| 1 |
//===- Delta.cpp - Delta Debugging Algorithm Implementation ---------------===// |
| 2 |
// |
| 3 |
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 |
// See https://llvm.org/LICENSE.txt for license information. |
| 5 |
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 |
// |
| 7 |
//===----------------------------------------------------------------------===// |
| 8 |
// |
| 9 |
// This file contains the implementation for the Delta Debugging Algorithm: |
| 10 |
// it splits a given set of Targets (i.e. Functions, Instructions, BBs, etc.) |
| 11 |
// into chunks and tries to reduce the number chunks that are interesting. |
| 12 |
// |
| 13 |
//===----------------------------------------------------------------------===// |
| 14 |
|
| 15 |
#include "Delta.h" |
| 16 |
#include "ReducerWorkItem.h" |
| 17 |
#include "TestRunner.h" |
| 18 |
#include "Utils.h" |
| 19 |
#include "llvm/ADT/STLExtras.h" |
| 20 |
#include "llvm/Analysis/ModuleSummaryAnalysis.h" |
| 21 |
#include "llvm/Analysis/ProfileSummaryInfo.h" |
| 22 |
#include "llvm/Bitcode/BitcodeReader.h" |
| 23 |
#include "llvm/Bitcode/BitcodeWriter.h" |
| 24 |
#include "llvm/CodeGen/MachineFunction.h" |
| 25 |
#include "llvm/IR/Module.h" |
| 26 |
#include "llvm/IR/Verifier.h" |
| 27 |
#include "llvm/MC/TargetRegistry.h" |
| 28 |
#include "llvm/Support/CommandLine.h" |
| 29 |
#include "llvm/Support/MemoryBufferRef.h" |
| 30 |
#include "llvm/Support/ThreadPool.h" |
| 31 |
#include <fstream> |
| 32 |
#include <set> |
| 33 |
|
| 34 |
using namespace llvm; |
| 35 |
|
| 36 |
extern cl::OptionCategory LLVMReduceOptions; |
| 37 |
|
| 38 |
static cl::opt<bool> AbortOnInvalidReduction( |
| 39 |
"abort-on-invalid-reduction", |
| 40 |
cl::desc("Abort if any reduction results in invalid IR"), |
| 41 |
cl::cat(LLVMReduceOptions)); |
| 42 |
|
| 43 |
static cl::opt<unsigned int> StartingGranularityLevel( |
| 44 |
"starting-granularity-level", |
| 45 |
cl::desc("Number of times to divide chunks prior to first test"), |
| 46 |
cl::cat(LLVMReduceOptions)); |
| 47 |
|
| 48 |
#ifdef LLVM_ENABLE_THREADS |
| 49 |
static cl::opt<unsigned> NumJobs( |
| 50 |
"j", |
| 51 |
cl::desc("Maximum number of threads to use to process chunks. Set to 1 to " |
| 52 |
"disable parallelism."), |
| 53 |
cl::init(1), cl::cat(LLVMReduceOptions)); |
| 54 |
#else |
| 55 |
unsigned NumJobs = 1; |
| 56 |
#endif |
| 57 |
|
| 58 |
/// Splits Chunks in half and prints them. |
| 59 |
/// If unable to split (when chunk size is 1) returns false. |
| 60 |
static bool increaseGranularity(std::vector<Chunk> &Chunks) { |
| 61 |
if (Verbose) |
| 62 |
errs() << "Increasing granularity..."; |
| 63 |
std::vector<Chunk> NewChunks; |
| 64 |
bool SplitAny = false; |
| 65 |
|
| 66 |
for (Chunk C : Chunks) { |
| 67 |
if (C.End - C.Begin == 0) |
| 68 |
NewChunks.push_back(C); |
| 69 |
else { |
| 70 |
int Half = (C.Begin + C.End) / 2; |
| 71 |
NewChunks.push_back({C.Begin, Half}); |
| 72 |
NewChunks.push_back({Half + 1, C.End}); |
| 73 |
SplitAny = true; |
| 74 |
} |
| 75 |
} |
| 76 |
if (SplitAny) { |
| 77 |
Chunks = NewChunks; |
| 78 |
if (Verbose) { |
| 79 |
errs() << "Success! " << NewChunks.size() << " New Chunks:\n"; |
| 80 |
for (auto C : Chunks) { |
| 81 |
errs() << '\t'; |
| 82 |
C.print(); |
| 83 |
errs() << '\n'; |
| 84 |
} |
| 85 |
} |
| 86 |
} |
| 87 |
return SplitAny; |
| 88 |
} |
| 89 |
|
| 90 |
// Check if \p ChunkToCheckForUninterestingness is interesting. Returns the |
| 91 |
// modified module if the chunk resulted in a reduction. |
| 92 |
static std::unique_ptr<ReducerWorkItem> |
| 93 |
CheckChunk(const Chunk ChunkToCheckForUninterestingness, |
| 94 |
std::unique_ptr<ReducerWorkItem> Clone, const TestRunner &Test, |
| 95 |
ReductionFunc ExtractChunksFromModule, |
| 96 |
const DenseSet<Chunk> &UninterestingChunks, |
| 97 |
const std::vector<Chunk> &ChunksStillConsideredInteresting) { |
| 98 |
// Take all of ChunksStillConsideredInteresting chunks, except those we've |
| 99 |
// already deemed uninteresting (UninterestingChunks) but didn't remove |
| 100 |
// from ChunksStillConsideredInteresting yet, and additionally ignore |
| 101 |
// ChunkToCheckForUninterestingness chunk. |
| 102 |
std::vector<Chunk> CurrentChunks; |
| 103 |
CurrentChunks.reserve(ChunksStillConsideredInteresting.size() - |
| 104 |
UninterestingChunks.size() - 1); |
| 105 |
copy_if(ChunksStillConsideredInteresting, std::back_inserter(CurrentChunks), |
| 106 |
[&](const Chunk &C) { |
| 107 |
return C != ChunkToCheckForUninterestingness && |
| 108 |
!UninterestingChunks.count(C); |
| 109 |
}); |
| 110 |
|
| 111 |
// Generate Module with only Targets inside Current Chunks |
| 112 |
Oracle O(CurrentChunks); |
| 113 |
ExtractChunksFromModule(O, *Clone); |
| 114 |
|
| 115 |
// Some reductions may result in invalid IR. Skip such reductions. |
| 116 |
if (Clone->verify(&errs())) { |
| 117 |
if (AbortOnInvalidReduction) { |
| 118 |
errs() << "Invalid reduction, aborting.\n"; |
| 119 |
Clone->print(errs()); |
| 120 |
exit(1); |
| 121 |
} |
| 122 |
if (Verbose) { |
| 123 |
errs() << " **** WARNING | reduction resulted in invalid module, " |
| 124 |
"skipping\n"; |
| 125 |
} |
| 126 |
return nullptr; |
| 127 |
} |
| 128 |
|
| 129 |
if (Verbose) { |
| 130 |
errs() << "Ignoring: "; |
| 131 |
ChunkToCheckForUninterestingness.print(); |
| 132 |
for (const Chunk &C : UninterestingChunks) |
| 133 |
C.print(); |
| 134 |
errs() << "\n"; |
| 135 |
} |
| 136 |
|
| 137 |
if (!Clone->isReduced(Test)) { |
| 138 |
// Program became non-reduced, so this chunk appears to be interesting. |
| 139 |
if (Verbose) |
| 140 |
errs() << "\n"; |
| 141 |
return nullptr; |
| 142 |
} |
| 143 |
return Clone; |
| 144 |
} |
| 145 |
|
| 146 |
static SmallString<0> ProcessChunkFromSerializedBitcode( |
| 147 |
const Chunk ChunkToCheckForUninterestingness, const TestRunner &Test, |
| 148 |
ReductionFunc ExtractChunksFromModule, |
| 149 |
const DenseSet<Chunk> &UninterestingChunks, |
| 150 |
ArrayRef<Chunk> ChunksStillConsideredInteresting, StringRef OriginalBC, |
| 151 |
std::atomic<bool> &AnyReduced) { |
| 152 |
LLVMContext Ctx; |
| 153 |
auto CloneMMM = std::make_unique<ReducerWorkItem>(); |
| 154 |
MemoryBufferRef Data(OriginalBC, "<bc file>"); |
| 155 |
CloneMMM->readBitcode(Data, Ctx, Test.getToolName()); |
| 156 |
|
| 157 |
SmallString<0> Result; |
| 158 |
if (std::unique_ptr<ReducerWorkItem> ChunkResult = |
| 159 |
CheckChunk(ChunkToCheckForUninterestingness, std::move(CloneMMM), |
| 160 |
Test, ExtractChunksFromModule, UninterestingChunks, |
| 161 |
ChunksStillConsideredInteresting)) { |
| 162 |
raw_svector_ostream BCOS(Result); |
| 163 |
ChunkResult->writeBitcode(BCOS); |
| 164 |
// Communicate that the task reduced a chunk. |
| 165 |
AnyReduced = true; |
| 166 |
} |
| 167 |
return Result; |
| 168 |
} |
| 169 |
|
| 170 |
using SharedTaskQueue = std::deque<std::shared_future<SmallString<0>>>; |
| 171 |
|
| 172 |
static void waitAndDiscardResultsBarrier(SharedTaskQueue &TaskQueue) { |
| 173 |
while (!TaskQueue.empty()) { |
| 174 |
auto &Future = TaskQueue.front(); |
| 175 |
Future.wait(); |
| 176 |
TaskQueue.pop_front(); |
| 177 |
} |
| 178 |
} |
| 179 |
|
| 180 |
/// Runs the Delta Debugging algorithm, splits the code into chunks and |
| 181 |
/// reduces the amount of chunks that are considered interesting by the |
| 182 |
/// given test. The number of chunks is determined by a preliminary run of the |
| 183 |
/// reduction pass where no change must be made to the module. |
| 184 |
void llvm::runDeltaPass(TestRunner &Test, ReductionFunc ExtractChunksFromModule, |
| 185 |
StringRef Message) { |
| 186 |
assert(!Test.getProgram().verify(&errs()) && |
| 187 |
"input module is broken before making changes"); |
| 188 |
errs() << "*** " << Message << "...\n"; |
| 189 |
|
| 190 |
int Targets; |
| 191 |
{ |
| 192 |
// Count the number of chunks by counting the number of calls to |
| 193 |
// Oracle::shouldKeep() but always returning true so no changes are |
| 194 |
// made. |
| 195 |
std::vector<Chunk> AllChunks = {{0, INT_MAX}}; |
| 196 |
Oracle Counter(AllChunks); |
| 197 |
ExtractChunksFromModule(Counter, Test.getProgram()); |
| 198 |
Targets = Counter.count(); |
| 199 |
|
| 200 |
assert(!Test.getProgram().verify(&errs()) && |
| 201 |
"input module is broken after counting chunks"); |
| 202 |
assert(Test.getProgram().isReduced(Test) && |
| 203 |
"input module no longer interesting after counting chunks"); |
| 204 |
|
| 205 |
#ifndef NDEBUG |
| 206 |
// Make sure that the number of chunks does not change as we reduce. |
| 207 |
std::vector<Chunk> NoChunks = {{0, INT_MAX}}; |
| 208 |
Oracle NoChunksCounter(NoChunks); |
| 209 |
std::unique_ptr<ReducerWorkItem> Clone = |
| 210 |
Test.getProgram().clone(Test.getTargetMachine()); |
| 211 |
ExtractChunksFromModule(NoChunksCounter, *Clone); |
| 212 |
assert(Targets == NoChunksCounter.count() && |
| 213 |
"number of chunks changes when reducing"); |
| 214 |
#endif |
| 215 |
} |
| 216 |
if (!Targets) { |
| 217 |
if (Verbose) |
| 218 |
errs() << "\nNothing to reduce\n"; |
| 219 |
errs() << "----------------------------\n"; |
| 220 |
return; |
| 221 |
} |
| 222 |
|
| 223 |
std::vector<Chunk> ChunksStillConsideredInteresting = {{0, Targets - 1}}; |
| 224 |
std::unique_ptr<ReducerWorkItem> ReducedProgram; |
| 225 |
|
| 226 |
for (unsigned int Level = 0; Level < StartingGranularityLevel; Level++) { |
| 227 |
increaseGranularity(ChunksStillConsideredInteresting); |
| 228 |
} |
| 229 |
|
| 230 |
std::atomic<bool> AnyReduced; |
| 231 |
std::unique_ptr<ThreadPool> ChunkThreadPoolPtr; |
| 232 |
if (NumJobs > 1) |
| 233 |
ChunkThreadPoolPtr = |
| 234 |
std::make_unique<ThreadPool>(hardware_concurrency(NumJobs)); |
| 235 |
|
| 236 |
bool FoundAtLeastOneNewUninterestingChunkWithCurrentGranularity; |
| 237 |
do { |
| 238 |
FoundAtLeastOneNewUninterestingChunkWithCurrentGranularity = false; |
| 239 |
|
| 240 |
DenseSet<Chunk> UninterestingChunks; |
| 241 |
|
| 242 |
// When running with more than one thread, serialize the original bitcode |
| 243 |
// to OriginalBC. |
| 244 |
SmallString<0> OriginalBC; |
| 245 |
if (NumJobs > 1) { |
| 246 |
raw_svector_ostream BCOS(OriginalBC); |
| 247 |
Test.getProgram().writeBitcode(BCOS); |
| 248 |
} |
| 249 |
|
| 250 |
SharedTaskQueue TaskQueue; |
| 251 |
for (auto I = ChunksStillConsideredInteresting.rbegin(), |
| 252 |
E = ChunksStillConsideredInteresting.rend(); |
| 253 |
I != E; ++I) { |
| 254 |
std::unique_ptr<ReducerWorkItem> Result = nullptr; |
| 255 |
unsigned WorkLeft = std::distance(I, E); |
| 256 |
|
| 257 |
// Run in parallel mode, if the user requested more than one thread and |
| 258 |
// there are at least a few chunks to process. |
| 259 |
if (NumJobs > 1 && WorkLeft > 1) { |
| 260 |
unsigned NumInitialTasks = std::min(WorkLeft, unsigned(NumJobs)); |
| 261 |
unsigned NumChunksProcessed = 0; |
| 262 |
|
| 263 |
ThreadPool &ChunkThreadPool = *ChunkThreadPoolPtr; |
| 264 |
TaskQueue.clear(); |
| 265 |
|
| 266 |
AnyReduced = false; |
| 267 |
// Queue jobs to process NumInitialTasks chunks in parallel using |
| 268 |
// ChunkThreadPool. When the tasks are added to the pool, parse the |
| 269 |
// original module from OriginalBC with a fresh LLVMContext object. This |
| 270 |
// ensures that the cloned module of each task uses an independent |
| 271 |
// LLVMContext object. If a task reduces the input, serialize the result |
| 272 |
// back in the corresponding Result element. |
| 273 |
for (unsigned J = 0; J < NumInitialTasks; ++J) { |
| 274 |
Chunk ChunkToCheck = *(I + J); |
| 275 |
TaskQueue.emplace_back(ChunkThreadPool.async( |
| 276 |
ProcessChunkFromSerializedBitcode, ChunkToCheck, std::ref(Test), |
| 277 |
ExtractChunksFromModule, UninterestingChunks, |
| 278 |
ChunksStillConsideredInteresting, OriginalBC, |
| 279 |
std::ref(AnyReduced))); |
| 280 |
} |
| 281 |
|
| 282 |
// Start processing results of the queued tasks. We wait for the first |
| 283 |
// task in the queue to finish. If it reduced a chunk, we parse the |
| 284 |
// result and exit the loop. |
| 285 |
// Otherwise we will try to schedule a new task, if |
| 286 |
// * no other pending job reduced a chunk and |
| 287 |
// * we have not reached the end of the chunk. |
| 288 |
while (!TaskQueue.empty()) { |
| 289 |
auto &Future = TaskQueue.front(); |
| 290 |
Future.wait(); |
| 291 |
|
| 292 |
NumChunksProcessed++; |
| 293 |
SmallString<0> Res = Future.get(); |
| 294 |
TaskQueue.pop_front(); |
| 295 |
if (Res.empty()) { |
| 296 |
unsigned NumScheduledTasks = NumChunksProcessed + TaskQueue.size(); |
| 297 |
if (!AnyReduced && I + NumScheduledTasks != E) { |
| 298 |
Chunk ChunkToCheck = *(I + NumScheduledTasks); |
| 299 |
TaskQueue.emplace_back(ChunkThreadPool.async( |
| 300 |
ProcessChunkFromSerializedBitcode, ChunkToCheck, |
| 301 |
std::ref(Test), ExtractChunksFromModule, UninterestingChunks, |
| 302 |
ChunksStillConsideredInteresting, OriginalBC, |
| 303 |
std::ref(AnyReduced))); |
| 304 |
} |
| 305 |
continue; |
| 306 |
} |
| 307 |
|
| 308 |
Result = std::make_unique<ReducerWorkItem>(); |
| 309 |
MemoryBufferRef Data(StringRef(Res), "<bc file>"); |
| 310 |
Result->readBitcode(Data, Test.getProgram().M->getContext(), |
| 311 |
Test.getToolName()); |
| 312 |
break; |
| 313 |
} |
| 314 |
|
| 315 |
// If we broke out of the loop, we still need to wait for everything to |
| 316 |
// avoid race access to the chunk set. |
| 317 |
// |
| 318 |
// TODO: Create a way to kill remaining items we're ignoring; they could |
| 319 |
// take a long time. |
| 320 |
waitAndDiscardResultsBarrier(TaskQueue); |
| 321 |
|
| 322 |
// Forward I to the last chunk processed in parallel. |
| 323 |
I += NumChunksProcessed - 1; |
| 324 |
} else { |
| 325 |
Result = |
| 326 |
CheckChunk(*I, Test.getProgram().clone(Test.getTargetMachine()), |
| 327 |
Test, ExtractChunksFromModule, UninterestingChunks, |
| 328 |
ChunksStillConsideredInteresting); |
| 329 |
} |
| 330 |
|
| 331 |
if (!Result) |
| 332 |
continue; |
| 333 |
|
| 334 |
const Chunk ChunkToCheckForUninterestingness = *I; |
| 335 |
FoundAtLeastOneNewUninterestingChunkWithCurrentGranularity = true; |
| 336 |
UninterestingChunks.insert(ChunkToCheckForUninterestingness); |
| 337 |
ReducedProgram = std::move(Result); |
| 338 |
|
| 339 |
// FIXME: Report meaningful progress info |
| 340 |
Test.writeOutput(" **** SUCCESS | Saved new best reduction to "); |
| 341 |
} |
| 342 |
// Delete uninteresting chunks |
| 343 |
erase_if(ChunksStillConsideredInteresting, |
| 344 |
[&UninterestingChunks](const Chunk &C) { |
| 345 |
return UninterestingChunks.count(C); |
| 346 |
}); |
| 347 |
} while (!ChunksStillConsideredInteresting.empty() && |
| 348 |
(FoundAtLeastOneNewUninterestingChunkWithCurrentGranularity || |
| 349 |
increaseGranularity(ChunksStillConsideredInteresting))); |
| 350 |
|
| 351 |
// If we reduced the testcase replace it |
| 352 |
if (ReducedProgram) |
| 353 |
Test.setProgram(std::move(ReducedProgram)); |
| 354 |
if (Verbose) |
| 355 |
errs() << "Couldn't increase anymore.\n"; |
| 356 |
errs() << "----------------------------\n"; |
| 357 |
} |