| File: | llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp |
| Warning: | line 2505, column 19 Value stored to 'Current' during its initialization is never read |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 1 | //===- DeadStoreElimination.cpp - Fast Dead Store Elimination -------------===// |
| 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 implements a trivial dead store elimination that only considers |
| 10 | // basic-block local redundant stores. |
| 11 | // |
| 12 | // FIXME: This should eventually be extended to be a post-dominator tree |
| 13 | // traversal. Doing so would be pretty trivial. |
| 14 | // |
| 15 | //===----------------------------------------------------------------------===// |
| 16 | |
| 17 | #include "llvm/Transforms/Scalar/DeadStoreElimination.h" |
| 18 | #include "llvm/ADT/APInt.h" |
| 19 | #include "llvm/ADT/DenseMap.h" |
| 20 | #include "llvm/ADT/MapVector.h" |
| 21 | #include "llvm/ADT/PostOrderIterator.h" |
| 22 | #include "llvm/ADT/SetVector.h" |
| 23 | #include "llvm/ADT/SmallPtrSet.h" |
| 24 | #include "llvm/ADT/SmallVector.h" |
| 25 | #include "llvm/ADT/Statistic.h" |
| 26 | #include "llvm/ADT/StringRef.h" |
| 27 | #include "llvm/Analysis/AliasAnalysis.h" |
| 28 | #include "llvm/Analysis/CaptureTracking.h" |
| 29 | #include "llvm/Analysis/GlobalsModRef.h" |
| 30 | #include "llvm/Analysis/MemoryBuiltins.h" |
| 31 | #include "llvm/Analysis/MemoryDependenceAnalysis.h" |
| 32 | #include "llvm/Analysis/MemoryLocation.h" |
| 33 | #include "llvm/Analysis/MemorySSA.h" |
| 34 | #include "llvm/Analysis/MemorySSAUpdater.h" |
| 35 | #include "llvm/Analysis/PostDominators.h" |
| 36 | #include "llvm/Analysis/TargetLibraryInfo.h" |
| 37 | #include "llvm/Analysis/ValueTracking.h" |
| 38 | #include "llvm/IR/Argument.h" |
| 39 | #include "llvm/IR/BasicBlock.h" |
| 40 | #include "llvm/IR/Constant.h" |
| 41 | #include "llvm/IR/Constants.h" |
| 42 | #include "llvm/IR/DataLayout.h" |
| 43 | #include "llvm/IR/Dominators.h" |
| 44 | #include "llvm/IR/Function.h" |
| 45 | #include "llvm/IR/InstIterator.h" |
| 46 | #include "llvm/IR/InstrTypes.h" |
| 47 | #include "llvm/IR/Instruction.h" |
| 48 | #include "llvm/IR/Instructions.h" |
| 49 | #include "llvm/IR/IntrinsicInst.h" |
| 50 | #include "llvm/IR/Intrinsics.h" |
| 51 | #include "llvm/IR/LLVMContext.h" |
| 52 | #include "llvm/IR/Module.h" |
| 53 | #include "llvm/IR/PassManager.h" |
| 54 | #include "llvm/IR/PatternMatch.h" |
| 55 | #include "llvm/IR/Value.h" |
| 56 | #include "llvm/InitializePasses.h" |
| 57 | #include "llvm/Pass.h" |
| 58 | #include "llvm/Support/Casting.h" |
| 59 | #include "llvm/Support/CommandLine.h" |
| 60 | #include "llvm/Support/Debug.h" |
| 61 | #include "llvm/Support/DebugCounter.h" |
| 62 | #include "llvm/Support/ErrorHandling.h" |
| 63 | #include "llvm/Support/MathExtras.h" |
| 64 | #include "llvm/Support/raw_ostream.h" |
| 65 | #include "llvm/Transforms/Scalar.h" |
| 66 | #include "llvm/Transforms/Utils/AssumeBundleBuilder.h" |
| 67 | #include "llvm/Transforms/Utils/Local.h" |
| 68 | #include <algorithm> |
| 69 | #include <cassert> |
| 70 | #include <cstddef> |
| 71 | #include <cstdint> |
| 72 | #include <iterator> |
| 73 | #include <map> |
| 74 | #include <utility> |
| 75 | |
| 76 | using namespace llvm; |
| 77 | using namespace PatternMatch; |
| 78 | |
| 79 | #define DEBUG_TYPE"dse" "dse" |
| 80 | |
| 81 | STATISTIC(NumRemainingStores, "Number of stores remaining after DSE")static llvm::Statistic NumRemainingStores = {"dse", "NumRemainingStores" , "Number of stores remaining after DSE"}; |
| 82 | STATISTIC(NumRedundantStores, "Number of redundant stores deleted")static llvm::Statistic NumRedundantStores = {"dse", "NumRedundantStores" , "Number of redundant stores deleted"}; |
| 83 | STATISTIC(NumFastStores, "Number of stores deleted")static llvm::Statistic NumFastStores = {"dse", "NumFastStores" , "Number of stores deleted"}; |
| 84 | STATISTIC(NumFastOther, "Number of other instrs removed")static llvm::Statistic NumFastOther = {"dse", "NumFastOther", "Number of other instrs removed"}; |
| 85 | STATISTIC(NumCompletePartials, "Number of stores dead by later partials")static llvm::Statistic NumCompletePartials = {"dse", "NumCompletePartials" , "Number of stores dead by later partials"}; |
| 86 | STATISTIC(NumModifiedStores, "Number of stores modified")static llvm::Statistic NumModifiedStores = {"dse", "NumModifiedStores" , "Number of stores modified"}; |
| 87 | STATISTIC(NumCFGChecks, "Number of stores modified")static llvm::Statistic NumCFGChecks = {"dse", "NumCFGChecks", "Number of stores modified"}; |
| 88 | STATISTIC(NumCFGTries, "Number of stores modified")static llvm::Statistic NumCFGTries = {"dse", "NumCFGTries", "Number of stores modified" }; |
| 89 | STATISTIC(NumCFGSuccess, "Number of stores modified")static llvm::Statistic NumCFGSuccess = {"dse", "NumCFGSuccess" , "Number of stores modified"}; |
| 90 | STATISTIC(NumGetDomMemoryDefPassed,static llvm::Statistic NumGetDomMemoryDefPassed = {"dse", "NumGetDomMemoryDefPassed" , "Number of times a valid candidate is returned from getDomMemoryDef" } |
| 91 | "Number of times a valid candidate is returned from getDomMemoryDef")static llvm::Statistic NumGetDomMemoryDefPassed = {"dse", "NumGetDomMemoryDefPassed" , "Number of times a valid candidate is returned from getDomMemoryDef" }; |
| 92 | STATISTIC(NumDomMemDefChecks,static llvm::Statistic NumDomMemDefChecks = {"dse", "NumDomMemDefChecks" , "Number iterations check for reads in getDomMemoryDef"} |
| 93 | "Number iterations check for reads in getDomMemoryDef")static llvm::Statistic NumDomMemDefChecks = {"dse", "NumDomMemDefChecks" , "Number iterations check for reads in getDomMemoryDef"}; |
| 94 | |
| 95 | DEBUG_COUNTER(MemorySSACounter, "dse-memoryssa",static const unsigned MemorySSACounter = DebugCounter::registerCounter ("dse-memoryssa", "Controls which MemoryDefs are eliminated." ) |
| 96 | "Controls which MemoryDefs are eliminated.")static const unsigned MemorySSACounter = DebugCounter::registerCounter ("dse-memoryssa", "Controls which MemoryDefs are eliminated." ); |
| 97 | |
| 98 | static cl::opt<bool> |
| 99 | EnablePartialOverwriteTracking("enable-dse-partial-overwrite-tracking", |
| 100 | cl::init(true), cl::Hidden, |
| 101 | cl::desc("Enable partial-overwrite tracking in DSE")); |
| 102 | |
| 103 | static cl::opt<bool> |
| 104 | EnablePartialStoreMerging("enable-dse-partial-store-merging", |
| 105 | cl::init(true), cl::Hidden, |
| 106 | cl::desc("Enable partial store merging in DSE")); |
| 107 | |
| 108 | static cl::opt<bool> |
| 109 | EnableMemorySSA("enable-dse-memoryssa", cl::init(true), cl::Hidden, |
| 110 | cl::desc("Use the new MemorySSA-backed DSE.")); |
| 111 | |
| 112 | static cl::opt<unsigned> |
| 113 | MemorySSAScanLimit("dse-memoryssa-scanlimit", cl::init(150), cl::Hidden, |
| 114 | cl::desc("The number of memory instructions to scan for " |
| 115 | "dead store elimination (default = 100)")); |
| 116 | static cl::opt<unsigned> MemorySSAUpwardsStepLimit( |
| 117 | "dse-memoryssa-walklimit", cl::init(90), cl::Hidden, |
| 118 | cl::desc("The maximum number of steps while walking upwards to find " |
| 119 | "MemoryDefs that may be killed (default = 90)")); |
| 120 | |
| 121 | static cl::opt<unsigned> MemorySSAPartialStoreLimit( |
| 122 | "dse-memoryssa-partial-store-limit", cl::init(5), cl::Hidden, |
| 123 | cl::desc("The maximum number candidates that only partially overwrite the " |
| 124 | "killing MemoryDef to consider" |
| 125 | " (default = 5)")); |
| 126 | |
| 127 | static cl::opt<unsigned> MemorySSADefsPerBlockLimit( |
| 128 | "dse-memoryssa-defs-per-block-limit", cl::init(5000), cl::Hidden, |
| 129 | cl::desc("The number of MemoryDefs we consider as candidates to eliminated " |
| 130 | "other stores per basic block (default = 5000)")); |
| 131 | |
| 132 | static cl::opt<unsigned> MemorySSASameBBStepCost( |
| 133 | "dse-memoryssa-samebb-cost", cl::init(1), cl::Hidden, |
| 134 | cl::desc( |
| 135 | "The cost of a step in the same basic block as the killing MemoryDef" |
| 136 | "(default = 1)")); |
| 137 | |
| 138 | static cl::opt<unsigned> |
| 139 | MemorySSAOtherBBStepCost("dse-memoryssa-otherbb-cost", cl::init(5), |
| 140 | cl::Hidden, |
| 141 | cl::desc("The cost of a step in a different basic " |
| 142 | "block than the killing MemoryDef" |
| 143 | "(default = 5)")); |
| 144 | |
| 145 | static cl::opt<unsigned> MemorySSAPathCheckLimit( |
| 146 | "dse-memoryssa-path-check-limit", cl::init(50), cl::Hidden, |
| 147 | cl::desc("The maximum number of blocks to check when trying to prove that " |
| 148 | "all paths to an exit go through a killing block (default = 50)")); |
| 149 | |
| 150 | //===----------------------------------------------------------------------===// |
| 151 | // Helper functions |
| 152 | //===----------------------------------------------------------------------===// |
| 153 | using OverlapIntervalsTy = std::map<int64_t, int64_t>; |
| 154 | using InstOverlapIntervalsTy = DenseMap<Instruction *, OverlapIntervalsTy>; |
| 155 | |
| 156 | /// Delete this instruction. Before we do, go through and zero out all the |
| 157 | /// operands of this instruction. If any of them become dead, delete them and |
| 158 | /// the computation tree that feeds them. |
| 159 | /// If ValueSet is non-null, remove any deleted instructions from it as well. |
| 160 | static void |
| 161 | deleteDeadInstruction(Instruction *I, BasicBlock::iterator *BBI, |
| 162 | MemoryDependenceResults &MD, const TargetLibraryInfo &TLI, |
| 163 | InstOverlapIntervalsTy &IOL, |
| 164 | MapVector<Instruction *, bool> &ThrowableInst, |
| 165 | SmallSetVector<const Value *, 16> *ValueSet = nullptr) { |
| 166 | SmallVector<Instruction*, 32> NowDeadInsts; |
| 167 | |
| 168 | NowDeadInsts.push_back(I); |
| 169 | --NumFastOther; |
| 170 | |
| 171 | // Keeping the iterator straight is a pain, so we let this routine tell the |
| 172 | // caller what the next instruction is after we're done mucking about. |
| 173 | BasicBlock::iterator NewIter = *BBI; |
| 174 | |
| 175 | // Before we touch this instruction, remove it from memdep! |
| 176 | do { |
| 177 | Instruction *DeadInst = NowDeadInsts.pop_back_val(); |
| 178 | // Mark the DeadInst as dead in the list of throwable instructions. |
| 179 | auto It = ThrowableInst.find(DeadInst); |
| 180 | if (It != ThrowableInst.end()) |
| 181 | ThrowableInst[It->first] = false; |
| 182 | ++NumFastOther; |
| 183 | |
| 184 | // Try to preserve debug information attached to the dead instruction. |
| 185 | salvageDebugInfo(*DeadInst); |
| 186 | salvageKnowledge(DeadInst); |
| 187 | |
| 188 | // This instruction is dead, zap it, in stages. Start by removing it from |
| 189 | // MemDep, which needs to know the operands and needs it to be in the |
| 190 | // function. |
| 191 | MD.removeInstruction(DeadInst); |
| 192 | |
| 193 | for (unsigned op = 0, e = DeadInst->getNumOperands(); op != e; ++op) { |
| 194 | Value *Op = DeadInst->getOperand(op); |
| 195 | DeadInst->setOperand(op, nullptr); |
| 196 | |
| 197 | // If this operand just became dead, add it to the NowDeadInsts list. |
| 198 | if (!Op->use_empty()) continue; |
| 199 | |
| 200 | if (Instruction *OpI = dyn_cast<Instruction>(Op)) |
| 201 | if (isInstructionTriviallyDead(OpI, &TLI)) |
| 202 | NowDeadInsts.push_back(OpI); |
| 203 | } |
| 204 | |
| 205 | if (ValueSet) ValueSet->remove(DeadInst); |
| 206 | IOL.erase(DeadInst); |
| 207 | |
| 208 | if (NewIter == DeadInst->getIterator()) |
| 209 | NewIter = DeadInst->eraseFromParent(); |
| 210 | else |
| 211 | DeadInst->eraseFromParent(); |
| 212 | } while (!NowDeadInsts.empty()); |
| 213 | *BBI = NewIter; |
| 214 | // Pop dead entries from back of ThrowableInst till we find an alive entry. |
| 215 | while (!ThrowableInst.empty() && !ThrowableInst.back().second) |
| 216 | ThrowableInst.pop_back(); |
| 217 | } |
| 218 | |
| 219 | /// Does this instruction write some memory? This only returns true for things |
| 220 | /// that we can analyze with other helpers below. |
| 221 | static bool hasAnalyzableMemoryWrite(Instruction *I, |
| 222 | const TargetLibraryInfo &TLI) { |
| 223 | if (isa<StoreInst>(I)) |
| 224 | return true; |
| 225 | if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { |
| 226 | switch (II->getIntrinsicID()) { |
| 227 | default: |
| 228 | return false; |
| 229 | case Intrinsic::memset: |
| 230 | case Intrinsic::memmove: |
| 231 | case Intrinsic::memcpy: |
| 232 | case Intrinsic::memcpy_inline: |
| 233 | case Intrinsic::memcpy_element_unordered_atomic: |
| 234 | case Intrinsic::memmove_element_unordered_atomic: |
| 235 | case Intrinsic::memset_element_unordered_atomic: |
| 236 | case Intrinsic::init_trampoline: |
| 237 | case Intrinsic::lifetime_end: |
| 238 | case Intrinsic::masked_store: |
| 239 | return true; |
| 240 | } |
| 241 | } |
| 242 | if (auto *CB = dyn_cast<CallBase>(I)) { |
| 243 | LibFunc LF; |
| 244 | if (TLI.getLibFunc(*CB, LF) && TLI.has(LF)) { |
| 245 | switch (LF) { |
| 246 | case LibFunc_strcpy: |
| 247 | case LibFunc_strncpy: |
| 248 | case LibFunc_strcat: |
| 249 | case LibFunc_strncat: |
| 250 | return true; |
| 251 | default: |
| 252 | return false; |
| 253 | } |
| 254 | } |
| 255 | } |
| 256 | return false; |
| 257 | } |
| 258 | |
| 259 | /// Return a Location stored to by the specified instruction. If isRemovable |
| 260 | /// returns true, this function and getLocForRead completely describe the memory |
| 261 | /// operations for this instruction. |
| 262 | static MemoryLocation getLocForWrite(Instruction *Inst, |
| 263 | const TargetLibraryInfo &TLI) { |
| 264 | if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) |
| 265 | return MemoryLocation::get(SI); |
| 266 | |
| 267 | // memcpy/memmove/memset. |
| 268 | if (auto *MI = dyn_cast<AnyMemIntrinsic>(Inst)) |
| 269 | return MemoryLocation::getForDest(MI); |
| 270 | |
| 271 | if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) { |
| 272 | switch (II->getIntrinsicID()) { |
| 273 | default: |
| 274 | return MemoryLocation(); // Unhandled intrinsic. |
| 275 | case Intrinsic::init_trampoline: |
| 276 | return MemoryLocation::getAfter(II->getArgOperand(0)); |
| 277 | case Intrinsic::masked_store: |
| 278 | return MemoryLocation::getForArgument(II, 1, TLI); |
| 279 | case Intrinsic::lifetime_end: { |
| 280 | uint64_t Len = cast<ConstantInt>(II->getArgOperand(0))->getZExtValue(); |
| 281 | return MemoryLocation(II->getArgOperand(1), Len); |
| 282 | } |
| 283 | } |
| 284 | } |
| 285 | if (auto *CB = dyn_cast<CallBase>(Inst)) |
| 286 | // All the supported TLI functions so far happen to have dest as their |
| 287 | // first argument. |
| 288 | return MemoryLocation::getAfter(CB->getArgOperand(0)); |
| 289 | return MemoryLocation(); |
| 290 | } |
| 291 | |
| 292 | /// Return the location read by the specified "hasAnalyzableMemoryWrite" |
| 293 | /// instruction if any. |
| 294 | static MemoryLocation getLocForRead(Instruction *Inst, |
| 295 | const TargetLibraryInfo &TLI) { |
| 296 | assert(hasAnalyzableMemoryWrite(Inst, TLI) && "Unknown instruction case")((hasAnalyzableMemoryWrite(Inst, TLI) && "Unknown instruction case" ) ? static_cast<void> (0) : __assert_fail ("hasAnalyzableMemoryWrite(Inst, TLI) && \"Unknown instruction case\"" , "/build/llvm-toolchain-snapshot-13~++20210301100612+564f5b0734bd/llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp" , 296, __PRETTY_FUNCTION__)); |
| 297 | |
| 298 | // The only instructions that both read and write are the mem transfer |
| 299 | // instructions (memcpy/memmove). |
| 300 | if (auto *MTI = dyn_cast<AnyMemTransferInst>(Inst)) |
| 301 | return MemoryLocation::getForSource(MTI); |
| 302 | return MemoryLocation(); |
| 303 | } |
| 304 | |
| 305 | /// If the value of this instruction and the memory it writes to is unused, may |
| 306 | /// we delete this instruction? |
| 307 | static bool isRemovable(Instruction *I) { |
| 308 | // Don't remove volatile/atomic stores. |
| 309 | if (StoreInst *SI = dyn_cast<StoreInst>(I)) |
| 310 | return SI->isUnordered(); |
| 311 | |
| 312 | if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { |
| 313 | switch (II->getIntrinsicID()) { |
| 314 | default: llvm_unreachable("doesn't pass 'hasAnalyzableMemoryWrite' predicate")::llvm::llvm_unreachable_internal("doesn't pass 'hasAnalyzableMemoryWrite' predicate" , "/build/llvm-toolchain-snapshot-13~++20210301100612+564f5b0734bd/llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp" , 314); |
| 315 | case Intrinsic::lifetime_end: |
| 316 | // Never remove dead lifetime_end's, e.g. because it is followed by a |
| 317 | // free. |
| 318 | return false; |
| 319 | case Intrinsic::init_trampoline: |
| 320 | // Always safe to remove init_trampoline. |
| 321 | return true; |
| 322 | case Intrinsic::memset: |
| 323 | case Intrinsic::memmove: |
| 324 | case Intrinsic::memcpy: |
| 325 | case Intrinsic::memcpy_inline: |
| 326 | // Don't remove volatile memory intrinsics. |
| 327 | return !cast<MemIntrinsic>(II)->isVolatile(); |
| 328 | case Intrinsic::memcpy_element_unordered_atomic: |
| 329 | case Intrinsic::memmove_element_unordered_atomic: |
| 330 | case Intrinsic::memset_element_unordered_atomic: |
| 331 | case Intrinsic::masked_store: |
| 332 | return true; |
| 333 | } |
| 334 | } |
| 335 | |
| 336 | // note: only get here for calls with analyzable writes - i.e. libcalls |
| 337 | if (auto *CB = dyn_cast<CallBase>(I)) |
| 338 | return CB->use_empty(); |
| 339 | |
| 340 | return false; |
| 341 | } |
| 342 | |
| 343 | /// Returns true if the end of this instruction can be safely shortened in |
| 344 | /// length. |
| 345 | static bool isShortenableAtTheEnd(Instruction *I) { |
| 346 | // Don't shorten stores for now |
| 347 | if (isa<StoreInst>(I)) |
| 348 | return false; |
| 349 | |
| 350 | if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { |
| 351 | switch (II->getIntrinsicID()) { |
| 352 | default: return false; |
| 353 | case Intrinsic::memset: |
| 354 | case Intrinsic::memcpy: |
| 355 | case Intrinsic::memcpy_element_unordered_atomic: |
| 356 | case Intrinsic::memset_element_unordered_atomic: |
| 357 | // Do shorten memory intrinsics. |
| 358 | // FIXME: Add memmove if it's also safe to transform. |
| 359 | return true; |
| 360 | } |
| 361 | } |
| 362 | |
| 363 | // Don't shorten libcalls calls for now. |
| 364 | |
| 365 | return false; |
| 366 | } |
| 367 | |
| 368 | /// Returns true if the beginning of this instruction can be safely shortened |
| 369 | /// in length. |
| 370 | static bool isShortenableAtTheBeginning(Instruction *I) { |
| 371 | // FIXME: Handle only memset for now. Supporting memcpy/memmove should be |
| 372 | // easily done by offsetting the source address. |
| 373 | return isa<AnyMemSetInst>(I); |
| 374 | } |
| 375 | |
| 376 | /// Return the pointer that is being written to. |
| 377 | static Value *getStoredPointerOperand(Instruction *I, |
| 378 | const TargetLibraryInfo &TLI) { |
| 379 | //TODO: factor this to reuse getLocForWrite |
| 380 | MemoryLocation Loc = getLocForWrite(I, TLI); |
| 381 | assert(Loc.Ptr &&((Loc.Ptr && "unable to find pointer written for analyzable instruction?" ) ? static_cast<void> (0) : __assert_fail ("Loc.Ptr && \"unable to find pointer written for analyzable instruction?\"" , "/build/llvm-toolchain-snapshot-13~++20210301100612+564f5b0734bd/llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp" , 382, __PRETTY_FUNCTION__)) |
| 382 | "unable to find pointer written for analyzable instruction?")((Loc.Ptr && "unable to find pointer written for analyzable instruction?" ) ? static_cast<void> (0) : __assert_fail ("Loc.Ptr && \"unable to find pointer written for analyzable instruction?\"" , "/build/llvm-toolchain-snapshot-13~++20210301100612+564f5b0734bd/llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp" , 382, __PRETTY_FUNCTION__)); |
| 383 | // TODO: most APIs don't expect const Value * |
| 384 | return const_cast<Value*>(Loc.Ptr); |
| 385 | } |
| 386 | |
| 387 | static uint64_t getPointerSize(const Value *V, const DataLayout &DL, |
| 388 | const TargetLibraryInfo &TLI, |
| 389 | const Function *F) { |
| 390 | uint64_t Size; |
| 391 | ObjectSizeOpts Opts; |
| 392 | Opts.NullIsUnknownSize = NullPointerIsDefined(F); |
| 393 | |
| 394 | if (getObjectSize(V, Size, DL, &TLI, Opts)) |
| 395 | return Size; |
| 396 | return MemoryLocation::UnknownSize; |
| 397 | } |
| 398 | |
| 399 | namespace { |
| 400 | |
| 401 | enum OverwriteResult { |
| 402 | OW_Begin, |
| 403 | OW_Complete, |
| 404 | OW_End, |
| 405 | OW_PartialEarlierWithFullLater, |
| 406 | OW_MaybePartial, |
| 407 | OW_Unknown |
| 408 | }; |
| 409 | |
| 410 | } // end anonymous namespace |
| 411 | |
| 412 | /// Check if two instruction are masked stores that completely |
| 413 | /// overwrite one another. More specifically, \p Later has to |
| 414 | /// overwrite \p Earlier. |
| 415 | template <typename AATy> |
| 416 | static OverwriteResult isMaskedStoreOverwrite(const Instruction *Later, |
| 417 | const Instruction *Earlier, |
| 418 | AATy &AA) { |
| 419 | const auto *IIL = dyn_cast<IntrinsicInst>(Later); |
| 420 | const auto *IIE = dyn_cast<IntrinsicInst>(Earlier); |
| 421 | if (IIL == nullptr || IIE == nullptr) |
| 422 | return OW_Unknown; |
| 423 | if (IIL->getIntrinsicID() != Intrinsic::masked_store || |
| 424 | IIE->getIntrinsicID() != Intrinsic::masked_store) |
| 425 | return OW_Unknown; |
| 426 | // Pointers. |
| 427 | Value *LP = IIL->getArgOperand(1)->stripPointerCasts(); |
| 428 | Value *EP = IIE->getArgOperand(1)->stripPointerCasts(); |
| 429 | if (LP != EP && !AA.isMustAlias(LP, EP)) |
| 430 | return OW_Unknown; |
| 431 | // Masks. |
| 432 | // TODO: check that Later's mask is a superset of the Earlier's mask. |
| 433 | if (IIL->getArgOperand(3) != IIE->getArgOperand(3)) |
| 434 | return OW_Unknown; |
| 435 | return OW_Complete; |
| 436 | } |
| 437 | |
| 438 | /// Return 'OW_Complete' if a store to the 'Later' location (by \p LaterI |
| 439 | /// instruction) completely overwrites a store to the 'Earlier' location. |
| 440 | /// (by \p EarlierI instruction). |
| 441 | /// Return OW_MaybePartial if \p Later does not completely overwrite |
| 442 | /// \p Earlier, but they both write to the same underlying object. In that |
| 443 | /// case, use isPartialOverwrite to check if \p Later partially overwrites |
| 444 | /// \p Earlier. Returns 'OW_Unknown' if nothing can be determined. |
| 445 | template <typename AATy> |
| 446 | static OverwriteResult |
| 447 | isOverwrite(const Instruction *LaterI, const Instruction *EarlierI, |
| 448 | const MemoryLocation &Later, const MemoryLocation &Earlier, |
| 449 | const DataLayout &DL, const TargetLibraryInfo &TLI, |
| 450 | int64_t &EarlierOff, int64_t &LaterOff, AATy &AA, |
| 451 | const Function *F) { |
| 452 | // FIXME: Vet that this works for size upper-bounds. Seems unlikely that we'll |
| 453 | // get imprecise values here, though (except for unknown sizes). |
| 454 | if (!Later.Size.isPrecise() || !Earlier.Size.isPrecise()) { |
| 455 | // Masked stores have imprecise locations, but we can reason about them |
| 456 | // to some extent. |
| 457 | return isMaskedStoreOverwrite(LaterI, EarlierI, AA); |
| 458 | } |
| 459 | |
| 460 | const uint64_t LaterSize = Later.Size.getValue(); |
| 461 | const uint64_t EarlierSize = Earlier.Size.getValue(); |
| 462 | |
| 463 | const Value *P1 = Earlier.Ptr->stripPointerCasts(); |
| 464 | const Value *P2 = Later.Ptr->stripPointerCasts(); |
| 465 | |
| 466 | // If the start pointers are the same, we just have to compare sizes to see if |
| 467 | // the later store was larger than the earlier store. |
| 468 | if (P1 == P2 || AA.isMustAlias(P1, P2)) { |
| 469 | // Make sure that the Later size is >= the Earlier size. |
| 470 | if (LaterSize >= EarlierSize) |
| 471 | return OW_Complete; |
| 472 | } |
| 473 | |
| 474 | // Check to see if the later store is to the entire object (either a global, |
| 475 | // an alloca, or a byval/inalloca argument). If so, then it clearly |
| 476 | // overwrites any other store to the same object. |
| 477 | const Value *UO1 = getUnderlyingObject(P1), *UO2 = getUnderlyingObject(P2); |
| 478 | |
| 479 | // If we can't resolve the same pointers to the same object, then we can't |
| 480 | // analyze them at all. |
| 481 | if (UO1 != UO2) |
| 482 | return OW_Unknown; |
| 483 | |
| 484 | // If the "Later" store is to a recognizable object, get its size. |
| 485 | uint64_t ObjectSize = getPointerSize(UO2, DL, TLI, F); |
| 486 | if (ObjectSize != MemoryLocation::UnknownSize) |
| 487 | if (ObjectSize == LaterSize && ObjectSize >= EarlierSize) |
| 488 | return OW_Complete; |
| 489 | |
| 490 | // Okay, we have stores to two completely different pointers. Try to |
| 491 | // decompose the pointer into a "base + constant_offset" form. If the base |
| 492 | // pointers are equal, then we can reason about the two stores. |
| 493 | EarlierOff = 0; |
| 494 | LaterOff = 0; |
| 495 | const Value *BP1 = GetPointerBaseWithConstantOffset(P1, EarlierOff, DL); |
| 496 | const Value *BP2 = GetPointerBaseWithConstantOffset(P2, LaterOff, DL); |
| 497 | |
| 498 | // If the base pointers still differ, we have two completely different stores. |
| 499 | if (BP1 != BP2) |
| 500 | return OW_Unknown; |
| 501 | |
| 502 | // The later access completely overlaps the earlier store if and only if |
| 503 | // both start and end of the earlier one is "inside" the later one: |
| 504 | // |<->|--earlier--|<->| |
| 505 | // |-------later-------| |
| 506 | // Accesses may overlap if and only if start of one of them is "inside" |
| 507 | // another one: |
| 508 | // |<->|--earlier--|<----->| |
| 509 | // |-------later-------| |
| 510 | // OR |
| 511 | // |----- earlier -----| |
| 512 | // |<->|---later---|<----->| |
| 513 | // |
| 514 | // We have to be careful here as *Off is signed while *.Size is unsigned. |
| 515 | |
| 516 | // Check if the earlier access starts "not before" the later one. |
| 517 | if (EarlierOff >= LaterOff) { |
| 518 | // If the earlier access ends "not after" the later access then the earlier |
| 519 | // one is completely overwritten by the later one. |
| 520 | if (uint64_t(EarlierOff - LaterOff) + EarlierSize <= LaterSize) |
| 521 | return OW_Complete; |
| 522 | // If start of the earlier access is "before" end of the later access then |
| 523 | // accesses overlap. |
| 524 | else if ((uint64_t)(EarlierOff - LaterOff) < LaterSize) |
| 525 | return OW_MaybePartial; |
| 526 | } |
| 527 | // If start of the later access is "before" end of the earlier access then |
| 528 | // accesses overlap. |
| 529 | else if ((uint64_t)(LaterOff - EarlierOff) < EarlierSize) { |
| 530 | return OW_MaybePartial; |
| 531 | } |
| 532 | |
| 533 | // Can reach here only if accesses are known not to overlap. There is no |
| 534 | // dedicated code to indicate no overlap so signal "unknown". |
| 535 | return OW_Unknown; |
| 536 | } |
| 537 | |
| 538 | /// Return 'OW_Complete' if a store to the 'Later' location completely |
| 539 | /// overwrites a store to the 'Earlier' location, 'OW_End' if the end of the |
| 540 | /// 'Earlier' location is completely overwritten by 'Later', 'OW_Begin' if the |
| 541 | /// beginning of the 'Earlier' location is overwritten by 'Later'. |
| 542 | /// 'OW_PartialEarlierWithFullLater' means that an earlier (big) store was |
| 543 | /// overwritten by a latter (smaller) store which doesn't write outside the big |
| 544 | /// store's memory locations. Returns 'OW_Unknown' if nothing can be determined. |
| 545 | /// NOTE: This function must only be called if both \p Later and \p Earlier |
| 546 | /// write to the same underlying object with valid \p EarlierOff and \p |
| 547 | /// LaterOff. |
| 548 | static OverwriteResult isPartialOverwrite(const MemoryLocation &Later, |
| 549 | const MemoryLocation &Earlier, |
| 550 | int64_t EarlierOff, int64_t LaterOff, |
| 551 | Instruction *DepWrite, |
| 552 | InstOverlapIntervalsTy &IOL) { |
| 553 | const uint64_t LaterSize = Later.Size.getValue(); |
| 554 | const uint64_t EarlierSize = Earlier.Size.getValue(); |
| 555 | // We may now overlap, although the overlap is not complete. There might also |
| 556 | // be other incomplete overlaps, and together, they might cover the complete |
| 557 | // earlier write. |
| 558 | // Note: The correctness of this logic depends on the fact that this function |
| 559 | // is not even called providing DepWrite when there are any intervening reads. |
| 560 | if (EnablePartialOverwriteTracking && |
| 561 | LaterOff < int64_t(EarlierOff + EarlierSize) && |
| 562 | int64_t(LaterOff + LaterSize) >= EarlierOff) { |
| 563 | |
| 564 | // Insert our part of the overlap into the map. |
| 565 | auto &IM = IOL[DepWrite]; |
| 566 | LLVM_DEBUG(dbgs() << "DSE: Partial overwrite: Earlier [" << EarlierOffdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Partial overwrite: Earlier [" << EarlierOff << ", " << int64_t(EarlierOff + EarlierSize) << ") Later [" << LaterOff << ", " << int64_t(LaterOff + LaterSize) << ")\n"; } } while (false) |
| 567 | << ", " << int64_t(EarlierOff + EarlierSize)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Partial overwrite: Earlier [" << EarlierOff << ", " << int64_t(EarlierOff + EarlierSize) << ") Later [" << LaterOff << ", " << int64_t(LaterOff + LaterSize) << ")\n"; } } while (false) |
| 568 | << ") Later [" << LaterOff << ", "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Partial overwrite: Earlier [" << EarlierOff << ", " << int64_t(EarlierOff + EarlierSize) << ") Later [" << LaterOff << ", " << int64_t(LaterOff + LaterSize) << ")\n"; } } while (false) |
| 569 | << int64_t(LaterOff + LaterSize) << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Partial overwrite: Earlier [" << EarlierOff << ", " << int64_t(EarlierOff + EarlierSize) << ") Later [" << LaterOff << ", " << int64_t(LaterOff + LaterSize) << ")\n"; } } while (false); |
| 570 | |
| 571 | // Make sure that we only insert non-overlapping intervals and combine |
| 572 | // adjacent intervals. The intervals are stored in the map with the ending |
| 573 | // offset as the key (in the half-open sense) and the starting offset as |
| 574 | // the value. |
| 575 | int64_t LaterIntStart = LaterOff, LaterIntEnd = LaterOff + LaterSize; |
| 576 | |
| 577 | // Find any intervals ending at, or after, LaterIntStart which start |
| 578 | // before LaterIntEnd. |
| 579 | auto ILI = IM.lower_bound(LaterIntStart); |
| 580 | if (ILI != IM.end() && ILI->second <= LaterIntEnd) { |
| 581 | // This existing interval is overlapped with the current store somewhere |
| 582 | // in [LaterIntStart, LaterIntEnd]. Merge them by erasing the existing |
| 583 | // intervals and adjusting our start and end. |
| 584 | LaterIntStart = std::min(LaterIntStart, ILI->second); |
| 585 | LaterIntEnd = std::max(LaterIntEnd, ILI->first); |
| 586 | ILI = IM.erase(ILI); |
| 587 | |
| 588 | // Continue erasing and adjusting our end in case other previous |
| 589 | // intervals are also overlapped with the current store. |
| 590 | // |
| 591 | // |--- ealier 1 ---| |--- ealier 2 ---| |
| 592 | // |------- later---------| |
| 593 | // |
| 594 | while (ILI != IM.end() && ILI->second <= LaterIntEnd) { |
| 595 | assert(ILI->second > LaterIntStart && "Unexpected interval")((ILI->second > LaterIntStart && "Unexpected interval" ) ? static_cast<void> (0) : __assert_fail ("ILI->second > LaterIntStart && \"Unexpected interval\"" , "/build/llvm-toolchain-snapshot-13~++20210301100612+564f5b0734bd/llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp" , 595, __PRETTY_FUNCTION__)); |
| 596 | LaterIntEnd = std::max(LaterIntEnd, ILI->first); |
| 597 | ILI = IM.erase(ILI); |
| 598 | } |
| 599 | } |
| 600 | |
| 601 | IM[LaterIntEnd] = LaterIntStart; |
| 602 | |
| 603 | ILI = IM.begin(); |
| 604 | if (ILI->second <= EarlierOff && |
| 605 | ILI->first >= int64_t(EarlierOff + EarlierSize)) { |
| 606 | LLVM_DEBUG(dbgs() << "DSE: Full overwrite from partials: Earlier ["do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Full overwrite from partials: Earlier [" << EarlierOff << ", " << int64_t(EarlierOff + EarlierSize) << ") Composite Later [" << ILI-> second << ", " << ILI->first << ")\n"; } } while (false) |
| 607 | << EarlierOff << ", "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Full overwrite from partials: Earlier [" << EarlierOff << ", " << int64_t(EarlierOff + EarlierSize) << ") Composite Later [" << ILI-> second << ", " << ILI->first << ")\n"; } } while (false) |
| 608 | << int64_t(EarlierOff + EarlierSize)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Full overwrite from partials: Earlier [" << EarlierOff << ", " << int64_t(EarlierOff + EarlierSize) << ") Composite Later [" << ILI-> second << ", " << ILI->first << ")\n"; } } while (false) |
| 609 | << ") Composite Later [" << ILI->second << ", "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Full overwrite from partials: Earlier [" << EarlierOff << ", " << int64_t(EarlierOff + EarlierSize) << ") Composite Later [" << ILI-> second << ", " << ILI->first << ")\n"; } } while (false) |
| 610 | << ILI->first << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Full overwrite from partials: Earlier [" << EarlierOff << ", " << int64_t(EarlierOff + EarlierSize) << ") Composite Later [" << ILI-> second << ", " << ILI->first << ")\n"; } } while (false); |
| 611 | ++NumCompletePartials; |
| 612 | return OW_Complete; |
| 613 | } |
| 614 | } |
| 615 | |
| 616 | // Check for an earlier store which writes to all the memory locations that |
| 617 | // the later store writes to. |
| 618 | if (EnablePartialStoreMerging && LaterOff >= EarlierOff && |
| 619 | int64_t(EarlierOff + EarlierSize) > LaterOff && |
| 620 | uint64_t(LaterOff - EarlierOff) + LaterSize <= EarlierSize) { |
| 621 | LLVM_DEBUG(dbgs() << "DSE: Partial overwrite an earlier load ["do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Partial overwrite an earlier load [" << EarlierOff << ", " << int64_t(EarlierOff + EarlierSize) << ") by a later store [" << LaterOff << ", " << int64_t(LaterOff + LaterSize) << ")\n"; } } while (false) |
| 622 | << EarlierOff << ", "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Partial overwrite an earlier load [" << EarlierOff << ", " << int64_t(EarlierOff + EarlierSize) << ") by a later store [" << LaterOff << ", " << int64_t(LaterOff + LaterSize) << ")\n"; } } while (false) |
| 623 | << int64_t(EarlierOff + EarlierSize)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Partial overwrite an earlier load [" << EarlierOff << ", " << int64_t(EarlierOff + EarlierSize) << ") by a later store [" << LaterOff << ", " << int64_t(LaterOff + LaterSize) << ")\n"; } } while (false) |
| 624 | << ") by a later store [" << LaterOff << ", "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Partial overwrite an earlier load [" << EarlierOff << ", " << int64_t(EarlierOff + EarlierSize) << ") by a later store [" << LaterOff << ", " << int64_t(LaterOff + LaterSize) << ")\n"; } } while (false) |
| 625 | << int64_t(LaterOff + LaterSize) << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Partial overwrite an earlier load [" << EarlierOff << ", " << int64_t(EarlierOff + EarlierSize) << ") by a later store [" << LaterOff << ", " << int64_t(LaterOff + LaterSize) << ")\n"; } } while (false); |
| 626 | // TODO: Maybe come up with a better name? |
| 627 | return OW_PartialEarlierWithFullLater; |
| 628 | } |
| 629 | |
| 630 | // Another interesting case is if the later store overwrites the end of the |
| 631 | // earlier store. |
| 632 | // |
| 633 | // |--earlier--| |
| 634 | // |-- later --| |
| 635 | // |
| 636 | // In this case we may want to trim the size of earlier to avoid generating |
| 637 | // writes to addresses which will definitely be overwritten later |
| 638 | if (!EnablePartialOverwriteTracking && |
| 639 | (LaterOff > EarlierOff && LaterOff < int64_t(EarlierOff + EarlierSize) && |
| 640 | int64_t(LaterOff + LaterSize) >= int64_t(EarlierOff + EarlierSize))) |
| 641 | return OW_End; |
| 642 | |
| 643 | // Finally, we also need to check if the later store overwrites the beginning |
| 644 | // of the earlier store. |
| 645 | // |
| 646 | // |--earlier--| |
| 647 | // |-- later --| |
| 648 | // |
| 649 | // In this case we may want to move the destination address and trim the size |
| 650 | // of earlier to avoid generating writes to addresses which will definitely |
| 651 | // be overwritten later. |
| 652 | if (!EnablePartialOverwriteTracking && |
| 653 | (LaterOff <= EarlierOff && int64_t(LaterOff + LaterSize) > EarlierOff)) { |
| 654 | assert(int64_t(LaterOff + LaterSize) < int64_t(EarlierOff + EarlierSize) &&((int64_t(LaterOff + LaterSize) < int64_t(EarlierOff + EarlierSize ) && "Expect to be handled as OW_Complete") ? static_cast <void> (0) : __assert_fail ("int64_t(LaterOff + LaterSize) < int64_t(EarlierOff + EarlierSize) && \"Expect to be handled as OW_Complete\"" , "/build/llvm-toolchain-snapshot-13~++20210301100612+564f5b0734bd/llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp" , 655, __PRETTY_FUNCTION__)) |
| 655 | "Expect to be handled as OW_Complete")((int64_t(LaterOff + LaterSize) < int64_t(EarlierOff + EarlierSize ) && "Expect to be handled as OW_Complete") ? static_cast <void> (0) : __assert_fail ("int64_t(LaterOff + LaterSize) < int64_t(EarlierOff + EarlierSize) && \"Expect to be handled as OW_Complete\"" , "/build/llvm-toolchain-snapshot-13~++20210301100612+564f5b0734bd/llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp" , 655, __PRETTY_FUNCTION__)); |
| 656 | return OW_Begin; |
| 657 | } |
| 658 | // Otherwise, they don't completely overlap. |
| 659 | return OW_Unknown; |
| 660 | } |
| 661 | |
| 662 | /// If 'Inst' might be a self read (i.e. a noop copy of a |
| 663 | /// memory region into an identical pointer) then it doesn't actually make its |
| 664 | /// input dead in the traditional sense. Consider this case: |
| 665 | /// |
| 666 | /// memmove(A <- B) |
| 667 | /// memmove(A <- A) |
| 668 | /// |
| 669 | /// In this case, the second store to A does not make the first store to A dead. |
| 670 | /// The usual situation isn't an explicit A<-A store like this (which can be |
| 671 | /// trivially removed) but a case where two pointers may alias. |
| 672 | /// |
| 673 | /// This function detects when it is unsafe to remove a dependent instruction |
| 674 | /// because the DSE inducing instruction may be a self-read. |
| 675 | static bool isPossibleSelfRead(Instruction *Inst, |
| 676 | const MemoryLocation &InstStoreLoc, |
| 677 | Instruction *DepWrite, |
| 678 | const TargetLibraryInfo &TLI, |
| 679 | AliasAnalysis &AA) { |
| 680 | // Self reads can only happen for instructions that read memory. Get the |
| 681 | // location read. |
| 682 | MemoryLocation InstReadLoc = getLocForRead(Inst, TLI); |
| 683 | if (!InstReadLoc.Ptr) |
| 684 | return false; // Not a reading instruction. |
| 685 | |
| 686 | // If the read and written loc obviously don't alias, it isn't a read. |
| 687 | if (AA.isNoAlias(InstReadLoc, InstStoreLoc)) |
| 688 | return false; |
| 689 | |
| 690 | if (isa<AnyMemCpyInst>(Inst)) { |
| 691 | // LLVM's memcpy overlap semantics are not fully fleshed out (see PR11763) |
| 692 | // but in practice memcpy(A <- B) either means that A and B are disjoint or |
| 693 | // are equal (i.e. there are not partial overlaps). Given that, if we have: |
| 694 | // |
| 695 | // memcpy/memmove(A <- B) // DepWrite |
| 696 | // memcpy(A <- B) // Inst |
| 697 | // |
| 698 | // with Inst reading/writing a >= size than DepWrite, we can reason as |
| 699 | // follows: |
| 700 | // |
| 701 | // - If A == B then both the copies are no-ops, so the DepWrite can be |
| 702 | // removed. |
| 703 | // - If A != B then A and B are disjoint locations in Inst. Since |
| 704 | // Inst.size >= DepWrite.size A and B are disjoint in DepWrite too. |
| 705 | // Therefore DepWrite can be removed. |
| 706 | MemoryLocation DepReadLoc = getLocForRead(DepWrite, TLI); |
| 707 | |
| 708 | if (DepReadLoc.Ptr && AA.isMustAlias(InstReadLoc.Ptr, DepReadLoc.Ptr)) |
| 709 | return false; |
| 710 | } |
| 711 | |
| 712 | // If DepWrite doesn't read memory or if we can't prove it is a must alias, |
| 713 | // then it can't be considered dead. |
| 714 | return true; |
| 715 | } |
| 716 | |
| 717 | /// Returns true if the memory which is accessed by the second instruction is not |
| 718 | /// modified between the first and the second instruction. |
| 719 | /// Precondition: Second instruction must be dominated by the first |
| 720 | /// instruction. |
| 721 | template <typename AATy> |
| 722 | static bool |
| 723 | memoryIsNotModifiedBetween(Instruction *FirstI, Instruction *SecondI, AATy &AA, |
| 724 | const DataLayout &DL, DominatorTree *DT) { |
| 725 | // Do a backwards scan through the CFG from SecondI to FirstI. Look for |
| 726 | // instructions which can modify the memory location accessed by SecondI. |
| 727 | // |
| 728 | // While doing the walk keep track of the address to check. It might be |
| 729 | // different in different basic blocks due to PHI translation. |
| 730 | using BlockAddressPair = std::pair<BasicBlock *, PHITransAddr>; |
| 731 | SmallVector<BlockAddressPair, 16> WorkList; |
| 732 | // Keep track of the address we visited each block with. Bail out if we |
| 733 | // visit a block with different addresses. |
| 734 | DenseMap<BasicBlock *, Value *> Visited; |
| 735 | |
| 736 | BasicBlock::iterator FirstBBI(FirstI); |
| 737 | ++FirstBBI; |
| 738 | BasicBlock::iterator SecondBBI(SecondI); |
| 739 | BasicBlock *FirstBB = FirstI->getParent(); |
| 740 | BasicBlock *SecondBB = SecondI->getParent(); |
| 741 | MemoryLocation MemLoc = MemoryLocation::get(SecondI); |
| 742 | auto *MemLocPtr = const_cast<Value *>(MemLoc.Ptr); |
| 743 | |
| 744 | // Start checking the SecondBB. |
| 745 | WorkList.push_back( |
| 746 | std::make_pair(SecondBB, PHITransAddr(MemLocPtr, DL, nullptr))); |
| 747 | bool isFirstBlock = true; |
| 748 | |
| 749 | // Check all blocks going backward until we reach the FirstBB. |
| 750 | while (!WorkList.empty()) { |
| 751 | BlockAddressPair Current = WorkList.pop_back_val(); |
| 752 | BasicBlock *B = Current.first; |
| 753 | PHITransAddr &Addr = Current.second; |
| 754 | Value *Ptr = Addr.getAddr(); |
| 755 | |
| 756 | // Ignore instructions before FirstI if this is the FirstBB. |
| 757 | BasicBlock::iterator BI = (B == FirstBB ? FirstBBI : B->begin()); |
| 758 | |
| 759 | BasicBlock::iterator EI; |
| 760 | if (isFirstBlock) { |
| 761 | // Ignore instructions after SecondI if this is the first visit of SecondBB. |
| 762 | assert(B == SecondBB && "first block is not the store block")((B == SecondBB && "first block is not the store block" ) ? static_cast<void> (0) : __assert_fail ("B == SecondBB && \"first block is not the store block\"" , "/build/llvm-toolchain-snapshot-13~++20210301100612+564f5b0734bd/llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp" , 762, __PRETTY_FUNCTION__)); |
| 763 | EI = SecondBBI; |
| 764 | isFirstBlock = false; |
| 765 | } else { |
| 766 | // It's not SecondBB or (in case of a loop) the second visit of SecondBB. |
| 767 | // In this case we also have to look at instructions after SecondI. |
| 768 | EI = B->end(); |
| 769 | } |
| 770 | for (; BI != EI; ++BI) { |
| 771 | Instruction *I = &*BI; |
| 772 | if (I->mayWriteToMemory() && I != SecondI) |
| 773 | if (isModSet(AA.getModRefInfo(I, MemLoc.getWithNewPtr(Ptr)))) |
| 774 | return false; |
| 775 | } |
| 776 | if (B != FirstBB) { |
| 777 | assert(B != &FirstBB->getParent()->getEntryBlock() &&((B != &FirstBB->getParent()->getEntryBlock() && "Should not hit the entry block because SI must be dominated by LI" ) ? static_cast<void> (0) : __assert_fail ("B != &FirstBB->getParent()->getEntryBlock() && \"Should not hit the entry block because SI must be dominated by LI\"" , "/build/llvm-toolchain-snapshot-13~++20210301100612+564f5b0734bd/llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp" , 778, __PRETTY_FUNCTION__)) |
| 778 | "Should not hit the entry block because SI must be dominated by LI")((B != &FirstBB->getParent()->getEntryBlock() && "Should not hit the entry block because SI must be dominated by LI" ) ? static_cast<void> (0) : __assert_fail ("B != &FirstBB->getParent()->getEntryBlock() && \"Should not hit the entry block because SI must be dominated by LI\"" , "/build/llvm-toolchain-snapshot-13~++20210301100612+564f5b0734bd/llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp" , 778, __PRETTY_FUNCTION__)); |
| 779 | for (BasicBlock *Pred : predecessors(B)) { |
| 780 | PHITransAddr PredAddr = Addr; |
| 781 | if (PredAddr.NeedsPHITranslationFromBlock(B)) { |
| 782 | if (!PredAddr.IsPotentiallyPHITranslatable()) |
| 783 | return false; |
| 784 | if (PredAddr.PHITranslateValue(B, Pred, DT, false)) |
| 785 | return false; |
| 786 | } |
| 787 | Value *TranslatedPtr = PredAddr.getAddr(); |
| 788 | auto Inserted = Visited.insert(std::make_pair(Pred, TranslatedPtr)); |
| 789 | if (!Inserted.second) { |
| 790 | // We already visited this block before. If it was with a different |
| 791 | // address - bail out! |
| 792 | if (TranslatedPtr != Inserted.first->second) |
| 793 | return false; |
| 794 | // ... otherwise just skip it. |
| 795 | continue; |
| 796 | } |
| 797 | WorkList.push_back(std::make_pair(Pred, PredAddr)); |
| 798 | } |
| 799 | } |
| 800 | } |
| 801 | return true; |
| 802 | } |
| 803 | |
| 804 | /// Find all blocks that will unconditionally lead to the block BB and append |
| 805 | /// them to F. |
| 806 | static void findUnconditionalPreds(SmallVectorImpl<BasicBlock *> &Blocks, |
| 807 | BasicBlock *BB, DominatorTree *DT) { |
| 808 | for (BasicBlock *Pred : predecessors(BB)) { |
| 809 | if (Pred == BB) continue; |
| 810 | Instruction *PredTI = Pred->getTerminator(); |
| 811 | if (PredTI->getNumSuccessors() != 1) |
| 812 | continue; |
| 813 | |
| 814 | if (DT->isReachableFromEntry(Pred)) |
| 815 | Blocks.push_back(Pred); |
| 816 | } |
| 817 | } |
| 818 | |
| 819 | /// Handle frees of entire structures whose dependency is a store |
| 820 | /// to a field of that structure. |
| 821 | static bool handleFree(CallInst *F, AliasAnalysis *AA, |
| 822 | MemoryDependenceResults *MD, DominatorTree *DT, |
| 823 | const TargetLibraryInfo *TLI, |
| 824 | InstOverlapIntervalsTy &IOL, |
| 825 | MapVector<Instruction *, bool> &ThrowableInst) { |
| 826 | bool MadeChange = false; |
| 827 | |
| 828 | MemoryLocation Loc = MemoryLocation::getAfter(F->getOperand(0)); |
| 829 | SmallVector<BasicBlock *, 16> Blocks; |
| 830 | Blocks.push_back(F->getParent()); |
| 831 | |
| 832 | while (!Blocks.empty()) { |
| 833 | BasicBlock *BB = Blocks.pop_back_val(); |
| 834 | Instruction *InstPt = BB->getTerminator(); |
| 835 | if (BB == F->getParent()) InstPt = F; |
| 836 | |
| 837 | MemDepResult Dep = |
| 838 | MD->getPointerDependencyFrom(Loc, false, InstPt->getIterator(), BB); |
| 839 | while (Dep.isDef() || Dep.isClobber()) { |
| 840 | Instruction *Dependency = Dep.getInst(); |
| 841 | if (!hasAnalyzableMemoryWrite(Dependency, *TLI) || |
| 842 | !isRemovable(Dependency)) |
| 843 | break; |
| 844 | |
| 845 | Value *DepPointer = |
| 846 | getUnderlyingObject(getStoredPointerOperand(Dependency, *TLI)); |
| 847 | |
| 848 | // Check for aliasing. |
| 849 | if (!AA->isMustAlias(F->getArgOperand(0), DepPointer)) |
| 850 | break; |
| 851 | |
| 852 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Dead Store to soon to be freed memory:\n DEAD: " << *Dependency << '\n'; } } while (false) |
| 853 | dbgs() << "DSE: Dead Store to soon to be freed memory:\n DEAD: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Dead Store to soon to be freed memory:\n DEAD: " << *Dependency << '\n'; } } while (false) |
| 854 | << *Dependency << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Dead Store to soon to be freed memory:\n DEAD: " << *Dependency << '\n'; } } while (false); |
| 855 | |
| 856 | // DCE instructions only used to calculate that store. |
| 857 | BasicBlock::iterator BBI(Dependency); |
| 858 | deleteDeadInstruction(Dependency, &BBI, *MD, *TLI, IOL, |
| 859 | ThrowableInst); |
| 860 | ++NumFastStores; |
| 861 | MadeChange = true; |
| 862 | |
| 863 | // Inst's old Dependency is now deleted. Compute the next dependency, |
| 864 | // which may also be dead, as in |
| 865 | // s[0] = 0; |
| 866 | // s[1] = 0; // This has just been deleted. |
| 867 | // free(s); |
| 868 | Dep = MD->getPointerDependencyFrom(Loc, false, BBI, BB); |
| 869 | } |
| 870 | |
| 871 | if (Dep.isNonLocal()) |
| 872 | findUnconditionalPreds(Blocks, BB, DT); |
| 873 | } |
| 874 | |
| 875 | return MadeChange; |
| 876 | } |
| 877 | |
| 878 | /// Check to see if the specified location may alias any of the stack objects in |
| 879 | /// the DeadStackObjects set. If so, they become live because the location is |
| 880 | /// being loaded. |
| 881 | static void removeAccessedObjects(const MemoryLocation &LoadedLoc, |
| 882 | SmallSetVector<const Value *, 16> &DeadStackObjects, |
| 883 | const DataLayout &DL, AliasAnalysis *AA, |
| 884 | const TargetLibraryInfo *TLI, |
| 885 | const Function *F) { |
| 886 | const Value *UnderlyingPointer = getUnderlyingObject(LoadedLoc.Ptr); |
| 887 | |
| 888 | // A constant can't be in the dead pointer set. |
| 889 | if (isa<Constant>(UnderlyingPointer)) |
| 890 | return; |
| 891 | |
| 892 | // If the kill pointer can be easily reduced to an alloca, don't bother doing |
| 893 | // extraneous AA queries. |
| 894 | if (isa<AllocaInst>(UnderlyingPointer) || isa<Argument>(UnderlyingPointer)) { |
| 895 | DeadStackObjects.remove(UnderlyingPointer); |
| 896 | return; |
| 897 | } |
| 898 | |
| 899 | // Remove objects that could alias LoadedLoc. |
| 900 | DeadStackObjects.remove_if([&](const Value *I) { |
| 901 | // See if the loaded location could alias the stack location. |
| 902 | MemoryLocation StackLoc(I, getPointerSize(I, DL, *TLI, F)); |
| 903 | return !AA->isNoAlias(StackLoc, LoadedLoc); |
| 904 | }); |
| 905 | } |
| 906 | |
| 907 | /// Remove dead stores to stack-allocated locations in the function end block. |
| 908 | /// Ex: |
| 909 | /// %A = alloca i32 |
| 910 | /// ... |
| 911 | /// store i32 1, i32* %A |
| 912 | /// ret void |
| 913 | static bool handleEndBlock(BasicBlock &BB, AliasAnalysis *AA, |
| 914 | MemoryDependenceResults *MD, |
| 915 | const TargetLibraryInfo *TLI, |
| 916 | InstOverlapIntervalsTy &IOL, |
| 917 | MapVector<Instruction *, bool> &ThrowableInst) { |
| 918 | bool MadeChange = false; |
| 919 | |
| 920 | // Keep track of all of the stack objects that are dead at the end of the |
| 921 | // function. |
| 922 | SmallSetVector<const Value*, 16> DeadStackObjects; |
| 923 | |
| 924 | // Find all of the alloca'd pointers in the entry block. |
| 925 | BasicBlock &Entry = BB.getParent()->front(); |
| 926 | for (Instruction &I : Entry) { |
| 927 | if (isa<AllocaInst>(&I)) |
| 928 | DeadStackObjects.insert(&I); |
| 929 | |
| 930 | // Okay, so these are dead heap objects, but if the pointer never escapes |
| 931 | // then it's leaked by this function anyways. |
| 932 | else if (isAllocLikeFn(&I, TLI) && !PointerMayBeCaptured(&I, true, true)) |
| 933 | DeadStackObjects.insert(&I); |
| 934 | } |
| 935 | |
| 936 | // Treat byval or inalloca arguments the same, stores to them are dead at the |
| 937 | // end of the function. |
| 938 | for (Argument &AI : BB.getParent()->args()) |
| 939 | if (AI.hasPassPointeeByValueCopyAttr()) |
| 940 | DeadStackObjects.insert(&AI); |
| 941 | |
| 942 | const DataLayout &DL = BB.getModule()->getDataLayout(); |
| 943 | |
| 944 | // Scan the basic block backwards |
| 945 | for (BasicBlock::iterator BBI = BB.end(); BBI != BB.begin(); ){ |
| 946 | --BBI; |
| 947 | |
| 948 | // If we find a store, check to see if it points into a dead stack value. |
| 949 | if (hasAnalyzableMemoryWrite(&*BBI, *TLI) && isRemovable(&*BBI)) { |
| 950 | // See through pointer-to-pointer bitcasts |
| 951 | SmallVector<const Value *, 4> Pointers; |
| 952 | getUnderlyingObjects(getStoredPointerOperand(&*BBI, *TLI), Pointers); |
| 953 | |
| 954 | // Stores to stack values are valid candidates for removal. |
| 955 | bool AllDead = true; |
| 956 | for (const Value *Pointer : Pointers) |
| 957 | if (!DeadStackObjects.count(Pointer)) { |
| 958 | AllDead = false; |
| 959 | break; |
| 960 | } |
| 961 | |
| 962 | if (AllDead) { |
| 963 | Instruction *Dead = &*BBI; |
| 964 | |
| 965 | LLVM_DEBUG(dbgs() << "DSE: Dead Store at End of Block:\n DEAD: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Dead Store at End of Block:\n DEAD: " << *Dead << "\n Objects: "; for (SmallVectorImpl <const Value *>::iterator I = Pointers.begin(), E = Pointers .end(); I != E; ++I) { dbgs() << **I; if (std::next(I) != E) dbgs() << ", "; } dbgs() << '\n'; } } while ( false) |
| 966 | << *Dead << "\n Objects: ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Dead Store at End of Block:\n DEAD: " << *Dead << "\n Objects: "; for (SmallVectorImpl <const Value *>::iterator I = Pointers.begin(), E = Pointers .end(); I != E; ++I) { dbgs() << **I; if (std::next(I) != E) dbgs() << ", "; } dbgs() << '\n'; } } while ( false) |
| 967 | for (SmallVectorImpl<const Value *>::iterator I =do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Dead Store at End of Block:\n DEAD: " << *Dead << "\n Objects: "; for (SmallVectorImpl <const Value *>::iterator I = Pointers.begin(), E = Pointers .end(); I != E; ++I) { dbgs() << **I; if (std::next(I) != E) dbgs() << ", "; } dbgs() << '\n'; } } while ( false) |
| 968 | Pointers.begin(),do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Dead Store at End of Block:\n DEAD: " << *Dead << "\n Objects: "; for (SmallVectorImpl <const Value *>::iterator I = Pointers.begin(), E = Pointers .end(); I != E; ++I) { dbgs() << **I; if (std::next(I) != E) dbgs() << ", "; } dbgs() << '\n'; } } while ( false) |
| 969 | E = Pointers.end();do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Dead Store at End of Block:\n DEAD: " << *Dead << "\n Objects: "; for (SmallVectorImpl <const Value *>::iterator I = Pointers.begin(), E = Pointers .end(); I != E; ++I) { dbgs() << **I; if (std::next(I) != E) dbgs() << ", "; } dbgs() << '\n'; } } while ( false) |
| 970 | I != E; ++I) {do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Dead Store at End of Block:\n DEAD: " << *Dead << "\n Objects: "; for (SmallVectorImpl <const Value *>::iterator I = Pointers.begin(), E = Pointers .end(); I != E; ++I) { dbgs() << **I; if (std::next(I) != E) dbgs() << ", "; } dbgs() << '\n'; } } while ( false) |
| 971 | dbgs() << **I;do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Dead Store at End of Block:\n DEAD: " << *Dead << "\n Objects: "; for (SmallVectorImpl <const Value *>::iterator I = Pointers.begin(), E = Pointers .end(); I != E; ++I) { dbgs() << **I; if (std::next(I) != E) dbgs() << ", "; } dbgs() << '\n'; } } while ( false) |
| 972 | if (std::next(I) != E)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Dead Store at End of Block:\n DEAD: " << *Dead << "\n Objects: "; for (SmallVectorImpl <const Value *>::iterator I = Pointers.begin(), E = Pointers .end(); I != E; ++I) { dbgs() << **I; if (std::next(I) != E) dbgs() << ", "; } dbgs() << '\n'; } } while ( false) |
| 973 | dbgs() << ", ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Dead Store at End of Block:\n DEAD: " << *Dead << "\n Objects: "; for (SmallVectorImpl <const Value *>::iterator I = Pointers.begin(), E = Pointers .end(); I != E; ++I) { dbgs() << **I; if (std::next(I) != E) dbgs() << ", "; } dbgs() << '\n'; } } while ( false) |
| 974 | } dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Dead Store at End of Block:\n DEAD: " << *Dead << "\n Objects: "; for (SmallVectorImpl <const Value *>::iterator I = Pointers.begin(), E = Pointers .end(); I != E; ++I) { dbgs() << **I; if (std::next(I) != E) dbgs() << ", "; } dbgs() << '\n'; } } while ( false) |
| 975 | << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Dead Store at End of Block:\n DEAD: " << *Dead << "\n Objects: "; for (SmallVectorImpl <const Value *>::iterator I = Pointers.begin(), E = Pointers .end(); I != E; ++I) { dbgs() << **I; if (std::next(I) != E) dbgs() << ", "; } dbgs() << '\n'; } } while ( false); |
| 976 | |
| 977 | // DCE instructions only used to calculate that store. |
| 978 | deleteDeadInstruction(Dead, &BBI, *MD, *TLI, IOL, ThrowableInst, |
| 979 | &DeadStackObjects); |
| 980 | ++NumFastStores; |
| 981 | MadeChange = true; |
| 982 | continue; |
| 983 | } |
| 984 | } |
| 985 | |
| 986 | // Remove any dead non-memory-mutating instructions. |
| 987 | if (isInstructionTriviallyDead(&*BBI, TLI)) { |
| 988 | LLVM_DEBUG(dbgs() << "DSE: Removing trivially dead instruction:\n DEAD: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Removing trivially dead instruction:\n DEAD: " << *&*BBI << '\n'; } } while (false) |
| 989 | << *&*BBI << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Removing trivially dead instruction:\n DEAD: " << *&*BBI << '\n'; } } while (false); |
| 990 | deleteDeadInstruction(&*BBI, &BBI, *MD, *TLI, IOL, ThrowableInst, |
| 991 | &DeadStackObjects); |
| 992 | ++NumFastOther; |
| 993 | MadeChange = true; |
| 994 | continue; |
| 995 | } |
| 996 | |
| 997 | if (isa<AllocaInst>(BBI)) { |
| 998 | // Remove allocas from the list of dead stack objects; there can't be |
| 999 | // any references before the definition. |
| 1000 | DeadStackObjects.remove(&*BBI); |
| 1001 | continue; |
| 1002 | } |
| 1003 | |
| 1004 | if (auto *Call = dyn_cast<CallBase>(&*BBI)) { |
| 1005 | // Remove allocation function calls from the list of dead stack objects; |
| 1006 | // there can't be any references before the definition. |
| 1007 | if (isAllocLikeFn(&*BBI, TLI)) |
| 1008 | DeadStackObjects.remove(&*BBI); |
| 1009 | |
| 1010 | // If this call does not access memory, it can't be loading any of our |
| 1011 | // pointers. |
| 1012 | if (AA->doesNotAccessMemory(Call)) |
| 1013 | continue; |
| 1014 | |
| 1015 | // If the call might load from any of our allocas, then any store above |
| 1016 | // the call is live. |
| 1017 | DeadStackObjects.remove_if([&](const Value *I) { |
| 1018 | // See if the call site touches the value. |
| 1019 | return isRefSet(AA->getModRefInfo( |
| 1020 | Call, I, getPointerSize(I, DL, *TLI, BB.getParent()))); |
| 1021 | }); |
| 1022 | |
| 1023 | // If all of the allocas were clobbered by the call then we're not going |
| 1024 | // to find anything else to process. |
| 1025 | if (DeadStackObjects.empty()) |
| 1026 | break; |
| 1027 | |
| 1028 | continue; |
| 1029 | } |
| 1030 | |
| 1031 | // We can remove the dead stores, irrespective of the fence and its ordering |
| 1032 | // (release/acquire/seq_cst). Fences only constraints the ordering of |
| 1033 | // already visible stores, it does not make a store visible to other |
| 1034 | // threads. So, skipping over a fence does not change a store from being |
| 1035 | // dead. |
| 1036 | if (isa<FenceInst>(*BBI)) |
| 1037 | continue; |
| 1038 | |
| 1039 | MemoryLocation LoadedLoc; |
| 1040 | |
| 1041 | // If we encounter a use of the pointer, it is no longer considered dead |
| 1042 | if (LoadInst *L = dyn_cast<LoadInst>(BBI)) { |
| 1043 | if (!L->isUnordered()) // Be conservative with atomic/volatile load |
| 1044 | break; |
| 1045 | LoadedLoc = MemoryLocation::get(L); |
| 1046 | } else if (VAArgInst *V = dyn_cast<VAArgInst>(BBI)) { |
| 1047 | LoadedLoc = MemoryLocation::get(V); |
| 1048 | } else if (!BBI->mayReadFromMemory()) { |
| 1049 | // Instruction doesn't read memory. Note that stores that weren't removed |
| 1050 | // above will hit this case. |
| 1051 | continue; |
| 1052 | } else { |
| 1053 | // Unknown inst; assume it clobbers everything. |
| 1054 | break; |
| 1055 | } |
| 1056 | |
| 1057 | // Remove any allocas from the DeadPointer set that are loaded, as this |
| 1058 | // makes any stores above the access live. |
| 1059 | removeAccessedObjects(LoadedLoc, DeadStackObjects, DL, AA, TLI, BB.getParent()); |
| 1060 | |
| 1061 | // If all of the allocas were clobbered by the access then we're not going |
| 1062 | // to find anything else to process. |
| 1063 | if (DeadStackObjects.empty()) |
| 1064 | break; |
| 1065 | } |
| 1066 | |
| 1067 | return MadeChange; |
| 1068 | } |
| 1069 | |
| 1070 | static bool tryToShorten(Instruction *EarlierWrite, int64_t &EarlierOffset, |
| 1071 | uint64_t &EarlierSize, int64_t LaterOffset, |
| 1072 | uint64_t LaterSize, bool IsOverwriteEnd) { |
| 1073 | // TODO: base this on the target vector size so that if the earlier |
| 1074 | // store was too small to get vector writes anyway then its likely |
| 1075 | // a good idea to shorten it |
| 1076 | // Power of 2 vector writes are probably always a bad idea to optimize |
| 1077 | // as any store/memset/memcpy is likely using vector instructions so |
| 1078 | // shortening it to not vector size is likely to be slower |
| 1079 | auto *EarlierIntrinsic = cast<AnyMemIntrinsic>(EarlierWrite); |
| 1080 | unsigned EarlierWriteAlign = EarlierIntrinsic->getDestAlignment(); |
| 1081 | if (!IsOverwriteEnd) |
| 1082 | LaterOffset = int64_t(LaterOffset + LaterSize); |
| 1083 | |
| 1084 | if (!(isPowerOf2_64(LaterOffset) && EarlierWriteAlign <= LaterOffset) && |
| 1085 | !((EarlierWriteAlign != 0) && LaterOffset % EarlierWriteAlign == 0)) |
| 1086 | return false; |
| 1087 | |
| 1088 | int64_t NewLength = IsOverwriteEnd |
| 1089 | ? LaterOffset - EarlierOffset |
| 1090 | : EarlierSize - (LaterOffset - EarlierOffset); |
| 1091 | |
| 1092 | if (auto *AMI = dyn_cast<AtomicMemIntrinsic>(EarlierWrite)) { |
| 1093 | // When shortening an atomic memory intrinsic, the newly shortened |
| 1094 | // length must remain an integer multiple of the element size. |
| 1095 | const uint32_t ElementSize = AMI->getElementSizeInBytes(); |
| 1096 | if (0 != NewLength % ElementSize) |
| 1097 | return false; |
| 1098 | } |
| 1099 | |
| 1100 | LLVM_DEBUG(dbgs() << "DSE: Remove Dead Store:\n OW "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Remove Dead Store:\n OW " << (IsOverwriteEnd ? "END" : "BEGIN") << ": " << *EarlierWrite << "\n KILLER (offset " << LaterOffset << ", " << EarlierSize << ")\n"; } } while (false) |
| 1101 | << (IsOverwriteEnd ? "END" : "BEGIN") << ": "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Remove Dead Store:\n OW " << (IsOverwriteEnd ? "END" : "BEGIN") << ": " << *EarlierWrite << "\n KILLER (offset " << LaterOffset << ", " << EarlierSize << ")\n"; } } while (false) |
| 1102 | << *EarlierWrite << "\n KILLER (offset " << LaterOffsetdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Remove Dead Store:\n OW " << (IsOverwriteEnd ? "END" : "BEGIN") << ": " << *EarlierWrite << "\n KILLER (offset " << LaterOffset << ", " << EarlierSize << ")\n"; } } while (false) |
| 1103 | << ", " << EarlierSize << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Remove Dead Store:\n OW " << (IsOverwriteEnd ? "END" : "BEGIN") << ": " << *EarlierWrite << "\n KILLER (offset " << LaterOffset << ", " << EarlierSize << ")\n"; } } while (false); |
| 1104 | |
| 1105 | Value *EarlierWriteLength = EarlierIntrinsic->getLength(); |
| 1106 | Value *TrimmedLength = |
| 1107 | ConstantInt::get(EarlierWriteLength->getType(), NewLength); |
| 1108 | EarlierIntrinsic->setLength(TrimmedLength); |
| 1109 | |
| 1110 | EarlierSize = NewLength; |
| 1111 | if (!IsOverwriteEnd) { |
| 1112 | int64_t OffsetMoved = (LaterOffset - EarlierOffset); |
| 1113 | Value *Indices[1] = { |
| 1114 | ConstantInt::get(EarlierWriteLength->getType(), OffsetMoved)}; |
| 1115 | GetElementPtrInst *NewDestGEP = GetElementPtrInst::CreateInBounds( |
| 1116 | EarlierIntrinsic->getRawDest()->getType()->getPointerElementType(), |
| 1117 | EarlierIntrinsic->getRawDest(), Indices, "", EarlierWrite); |
| 1118 | NewDestGEP->setDebugLoc(EarlierIntrinsic->getDebugLoc()); |
| 1119 | EarlierIntrinsic->setDest(NewDestGEP); |
| 1120 | EarlierOffset = EarlierOffset + OffsetMoved; |
| 1121 | } |
| 1122 | return true; |
| 1123 | } |
| 1124 | |
| 1125 | static bool tryToShortenEnd(Instruction *EarlierWrite, |
| 1126 | OverlapIntervalsTy &IntervalMap, |
| 1127 | int64_t &EarlierStart, uint64_t &EarlierSize) { |
| 1128 | if (IntervalMap.empty() || !isShortenableAtTheEnd(EarlierWrite)) |
| 1129 | return false; |
| 1130 | |
| 1131 | OverlapIntervalsTy::iterator OII = --IntervalMap.end(); |
| 1132 | int64_t LaterStart = OII->second; |
| 1133 | uint64_t LaterSize = OII->first - LaterStart; |
| 1134 | |
| 1135 | assert(OII->first - LaterStart >= 0 && "Size expected to be positive")((OII->first - LaterStart >= 0 && "Size expected to be positive" ) ? static_cast<void> (0) : __assert_fail ("OII->first - LaterStart >= 0 && \"Size expected to be positive\"" , "/build/llvm-toolchain-snapshot-13~++20210301100612+564f5b0734bd/llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp" , 1135, __PRETTY_FUNCTION__)); |
| 1136 | |
| 1137 | if (LaterStart > EarlierStart && |
| 1138 | // Note: "LaterStart - EarlierStart" is known to be positive due to |
| 1139 | // preceding check. |
| 1140 | (uint64_t)(LaterStart - EarlierStart) < EarlierSize && |
| 1141 | // Note: "EarlierSize - (uint64_t)(LaterStart - EarlierStart)" is known to |
| 1142 | // be non negative due to preceding checks. |
| 1143 | LaterSize >= EarlierSize - (uint64_t)(LaterStart - EarlierStart)) { |
| 1144 | if (tryToShorten(EarlierWrite, EarlierStart, EarlierSize, LaterStart, |
| 1145 | LaterSize, true)) { |
| 1146 | IntervalMap.erase(OII); |
| 1147 | return true; |
| 1148 | } |
| 1149 | } |
| 1150 | return false; |
| 1151 | } |
| 1152 | |
| 1153 | static bool tryToShortenBegin(Instruction *EarlierWrite, |
| 1154 | OverlapIntervalsTy &IntervalMap, |
| 1155 | int64_t &EarlierStart, uint64_t &EarlierSize) { |
| 1156 | if (IntervalMap.empty() || !isShortenableAtTheBeginning(EarlierWrite)) |
| 1157 | return false; |
| 1158 | |
| 1159 | OverlapIntervalsTy::iterator OII = IntervalMap.begin(); |
| 1160 | int64_t LaterStart = OII->second; |
| 1161 | uint64_t LaterSize = OII->first - LaterStart; |
| 1162 | |
| 1163 | assert(OII->first - LaterStart >= 0 && "Size expected to be positive")((OII->first - LaterStart >= 0 && "Size expected to be positive" ) ? static_cast<void> (0) : __assert_fail ("OII->first - LaterStart >= 0 && \"Size expected to be positive\"" , "/build/llvm-toolchain-snapshot-13~++20210301100612+564f5b0734bd/llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp" , 1163, __PRETTY_FUNCTION__)); |
| 1164 | |
| 1165 | if (LaterStart <= EarlierStart && |
| 1166 | // Note: "EarlierStart - LaterStart" is known to be non negative due to |
| 1167 | // preceding check. |
| 1168 | LaterSize > (uint64_t)(EarlierStart - LaterStart)) { |
| 1169 | // Note: "LaterSize - (uint64_t)(EarlierStart - LaterStart)" is known to be |
| 1170 | // positive due to preceding checks. |
| 1171 | assert(LaterSize - (uint64_t)(EarlierStart - LaterStart) < EarlierSize &&((LaterSize - (uint64_t)(EarlierStart - LaterStart) < EarlierSize && "Should have been handled as OW_Complete") ? static_cast <void> (0) : __assert_fail ("LaterSize - (uint64_t)(EarlierStart - LaterStart) < EarlierSize && \"Should have been handled as OW_Complete\"" , "/build/llvm-toolchain-snapshot-13~++20210301100612+564f5b0734bd/llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp" , 1172, __PRETTY_FUNCTION__)) |
| 1172 | "Should have been handled as OW_Complete")((LaterSize - (uint64_t)(EarlierStart - LaterStart) < EarlierSize && "Should have been handled as OW_Complete") ? static_cast <void> (0) : __assert_fail ("LaterSize - (uint64_t)(EarlierStart - LaterStart) < EarlierSize && \"Should have been handled as OW_Complete\"" , "/build/llvm-toolchain-snapshot-13~++20210301100612+564f5b0734bd/llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp" , 1172, __PRETTY_FUNCTION__)); |
| 1173 | if (tryToShorten(EarlierWrite, EarlierStart, EarlierSize, LaterStart, |
| 1174 | LaterSize, false)) { |
| 1175 | IntervalMap.erase(OII); |
| 1176 | return true; |
| 1177 | } |
| 1178 | } |
| 1179 | return false; |
| 1180 | } |
| 1181 | |
| 1182 | static bool removePartiallyOverlappedStores(const DataLayout &DL, |
| 1183 | InstOverlapIntervalsTy &IOL, |
| 1184 | const TargetLibraryInfo &TLI) { |
| 1185 | bool Changed = false; |
| 1186 | for (auto OI : IOL) { |
| 1187 | Instruction *EarlierWrite = OI.first; |
| 1188 | MemoryLocation Loc = getLocForWrite(EarlierWrite, TLI); |
| 1189 | assert(isRemovable(EarlierWrite) && "Expect only removable instruction")((isRemovable(EarlierWrite) && "Expect only removable instruction" ) ? static_cast<void> (0) : __assert_fail ("isRemovable(EarlierWrite) && \"Expect only removable instruction\"" , "/build/llvm-toolchain-snapshot-13~++20210301100612+564f5b0734bd/llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp" , 1189, __PRETTY_FUNCTION__)); |
| 1190 | |
| 1191 | const Value *Ptr = Loc.Ptr->stripPointerCasts(); |
| 1192 | int64_t EarlierStart = 0; |
| 1193 | uint64_t EarlierSize = Loc.Size.getValue(); |
| 1194 | GetPointerBaseWithConstantOffset(Ptr, EarlierStart, DL); |
| 1195 | OverlapIntervalsTy &IntervalMap = OI.second; |
| 1196 | Changed |= |
| 1197 | tryToShortenEnd(EarlierWrite, IntervalMap, EarlierStart, EarlierSize); |
| 1198 | if (IntervalMap.empty()) |
| 1199 | continue; |
| 1200 | Changed |= |
| 1201 | tryToShortenBegin(EarlierWrite, IntervalMap, EarlierStart, EarlierSize); |
| 1202 | } |
| 1203 | return Changed; |
| 1204 | } |
| 1205 | |
| 1206 | static bool eliminateNoopStore(Instruction *Inst, BasicBlock::iterator &BBI, |
| 1207 | AliasAnalysis *AA, MemoryDependenceResults *MD, |
| 1208 | const DataLayout &DL, |
| 1209 | const TargetLibraryInfo *TLI, |
| 1210 | InstOverlapIntervalsTy &IOL, |
| 1211 | MapVector<Instruction *, bool> &ThrowableInst, |
| 1212 | DominatorTree *DT) { |
| 1213 | // Must be a store instruction. |
| 1214 | StoreInst *SI = dyn_cast<StoreInst>(Inst); |
| 1215 | if (!SI) |
| 1216 | return false; |
| 1217 | |
| 1218 | // If we're storing the same value back to a pointer that we just loaded from, |
| 1219 | // then the store can be removed. |
| 1220 | if (LoadInst *DepLoad = dyn_cast<LoadInst>(SI->getValueOperand())) { |
| 1221 | if (SI->getPointerOperand() == DepLoad->getPointerOperand() && |
| 1222 | isRemovable(SI) && |
| 1223 | memoryIsNotModifiedBetween(DepLoad, SI, *AA, DL, DT)) { |
| 1224 | |
| 1225 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Remove Store Of Load from same pointer:\n LOAD: " << *DepLoad << "\n STORE: " << *SI << '\n'; } } while (false) |
| 1226 | dbgs() << "DSE: Remove Store Of Load from same pointer:\n LOAD: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Remove Store Of Load from same pointer:\n LOAD: " << *DepLoad << "\n STORE: " << *SI << '\n'; } } while (false) |
| 1227 | << *DepLoad << "\n STORE: " << *SI << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Remove Store Of Load from same pointer:\n LOAD: " << *DepLoad << "\n STORE: " << *SI << '\n'; } } while (false); |
| 1228 | |
| 1229 | deleteDeadInstruction(SI, &BBI, *MD, *TLI, IOL, ThrowableInst); |
| 1230 | ++NumRedundantStores; |
| 1231 | return true; |
| 1232 | } |
| 1233 | } |
| 1234 | |
| 1235 | // Remove null stores into the calloc'ed objects |
| 1236 | Constant *StoredConstant = dyn_cast<Constant>(SI->getValueOperand()); |
| 1237 | if (StoredConstant && StoredConstant->isNullValue() && isRemovable(SI)) { |
| 1238 | Instruction *UnderlyingPointer = |
| 1239 | dyn_cast<Instruction>(getUnderlyingObject(SI->getPointerOperand())); |
| 1240 | |
| 1241 | if (UnderlyingPointer && isCallocLikeFn(UnderlyingPointer, TLI) && |
| 1242 | memoryIsNotModifiedBetween(UnderlyingPointer, SI, *AA, DL, DT)) { |
| 1243 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Remove null store to the calloc'ed object:\n DEAD: " << *Inst << "\n OBJECT: " << *UnderlyingPointer << '\n'; } } while (false) |
| 1244 | dbgs() << "DSE: Remove null store to the calloc'ed object:\n DEAD: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Remove null store to the calloc'ed object:\n DEAD: " << *Inst << "\n OBJECT: " << *UnderlyingPointer << '\n'; } } while (false) |
| 1245 | << *Inst << "\n OBJECT: " << *UnderlyingPointer << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Remove null store to the calloc'ed object:\n DEAD: " << *Inst << "\n OBJECT: " << *UnderlyingPointer << '\n'; } } while (false); |
| 1246 | |
| 1247 | deleteDeadInstruction(SI, &BBI, *MD, *TLI, IOL, ThrowableInst); |
| 1248 | ++NumRedundantStores; |
| 1249 | return true; |
| 1250 | } |
| 1251 | } |
| 1252 | return false; |
| 1253 | } |
| 1254 | |
| 1255 | template <typename AATy> |
| 1256 | static Constant *tryToMergePartialOverlappingStores( |
| 1257 | StoreInst *Earlier, StoreInst *Later, int64_t InstWriteOffset, |
| 1258 | int64_t DepWriteOffset, const DataLayout &DL, AATy &AA, DominatorTree *DT) { |
| 1259 | |
| 1260 | if (Earlier && isa<ConstantInt>(Earlier->getValueOperand()) && |
| 1261 | DL.typeSizeEqualsStoreSize(Earlier->getValueOperand()->getType()) && |
| 1262 | Later && isa<ConstantInt>(Later->getValueOperand()) && |
| 1263 | DL.typeSizeEqualsStoreSize(Later->getValueOperand()->getType()) && |
| 1264 | memoryIsNotModifiedBetween(Earlier, Later, AA, DL, DT)) { |
| 1265 | // If the store we find is: |
| 1266 | // a) partially overwritten by the store to 'Loc' |
| 1267 | // b) the later store is fully contained in the earlier one and |
| 1268 | // c) they both have a constant value |
| 1269 | // d) none of the two stores need padding |
| 1270 | // Merge the two stores, replacing the earlier store's value with a |
| 1271 | // merge of both values. |
| 1272 | // TODO: Deal with other constant types (vectors, etc), and probably |
| 1273 | // some mem intrinsics (if needed) |
| 1274 | |
| 1275 | APInt EarlierValue = |
| 1276 | cast<ConstantInt>(Earlier->getValueOperand())->getValue(); |
| 1277 | APInt LaterValue = cast<ConstantInt>(Later->getValueOperand())->getValue(); |
| 1278 | unsigned LaterBits = LaterValue.getBitWidth(); |
| 1279 | assert(EarlierValue.getBitWidth() > LaterValue.getBitWidth())((EarlierValue.getBitWidth() > LaterValue.getBitWidth()) ? static_cast<void> (0) : __assert_fail ("EarlierValue.getBitWidth() > LaterValue.getBitWidth()" , "/build/llvm-toolchain-snapshot-13~++20210301100612+564f5b0734bd/llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp" , 1279, __PRETTY_FUNCTION__)); |
| 1280 | LaterValue = LaterValue.zext(EarlierValue.getBitWidth()); |
| 1281 | |
| 1282 | // Offset of the smaller store inside the larger store |
| 1283 | unsigned BitOffsetDiff = (InstWriteOffset - DepWriteOffset) * 8; |
| 1284 | unsigned LShiftAmount = DL.isBigEndian() ? EarlierValue.getBitWidth() - |
| 1285 | BitOffsetDiff - LaterBits |
| 1286 | : BitOffsetDiff; |
| 1287 | APInt Mask = APInt::getBitsSet(EarlierValue.getBitWidth(), LShiftAmount, |
| 1288 | LShiftAmount + LaterBits); |
| 1289 | // Clear the bits we'll be replacing, then OR with the smaller |
| 1290 | // store, shifted appropriately. |
| 1291 | APInt Merged = (EarlierValue & ~Mask) | (LaterValue << LShiftAmount); |
| 1292 | LLVM_DEBUG(dbgs() << "DSE: Merge Stores:\n Earlier: " << *Earlierdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Merge Stores:\n Earlier: " << *Earlier << "\n Later: " << *Later << "\n Merged Value: " << Merged << '\n'; } } while (false) |
| 1293 | << "\n Later: " << *Laterdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Merge Stores:\n Earlier: " << *Earlier << "\n Later: " << *Later << "\n Merged Value: " << Merged << '\n'; } } while (false) |
| 1294 | << "\n Merged Value: " << Merged << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Merge Stores:\n Earlier: " << *Earlier << "\n Later: " << *Later << "\n Merged Value: " << Merged << '\n'; } } while (false); |
| 1295 | return ConstantInt::get(Earlier->getValueOperand()->getType(), Merged); |
| 1296 | } |
| 1297 | return nullptr; |
| 1298 | } |
| 1299 | |
| 1300 | static bool eliminateDeadStores(BasicBlock &BB, AliasAnalysis *AA, |
| 1301 | MemoryDependenceResults *MD, DominatorTree *DT, |
| 1302 | const TargetLibraryInfo *TLI) { |
| 1303 | const DataLayout &DL = BB.getModule()->getDataLayout(); |
| 1304 | bool MadeChange = false; |
| 1305 | |
| 1306 | MapVector<Instruction *, bool> ThrowableInst; |
| 1307 | |
| 1308 | // A map of interval maps representing partially-overwritten value parts. |
| 1309 | InstOverlapIntervalsTy IOL; |
| 1310 | |
| 1311 | // Do a top-down walk on the BB. |
| 1312 | for (BasicBlock::iterator BBI = BB.begin(), BBE = BB.end(); BBI != BBE; ) { |
| 1313 | // Handle 'free' calls specially. |
| 1314 | if (CallInst *F = isFreeCall(&*BBI, TLI)) { |
| 1315 | MadeChange |= handleFree(F, AA, MD, DT, TLI, IOL, ThrowableInst); |
| 1316 | // Increment BBI after handleFree has potentially deleted instructions. |
| 1317 | // This ensures we maintain a valid iterator. |
| 1318 | ++BBI; |
| 1319 | continue; |
| 1320 | } |
| 1321 | |
| 1322 | Instruction *Inst = &*BBI++; |
| 1323 | |
| 1324 | if (Inst->mayThrow()) { |
| 1325 | ThrowableInst[Inst] = true; |
| 1326 | continue; |
| 1327 | } |
| 1328 | |
| 1329 | // Check to see if Inst writes to memory. If not, continue. |
| 1330 | if (!hasAnalyzableMemoryWrite(Inst, *TLI)) |
| 1331 | continue; |
| 1332 | |
| 1333 | // eliminateNoopStore will update in iterator, if necessary. |
| 1334 | if (eliminateNoopStore(Inst, BBI, AA, MD, DL, TLI, IOL, |
| 1335 | ThrowableInst, DT)) { |
| 1336 | MadeChange = true; |
| 1337 | continue; |
| 1338 | } |
| 1339 | |
| 1340 | // If we find something that writes memory, get its memory dependence. |
| 1341 | MemDepResult InstDep = MD->getDependency(Inst); |
| 1342 | |
| 1343 | // Ignore any store where we can't find a local dependence. |
| 1344 | // FIXME: cross-block DSE would be fun. :) |
| 1345 | if (!InstDep.isDef() && !InstDep.isClobber()) |
| 1346 | continue; |
| 1347 | |
| 1348 | // Figure out what location is being stored to. |
| 1349 | MemoryLocation Loc = getLocForWrite(Inst, *TLI); |
| 1350 | |
| 1351 | // If we didn't get a useful location, fail. |
| 1352 | if (!Loc.Ptr) |
| 1353 | continue; |
| 1354 | |
| 1355 | // Loop until we find a store we can eliminate or a load that |
| 1356 | // invalidates the analysis. Without an upper bound on the number of |
| 1357 | // instructions examined, this analysis can become very time-consuming. |
| 1358 | // However, the potential gain diminishes as we process more instructions |
| 1359 | // without eliminating any of them. Therefore, we limit the number of |
| 1360 | // instructions we look at. |
| 1361 | auto Limit = MD->getDefaultBlockScanLimit(); |
| 1362 | while (InstDep.isDef() || InstDep.isClobber()) { |
| 1363 | // Get the memory clobbered by the instruction we depend on. MemDep will |
| 1364 | // skip any instructions that 'Loc' clearly doesn't interact with. If we |
| 1365 | // end up depending on a may- or must-aliased load, then we can't optimize |
| 1366 | // away the store and we bail out. However, if we depend on something |
| 1367 | // that overwrites the memory location we *can* potentially optimize it. |
| 1368 | // |
| 1369 | // Find out what memory location the dependent instruction stores. |
| 1370 | Instruction *DepWrite = InstDep.getInst(); |
| 1371 | if (!hasAnalyzableMemoryWrite(DepWrite, *TLI)) |
| 1372 | break; |
| 1373 | MemoryLocation DepLoc = getLocForWrite(DepWrite, *TLI); |
| 1374 | // If we didn't get a useful location, or if it isn't a size, bail out. |
| 1375 | if (!DepLoc.Ptr) |
| 1376 | break; |
| 1377 | |
| 1378 | // Find the last throwable instruction not removed by call to |
| 1379 | // deleteDeadInstruction. |
| 1380 | Instruction *LastThrowing = nullptr; |
| 1381 | if (!ThrowableInst.empty()) |
| 1382 | LastThrowing = ThrowableInst.back().first; |
| 1383 | |
| 1384 | // Make sure we don't look past a call which might throw. This is an |
| 1385 | // issue because MemoryDependenceAnalysis works in the wrong direction: |
| 1386 | // it finds instructions which dominate the current instruction, rather than |
| 1387 | // instructions which are post-dominated by the current instruction. |
| 1388 | // |
| 1389 | // If the underlying object is a non-escaping memory allocation, any store |
| 1390 | // to it is dead along the unwind edge. Otherwise, we need to preserve |
| 1391 | // the store. |
| 1392 | if (LastThrowing && DepWrite->comesBefore(LastThrowing)) { |
| 1393 | const Value *Underlying = getUnderlyingObject(DepLoc.Ptr); |
| 1394 | bool IsStoreDeadOnUnwind = isa<AllocaInst>(Underlying); |
| 1395 | if (!IsStoreDeadOnUnwind) { |
| 1396 | // We're looking for a call to an allocation function |
| 1397 | // where the allocation doesn't escape before the last |
| 1398 | // throwing instruction; PointerMayBeCaptured |
| 1399 | // reasonably fast approximation. |
| 1400 | IsStoreDeadOnUnwind = isAllocLikeFn(Underlying, TLI) && |
| 1401 | !PointerMayBeCaptured(Underlying, false, true); |
| 1402 | } |
| 1403 | if (!IsStoreDeadOnUnwind) |
| 1404 | break; |
| 1405 | } |
| 1406 | |
| 1407 | // If we find a write that is a) removable (i.e., non-volatile), b) is |
| 1408 | // completely obliterated by the store to 'Loc', and c) which we know that |
| 1409 | // 'Inst' doesn't load from, then we can remove it. |
| 1410 | // Also try to merge two stores if a later one only touches memory written |
| 1411 | // to by the earlier one. |
| 1412 | if (isRemovable(DepWrite) && |
| 1413 | !isPossibleSelfRead(Inst, Loc, DepWrite, *TLI, *AA)) { |
| 1414 | int64_t InstWriteOffset, DepWriteOffset; |
| 1415 | OverwriteResult OR = isOverwrite(Inst, DepWrite, Loc, DepLoc, DL, *TLI, |
| 1416 | DepWriteOffset, InstWriteOffset, *AA, |
| 1417 | BB.getParent()); |
| 1418 | if (OR == OW_MaybePartial) |
| 1419 | OR = isPartialOverwrite(Loc, DepLoc, DepWriteOffset, InstWriteOffset, |
| 1420 | DepWrite, IOL); |
| 1421 | |
| 1422 | if (OR == OW_Complete) { |
| 1423 | LLVM_DEBUG(dbgs() << "DSE: Remove Dead Store:\n DEAD: " << *DepWritedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Remove Dead Store:\n DEAD: " << *DepWrite << "\n KILLER: " << *Inst << '\n'; } } while (false) |
| 1424 | << "\n KILLER: " << *Inst << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Remove Dead Store:\n DEAD: " << *DepWrite << "\n KILLER: " << *Inst << '\n'; } } while (false); |
| 1425 | |
| 1426 | // Delete the store and now-dead instructions that feed it. |
| 1427 | deleteDeadInstruction(DepWrite, &BBI, *MD, *TLI, IOL, |
| 1428 | ThrowableInst); |
| 1429 | ++NumFastStores; |
| 1430 | MadeChange = true; |
| 1431 | |
| 1432 | // We erased DepWrite; start over. |
| 1433 | InstDep = MD->getDependency(Inst); |
| 1434 | continue; |
| 1435 | } else if ((OR == OW_End && isShortenableAtTheEnd(DepWrite)) || |
| 1436 | ((OR == OW_Begin && |
| 1437 | isShortenableAtTheBeginning(DepWrite)))) { |
| 1438 | assert(!EnablePartialOverwriteTracking && "Do not expect to perform "((!EnablePartialOverwriteTracking && "Do not expect to perform " "when partial-overwrite " "tracking is enabled") ? static_cast <void> (0) : __assert_fail ("!EnablePartialOverwriteTracking && \"Do not expect to perform \" \"when partial-overwrite \" \"tracking is enabled\"" , "/build/llvm-toolchain-snapshot-13~++20210301100612+564f5b0734bd/llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp" , 1440, __PRETTY_FUNCTION__)) |
| 1439 | "when partial-overwrite "((!EnablePartialOverwriteTracking && "Do not expect to perform " "when partial-overwrite " "tracking is enabled") ? static_cast <void> (0) : __assert_fail ("!EnablePartialOverwriteTracking && \"Do not expect to perform \" \"when partial-overwrite \" \"tracking is enabled\"" , "/build/llvm-toolchain-snapshot-13~++20210301100612+564f5b0734bd/llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp" , 1440, __PRETTY_FUNCTION__)) |
| 1440 | "tracking is enabled")((!EnablePartialOverwriteTracking && "Do not expect to perform " "when partial-overwrite " "tracking is enabled") ? static_cast <void> (0) : __assert_fail ("!EnablePartialOverwriteTracking && \"Do not expect to perform \" \"when partial-overwrite \" \"tracking is enabled\"" , "/build/llvm-toolchain-snapshot-13~++20210301100612+564f5b0734bd/llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp" , 1440, __PRETTY_FUNCTION__)); |
| 1441 | // The overwrite result is known, so these must be known, too. |
| 1442 | uint64_t EarlierSize = DepLoc.Size.getValue(); |
| 1443 | uint64_t LaterSize = Loc.Size.getValue(); |
| 1444 | bool IsOverwriteEnd = (OR == OW_End); |
| 1445 | MadeChange |= tryToShorten(DepWrite, DepWriteOffset, EarlierSize, |
| 1446 | InstWriteOffset, LaterSize, IsOverwriteEnd); |
| 1447 | } else if (EnablePartialStoreMerging && |
| 1448 | OR == OW_PartialEarlierWithFullLater) { |
| 1449 | auto *Earlier = dyn_cast<StoreInst>(DepWrite); |
| 1450 | auto *Later = dyn_cast<StoreInst>(Inst); |
| 1451 | if (Constant *C = tryToMergePartialOverlappingStores( |
| 1452 | Earlier, Later, InstWriteOffset, DepWriteOffset, DL, *AA, |
| 1453 | DT)) { |
| 1454 | auto *SI = new StoreInst( |
| 1455 | C, Earlier->getPointerOperand(), false, Earlier->getAlign(), |
| 1456 | Earlier->getOrdering(), Earlier->getSyncScopeID(), DepWrite); |
| 1457 | |
| 1458 | unsigned MDToKeep[] = {LLVMContext::MD_dbg, LLVMContext::MD_tbaa, |
| 1459 | LLVMContext::MD_alias_scope, |
| 1460 | LLVMContext::MD_noalias, |
| 1461 | LLVMContext::MD_nontemporal}; |
| 1462 | SI->copyMetadata(*DepWrite, MDToKeep); |
| 1463 | ++NumModifiedStores; |
| 1464 | |
| 1465 | // Delete the old stores and now-dead instructions that feed them. |
| 1466 | deleteDeadInstruction(Inst, &BBI, *MD, *TLI, IOL, |
| 1467 | ThrowableInst); |
| 1468 | deleteDeadInstruction(DepWrite, &BBI, *MD, *TLI, IOL, |
| 1469 | ThrowableInst); |
| 1470 | MadeChange = true; |
| 1471 | |
| 1472 | // We erased DepWrite and Inst (Loc); start over. |
| 1473 | break; |
| 1474 | } |
| 1475 | } |
| 1476 | } |
| 1477 | |
| 1478 | // If this is a may-aliased store that is clobbering the store value, we |
| 1479 | // can keep searching past it for another must-aliased pointer that stores |
| 1480 | // to the same location. For example, in: |
| 1481 | // store -> P |
| 1482 | // store -> Q |
| 1483 | // store -> P |
| 1484 | // we can remove the first store to P even though we don't know if P and Q |
| 1485 | // alias. |
| 1486 | if (DepWrite == &BB.front()) break; |
| 1487 | |
| 1488 | // Can't look past this instruction if it might read 'Loc'. |
| 1489 | if (isRefSet(AA->getModRefInfo(DepWrite, Loc))) |
| 1490 | break; |
| 1491 | |
| 1492 | InstDep = MD->getPointerDependencyFrom(Loc, /*isLoad=*/ false, |
| 1493 | DepWrite->getIterator(), &BB, |
| 1494 | /*QueryInst=*/ nullptr, &Limit); |
| 1495 | } |
| 1496 | } |
| 1497 | |
| 1498 | if (EnablePartialOverwriteTracking) |
| 1499 | MadeChange |= removePartiallyOverlappedStores(DL, IOL, *TLI); |
| 1500 | |
| 1501 | // If this block ends in a return, unwind, or unreachable, all allocas are |
| 1502 | // dead at its end, which means stores to them are also dead. |
| 1503 | if (BB.getTerminator()->getNumSuccessors() == 0) |
| 1504 | MadeChange |= handleEndBlock(BB, AA, MD, TLI, IOL, ThrowableInst); |
| 1505 | |
| 1506 | return MadeChange; |
| 1507 | } |
| 1508 | |
| 1509 | static bool eliminateDeadStores(Function &F, AliasAnalysis *AA, |
| 1510 | MemoryDependenceResults *MD, DominatorTree *DT, |
| 1511 | const TargetLibraryInfo *TLI) { |
| 1512 | bool MadeChange = false; |
| 1513 | for (BasicBlock &BB : F) |
| 1514 | // Only check non-dead blocks. Dead blocks may have strange pointer |
| 1515 | // cycles that will confuse alias analysis. |
| 1516 | if (DT->isReachableFromEntry(&BB)) |
| 1517 | MadeChange |= eliminateDeadStores(BB, AA, MD, DT, TLI); |
| 1518 | |
| 1519 | return MadeChange; |
| 1520 | } |
| 1521 | |
| 1522 | namespace { |
| 1523 | //============================================================================= |
| 1524 | // MemorySSA backed dead store elimination. |
| 1525 | // |
| 1526 | // The code below implements dead store elimination using MemorySSA. It uses |
| 1527 | // the following general approach: given a MemoryDef, walk upwards to find |
| 1528 | // clobbering MemoryDefs that may be killed by the starting def. Then check |
| 1529 | // that there are no uses that may read the location of the original MemoryDef |
| 1530 | // in between both MemoryDefs. A bit more concretely: |
| 1531 | // |
| 1532 | // For all MemoryDefs StartDef: |
| 1533 | // 1. Get the next dominating clobbering MemoryDef (EarlierAccess) by walking |
| 1534 | // upwards. |
| 1535 | // 2. Check that there are no reads between EarlierAccess and the StartDef by |
| 1536 | // checking all uses starting at EarlierAccess and walking until we see |
| 1537 | // StartDef. |
| 1538 | // 3. For each found CurrentDef, check that: |
| 1539 | // 1. There are no barrier instructions between CurrentDef and StartDef (like |
| 1540 | // throws or stores with ordering constraints). |
| 1541 | // 2. StartDef is executed whenever CurrentDef is executed. |
| 1542 | // 3. StartDef completely overwrites CurrentDef. |
| 1543 | // 4. Erase CurrentDef from the function and MemorySSA. |
| 1544 | |
| 1545 | // Returns true if \p I is an intrisnic that does not read or write memory. |
| 1546 | bool isNoopIntrinsic(Instruction *I) { |
| 1547 | if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { |
| 1548 | switch (II->getIntrinsicID()) { |
| 1549 | case Intrinsic::lifetime_start: |
| 1550 | case Intrinsic::lifetime_end: |
| 1551 | case Intrinsic::invariant_end: |
| 1552 | case Intrinsic::launder_invariant_group: |
| 1553 | case Intrinsic::assume: |
| 1554 | return true; |
| 1555 | case Intrinsic::dbg_addr: |
| 1556 | case Intrinsic::dbg_declare: |
| 1557 | case Intrinsic::dbg_label: |
| 1558 | case Intrinsic::dbg_value: |
| 1559 | llvm_unreachable("Intrinsic should not be modeled in MemorySSA")::llvm::llvm_unreachable_internal("Intrinsic should not be modeled in MemorySSA" , "/build/llvm-toolchain-snapshot-13~++20210301100612+564f5b0734bd/llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp" , 1559); |
| 1560 | default: |
| 1561 | return false; |
| 1562 | } |
| 1563 | } |
| 1564 | return false; |
| 1565 | } |
| 1566 | |
| 1567 | // Check if we can ignore \p D for DSE. |
| 1568 | bool canSkipDef(MemoryDef *D, bool DefVisibleToCaller) { |
| 1569 | Instruction *DI = D->getMemoryInst(); |
| 1570 | // Calls that only access inaccessible memory cannot read or write any memory |
| 1571 | // locations we consider for elimination. |
| 1572 | if (auto *CB = dyn_cast<CallBase>(DI)) |
| 1573 | if (CB->onlyAccessesInaccessibleMemory()) |
| 1574 | return true; |
| 1575 | |
| 1576 | // We can eliminate stores to locations not visible to the caller across |
| 1577 | // throwing instructions. |
| 1578 | if (DI->mayThrow() && !DefVisibleToCaller) |
| 1579 | return true; |
| 1580 | |
| 1581 | // We can remove the dead stores, irrespective of the fence and its ordering |
| 1582 | // (release/acquire/seq_cst). Fences only constraints the ordering of |
| 1583 | // already visible stores, it does not make a store visible to other |
| 1584 | // threads. So, skipping over a fence does not change a store from being |
| 1585 | // dead. |
| 1586 | if (isa<FenceInst>(DI)) |
| 1587 | return true; |
| 1588 | |
| 1589 | // Skip intrinsics that do not really read or modify memory. |
| 1590 | if (isNoopIntrinsic(D->getMemoryInst())) |
| 1591 | return true; |
| 1592 | |
| 1593 | return false; |
| 1594 | } |
| 1595 | |
| 1596 | struct DSEState { |
| 1597 | Function &F; |
| 1598 | AliasAnalysis &AA; |
| 1599 | |
| 1600 | /// The single BatchAA instance that is used to cache AA queries. It will |
| 1601 | /// not be invalidated over the whole run. This is safe, because: |
| 1602 | /// 1. Only memory writes are removed, so the alias cache for memory |
| 1603 | /// locations remains valid. |
| 1604 | /// 2. No new instructions are added (only instructions removed), so cached |
| 1605 | /// information for a deleted value cannot be accessed by a re-used new |
| 1606 | /// value pointer. |
| 1607 | BatchAAResults BatchAA; |
| 1608 | |
| 1609 | MemorySSA &MSSA; |
| 1610 | DominatorTree &DT; |
| 1611 | PostDominatorTree &PDT; |
| 1612 | const TargetLibraryInfo &TLI; |
| 1613 | const DataLayout &DL; |
| 1614 | |
| 1615 | // All MemoryDefs that potentially could kill other MemDefs. |
| 1616 | SmallVector<MemoryDef *, 64> MemDefs; |
| 1617 | // Any that should be skipped as they are already deleted |
| 1618 | SmallPtrSet<MemoryAccess *, 4> SkipStores; |
| 1619 | // Keep track of all of the objects that are invisible to the caller before |
| 1620 | // the function returns. |
| 1621 | // SmallPtrSet<const Value *, 16> InvisibleToCallerBeforeRet; |
| 1622 | DenseMap<const Value *, bool> InvisibleToCallerBeforeRet; |
| 1623 | // Keep track of all of the objects that are invisible to the caller after |
| 1624 | // the function returns. |
| 1625 | DenseMap<const Value *, bool> InvisibleToCallerAfterRet; |
| 1626 | // Keep track of blocks with throwing instructions not modeled in MemorySSA. |
| 1627 | SmallPtrSet<BasicBlock *, 16> ThrowingBlocks; |
| 1628 | // Post-order numbers for each basic block. Used to figure out if memory |
| 1629 | // accesses are executed before another access. |
| 1630 | DenseMap<BasicBlock *, unsigned> PostOrderNumbers; |
| 1631 | |
| 1632 | /// Keep track of instructions (partly) overlapping with killing MemoryDefs per |
| 1633 | /// basic block. |
| 1634 | DenseMap<BasicBlock *, InstOverlapIntervalsTy> IOLs; |
| 1635 | |
| 1636 | DSEState(Function &F, AliasAnalysis &AA, MemorySSA &MSSA, DominatorTree &DT, |
| 1637 | PostDominatorTree &PDT, const TargetLibraryInfo &TLI) |
| 1638 | : F(F), AA(AA), BatchAA(AA), MSSA(MSSA), DT(DT), PDT(PDT), TLI(TLI), |
| 1639 | DL(F.getParent()->getDataLayout()) {} |
| 1640 | |
| 1641 | static DSEState get(Function &F, AliasAnalysis &AA, MemorySSA &MSSA, |
| 1642 | DominatorTree &DT, PostDominatorTree &PDT, |
| 1643 | const TargetLibraryInfo &TLI) { |
| 1644 | DSEState State(F, AA, MSSA, DT, PDT, TLI); |
| 1645 | // Collect blocks with throwing instructions not modeled in MemorySSA and |
| 1646 | // alloc-like objects. |
| 1647 | unsigned PO = 0; |
| 1648 | for (BasicBlock *BB : post_order(&F)) { |
| 1649 | State.PostOrderNumbers[BB] = PO++; |
| 1650 | for (Instruction &I : *BB) { |
| 1651 | MemoryAccess *MA = MSSA.getMemoryAccess(&I); |
| 1652 | if (I.mayThrow() && !MA) |
| 1653 | State.ThrowingBlocks.insert(I.getParent()); |
| 1654 | |
| 1655 | auto *MD = dyn_cast_or_null<MemoryDef>(MA); |
| 1656 | if (MD && State.MemDefs.size() < MemorySSADefsPerBlockLimit && |
| 1657 | (State.getLocForWriteEx(&I) || State.isMemTerminatorInst(&I))) |
| 1658 | State.MemDefs.push_back(MD); |
| 1659 | } |
| 1660 | } |
| 1661 | |
| 1662 | // Treat byval or inalloca arguments the same as Allocas, stores to them are |
| 1663 | // dead at the end of the function. |
| 1664 | for (Argument &AI : F.args()) |
| 1665 | if (AI.hasPassPointeeByValueCopyAttr()) { |
| 1666 | // For byval, the caller doesn't know the address of the allocation. |
| 1667 | if (AI.hasByValAttr()) |
| 1668 | State.InvisibleToCallerBeforeRet.insert({&AI, true}); |
| 1669 | State.InvisibleToCallerAfterRet.insert({&AI, true}); |
| 1670 | } |
| 1671 | |
| 1672 | return State; |
| 1673 | } |
| 1674 | |
| 1675 | bool isInvisibleToCallerAfterRet(const Value *V) { |
| 1676 | if (isa<AllocaInst>(V)) |
| 1677 | return true; |
| 1678 | auto I = InvisibleToCallerAfterRet.insert({V, false}); |
| 1679 | if (I.second) { |
| 1680 | if (!isInvisibleToCallerBeforeRet(V)) { |
| 1681 | I.first->second = false; |
| 1682 | } else { |
| 1683 | auto *Inst = dyn_cast<Instruction>(V); |
| 1684 | if (Inst && isAllocLikeFn(Inst, &TLI)) |
| 1685 | I.first->second = !PointerMayBeCaptured(V, true, false); |
| 1686 | } |
| 1687 | } |
| 1688 | return I.first->second; |
| 1689 | } |
| 1690 | |
| 1691 | bool isInvisibleToCallerBeforeRet(const Value *V) { |
| 1692 | if (isa<AllocaInst>(V)) |
| 1693 | return true; |
| 1694 | auto I = InvisibleToCallerBeforeRet.insert({V, false}); |
| 1695 | if (I.second) { |
| 1696 | auto *Inst = dyn_cast<Instruction>(V); |
| 1697 | if (Inst && isAllocLikeFn(Inst, &TLI)) |
| 1698 | // NOTE: This could be made more precise by PointerMayBeCapturedBefore |
| 1699 | // with the killing MemoryDef. But we refrain from doing so for now to |
| 1700 | // limit compile-time and this does not cause any changes to the number |
| 1701 | // of stores removed on a large test set in practice. |
| 1702 | I.first->second = !PointerMayBeCaptured(V, false, true); |
| 1703 | } |
| 1704 | return I.first->second; |
| 1705 | } |
| 1706 | |
| 1707 | Optional<MemoryLocation> getLocForWriteEx(Instruction *I) const { |
| 1708 | if (!I->mayWriteToMemory()) |
| 1709 | return None; |
| 1710 | |
| 1711 | if (auto *MTI = dyn_cast<AnyMemIntrinsic>(I)) |
| 1712 | return {MemoryLocation::getForDest(MTI)}; |
| 1713 | |
| 1714 | if (auto *CB = dyn_cast<CallBase>(I)) { |
| 1715 | // If the functions may write to memory we do not know about, bail out. |
| 1716 | if (!CB->onlyAccessesArgMemory() && |
| 1717 | !CB->onlyAccessesInaccessibleMemOrArgMem()) |
| 1718 | return None; |
| 1719 | |
| 1720 | LibFunc LF; |
| 1721 | if (TLI.getLibFunc(*CB, LF) && TLI.has(LF)) { |
| 1722 | switch (LF) { |
| 1723 | case LibFunc_strcpy: |
| 1724 | case LibFunc_strncpy: |
| 1725 | case LibFunc_strcat: |
| 1726 | case LibFunc_strncat: |
| 1727 | return {MemoryLocation::getAfter(CB->getArgOperand(0))}; |
| 1728 | default: |
| 1729 | break; |
| 1730 | } |
| 1731 | } |
| 1732 | switch (CB->getIntrinsicID()) { |
| 1733 | case Intrinsic::init_trampoline: |
| 1734 | return {MemoryLocation::getAfter(CB->getArgOperand(0))}; |
| 1735 | case Intrinsic::masked_store: |
| 1736 | return {MemoryLocation::getForArgument(CB, 1, TLI)}; |
| 1737 | default: |
| 1738 | break; |
| 1739 | } |
| 1740 | return None; |
| 1741 | } |
| 1742 | |
| 1743 | return MemoryLocation::getOrNone(I); |
| 1744 | } |
| 1745 | |
| 1746 | /// Returns true if \p UseInst completely overwrites \p DefLoc |
| 1747 | /// (stored by \p DefInst). |
| 1748 | bool isCompleteOverwrite(const MemoryLocation &DefLoc, Instruction *DefInst, |
| 1749 | Instruction *UseInst) { |
| 1750 | // UseInst has a MemoryDef associated in MemorySSA. It's possible for a |
| 1751 | // MemoryDef to not write to memory, e.g. a volatile load is modeled as a |
| 1752 | // MemoryDef. |
| 1753 | if (!UseInst->mayWriteToMemory()) |
| 1754 | return false; |
| 1755 | |
| 1756 | if (auto *CB = dyn_cast<CallBase>(UseInst)) |
| 1757 | if (CB->onlyAccessesInaccessibleMemory()) |
| 1758 | return false; |
| 1759 | |
| 1760 | int64_t InstWriteOffset, DepWriteOffset; |
| 1761 | if (auto CC = getLocForWriteEx(UseInst)) |
| 1762 | return isOverwrite(UseInst, DefInst, *CC, DefLoc, DL, TLI, DepWriteOffset, |
| 1763 | InstWriteOffset, BatchAA, &F) == OW_Complete; |
| 1764 | return false; |
| 1765 | } |
| 1766 | |
| 1767 | /// Returns true if \p Def is not read before returning from the function. |
| 1768 | bool isWriteAtEndOfFunction(MemoryDef *Def) { |
| 1769 | LLVM_DEBUG(dbgs() << " Check if def " << *Def << " ("do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << " Check if def " << *Def << " (" << *Def->getMemoryInst() << ") is at the end the function \n" ; } } while (false) |
| 1770 | << *Def->getMemoryInst()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << " Check if def " << *Def << " (" << *Def->getMemoryInst() << ") is at the end the function \n" ; } } while (false) |
| 1771 | << ") is at the end the function \n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << " Check if def " << *Def << " (" << *Def->getMemoryInst() << ") is at the end the function \n" ; } } while (false); |
| 1772 | |
| 1773 | auto MaybeLoc = getLocForWriteEx(Def->getMemoryInst()); |
| 1774 | if (!MaybeLoc) { |
| 1775 | LLVM_DEBUG(dbgs() << " ... could not get location for write.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << " ... could not get location for write.\n" ; } } while (false); |
| 1776 | return false; |
| 1777 | } |
| 1778 | |
| 1779 | SmallVector<MemoryAccess *, 4> WorkList; |
| 1780 | SmallPtrSet<MemoryAccess *, 8> Visited; |
| 1781 | auto PushMemUses = [&WorkList, &Visited](MemoryAccess *Acc) { |
| 1782 | if (!Visited.insert(Acc).second) |
| 1783 | return; |
| 1784 | for (Use &U : Acc->uses()) |
| 1785 | WorkList.push_back(cast<MemoryAccess>(U.getUser())); |
| 1786 | }; |
| 1787 | PushMemUses(Def); |
| 1788 | for (unsigned I = 0; I < WorkList.size(); I++) { |
| 1789 | if (WorkList.size() >= MemorySSAScanLimit) { |
| 1790 | LLVM_DEBUG(dbgs() << " ... hit exploration limit.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << " ... hit exploration limit.\n"; } } while (false); |
| 1791 | return false; |
| 1792 | } |
| 1793 | |
| 1794 | MemoryAccess *UseAccess = WorkList[I]; |
| 1795 | // Simply adding the users of MemoryPhi to the worklist is not enough, |
| 1796 | // because we might miss read clobbers in different iterations of a loop, |
| 1797 | // for example. |
| 1798 | // TODO: Add support for phi translation to handle the loop case. |
| 1799 | if (isa<MemoryPhi>(UseAccess)) |
| 1800 | return false; |
| 1801 | |
| 1802 | // TODO: Checking for aliasing is expensive. Consider reducing the amount |
| 1803 | // of times this is called and/or caching it. |
| 1804 | Instruction *UseInst = cast<MemoryUseOrDef>(UseAccess)->getMemoryInst(); |
| 1805 | if (isReadClobber(*MaybeLoc, UseInst)) { |
| 1806 | LLVM_DEBUG(dbgs() << " ... hit read clobber " << *UseInst << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << " ... hit read clobber " << *UseInst << ".\n"; } } while (false); |
| 1807 | return false; |
| 1808 | } |
| 1809 | |
| 1810 | if (MemoryDef *UseDef = dyn_cast<MemoryDef>(UseAccess)) |
| 1811 | PushMemUses(UseDef); |
| 1812 | } |
| 1813 | return true; |
| 1814 | } |
| 1815 | |
| 1816 | /// If \p I is a memory terminator like llvm.lifetime.end or free, return a |
| 1817 | /// pair with the MemoryLocation terminated by \p I and a boolean flag |
| 1818 | /// indicating whether \p I is a free-like call. |
| 1819 | Optional<std::pair<MemoryLocation, bool>> |
| 1820 | getLocForTerminator(Instruction *I) const { |
| 1821 | uint64_t Len; |
| 1822 | Value *Ptr; |
| 1823 | if (match(I, m_Intrinsic<Intrinsic::lifetime_end>(m_ConstantInt(Len), |
| 1824 | m_Value(Ptr)))) |
| 1825 | return {std::make_pair(MemoryLocation(Ptr, Len), false)}; |
| 1826 | |
| 1827 | if (auto *CB = dyn_cast<CallBase>(I)) { |
| 1828 | if (isFreeCall(I, &TLI)) |
| 1829 | return {std::make_pair(MemoryLocation::getAfter(CB->getArgOperand(0)), |
| 1830 | true)}; |
| 1831 | } |
| 1832 | |
| 1833 | return None; |
| 1834 | } |
| 1835 | |
| 1836 | /// Returns true if \p I is a memory terminator instruction like |
| 1837 | /// llvm.lifetime.end or free. |
| 1838 | bool isMemTerminatorInst(Instruction *I) const { |
| 1839 | IntrinsicInst *II = dyn_cast<IntrinsicInst>(I); |
| 1840 | return (II && II->getIntrinsicID() == Intrinsic::lifetime_end) || |
| 1841 | isFreeCall(I, &TLI); |
| 1842 | } |
| 1843 | |
| 1844 | /// Returns true if \p MaybeTerm is a memory terminator for \p Loc from |
| 1845 | /// instruction \p AccessI. |
| 1846 | bool isMemTerminator(const MemoryLocation &Loc, Instruction *AccessI, |
| 1847 | Instruction *MaybeTerm) { |
| 1848 | Optional<std::pair<MemoryLocation, bool>> MaybeTermLoc = |
| 1849 | getLocForTerminator(MaybeTerm); |
| 1850 | |
| 1851 | if (!MaybeTermLoc) |
| 1852 | return false; |
| 1853 | |
| 1854 | // If the terminator is a free-like call, all accesses to the underlying |
| 1855 | // object can be considered terminated. |
| 1856 | if (getUnderlyingObject(Loc.Ptr) != |
| 1857 | getUnderlyingObject(MaybeTermLoc->first.Ptr)) |
| 1858 | return false; |
| 1859 | |
| 1860 | auto TermLoc = MaybeTermLoc->first; |
| 1861 | if (MaybeTermLoc->second) { |
| 1862 | const Value *LocUO = getUnderlyingObject(Loc.Ptr); |
| 1863 | return BatchAA.isMustAlias(TermLoc.Ptr, LocUO); |
| 1864 | } |
| 1865 | int64_t InstWriteOffset, DepWriteOffset; |
| 1866 | return isOverwrite(MaybeTerm, AccessI, TermLoc, Loc, DL, TLI, |
| 1867 | DepWriteOffset, InstWriteOffset, BatchAA, |
| 1868 | &F) == OW_Complete; |
| 1869 | } |
| 1870 | |
| 1871 | // Returns true if \p Use may read from \p DefLoc. |
| 1872 | bool isReadClobber(const MemoryLocation &DefLoc, Instruction *UseInst) { |
| 1873 | if (isNoopIntrinsic(UseInst)) |
| 1874 | return false; |
| 1875 | |
| 1876 | // Monotonic or weaker atomic stores can be re-ordered and do not need to be |
| 1877 | // treated as read clobber. |
| 1878 | if (auto SI = dyn_cast<StoreInst>(UseInst)) |
| 1879 | return isStrongerThan(SI->getOrdering(), AtomicOrdering::Monotonic); |
| 1880 | |
| 1881 | if (!UseInst->mayReadFromMemory()) |
| 1882 | return false; |
| 1883 | |
| 1884 | if (auto *CB = dyn_cast<CallBase>(UseInst)) |
| 1885 | if (CB->onlyAccessesInaccessibleMemory()) |
| 1886 | return false; |
| 1887 | |
| 1888 | // NOTE: For calls, the number of stores removed could be slightly improved |
| 1889 | // by using AA.callCapturesBefore(UseInst, DefLoc, &DT), but that showed to |
| 1890 | // be expensive compared to the benefits in practice. For now, avoid more |
| 1891 | // expensive analysis to limit compile-time. |
| 1892 | return isRefSet(BatchAA.getModRefInfo(UseInst, DefLoc)); |
| 1893 | } |
| 1894 | |
| 1895 | /// Returns true if \p Ptr is guaranteed to be loop invariant for any possible |
| 1896 | /// loop. In particular, this guarantees that it only references a single |
| 1897 | /// MemoryLocation during execution of the containing function. |
| 1898 | bool IsGuaranteedLoopInvariant(Value *Ptr) { |
| 1899 | auto IsGuaranteedLoopInvariantBase = [this](Value *Ptr) { |
| 1900 | Ptr = Ptr->stripPointerCasts(); |
| 1901 | if (auto *I = dyn_cast<Instruction>(Ptr)) { |
| 1902 | if (isa<AllocaInst>(Ptr)) |
| 1903 | return true; |
| 1904 | |
| 1905 | if (isAllocLikeFn(I, &TLI)) |
| 1906 | return true; |
| 1907 | |
| 1908 | return false; |
| 1909 | } |
| 1910 | return true; |
| 1911 | }; |
| 1912 | |
| 1913 | Ptr = Ptr->stripPointerCasts(); |
| 1914 | if (auto *I = dyn_cast<Instruction>(Ptr)) { |
| 1915 | if (I->getParent() == &I->getFunction()->getEntryBlock()) { |
| 1916 | return true; |
| 1917 | } |
| 1918 | } |
| 1919 | if (auto *GEP = dyn_cast<GEPOperator>(Ptr)) { |
| 1920 | return IsGuaranteedLoopInvariantBase(GEP->getPointerOperand()) && |
| 1921 | GEP->hasAllConstantIndices(); |
| 1922 | } |
| 1923 | return IsGuaranteedLoopInvariantBase(Ptr); |
| 1924 | } |
| 1925 | |
| 1926 | // Find a MemoryDef writing to \p DefLoc and dominating \p StartAccess, with |
| 1927 | // no read access between them or on any other path to a function exit block |
| 1928 | // if \p DefLoc is not accessible after the function returns. If there is no |
| 1929 | // such MemoryDef, return None. The returned value may not (completely) |
| 1930 | // overwrite \p DefLoc. Currently we bail out when we encounter an aliasing |
| 1931 | // MemoryUse (read). |
| 1932 | Optional<MemoryAccess *> |
| 1933 | getDomMemoryDef(MemoryDef *KillingDef, MemoryAccess *StartAccess, |
| 1934 | const MemoryLocation &DefLoc, const Value *DefUO, |
| 1935 | unsigned &ScanLimit, unsigned &WalkerStepLimit, |
| 1936 | bool IsMemTerm, unsigned &PartialLimit) { |
| 1937 | if (ScanLimit == 0 || WalkerStepLimit == 0) { |
| 1938 | LLVM_DEBUG(dbgs() << "\n ... hit scan limit\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "\n ... hit scan limit\n"; } } while (false); |
| 1939 | return None; |
| 1940 | } |
| 1941 | |
| 1942 | MemoryAccess *Current = StartAccess; |
| 1943 | Instruction *KillingI = KillingDef->getMemoryInst(); |
| 1944 | bool StepAgain; |
| 1945 | LLVM_DEBUG(dbgs() << " trying to get dominating access\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << " trying to get dominating access\n" ; } } while (false); |
| 1946 | |
| 1947 | // Find the next clobbering Mod access for DefLoc, starting at StartAccess. |
| 1948 | Optional<MemoryLocation> CurrentLoc; |
| 1949 | do { |
| 1950 | StepAgain = false; |
| 1951 | LLVM_DEBUG({do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { { dbgs() << " visiting " << *Current ; if (!MSSA.isLiveOnEntryDef(Current) && isa<MemoryUseOrDef >(Current)) dbgs() << " (" << *cast<MemoryUseOrDef >(Current)->getMemoryInst() << ")"; dbgs() << "\n"; }; } } while (false) |
| 1952 | dbgs() << " visiting " << *Current;do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { { dbgs() << " visiting " << *Current ; if (!MSSA.isLiveOnEntryDef(Current) && isa<MemoryUseOrDef >(Current)) dbgs() << " (" << *cast<MemoryUseOrDef >(Current)->getMemoryInst() << ")"; dbgs() << "\n"; }; } } while (false) |
| 1953 | if (!MSSA.isLiveOnEntryDef(Current) && isa<MemoryUseOrDef>(Current))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { { dbgs() << " visiting " << *Current ; if (!MSSA.isLiveOnEntryDef(Current) && isa<MemoryUseOrDef >(Current)) dbgs() << " (" << *cast<MemoryUseOrDef >(Current)->getMemoryInst() << ")"; dbgs() << "\n"; }; } } while (false) |
| 1954 | dbgs() << " (" << *cast<MemoryUseOrDef>(Current)->getMemoryInst()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { { dbgs() << " visiting " << *Current ; if (!MSSA.isLiveOnEntryDef(Current) && isa<MemoryUseOrDef >(Current)) dbgs() << " (" << *cast<MemoryUseOrDef >(Current)->getMemoryInst() << ")"; dbgs() << "\n"; }; } } while (false) |
| 1955 | << ")";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { { dbgs() << " visiting " << *Current ; if (!MSSA.isLiveOnEntryDef(Current) && isa<MemoryUseOrDef >(Current)) dbgs() << " (" << *cast<MemoryUseOrDef >(Current)->getMemoryInst() << ")"; dbgs() << "\n"; }; } } while (false) |
| 1956 | dbgs() << "\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { { dbgs() << " visiting " << *Current ; if (!MSSA.isLiveOnEntryDef(Current) && isa<MemoryUseOrDef >(Current)) dbgs() << " (" << *cast<MemoryUseOrDef >(Current)->getMemoryInst() << ")"; dbgs() << "\n"; }; } } while (false) |
| 1957 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { { dbgs() << " visiting " << *Current ; if (!MSSA.isLiveOnEntryDef(Current) && isa<MemoryUseOrDef >(Current)) dbgs() << " (" << *cast<MemoryUseOrDef >(Current)->getMemoryInst() << ")"; dbgs() << "\n"; }; } } while (false); |
| 1958 | |
| 1959 | // Reached TOP. |
| 1960 | if (MSSA.isLiveOnEntryDef(Current)) { |
| 1961 | LLVM_DEBUG(dbgs() << " ... found LiveOnEntryDef\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << " ... found LiveOnEntryDef\n"; } } while (false); |
| 1962 | return None; |
| 1963 | } |
| 1964 | |
| 1965 | // Cost of a step. Accesses in the same block are more likely to be valid |
| 1966 | // candidates for elimination, hence consider them cheaper. |
| 1967 | unsigned StepCost = KillingDef->getBlock() == Current->getBlock() |
| 1968 | ? MemorySSASameBBStepCost |
| 1969 | : MemorySSAOtherBBStepCost; |
| 1970 | if (WalkerStepLimit <= StepCost) { |
| 1971 | LLVM_DEBUG(dbgs() << " ... hit walker step limit\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << " ... hit walker step limit\n"; } } while (false); |
| 1972 | return None; |
| 1973 | } |
| 1974 | WalkerStepLimit -= StepCost; |
| 1975 | |
| 1976 | // Return for MemoryPhis. They cannot be eliminated directly and the |
| 1977 | // caller is responsible for traversing them. |
| 1978 | if (isa<MemoryPhi>(Current)) { |
| 1979 | LLVM_DEBUG(dbgs() << " ... found MemoryPhi\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << " ... found MemoryPhi\n"; } } while (false); |
| 1980 | return Current; |
| 1981 | } |
| 1982 | |
| 1983 | // Below, check if CurrentDef is a valid candidate to be eliminated by |
| 1984 | // KillingDef. If it is not, check the next candidate. |
| 1985 | MemoryDef *CurrentDef = cast<MemoryDef>(Current); |
| 1986 | Instruction *CurrentI = CurrentDef->getMemoryInst(); |
| 1987 | |
| 1988 | if (canSkipDef(CurrentDef, !isInvisibleToCallerBeforeRet(DefUO))) { |
| 1989 | StepAgain = true; |
| 1990 | Current = CurrentDef->getDefiningAccess(); |
| 1991 | continue; |
| 1992 | } |
| 1993 | |
| 1994 | // Before we try to remove anything, check for any extra throwing |
| 1995 | // instructions that block us from DSEing |
| 1996 | if (mayThrowBetween(KillingI, CurrentI, DefUO)) { |
| 1997 | LLVM_DEBUG(dbgs() << " ... skip, may throw!\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << " ... skip, may throw!\n"; } } while (false); |
| 1998 | return None; |
| 1999 | } |
| 2000 | |
| 2001 | // Check for anything that looks like it will be a barrier to further |
| 2002 | // removal |
| 2003 | if (isDSEBarrier(DefUO, CurrentI)) { |
| 2004 | LLVM_DEBUG(dbgs() << " ... skip, barrier\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << " ... skip, barrier\n"; } } while (false); |
| 2005 | return None; |
| 2006 | } |
| 2007 | |
| 2008 | // If Current is known to be on path that reads DefLoc or is a read |
| 2009 | // clobber, bail out, as the path is not profitable. We skip this check |
| 2010 | // for intrinsic calls, because the code knows how to handle memcpy |
| 2011 | // intrinsics. |
| 2012 | if (!isa<IntrinsicInst>(CurrentI) && isReadClobber(DefLoc, CurrentI)) |
| 2013 | return None; |
| 2014 | |
| 2015 | // Quick check if there are direct uses that are read-clobbers. |
| 2016 | if (any_of(Current->uses(), [this, &DefLoc, StartAccess](Use &U) { |
| 2017 | if (auto *UseOrDef = dyn_cast<MemoryUseOrDef>(U.getUser())) |
| 2018 | return !MSSA.dominates(StartAccess, UseOrDef) && |
| 2019 | isReadClobber(DefLoc, UseOrDef->getMemoryInst()); |
| 2020 | return false; |
| 2021 | })) { |
| 2022 | LLVM_DEBUG(dbgs() << " ... found a read clobber\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << " ... found a read clobber\n"; } } while (false); |
| 2023 | return None; |
| 2024 | } |
| 2025 | |
| 2026 | // If Current cannot be analyzed or is not removable, check the next |
| 2027 | // candidate. |
| 2028 | if (!hasAnalyzableMemoryWrite(CurrentI, TLI) || !isRemovable(CurrentI)) { |
| 2029 | StepAgain = true; |
| 2030 | Current = CurrentDef->getDefiningAccess(); |
| 2031 | continue; |
| 2032 | } |
| 2033 | |
| 2034 | // If Current does not have an analyzable write location, skip it |
| 2035 | CurrentLoc = getLocForWriteEx(CurrentI); |
| 2036 | if (!CurrentLoc) { |
| 2037 | StepAgain = true; |
| 2038 | Current = CurrentDef->getDefiningAccess(); |
| 2039 | continue; |
| 2040 | } |
| 2041 | |
| 2042 | // AliasAnalysis does not account for loops. Limit elimination to |
| 2043 | // candidates for which we can guarantee they always store to the same |
| 2044 | // memory location and not multiple locations in a loop. |
| 2045 | if (Current->getBlock() != KillingDef->getBlock() && |
| 2046 | !IsGuaranteedLoopInvariant(const_cast<Value *>(CurrentLoc->Ptr))) { |
| 2047 | StepAgain = true; |
| 2048 | Current = CurrentDef->getDefiningAccess(); |
| 2049 | WalkerStepLimit -= 1; |
| 2050 | continue; |
| 2051 | } |
| 2052 | |
| 2053 | if (IsMemTerm) { |
| 2054 | // If the killing def is a memory terminator (e.g. lifetime.end), check |
| 2055 | // the next candidate if the current Current does not write the same |
| 2056 | // underlying object as the terminator. |
| 2057 | if (!isMemTerminator(*CurrentLoc, CurrentI, KillingI)) { |
| 2058 | StepAgain = true; |
| 2059 | Current = CurrentDef->getDefiningAccess(); |
| 2060 | } |
| 2061 | continue; |
| 2062 | } else { |
| 2063 | int64_t InstWriteOffset, DepWriteOffset; |
| 2064 | auto OR = isOverwrite(KillingI, CurrentI, DefLoc, *CurrentLoc, DL, TLI, |
| 2065 | DepWriteOffset, InstWriteOffset, BatchAA, &F); |
| 2066 | // If Current does not write to the same object as KillingDef, check |
| 2067 | // the next candidate. |
| 2068 | if (OR == OW_Unknown) { |
| 2069 | StepAgain = true; |
| 2070 | Current = CurrentDef->getDefiningAccess(); |
| 2071 | } else if (OR == OW_MaybePartial) { |
| 2072 | // If KillingDef only partially overwrites Current, check the next |
| 2073 | // candidate if the partial step limit is exceeded. This aggressively |
| 2074 | // limits the number of candidates for partial store elimination, |
| 2075 | // which are less likely to be removable in the end. |
| 2076 | if (PartialLimit <= 1) { |
| 2077 | StepAgain = true; |
| 2078 | Current = CurrentDef->getDefiningAccess(); |
| 2079 | WalkerStepLimit -= 1; |
| 2080 | continue; |
| 2081 | } |
| 2082 | PartialLimit -= 1; |
| 2083 | } |
| 2084 | } |
| 2085 | } while (StepAgain); |
| 2086 | |
| 2087 | // Accesses to objects accessible after the function returns can only be |
| 2088 | // eliminated if the access is killed along all paths to the exit. Collect |
| 2089 | // the blocks with killing (=completely overwriting MemoryDefs) and check if |
| 2090 | // they cover all paths from EarlierAccess to any function exit. |
| 2091 | SmallPtrSet<Instruction *, 16> KillingDefs; |
| 2092 | KillingDefs.insert(KillingDef->getMemoryInst()); |
| 2093 | MemoryAccess *EarlierAccess = Current; |
| 2094 | Instruction *EarlierMemInst = |
| 2095 | cast<MemoryDef>(EarlierAccess)->getMemoryInst(); |
| 2096 | LLVM_DEBUG(dbgs() << " Checking for reads of " << *EarlierAccess << " ("do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << " Checking for reads of " << *EarlierAccess << " (" << *EarlierMemInst << ")\n"; } } while (false) |
| 2097 | << *EarlierMemInst << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << " Checking for reads of " << *EarlierAccess << " (" << *EarlierMemInst << ")\n"; } } while (false); |
| 2098 | |
| 2099 | SmallSetVector<MemoryAccess *, 32> WorkList; |
| 2100 | auto PushMemUses = [&WorkList](MemoryAccess *Acc) { |
| 2101 | for (Use &U : Acc->uses()) |
| 2102 | WorkList.insert(cast<MemoryAccess>(U.getUser())); |
| 2103 | }; |
| 2104 | PushMemUses(EarlierAccess); |
| 2105 | |
| 2106 | // Optimistically collect all accesses for reads. If we do not find any |
| 2107 | // read clobbers, add them to the cache. |
| 2108 | SmallPtrSet<MemoryAccess *, 16> KnownNoReads; |
| 2109 | if (!EarlierMemInst->mayReadFromMemory()) |
| 2110 | KnownNoReads.insert(EarlierAccess); |
| 2111 | // Check if EarlierDef may be read. |
| 2112 | for (unsigned I = 0; I < WorkList.size(); I++) { |
| 2113 | MemoryAccess *UseAccess = WorkList[I]; |
| 2114 | |
| 2115 | LLVM_DEBUG(dbgs() << " " << *UseAccess)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << " " << *UseAccess; } } while (false); |
| 2116 | // Bail out if the number of accesses to check exceeds the scan limit. |
| 2117 | if (ScanLimit < (WorkList.size() - I)) { |
| 2118 | LLVM_DEBUG(dbgs() << "\n ... hit scan limit\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "\n ... hit scan limit\n"; } } while (false); |
| 2119 | return None; |
| 2120 | } |
| 2121 | --ScanLimit; |
| 2122 | NumDomMemDefChecks++; |
| 2123 | KnownNoReads.insert(UseAccess); |
| 2124 | |
| 2125 | if (isa<MemoryPhi>(UseAccess)) { |
| 2126 | if (any_of(KillingDefs, [this, UseAccess](Instruction *KI) { |
| 2127 | return DT.properlyDominates(KI->getParent(), |
| 2128 | UseAccess->getBlock()); |
| 2129 | })) { |
| 2130 | LLVM_DEBUG(dbgs() << " ... skipping, dominated by killing block\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << " ... skipping, dominated by killing block\n" ; } } while (false); |
| 2131 | continue; |
| 2132 | } |
| 2133 | LLVM_DEBUG(dbgs() << "\n ... adding PHI uses\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "\n ... adding PHI uses\n"; } } while (false); |
| 2134 | PushMemUses(UseAccess); |
| 2135 | continue; |
| 2136 | } |
| 2137 | |
| 2138 | Instruction *UseInst = cast<MemoryUseOrDef>(UseAccess)->getMemoryInst(); |
| 2139 | LLVM_DEBUG(dbgs() << " (" << *UseInst << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << " (" << *UseInst << ")\n" ; } } while (false); |
| 2140 | |
| 2141 | if (any_of(KillingDefs, [this, UseInst](Instruction *KI) { |
| 2142 | return DT.dominates(KI, UseInst); |
| 2143 | })) { |
| 2144 | LLVM_DEBUG(dbgs() << " ... skipping, dominated by killing def\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << " ... skipping, dominated by killing def\n" ; } } while (false); |
| 2145 | continue; |
| 2146 | } |
| 2147 | |
| 2148 | // A memory terminator kills all preceeding MemoryDefs and all succeeding |
| 2149 | // MemoryAccesses. We do not have to check it's users. |
| 2150 | if (isMemTerminator(*CurrentLoc, EarlierMemInst, UseInst)) { |
| 2151 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << " ... skipping, memterminator invalidates following accesses\n" ; } } while (false) |
| 2152 | dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << " ... skipping, memterminator invalidates following accesses\n" ; } } while (false) |
| 2153 | << " ... skipping, memterminator invalidates following accesses\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << " ... skipping, memterminator invalidates following accesses\n" ; } } while (false); |
| 2154 | continue; |
| 2155 | } |
| 2156 | |
| 2157 | if (isNoopIntrinsic(cast<MemoryUseOrDef>(UseAccess)->getMemoryInst())) { |
| 2158 | LLVM_DEBUG(dbgs() << " ... adding uses of intrinsic\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << " ... adding uses of intrinsic\n" ; } } while (false); |
| 2159 | PushMemUses(UseAccess); |
| 2160 | continue; |
| 2161 | } |
| 2162 | |
| 2163 | if (UseInst->mayThrow() && !isInvisibleToCallerBeforeRet(DefUO)) { |
| 2164 | LLVM_DEBUG(dbgs() << " ... found throwing instruction\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << " ... found throwing instruction\n" ; } } while (false); |
| 2165 | return None; |
| 2166 | } |
| 2167 | |
| 2168 | // Uses which may read the original MemoryDef mean we cannot eliminate the |
| 2169 | // original MD. Stop walk. |
| 2170 | if (isReadClobber(*CurrentLoc, UseInst)) { |
| 2171 | LLVM_DEBUG(dbgs() << " ... found read clobber\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << " ... found read clobber\n"; } } while (false); |
| 2172 | return None; |
| 2173 | } |
| 2174 | |
| 2175 | // For the KillingDef and EarlierAccess we only have to check if it reads |
| 2176 | // the memory location. |
| 2177 | // TODO: It would probably be better to check for self-reads before |
| 2178 | // calling the function. |
| 2179 | if (KillingDef == UseAccess || EarlierAccess == UseAccess) { |
| 2180 | LLVM_DEBUG(dbgs() << " ... skipping killing def/dom access\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << " ... skipping killing def/dom access\n" ; } } while (false); |
| 2181 | continue; |
| 2182 | } |
| 2183 | |
| 2184 | // Check all uses for MemoryDefs, except for defs completely overwriting |
| 2185 | // the original location. Otherwise we have to check uses of *all* |
| 2186 | // MemoryDefs we discover, including non-aliasing ones. Otherwise we might |
| 2187 | // miss cases like the following |
| 2188 | // 1 = Def(LoE) ; <----- EarlierDef stores [0,1] |
| 2189 | // 2 = Def(1) ; (2, 1) = NoAlias, stores [2,3] |
| 2190 | // Use(2) ; MayAlias 2 *and* 1, loads [0, 3]. |
| 2191 | // (The Use points to the *first* Def it may alias) |
| 2192 | // 3 = Def(1) ; <---- Current (3, 2) = NoAlias, (3,1) = MayAlias, |
| 2193 | // stores [0,1] |
| 2194 | if (MemoryDef *UseDef = dyn_cast<MemoryDef>(UseAccess)) { |
| 2195 | if (isCompleteOverwrite(*CurrentLoc, EarlierMemInst, UseInst)) { |
| 2196 | if (!isInvisibleToCallerAfterRet(DefUO) && |
| 2197 | UseAccess != EarlierAccess) { |
| 2198 | BasicBlock *MaybeKillingBlock = UseInst->getParent(); |
| 2199 | if (PostOrderNumbers.find(MaybeKillingBlock)->second < |
| 2200 | PostOrderNumbers.find(EarlierAccess->getBlock())->second) { |
| 2201 | |
| 2202 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << " ... found killing def " << *UseInst << "\n"; } } while (false) |
| 2203 | << " ... found killing def " << *UseInst << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << " ... found killing def " << *UseInst << "\n"; } } while (false); |
| 2204 | KillingDefs.insert(UseInst); |
| 2205 | } |
| 2206 | } |
| 2207 | } else |
| 2208 | PushMemUses(UseDef); |
| 2209 | } |
| 2210 | } |
| 2211 | |
| 2212 | // For accesses to locations visible after the function returns, make sure |
| 2213 | // that the location is killed (=overwritten) along all paths from |
| 2214 | // EarlierAccess to the exit. |
| 2215 | if (!isInvisibleToCallerAfterRet(DefUO)) { |
| 2216 | SmallPtrSet<BasicBlock *, 16> KillingBlocks; |
| 2217 | for (Instruction *KD : KillingDefs) |
| 2218 | KillingBlocks.insert(KD->getParent()); |
| 2219 | assert(!KillingBlocks.empty() &&((!KillingBlocks.empty() && "Expected at least a single killing block" ) ? static_cast<void> (0) : __assert_fail ("!KillingBlocks.empty() && \"Expected at least a single killing block\"" , "/build/llvm-toolchain-snapshot-13~++20210301100612+564f5b0734bd/llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp" , 2220, __PRETTY_FUNCTION__)) |
| 2220 | "Expected at least a single killing block")((!KillingBlocks.empty() && "Expected at least a single killing block" ) ? static_cast<void> (0) : __assert_fail ("!KillingBlocks.empty() && \"Expected at least a single killing block\"" , "/build/llvm-toolchain-snapshot-13~++20210301100612+564f5b0734bd/llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp" , 2220, __PRETTY_FUNCTION__)); |
| 2221 | |
| 2222 | // Find the common post-dominator of all killing blocks. |
| 2223 | BasicBlock *CommonPred = *KillingBlocks.begin(); |
| 2224 | for (auto I = std::next(KillingBlocks.begin()), E = KillingBlocks.end(); |
| 2225 | I != E; I++) { |
| 2226 | if (!CommonPred) |
| 2227 | break; |
| 2228 | CommonPred = PDT.findNearestCommonDominator(CommonPred, *I); |
| 2229 | } |
| 2230 | |
| 2231 | // If CommonPred is in the set of killing blocks, just check if it |
| 2232 | // post-dominates EarlierAccess. |
| 2233 | if (KillingBlocks.count(CommonPred)) { |
| 2234 | if (PDT.dominates(CommonPred, EarlierAccess->getBlock())) |
| 2235 | return {EarlierAccess}; |
| 2236 | return None; |
| 2237 | } |
| 2238 | |
| 2239 | // If the common post-dominator does not post-dominate EarlierAccess, |
| 2240 | // there is a path from EarlierAccess to an exit not going through a |
| 2241 | // killing block. |
| 2242 | if (PDT.dominates(CommonPred, EarlierAccess->getBlock())) { |
| 2243 | SetVector<BasicBlock *> WorkList; |
| 2244 | |
| 2245 | // If CommonPred is null, there are multiple exits from the function. |
| 2246 | // They all have to be added to the worklist. |
| 2247 | if (CommonPred) |
| 2248 | WorkList.insert(CommonPred); |
| 2249 | else |
| 2250 | for (BasicBlock *R : PDT.roots()) |
| 2251 | WorkList.insert(R); |
| 2252 | |
| 2253 | NumCFGTries++; |
| 2254 | // Check if all paths starting from an exit node go through one of the |
| 2255 | // killing blocks before reaching EarlierAccess. |
| 2256 | for (unsigned I = 0; I < WorkList.size(); I++) { |
| 2257 | NumCFGChecks++; |
| 2258 | BasicBlock *Current = WorkList[I]; |
| 2259 | if (KillingBlocks.count(Current)) |
| 2260 | continue; |
| 2261 | if (Current == EarlierAccess->getBlock()) |
| 2262 | return None; |
| 2263 | |
| 2264 | // EarlierAccess is reachable from the entry, so we don't have to |
| 2265 | // explore unreachable blocks further. |
| 2266 | if (!DT.isReachableFromEntry(Current)) |
| 2267 | continue; |
| 2268 | |
| 2269 | for (BasicBlock *Pred : predecessors(Current)) |
| 2270 | WorkList.insert(Pred); |
| 2271 | |
| 2272 | if (WorkList.size() >= MemorySSAPathCheckLimit) |
| 2273 | return None; |
| 2274 | } |
| 2275 | NumCFGSuccess++; |
| 2276 | return {EarlierAccess}; |
| 2277 | } |
| 2278 | return None; |
| 2279 | } |
| 2280 | |
| 2281 | // No aliasing MemoryUses of EarlierAccess found, EarlierAccess is |
| 2282 | // potentially dead. |
| 2283 | return {EarlierAccess}; |
| 2284 | } |
| 2285 | |
| 2286 | // Delete dead memory defs |
| 2287 | void deleteDeadInstruction(Instruction *SI) { |
| 2288 | MemorySSAUpdater Updater(&MSSA); |
| 2289 | SmallVector<Instruction *, 32> NowDeadInsts; |
| 2290 | NowDeadInsts.push_back(SI); |
| 2291 | --NumFastOther; |
| 2292 | |
| 2293 | while (!NowDeadInsts.empty()) { |
| 2294 | Instruction *DeadInst = NowDeadInsts.pop_back_val(); |
| 2295 | ++NumFastOther; |
| 2296 | |
| 2297 | // Try to preserve debug information attached to the dead instruction. |
| 2298 | salvageDebugInfo(*DeadInst); |
| 2299 | salvageKnowledge(DeadInst); |
| 2300 | |
| 2301 | // Remove the Instruction from MSSA. |
| 2302 | if (MemoryAccess *MA = MSSA.getMemoryAccess(DeadInst)) { |
| 2303 | if (MemoryDef *MD = dyn_cast<MemoryDef>(MA)) { |
| 2304 | SkipStores.insert(MD); |
| 2305 | } |
| 2306 | Updater.removeMemoryAccess(MA); |
| 2307 | } |
| 2308 | |
| 2309 | auto I = IOLs.find(DeadInst->getParent()); |
| 2310 | if (I != IOLs.end()) |
| 2311 | I->second.erase(DeadInst); |
| 2312 | // Remove its operands |
| 2313 | for (Use &O : DeadInst->operands()) |
| 2314 | if (Instruction *OpI = dyn_cast<Instruction>(O)) { |
| 2315 | O = nullptr; |
| 2316 | if (isInstructionTriviallyDead(OpI, &TLI)) |
| 2317 | NowDeadInsts.push_back(OpI); |
| 2318 | } |
| 2319 | |
| 2320 | DeadInst->eraseFromParent(); |
| 2321 | } |
| 2322 | } |
| 2323 | |
| 2324 | // Check for any extra throws between SI and NI that block DSE. This only |
| 2325 | // checks extra maythrows (those that aren't MemoryDef's). MemoryDef that may |
| 2326 | // throw are handled during the walk from one def to the next. |
| 2327 | bool mayThrowBetween(Instruction *SI, Instruction *NI, |
| 2328 | const Value *SILocUnd) { |
| 2329 | // First see if we can ignore it by using the fact that SI is an |
| 2330 | // alloca/alloca like object that is not visible to the caller during |
| 2331 | // execution of the function. |
| 2332 | if (SILocUnd && isInvisibleToCallerBeforeRet(SILocUnd)) |
| 2333 | return false; |
| 2334 | |
| 2335 | if (SI->getParent() == NI->getParent()) |
| 2336 | return ThrowingBlocks.count(SI->getParent()); |
| 2337 | return !ThrowingBlocks.empty(); |
| 2338 | } |
| 2339 | |
| 2340 | // Check if \p NI acts as a DSE barrier for \p SI. The following instructions |
| 2341 | // act as barriers: |
| 2342 | // * A memory instruction that may throw and \p SI accesses a non-stack |
| 2343 | // object. |
| 2344 | // * Atomic stores stronger that monotonic. |
| 2345 | bool isDSEBarrier(const Value *SILocUnd, Instruction *NI) { |
| 2346 | // If NI may throw it acts as a barrier, unless we are to an alloca/alloca |
| 2347 | // like object that does not escape. |
| 2348 | if (NI->mayThrow() && !isInvisibleToCallerBeforeRet(SILocUnd)) |
| 2349 | return true; |
| 2350 | |
| 2351 | // If NI is an atomic load/store stronger than monotonic, do not try to |
| 2352 | // eliminate/reorder it. |
| 2353 | if (NI->isAtomic()) { |
| 2354 | if (auto *LI = dyn_cast<LoadInst>(NI)) |
| 2355 | return isStrongerThanMonotonic(LI->getOrdering()); |
| 2356 | if (auto *SI = dyn_cast<StoreInst>(NI)) |
| 2357 | return isStrongerThanMonotonic(SI->getOrdering()); |
| 2358 | if (auto *ARMW = dyn_cast<AtomicRMWInst>(NI)) |
| 2359 | return isStrongerThanMonotonic(ARMW->getOrdering()); |
| 2360 | if (auto *CmpXchg = dyn_cast<AtomicCmpXchgInst>(NI)) |
| 2361 | return isStrongerThanMonotonic(CmpXchg->getSuccessOrdering()) || |
| 2362 | isStrongerThanMonotonic(CmpXchg->getFailureOrdering()); |
| 2363 | llvm_unreachable("other instructions should be skipped in MemorySSA")::llvm::llvm_unreachable_internal("other instructions should be skipped in MemorySSA" , "/build/llvm-toolchain-snapshot-13~++20210301100612+564f5b0734bd/llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp" , 2363); |
| 2364 | } |
| 2365 | return false; |
| 2366 | } |
| 2367 | |
| 2368 | /// Eliminate writes to objects that are not visible in the caller and are not |
| 2369 | /// accessed before returning from the function. |
| 2370 | bool eliminateDeadWritesAtEndOfFunction() { |
| 2371 | bool MadeChange = false; |
| 2372 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "Trying to eliminate MemoryDefs at the end of the function\n" ; } } while (false) |
| 2373 | dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "Trying to eliminate MemoryDefs at the end of the function\n" ; } } while (false) |
| 2374 | << "Trying to eliminate MemoryDefs at the end of the function\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "Trying to eliminate MemoryDefs at the end of the function\n" ; } } while (false); |
| 2375 | for (int I = MemDefs.size() - 1; I >= 0; I--) { |
| 2376 | MemoryDef *Def = MemDefs[I]; |
| 2377 | if (SkipStores.contains(Def) || !isRemovable(Def->getMemoryInst())) |
| 2378 | continue; |
| 2379 | |
| 2380 | Instruction *DefI = Def->getMemoryInst(); |
| 2381 | SmallVector<const Value *, 4> Pointers; |
| 2382 | auto DefLoc = getLocForWriteEx(DefI); |
| 2383 | if (!DefLoc) |
| 2384 | continue; |
| 2385 | |
| 2386 | // NOTE: Currently eliminating writes at the end of a function is limited |
| 2387 | // to MemoryDefs with a single underlying object, to save compile-time. In |
| 2388 | // practice it appears the case with multiple underlying objects is very |
| 2389 | // uncommon. If it turns out to be important, we can use |
| 2390 | // getUnderlyingObjects here instead. |
| 2391 | const Value *UO = getUnderlyingObject(DefLoc->Ptr); |
| 2392 | if (!UO || !isInvisibleToCallerAfterRet(UO)) |
| 2393 | continue; |
| 2394 | |
| 2395 | if (isWriteAtEndOfFunction(Def)) { |
| 2396 | // See through pointer-to-pointer bitcasts |
| 2397 | LLVM_DEBUG(dbgs() << " ... MemoryDef is not accessed until the end "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << " ... MemoryDef is not accessed until the end " "of the function\n"; } } while (false) |
| 2398 | "of the function\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << " ... MemoryDef is not accessed until the end " "of the function\n"; } } while (false); |
| 2399 | deleteDeadInstruction(DefI); |
| 2400 | ++NumFastStores; |
| 2401 | MadeChange = true; |
| 2402 | } |
| 2403 | } |
| 2404 | return MadeChange; |
| 2405 | } |
| 2406 | |
| 2407 | /// \returns true if \p Def is a no-op store, either because it |
| 2408 | /// directly stores back a loaded value or stores zero to a calloced object. |
| 2409 | bool storeIsNoop(MemoryDef *Def, const MemoryLocation &DefLoc, |
| 2410 | const Value *DefUO) { |
| 2411 | StoreInst *Store = dyn_cast<StoreInst>(Def->getMemoryInst()); |
| 2412 | if (!Store) |
| 2413 | return false; |
| 2414 | |
| 2415 | if (auto *LoadI = dyn_cast<LoadInst>(Store->getOperand(0))) { |
| 2416 | if (LoadI->getPointerOperand() == Store->getOperand(1)) { |
| 2417 | // Get the defining access for the load. |
| 2418 | auto *LoadAccess = MSSA.getMemoryAccess(LoadI)->getDefiningAccess(); |
| 2419 | // Fast path: the defining accesses are the same. |
| 2420 | if (LoadAccess == Def->getDefiningAccess()) |
| 2421 | return true; |
| 2422 | |
| 2423 | // Look through phi accesses. Recursively scan all phi accesses by |
| 2424 | // adding them to a worklist. Bail when we run into a memory def that |
| 2425 | // does not match LoadAccess. |
| 2426 | SetVector<MemoryAccess *> ToCheck; |
| 2427 | MemoryAccess *Current = |
| 2428 | MSSA.getWalker()->getClobberingMemoryAccess(Def); |
| 2429 | // We don't want to bail when we run into the store memory def. But, |
| 2430 | // the phi access may point to it. So, pretend like we've already |
| 2431 | // checked it. |
| 2432 | ToCheck.insert(Def); |
| 2433 | ToCheck.insert(Current); |
| 2434 | // Start at current (1) to simulate already having checked Def. |
| 2435 | for (unsigned I = 1; I < ToCheck.size(); ++I) { |
| 2436 | Current = ToCheck[I]; |
| 2437 | if (auto PhiAccess = dyn_cast<MemoryPhi>(Current)) { |
| 2438 | // Check all the operands. |
| 2439 | for (auto &Use : PhiAccess->incoming_values()) |
| 2440 | ToCheck.insert(cast<MemoryAccess>(&Use)); |
| 2441 | continue; |
| 2442 | } |
| 2443 | |
| 2444 | // If we found a memory def, bail. This happens when we have an |
| 2445 | // unrelated write in between an otherwise noop store. |
| 2446 | assert(isa<MemoryDef>(Current) &&((isa<MemoryDef>(Current) && "Only MemoryDefs should reach here." ) ? static_cast<void> (0) : __assert_fail ("isa<MemoryDef>(Current) && \"Only MemoryDefs should reach here.\"" , "/build/llvm-toolchain-snapshot-13~++20210301100612+564f5b0734bd/llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp" , 2447, __PRETTY_FUNCTION__)) |
| 2447 | "Only MemoryDefs should reach here.")((isa<MemoryDef>(Current) && "Only MemoryDefs should reach here." ) ? static_cast<void> (0) : __assert_fail ("isa<MemoryDef>(Current) && \"Only MemoryDefs should reach here.\"" , "/build/llvm-toolchain-snapshot-13~++20210301100612+564f5b0734bd/llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp" , 2447, __PRETTY_FUNCTION__)); |
| 2448 | // TODO: Skip no alias MemoryDefs that have no aliasing reads. |
| 2449 | // We are searching for the definition of the store's destination. |
| 2450 | // So, if that is the same definition as the load, then this is a |
| 2451 | // noop. Otherwise, fail. |
| 2452 | if (LoadAccess != Current) |
| 2453 | return false; |
| 2454 | } |
| 2455 | return true; |
| 2456 | } |
| 2457 | } |
| 2458 | |
| 2459 | Constant *StoredConstant = dyn_cast<Constant>(Store->getOperand(0)); |
| 2460 | if (StoredConstant && StoredConstant->isNullValue()) { |
| 2461 | auto *DefUOInst = dyn_cast<Instruction>(DefUO); |
| 2462 | if (DefUOInst && isCallocLikeFn(DefUOInst, &TLI)) { |
| 2463 | auto *UnderlyingDef = cast<MemoryDef>(MSSA.getMemoryAccess(DefUOInst)); |
| 2464 | // If UnderlyingDef is the clobbering access of Def, no instructions |
| 2465 | // between them can modify the memory location. |
| 2466 | auto *ClobberDef = |
| 2467 | MSSA.getSkipSelfWalker()->getClobberingMemoryAccess(Def); |
| 2468 | return UnderlyingDef == ClobberDef; |
| 2469 | } |
| 2470 | } |
| 2471 | return false; |
| 2472 | } |
| 2473 | }; |
| 2474 | |
| 2475 | bool eliminateDeadStoresMemorySSA(Function &F, AliasAnalysis &AA, |
| 2476 | MemorySSA &MSSA, DominatorTree &DT, |
| 2477 | PostDominatorTree &PDT, |
| 2478 | const TargetLibraryInfo &TLI) { |
| 2479 | bool MadeChange = false; |
| 2480 | |
| 2481 | DSEState State = DSEState::get(F, AA, MSSA, DT, PDT, TLI); |
| 2482 | // For each store: |
| 2483 | for (unsigned I = 0; I < State.MemDefs.size(); I++) { |
| 2484 | MemoryDef *KillingDef = State.MemDefs[I]; |
| 2485 | if (State.SkipStores.count(KillingDef)) |
| 2486 | continue; |
| 2487 | Instruction *SI = KillingDef->getMemoryInst(); |
| 2488 | |
| 2489 | Optional<MemoryLocation> MaybeSILoc; |
| 2490 | if (State.isMemTerminatorInst(SI)) |
| 2491 | MaybeSILoc = State.getLocForTerminator(SI).map( |
| 2492 | [](const std::pair<MemoryLocation, bool> &P) { return P.first; }); |
| 2493 | else |
| 2494 | MaybeSILoc = State.getLocForWriteEx(SI); |
| 2495 | |
| 2496 | if (!MaybeSILoc) { |
| 2497 | LLVM_DEBUG(dbgs() << "Failed to find analyzable write location for "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "Failed to find analyzable write location for " << *SI << "\n"; } } while (false) |
| 2498 | << *SI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "Failed to find analyzable write location for " << *SI << "\n"; } } while (false); |
| 2499 | continue; |
| 2500 | } |
| 2501 | MemoryLocation SILoc = *MaybeSILoc; |
| 2502 | assert(SILoc.Ptr && "SILoc should not be null")((SILoc.Ptr && "SILoc should not be null") ? static_cast <void> (0) : __assert_fail ("SILoc.Ptr && \"SILoc should not be null\"" , "/build/llvm-toolchain-snapshot-13~++20210301100612+564f5b0734bd/llvm/lib/Transforms/Scalar/DeadStoreElimination.cpp" , 2502, __PRETTY_FUNCTION__)); |
| 2503 | const Value *SILocUnd = getUnderlyingObject(SILoc.Ptr); |
| 2504 | |
| 2505 | MemoryAccess *Current = KillingDef; |
Value stored to 'Current' during its initialization is never read | |
| 2506 | LLVM_DEBUG(dbgs() << "Trying to eliminate MemoryDefs killed by "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "Trying to eliminate MemoryDefs killed by " << *KillingDef << " (" << *SI << ")\n" ; } } while (false) |
| 2507 | << *KillingDef << " (" << *SI << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "Trying to eliminate MemoryDefs killed by " << *KillingDef << " (" << *SI << ")\n" ; } } while (false); |
| 2508 | |
| 2509 | unsigned ScanLimit = MemorySSAScanLimit; |
| 2510 | unsigned WalkerStepLimit = MemorySSAUpwardsStepLimit; |
| 2511 | unsigned PartialLimit = MemorySSAPartialStoreLimit; |
| 2512 | // Worklist of MemoryAccesses that may be killed by KillingDef. |
| 2513 | SetVector<MemoryAccess *> ToCheck; |
| 2514 | |
| 2515 | if (SILocUnd) |
| 2516 | ToCheck.insert(KillingDef->getDefiningAccess()); |
| 2517 | |
| 2518 | bool Shortend = false; |
| 2519 | bool IsMemTerm = State.isMemTerminatorInst(SI); |
| 2520 | // Check if MemoryAccesses in the worklist are killed by KillingDef. |
| 2521 | for (unsigned I = 0; I < ToCheck.size(); I++) { |
| 2522 | Current = ToCheck[I]; |
| 2523 | if (State.SkipStores.count(Current)) |
| 2524 | continue; |
| 2525 | |
| 2526 | Optional<MemoryAccess *> Next = State.getDomMemoryDef( |
| 2527 | KillingDef, Current, SILoc, SILocUnd, ScanLimit, WalkerStepLimit, |
| 2528 | IsMemTerm, PartialLimit); |
| 2529 | |
| 2530 | if (!Next) { |
| 2531 | LLVM_DEBUG(dbgs() << " finished walk\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << " finished walk\n"; } } while (false ); |
| 2532 | continue; |
| 2533 | } |
| 2534 | |
| 2535 | MemoryAccess *EarlierAccess = *Next; |
| 2536 | LLVM_DEBUG(dbgs() << " Checking if we can kill " << *EarlierAccess)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << " Checking if we can kill " << *EarlierAccess; } } while (false); |
| 2537 | if (isa<MemoryPhi>(EarlierAccess)) { |
| 2538 | LLVM_DEBUG(dbgs() << "\n ... adding incoming values to worklist\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "\n ... adding incoming values to worklist\n" ; } } while (false); |
| 2539 | for (Value *V : cast<MemoryPhi>(EarlierAccess)->incoming_values()) { |
| 2540 | MemoryAccess *IncomingAccess = cast<MemoryAccess>(V); |
| 2541 | BasicBlock *IncomingBlock = IncomingAccess->getBlock(); |
| 2542 | BasicBlock *PhiBlock = EarlierAccess->getBlock(); |
| 2543 | |
| 2544 | // We only consider incoming MemoryAccesses that come before the |
| 2545 | // MemoryPhi. Otherwise we could discover candidates that do not |
| 2546 | // strictly dominate our starting def. |
| 2547 | if (State.PostOrderNumbers[IncomingBlock] > |
| 2548 | State.PostOrderNumbers[PhiBlock]) |
| 2549 | ToCheck.insert(IncomingAccess); |
| 2550 | } |
| 2551 | continue; |
| 2552 | } |
| 2553 | auto *NextDef = cast<MemoryDef>(EarlierAccess); |
| 2554 | Instruction *NI = NextDef->getMemoryInst(); |
| 2555 | LLVM_DEBUG(dbgs() << " (" << *NI << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << " (" << *NI << ")\n"; } } while (false); |
| 2556 | ToCheck.insert(NextDef->getDefiningAccess()); |
| 2557 | NumGetDomMemoryDefPassed++; |
| 2558 | |
| 2559 | if (!DebugCounter::shouldExecute(MemorySSACounter)) |
| 2560 | continue; |
| 2561 | |
| 2562 | MemoryLocation NILoc = *State.getLocForWriteEx(NI); |
| 2563 | |
| 2564 | if (IsMemTerm) { |
| 2565 | const Value *NIUnd = getUnderlyingObject(NILoc.Ptr); |
| 2566 | if (SILocUnd != NIUnd) |
| 2567 | continue; |
| 2568 | LLVM_DEBUG(dbgs() << "DSE: Remove Dead Store:\n DEAD: " << *NIdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Remove Dead Store:\n DEAD: " << *NI << "\n KILLER: " << *SI << '\n' ; } } while (false) |
| 2569 | << "\n KILLER: " << *SI << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Remove Dead Store:\n DEAD: " << *NI << "\n KILLER: " << *SI << '\n' ; } } while (false); |
| 2570 | State.deleteDeadInstruction(NI); |
| 2571 | ++NumFastStores; |
| 2572 | MadeChange = true; |
| 2573 | } else { |
| 2574 | // Check if NI overwrites SI. |
| 2575 | int64_t InstWriteOffset, DepWriteOffset; |
| 2576 | OverwriteResult OR = |
| 2577 | isOverwrite(SI, NI, SILoc, NILoc, State.DL, TLI, DepWriteOffset, |
| 2578 | InstWriteOffset, State.BatchAA, &F); |
| 2579 | if (OR == OW_MaybePartial) { |
| 2580 | auto Iter = State.IOLs.insert( |
| 2581 | std::make_pair<BasicBlock *, InstOverlapIntervalsTy>( |
| 2582 | NI->getParent(), InstOverlapIntervalsTy())); |
| 2583 | auto &IOL = Iter.first->second; |
| 2584 | OR = isPartialOverwrite(SILoc, NILoc, DepWriteOffset, InstWriteOffset, |
| 2585 | NI, IOL); |
| 2586 | } |
| 2587 | |
| 2588 | if (EnablePartialStoreMerging && OR == OW_PartialEarlierWithFullLater) { |
| 2589 | auto *Earlier = dyn_cast<StoreInst>(NI); |
| 2590 | auto *Later = dyn_cast<StoreInst>(SI); |
| 2591 | // We are re-using tryToMergePartialOverlappingStores, which requires |
| 2592 | // Earlier to domiante Later. |
| 2593 | // TODO: implement tryToMergeParialOverlappingStores using MemorySSA. |
| 2594 | if (Earlier && Later && DT.dominates(Earlier, Later)) { |
| 2595 | if (Constant *Merged = tryToMergePartialOverlappingStores( |
| 2596 | Earlier, Later, InstWriteOffset, DepWriteOffset, State.DL, |
| 2597 | State.BatchAA, &DT)) { |
| 2598 | |
| 2599 | // Update stored value of earlier store to merged constant. |
| 2600 | Earlier->setOperand(0, Merged); |
| 2601 | ++NumModifiedStores; |
| 2602 | MadeChange = true; |
| 2603 | |
| 2604 | Shortend = true; |
| 2605 | // Remove later store and remove any outstanding overlap intervals |
| 2606 | // for the updated store. |
| 2607 | State.deleteDeadInstruction(Later); |
| 2608 | auto I = State.IOLs.find(Earlier->getParent()); |
| 2609 | if (I != State.IOLs.end()) |
| 2610 | I->second.erase(Earlier); |
| 2611 | break; |
| 2612 | } |
| 2613 | } |
| 2614 | } |
| 2615 | |
| 2616 | if (OR == OW_Complete) { |
| 2617 | LLVM_DEBUG(dbgs() << "DSE: Remove Dead Store:\n DEAD: " << *NIdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Remove Dead Store:\n DEAD: " << *NI << "\n KILLER: " << *SI << '\n' ; } } while (false) |
| 2618 | << "\n KILLER: " << *SI << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Remove Dead Store:\n DEAD: " << *NI << "\n KILLER: " << *SI << '\n' ; } } while (false); |
| 2619 | State.deleteDeadInstruction(NI); |
| 2620 | ++NumFastStores; |
| 2621 | MadeChange = true; |
| 2622 | } |
| 2623 | } |
| 2624 | } |
| 2625 | |
| 2626 | // Check if the store is a no-op. |
| 2627 | if (!Shortend && isRemovable(SI) && |
| 2628 | State.storeIsNoop(KillingDef, SILoc, SILocUnd)) { |
| 2629 | LLVM_DEBUG(dbgs() << "DSE: Remove No-Op Store:\n DEAD: " << *SI << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("dse")) { dbgs() << "DSE: Remove No-Op Store:\n DEAD: " << *SI << '\n'; } } while (false); |
| 2630 | State.deleteDeadInstruction(SI); |
| 2631 | NumRedundantStores++; |
| 2632 | MadeChange = true; |
| 2633 | continue; |
| 2634 | } |
| 2635 | } |
| 2636 | |
| 2637 | if (EnablePartialOverwriteTracking) |
| 2638 | for (auto &KV : State.IOLs) |
| 2639 | MadeChange |= removePartiallyOverlappedStores(State.DL, KV.second, TLI); |
| 2640 | |
| 2641 | MadeChange |= State.eliminateDeadWritesAtEndOfFunction(); |
| 2642 | return MadeChange; |
| 2643 | } |
| 2644 | } // end anonymous namespace |
| 2645 | |
| 2646 | //===----------------------------------------------------------------------===// |
| 2647 | // DSE Pass |
| 2648 | //===----------------------------------------------------------------------===// |
| 2649 | PreservedAnalyses DSEPass::run(Function &F, FunctionAnalysisManager &AM) { |
| 2650 | AliasAnalysis &AA = AM.getResult<AAManager>(F); |
| 2651 | const TargetLibraryInfo &TLI = AM.getResult<TargetLibraryAnalysis>(F); |
| 2652 | DominatorTree &DT = AM.getResult<DominatorTreeAnalysis>(F); |
| 2653 | |
| 2654 | bool Changed = false; |
| 2655 | if (EnableMemorySSA) { |
| 2656 | MemorySSA &MSSA = AM.getResult<MemorySSAAnalysis>(F).getMSSA(); |
| 2657 | PostDominatorTree &PDT = AM.getResult<PostDominatorTreeAnalysis>(F); |
| 2658 | |
| 2659 | Changed = eliminateDeadStoresMemorySSA(F, AA, MSSA, DT, PDT, TLI); |
| 2660 | } else { |
| 2661 | MemoryDependenceResults &MD = AM.getResult<MemoryDependenceAnalysis>(F); |
| 2662 | |
| 2663 | Changed = eliminateDeadStores(F, &AA, &MD, &DT, &TLI); |
| 2664 | } |
| 2665 | |
| 2666 | #ifdef LLVM_ENABLE_STATS1 |
| 2667 | if (AreStatisticsEnabled()) |
| 2668 | for (auto &I : instructions(F)) |
| 2669 | NumRemainingStores += isa<StoreInst>(&I); |
| 2670 | #endif |
| 2671 | |
| 2672 | if (!Changed) |
| 2673 | return PreservedAnalyses::all(); |
| 2674 | |
| 2675 | PreservedAnalyses PA; |
| 2676 | PA.preserveSet<CFGAnalyses>(); |
| 2677 | PA.preserve<GlobalsAA>(); |
| 2678 | if (EnableMemorySSA) |
| 2679 | PA.preserve<MemorySSAAnalysis>(); |
| 2680 | else |
| 2681 | PA.preserve<MemoryDependenceAnalysis>(); |
| 2682 | return PA; |
| 2683 | } |
| 2684 | |
| 2685 | namespace { |
| 2686 | |
| 2687 | /// A legacy pass for the legacy pass manager that wraps \c DSEPass. |
| 2688 | class DSELegacyPass : public FunctionPass { |
| 2689 | public: |
| 2690 | static char ID; // Pass identification, replacement for typeid |
| 2691 | |
| 2692 | DSELegacyPass() : FunctionPass(ID) { |
| 2693 | initializeDSELegacyPassPass(*PassRegistry::getPassRegistry()); |
| 2694 | } |
| 2695 | |
| 2696 | bool runOnFunction(Function &F) override { |
| 2697 | if (skipFunction(F)) |
| 2698 | return false; |
| 2699 | |
| 2700 | AliasAnalysis &AA = getAnalysis<AAResultsWrapperPass>().getAAResults(); |
| 2701 | DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree(); |
| 2702 | const TargetLibraryInfo &TLI = |
| 2703 | getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F); |
| 2704 | |
| 2705 | bool Changed = false; |
| 2706 | if (EnableMemorySSA) { |
| 2707 | MemorySSA &MSSA = getAnalysis<MemorySSAWrapperPass>().getMSSA(); |
| 2708 | PostDominatorTree &PDT = |
| 2709 | getAnalysis<PostDominatorTreeWrapperPass>().getPostDomTree(); |
| 2710 | |
| 2711 | Changed = eliminateDeadStoresMemorySSA(F, AA, MSSA, DT, PDT, TLI); |
| 2712 | } else { |
| 2713 | MemoryDependenceResults &MD = |
| 2714 | getAnalysis<MemoryDependenceWrapperPass>().getMemDep(); |
| 2715 | |
| 2716 | Changed = eliminateDeadStores(F, &AA, &MD, &DT, &TLI); |
| 2717 | } |
| 2718 | |
| 2719 | #ifdef LLVM_ENABLE_STATS1 |
| 2720 | if (AreStatisticsEnabled()) |
| 2721 | for (auto &I : instructions(F)) |
| 2722 | NumRemainingStores += isa<StoreInst>(&I); |
| 2723 | #endif |
| 2724 | |
| 2725 | return Changed; |
| 2726 | } |
| 2727 | |
| 2728 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
| 2729 | AU.setPreservesCFG(); |
| 2730 | AU.addRequired<AAResultsWrapperPass>(); |
| 2731 | AU.addRequired<TargetLibraryInfoWrapperPass>(); |
| 2732 | AU.addPreserved<GlobalsAAWrapperPass>(); |
| 2733 | AU.addRequired<DominatorTreeWrapperPass>(); |
| 2734 | AU.addPreserved<DominatorTreeWrapperPass>(); |
| 2735 | |
| 2736 | if (EnableMemorySSA) { |
| 2737 | AU.addRequired<PostDominatorTreeWrapperPass>(); |
| 2738 | AU.addRequired<MemorySSAWrapperPass>(); |
| 2739 | AU.addPreserved<PostDominatorTreeWrapperPass>(); |
| 2740 | AU.addPreserved<MemorySSAWrapperPass>(); |
| 2741 | } else { |
| 2742 | AU.addRequired<MemoryDependenceWrapperPass>(); |
| 2743 | AU.addPreserved<MemoryDependenceWrapperPass>(); |
| 2744 | } |
| 2745 | } |
| 2746 | }; |
| 2747 | |
| 2748 | } // end anonymous namespace |
| 2749 | |
| 2750 | char DSELegacyPass::ID = 0; |
| 2751 | |
| 2752 | INITIALIZE_PASS_BEGIN(DSELegacyPass, "dse", "Dead Store Elimination", false,static void *initializeDSELegacyPassPassOnce(PassRegistry & Registry) { |
| 2753 | false)static void *initializeDSELegacyPassPassOnce(PassRegistry & Registry) { |
| 2754 | INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry); |
| 2755 | INITIALIZE_PASS_DEPENDENCY(PostDominatorTreeWrapperPass)initializePostDominatorTreeWrapperPassPass(Registry); |
| 2756 | INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)initializeAAResultsWrapperPassPass(Registry); |
| 2757 | INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)initializeGlobalsAAWrapperPassPass(Registry); |
| 2758 | INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)initializeMemorySSAWrapperPassPass(Registry); |
| 2759 | INITIALIZE_PASS_DEPENDENCY(MemoryDependenceWrapperPass)initializeMemoryDependenceWrapperPassPass(Registry); |
| 2760 | INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)initializeTargetLibraryInfoWrapperPassPass(Registry); |
| 2761 | INITIALIZE_PASS_END(DSELegacyPass, "dse", "Dead Store Elimination", false,PassInfo *PI = new PassInfo( "Dead Store Elimination", "dse", &DSELegacyPass::ID, PassInfo::NormalCtor_t(callDefaultCtor <DSELegacyPass>), false, false); Registry.registerPass( *PI, true); return PI; } static llvm::once_flag InitializeDSELegacyPassPassFlag ; void llvm::initializeDSELegacyPassPass(PassRegistry &Registry ) { llvm::call_once(InitializeDSELegacyPassPassFlag, initializeDSELegacyPassPassOnce , std::ref(Registry)); } |
| 2762 | false)PassInfo *PI = new PassInfo( "Dead Store Elimination", "dse", &DSELegacyPass::ID, PassInfo::NormalCtor_t(callDefaultCtor <DSELegacyPass>), false, false); Registry.registerPass( *PI, true); return PI; } static llvm::once_flag InitializeDSELegacyPassPassFlag ; void llvm::initializeDSELegacyPassPass(PassRegistry &Registry ) { llvm::call_once(InitializeDSELegacyPassPassFlag, initializeDSELegacyPassPassOnce , std::ref(Registry)); } |
| 2763 | |
| 2764 | FunctionPass *llvm::createDeadStoreEliminationPass() { |
| 2765 | return new DSELegacyPass(); |
| 2766 | } |