1 | //===-- ThreadSanitizer.cpp - race detector -------------------------------===// |
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 is a part of ThreadSanitizer, a race detector. |
10 | // |
11 | // The tool is under development, for the details about previous versions see |
12 | // http://code.google.com/p/data-race-test |
13 | // |
14 | // The instrumentation phase is quite simple: |
15 | // - Insert calls to run-time library before every memory access. |
16 | // - Optimizations may apply to avoid instrumenting some of the accesses. |
17 | // - Insert calls at function entry/exit. |
18 | // The rest is handled by the run-time library. |
19 | //===----------------------------------------------------------------------===// |
20 | |
21 | #include "llvm/Transforms/Instrumentation/ThreadSanitizer.h" |
22 | #include "llvm/ADT/DenseMap.h" |
23 | #include "llvm/ADT/SmallString.h" |
24 | #include "llvm/ADT/SmallVector.h" |
25 | #include "llvm/ADT/Statistic.h" |
26 | #include "llvm/ADT/StringExtras.h" |
27 | #include "llvm/Analysis/CaptureTracking.h" |
28 | #include "llvm/Analysis/TargetLibraryInfo.h" |
29 | #include "llvm/Analysis/ValueTracking.h" |
30 | #include "llvm/IR/DataLayout.h" |
31 | #include "llvm/IR/Function.h" |
32 | #include "llvm/IR/IRBuilder.h" |
33 | #include "llvm/IR/Instructions.h" |
34 | #include "llvm/IR/IntrinsicInst.h" |
35 | #include "llvm/IR/Intrinsics.h" |
36 | #include "llvm/IR/LLVMContext.h" |
37 | #include "llvm/IR/Metadata.h" |
38 | #include "llvm/IR/Module.h" |
39 | #include "llvm/IR/Type.h" |
40 | #include "llvm/ProfileData/InstrProf.h" |
41 | #include "llvm/Support/CommandLine.h" |
42 | #include "llvm/Support/Debug.h" |
43 | #include "llvm/Support/MathExtras.h" |
44 | #include "llvm/Support/raw_ostream.h" |
45 | #include "llvm/Transforms/Instrumentation.h" |
46 | #include "llvm/Transforms/Utils/EscapeEnumerator.h" |
47 | #include "llvm/Transforms/Utils/Local.h" |
48 | #include "llvm/Transforms/Utils/ModuleUtils.h" |
49 | |
50 | using namespace llvm; |
51 | |
52 | #define DEBUG_TYPE "tsan" |
53 | |
54 | static cl::opt<bool> ClInstrumentMemoryAccesses( |
55 | "tsan-instrument-memory-accesses" , cl::init(Val: true), |
56 | cl::desc("Instrument memory accesses" ), cl::Hidden); |
57 | static cl::opt<bool> |
58 | ClInstrumentFuncEntryExit("tsan-instrument-func-entry-exit" , cl::init(Val: true), |
59 | cl::desc("Instrument function entry and exit" ), |
60 | cl::Hidden); |
61 | static cl::opt<bool> ClHandleCxxExceptions( |
62 | "tsan-handle-cxx-exceptions" , cl::init(Val: true), |
63 | cl::desc("Handle C++ exceptions (insert cleanup blocks for unwinding)" ), |
64 | cl::Hidden); |
65 | static cl::opt<bool> ClInstrumentAtomics("tsan-instrument-atomics" , |
66 | cl::init(Val: true), |
67 | cl::desc("Instrument atomics" ), |
68 | cl::Hidden); |
69 | static cl::opt<bool> ClInstrumentMemIntrinsics( |
70 | "tsan-instrument-memintrinsics" , cl::init(Val: true), |
71 | cl::desc("Instrument memintrinsics (memset/memcpy/memmove)" ), cl::Hidden); |
72 | static cl::opt<bool> ClDistinguishVolatile( |
73 | "tsan-distinguish-volatile" , cl::init(Val: false), |
74 | cl::desc("Emit special instrumentation for accesses to volatiles" ), |
75 | cl::Hidden); |
76 | static cl::opt<bool> ClInstrumentReadBeforeWrite( |
77 | "tsan-instrument-read-before-write" , cl::init(Val: false), |
78 | cl::desc("Do not eliminate read instrumentation for read-before-writes" ), |
79 | cl::Hidden); |
80 | static cl::opt<bool> ClCompoundReadBeforeWrite( |
81 | "tsan-compound-read-before-write" , cl::init(Val: false), |
82 | cl::desc("Emit special compound instrumentation for reads-before-writes" ), |
83 | cl::Hidden); |
84 | |
85 | STATISTIC(NumInstrumentedReads, "Number of instrumented reads" ); |
86 | STATISTIC(NumInstrumentedWrites, "Number of instrumented writes" ); |
87 | STATISTIC(NumOmittedReadsBeforeWrite, |
88 | "Number of reads ignored due to following writes" ); |
89 | STATISTIC(NumAccessesWithBadSize, "Number of accesses with bad size" ); |
90 | STATISTIC(NumInstrumentedVtableWrites, "Number of vtable ptr writes" ); |
91 | STATISTIC(NumInstrumentedVtableReads, "Number of vtable ptr reads" ); |
92 | STATISTIC(NumOmittedReadsFromConstantGlobals, |
93 | "Number of reads from constant globals" ); |
94 | STATISTIC(NumOmittedReadsFromVtable, "Number of vtable reads" ); |
95 | STATISTIC(NumOmittedNonCaptured, "Number of accesses ignored due to capturing" ); |
96 | |
97 | const char kTsanModuleCtorName[] = "tsan.module_ctor" ; |
98 | const char kTsanInitName[] = "__tsan_init" ; |
99 | |
100 | namespace { |
101 | |
102 | /// ThreadSanitizer: instrument the code in module to find races. |
103 | /// |
104 | /// Instantiating ThreadSanitizer inserts the tsan runtime library API function |
105 | /// declarations into the module if they don't exist already. Instantiating |
106 | /// ensures the __tsan_init function is in the list of global constructors for |
107 | /// the module. |
108 | struct ThreadSanitizer { |
109 | ThreadSanitizer() { |
110 | // Check options and warn user. |
111 | if (ClInstrumentReadBeforeWrite && ClCompoundReadBeforeWrite) { |
112 | errs() |
113 | << "warning: Option -tsan-compound-read-before-write has no effect " |
114 | "when -tsan-instrument-read-before-write is set.\n" ; |
115 | } |
116 | } |
117 | |
118 | bool sanitizeFunction(Function &F, const TargetLibraryInfo &TLI); |
119 | |
120 | private: |
121 | // Internal Instruction wrapper that contains more information about the |
122 | // Instruction from prior analysis. |
123 | struct InstructionInfo { |
124 | // Instrumentation emitted for this instruction is for a compounded set of |
125 | // read and write operations in the same basic block. |
126 | static constexpr unsigned kCompoundRW = (1U << 0); |
127 | |
128 | explicit InstructionInfo(Instruction *Inst) : Inst(Inst) {} |
129 | |
130 | Instruction *Inst; |
131 | unsigned Flags = 0; |
132 | }; |
133 | |
134 | void initialize(Module &M, const TargetLibraryInfo &TLI); |
135 | bool instrumentLoadOrStore(const InstructionInfo &II, const DataLayout &DL); |
136 | bool instrumentAtomic(Instruction *I, const DataLayout &DL); |
137 | bool instrumentMemIntrinsic(Instruction *I); |
138 | void chooseInstructionsToInstrument(SmallVectorImpl<Instruction *> &Local, |
139 | SmallVectorImpl<InstructionInfo> &All, |
140 | const DataLayout &DL); |
141 | bool addrPointsToConstantData(Value *Addr); |
142 | int getMemoryAccessFuncIndex(Type *OrigTy, Value *Addr, const DataLayout &DL); |
143 | void InsertRuntimeIgnores(Function &F); |
144 | |
145 | Type *IntptrTy; |
146 | FunctionCallee TsanFuncEntry; |
147 | FunctionCallee TsanFuncExit; |
148 | FunctionCallee TsanIgnoreBegin; |
149 | FunctionCallee TsanIgnoreEnd; |
150 | // Accesses sizes are powers of two: 1, 2, 4, 8, 16. |
151 | static const size_t kNumberOfAccessSizes = 5; |
152 | FunctionCallee TsanRead[kNumberOfAccessSizes]; |
153 | FunctionCallee TsanWrite[kNumberOfAccessSizes]; |
154 | FunctionCallee TsanUnalignedRead[kNumberOfAccessSizes]; |
155 | FunctionCallee TsanUnalignedWrite[kNumberOfAccessSizes]; |
156 | FunctionCallee TsanVolatileRead[kNumberOfAccessSizes]; |
157 | FunctionCallee TsanVolatileWrite[kNumberOfAccessSizes]; |
158 | FunctionCallee TsanUnalignedVolatileRead[kNumberOfAccessSizes]; |
159 | FunctionCallee TsanUnalignedVolatileWrite[kNumberOfAccessSizes]; |
160 | FunctionCallee TsanCompoundRW[kNumberOfAccessSizes]; |
161 | FunctionCallee TsanUnalignedCompoundRW[kNumberOfAccessSizes]; |
162 | FunctionCallee TsanAtomicLoad[kNumberOfAccessSizes]; |
163 | FunctionCallee TsanAtomicStore[kNumberOfAccessSizes]; |
164 | FunctionCallee TsanAtomicRMW[AtomicRMWInst::LAST_BINOP + 1] |
165 | [kNumberOfAccessSizes]; |
166 | FunctionCallee TsanAtomicCAS[kNumberOfAccessSizes]; |
167 | FunctionCallee TsanAtomicThreadFence; |
168 | FunctionCallee TsanAtomicSignalFence; |
169 | FunctionCallee TsanVptrUpdate; |
170 | FunctionCallee TsanVptrLoad; |
171 | FunctionCallee MemmoveFn, MemcpyFn, MemsetFn; |
172 | }; |
173 | |
174 | void insertModuleCtor(Module &M) { |
175 | getOrCreateSanitizerCtorAndInitFunctions( |
176 | M, CtorName: kTsanModuleCtorName, InitName: kTsanInitName, /*InitArgTypes=*/{}, |
177 | /*InitArgs=*/{}, |
178 | // This callback is invoked when the functions are created the first |
179 | // time. Hook them into the global ctors list in that case: |
180 | FunctionsCreatedCallback: [&](Function *Ctor, FunctionCallee) { appendToGlobalCtors(M, F: Ctor, Priority: 0); }); |
181 | } |
182 | } // namespace |
183 | |
184 | PreservedAnalyses ThreadSanitizerPass::run(Function &F, |
185 | FunctionAnalysisManager &FAM) { |
186 | ThreadSanitizer TSan; |
187 | if (TSan.sanitizeFunction(F, TLI: FAM.getResult<TargetLibraryAnalysis>(IR&: F))) |
188 | return PreservedAnalyses::none(); |
189 | return PreservedAnalyses::all(); |
190 | } |
191 | |
192 | PreservedAnalyses ModuleThreadSanitizerPass::run(Module &M, |
193 | ModuleAnalysisManager &MAM) { |
194 | insertModuleCtor(M); |
195 | return PreservedAnalyses::none(); |
196 | } |
197 | void ThreadSanitizer::initialize(Module &M, const TargetLibraryInfo &TLI) { |
198 | const DataLayout &DL = M.getDataLayout(); |
199 | LLVMContext &Ctx = M.getContext(); |
200 | IntptrTy = DL.getIntPtrType(C&: Ctx); |
201 | |
202 | IRBuilder<> IRB(Ctx); |
203 | AttributeList Attr; |
204 | Attr = Attr.addFnAttribute(C&: Ctx, Kind: Attribute::NoUnwind); |
205 | // Initialize the callbacks. |
206 | TsanFuncEntry = M.getOrInsertFunction(Name: "__tsan_func_entry" , AttributeList: Attr, |
207 | RetTy: IRB.getVoidTy(), Args: IRB.getPtrTy()); |
208 | TsanFuncExit = |
209 | M.getOrInsertFunction(Name: "__tsan_func_exit" , AttributeList: Attr, RetTy: IRB.getVoidTy()); |
210 | TsanIgnoreBegin = M.getOrInsertFunction(Name: "__tsan_ignore_thread_begin" , AttributeList: Attr, |
211 | RetTy: IRB.getVoidTy()); |
212 | TsanIgnoreEnd = |
213 | M.getOrInsertFunction(Name: "__tsan_ignore_thread_end" , AttributeList: Attr, RetTy: IRB.getVoidTy()); |
214 | IntegerType *OrdTy = IRB.getInt32Ty(); |
215 | for (size_t i = 0; i < kNumberOfAccessSizes; ++i) { |
216 | const unsigned ByteSize = 1U << i; |
217 | const unsigned BitSize = ByteSize * 8; |
218 | std::string ByteSizeStr = utostr(X: ByteSize); |
219 | std::string BitSizeStr = utostr(X: BitSize); |
220 | SmallString<32> ReadName("__tsan_read" + ByteSizeStr); |
221 | TsanRead[i] = M.getOrInsertFunction(Name: ReadName, AttributeList: Attr, RetTy: IRB.getVoidTy(), |
222 | Args: IRB.getPtrTy()); |
223 | |
224 | SmallString<32> WriteName("__tsan_write" + ByteSizeStr); |
225 | TsanWrite[i] = M.getOrInsertFunction(Name: WriteName, AttributeList: Attr, RetTy: IRB.getVoidTy(), |
226 | Args: IRB.getPtrTy()); |
227 | |
228 | SmallString<64> UnalignedReadName("__tsan_unaligned_read" + ByteSizeStr); |
229 | TsanUnalignedRead[i] = M.getOrInsertFunction( |
230 | Name: UnalignedReadName, AttributeList: Attr, RetTy: IRB.getVoidTy(), Args: IRB.getPtrTy()); |
231 | |
232 | SmallString<64> UnalignedWriteName("__tsan_unaligned_write" + ByteSizeStr); |
233 | TsanUnalignedWrite[i] = M.getOrInsertFunction( |
234 | Name: UnalignedWriteName, AttributeList: Attr, RetTy: IRB.getVoidTy(), Args: IRB.getPtrTy()); |
235 | |
236 | SmallString<64> VolatileReadName("__tsan_volatile_read" + ByteSizeStr); |
237 | TsanVolatileRead[i] = M.getOrInsertFunction( |
238 | Name: VolatileReadName, AttributeList: Attr, RetTy: IRB.getVoidTy(), Args: IRB.getPtrTy()); |
239 | |
240 | SmallString<64> VolatileWriteName("__tsan_volatile_write" + ByteSizeStr); |
241 | TsanVolatileWrite[i] = M.getOrInsertFunction( |
242 | Name: VolatileWriteName, AttributeList: Attr, RetTy: IRB.getVoidTy(), Args: IRB.getPtrTy()); |
243 | |
244 | SmallString<64> UnalignedVolatileReadName("__tsan_unaligned_volatile_read" + |
245 | ByteSizeStr); |
246 | TsanUnalignedVolatileRead[i] = M.getOrInsertFunction( |
247 | Name: UnalignedVolatileReadName, AttributeList: Attr, RetTy: IRB.getVoidTy(), Args: IRB.getPtrTy()); |
248 | |
249 | SmallString<64> UnalignedVolatileWriteName( |
250 | "__tsan_unaligned_volatile_write" + ByteSizeStr); |
251 | TsanUnalignedVolatileWrite[i] = M.getOrInsertFunction( |
252 | Name: UnalignedVolatileWriteName, AttributeList: Attr, RetTy: IRB.getVoidTy(), Args: IRB.getPtrTy()); |
253 | |
254 | SmallString<64> CompoundRWName("__tsan_read_write" + ByteSizeStr); |
255 | TsanCompoundRW[i] = M.getOrInsertFunction( |
256 | Name: CompoundRWName, AttributeList: Attr, RetTy: IRB.getVoidTy(), Args: IRB.getPtrTy()); |
257 | |
258 | SmallString<64> UnalignedCompoundRWName("__tsan_unaligned_read_write" + |
259 | ByteSizeStr); |
260 | TsanUnalignedCompoundRW[i] = M.getOrInsertFunction( |
261 | Name: UnalignedCompoundRWName, AttributeList: Attr, RetTy: IRB.getVoidTy(), Args: IRB.getPtrTy()); |
262 | |
263 | Type *Ty = Type::getIntNTy(C&: Ctx, N: BitSize); |
264 | Type *PtrTy = PointerType::get(C&: Ctx, AddressSpace: 0); |
265 | SmallString<32> AtomicLoadName("__tsan_atomic" + BitSizeStr + "_load" ); |
266 | TsanAtomicLoad[i] = |
267 | M.getOrInsertFunction(Name: AtomicLoadName, |
268 | AttributeList: TLI.getAttrList(C: &Ctx, ArgNos: {1}, /*Signed=*/true, |
269 | /*Ret=*/BitSize <= 32, AL: Attr), |
270 | RetTy: Ty, Args: PtrTy, Args: OrdTy); |
271 | |
272 | // Args of type Ty need extension only when BitSize is 32 or less. |
273 | using Idxs = std::vector<unsigned>; |
274 | Idxs Idxs2Or12 ((BitSize <= 32) ? Idxs({1, 2}) : Idxs({2})); |
275 | Idxs Idxs34Or1234((BitSize <= 32) ? Idxs({1, 2, 3, 4}) : Idxs({3, 4})); |
276 | SmallString<32> AtomicStoreName("__tsan_atomic" + BitSizeStr + "_store" ); |
277 | TsanAtomicStore[i] = M.getOrInsertFunction( |
278 | Name: AtomicStoreName, |
279 | AttributeList: TLI.getAttrList(C: &Ctx, ArgNos: Idxs2Or12, /*Signed=*/true, /*Ret=*/false, AL: Attr), |
280 | RetTy: IRB.getVoidTy(), Args: PtrTy, Args: Ty, Args: OrdTy); |
281 | |
282 | for (unsigned Op = AtomicRMWInst::FIRST_BINOP; |
283 | Op <= AtomicRMWInst::LAST_BINOP; ++Op) { |
284 | TsanAtomicRMW[Op][i] = nullptr; |
285 | const char *NamePart = nullptr; |
286 | if (Op == AtomicRMWInst::Xchg) |
287 | NamePart = "_exchange" ; |
288 | else if (Op == AtomicRMWInst::Add) |
289 | NamePart = "_fetch_add" ; |
290 | else if (Op == AtomicRMWInst::Sub) |
291 | NamePart = "_fetch_sub" ; |
292 | else if (Op == AtomicRMWInst::And) |
293 | NamePart = "_fetch_and" ; |
294 | else if (Op == AtomicRMWInst::Or) |
295 | NamePart = "_fetch_or" ; |
296 | else if (Op == AtomicRMWInst::Xor) |
297 | NamePart = "_fetch_xor" ; |
298 | else if (Op == AtomicRMWInst::Nand) |
299 | NamePart = "_fetch_nand" ; |
300 | else |
301 | continue; |
302 | SmallString<32> RMWName("__tsan_atomic" + itostr(X: BitSize) + NamePart); |
303 | TsanAtomicRMW[Op][i] = M.getOrInsertFunction( |
304 | Name: RMWName, |
305 | AttributeList: TLI.getAttrList(C: &Ctx, ArgNos: Idxs2Or12, /*Signed=*/true, |
306 | /*Ret=*/BitSize <= 32, AL: Attr), |
307 | RetTy: Ty, Args: PtrTy, Args: Ty, Args: OrdTy); |
308 | } |
309 | |
310 | SmallString<32> AtomicCASName("__tsan_atomic" + BitSizeStr + |
311 | "_compare_exchange_val" ); |
312 | TsanAtomicCAS[i] = M.getOrInsertFunction( |
313 | Name: AtomicCASName, |
314 | AttributeList: TLI.getAttrList(C: &Ctx, ArgNos: Idxs34Or1234, /*Signed=*/true, |
315 | /*Ret=*/BitSize <= 32, AL: Attr), |
316 | RetTy: Ty, Args: PtrTy, Args: Ty, Args: Ty, Args: OrdTy, Args: OrdTy); |
317 | } |
318 | TsanVptrUpdate = |
319 | M.getOrInsertFunction(Name: "__tsan_vptr_update" , AttributeList: Attr, RetTy: IRB.getVoidTy(), |
320 | Args: IRB.getPtrTy(), Args: IRB.getPtrTy()); |
321 | TsanVptrLoad = M.getOrInsertFunction(Name: "__tsan_vptr_read" , AttributeList: Attr, |
322 | RetTy: IRB.getVoidTy(), Args: IRB.getPtrTy()); |
323 | TsanAtomicThreadFence = M.getOrInsertFunction( |
324 | Name: "__tsan_atomic_thread_fence" , |
325 | AttributeList: TLI.getAttrList(C: &Ctx, ArgNos: {0}, /*Signed=*/true, /*Ret=*/false, AL: Attr), |
326 | RetTy: IRB.getVoidTy(), Args: OrdTy); |
327 | |
328 | TsanAtomicSignalFence = M.getOrInsertFunction( |
329 | Name: "__tsan_atomic_signal_fence" , |
330 | AttributeList: TLI.getAttrList(C: &Ctx, ArgNos: {0}, /*Signed=*/true, /*Ret=*/false, AL: Attr), |
331 | RetTy: IRB.getVoidTy(), Args: OrdTy); |
332 | |
333 | MemmoveFn = |
334 | M.getOrInsertFunction(Name: "__tsan_memmove" , AttributeList: Attr, RetTy: IRB.getPtrTy(), |
335 | Args: IRB.getPtrTy(), Args: IRB.getPtrTy(), Args: IntptrTy); |
336 | MemcpyFn = |
337 | M.getOrInsertFunction(Name: "__tsan_memcpy" , AttributeList: Attr, RetTy: IRB.getPtrTy(), |
338 | Args: IRB.getPtrTy(), Args: IRB.getPtrTy(), Args: IntptrTy); |
339 | MemsetFn = M.getOrInsertFunction( |
340 | Name: "__tsan_memset" , |
341 | AttributeList: TLI.getAttrList(C: &Ctx, ArgNos: {1}, /*Signed=*/true, /*Ret=*/false, AL: Attr), |
342 | RetTy: IRB.getPtrTy(), Args: IRB.getPtrTy(), Args: IRB.getInt32Ty(), Args: IntptrTy); |
343 | } |
344 | |
345 | static bool isVtableAccess(Instruction *I) { |
346 | if (MDNode *Tag = I->getMetadata(KindID: LLVMContext::MD_tbaa)) |
347 | return Tag->isTBAAVtableAccess(); |
348 | return false; |
349 | } |
350 | |
351 | // Do not instrument known races/"benign races" that come from compiler |
352 | // instrumentatin. The user has no way of suppressing them. |
353 | static bool shouldInstrumentReadWriteFromAddress(const Module *M, Value *Addr) { |
354 | // Peel off GEPs and BitCasts. |
355 | Addr = Addr->stripInBoundsOffsets(); |
356 | |
357 | if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Val: Addr)) { |
358 | if (GV->hasSection()) { |
359 | StringRef SectionName = GV->getSection(); |
360 | // Check if the global is in the PGO counters section. |
361 | auto OF = Triple(M->getTargetTriple()).getObjectFormat(); |
362 | if (SectionName.ends_with( |
363 | Suffix: getInstrProfSectionName(IPSK: IPSK_cnts, OF, /*AddSegmentInfo=*/false))) |
364 | return false; |
365 | } |
366 | } |
367 | |
368 | // Do not instrument accesses from different address spaces; we cannot deal |
369 | // with them. |
370 | if (Addr) { |
371 | Type *PtrTy = cast<PointerType>(Val: Addr->getType()->getScalarType()); |
372 | if (PtrTy->getPointerAddressSpace() != 0) |
373 | return false; |
374 | } |
375 | |
376 | return true; |
377 | } |
378 | |
379 | bool ThreadSanitizer::addrPointsToConstantData(Value *Addr) { |
380 | // If this is a GEP, just analyze its pointer operand. |
381 | if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Val: Addr)) |
382 | Addr = GEP->getPointerOperand(); |
383 | |
384 | if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Val: Addr)) { |
385 | if (GV->isConstant()) { |
386 | // Reads from constant globals can not race with any writes. |
387 | NumOmittedReadsFromConstantGlobals++; |
388 | return true; |
389 | } |
390 | } else if (LoadInst *L = dyn_cast<LoadInst>(Val: Addr)) { |
391 | if (isVtableAccess(I: L)) { |
392 | // Reads from a vtable pointer can not race with any writes. |
393 | NumOmittedReadsFromVtable++; |
394 | return true; |
395 | } |
396 | } |
397 | return false; |
398 | } |
399 | |
400 | // Instrumenting some of the accesses may be proven redundant. |
401 | // Currently handled: |
402 | // - read-before-write (within same BB, no calls between) |
403 | // - not captured variables |
404 | // |
405 | // We do not handle some of the patterns that should not survive |
406 | // after the classic compiler optimizations. |
407 | // E.g. two reads from the same temp should be eliminated by CSE, |
408 | // two writes should be eliminated by DSE, etc. |
409 | // |
410 | // 'Local' is a vector of insns within the same BB (no calls between). |
411 | // 'All' is a vector of insns that will be instrumented. |
412 | void ThreadSanitizer::chooseInstructionsToInstrument( |
413 | SmallVectorImpl<Instruction *> &Local, |
414 | SmallVectorImpl<InstructionInfo> &All, const DataLayout &DL) { |
415 | DenseMap<Value *, size_t> WriteTargets; // Map of addresses to index in All |
416 | // Iterate from the end. |
417 | for (Instruction *I : reverse(C&: Local)) { |
418 | const bool IsWrite = isa<StoreInst>(Val: *I); |
419 | Value *Addr = IsWrite ? cast<StoreInst>(Val: I)->getPointerOperand() |
420 | : cast<LoadInst>(Val: I)->getPointerOperand(); |
421 | |
422 | if (!shouldInstrumentReadWriteFromAddress(M: I->getModule(), Addr)) |
423 | continue; |
424 | |
425 | if (!IsWrite) { |
426 | const auto WriteEntry = WriteTargets.find(Val: Addr); |
427 | if (!ClInstrumentReadBeforeWrite && WriteEntry != WriteTargets.end()) { |
428 | auto &WI = All[WriteEntry->second]; |
429 | // If we distinguish volatile accesses and if either the read or write |
430 | // is volatile, do not omit any instrumentation. |
431 | const bool AnyVolatile = |
432 | ClDistinguishVolatile && (cast<LoadInst>(Val: I)->isVolatile() || |
433 | cast<StoreInst>(Val: WI.Inst)->isVolatile()); |
434 | if (!AnyVolatile) { |
435 | // We will write to this temp, so no reason to analyze the read. |
436 | // Mark the write instruction as compound. |
437 | WI.Flags |= InstructionInfo::kCompoundRW; |
438 | NumOmittedReadsBeforeWrite++; |
439 | continue; |
440 | } |
441 | } |
442 | |
443 | if (addrPointsToConstantData(Addr)) { |
444 | // Addr points to some constant data -- it can not race with any writes. |
445 | continue; |
446 | } |
447 | } |
448 | |
449 | if (isa<AllocaInst>(Val: getUnderlyingObject(V: Addr)) && |
450 | !PointerMayBeCaptured(V: Addr, ReturnCaptures: true, StoreCaptures: true)) { |
451 | // The variable is addressable but not captured, so it cannot be |
452 | // referenced from a different thread and participate in a data race |
453 | // (see llvm/Analysis/CaptureTracking.h for details). |
454 | NumOmittedNonCaptured++; |
455 | continue; |
456 | } |
457 | |
458 | // Instrument this instruction. |
459 | All.emplace_back(Args&: I); |
460 | if (IsWrite) { |
461 | // For read-before-write and compound instrumentation we only need one |
462 | // write target, and we can override any previous entry if it exists. |
463 | WriteTargets[Addr] = All.size() - 1; |
464 | } |
465 | } |
466 | Local.clear(); |
467 | } |
468 | |
469 | static bool isTsanAtomic(const Instruction *I) { |
470 | // TODO: Ask TTI whether synchronization scope is between threads. |
471 | auto SSID = getAtomicSyncScopeID(I); |
472 | if (!SSID) |
473 | return false; |
474 | if (isa<LoadInst>(Val: I) || isa<StoreInst>(Val: I)) |
475 | return *SSID != SyncScope::SingleThread; |
476 | return true; |
477 | } |
478 | |
479 | void ThreadSanitizer::InsertRuntimeIgnores(Function &F) { |
480 | InstrumentationIRBuilder IRB(F.getEntryBlock().getFirstNonPHI()); |
481 | IRB.CreateCall(Callee: TsanIgnoreBegin); |
482 | EscapeEnumerator EE(F, "tsan_ignore_cleanup" , ClHandleCxxExceptions); |
483 | while (IRBuilder<> *AtExit = EE.Next()) { |
484 | InstrumentationIRBuilder::ensureDebugInfo(IRB&: *AtExit, F); |
485 | AtExit->CreateCall(Callee: TsanIgnoreEnd); |
486 | } |
487 | } |
488 | |
489 | bool ThreadSanitizer::sanitizeFunction(Function &F, |
490 | const TargetLibraryInfo &TLI) { |
491 | // This is required to prevent instrumenting call to __tsan_init from within |
492 | // the module constructor. |
493 | if (F.getName() == kTsanModuleCtorName) |
494 | return false; |
495 | // Naked functions can not have prologue/epilogue |
496 | // (__tsan_func_entry/__tsan_func_exit) generated, so don't instrument them at |
497 | // all. |
498 | if (F.hasFnAttribute(Kind: Attribute::Naked)) |
499 | return false; |
500 | |
501 | // __attribute__(disable_sanitizer_instrumentation) prevents all kinds of |
502 | // instrumentation. |
503 | if (F.hasFnAttribute(Kind: Attribute::DisableSanitizerInstrumentation)) |
504 | return false; |
505 | |
506 | initialize(M&: *F.getParent(), TLI); |
507 | SmallVector<InstructionInfo, 8> AllLoadsAndStores; |
508 | SmallVector<Instruction*, 8> LocalLoadsAndStores; |
509 | SmallVector<Instruction*, 8> AtomicAccesses; |
510 | SmallVector<Instruction*, 8> MemIntrinCalls; |
511 | bool Res = false; |
512 | bool HasCalls = false; |
513 | bool SanitizeFunction = F.hasFnAttribute(Kind: Attribute::SanitizeThread); |
514 | const DataLayout &DL = F.getDataLayout(); |
515 | |
516 | // Traverse all instructions, collect loads/stores/returns, check for calls. |
517 | for (auto &BB : F) { |
518 | for (auto &Inst : BB) { |
519 | // Skip instructions inserted by another instrumentation. |
520 | if (Inst.hasMetadata(KindID: LLVMContext::MD_nosanitize)) |
521 | continue; |
522 | if (isTsanAtomic(I: &Inst)) |
523 | AtomicAccesses.push_back(Elt: &Inst); |
524 | else if (isa<LoadInst>(Val: Inst) || isa<StoreInst>(Val: Inst)) |
525 | LocalLoadsAndStores.push_back(Elt: &Inst); |
526 | else if ((isa<CallInst>(Val: Inst) && !isa<DbgInfoIntrinsic>(Val: Inst)) || |
527 | isa<InvokeInst>(Val: Inst)) { |
528 | if (CallInst *CI = dyn_cast<CallInst>(Val: &Inst)) |
529 | maybeMarkSanitizerLibraryCallNoBuiltin(CI, TLI: &TLI); |
530 | if (isa<MemIntrinsic>(Val: Inst)) |
531 | MemIntrinCalls.push_back(Elt: &Inst); |
532 | HasCalls = true; |
533 | chooseInstructionsToInstrument(Local&: LocalLoadsAndStores, All&: AllLoadsAndStores, |
534 | DL); |
535 | } |
536 | } |
537 | chooseInstructionsToInstrument(Local&: LocalLoadsAndStores, All&: AllLoadsAndStores, DL); |
538 | } |
539 | |
540 | // We have collected all loads and stores. |
541 | // FIXME: many of these accesses do not need to be checked for races |
542 | // (e.g. variables that do not escape, etc). |
543 | |
544 | // Instrument memory accesses only if we want to report bugs in the function. |
545 | if (ClInstrumentMemoryAccesses && SanitizeFunction) |
546 | for (const auto &II : AllLoadsAndStores) { |
547 | Res |= instrumentLoadOrStore(II, DL); |
548 | } |
549 | |
550 | // Instrument atomic memory accesses in any case (they can be used to |
551 | // implement synchronization). |
552 | if (ClInstrumentAtomics) |
553 | for (auto *Inst : AtomicAccesses) { |
554 | Res |= instrumentAtomic(I: Inst, DL); |
555 | } |
556 | |
557 | if (ClInstrumentMemIntrinsics && SanitizeFunction) |
558 | for (auto *Inst : MemIntrinCalls) { |
559 | Res |= instrumentMemIntrinsic(I: Inst); |
560 | } |
561 | |
562 | if (F.hasFnAttribute(Kind: "sanitize_thread_no_checking_at_run_time" )) { |
563 | assert(!F.hasFnAttribute(Attribute::SanitizeThread)); |
564 | if (HasCalls) |
565 | InsertRuntimeIgnores(F); |
566 | } |
567 | |
568 | // Instrument function entry/exit points if there were instrumented accesses. |
569 | if ((Res || HasCalls) && ClInstrumentFuncEntryExit) { |
570 | InstrumentationIRBuilder IRB(F.getEntryBlock().getFirstNonPHI()); |
571 | Value *ReturnAddress = IRB.CreateCall( |
572 | Callee: Intrinsic::getDeclaration(M: F.getParent(), id: Intrinsic::returnaddress), |
573 | Args: IRB.getInt32(C: 0)); |
574 | IRB.CreateCall(Callee: TsanFuncEntry, Args: ReturnAddress); |
575 | |
576 | EscapeEnumerator EE(F, "tsan_cleanup" , ClHandleCxxExceptions); |
577 | while (IRBuilder<> *AtExit = EE.Next()) { |
578 | InstrumentationIRBuilder::ensureDebugInfo(IRB&: *AtExit, F); |
579 | AtExit->CreateCall(Callee: TsanFuncExit, Args: {}); |
580 | } |
581 | Res = true; |
582 | } |
583 | return Res; |
584 | } |
585 | |
586 | bool ThreadSanitizer::instrumentLoadOrStore(const InstructionInfo &II, |
587 | const DataLayout &DL) { |
588 | InstrumentationIRBuilder IRB(II.Inst); |
589 | const bool IsWrite = isa<StoreInst>(Val: *II.Inst); |
590 | Value *Addr = IsWrite ? cast<StoreInst>(Val: II.Inst)->getPointerOperand() |
591 | : cast<LoadInst>(Val: II.Inst)->getPointerOperand(); |
592 | Type *OrigTy = getLoadStoreType(I: II.Inst); |
593 | |
594 | // swifterror memory addresses are mem2reg promoted by instruction selection. |
595 | // As such they cannot have regular uses like an instrumentation function and |
596 | // it makes no sense to track them as memory. |
597 | if (Addr->isSwiftError()) |
598 | return false; |
599 | |
600 | int Idx = getMemoryAccessFuncIndex(OrigTy, Addr, DL); |
601 | if (Idx < 0) |
602 | return false; |
603 | if (IsWrite && isVtableAccess(I: II.Inst)) { |
604 | LLVM_DEBUG(dbgs() << " VPTR : " << *II.Inst << "\n" ); |
605 | Value *StoredValue = cast<StoreInst>(Val: II.Inst)->getValueOperand(); |
606 | // StoredValue may be a vector type if we are storing several vptrs at once. |
607 | // In this case, just take the first element of the vector since this is |
608 | // enough to find vptr races. |
609 | if (isa<VectorType>(Val: StoredValue->getType())) |
610 | StoredValue = IRB.CreateExtractElement( |
611 | Vec: StoredValue, Idx: ConstantInt::get(Ty: IRB.getInt32Ty(), V: 0)); |
612 | if (StoredValue->getType()->isIntegerTy()) |
613 | StoredValue = IRB.CreateIntToPtr(V: StoredValue, DestTy: IRB.getPtrTy()); |
614 | // Call TsanVptrUpdate. |
615 | IRB.CreateCall(Callee: TsanVptrUpdate, Args: {Addr, StoredValue}); |
616 | NumInstrumentedVtableWrites++; |
617 | return true; |
618 | } |
619 | if (!IsWrite && isVtableAccess(I: II.Inst)) { |
620 | IRB.CreateCall(Callee: TsanVptrLoad, Args: Addr); |
621 | NumInstrumentedVtableReads++; |
622 | return true; |
623 | } |
624 | |
625 | const Align Alignment = IsWrite ? cast<StoreInst>(Val: II.Inst)->getAlign() |
626 | : cast<LoadInst>(Val: II.Inst)->getAlign(); |
627 | const bool IsCompoundRW = |
628 | ClCompoundReadBeforeWrite && (II.Flags & InstructionInfo::kCompoundRW); |
629 | const bool IsVolatile = ClDistinguishVolatile && |
630 | (IsWrite ? cast<StoreInst>(Val: II.Inst)->isVolatile() |
631 | : cast<LoadInst>(Val: II.Inst)->isVolatile()); |
632 | assert((!IsVolatile || !IsCompoundRW) && "Compound volatile invalid!" ); |
633 | |
634 | const uint32_t TypeSize = DL.getTypeStoreSizeInBits(Ty: OrigTy); |
635 | FunctionCallee OnAccessFunc = nullptr; |
636 | if (Alignment >= Align(8) || (Alignment.value() % (TypeSize / 8)) == 0) { |
637 | if (IsCompoundRW) |
638 | OnAccessFunc = TsanCompoundRW[Idx]; |
639 | else if (IsVolatile) |
640 | OnAccessFunc = IsWrite ? TsanVolatileWrite[Idx] : TsanVolatileRead[Idx]; |
641 | else |
642 | OnAccessFunc = IsWrite ? TsanWrite[Idx] : TsanRead[Idx]; |
643 | } else { |
644 | if (IsCompoundRW) |
645 | OnAccessFunc = TsanUnalignedCompoundRW[Idx]; |
646 | else if (IsVolatile) |
647 | OnAccessFunc = IsWrite ? TsanUnalignedVolatileWrite[Idx] |
648 | : TsanUnalignedVolatileRead[Idx]; |
649 | else |
650 | OnAccessFunc = IsWrite ? TsanUnalignedWrite[Idx] : TsanUnalignedRead[Idx]; |
651 | } |
652 | IRB.CreateCall(Callee: OnAccessFunc, Args: Addr); |
653 | if (IsCompoundRW || IsWrite) |
654 | NumInstrumentedWrites++; |
655 | if (IsCompoundRW || !IsWrite) |
656 | NumInstrumentedReads++; |
657 | return true; |
658 | } |
659 | |
660 | static ConstantInt *createOrdering(IRBuilder<> *IRB, AtomicOrdering ord) { |
661 | uint32_t v = 0; |
662 | switch (ord) { |
663 | case AtomicOrdering::NotAtomic: |
664 | llvm_unreachable("unexpected atomic ordering!" ); |
665 | case AtomicOrdering::Unordered: [[fallthrough]]; |
666 | case AtomicOrdering::Monotonic: v = 0; break; |
667 | // Not specified yet: |
668 | // case AtomicOrdering::Consume: v = 1; break; |
669 | case AtomicOrdering::Acquire: v = 2; break; |
670 | case AtomicOrdering::Release: v = 3; break; |
671 | case AtomicOrdering::AcquireRelease: v = 4; break; |
672 | case AtomicOrdering::SequentiallyConsistent: v = 5; break; |
673 | } |
674 | return IRB->getInt32(C: v); |
675 | } |
676 | |
677 | // If a memset intrinsic gets inlined by the code gen, we will miss races on it. |
678 | // So, we either need to ensure the intrinsic is not inlined, or instrument it. |
679 | // We do not instrument memset/memmove/memcpy intrinsics (too complicated), |
680 | // instead we simply replace them with regular function calls, which are then |
681 | // intercepted by the run-time. |
682 | // Since tsan is running after everyone else, the calls should not be |
683 | // replaced back with intrinsics. If that becomes wrong at some point, |
684 | // we will need to call e.g. __tsan_memset to avoid the intrinsics. |
685 | bool ThreadSanitizer::instrumentMemIntrinsic(Instruction *I) { |
686 | InstrumentationIRBuilder IRB(I); |
687 | if (MemSetInst *M = dyn_cast<MemSetInst>(Val: I)) { |
688 | Value *Cast1 = IRB.CreateIntCast(V: M->getArgOperand(i: 1), DestTy: IRB.getInt32Ty(), isSigned: false); |
689 | Value *Cast2 = IRB.CreateIntCast(V: M->getArgOperand(i: 2), DestTy: IntptrTy, isSigned: false); |
690 | IRB.CreateCall( |
691 | Callee: MemsetFn, |
692 | Args: {M->getArgOperand(i: 0), |
693 | Cast1, |
694 | Cast2}); |
695 | I->eraseFromParent(); |
696 | } else if (MemTransferInst *M = dyn_cast<MemTransferInst>(Val: I)) { |
697 | IRB.CreateCall( |
698 | Callee: isa<MemCpyInst>(Val: M) ? MemcpyFn : MemmoveFn, |
699 | Args: {M->getArgOperand(i: 0), |
700 | M->getArgOperand(i: 1), |
701 | IRB.CreateIntCast(V: M->getArgOperand(i: 2), DestTy: IntptrTy, isSigned: false)}); |
702 | I->eraseFromParent(); |
703 | } |
704 | return false; |
705 | } |
706 | |
707 | // Both llvm and ThreadSanitizer atomic operations are based on C++11/C1x |
708 | // standards. For background see C++11 standard. A slightly older, publicly |
709 | // available draft of the standard (not entirely up-to-date, but close enough |
710 | // for casual browsing) is available here: |
711 | // http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2011/n3242.pdf |
712 | // The following page contains more background information: |
713 | // http://www.hpl.hp.com/personal/Hans_Boehm/c++mm/ |
714 | |
715 | bool ThreadSanitizer::instrumentAtomic(Instruction *I, const DataLayout &DL) { |
716 | InstrumentationIRBuilder IRB(I); |
717 | if (LoadInst *LI = dyn_cast<LoadInst>(Val: I)) { |
718 | Value *Addr = LI->getPointerOperand(); |
719 | Type *OrigTy = LI->getType(); |
720 | int Idx = getMemoryAccessFuncIndex(OrigTy, Addr, DL); |
721 | if (Idx < 0) |
722 | return false; |
723 | Value *Args[] = {Addr, |
724 | createOrdering(IRB: &IRB, ord: LI->getOrdering())}; |
725 | Value *C = IRB.CreateCall(Callee: TsanAtomicLoad[Idx], Args); |
726 | Value *Cast = IRB.CreateBitOrPointerCast(V: C, DestTy: OrigTy); |
727 | I->replaceAllUsesWith(V: Cast); |
728 | } else if (StoreInst *SI = dyn_cast<StoreInst>(Val: I)) { |
729 | Value *Addr = SI->getPointerOperand(); |
730 | int Idx = |
731 | getMemoryAccessFuncIndex(OrigTy: SI->getValueOperand()->getType(), Addr, DL); |
732 | if (Idx < 0) |
733 | return false; |
734 | const unsigned ByteSize = 1U << Idx; |
735 | const unsigned BitSize = ByteSize * 8; |
736 | Type *Ty = Type::getIntNTy(C&: IRB.getContext(), N: BitSize); |
737 | Value *Args[] = {Addr, |
738 | IRB.CreateBitOrPointerCast(V: SI->getValueOperand(), DestTy: Ty), |
739 | createOrdering(IRB: &IRB, ord: SI->getOrdering())}; |
740 | IRB.CreateCall(Callee: TsanAtomicStore[Idx], Args); |
741 | SI->eraseFromParent(); |
742 | } else if (AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(Val: I)) { |
743 | Value *Addr = RMWI->getPointerOperand(); |
744 | int Idx = |
745 | getMemoryAccessFuncIndex(OrigTy: RMWI->getValOperand()->getType(), Addr, DL); |
746 | if (Idx < 0) |
747 | return false; |
748 | FunctionCallee F = TsanAtomicRMW[RMWI->getOperation()][Idx]; |
749 | if (!F) |
750 | return false; |
751 | const unsigned ByteSize = 1U << Idx; |
752 | const unsigned BitSize = ByteSize * 8; |
753 | Type *Ty = Type::getIntNTy(C&: IRB.getContext(), N: BitSize); |
754 | Value *Val = RMWI->getValOperand(); |
755 | Value *Args[] = {Addr, IRB.CreateBitOrPointerCast(V: Val, DestTy: Ty), |
756 | createOrdering(IRB: &IRB, ord: RMWI->getOrdering())}; |
757 | Value *C = IRB.CreateCall(Callee: F, Args); |
758 | I->replaceAllUsesWith(V: IRB.CreateBitOrPointerCast(V: C, DestTy: Val->getType())); |
759 | I->eraseFromParent(); |
760 | } else if (AtomicCmpXchgInst *CASI = dyn_cast<AtomicCmpXchgInst>(Val: I)) { |
761 | Value *Addr = CASI->getPointerOperand(); |
762 | Type *OrigOldValTy = CASI->getNewValOperand()->getType(); |
763 | int Idx = getMemoryAccessFuncIndex(OrigTy: OrigOldValTy, Addr, DL); |
764 | if (Idx < 0) |
765 | return false; |
766 | const unsigned ByteSize = 1U << Idx; |
767 | const unsigned BitSize = ByteSize * 8; |
768 | Type *Ty = Type::getIntNTy(C&: IRB.getContext(), N: BitSize); |
769 | Value *CmpOperand = |
770 | IRB.CreateBitOrPointerCast(V: CASI->getCompareOperand(), DestTy: Ty); |
771 | Value *NewOperand = |
772 | IRB.CreateBitOrPointerCast(V: CASI->getNewValOperand(), DestTy: Ty); |
773 | Value *Args[] = {Addr, |
774 | CmpOperand, |
775 | NewOperand, |
776 | createOrdering(IRB: &IRB, ord: CASI->getSuccessOrdering()), |
777 | createOrdering(IRB: &IRB, ord: CASI->getFailureOrdering())}; |
778 | CallInst *C = IRB.CreateCall(Callee: TsanAtomicCAS[Idx], Args); |
779 | Value *Success = IRB.CreateICmpEQ(LHS: C, RHS: CmpOperand); |
780 | Value *OldVal = C; |
781 | if (Ty != OrigOldValTy) { |
782 | // The value is a pointer, so we need to cast the return value. |
783 | OldVal = IRB.CreateIntToPtr(V: C, DestTy: OrigOldValTy); |
784 | } |
785 | |
786 | Value *Res = |
787 | IRB.CreateInsertValue(Agg: PoisonValue::get(T: CASI->getType()), Val: OldVal, Idxs: 0); |
788 | Res = IRB.CreateInsertValue(Agg: Res, Val: Success, Idxs: 1); |
789 | |
790 | I->replaceAllUsesWith(V: Res); |
791 | I->eraseFromParent(); |
792 | } else if (FenceInst *FI = dyn_cast<FenceInst>(Val: I)) { |
793 | Value *Args[] = {createOrdering(IRB: &IRB, ord: FI->getOrdering())}; |
794 | FunctionCallee F = FI->getSyncScopeID() == SyncScope::SingleThread |
795 | ? TsanAtomicSignalFence |
796 | : TsanAtomicThreadFence; |
797 | IRB.CreateCall(Callee: F, Args); |
798 | FI->eraseFromParent(); |
799 | } |
800 | return true; |
801 | } |
802 | |
803 | int ThreadSanitizer::getMemoryAccessFuncIndex(Type *OrigTy, Value *Addr, |
804 | const DataLayout &DL) { |
805 | assert(OrigTy->isSized()); |
806 | if (OrigTy->isScalableTy()) { |
807 | // FIXME: support vscale. |
808 | return -1; |
809 | } |
810 | uint32_t TypeSize = DL.getTypeStoreSizeInBits(Ty: OrigTy); |
811 | if (TypeSize != 8 && TypeSize != 16 && |
812 | TypeSize != 32 && TypeSize != 64 && TypeSize != 128) { |
813 | NumAccessesWithBadSize++; |
814 | // Ignore all unusual sizes. |
815 | return -1; |
816 | } |
817 | size_t Idx = llvm::countr_zero(Val: TypeSize / 8); |
818 | assert(Idx < kNumberOfAccessSizes); |
819 | return Idx; |
820 | } |
821 | |