1 //===- Loads.cpp - Local load analysis ------------------------------------===// 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 defines simple local analyses for load instructions. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "llvm/Analysis/Loads.h" 15 #include "llvm/Analysis/AliasAnalysis.h" 16 #include "llvm/Analysis/ValueTracking.h" 17 #include "llvm/IR/DataLayout.h" 18 #include "llvm/IR/GlobalAlias.h" 19 #include "llvm/IR/GlobalVariable.h" 20 #include "llvm/IR/IntrinsicInst.h" 21 #include "llvm/IR/LLVMContext.h" 22 #include "llvm/IR/Module.h" 23 #include "llvm/IR/Operator.h" 24 #include "llvm/IR/Statepoint.h" 25 26 using namespace llvm; 27 28 static bool isAligned(const Value *Base, const APInt &Offset, unsigned Align, 29 const DataLayout &DL) { 30 APInt BaseAlign(Offset.getBitWidth(), Base->getPointerAlignment(DL)); 31 32 if (!BaseAlign) { 33 Type *Ty = Base->getType()->getPointerElementType(); 34 if (!Ty->isSized()) 35 return false; 36 BaseAlign = DL.getABITypeAlignment(Ty); 37 } 38 39 APInt Alignment(Offset.getBitWidth(), Align); 40 41 assert(Alignment.isPowerOf2() && "must be a power of 2!"); 42 return BaseAlign.uge(Alignment) && !(Offset & (Alignment-1)); 43 } 44 45 static bool isAligned(const Value *Base, unsigned Align, const DataLayout &DL) { 46 Type *Ty = Base->getType(); 47 assert(Ty->isSized() && "must be sized"); 48 APInt Offset(DL.getTypeStoreSizeInBits(Ty), 0); 49 return isAligned(Base, Offset, Align, DL); 50 } 51 52 /// Test if V is always a pointer to allocated and suitably aligned memory for 53 /// a simple load or store. 54 static bool isDereferenceableAndAlignedPointer( 55 const Value *V, unsigned Align, const APInt &Size, const DataLayout &DL, 56 const Instruction *CtxI, const DominatorTree *DT, 57 SmallPtrSetImpl<const Value *> &Visited) { 58 // Note that it is not safe to speculate into a malloc'd region because 59 // malloc may return null. 60 61 // bitcast instructions are no-ops as far as dereferenceability is concerned. 62 if (const BitCastOperator *BC = dyn_cast<BitCastOperator>(V)) 63 return isDereferenceableAndAlignedPointer(BC->getOperand(0), Align, Size, 64 DL, CtxI, DT, Visited); 65 66 bool CheckForNonNull = false; 67 APInt KnownDerefBytes(Size.getBitWidth(), 68 V->getPointerDereferenceableBytes(DL, CheckForNonNull)); 69 if (KnownDerefBytes.getBoolValue()) { 70 if (KnownDerefBytes.uge(Size)) 71 if (!CheckForNonNull || isKnownNonNullAt(V, CtxI, DT)) 72 return isAligned(V, Align, DL); 73 } 74 75 // For GEPs, determine if the indexing lands within the allocated object. 76 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) { 77 const Value *Base = GEP->getPointerOperand(); 78 79 APInt Offset(DL.getPointerTypeSizeInBits(GEP->getType()), 0); 80 if (!GEP->accumulateConstantOffset(DL, Offset) || Offset.isNegative() || 81 !Offset.urem(APInt(Offset.getBitWidth(), Align)).isMinValue()) 82 return false; 83 84 // If the base pointer is dereferenceable for Offset+Size bytes, then the 85 // GEP (== Base + Offset) is dereferenceable for Size bytes. If the base 86 // pointer is aligned to Align bytes, and the Offset is divisible by Align 87 // then the GEP (== Base + Offset == k_0 * Align + k_1 * Align) is also 88 // aligned to Align bytes. 89 90 return Visited.insert(Base).second && 91 isDereferenceableAndAlignedPointer(Base, Align, Offset + Size, DL, 92 CtxI, DT, Visited); 93 } 94 95 // For gc.relocate, look through relocations 96 if (const GCRelocateInst *RelocateInst = dyn_cast<GCRelocateInst>(V)) 97 return isDereferenceableAndAlignedPointer( 98 RelocateInst->getDerivedPtr(), Align, Size, DL, CtxI, DT, Visited); 99 100 if (const AddrSpaceCastInst *ASC = dyn_cast<AddrSpaceCastInst>(V)) 101 return isDereferenceableAndAlignedPointer(ASC->getOperand(0), Align, Size, 102 DL, CtxI, DT, Visited); 103 104 if (auto CS = ImmutableCallSite(V)) 105 if (const Value *RV = CS.getReturnedArgOperand()) 106 return isDereferenceableAndAlignedPointer(RV, Align, Size, DL, CtxI, DT, 107 Visited); 108 109 // If we don't know, assume the worst. 110 return false; 111 } 112 113 bool llvm::isDereferenceableAndAlignedPointer(const Value *V, unsigned Align, 114 const DataLayout &DL, 115 const Instruction *CtxI, 116 const DominatorTree *DT) { 117 // When dereferenceability information is provided by a dereferenceable 118 // attribute, we know exactly how many bytes are dereferenceable. If we can 119 // determine the exact offset to the attributed variable, we can use that 120 // information here. 121 Type *VTy = V->getType(); 122 Type *Ty = VTy->getPointerElementType(); 123 124 // Require ABI alignment for loads without alignment specification 125 if (Align == 0) 126 Align = DL.getABITypeAlignment(Ty); 127 128 if (!Ty->isSized()) 129 return false; 130 131 SmallPtrSet<const Value *, 32> Visited; 132 return ::isDereferenceableAndAlignedPointer( 133 V, Align, APInt(DL.getTypeSizeInBits(VTy), DL.getTypeStoreSize(Ty)), DL, 134 CtxI, DT, Visited); 135 } 136 137 bool llvm::isDereferenceablePointer(const Value *V, const DataLayout &DL, 138 const Instruction *CtxI, 139 const DominatorTree *DT) { 140 return isDereferenceableAndAlignedPointer(V, 1, DL, CtxI, DT); 141 } 142 143 /// \brief Test if A and B will obviously have the same value. 144 /// 145 /// This includes recognizing that %t0 and %t1 will have the same 146 /// value in code like this: 147 /// \code 148 /// %t0 = getelementptr \@a, 0, 3 149 /// store i32 0, i32* %t0 150 /// %t1 = getelementptr \@a, 0, 3 151 /// %t2 = load i32* %t1 152 /// \endcode 153 /// 154 static bool AreEquivalentAddressValues(const Value *A, const Value *B) { 155 // Test if the values are trivially equivalent. 156 if (A == B) 157 return true; 158 159 // Test if the values come from identical arithmetic instructions. 160 // Use isIdenticalToWhenDefined instead of isIdenticalTo because 161 // this function is only used when one address use dominates the 162 // other, which means that they'll always either have the same 163 // value or one of them will have an undefined value. 164 if (isa<BinaryOperator>(A) || isa<CastInst>(A) || isa<PHINode>(A) || 165 isa<GetElementPtrInst>(A)) 166 if (const Instruction *BI = dyn_cast<Instruction>(B)) 167 if (cast<Instruction>(A)->isIdenticalToWhenDefined(BI)) 168 return true; 169 170 // Otherwise they may not be equivalent. 171 return false; 172 } 173 174 /// \brief Check if executing a load of this pointer value cannot trap. 175 /// 176 /// If DT and ScanFrom are specified this method performs context-sensitive 177 /// analysis and returns true if it is safe to load immediately before ScanFrom. 178 /// 179 /// If it is not obviously safe to load from the specified pointer, we do 180 /// a quick local scan of the basic block containing \c ScanFrom, to determine 181 /// if the address is already accessed. 182 /// 183 /// This uses the pointee type to determine how many bytes need to be safe to 184 /// load from the pointer. 185 bool llvm::isSafeToLoadUnconditionally(Value *V, unsigned Align, 186 const DataLayout &DL, 187 Instruction *ScanFrom, 188 const DominatorTree *DT) { 189 // Zero alignment means that the load has the ABI alignment for the target 190 if (Align == 0) 191 Align = DL.getABITypeAlignment(V->getType()->getPointerElementType()); 192 assert(isPowerOf2_32(Align)); 193 194 // If DT is not specified we can't make context-sensitive query 195 const Instruction* CtxI = DT ? ScanFrom : nullptr; 196 if (isDereferenceableAndAlignedPointer(V, Align, DL, CtxI, DT)) 197 return true; 198 199 int64_t ByteOffset = 0; 200 Value *Base = V; 201 Base = GetPointerBaseWithConstantOffset(V, ByteOffset, DL); 202 203 if (ByteOffset < 0) // out of bounds 204 return false; 205 206 Type *BaseType = nullptr; 207 unsigned BaseAlign = 0; 208 if (const AllocaInst *AI = dyn_cast<AllocaInst>(Base)) { 209 // An alloca is safe to load from as load as it is suitably aligned. 210 BaseType = AI->getAllocatedType(); 211 BaseAlign = AI->getAlignment(); 212 } else if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(Base)) { 213 // Global variables are not necessarily safe to load from if they are 214 // interposed arbitrarily. Their size may change or they may be weak and 215 // require a test to determine if they were in fact provided. 216 if (!GV->isInterposable()) { 217 BaseType = GV->getType()->getElementType(); 218 BaseAlign = GV->getAlignment(); 219 } 220 } 221 222 PointerType *AddrTy = cast<PointerType>(V->getType()); 223 uint64_t LoadSize = DL.getTypeStoreSize(AddrTy->getElementType()); 224 225 // If we found a base allocated type from either an alloca or global variable, 226 // try to see if we are definitively within the allocated region. We need to 227 // know the size of the base type and the loaded type to do anything in this 228 // case. 229 if (BaseType && BaseType->isSized()) { 230 if (BaseAlign == 0) 231 BaseAlign = DL.getPrefTypeAlignment(BaseType); 232 233 if (Align <= BaseAlign) { 234 // Check if the load is within the bounds of the underlying object. 235 if (ByteOffset + LoadSize <= DL.getTypeAllocSize(BaseType) && 236 ((ByteOffset % Align) == 0)) 237 return true; 238 } 239 } 240 241 if (!ScanFrom) 242 return false; 243 244 // Otherwise, be a little bit aggressive by scanning the local block where we 245 // want to check to see if the pointer is already being loaded or stored 246 // from/to. If so, the previous load or store would have already trapped, 247 // so there is no harm doing an extra load (also, CSE will later eliminate 248 // the load entirely). 249 BasicBlock::iterator BBI = ScanFrom->getIterator(), 250 E = ScanFrom->getParent()->begin(); 251 252 // We can at least always strip pointer casts even though we can't use the 253 // base here. 254 V = V->stripPointerCasts(); 255 256 while (BBI != E) { 257 --BBI; 258 259 // If we see a free or a call which may write to memory (i.e. which might do 260 // a free) the pointer could be marked invalid. 261 if (isa<CallInst>(BBI) && BBI->mayWriteToMemory() && 262 !isa<DbgInfoIntrinsic>(BBI)) 263 return false; 264 265 Value *AccessedPtr; 266 unsigned AccessedAlign; 267 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) { 268 AccessedPtr = LI->getPointerOperand(); 269 AccessedAlign = LI->getAlignment(); 270 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) { 271 AccessedPtr = SI->getPointerOperand(); 272 AccessedAlign = SI->getAlignment(); 273 } else 274 continue; 275 276 Type *AccessedTy = AccessedPtr->getType()->getPointerElementType(); 277 if (AccessedAlign == 0) 278 AccessedAlign = DL.getABITypeAlignment(AccessedTy); 279 if (AccessedAlign < Align) 280 continue; 281 282 // Handle trivial cases. 283 if (AccessedPtr == V) 284 return true; 285 286 if (AreEquivalentAddressValues(AccessedPtr->stripPointerCasts(), V) && 287 LoadSize <= DL.getTypeStoreSize(AccessedTy)) 288 return true; 289 } 290 return false; 291 } 292 293 /// DefMaxInstsToScan - the default number of maximum instructions 294 /// to scan in the block, used by FindAvailableLoadedValue(). 295 /// FindAvailableLoadedValue() was introduced in r60148, to improve jump 296 /// threading in part by eliminating partially redundant loads. 297 /// At that point, the value of MaxInstsToScan was already set to '6' 298 /// without documented explanation. 299 cl::opt<unsigned> 300 llvm::DefMaxInstsToScan("available-load-scan-limit", cl::init(6), cl::Hidden, 301 cl::desc("Use this to specify the default maximum number of instructions " 302 "to scan backward from a given instruction, when searching for " 303 "available loaded value")); 304 305 Value *llvm::FindAvailableLoadedValue(LoadInst *Load, BasicBlock *ScanBB, 306 BasicBlock::iterator &ScanFrom, 307 unsigned MaxInstsToScan, 308 AliasAnalysis *AA, AAMDNodes *AATags, 309 bool *IsLoadCSE) { 310 if (MaxInstsToScan == 0) 311 MaxInstsToScan = ~0U; 312 313 Value *Ptr = Load->getPointerOperand(); 314 Type *AccessTy = Load->getType(); 315 316 // We can never remove a volatile load 317 if (Load->isVolatile()) 318 return nullptr; 319 320 // Anything stronger than unordered is currently unimplemented. 321 if (!Load->isUnordered()) 322 return nullptr; 323 324 const DataLayout &DL = ScanBB->getModule()->getDataLayout(); 325 326 // Try to get the store size for the type. 327 uint64_t AccessSize = DL.getTypeStoreSize(AccessTy); 328 329 Value *StrippedPtr = Ptr->stripPointerCasts(); 330 331 while (ScanFrom != ScanBB->begin()) { 332 // We must ignore debug info directives when counting (otherwise they 333 // would affect codegen). 334 Instruction *Inst = &*--ScanFrom; 335 if (isa<DbgInfoIntrinsic>(Inst)) 336 continue; 337 338 // Restore ScanFrom to expected value in case next test succeeds 339 ScanFrom++; 340 341 // Don't scan huge blocks. 342 if (MaxInstsToScan-- == 0) 343 return nullptr; 344 345 --ScanFrom; 346 // If this is a load of Ptr, the loaded value is available. 347 // (This is true even if the load is volatile or atomic, although 348 // those cases are unlikely.) 349 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) 350 if (AreEquivalentAddressValues( 351 LI->getPointerOperand()->stripPointerCasts(), StrippedPtr) && 352 CastInst::isBitOrNoopPointerCastable(LI->getType(), AccessTy, DL)) { 353 354 // We can value forward from an atomic to a non-atomic, but not the 355 // other way around. 356 if (LI->isAtomic() < Load->isAtomic()) 357 return nullptr; 358 359 if (AATags) 360 LI->getAAMetadata(*AATags); 361 if (IsLoadCSE) 362 *IsLoadCSE = true; 363 return LI; 364 } 365 366 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 367 Value *StorePtr = SI->getPointerOperand()->stripPointerCasts(); 368 // If this is a store through Ptr, the value is available! 369 // (This is true even if the store is volatile or atomic, although 370 // those cases are unlikely.) 371 if (AreEquivalentAddressValues(StorePtr, StrippedPtr) && 372 CastInst::isBitOrNoopPointerCastable(SI->getValueOperand()->getType(), 373 AccessTy, DL)) { 374 375 // We can value forward from an atomic to a non-atomic, but not the 376 // other way around. 377 if (SI->isAtomic() < Load->isAtomic()) 378 return nullptr; 379 380 if (AATags) 381 SI->getAAMetadata(*AATags); 382 return SI->getOperand(0); 383 } 384 385 // If both StrippedPtr and StorePtr reach all the way to an alloca or 386 // global and they are different, ignore the store. This is a trivial form 387 // of alias analysis that is important for reg2mem'd code. 388 if ((isa<AllocaInst>(StrippedPtr) || isa<GlobalVariable>(StrippedPtr)) && 389 (isa<AllocaInst>(StorePtr) || isa<GlobalVariable>(StorePtr)) && 390 StrippedPtr != StorePtr) 391 continue; 392 393 // If we have alias analysis and it says the store won't modify the loaded 394 // value, ignore the store. 395 if (AA && (AA->getModRefInfo(SI, StrippedPtr, AccessSize) & MRI_Mod) == 0) 396 continue; 397 398 // Otherwise the store that may or may not alias the pointer, bail out. 399 ++ScanFrom; 400 return nullptr; 401 } 402 403 // If this is some other instruction that may clobber Ptr, bail out. 404 if (Inst->mayWriteToMemory()) { 405 // If alias analysis claims that it really won't modify the load, 406 // ignore it. 407 if (AA && 408 (AA->getModRefInfo(Inst, StrippedPtr, AccessSize) & MRI_Mod) == 0) 409 continue; 410 411 // May modify the pointer, bail out. 412 ++ScanFrom; 413 return nullptr; 414 } 415 } 416 417 // Got to the start of the block, we didn't find it, but are done for this 418 // block. 419 return nullptr; 420 } 421
