 // We know that this method is only called when the mem transfer fully 413 // provides the bits for the load. 414 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) { 415 auto *Val = dyn_cast<ConstantInt>(MSI->getValue()); 416 if (!Val) 417 return nullptr; 418 419 Val = ConstantInt::get(Ctx, APInt::getSplat(LoadSize * 8, Val->getValue())); 420 return ConstantFoldLoadFromConst(Val, LoadTy, DL); 421 } 422 423 // Otherwise, this is a memcpy/memmove from a constant global. 424 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst); 425 Constant *Src = cast<Constant>(MTI->getSource()); 426 unsigned IndexSize = DL.getIndexTypeSizeInBits(Src->getType()); 427 return ConstantFoldLoadFromConstPtr(Src, LoadTy, APInt(IndexSize, Offset), 428 DL); 429 } 430 } // namespace VNCoercion 431 } // namespace llvm 432