1 //===-- MemorySanitizer.cpp - detector of uninitialized reads -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 /// This file is a part of MemorySanitizer, a detector of uninitialized
13 /// The algorithm of the tool is similar to Memcheck
14 /// (http://goo.gl/QKbem). We associate a few shadow bits with every
15 /// byte of the application memory, poison the shadow of the malloc-ed
16 /// or alloca-ed memory, load the shadow bits on every memory read,
17 /// propagate the shadow bits through some of the arithmetic
18 /// instruction (including MOV), store the shadow bits on every memory
19 /// write, report a bug on some other instructions (e.g. JMP) if the
20 /// associated shadow is poisoned.
22 /// But there are differences too. The first and the major one:
23 /// compiler instrumentation instead of binary instrumentation. This
24 /// gives us much better register allocation, possible compiler
25 /// optimizations and a fast start-up. But this brings the major issue
26 /// as well: msan needs to see all program events, including system
27 /// calls and reads/writes in system libraries, so we either need to
28 /// compile *everything* with msan or use a binary translation
29 /// component (e.g. DynamoRIO) to instrument pre-built libraries.
30 /// Another difference from Memcheck is that we use 8 shadow bits per
31 /// byte of application memory and use a direct shadow mapping. This
32 /// greatly simplifies the instrumentation code and avoids races on
33 /// shadow updates (Memcheck is single-threaded so races are not a
34 /// concern there. Memcheck uses 2 shadow bits per byte with a slow
35 /// path storage that uses 8 bits per byte).
37 /// The default value of shadow is 0, which means "clean" (not poisoned).
39 /// Every module initializer should call __msan_init to ensure that the
40 /// shadow memory is ready. On error, __msan_warning is called. Since
41 /// parameters and return values may be passed via registers, we have a
42 /// specialized thread-local shadow for return values
43 /// (__msan_retval_tls) and parameters (__msan_param_tls).
47 /// MemorySanitizer can track origins (allocation points) of all uninitialized
48 /// values. This behavior is controlled with a flag (msan-track-origins) and is
49 /// disabled by default.
51 /// Origins are 4-byte values created and interpreted by the runtime library.
52 /// They are stored in a second shadow mapping, one 4-byte value for 4 bytes
53 /// of application memory. Propagation of origins is basically a bunch of
54 /// "select" instructions that pick the origin of a dirty argument, if an
55 /// instruction has one.
57 /// Every 4 aligned, consecutive bytes of application memory have one origin
58 /// value associated with them. If these bytes contain uninitialized data
59 /// coming from 2 different allocations, the last store wins. Because of this,
60 /// MemorySanitizer reports can show unrelated origins, but this is unlikely in
63 /// Origins are meaningless for fully initialized values, so MemorySanitizer
64 /// avoids storing origin to memory when a fully initialized value is stored.
65 /// This way it avoids needless overwritting origin of the 4-byte region on
66 /// a short (i.e. 1 byte) clean store, and it is also good for performance.
70 /// Ideally, every atomic store of application value should update the
71 /// corresponding shadow location in an atomic way. Unfortunately, atomic store
72 /// of two disjoint locations can not be done without severe slowdown.
74 /// Therefore, we implement an approximation that may err on the safe side.
75 /// In this implementation, every atomically accessed location in the program
76 /// may only change from (partially) uninitialized to fully initialized, but
77 /// not the other way around. We load the shadow _after_ the application load,
78 /// and we store the shadow _before_ the app store. Also, we always store clean
79 /// shadow (if the application store is atomic). This way, if the store-load
80 /// pair constitutes a happens-before arc, shadow store and load are correctly
81 /// ordered such that the load will get either the value that was stored, or
82 /// some later value (which is always clean).
84 /// This does not work very well with Compare-And-Swap (CAS) and
85 /// Read-Modify-Write (RMW) operations. To follow the above logic, CAS and RMW
86 /// must store the new shadow before the app operation, and load the shadow
87 /// after the app operation. Computers don't work this way. Current
88 /// implementation ignores the load aspect of CAS/RMW, always returning a clean
89 /// value. It implements the store part as a simple atomic store by storing a
92 //===----------------------------------------------------------------------===//
94 #include "llvm/ADT/DepthFirstIterator.h"
95 #include "llvm/ADT/SmallString.h"
96 #include "llvm/ADT/SmallVector.h"
97 #include "llvm/ADT/StringExtras.h"
98 #include "llvm/ADT/Triple.h"
99 #include "llvm/IR/DataLayout.h"
100 #include "llvm/IR/Function.h"
101 #include "llvm/IR/IRBuilder.h"
102 #include "llvm/IR/InlineAsm.h"
103 #include "llvm/IR/InstVisitor.h"
104 #include "llvm/IR/IntrinsicInst.h"
105 #include "llvm/IR/LLVMContext.h"
106 #include "llvm/IR/MDBuilder.h"
107 #include "llvm/IR/Module.h"
108 #include "llvm/IR/Type.h"
109 #include "llvm/IR/ValueMap.h"
110 #include "llvm/Support/CommandLine.h"
111 #include "llvm/Support/Debug.h"
112 #include "llvm/Support/raw_ostream.h"
113 #include "llvm/Transforms/Instrumentation.h"
114 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
115 #include "llvm/Transforms/Utils/Local.h"
116 #include "llvm/Transforms/Utils/ModuleUtils.h"
118 using namespace llvm;
120 #define DEBUG_TYPE "msan"
122 static const unsigned kOriginSize = 4;
123 static const unsigned kMinOriginAlignment = 4;
124 static const unsigned kShadowTLSAlignment = 8;
126 // These constants must be kept in sync with the ones in msan.h.
127 static const unsigned kParamTLSSize = 800;
128 static const unsigned kRetvalTLSSize = 800;
130 // Accesses sizes are powers of two: 1, 2, 4, 8.
131 static const size_t kNumberOfAccessSizes = 4;
133 /// \brief Track origins of uninitialized values.
135 /// Adds a section to MemorySanitizer report that points to the allocation
136 /// (stack or heap) the uninitialized bits came from originally.
137 static cl::opt<int> ClTrackOrigins("msan-track-origins",
138 cl::desc("Track origins (allocation sites) of poisoned memory"),
139 cl::Hidden, cl::init(0));
140 static cl::opt<bool> ClKeepGoing("msan-keep-going",
141 cl::desc("keep going after reporting a UMR"),
142 cl::Hidden, cl::init(false));
143 static cl::opt<bool> ClPoisonStack("msan-poison-stack",
144 cl::desc("poison uninitialized stack variables"),
145 cl::Hidden, cl::init(true));
146 static cl::opt<bool> ClPoisonStackWithCall("msan-poison-stack-with-call",
147 cl::desc("poison uninitialized stack variables with a call"),
148 cl::Hidden, cl::init(false));
149 static cl::opt<int> ClPoisonStackPattern("msan-poison-stack-pattern",
150 cl::desc("poison uninitialized stack variables with the given pattern"),
151 cl::Hidden, cl::init(0xff));
152 static cl::opt<bool> ClPoisonUndef("msan-poison-undef",
153 cl::desc("poison undef temps"),
154 cl::Hidden, cl::init(true));
156 static cl::opt<bool> ClHandleICmp("msan-handle-icmp",
157 cl::desc("propagate shadow through ICmpEQ and ICmpNE"),
158 cl::Hidden, cl::init(true));
160 static cl::opt<bool> ClHandleICmpExact("msan-handle-icmp-exact",
161 cl::desc("exact handling of relational integer ICmp"),
162 cl::Hidden, cl::init(false));
164 // This flag controls whether we check the shadow of the address
165 // operand of load or store. Such bugs are very rare, since load from
166 // a garbage address typically results in SEGV, but still happen
167 // (e.g. only lower bits of address are garbage, or the access happens
168 // early at program startup where malloc-ed memory is more likely to
169 // be zeroed. As of 2012-08-28 this flag adds 20% slowdown.
170 static cl::opt<bool> ClCheckAccessAddress("msan-check-access-address",
171 cl::desc("report accesses through a pointer which has poisoned shadow"),
172 cl::Hidden, cl::init(true));
174 static cl::opt<bool> ClDumpStrictInstructions("msan-dump-strict-instructions",
175 cl::desc("print out instructions with default strict semantics"),
176 cl::Hidden, cl::init(false));
178 static cl::opt<int> ClInstrumentationWithCallThreshold(
179 "msan-instrumentation-with-call-threshold",
181 "If the function being instrumented requires more than "
182 "this number of checks and origin stores, use callbacks instead of "
183 "inline checks (-1 means never use callbacks)."),
184 cl::Hidden, cl::init(3500));
186 // This is an experiment to enable handling of cases where shadow is a non-zero
187 // compile-time constant. For some unexplainable reason they were silently
188 // ignored in the instrumentation.
189 static cl::opt<bool> ClCheckConstantShadow("msan-check-constant-shadow",
190 cl::desc("Insert checks for constant shadow values"),
191 cl::Hidden, cl::init(false));
193 // This is off by default because of a bug in gold:
194 // https://sourceware.org/bugzilla/show_bug.cgi?id=19002
195 static cl::opt<bool> ClWithComdat("msan-with-comdat",
196 cl::desc("Place MSan constructors in comdat sections"),
197 cl::Hidden, cl::init(false));
199 static const char *const kMsanModuleCtorName = "msan.module_ctor";
200 static const char *const kMsanInitName = "__msan_init";
204 // Memory map parameters used in application-to-shadow address calculation.
205 // Offset = (Addr & ~AndMask) ^ XorMask
206 // Shadow = ShadowBase + Offset
207 // Origin = OriginBase + Offset
208 struct MemoryMapParams {
215 struct PlatformMemoryMapParams {
216 const MemoryMapParams *bits32;
217 const MemoryMapParams *bits64;
221 static const MemoryMapParams Linux_I386_MemoryMapParams = {
222 0x000080000000, // AndMask
223 0, // XorMask (not used)
224 0, // ShadowBase (not used)
225 0x000040000000, // OriginBase
229 static const MemoryMapParams Linux_X86_64_MemoryMapParams = {
230 #ifdef MSAN_LINUX_X86_64_OLD_MAPPING
231 0x400000000000, // AndMask
232 0, // XorMask (not used)
233 0, // ShadowBase (not used)
234 0x200000000000, // OriginBase
236 0, // AndMask (not used)
237 0x500000000000, // XorMask
238 0, // ShadowBase (not used)
239 0x100000000000, // OriginBase
244 static const MemoryMapParams Linux_MIPS64_MemoryMapParams = {
245 0, // AndMask (not used)
246 0x008000000000, // XorMask
247 0, // ShadowBase (not used)
248 0x002000000000, // OriginBase
252 static const MemoryMapParams Linux_PowerPC64_MemoryMapParams = {
253 0x200000000000, // AndMask
254 0x100000000000, // XorMask
255 0x080000000000, // ShadowBase
256 0x1C0000000000, // OriginBase
260 static const MemoryMapParams Linux_AArch64_MemoryMapParams = {
261 0, // AndMask (not used)
262 0x06000000000, // XorMask
263 0, // ShadowBase (not used)
264 0x01000000000, // OriginBase
268 static const MemoryMapParams FreeBSD_I386_MemoryMapParams = {
269 0x000180000000, // AndMask
270 0x000040000000, // XorMask
271 0x000020000000, // ShadowBase
272 0x000700000000, // OriginBase
276 static const MemoryMapParams FreeBSD_X86_64_MemoryMapParams = {
277 0xc00000000000, // AndMask
278 0x200000000000, // XorMask
279 0x100000000000, // ShadowBase
280 0x380000000000, // OriginBase
283 static const PlatformMemoryMapParams Linux_X86_MemoryMapParams = {
284 &Linux_I386_MemoryMapParams,
285 &Linux_X86_64_MemoryMapParams,
288 static const PlatformMemoryMapParams Linux_MIPS_MemoryMapParams = {
290 &Linux_MIPS64_MemoryMapParams,
293 static const PlatformMemoryMapParams Linux_PowerPC_MemoryMapParams = {
295 &Linux_PowerPC64_MemoryMapParams,
298 static const PlatformMemoryMapParams Linux_ARM_MemoryMapParams = {
300 &Linux_AArch64_MemoryMapParams,
303 static const PlatformMemoryMapParams FreeBSD_X86_MemoryMapParams = {
304 &FreeBSD_I386_MemoryMapParams,
305 &FreeBSD_X86_64_MemoryMapParams,
308 /// \brief An instrumentation pass implementing detection of uninitialized
311 /// MemorySanitizer: instrument the code in module to find
312 /// uninitialized reads.
313 class MemorySanitizer : public FunctionPass {
315 MemorySanitizer(int TrackOrigins = 0, bool Recover = false)
317 TrackOrigins(std::max(TrackOrigins, (int)ClTrackOrigins)),
318 Recover(Recover || ClKeepGoing),
319 WarningFn(nullptr) {}
320 StringRef getPassName() const override { return "MemorySanitizer"; }
321 void getAnalysisUsage(AnalysisUsage &AU) const override {
322 AU.addRequired<TargetLibraryInfoWrapperPass>();
324 bool runOnFunction(Function &F) override;
325 bool doInitialization(Module &M) override;
326 static char ID; // Pass identification, replacement for typeid.
329 void initializeCallbacks(Module &M);
331 /// \brief Track origins (allocation points) of uninitialized values.
338 /// \brief Thread-local shadow storage for function parameters.
339 GlobalVariable *ParamTLS;
340 /// \brief Thread-local origin storage for function parameters.
341 GlobalVariable *ParamOriginTLS;
342 /// \brief Thread-local shadow storage for function return value.
343 GlobalVariable *RetvalTLS;
344 /// \brief Thread-local origin storage for function return value.
345 GlobalVariable *RetvalOriginTLS;
346 /// \brief Thread-local shadow storage for in-register va_arg function
347 /// parameters (x86_64-specific).
348 GlobalVariable *VAArgTLS;
349 /// \brief Thread-local shadow storage for va_arg overflow area
350 /// (x86_64-specific).
351 GlobalVariable *VAArgOverflowSizeTLS;
352 /// \brief Thread-local space used to pass origin value to the UMR reporting
354 GlobalVariable *OriginTLS;
356 /// \brief The run-time callback to print a warning.
358 // These arrays are indexed by log2(AccessSize).
359 Value *MaybeWarningFn[kNumberOfAccessSizes];
360 Value *MaybeStoreOriginFn[kNumberOfAccessSizes];
362 /// \brief Run-time helper that generates a new origin value for a stack
364 Value *MsanSetAllocaOrigin4Fn;
365 /// \brief Run-time helper that poisons stack on function entry.
366 Value *MsanPoisonStackFn;
367 /// \brief Run-time helper that records a store (or any event) of an
368 /// uninitialized value and returns an updated origin id encoding this info.
369 Value *MsanChainOriginFn;
370 /// \brief MSan runtime replacements for memmove, memcpy and memset.
371 Value *MemmoveFn, *MemcpyFn, *MemsetFn;
373 /// \brief Memory map parameters used in application-to-shadow calculation.
374 const MemoryMapParams *MapParams;
376 MDNode *ColdCallWeights;
377 /// \brief Branch weights for origin store.
378 MDNode *OriginStoreWeights;
379 /// \brief An empty volatile inline asm that prevents callback merge.
381 Function *MsanCtorFunction;
383 friend struct MemorySanitizerVisitor;
384 friend struct VarArgAMD64Helper;
385 friend struct VarArgMIPS64Helper;
386 friend struct VarArgAArch64Helper;
387 friend struct VarArgPowerPC64Helper;
389 } // anonymous namespace
391 char MemorySanitizer::ID = 0;
392 INITIALIZE_PASS_BEGIN(
393 MemorySanitizer, "msan",
394 "MemorySanitizer: detects uninitialized reads.", false, false)
395 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
397 MemorySanitizer, "msan",
398 "MemorySanitizer: detects uninitialized reads.", false, false)
400 FunctionPass *llvm::createMemorySanitizerPass(int TrackOrigins, bool Recover) {
401 return new MemorySanitizer(TrackOrigins, Recover);
404 /// \brief Create a non-const global initialized with the given string.
406 /// Creates a writable global for Str so that we can pass it to the
407 /// run-time lib. Runtime uses first 4 bytes of the string to store the
408 /// frame ID, so the string needs to be mutable.
409 static GlobalVariable *createPrivateNonConstGlobalForString(Module &M,
411 Constant *StrConst = ConstantDataArray::getString(M.getContext(), Str);
412 return new GlobalVariable(M, StrConst->getType(), /*isConstant=*/false,
413 GlobalValue::PrivateLinkage, StrConst, "");
416 /// \brief Insert extern declaration of runtime-provided functions and globals.
417 void MemorySanitizer::initializeCallbacks(Module &M) {
418 // Only do this once.
423 // Create the callback.
424 // FIXME: this function should have "Cold" calling conv,
425 // which is not yet implemented.
426 StringRef WarningFnName = Recover ? "__msan_warning"
427 : "__msan_warning_noreturn";
428 WarningFn = M.getOrInsertFunction(WarningFnName, IRB.getVoidTy());
430 for (size_t AccessSizeIndex = 0; AccessSizeIndex < kNumberOfAccessSizes;
432 unsigned AccessSize = 1 << AccessSizeIndex;
433 std::string FunctionName = "__msan_maybe_warning_" + itostr(AccessSize);
434 MaybeWarningFn[AccessSizeIndex] = M.getOrInsertFunction(
435 FunctionName, IRB.getVoidTy(), IRB.getIntNTy(AccessSize * 8),
438 FunctionName = "__msan_maybe_store_origin_" + itostr(AccessSize);
439 MaybeStoreOriginFn[AccessSizeIndex] = M.getOrInsertFunction(
440 FunctionName, IRB.getVoidTy(), IRB.getIntNTy(AccessSize * 8),
441 IRB.getInt8PtrTy(), IRB.getInt32Ty());
444 MsanSetAllocaOrigin4Fn = M.getOrInsertFunction(
445 "__msan_set_alloca_origin4", IRB.getVoidTy(), IRB.getInt8PtrTy(), IntptrTy,
446 IRB.getInt8PtrTy(), IntptrTy);
448 M.getOrInsertFunction("__msan_poison_stack", IRB.getVoidTy(),
449 IRB.getInt8PtrTy(), IntptrTy);
450 MsanChainOriginFn = M.getOrInsertFunction(
451 "__msan_chain_origin", IRB.getInt32Ty(), IRB.getInt32Ty());
452 MemmoveFn = M.getOrInsertFunction(
453 "__msan_memmove", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
454 IRB.getInt8PtrTy(), IntptrTy);
455 MemcpyFn = M.getOrInsertFunction(
456 "__msan_memcpy", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
458 MemsetFn = M.getOrInsertFunction(
459 "__msan_memset", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt32Ty(),
463 RetvalTLS = new GlobalVariable(
464 M, ArrayType::get(IRB.getInt64Ty(), kRetvalTLSSize / 8), false,
465 GlobalVariable::ExternalLinkage, nullptr, "__msan_retval_tls", nullptr,
466 GlobalVariable::InitialExecTLSModel);
467 RetvalOriginTLS = new GlobalVariable(
468 M, OriginTy, false, GlobalVariable::ExternalLinkage, nullptr,
469 "__msan_retval_origin_tls", nullptr, GlobalVariable::InitialExecTLSModel);
471 ParamTLS = new GlobalVariable(
472 M, ArrayType::get(IRB.getInt64Ty(), kParamTLSSize / 8), false,
473 GlobalVariable::ExternalLinkage, nullptr, "__msan_param_tls", nullptr,
474 GlobalVariable::InitialExecTLSModel);
475 ParamOriginTLS = new GlobalVariable(
476 M, ArrayType::get(OriginTy, kParamTLSSize / 4), false,
477 GlobalVariable::ExternalLinkage, nullptr, "__msan_param_origin_tls",
478 nullptr, GlobalVariable::InitialExecTLSModel);
480 VAArgTLS = new GlobalVariable(
481 M, ArrayType::get(IRB.getInt64Ty(), kParamTLSSize / 8), false,
482 GlobalVariable::ExternalLinkage, nullptr, "__msan_va_arg_tls", nullptr,
483 GlobalVariable::InitialExecTLSModel);
484 VAArgOverflowSizeTLS = new GlobalVariable(
485 M, IRB.getInt64Ty(), false, GlobalVariable::ExternalLinkage, nullptr,
486 "__msan_va_arg_overflow_size_tls", nullptr,
487 GlobalVariable::InitialExecTLSModel);
488 OriginTLS = new GlobalVariable(
489 M, IRB.getInt32Ty(), false, GlobalVariable::ExternalLinkage, nullptr,
490 "__msan_origin_tls", nullptr, GlobalVariable::InitialExecTLSModel);
492 // We insert an empty inline asm after __msan_report* to avoid callback merge.
493 EmptyAsm = InlineAsm::get(FunctionType::get(IRB.getVoidTy(), false),
494 StringRef(""), StringRef(""),
495 /*hasSideEffects=*/true);
498 /// \brief Module-level initialization.
500 /// inserts a call to __msan_init to the module's constructor list.
501 bool MemorySanitizer::doInitialization(Module &M) {
502 auto &DL = M.getDataLayout();
504 Triple TargetTriple(M.getTargetTriple());
505 switch (TargetTriple.getOS()) {
506 case Triple::FreeBSD:
507 switch (TargetTriple.getArch()) {
509 MapParams = FreeBSD_X86_MemoryMapParams.bits64;
512 MapParams = FreeBSD_X86_MemoryMapParams.bits32;
515 report_fatal_error("unsupported architecture");
519 switch (TargetTriple.getArch()) {
521 MapParams = Linux_X86_MemoryMapParams.bits64;
524 MapParams = Linux_X86_MemoryMapParams.bits32;
527 case Triple::mips64el:
528 MapParams = Linux_MIPS_MemoryMapParams.bits64;
531 case Triple::ppc64le:
532 MapParams = Linux_PowerPC_MemoryMapParams.bits64;
534 case Triple::aarch64:
535 case Triple::aarch64_be:
536 MapParams = Linux_ARM_MemoryMapParams.bits64;
539 report_fatal_error("unsupported architecture");
543 report_fatal_error("unsupported operating system");
546 C = &(M.getContext());
548 IntptrTy = IRB.getIntPtrTy(DL);
549 OriginTy = IRB.getInt32Ty();
551 ColdCallWeights = MDBuilder(*C).createBranchWeights(1, 1000);
552 OriginStoreWeights = MDBuilder(*C).createBranchWeights(1, 1000);
554 std::tie(MsanCtorFunction, std::ignore) =
555 createSanitizerCtorAndInitFunctions(M, kMsanModuleCtorName, kMsanInitName,
559 Comdat *MsanCtorComdat = M.getOrInsertComdat(kMsanModuleCtorName);
560 MsanCtorFunction->setComdat(MsanCtorComdat);
561 appendToGlobalCtors(M, MsanCtorFunction, 0, MsanCtorFunction);
563 appendToGlobalCtors(M, MsanCtorFunction, 0);
568 new GlobalVariable(M, IRB.getInt32Ty(), true, GlobalValue::WeakODRLinkage,
569 IRB.getInt32(TrackOrigins), "__msan_track_origins");
572 new GlobalVariable(M, IRB.getInt32Ty(), true, GlobalValue::WeakODRLinkage,
573 IRB.getInt32(Recover), "__msan_keep_going");
580 /// \brief A helper class that handles instrumentation of VarArg
581 /// functions on a particular platform.
583 /// Implementations are expected to insert the instrumentation
584 /// necessary to propagate argument shadow through VarArg function
585 /// calls. Visit* methods are called during an InstVisitor pass over
586 /// the function, and should avoid creating new basic blocks. A new
587 /// instance of this class is created for each instrumented function.
588 struct VarArgHelper {
589 /// \brief Visit a CallSite.
590 virtual void visitCallSite(CallSite &CS, IRBuilder<> &IRB) = 0;
592 /// \brief Visit a va_start call.
593 virtual void visitVAStartInst(VAStartInst &I) = 0;
595 /// \brief Visit a va_copy call.
596 virtual void visitVACopyInst(VACopyInst &I) = 0;
598 /// \brief Finalize function instrumentation.
600 /// This method is called after visiting all interesting (see above)
601 /// instructions in a function.
602 virtual void finalizeInstrumentation() = 0;
604 virtual ~VarArgHelper() {}
607 struct MemorySanitizerVisitor;
610 CreateVarArgHelper(Function &Func, MemorySanitizer &Msan,
611 MemorySanitizerVisitor &Visitor);
613 unsigned TypeSizeToSizeIndex(unsigned TypeSize) {
614 if (TypeSize <= 8) return 0;
615 return Log2_32_Ceil((TypeSize + 7) / 8);
618 /// This class does all the work for a given function. Store and Load
619 /// instructions store and load corresponding shadow and origin
620 /// values. Most instructions propagate shadow from arguments to their
621 /// return values. Certain instructions (most importantly, BranchInst)
622 /// test their argument shadow and print reports (with a runtime call) if it's
624 struct MemorySanitizerVisitor : public InstVisitor<MemorySanitizerVisitor> {
627 SmallVector<PHINode *, 16> ShadowPHINodes, OriginPHINodes;
628 ValueMap<Value*, Value*> ShadowMap, OriginMap;
629 std::unique_ptr<VarArgHelper> VAHelper;
630 const TargetLibraryInfo *TLI;
632 // The following flags disable parts of MSan instrumentation based on
633 // blacklist contents and command-line options.
635 bool PropagateShadow;
638 bool CheckReturnValue;
640 struct ShadowOriginAndInsertPoint {
643 Instruction *OrigIns;
644 ShadowOriginAndInsertPoint(Value *S, Value *O, Instruction *I)
645 : Shadow(S), Origin(O), OrigIns(I) { }
647 SmallVector<ShadowOriginAndInsertPoint, 16> InstrumentationList;
648 SmallVector<StoreInst *, 16> StoreList;
650 MemorySanitizerVisitor(Function &F, MemorySanitizer &MS)
651 : F(F), MS(MS), VAHelper(CreateVarArgHelper(F, MS, *this)) {
652 bool SanitizeFunction = F.hasFnAttribute(Attribute::SanitizeMemory);
653 InsertChecks = SanitizeFunction;
654 PropagateShadow = SanitizeFunction;
655 PoisonStack = SanitizeFunction && ClPoisonStack;
656 PoisonUndef = SanitizeFunction && ClPoisonUndef;
657 // FIXME: Consider using SpecialCaseList to specify a list of functions that
658 // must always return fully initialized values. For now, we hardcode "main".
659 CheckReturnValue = SanitizeFunction && (F.getName() == "main");
660 TLI = &MS.getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
662 DEBUG(if (!InsertChecks)
663 dbgs() << "MemorySanitizer is not inserting checks into '"
664 << F.getName() << "'\n");
667 Value *updateOrigin(Value *V, IRBuilder<> &IRB) {
668 if (MS.TrackOrigins <= 1) return V;
669 return IRB.CreateCall(MS.MsanChainOriginFn, V);
672 Value *originToIntptr(IRBuilder<> &IRB, Value *Origin) {
673 const DataLayout &DL = F.getParent()->getDataLayout();
674 unsigned IntptrSize = DL.getTypeStoreSize(MS.IntptrTy);
675 if (IntptrSize == kOriginSize) return Origin;
676 assert(IntptrSize == kOriginSize * 2);
677 Origin = IRB.CreateIntCast(Origin, MS.IntptrTy, /* isSigned */ false);
678 return IRB.CreateOr(Origin, IRB.CreateShl(Origin, kOriginSize * 8));
681 /// \brief Fill memory range with the given origin value.
682 void paintOrigin(IRBuilder<> &IRB, Value *Origin, Value *OriginPtr,
683 unsigned Size, unsigned Alignment) {
684 const DataLayout &DL = F.getParent()->getDataLayout();
685 unsigned IntptrAlignment = DL.getABITypeAlignment(MS.IntptrTy);
686 unsigned IntptrSize = DL.getTypeStoreSize(MS.IntptrTy);
687 assert(IntptrAlignment >= kMinOriginAlignment);
688 assert(IntptrSize >= kOriginSize);
691 unsigned CurrentAlignment = Alignment;
692 if (Alignment >= IntptrAlignment && IntptrSize > kOriginSize) {
693 Value *IntptrOrigin = originToIntptr(IRB, Origin);
694 Value *IntptrOriginPtr =
695 IRB.CreatePointerCast(OriginPtr, PointerType::get(MS.IntptrTy, 0));
696 for (unsigned i = 0; i < Size / IntptrSize; ++i) {
697 Value *Ptr = i ? IRB.CreateConstGEP1_32(MS.IntptrTy, IntptrOriginPtr, i)
699 IRB.CreateAlignedStore(IntptrOrigin, Ptr, CurrentAlignment);
700 Ofs += IntptrSize / kOriginSize;
701 CurrentAlignment = IntptrAlignment;
705 for (unsigned i = Ofs; i < (Size + kOriginSize - 1) / kOriginSize; ++i) {
707 i ? IRB.CreateConstGEP1_32(nullptr, OriginPtr, i) : OriginPtr;
708 IRB.CreateAlignedStore(Origin, GEP, CurrentAlignment);
709 CurrentAlignment = kMinOriginAlignment;
713 void storeOrigin(IRBuilder<> &IRB, Value *Addr, Value *Shadow, Value *Origin,
714 unsigned Alignment, bool AsCall) {
715 const DataLayout &DL = F.getParent()->getDataLayout();
716 unsigned OriginAlignment = std::max(kMinOriginAlignment, Alignment);
717 unsigned StoreSize = DL.getTypeStoreSize(Shadow->getType());
718 if (Shadow->getType()->isAggregateType()) {
719 paintOrigin(IRB, updateOrigin(Origin, IRB),
720 getOriginPtr(Addr, IRB, Alignment), StoreSize,
723 Value *ConvertedShadow = convertToShadowTyNoVec(Shadow, IRB);
724 Constant *ConstantShadow = dyn_cast_or_null<Constant>(ConvertedShadow);
725 if (ConstantShadow) {
726 if (ClCheckConstantShadow && !ConstantShadow->isZeroValue())
727 paintOrigin(IRB, updateOrigin(Origin, IRB),
728 getOriginPtr(Addr, IRB, Alignment), StoreSize,
733 unsigned TypeSizeInBits =
734 DL.getTypeSizeInBits(ConvertedShadow->getType());
735 unsigned SizeIndex = TypeSizeToSizeIndex(TypeSizeInBits);
736 if (AsCall && SizeIndex < kNumberOfAccessSizes) {
737 Value *Fn = MS.MaybeStoreOriginFn[SizeIndex];
738 Value *ConvertedShadow2 = IRB.CreateZExt(
739 ConvertedShadow, IRB.getIntNTy(8 * (1 << SizeIndex)));
740 IRB.CreateCall(Fn, {ConvertedShadow2,
741 IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()),
744 Value *Cmp = IRB.CreateICmpNE(
745 ConvertedShadow, getCleanShadow(ConvertedShadow), "_mscmp");
746 Instruction *CheckTerm = SplitBlockAndInsertIfThen(
747 Cmp, &*IRB.GetInsertPoint(), false, MS.OriginStoreWeights);
748 IRBuilder<> IRBNew(CheckTerm);
749 paintOrigin(IRBNew, updateOrigin(Origin, IRBNew),
750 getOriginPtr(Addr, IRBNew, Alignment), StoreSize,
756 void materializeStores(bool InstrumentWithCalls) {
757 for (StoreInst *SI : StoreList) {
759 Value *Val = SI->getValueOperand();
760 Value *Addr = SI->getPointerOperand();
761 Value *Shadow = SI->isAtomic() ? getCleanShadow(Val) : getShadow(Val);
762 Value *ShadowPtr = getShadowPtr(Addr, Shadow->getType(), IRB);
765 IRB.CreateAlignedStore(Shadow, ShadowPtr, SI->getAlignment());
766 DEBUG(dbgs() << " STORE: " << *NewSI << "\n");
769 if (ClCheckAccessAddress)
770 insertShadowCheck(Addr, SI);
773 SI->setOrdering(addReleaseOrdering(SI->getOrdering()));
775 if (MS.TrackOrigins && !SI->isAtomic())
776 storeOrigin(IRB, Addr, Shadow, getOrigin(Val), SI->getAlignment(),
777 InstrumentWithCalls);
781 void materializeOneCheck(Instruction *OrigIns, Value *Shadow, Value *Origin,
783 IRBuilder<> IRB(OrigIns);
784 DEBUG(dbgs() << " SHAD0 : " << *Shadow << "\n");
785 Value *ConvertedShadow = convertToShadowTyNoVec(Shadow, IRB);
786 DEBUG(dbgs() << " SHAD1 : " << *ConvertedShadow << "\n");
788 Constant *ConstantShadow = dyn_cast_or_null<Constant>(ConvertedShadow);
789 if (ConstantShadow) {
790 if (ClCheckConstantShadow && !ConstantShadow->isZeroValue()) {
791 if (MS.TrackOrigins) {
792 IRB.CreateStore(Origin ? (Value *)Origin : (Value *)IRB.getInt32(0),
795 IRB.CreateCall(MS.WarningFn, {});
796 IRB.CreateCall(MS.EmptyAsm, {});
797 // FIXME: Insert UnreachableInst if !MS.Recover?
798 // This may invalidate some of the following checks and needs to be done
804 const DataLayout &DL = OrigIns->getModule()->getDataLayout();
806 unsigned TypeSizeInBits = DL.getTypeSizeInBits(ConvertedShadow->getType());
807 unsigned SizeIndex = TypeSizeToSizeIndex(TypeSizeInBits);
808 if (AsCall && SizeIndex < kNumberOfAccessSizes) {
809 Value *Fn = MS.MaybeWarningFn[SizeIndex];
810 Value *ConvertedShadow2 =
811 IRB.CreateZExt(ConvertedShadow, IRB.getIntNTy(8 * (1 << SizeIndex)));
812 IRB.CreateCall(Fn, {ConvertedShadow2, MS.TrackOrigins && Origin
814 : (Value *)IRB.getInt32(0)});
816 Value *Cmp = IRB.CreateICmpNE(ConvertedShadow,
817 getCleanShadow(ConvertedShadow), "_mscmp");
818 Instruction *CheckTerm = SplitBlockAndInsertIfThen(
820 /* Unreachable */ !MS.Recover, MS.ColdCallWeights);
822 IRB.SetInsertPoint(CheckTerm);
823 if (MS.TrackOrigins) {
824 IRB.CreateStore(Origin ? (Value *)Origin : (Value *)IRB.getInt32(0),
827 IRB.CreateCall(MS.WarningFn, {});
828 IRB.CreateCall(MS.EmptyAsm, {});
829 DEBUG(dbgs() << " CHECK: " << *Cmp << "\n");
833 void materializeChecks(bool InstrumentWithCalls) {
834 for (const auto &ShadowData : InstrumentationList) {
835 Instruction *OrigIns = ShadowData.OrigIns;
836 Value *Shadow = ShadowData.Shadow;
837 Value *Origin = ShadowData.Origin;
838 materializeOneCheck(OrigIns, Shadow, Origin, InstrumentWithCalls);
840 DEBUG(dbgs() << "DONE:\n" << F);
843 /// \brief Add MemorySanitizer instrumentation to a function.
844 bool runOnFunction() {
845 MS.initializeCallbacks(*F.getParent());
847 // In the presence of unreachable blocks, we may see Phi nodes with
848 // incoming nodes from such blocks. Since InstVisitor skips unreachable
849 // blocks, such nodes will not have any shadow value associated with them.
850 // It's easier to remove unreachable blocks than deal with missing shadow.
851 removeUnreachableBlocks(F);
853 // Iterate all BBs in depth-first order and create shadow instructions
854 // for all instructions (where applicable).
855 // For PHI nodes we create dummy shadow PHIs which will be finalized later.
856 for (BasicBlock *BB : depth_first(&F.getEntryBlock()))
860 // Finalize PHI nodes.
861 for (PHINode *PN : ShadowPHINodes) {
862 PHINode *PNS = cast<PHINode>(getShadow(PN));
863 PHINode *PNO = MS.TrackOrigins ? cast<PHINode>(getOrigin(PN)) : nullptr;
864 size_t NumValues = PN->getNumIncomingValues();
865 for (size_t v = 0; v < NumValues; v++) {
866 PNS->addIncoming(getShadow(PN, v), PN->getIncomingBlock(v));
867 if (PNO) PNO->addIncoming(getOrigin(PN, v), PN->getIncomingBlock(v));
871 VAHelper->finalizeInstrumentation();
873 bool InstrumentWithCalls = ClInstrumentationWithCallThreshold >= 0 &&
874 InstrumentationList.size() + StoreList.size() >
875 (unsigned)ClInstrumentationWithCallThreshold;
877 // Delayed instrumentation of StoreInst.
878 // This may add new checks to be inserted later.
879 materializeStores(InstrumentWithCalls);
881 // Insert shadow value checks.
882 materializeChecks(InstrumentWithCalls);
887 /// \brief Compute the shadow type that corresponds to a given Value.
888 Type *getShadowTy(Value *V) {
889 return getShadowTy(V->getType());
892 /// \brief Compute the shadow type that corresponds to a given Type.
893 Type *getShadowTy(Type *OrigTy) {
894 if (!OrigTy->isSized()) {
897 // For integer type, shadow is the same as the original type.
898 // This may return weird-sized types like i1.
899 if (IntegerType *IT = dyn_cast<IntegerType>(OrigTy))
901 const DataLayout &DL = F.getParent()->getDataLayout();
902 if (VectorType *VT = dyn_cast<VectorType>(OrigTy)) {
903 uint32_t EltSize = DL.getTypeSizeInBits(VT->getElementType());
904 return VectorType::get(IntegerType::get(*MS.C, EltSize),
905 VT->getNumElements());
907 if (ArrayType *AT = dyn_cast<ArrayType>(OrigTy)) {
908 return ArrayType::get(getShadowTy(AT->getElementType()),
909 AT->getNumElements());
911 if (StructType *ST = dyn_cast<StructType>(OrigTy)) {
912 SmallVector<Type*, 4> Elements;
913 for (unsigned i = 0, n = ST->getNumElements(); i < n; i++)
914 Elements.push_back(getShadowTy(ST->getElementType(i)));
915 StructType *Res = StructType::get(*MS.C, Elements, ST->isPacked());
916 DEBUG(dbgs() << "getShadowTy: " << *ST << " ===> " << *Res << "\n");
919 uint32_t TypeSize = DL.getTypeSizeInBits(OrigTy);
920 return IntegerType::get(*MS.C, TypeSize);
923 /// \brief Flatten a vector type.
924 Type *getShadowTyNoVec(Type *ty) {
925 if (VectorType *vt = dyn_cast<VectorType>(ty))
926 return IntegerType::get(*MS.C, vt->getBitWidth());
930 /// \brief Convert a shadow value to it's flattened variant.
931 Value *convertToShadowTyNoVec(Value *V, IRBuilder<> &IRB) {
932 Type *Ty = V->getType();
933 Type *NoVecTy = getShadowTyNoVec(Ty);
934 if (Ty == NoVecTy) return V;
935 return IRB.CreateBitCast(V, NoVecTy);
938 /// \brief Compute the integer shadow offset that corresponds to a given
939 /// application address.
941 /// Offset = (Addr & ~AndMask) ^ XorMask
942 Value *getShadowPtrOffset(Value *Addr, IRBuilder<> &IRB) {
943 Value *OffsetLong = IRB.CreatePointerCast(Addr, MS.IntptrTy);
945 uint64_t AndMask = MS.MapParams->AndMask;
948 IRB.CreateAnd(OffsetLong, ConstantInt::get(MS.IntptrTy, ~AndMask));
950 uint64_t XorMask = MS.MapParams->XorMask;
953 IRB.CreateXor(OffsetLong, ConstantInt::get(MS.IntptrTy, XorMask));
957 /// \brief Compute the shadow address that corresponds to a given application
960 /// Shadow = ShadowBase + Offset
961 Value *getShadowPtr(Value *Addr, Type *ShadowTy,
963 Value *ShadowLong = getShadowPtrOffset(Addr, IRB);
964 uint64_t ShadowBase = MS.MapParams->ShadowBase;
967 IRB.CreateAdd(ShadowLong,
968 ConstantInt::get(MS.IntptrTy, ShadowBase));
969 return IRB.CreateIntToPtr(ShadowLong, PointerType::get(ShadowTy, 0));
972 /// \brief Compute the origin address that corresponds to a given application
975 /// OriginAddr = (OriginBase + Offset) & ~3ULL
976 Value *getOriginPtr(Value *Addr, IRBuilder<> &IRB, unsigned Alignment) {
977 Value *OriginLong = getShadowPtrOffset(Addr, IRB);
978 uint64_t OriginBase = MS.MapParams->OriginBase;
981 IRB.CreateAdd(OriginLong,
982 ConstantInt::get(MS.IntptrTy, OriginBase));
983 if (Alignment < kMinOriginAlignment) {
984 uint64_t Mask = kMinOriginAlignment - 1;
985 OriginLong = IRB.CreateAnd(OriginLong,
986 ConstantInt::get(MS.IntptrTy, ~Mask));
988 return IRB.CreateIntToPtr(OriginLong,
989 PointerType::get(IRB.getInt32Ty(), 0));
992 /// \brief Compute the shadow address for a given function argument.
994 /// Shadow = ParamTLS+ArgOffset.
995 Value *getShadowPtrForArgument(Value *A, IRBuilder<> &IRB,
997 Value *Base = IRB.CreatePointerCast(MS.ParamTLS, MS.IntptrTy);
998 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
999 return IRB.CreateIntToPtr(Base, PointerType::get(getShadowTy(A), 0),
1003 /// \brief Compute the origin address for a given function argument.
1004 Value *getOriginPtrForArgument(Value *A, IRBuilder<> &IRB,
1006 if (!MS.TrackOrigins) return nullptr;
1007 Value *Base = IRB.CreatePointerCast(MS.ParamOriginTLS, MS.IntptrTy);
1008 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
1009 return IRB.CreateIntToPtr(Base, PointerType::get(MS.OriginTy, 0),
1013 /// \brief Compute the shadow address for a retval.
1014 Value *getShadowPtrForRetval(Value *A, IRBuilder<> &IRB) {
1015 Value *Base = IRB.CreatePointerCast(MS.RetvalTLS, MS.IntptrTy);
1016 return IRB.CreateIntToPtr(Base, PointerType::get(getShadowTy(A), 0),
1020 /// \brief Compute the origin address for a retval.
1021 Value *getOriginPtrForRetval(IRBuilder<> &IRB) {
1022 // We keep a single origin for the entire retval. Might be too optimistic.
1023 return MS.RetvalOriginTLS;
1026 /// \brief Set SV to be the shadow value for V.
1027 void setShadow(Value *V, Value *SV) {
1028 assert(!ShadowMap.count(V) && "Values may only have one shadow");
1029 ShadowMap[V] = PropagateShadow ? SV : getCleanShadow(V);
1032 /// \brief Set Origin to be the origin value for V.
1033 void setOrigin(Value *V, Value *Origin) {
1034 if (!MS.TrackOrigins) return;
1035 assert(!OriginMap.count(V) && "Values may only have one origin");
1036 DEBUG(dbgs() << "ORIGIN: " << *V << " ==> " << *Origin << "\n");
1037 OriginMap[V] = Origin;
1040 Constant *getCleanShadow(Type *OrigTy) {
1041 Type *ShadowTy = getShadowTy(OrigTy);
1044 return Constant::getNullValue(ShadowTy);
1047 /// \brief Create a clean shadow value for a given value.
1049 /// Clean shadow (all zeroes) means all bits of the value are defined
1051 Constant *getCleanShadow(Value *V) {
1052 return getCleanShadow(V->getType());
1055 /// \brief Create a dirty shadow of a given shadow type.
1056 Constant *getPoisonedShadow(Type *ShadowTy) {
1058 if (isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy))
1059 return Constant::getAllOnesValue(ShadowTy);
1060 if (ArrayType *AT = dyn_cast<ArrayType>(ShadowTy)) {
1061 SmallVector<Constant *, 4> Vals(AT->getNumElements(),
1062 getPoisonedShadow(AT->getElementType()));
1063 return ConstantArray::get(AT, Vals);
1065 if (StructType *ST = dyn_cast<StructType>(ShadowTy)) {
1066 SmallVector<Constant *, 4> Vals;
1067 for (unsigned i = 0, n = ST->getNumElements(); i < n; i++)
1068 Vals.push_back(getPoisonedShadow(ST->getElementType(i)));
1069 return ConstantStruct::get(ST, Vals);
1071 llvm_unreachable("Unexpected shadow type");
1074 /// \brief Create a dirty shadow for a given value.
1075 Constant *getPoisonedShadow(Value *V) {
1076 Type *ShadowTy = getShadowTy(V);
1079 return getPoisonedShadow(ShadowTy);
1082 /// \brief Create a clean (zero) origin.
1083 Value *getCleanOrigin() {
1084 return Constant::getNullValue(MS.OriginTy);
1087 /// \brief Get the shadow value for a given Value.
1089 /// This function either returns the value set earlier with setShadow,
1090 /// or extracts if from ParamTLS (for function arguments).
1091 Value *getShadow(Value *V) {
1092 if (!PropagateShadow) return getCleanShadow(V);
1093 if (Instruction *I = dyn_cast<Instruction>(V)) {
1094 // For instructions the shadow is already stored in the map.
1095 Value *Shadow = ShadowMap[V];
1097 DEBUG(dbgs() << "No shadow: " << *V << "\n" << *(I->getParent()));
1099 assert(Shadow && "No shadow for a value");
1103 if (UndefValue *U = dyn_cast<UndefValue>(V)) {
1104 Value *AllOnes = PoisonUndef ? getPoisonedShadow(V) : getCleanShadow(V);
1105 DEBUG(dbgs() << "Undef: " << *U << " ==> " << *AllOnes << "\n");
1109 if (Argument *A = dyn_cast<Argument>(V)) {
1110 // For arguments we compute the shadow on demand and store it in the map.
1111 Value **ShadowPtr = &ShadowMap[V];
1114 Function *F = A->getParent();
1115 IRBuilder<> EntryIRB(F->getEntryBlock().getFirstNonPHI());
1116 unsigned ArgOffset = 0;
1117 const DataLayout &DL = F->getParent()->getDataLayout();
1118 for (auto &FArg : F->args()) {
1119 if (!FArg.getType()->isSized()) {
1120 DEBUG(dbgs() << "Arg is not sized\n");
1125 ? DL.getTypeAllocSize(FArg.getType()->getPointerElementType())
1126 : DL.getTypeAllocSize(FArg.getType());
1128 bool Overflow = ArgOffset + Size > kParamTLSSize;
1129 Value *Base = getShadowPtrForArgument(&FArg, EntryIRB, ArgOffset);
1130 if (FArg.hasByValAttr()) {
1131 // ByVal pointer itself has clean shadow. We copy the actual
1132 // argument shadow to the underlying memory.
1133 // Figure out maximal valid memcpy alignment.
1134 unsigned ArgAlign = FArg.getParamAlignment();
1135 if (ArgAlign == 0) {
1136 Type *EltType = A->getType()->getPointerElementType();
1137 ArgAlign = DL.getABITypeAlignment(EltType);
1140 // ParamTLS overflow.
1141 EntryIRB.CreateMemSet(
1142 getShadowPtr(V, EntryIRB.getInt8Ty(), EntryIRB),
1143 Constant::getNullValue(EntryIRB.getInt8Ty()), Size, ArgAlign);
1145 unsigned CopyAlign = std::min(ArgAlign, kShadowTLSAlignment);
1146 Value *Cpy = EntryIRB.CreateMemCpy(
1147 getShadowPtr(V, EntryIRB.getInt8Ty(), EntryIRB), Base, Size,
1149 DEBUG(dbgs() << " ByValCpy: " << *Cpy << "\n");
1152 *ShadowPtr = getCleanShadow(V);
1155 // ParamTLS overflow.
1156 *ShadowPtr = getCleanShadow(V);
1159 EntryIRB.CreateAlignedLoad(Base, kShadowTLSAlignment);
1162 DEBUG(dbgs() << " ARG: " << FArg << " ==> " <<
1163 **ShadowPtr << "\n");
1164 if (MS.TrackOrigins && !Overflow) {
1166 getOriginPtrForArgument(&FArg, EntryIRB, ArgOffset);
1167 setOrigin(A, EntryIRB.CreateLoad(OriginPtr));
1169 setOrigin(A, getCleanOrigin());
1172 ArgOffset += alignTo(Size, kShadowTLSAlignment);
1174 assert(*ShadowPtr && "Could not find shadow for an argument");
1177 // For everything else the shadow is zero.
1178 return getCleanShadow(V);
1181 /// \brief Get the shadow for i-th argument of the instruction I.
1182 Value *getShadow(Instruction *I, int i) {
1183 return getShadow(I->getOperand(i));
1186 /// \brief Get the origin for a value.
1187 Value *getOrigin(Value *V) {
1188 if (!MS.TrackOrigins) return nullptr;
1189 if (!PropagateShadow) return getCleanOrigin();
1190 if (isa<Constant>(V)) return getCleanOrigin();
1191 assert((isa<Instruction>(V) || isa<Argument>(V)) &&
1192 "Unexpected value type in getOrigin()");
1193 Value *Origin = OriginMap[V];
1194 assert(Origin && "Missing origin");
1198 /// \brief Get the origin for i-th argument of the instruction I.
1199 Value *getOrigin(Instruction *I, int i) {
1200 return getOrigin(I->getOperand(i));
1203 /// \brief Remember the place where a shadow check should be inserted.
1205 /// This location will be later instrumented with a check that will print a
1206 /// UMR warning in runtime if the shadow value is not 0.
1207 void insertShadowCheck(Value *Shadow, Value *Origin, Instruction *OrigIns) {
1209 if (!InsertChecks) return;
1211 Type *ShadowTy = Shadow->getType();
1212 assert((isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy)) &&
1213 "Can only insert checks for integer and vector shadow types");
1215 InstrumentationList.push_back(
1216 ShadowOriginAndInsertPoint(Shadow, Origin, OrigIns));
1219 /// \brief Remember the place where a shadow check should be inserted.
1221 /// This location will be later instrumented with a check that will print a
1222 /// UMR warning in runtime if the value is not fully defined.
1223 void insertShadowCheck(Value *Val, Instruction *OrigIns) {
1225 Value *Shadow, *Origin;
1226 if (ClCheckConstantShadow) {
1227 Shadow = getShadow(Val);
1228 if (!Shadow) return;
1229 Origin = getOrigin(Val);
1231 Shadow = dyn_cast_or_null<Instruction>(getShadow(Val));
1232 if (!Shadow) return;
1233 Origin = dyn_cast_or_null<Instruction>(getOrigin(Val));
1235 insertShadowCheck(Shadow, Origin, OrigIns);
1238 AtomicOrdering addReleaseOrdering(AtomicOrdering a) {
1240 case AtomicOrdering::NotAtomic:
1241 return AtomicOrdering::NotAtomic;
1242 case AtomicOrdering::Unordered:
1243 case AtomicOrdering::Monotonic:
1244 case AtomicOrdering::Release:
1245 return AtomicOrdering::Release;
1246 case AtomicOrdering::Acquire:
1247 case AtomicOrdering::AcquireRelease:
1248 return AtomicOrdering::AcquireRelease;
1249 case AtomicOrdering::SequentiallyConsistent:
1250 return AtomicOrdering::SequentiallyConsistent;
1252 llvm_unreachable("Unknown ordering");
1255 AtomicOrdering addAcquireOrdering(AtomicOrdering a) {
1257 case AtomicOrdering::NotAtomic:
1258 return AtomicOrdering::NotAtomic;
1259 case AtomicOrdering::Unordered:
1260 case AtomicOrdering::Monotonic:
1261 case AtomicOrdering::Acquire:
1262 return AtomicOrdering::Acquire;
1263 case AtomicOrdering::Release:
1264 case AtomicOrdering::AcquireRelease:
1265 return AtomicOrdering::AcquireRelease;
1266 case AtomicOrdering::SequentiallyConsistent:
1267 return AtomicOrdering::SequentiallyConsistent;
1269 llvm_unreachable("Unknown ordering");
1272 // ------------------- Visitors.
1274 /// \brief Instrument LoadInst
1276 /// Loads the corresponding shadow and (optionally) origin.
1277 /// Optionally, checks that the load address is fully defined.
1278 void visitLoadInst(LoadInst &I) {
1279 assert(I.getType()->isSized() && "Load type must have size");
1280 IRBuilder<> IRB(I.getNextNode());
1281 Type *ShadowTy = getShadowTy(&I);
1282 Value *Addr = I.getPointerOperand();
1283 if (PropagateShadow && !I.getMetadata("nosanitize")) {
1284 Value *ShadowPtr = getShadowPtr(Addr, ShadowTy, IRB);
1286 IRB.CreateAlignedLoad(ShadowPtr, I.getAlignment(), "_msld"));
1288 setShadow(&I, getCleanShadow(&I));
1291 if (ClCheckAccessAddress)
1292 insertShadowCheck(I.getPointerOperand(), &I);
1295 I.setOrdering(addAcquireOrdering(I.getOrdering()));
1297 if (MS.TrackOrigins) {
1298 if (PropagateShadow) {
1299 unsigned Alignment = I.getAlignment();
1300 unsigned OriginAlignment = std::max(kMinOriginAlignment, Alignment);
1301 setOrigin(&I, IRB.CreateAlignedLoad(getOriginPtr(Addr, IRB, Alignment),
1304 setOrigin(&I, getCleanOrigin());
1309 /// \brief Instrument StoreInst
1311 /// Stores the corresponding shadow and (optionally) origin.
1312 /// Optionally, checks that the store address is fully defined.
1313 void visitStoreInst(StoreInst &I) {
1314 StoreList.push_back(&I);
1317 void handleCASOrRMW(Instruction &I) {
1318 assert(isa<AtomicRMWInst>(I) || isa<AtomicCmpXchgInst>(I));
1320 IRBuilder<> IRB(&I);
1321 Value *Addr = I.getOperand(0);
1322 Value *ShadowPtr = getShadowPtr(Addr, I.getType(), IRB);
1324 if (ClCheckAccessAddress)
1325 insertShadowCheck(Addr, &I);
1327 // Only test the conditional argument of cmpxchg instruction.
1328 // The other argument can potentially be uninitialized, but we can not
1329 // detect this situation reliably without possible false positives.
1330 if (isa<AtomicCmpXchgInst>(I))
1331 insertShadowCheck(I.getOperand(1), &I);
1333 IRB.CreateStore(getCleanShadow(&I), ShadowPtr);
1335 setShadow(&I, getCleanShadow(&I));
1336 setOrigin(&I, getCleanOrigin());
1339 void visitAtomicRMWInst(AtomicRMWInst &I) {
1341 I.setOrdering(addReleaseOrdering(I.getOrdering()));
1344 void visitAtomicCmpXchgInst(AtomicCmpXchgInst &I) {
1346 I.setSuccessOrdering(addReleaseOrdering(I.getSuccessOrdering()));
1349 // Vector manipulation.
1350 void visitExtractElementInst(ExtractElementInst &I) {
1351 insertShadowCheck(I.getOperand(1), &I);
1352 IRBuilder<> IRB(&I);
1353 setShadow(&I, IRB.CreateExtractElement(getShadow(&I, 0), I.getOperand(1),
1355 setOrigin(&I, getOrigin(&I, 0));
1358 void visitInsertElementInst(InsertElementInst &I) {
1359 insertShadowCheck(I.getOperand(2), &I);
1360 IRBuilder<> IRB(&I);
1361 setShadow(&I, IRB.CreateInsertElement(getShadow(&I, 0), getShadow(&I, 1),
1362 I.getOperand(2), "_msprop"));
1363 setOriginForNaryOp(I);
1366 void visitShuffleVectorInst(ShuffleVectorInst &I) {
1367 insertShadowCheck(I.getOperand(2), &I);
1368 IRBuilder<> IRB(&I);
1369 setShadow(&I, IRB.CreateShuffleVector(getShadow(&I, 0), getShadow(&I, 1),
1370 I.getOperand(2), "_msprop"));
1371 setOriginForNaryOp(I);
1375 void visitSExtInst(SExtInst &I) {
1376 IRBuilder<> IRB(&I);
1377 setShadow(&I, IRB.CreateSExt(getShadow(&I, 0), I.getType(), "_msprop"));
1378 setOrigin(&I, getOrigin(&I, 0));
1381 void visitZExtInst(ZExtInst &I) {
1382 IRBuilder<> IRB(&I);
1383 setShadow(&I, IRB.CreateZExt(getShadow(&I, 0), I.getType(), "_msprop"));
1384 setOrigin(&I, getOrigin(&I, 0));
1387 void visitTruncInst(TruncInst &I) {
1388 IRBuilder<> IRB(&I);
1389 setShadow(&I, IRB.CreateTrunc(getShadow(&I, 0), I.getType(), "_msprop"));
1390 setOrigin(&I, getOrigin(&I, 0));
1393 void visitBitCastInst(BitCastInst &I) {
1394 // Special case: if this is the bitcast (there is exactly 1 allowed) between
1395 // a musttail call and a ret, don't instrument. New instructions are not
1396 // allowed after a musttail call.
1397 if (auto *CI = dyn_cast<CallInst>(I.getOperand(0)))
1398 if (CI->isMustTailCall())
1400 IRBuilder<> IRB(&I);
1401 setShadow(&I, IRB.CreateBitCast(getShadow(&I, 0), getShadowTy(&I)));
1402 setOrigin(&I, getOrigin(&I, 0));
1405 void visitPtrToIntInst(PtrToIntInst &I) {
1406 IRBuilder<> IRB(&I);
1407 setShadow(&I, IRB.CreateIntCast(getShadow(&I, 0), getShadowTy(&I), false,
1408 "_msprop_ptrtoint"));
1409 setOrigin(&I, getOrigin(&I, 0));
1412 void visitIntToPtrInst(IntToPtrInst &I) {
1413 IRBuilder<> IRB(&I);
1414 setShadow(&I, IRB.CreateIntCast(getShadow(&I, 0), getShadowTy(&I), false,
1415 "_msprop_inttoptr"));
1416 setOrigin(&I, getOrigin(&I, 0));
1419 void visitFPToSIInst(CastInst& I) { handleShadowOr(I); }
1420 void visitFPToUIInst(CastInst& I) { handleShadowOr(I); }
1421 void visitSIToFPInst(CastInst& I) { handleShadowOr(I); }
1422 void visitUIToFPInst(CastInst& I) { handleShadowOr(I); }
1423 void visitFPExtInst(CastInst& I) { handleShadowOr(I); }
1424 void visitFPTruncInst(CastInst& I) { handleShadowOr(I); }
1426 /// \brief Propagate shadow for bitwise AND.
1428 /// This code is exact, i.e. if, for example, a bit in the left argument
1429 /// is defined and 0, then neither the value not definedness of the
1430 /// corresponding bit in B don't affect the resulting shadow.
1431 void visitAnd(BinaryOperator &I) {
1432 IRBuilder<> IRB(&I);
1433 // "And" of 0 and a poisoned value results in unpoisoned value.
1434 // 1&1 => 1; 0&1 => 0; p&1 => p;
1435 // 1&0 => 0; 0&0 => 0; p&0 => 0;
1436 // 1&p => p; 0&p => 0; p&p => p;
1437 // S = (S1 & S2) | (V1 & S2) | (S1 & V2)
1438 Value *S1 = getShadow(&I, 0);
1439 Value *S2 = getShadow(&I, 1);
1440 Value *V1 = I.getOperand(0);
1441 Value *V2 = I.getOperand(1);
1442 if (V1->getType() != S1->getType()) {
1443 V1 = IRB.CreateIntCast(V1, S1->getType(), false);
1444 V2 = IRB.CreateIntCast(V2, S2->getType(), false);
1446 Value *S1S2 = IRB.CreateAnd(S1, S2);
1447 Value *V1S2 = IRB.CreateAnd(V1, S2);
1448 Value *S1V2 = IRB.CreateAnd(S1, V2);
1449 setShadow(&I, IRB.CreateOr(S1S2, IRB.CreateOr(V1S2, S1V2)));
1450 setOriginForNaryOp(I);
1453 void visitOr(BinaryOperator &I) {
1454 IRBuilder<> IRB(&I);
1455 // "Or" of 1 and a poisoned value results in unpoisoned value.
1456 // 1|1 => 1; 0|1 => 1; p|1 => 1;
1457 // 1|0 => 1; 0|0 => 0; p|0 => p;
1458 // 1|p => 1; 0|p => p; p|p => p;
1459 // S = (S1 & S2) | (~V1 & S2) | (S1 & ~V2)
1460 Value *S1 = getShadow(&I, 0);
1461 Value *S2 = getShadow(&I, 1);
1462 Value *V1 = IRB.CreateNot(I.getOperand(0));
1463 Value *V2 = IRB.CreateNot(I.getOperand(1));
1464 if (V1->getType() != S1->getType()) {
1465 V1 = IRB.CreateIntCast(V1, S1->getType(), false);
1466 V2 = IRB.CreateIntCast(V2, S2->getType(), false);
1468 Value *S1S2 = IRB.CreateAnd(S1, S2);
1469 Value *V1S2 = IRB.CreateAnd(V1, S2);
1470 Value *S1V2 = IRB.CreateAnd(S1, V2);
1471 setShadow(&I, IRB.CreateOr(S1S2, IRB.CreateOr(V1S2, S1V2)));
1472 setOriginForNaryOp(I);
1475 /// \brief Default propagation of shadow and/or origin.
1477 /// This class implements the general case of shadow propagation, used in all
1478 /// cases where we don't know and/or don't care about what the operation
1479 /// actually does. It converts all input shadow values to a common type
1480 /// (extending or truncating as necessary), and bitwise OR's them.
1482 /// This is much cheaper than inserting checks (i.e. requiring inputs to be
1483 /// fully initialized), and less prone to false positives.
1485 /// This class also implements the general case of origin propagation. For a
1486 /// Nary operation, result origin is set to the origin of an argument that is
1487 /// not entirely initialized. If there is more than one such arguments, the
1488 /// rightmost of them is picked. It does not matter which one is picked if all
1489 /// arguments are initialized.
1490 template <bool CombineShadow>
1495 MemorySanitizerVisitor *MSV;
1498 Combiner(MemorySanitizerVisitor *MSV, IRBuilder<> &IRB) :
1499 Shadow(nullptr), Origin(nullptr), IRB(IRB), MSV(MSV) {}
1501 /// \brief Add a pair of shadow and origin values to the mix.
1502 Combiner &Add(Value *OpShadow, Value *OpOrigin) {
1503 if (CombineShadow) {
1508 OpShadow = MSV->CreateShadowCast(IRB, OpShadow, Shadow->getType());
1509 Shadow = IRB.CreateOr(Shadow, OpShadow, "_msprop");
1513 if (MSV->MS.TrackOrigins) {
1518 Constant *ConstOrigin = dyn_cast<Constant>(OpOrigin);
1519 // No point in adding something that might result in 0 origin value.
1520 if (!ConstOrigin || !ConstOrigin->isNullValue()) {
1521 Value *FlatShadow = MSV->convertToShadowTyNoVec(OpShadow, IRB);
1523 IRB.CreateICmpNE(FlatShadow, MSV->getCleanShadow(FlatShadow));
1524 Origin = IRB.CreateSelect(Cond, OpOrigin, Origin);
1531 /// \brief Add an application value to the mix.
1532 Combiner &Add(Value *V) {
1533 Value *OpShadow = MSV->getShadow(V);
1534 Value *OpOrigin = MSV->MS.TrackOrigins ? MSV->getOrigin(V) : nullptr;
1535 return Add(OpShadow, OpOrigin);
1538 /// \brief Set the current combined values as the given instruction's shadow
1540 void Done(Instruction *I) {
1541 if (CombineShadow) {
1543 Shadow = MSV->CreateShadowCast(IRB, Shadow, MSV->getShadowTy(I));
1544 MSV->setShadow(I, Shadow);
1546 if (MSV->MS.TrackOrigins) {
1548 MSV->setOrigin(I, Origin);
1553 typedef Combiner<true> ShadowAndOriginCombiner;
1554 typedef Combiner<false> OriginCombiner;
1556 /// \brief Propagate origin for arbitrary operation.
1557 void setOriginForNaryOp(Instruction &I) {
1558 if (!MS.TrackOrigins) return;
1559 IRBuilder<> IRB(&I);
1560 OriginCombiner OC(this, IRB);
1561 for (Instruction::op_iterator OI = I.op_begin(); OI != I.op_end(); ++OI)
1566 size_t VectorOrPrimitiveTypeSizeInBits(Type *Ty) {
1567 assert(!(Ty->isVectorTy() && Ty->getScalarType()->isPointerTy()) &&
1568 "Vector of pointers is not a valid shadow type");
1569 return Ty->isVectorTy() ?
1570 Ty->getVectorNumElements() * Ty->getScalarSizeInBits() :
1571 Ty->getPrimitiveSizeInBits();
1574 /// \brief Cast between two shadow types, extending or truncating as
1576 Value *CreateShadowCast(IRBuilder<> &IRB, Value *V, Type *dstTy,
1577 bool Signed = false) {
1578 Type *srcTy = V->getType();
1579 size_t srcSizeInBits = VectorOrPrimitiveTypeSizeInBits(srcTy);
1580 size_t dstSizeInBits = VectorOrPrimitiveTypeSizeInBits(dstTy);
1581 if (srcSizeInBits > 1 && dstSizeInBits == 1)
1582 return IRB.CreateICmpNE(V, getCleanShadow(V));
1584 if (dstTy->isIntegerTy() && srcTy->isIntegerTy())
1585 return IRB.CreateIntCast(V, dstTy, Signed);
1586 if (dstTy->isVectorTy() && srcTy->isVectorTy() &&
1587 dstTy->getVectorNumElements() == srcTy->getVectorNumElements())
1588 return IRB.CreateIntCast(V, dstTy, Signed);
1589 Value *V1 = IRB.CreateBitCast(V, Type::getIntNTy(*MS.C, srcSizeInBits));
1591 IRB.CreateIntCast(V1, Type::getIntNTy(*MS.C, dstSizeInBits), Signed);
1592 return IRB.CreateBitCast(V2, dstTy);
1593 // TODO: handle struct types.
1596 /// \brief Cast an application value to the type of its own shadow.
1597 Value *CreateAppToShadowCast(IRBuilder<> &IRB, Value *V) {
1598 Type *ShadowTy = getShadowTy(V);
1599 if (V->getType() == ShadowTy)
1601 if (V->getType()->isPtrOrPtrVectorTy())
1602 return IRB.CreatePtrToInt(V, ShadowTy);
1604 return IRB.CreateBitCast(V, ShadowTy);
1607 /// \brief Propagate shadow for arbitrary operation.
1608 void handleShadowOr(Instruction &I) {
1609 IRBuilder<> IRB(&I);
1610 ShadowAndOriginCombiner SC(this, IRB);
1611 for (Instruction::op_iterator OI = I.op_begin(); OI != I.op_end(); ++OI)
1616 // \brief Handle multiplication by constant.
1618 // Handle a special case of multiplication by constant that may have one or
1619 // more zeros in the lower bits. This makes corresponding number of lower bits
1620 // of the result zero as well. We model it by shifting the other operand
1621 // shadow left by the required number of bits. Effectively, we transform
1622 // (X * (A * 2**B)) to ((X << B) * A) and instrument (X << B) as (Sx << B).
1623 // We use multiplication by 2**N instead of shift to cover the case of
1624 // multiplication by 0, which may occur in some elements of a vector operand.
1625 void handleMulByConstant(BinaryOperator &I, Constant *ConstArg,
1627 Constant *ShadowMul;
1628 Type *Ty = ConstArg->getType();
1629 if (Ty->isVectorTy()) {
1630 unsigned NumElements = Ty->getVectorNumElements();
1631 Type *EltTy = Ty->getSequentialElementType();
1632 SmallVector<Constant *, 16> Elements;
1633 for (unsigned Idx = 0; Idx < NumElements; ++Idx) {
1634 if (ConstantInt *Elt =
1635 dyn_cast<ConstantInt>(ConstArg->getAggregateElement(Idx))) {
1636 const APInt &V = Elt->getValue();
1637 APInt V2 = APInt(V.getBitWidth(), 1) << V.countTrailingZeros();
1638 Elements.push_back(ConstantInt::get(EltTy, V2));
1640 Elements.push_back(ConstantInt::get(EltTy, 1));
1643 ShadowMul = ConstantVector::get(Elements);
1645 if (ConstantInt *Elt = dyn_cast<ConstantInt>(ConstArg)) {
1646 const APInt &V = Elt->getValue();
1647 APInt V2 = APInt(V.getBitWidth(), 1) << V.countTrailingZeros();
1648 ShadowMul = ConstantInt::get(Ty, V2);
1650 ShadowMul = ConstantInt::get(Ty, 1);
1654 IRBuilder<> IRB(&I);
1656 IRB.CreateMul(getShadow(OtherArg), ShadowMul, "msprop_mul_cst"));
1657 setOrigin(&I, getOrigin(OtherArg));
1660 void visitMul(BinaryOperator &I) {
1661 Constant *constOp0 = dyn_cast<Constant>(I.getOperand(0));
1662 Constant *constOp1 = dyn_cast<Constant>(I.getOperand(1));
1663 if (constOp0 && !constOp1)
1664 handleMulByConstant(I, constOp0, I.getOperand(1));
1665 else if (constOp1 && !constOp0)
1666 handleMulByConstant(I, constOp1, I.getOperand(0));
1671 void visitFAdd(BinaryOperator &I) { handleShadowOr(I); }
1672 void visitFSub(BinaryOperator &I) { handleShadowOr(I); }
1673 void visitFMul(BinaryOperator &I) { handleShadowOr(I); }
1674 void visitAdd(BinaryOperator &I) { handleShadowOr(I); }
1675 void visitSub(BinaryOperator &I) { handleShadowOr(I); }
1676 void visitXor(BinaryOperator &I) { handleShadowOr(I); }
1678 void handleDiv(Instruction &I) {
1679 IRBuilder<> IRB(&I);
1680 // Strict on the second argument.
1681 insertShadowCheck(I.getOperand(1), &I);
1682 setShadow(&I, getShadow(&I, 0));
1683 setOrigin(&I, getOrigin(&I, 0));
1686 void visitUDiv(BinaryOperator &I) { handleDiv(I); }
1687 void visitSDiv(BinaryOperator &I) { handleDiv(I); }
1688 void visitFDiv(BinaryOperator &I) { handleDiv(I); }
1689 void visitURem(BinaryOperator &I) { handleDiv(I); }
1690 void visitSRem(BinaryOperator &I) { handleDiv(I); }
1691 void visitFRem(BinaryOperator &I) { handleDiv(I); }
1693 /// \brief Instrument == and != comparisons.
1695 /// Sometimes the comparison result is known even if some of the bits of the
1696 /// arguments are not.
1697 void handleEqualityComparison(ICmpInst &I) {
1698 IRBuilder<> IRB(&I);
1699 Value *A = I.getOperand(0);
1700 Value *B = I.getOperand(1);
1701 Value *Sa = getShadow(A);
1702 Value *Sb = getShadow(B);
1704 // Get rid of pointers and vectors of pointers.
1705 // For ints (and vectors of ints), types of A and Sa match,
1706 // and this is a no-op.
1707 A = IRB.CreatePointerCast(A, Sa->getType());
1708 B = IRB.CreatePointerCast(B, Sb->getType());
1710 // A == B <==> (C = A^B) == 0
1711 // A != B <==> (C = A^B) != 0
1713 Value *C = IRB.CreateXor(A, B);
1714 Value *Sc = IRB.CreateOr(Sa, Sb);
1715 // Now dealing with i = (C == 0) comparison (or C != 0, does not matter now)
1716 // Result is defined if one of the following is true
1717 // * there is a defined 1 bit in C
1718 // * C is fully defined
1719 // Si = !(C & ~Sc) && Sc
1720 Value *Zero = Constant::getNullValue(Sc->getType());
1721 Value *MinusOne = Constant::getAllOnesValue(Sc->getType());
1723 IRB.CreateAnd(IRB.CreateICmpNE(Sc, Zero),
1725 IRB.CreateAnd(IRB.CreateXor(Sc, MinusOne), C), Zero));
1726 Si->setName("_msprop_icmp");
1728 setOriginForNaryOp(I);
1731 /// \brief Build the lowest possible value of V, taking into account V's
1732 /// uninitialized bits.
1733 Value *getLowestPossibleValue(IRBuilder<> &IRB, Value *A, Value *Sa,
1736 // Split shadow into sign bit and other bits.
1737 Value *SaOtherBits = IRB.CreateLShr(IRB.CreateShl(Sa, 1), 1);
1738 Value *SaSignBit = IRB.CreateXor(Sa, SaOtherBits);
1739 // Maximise the undefined shadow bit, minimize other undefined bits.
1741 IRB.CreateOr(IRB.CreateAnd(A, IRB.CreateNot(SaOtherBits)), SaSignBit);
1743 // Minimize undefined bits.
1744 return IRB.CreateAnd(A, IRB.CreateNot(Sa));
1748 /// \brief Build the highest possible value of V, taking into account V's
1749 /// uninitialized bits.
1750 Value *getHighestPossibleValue(IRBuilder<> &IRB, Value *A, Value *Sa,
1753 // Split shadow into sign bit and other bits.
1754 Value *SaOtherBits = IRB.CreateLShr(IRB.CreateShl(Sa, 1), 1);
1755 Value *SaSignBit = IRB.CreateXor(Sa, SaOtherBits);
1756 // Minimise the undefined shadow bit, maximise other undefined bits.
1758 IRB.CreateOr(IRB.CreateAnd(A, IRB.CreateNot(SaSignBit)), SaOtherBits);
1760 // Maximize undefined bits.
1761 return IRB.CreateOr(A, Sa);
1765 /// \brief Instrument relational comparisons.
1767 /// This function does exact shadow propagation for all relational
1768 /// comparisons of integers, pointers and vectors of those.
1769 /// FIXME: output seems suboptimal when one of the operands is a constant
1770 void handleRelationalComparisonExact(ICmpInst &I) {
1771 IRBuilder<> IRB(&I);
1772 Value *A = I.getOperand(0);
1773 Value *B = I.getOperand(1);
1774 Value *Sa = getShadow(A);
1775 Value *Sb = getShadow(B);
1777 // Get rid of pointers and vectors of pointers.
1778 // For ints (and vectors of ints), types of A and Sa match,
1779 // and this is a no-op.
1780 A = IRB.CreatePointerCast(A, Sa->getType());
1781 B = IRB.CreatePointerCast(B, Sb->getType());
1783 // Let [a0, a1] be the interval of possible values of A, taking into account
1784 // its undefined bits. Let [b0, b1] be the interval of possible values of B.
1785 // Then (A cmp B) is defined iff (a0 cmp b1) == (a1 cmp b0).
1786 bool IsSigned = I.isSigned();
1787 Value *S1 = IRB.CreateICmp(I.getPredicate(),
1788 getLowestPossibleValue(IRB, A, Sa, IsSigned),
1789 getHighestPossibleValue(IRB, B, Sb, IsSigned));
1790 Value *S2 = IRB.CreateICmp(I.getPredicate(),
1791 getHighestPossibleValue(IRB, A, Sa, IsSigned),
1792 getLowestPossibleValue(IRB, B, Sb, IsSigned));
1793 Value *Si = IRB.CreateXor(S1, S2);
1795 setOriginForNaryOp(I);
1798 /// \brief Instrument signed relational comparisons.
1800 /// Handle sign bit tests: x<0, x>=0, x<=-1, x>-1 by propagating the highest
1801 /// bit of the shadow. Everything else is delegated to handleShadowOr().
1802 void handleSignedRelationalComparison(ICmpInst &I) {
1804 Value *op = nullptr;
1805 CmpInst::Predicate pre;
1806 if ((constOp = dyn_cast<Constant>(I.getOperand(1)))) {
1807 op = I.getOperand(0);
1808 pre = I.getPredicate();
1809 } else if ((constOp = dyn_cast<Constant>(I.getOperand(0)))) {
1810 op = I.getOperand(1);
1811 pre = I.getSwappedPredicate();
1817 if ((constOp->isNullValue() &&
1818 (pre == CmpInst::ICMP_SLT || pre == CmpInst::ICMP_SGE)) ||
1819 (constOp->isAllOnesValue() &&
1820 (pre == CmpInst::ICMP_SGT || pre == CmpInst::ICMP_SLE))) {
1821 IRBuilder<> IRB(&I);
1822 Value *Shadow = IRB.CreateICmpSLT(getShadow(op), getCleanShadow(op),
1824 setShadow(&I, Shadow);
1825 setOrigin(&I, getOrigin(op));
1831 void visitICmpInst(ICmpInst &I) {
1832 if (!ClHandleICmp) {
1836 if (I.isEquality()) {
1837 handleEqualityComparison(I);
1841 assert(I.isRelational());
1842 if (ClHandleICmpExact) {
1843 handleRelationalComparisonExact(I);
1847 handleSignedRelationalComparison(I);
1851 assert(I.isUnsigned());
1852 if ((isa<Constant>(I.getOperand(0)) || isa<Constant>(I.getOperand(1)))) {
1853 handleRelationalComparisonExact(I);
1860 void visitFCmpInst(FCmpInst &I) {
1864 void handleShift(BinaryOperator &I) {
1865 IRBuilder<> IRB(&I);
1866 // If any of the S2 bits are poisoned, the whole thing is poisoned.
1867 // Otherwise perform the same shift on S1.
1868 Value *S1 = getShadow(&I, 0);
1869 Value *S2 = getShadow(&I, 1);
1870 Value *S2Conv = IRB.CreateSExt(IRB.CreateICmpNE(S2, getCleanShadow(S2)),
1872 Value *V2 = I.getOperand(1);
1873 Value *Shift = IRB.CreateBinOp(I.getOpcode(), S1, V2);
1874 setShadow(&I, IRB.CreateOr(Shift, S2Conv));
1875 setOriginForNaryOp(I);
1878 void visitShl(BinaryOperator &I) { handleShift(I); }
1879 void visitAShr(BinaryOperator &I) { handleShift(I); }
1880 void visitLShr(BinaryOperator &I) { handleShift(I); }
1882 /// \brief Instrument llvm.memmove
1884 /// At this point we don't know if llvm.memmove will be inlined or not.
1885 /// If we don't instrument it and it gets inlined,
1886 /// our interceptor will not kick in and we will lose the memmove.
1887 /// If we instrument the call here, but it does not get inlined,
1888 /// we will memove the shadow twice: which is bad in case
1889 /// of overlapping regions. So, we simply lower the intrinsic to a call.
1891 /// Similar situation exists for memcpy and memset.
1892 void visitMemMoveInst(MemMoveInst &I) {
1893 IRBuilder<> IRB(&I);
1896 {IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1897 IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()),
1898 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false)});
1899 I.eraseFromParent();
1902 // Similar to memmove: avoid copying shadow twice.
1903 // This is somewhat unfortunate as it may slowdown small constant memcpys.
1904 // FIXME: consider doing manual inline for small constant sizes and proper
1906 void visitMemCpyInst(MemCpyInst &I) {
1907 IRBuilder<> IRB(&I);
1910 {IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1911 IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()),
1912 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false)});
1913 I.eraseFromParent();
1917 void visitMemSetInst(MemSetInst &I) {
1918 IRBuilder<> IRB(&I);
1921 {IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1922 IRB.CreateIntCast(I.getArgOperand(1), IRB.getInt32Ty(), false),
1923 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false)});
1924 I.eraseFromParent();
1927 void visitVAStartInst(VAStartInst &I) {
1928 VAHelper->visitVAStartInst(I);
1931 void visitVACopyInst(VACopyInst &I) {
1932 VAHelper->visitVACopyInst(I);
1935 /// \brief Handle vector store-like intrinsics.
1937 /// Instrument intrinsics that look like a simple SIMD store: writes memory,
1938 /// has 1 pointer argument and 1 vector argument, returns void.
1939 bool handleVectorStoreIntrinsic(IntrinsicInst &I) {
1940 IRBuilder<> IRB(&I);
1941 Value* Addr = I.getArgOperand(0);
1942 Value *Shadow = getShadow(&I, 1);
1943 Value *ShadowPtr = getShadowPtr(Addr, Shadow->getType(), IRB);
1945 // We don't know the pointer alignment (could be unaligned SSE store!).
1946 // Have to assume to worst case.
1947 IRB.CreateAlignedStore(Shadow, ShadowPtr, 1);
1949 if (ClCheckAccessAddress)
1950 insertShadowCheck(Addr, &I);
1952 // FIXME: factor out common code from materializeStores
1953 if (MS.TrackOrigins)
1954 IRB.CreateStore(getOrigin(&I, 1), getOriginPtr(Addr, IRB, 1));
1958 /// \brief Handle vector load-like intrinsics.
1960 /// Instrument intrinsics that look like a simple SIMD load: reads memory,
1961 /// has 1 pointer argument, returns a vector.
1962 bool handleVectorLoadIntrinsic(IntrinsicInst &I) {
1963 IRBuilder<> IRB(&I);
1964 Value *Addr = I.getArgOperand(0);
1966 Type *ShadowTy = getShadowTy(&I);
1967 if (PropagateShadow) {
1968 Value *ShadowPtr = getShadowPtr(Addr, ShadowTy, IRB);
1969 // We don't know the pointer alignment (could be unaligned SSE load!).
1970 // Have to assume to worst case.
1971 setShadow(&I, IRB.CreateAlignedLoad(ShadowPtr, 1, "_msld"));
1973 setShadow(&I, getCleanShadow(&I));
1976 if (ClCheckAccessAddress)
1977 insertShadowCheck(Addr, &I);
1979 if (MS.TrackOrigins) {
1980 if (PropagateShadow)
1981 setOrigin(&I, IRB.CreateLoad(getOriginPtr(Addr, IRB, 1)));
1983 setOrigin(&I, getCleanOrigin());
1988 /// \brief Handle (SIMD arithmetic)-like intrinsics.
1990 /// Instrument intrinsics with any number of arguments of the same type,
1991 /// equal to the return type. The type should be simple (no aggregates or
1992 /// pointers; vectors are fine).
1993 /// Caller guarantees that this intrinsic does not access memory.
1994 bool maybeHandleSimpleNomemIntrinsic(IntrinsicInst &I) {
1995 Type *RetTy = I.getType();
1996 if (!(RetTy->isIntOrIntVectorTy() ||
1997 RetTy->isFPOrFPVectorTy() ||
1998 RetTy->isX86_MMXTy()))
2001 unsigned NumArgOperands = I.getNumArgOperands();
2003 for (unsigned i = 0; i < NumArgOperands; ++i) {
2004 Type *Ty = I.getArgOperand(i)->getType();
2009 IRBuilder<> IRB(&I);
2010 ShadowAndOriginCombiner SC(this, IRB);
2011 for (unsigned i = 0; i < NumArgOperands; ++i)
2012 SC.Add(I.getArgOperand(i));
2018 /// \brief Heuristically instrument unknown intrinsics.
2020 /// The main purpose of this code is to do something reasonable with all
2021 /// random intrinsics we might encounter, most importantly - SIMD intrinsics.
2022 /// We recognize several classes of intrinsics by their argument types and
2023 /// ModRefBehaviour and apply special intrumentation when we are reasonably
2024 /// sure that we know what the intrinsic does.
2026 /// We special-case intrinsics where this approach fails. See llvm.bswap
2027 /// handling as an example of that.
2028 bool handleUnknownIntrinsic(IntrinsicInst &I) {
2029 unsigned NumArgOperands = I.getNumArgOperands();
2030 if (NumArgOperands == 0)
2033 if (NumArgOperands == 2 &&
2034 I.getArgOperand(0)->getType()->isPointerTy() &&
2035 I.getArgOperand(1)->getType()->isVectorTy() &&
2036 I.getType()->isVoidTy() &&
2037 !I.onlyReadsMemory()) {
2038 // This looks like a vector store.
2039 return handleVectorStoreIntrinsic(I);
2042 if (NumArgOperands == 1 &&
2043 I.getArgOperand(0)->getType()->isPointerTy() &&
2044 I.getType()->isVectorTy() &&
2045 I.onlyReadsMemory()) {
2046 // This looks like a vector load.
2047 return handleVectorLoadIntrinsic(I);
2050 if (I.doesNotAccessMemory())
2051 if (maybeHandleSimpleNomemIntrinsic(I))
2054 // FIXME: detect and handle SSE maskstore/maskload
2058 void handleBswap(IntrinsicInst &I) {
2059 IRBuilder<> IRB(&I);
2060 Value *Op = I.getArgOperand(0);
2061 Type *OpType = Op->getType();
2062 Function *BswapFunc = Intrinsic::getDeclaration(
2063 F.getParent(), Intrinsic::bswap, makeArrayRef(&OpType, 1));
2064 setShadow(&I, IRB.CreateCall(BswapFunc, getShadow(Op)));
2065 setOrigin(&I, getOrigin(Op));
2068 // \brief Instrument vector convert instrinsic.
2070 // This function instruments intrinsics like cvtsi2ss:
2071 // %Out = int_xxx_cvtyyy(%ConvertOp)
2073 // %Out = int_xxx_cvtyyy(%CopyOp, %ConvertOp)
2074 // Intrinsic converts \p NumUsedElements elements of \p ConvertOp to the same
2075 // number \p Out elements, and (if has 2 arguments) copies the rest of the
2076 // elements from \p CopyOp.
2077 // In most cases conversion involves floating-point value which may trigger a
2078 // hardware exception when not fully initialized. For this reason we require
2079 // \p ConvertOp[0:NumUsedElements] to be fully initialized and trap otherwise.
2080 // We copy the shadow of \p CopyOp[NumUsedElements:] to \p
2081 // Out[NumUsedElements:]. This means that intrinsics without \p CopyOp always
2082 // return a fully initialized value.
2083 void handleVectorConvertIntrinsic(IntrinsicInst &I, int NumUsedElements) {
2084 IRBuilder<> IRB(&I);
2085 Value *CopyOp, *ConvertOp;
2087 switch (I.getNumArgOperands()) {
2089 assert(isa<ConstantInt>(I.getArgOperand(2)) && "Invalid rounding mode");
2092 CopyOp = I.getArgOperand(0);
2093 ConvertOp = I.getArgOperand(1);
2096 ConvertOp = I.getArgOperand(0);
2100 llvm_unreachable("Cvt intrinsic with unsupported number of arguments.");
2103 // The first *NumUsedElements* elements of ConvertOp are converted to the
2104 // same number of output elements. The rest of the output is copied from
2105 // CopyOp, or (if not available) filled with zeroes.
2106 // Combine shadow for elements of ConvertOp that are used in this operation,
2107 // and insert a check.
2108 // FIXME: consider propagating shadow of ConvertOp, at least in the case of
2109 // int->any conversion.
2110 Value *ConvertShadow = getShadow(ConvertOp);
2111 Value *AggShadow = nullptr;
2112 if (ConvertOp->getType()->isVectorTy()) {
2113 AggShadow = IRB.CreateExtractElement(
2114 ConvertShadow, ConstantInt::get(IRB.getInt32Ty(), 0));
2115 for (int i = 1; i < NumUsedElements; ++i) {
2116 Value *MoreShadow = IRB.CreateExtractElement(
2117 ConvertShadow, ConstantInt::get(IRB.getInt32Ty(), i));
2118 AggShadow = IRB.CreateOr(AggShadow, MoreShadow);
2121 AggShadow = ConvertShadow;
2123 assert(AggShadow->getType()->isIntegerTy());
2124 insertShadowCheck(AggShadow, getOrigin(ConvertOp), &I);
2126 // Build result shadow by zero-filling parts of CopyOp shadow that come from
2129 assert(CopyOp->getType() == I.getType());
2130 assert(CopyOp->getType()->isVectorTy());
2131 Value *ResultShadow = getShadow(CopyOp);
2132 Type *EltTy = ResultShadow->getType()->getVectorElementType();
2133 for (int i = 0; i < NumUsedElements; ++i) {
2134 ResultShadow = IRB.CreateInsertElement(
2135 ResultShadow, ConstantInt::getNullValue(EltTy),
2136 ConstantInt::get(IRB.getInt32Ty(), i));
2138 setShadow(&I, ResultShadow);
2139 setOrigin(&I, getOrigin(CopyOp));
2141 setShadow(&I, getCleanShadow(&I));
2142 setOrigin(&I, getCleanOrigin());
2146 // Given a scalar or vector, extract lower 64 bits (or less), and return all
2147 // zeroes if it is zero, and all ones otherwise.
2148 Value *Lower64ShadowExtend(IRBuilder<> &IRB, Value *S, Type *T) {
2149 if (S->getType()->isVectorTy())
2150 S = CreateShadowCast(IRB, S, IRB.getInt64Ty(), /* Signed */ true);
2151 assert(S->getType()->getPrimitiveSizeInBits() <= 64);
2152 Value *S2 = IRB.CreateICmpNE(S, getCleanShadow(S));
2153 return CreateShadowCast(IRB, S2, T, /* Signed */ true);
2156 // Given a vector, extract its first element, and return all
2157 // zeroes if it is zero, and all ones otherwise.
2158 Value *LowerElementShadowExtend(IRBuilder<> &IRB, Value *S, Type *T) {
2159 Value *S1 = IRB.CreateExtractElement(S, (uint64_t)0);
2160 Value *S2 = IRB.CreateICmpNE(S1, getCleanShadow(S1));
2161 return CreateShadowCast(IRB, S2, T, /* Signed */ true);
2164 Value *VariableShadowExtend(IRBuilder<> &IRB, Value *S) {
2165 Type *T = S->getType();
2166 assert(T->isVectorTy());
2167 Value *S2 = IRB.CreateICmpNE(S, getCleanShadow(S));
2168 return IRB.CreateSExt(S2, T);
2171 // \brief Instrument vector shift instrinsic.
2173 // This function instruments intrinsics like int_x86_avx2_psll_w.
2174 // Intrinsic shifts %In by %ShiftSize bits.
2175 // %ShiftSize may be a vector. In that case the lower 64 bits determine shift
2176 // size, and the rest is ignored. Behavior is defined even if shift size is
2177 // greater than register (or field) width.
2178 void handleVectorShiftIntrinsic(IntrinsicInst &I, bool Variable) {
2179 assert(I.getNumArgOperands() == 2);
2180 IRBuilder<> IRB(&I);
2181 // If any of the S2 bits are poisoned, the whole thing is poisoned.
2182 // Otherwise perform the same shift on S1.
2183 Value *S1 = getShadow(&I, 0);
2184 Value *S2 = getShadow(&I, 1);
2185 Value *S2Conv = Variable ? VariableShadowExtend(IRB, S2)
2186 : Lower64ShadowExtend(IRB, S2, getShadowTy(&I));
2187 Value *V1 = I.getOperand(0);
2188 Value *V2 = I.getOperand(1);
2189 Value *Shift = IRB.CreateCall(I.getCalledValue(),
2190 {IRB.CreateBitCast(S1, V1->getType()), V2});
2191 Shift = IRB.CreateBitCast(Shift, getShadowTy(&I));
2192 setShadow(&I, IRB.CreateOr(Shift, S2Conv));
2193 setOriginForNaryOp(I);
2196 // \brief Get an X86_MMX-sized vector type.
2197 Type *getMMXVectorTy(unsigned EltSizeInBits) {
2198 const unsigned X86_MMXSizeInBits = 64;
2199 return VectorType::get(IntegerType::get(*MS.C, EltSizeInBits),
2200 X86_MMXSizeInBits / EltSizeInBits);
2203 // \brief Returns a signed counterpart for an (un)signed-saturate-and-pack
2205 Intrinsic::ID getSignedPackIntrinsic(Intrinsic::ID id) {
2207 case llvm::Intrinsic::x86_sse2_packsswb_128:
2208 case llvm::Intrinsic::x86_sse2_packuswb_128:
2209 return llvm::Intrinsic::x86_sse2_packsswb_128;
2211 case llvm::Intrinsic::x86_sse2_packssdw_128:
2212 case llvm::Intrinsic::x86_sse41_packusdw:
2213 return llvm::Intrinsic::x86_sse2_packssdw_128;
2215 case llvm::Intrinsic::x86_avx2_packsswb:
2216 case llvm::Intrinsic::x86_avx2_packuswb:
2217 return llvm::Intrinsic::x86_avx2_packsswb;
2219 case llvm::Intrinsic::x86_avx2_packssdw:
2220 case llvm::Intrinsic::x86_avx2_packusdw:
2221 return llvm::Intrinsic::x86_avx2_packssdw;
2223 case llvm::Intrinsic::x86_mmx_packsswb:
2224 case llvm::Intrinsic::x86_mmx_packuswb:
2225 return llvm::Intrinsic::x86_mmx_packsswb;
2227 case llvm::Intrinsic::x86_mmx_packssdw:
2228 return llvm::Intrinsic::x86_mmx_packssdw;
2230 llvm_unreachable("unexpected intrinsic id");
2234 // \brief Instrument vector pack instrinsic.
2236 // This function instruments intrinsics like x86_mmx_packsswb, that
2237 // packs elements of 2 input vectors into half as many bits with saturation.
2238 // Shadow is propagated with the signed variant of the same intrinsic applied
2239 // to sext(Sa != zeroinitializer), sext(Sb != zeroinitializer).
2240 // EltSizeInBits is used only for x86mmx arguments.
2241 void handleVectorPackIntrinsic(IntrinsicInst &I, unsigned EltSizeInBits = 0) {
2242 assert(I.getNumArgOperands() == 2);
2243 bool isX86_MMX = I.getOperand(0)->getType()->isX86_MMXTy();
2244 IRBuilder<> IRB(&I);
2245 Value *S1 = getShadow(&I, 0);
2246 Value *S2 = getShadow(&I, 1);
2247 assert(isX86_MMX || S1->getType()->isVectorTy());
2249 // SExt and ICmpNE below must apply to individual elements of input vectors.
2250 // In case of x86mmx arguments, cast them to appropriate vector types and
2252 Type *T = isX86_MMX ? getMMXVectorTy(EltSizeInBits) : S1->getType();
2254 S1 = IRB.CreateBitCast(S1, T);
2255 S2 = IRB.CreateBitCast(S2, T);
2257 Value *S1_ext = IRB.CreateSExt(
2258 IRB.CreateICmpNE(S1, llvm::Constant::getNullValue(T)), T);
2259 Value *S2_ext = IRB.CreateSExt(
2260 IRB.CreateICmpNE(S2, llvm::Constant::getNullValue(T)), T);
2262 Type *X86_MMXTy = Type::getX86_MMXTy(*MS.C);
2263 S1_ext = IRB.CreateBitCast(S1_ext, X86_MMXTy);
2264 S2_ext = IRB.CreateBitCast(S2_ext, X86_MMXTy);
2267 Function *ShadowFn = Intrinsic::getDeclaration(
2268 F.getParent(), getSignedPackIntrinsic(I.getIntrinsicID()));
2271 IRB.CreateCall(ShadowFn, {S1_ext, S2_ext}, "_msprop_vector_pack");
2272 if (isX86_MMX) S = IRB.CreateBitCast(S, getShadowTy(&I));
2274 setOriginForNaryOp(I);
2277 // \brief Instrument sum-of-absolute-differencies intrinsic.
2278 void handleVectorSadIntrinsic(IntrinsicInst &I) {
2279 const unsigned SignificantBitsPerResultElement = 16;
2280 bool isX86_MMX = I.getOperand(0)->getType()->isX86_MMXTy();
2281 Type *ResTy = isX86_MMX ? IntegerType::get(*MS.C, 64) : I.getType();
2282 unsigned ZeroBitsPerResultElement =
2283 ResTy->getScalarSizeInBits() - SignificantBitsPerResultElement;
2285 IRBuilder<> IRB(&I);
2286 Value *S = IRB.CreateOr(getShadow(&I, 0), getShadow(&I, 1));
2287 S = IRB.CreateBitCast(S, ResTy);
2288 S = IRB.CreateSExt(IRB.CreateICmpNE(S, Constant::getNullValue(ResTy)),
2290 S = IRB.CreateLShr(S, ZeroBitsPerResultElement);
2291 S = IRB.CreateBitCast(S, getShadowTy(&I));
2293 setOriginForNaryOp(I);
2296 // \brief Instrument multiply-add intrinsic.
2297 void handleVectorPmaddIntrinsic(IntrinsicInst &I,
2298 unsigned EltSizeInBits = 0) {
2299 bool isX86_MMX = I.getOperand(0)->getType()->isX86_MMXTy();
2300 Type *ResTy = isX86_MMX ? getMMXVectorTy(EltSizeInBits * 2) : I.getType();
2301 IRBuilder<> IRB(&I);
2302 Value *S = IRB.CreateOr(getShadow(&I, 0), getShadow(&I, 1));
2303 S = IRB.CreateBitCast(S, ResTy);
2304 S = IRB.CreateSExt(IRB.CreateICmpNE(S, Constant::getNullValue(ResTy)),
2306 S = IRB.CreateBitCast(S, getShadowTy(&I));
2308 setOriginForNaryOp(I);
2311 // \brief Instrument compare-packed intrinsic.
2312 // Basically, an or followed by sext(icmp ne 0) to end up with all-zeros or
2314 void handleVectorComparePackedIntrinsic(IntrinsicInst &I) {
2315 IRBuilder<> IRB(&I);
2316 Type *ResTy = getShadowTy(&I);
2317 Value *S0 = IRB.CreateOr(getShadow(&I, 0), getShadow(&I, 1));
2318 Value *S = IRB.CreateSExt(
2319 IRB.CreateICmpNE(S0, Constant::getNullValue(ResTy)), ResTy);
2321 setOriginForNaryOp(I);
2324 // \brief Instrument compare-scalar intrinsic.
2325 // This handles both cmp* intrinsics which return the result in the first
2326 // element of a vector, and comi* which return the result as i32.
2327 void handleVectorCompareScalarIntrinsic(IntrinsicInst &I) {
2328 IRBuilder<> IRB(&I);
2329 Value *S0 = IRB.CreateOr(getShadow(&I, 0), getShadow(&I, 1));
2330 Value *S = LowerElementShadowExtend(IRB, S0, getShadowTy(&I));
2332 setOriginForNaryOp(I);
2335 void handleStmxcsr(IntrinsicInst &I) {
2336 IRBuilder<> IRB(&I);
2337 Value* Addr = I.getArgOperand(0);
2338 Type *Ty = IRB.getInt32Ty();
2339 Value *ShadowPtr = getShadowPtr(Addr, Ty, IRB);
2341 IRB.CreateStore(getCleanShadow(Ty),
2342 IRB.CreatePointerCast(ShadowPtr, Ty->getPointerTo()));
2344 if (ClCheckAccessAddress)
2345 insertShadowCheck(Addr, &I);
2348 void handleLdmxcsr(IntrinsicInst &I) {
2349 if (!InsertChecks) return;
2351 IRBuilder<> IRB(&I);
2352 Value *Addr = I.getArgOperand(0);
2353 Type *Ty = IRB.getInt32Ty();
2354 unsigned Alignment = 1;
2356 if (ClCheckAccessAddress)
2357 insertShadowCheck(Addr, &I);
2359 Value *Shadow = IRB.CreateAlignedLoad(getShadowPtr(Addr, Ty, IRB),
2360 Alignment, "_ldmxcsr");
2361 Value *Origin = MS.TrackOrigins
2362 ? IRB.CreateLoad(getOriginPtr(Addr, IRB, Alignment))
2364 insertShadowCheck(Shadow, Origin, &I);
2367 void visitIntrinsicInst(IntrinsicInst &I) {
2368 switch (I.getIntrinsicID()) {
2369 case llvm::Intrinsic::bswap:
2372 case llvm::Intrinsic::x86_sse_stmxcsr:
2375 case llvm::Intrinsic::x86_sse_ldmxcsr:
2378 case llvm::Intrinsic::x86_avx512_vcvtsd2usi64:
2379 case llvm::Intrinsic::x86_avx512_vcvtsd2usi32:
2380 case llvm::Intrinsic::x86_avx512_vcvtss2usi64:
2381 case llvm::Intrinsic::x86_avx512_vcvtss2usi32:
2382 case llvm::Intrinsic::x86_avx512_cvttss2usi64:
2383 case llvm::Intrinsic::x86_avx512_cvttss2usi:
2384 case llvm::Intrinsic::x86_avx512_cvttsd2usi64:
2385 case llvm::Intrinsic::x86_avx512_cvttsd2usi:
2386 case llvm::Intrinsic::x86_avx512_cvtusi2sd:
2387 case llvm::Intrinsic::x86_avx512_cvtusi2ss:
2388 case llvm::Intrinsic::x86_avx512_cvtusi642sd:
2389 case llvm::Intrinsic::x86_avx512_cvtusi642ss:
2390 case llvm::Intrinsic::x86_sse2_cvtsd2si64:
2391 case llvm::Intrinsic::x86_sse2_cvtsd2si:
2392 case llvm::Intrinsic::x86_sse2_cvtsd2ss:
2393 case llvm::Intrinsic::x86_sse2_cvtsi2sd:
2394 case llvm::Intrinsic::x86_sse2_cvtsi642sd:
2395 case llvm::Intrinsic::x86_sse2_cvtss2sd:
2396 case llvm::Intrinsic::x86_sse2_cvttsd2si64:
2397 case llvm::Intrinsic::x86_sse2_cvttsd2si:
2398 case llvm::Intrinsic::x86_sse_cvtsi2ss:
2399 case llvm::Intrinsic::x86_sse_cvtsi642ss:
2400 case llvm::Intrinsic::x86_sse_cvtss2si64:
2401 case llvm::Intrinsic::x86_sse_cvtss2si:
2402 case llvm::Intrinsic::x86_sse_cvttss2si64:
2403 case llvm::Intrinsic::x86_sse_cvttss2si:
2404 handleVectorConvertIntrinsic(I, 1);
2406 case llvm::Intrinsic::x86_sse_cvtps2pi:
2407 case llvm::Intrinsic::x86_sse_cvttps2pi:
2408 handleVectorConvertIntrinsic(I, 2);
2411 case llvm::Intrinsic::x86_avx512_psll_w_512:
2412 case llvm::Intrinsic::x86_avx512_psll_d_512:
2413 case llvm::Intrinsic::x86_avx512_psll_q_512:
2414 case llvm::Intrinsic::x86_avx512_pslli_w_512:
2415 case llvm::Intrinsic::x86_avx512_pslli_d_512:
2416 case llvm::Intrinsic::x86_avx512_pslli_q_512:
2417 case llvm::Intrinsic::x86_avx512_psrl_w_512:
2418 case llvm::Intrinsic::x86_avx512_psrl_d_512:
2419 case llvm::Intrinsic::x86_avx512_psrl_q_512:
2420 case llvm::Intrinsic::x86_avx512_psra_w_512:
2421 case llvm::Intrinsic::x86_avx512_psra_d_512:
2422 case llvm::Intrinsic::x86_avx512_psra_q_512:
2423 case llvm::Intrinsic::x86_avx512_psrli_w_512:
2424 case llvm::Intrinsic::x86_avx512_psrli_d_512:
2425 case llvm::Intrinsic::x86_avx512_psrli_q_512:
2426 case llvm::Intrinsic::x86_avx512_psrai_w_512:
2427 case llvm::Intrinsic::x86_avx512_psrai_d_512:
2428 case llvm::Intrinsic::x86_avx512_psrai_q_512:
2429 case llvm::Intrinsic::x86_avx512_psra_q_256:
2430 case llvm::Intrinsic::x86_avx512_psra_q_128:
2431 case llvm::Intrinsic::x86_avx512_psrai_q_256:
2432 case llvm::Intrinsic::x86_avx512_psrai_q_128:
2433 case llvm::Intrinsic::x86_avx2_psll_w:
2434 case llvm::Intrinsic::x86_avx2_psll_d:
2435 case llvm::Intrinsic::x86_avx2_psll_q:
2436 case llvm::Intrinsic::x86_avx2_pslli_w:
2437 case llvm::Intrinsic::x86_avx2_pslli_d:
2438 case llvm::Intrinsic::x86_avx2_pslli_q:
2439 case llvm::Intrinsic::x86_avx2_psrl_w:
2440 case llvm::Intrinsic::x86_avx2_psrl_d:
2441 case llvm::Intrinsic::x86_avx2_psrl_q:
2442 case llvm::Intrinsic::x86_avx2_psra_w:
2443 case llvm::Intrinsic::x86_avx2_psra_d:
2444 case llvm::Intrinsic::x86_avx2_psrli_w:
2445 case llvm::Intrinsic::x86_avx2_psrli_d:
2446 case llvm::Intrinsic::x86_avx2_psrli_q:
2447 case llvm::Intrinsic::x86_avx2_psrai_w:
2448 case llvm::Intrinsic::x86_avx2_psrai_d:
2449 case llvm::Intrinsic::x86_sse2_psll_w:
2450 case llvm::Intrinsic::x86_sse2_psll_d:
2451 case llvm::Intrinsic::x86_sse2_psll_q:
2452 case llvm::Intrinsic::x86_sse2_pslli_w:
2453 case llvm::Intrinsic::x86_sse2_pslli_d:
2454 case llvm::Intrinsic::x86_sse2_pslli_q:
2455 case llvm::Intrinsic::x86_sse2_psrl_w:
2456 case llvm::Intrinsic::x86_sse2_psrl_d:
2457 case llvm::Intrinsic::x86_sse2_psrl_q:
2458 case llvm::Intrinsic::x86_sse2_psra_w:
2459 case llvm::Intrinsic::x86_sse2_psra_d:
2460 case llvm::Intrinsic::x86_sse2_psrli_w:
2461 case llvm::Intrinsic::x86_sse2_psrli_d:
2462 case llvm::Intrinsic::x86_sse2_psrli_q:
2463 case llvm::Intrinsic::x86_sse2_psrai_w:
2464 case llvm::Intrinsic::x86_sse2_psrai_d:
2465 case llvm::Intrinsic::x86_mmx_psll_w:
2466 case llvm::Intrinsic::x86_mmx_psll_d:
2467 case llvm::Intrinsic::x86_mmx_psll_q:
2468 case llvm::Intrinsic::x86_mmx_pslli_w:
2469 case llvm::Intrinsic::x86_mmx_pslli_d:
2470 case llvm::Intrinsic::x86_mmx_pslli_q:
2471 case llvm::Intrinsic::x86_mmx_psrl_w:
2472 case llvm::Intrinsic::x86_mmx_psrl_d:
2473 case llvm::Intrinsic::x86_mmx_psrl_q:
2474 case llvm::Intrinsic::x86_mmx_psra_w:
2475 case llvm::Intrinsic::x86_mmx_psra_d:
2476 case llvm::Intrinsic::x86_mmx_psrli_w:
2477 case llvm::Intrinsic::x86_mmx_psrli_d:
2478 case llvm::Intrinsic::x86_mmx_psrli_q:
2479 case llvm::Intrinsic::x86_mmx_psrai_w:
2480 case llvm::Intrinsic::x86_mmx_psrai_d:
2481 handleVectorShiftIntrinsic(I, /* Variable */ false);
2483 case llvm::Intrinsic::x86_avx2_psllv_d:
2484 case llvm::Intrinsic::x86_avx2_psllv_d_256:
2485 case llvm::Intrinsic::x86_avx512_psllv_d_512:
2486 case llvm::Intrinsic::x86_avx2_psllv_q:
2487 case llvm::Intrinsic::x86_avx2_psllv_q_256:
2488 case llvm::Intrinsic::x86_avx512_psllv_q_512:
2489 case llvm::Intrinsic::x86_avx2_psrlv_d:
2490 case llvm::Intrinsic::x86_avx2_psrlv_d_256:
2491 case llvm::Intrinsic::x86_avx512_psrlv_d_512:
2492 case llvm::Intrinsic::x86_avx2_psrlv_q:
2493 case llvm::Intrinsic::x86_avx2_psrlv_q_256:
2494 case llvm::Intrinsic::x86_avx512_psrlv_q_512:
2495 case llvm::Intrinsic::x86_avx2_psrav_d:
2496 case llvm::Intrinsic::x86_avx2_psrav_d_256:
2497 case llvm::Intrinsic::x86_avx512_psrav_d_512:
2498 case llvm::Intrinsic::x86_avx512_psrav_q_128:
2499 case llvm::Intrinsic::x86_avx512_psrav_q_256:
2500 case llvm::Intrinsic::x86_avx512_psrav_q_512:
2501 handleVectorShiftIntrinsic(I, /* Variable */ true);
2504 case llvm::Intrinsic::x86_sse2_packsswb_128:
2505 case llvm::Intrinsic::x86_sse2_packssdw_128:
2506 case llvm::Intrinsic::x86_sse2_packuswb_128:
2507 case llvm::Intrinsic::x86_sse41_packusdw:
2508 case llvm::Intrinsic::x86_avx2_packsswb:
2509 case llvm::Intrinsic::x86_avx2_packssdw:
2510 case llvm::Intrinsic::x86_avx2_packuswb:
2511 case llvm::Intrinsic::x86_avx2_packusdw:
2512 handleVectorPackIntrinsic(I);
2515 case llvm::Intrinsic::x86_mmx_packsswb:
2516 case llvm::Intrinsic::x86_mmx_packuswb:
2517 handleVectorPackIntrinsic(I, 16);
2520 case llvm::Intrinsic::x86_mmx_packssdw:
2521 handleVectorPackIntrinsic(I, 32);
2524 case llvm::Intrinsic::x86_mmx_psad_bw:
2525 case llvm::Intrinsic::x86_sse2_psad_bw:
2526 case llvm::Intrinsic::x86_avx2_psad_bw:
2527 handleVectorSadIntrinsic(I);
2530 case llvm::Intrinsic::x86_sse2_pmadd_wd:
2531 case llvm::Intrinsic::x86_avx2_pmadd_wd:
2532 case llvm::Intrinsic::x86_ssse3_pmadd_ub_sw_128:
2533 case llvm::Intrinsic::x86_avx2_pmadd_ub_sw:
2534 handleVectorPmaddIntrinsic(I);
2537 case llvm::Intrinsic::x86_ssse3_pmadd_ub_sw:
2538 handleVectorPmaddIntrinsic(I, 8);
2541 case llvm::Intrinsic::x86_mmx_pmadd_wd:
2542 handleVectorPmaddIntrinsic(I, 16);
2545 case llvm::Intrinsic::x86_sse_cmp_ss:
2546 case llvm::Intrinsic::x86_sse2_cmp_sd:
2547 case llvm::Intrinsic::x86_sse_comieq_ss:
2548 case llvm::Intrinsic::x86_sse_comilt_ss:
2549 case llvm::Intrinsic::x86_sse_comile_ss:
2550 case llvm::Intrinsic::x86_sse_comigt_ss:
2551 case llvm::Intrinsic::x86_sse_comige_ss:
2552 case llvm::Intrinsic::x86_sse_comineq_ss:
2553 case llvm::Intrinsic::x86_sse_ucomieq_ss:
2554 case llvm::Intrinsic::x86_sse_ucomilt_ss:
2555 case llvm::Intrinsic::x86_sse_ucomile_ss:
2556 case llvm::Intrinsic::x86_sse_ucomigt_ss:
2557 case llvm::Intrinsic::x86_sse_ucomige_ss:
2558 case llvm::Intrinsic::x86_sse_ucomineq_ss:
2559 case llvm::Intrinsic::x86_sse2_comieq_sd:
2560 case llvm::Intrinsic::x86_sse2_comilt_sd:
2561 case llvm::Intrinsic::x86_sse2_comile_sd:
2562 case llvm::Intrinsic::x86_sse2_comigt_sd:
2563 case llvm::Intrinsic::x86_sse2_comige_sd:
2564 case llvm::Intrinsic::x86_sse2_comineq_sd:
2565 case llvm::Intrinsic::x86_sse2_ucomieq_sd:
2566 case llvm::Intrinsic::x86_sse2_ucomilt_sd:
2567 case llvm::Intrinsic::x86_sse2_ucomile_sd:
2568 case llvm::Intrinsic::x86_sse2_ucomigt_sd:
2569 case llvm::Intrinsic::x86_sse2_ucomige_sd:
2570 case llvm::Intrinsic::x86_sse2_ucomineq_sd:
2571 handleVectorCompareScalarIntrinsic(I);
2574 case llvm::Intrinsic::x86_sse_cmp_ps:
2575 case llvm::Intrinsic::x86_sse2_cmp_pd:
2576 // FIXME: For x86_avx_cmp_pd_256 and x86_avx_cmp_ps_256 this function
2577 // generates reasonably looking IR that fails in the backend with "Do not
2578 // know how to split the result of this operator!".
2579 handleVectorComparePackedIntrinsic(I);
2583 if (!handleUnknownIntrinsic(I))
2584 visitInstruction(I);
2589 void visitCallSite(CallSite CS) {
2590 Instruction &I = *CS.getInstruction();
2591 assert((CS.isCall() || CS.isInvoke()) && "Unknown type of CallSite");
2593 CallInst *Call = cast<CallInst>(&I);
2595 // For inline asm, do the usual thing: check argument shadow and mark all
2596 // outputs as clean. Note that any side effects of the inline asm that are
2597 // not immediately visible in its constraints are not handled.
2598 if (Call->isInlineAsm()) {
2599 visitInstruction(I);
2603 assert(!isa<IntrinsicInst>(&I) && "intrinsics are handled elsewhere");
2605 // We are going to insert code that relies on the fact that the callee
2606 // will become a non-readonly function after it is instrumented by us. To
2607 // prevent this code from being optimized out, mark that function
2608 // non-readonly in advance.
2609 if (Function *Func = Call->getCalledFunction()) {
2610 // Clear out readonly/readnone attributes.
2612 B.addAttribute(Attribute::ReadOnly)
2613 .addAttribute(Attribute::ReadNone);
2614 Func->removeAttributes(AttributeList::FunctionIndex, B);
2617 maybeMarkSanitizerLibraryCallNoBuiltin(Call, TLI);
2619 IRBuilder<> IRB(&I);
2621 unsigned ArgOffset = 0;
2622 DEBUG(dbgs() << " CallSite: " << I << "\n");
2623 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
2624 ArgIt != End; ++ArgIt) {
2626 unsigned i = ArgIt - CS.arg_begin();
2627 if (!A->getType()->isSized()) {
2628 DEBUG(dbgs() << "Arg " << i << " is not sized: " << I << "\n");
2632 Value *Store = nullptr;
2633 // Compute the Shadow for arg even if it is ByVal, because
2634 // in that case getShadow() will copy the actual arg shadow to
2635 // __msan_param_tls.
2636 Value *ArgShadow = getShadow(A);
2637 Value *ArgShadowBase = getShadowPtrForArgument(A, IRB, ArgOffset);
2638 DEBUG(dbgs() << " Arg#" << i << ": " << *A <<
2639 " Shadow: " << *ArgShadow << "\n");
2640 bool ArgIsInitialized = false;
2641 const DataLayout &DL = F.getParent()->getDataLayout();
2642 if (CS.paramHasAttr(i, Attribute::ByVal)) {
2643 assert(A->getType()->isPointerTy() &&
2644 "ByVal argument is not a pointer!");
2645 Size = DL.getTypeAllocSize(A->getType()->getPointerElementType());
2646 if (ArgOffset + Size > kParamTLSSize) break;
2647 unsigned ParamAlignment = CS.getParamAlignment(i);
2648 unsigned Alignment = std::min(ParamAlignment, kShadowTLSAlignment);
2649 Store = IRB.CreateMemCpy(ArgShadowBase,
2650 getShadowPtr(A, Type::getInt8Ty(*MS.C), IRB),
2653 Size = DL.getTypeAllocSize(A->getType());
2654 if (ArgOffset + Size > kParamTLSSize) break;
2655 Store = IRB.CreateAlignedStore(ArgShadow, ArgShadowBase,
2656 kShadowTLSAlignment);
2657 Constant *Cst = dyn_cast<Constant>(ArgShadow);
2658 if (Cst && Cst->isNullValue()) ArgIsInitialized = true;
2660 if (MS.TrackOrigins && !ArgIsInitialized)
2661 IRB.CreateStore(getOrigin(A),
2662 getOriginPtrForArgument(A, IRB, ArgOffset));
2664 assert(Size != 0 && Store != nullptr);
2665 DEBUG(dbgs() << " Param:" << *Store << "\n");
2666 ArgOffset += alignTo(Size, 8);
2668 DEBUG(dbgs() << " done with call args\n");
2671 cast<FunctionType>(CS.getCalledValue()->getType()->getContainedType(0));
2672 if (FT->isVarArg()) {
2673 VAHelper->visitCallSite(CS, IRB);
2676 // Now, get the shadow for the RetVal.
2677 if (!I.getType()->isSized()) return;
2678 // Don't emit the epilogue for musttail call returns.
2679 if (CS.isCall() && cast<CallInst>(&I)->isMustTailCall()) return;
2680 IRBuilder<> IRBBefore(&I);
2681 // Until we have full dynamic coverage, make sure the retval shadow is 0.
2682 Value *Base = getShadowPtrForRetval(&I, IRBBefore);
2683 IRBBefore.CreateAlignedStore(getCleanShadow(&I), Base, kShadowTLSAlignment);
2684 BasicBlock::iterator NextInsn;
2686 NextInsn = ++I.getIterator();
2687 assert(NextInsn != I.getParent()->end());
2689 BasicBlock *NormalDest = cast<InvokeInst>(&I)->getNormalDest();
2690 if (!NormalDest->getSinglePredecessor()) {
2691 // FIXME: this case is tricky, so we are just conservative here.
2692 // Perhaps we need to split the edge between this BB and NormalDest,
2693 // but a naive attempt to use SplitEdge leads to a crash.
2694 setShadow(&I, getCleanShadow(&I));
2695 setOrigin(&I, getCleanOrigin());
2698 NextInsn = NormalDest->getFirstInsertionPt();
2699 assert(NextInsn != NormalDest->end() &&
2700 "Could not find insertion point for retval shadow load");
2702 IRBuilder<> IRBAfter(&*NextInsn);
2703 Value *RetvalShadow =
2704 IRBAfter.CreateAlignedLoad(getShadowPtrForRetval(&I, IRBAfter),
2705 kShadowTLSAlignment, "_msret");
2706 setShadow(&I, RetvalShadow);
2707 if (MS.TrackOrigins)
2708 setOrigin(&I, IRBAfter.CreateLoad(getOriginPtrForRetval(IRBAfter)));
2711 bool isAMustTailRetVal(Value *RetVal) {
2712 if (auto *I = dyn_cast<BitCastInst>(RetVal)) {
2713 RetVal = I->getOperand(0);
2715 if (auto *I = dyn_cast<CallInst>(RetVal)) {
2716 return I->isMustTailCall();
2721 void visitReturnInst(ReturnInst &I) {
2722 IRBuilder<> IRB(&I);
2723 Value *RetVal = I.getReturnValue();
2724 if (!RetVal) return;
2725 // Don't emit the epilogue for musttail call returns.
2726 if (isAMustTailRetVal(RetVal)) return;
2727 Value *ShadowPtr = getShadowPtrForRetval(RetVal, IRB);
2728 if (CheckReturnValue) {
2729 insertShadowCheck(RetVal, &I);
2730 Value *Shadow = getCleanShadow(RetVal);
2731 IRB.CreateAlignedStore(Shadow, ShadowPtr, kShadowTLSAlignment);
2733 Value *Shadow = getShadow(RetVal);
2734 IRB.CreateAlignedStore(Shadow, ShadowPtr, kShadowTLSAlignment);
2735 if (MS.TrackOrigins)
2736 IRB.CreateStore(getOrigin(RetVal), getOriginPtrForRetval(IRB));
2740 void visitPHINode(PHINode &I) {
2741 IRBuilder<> IRB(&I);
2742 if (!PropagateShadow) {
2743 setShadow(&I, getCleanShadow(&I));
2744 setOrigin(&I, getCleanOrigin());
2748 ShadowPHINodes.push_back(&I);
2749 setShadow(&I, IRB.CreatePHI(getShadowTy(&I), I.getNumIncomingValues(),
2751 if (MS.TrackOrigins)
2752 setOrigin(&I, IRB.CreatePHI(MS.OriginTy, I.getNumIncomingValues(),
2756 void visitAllocaInst(AllocaInst &I) {
2757 setShadow(&I, getCleanShadow(&I));
2758 setOrigin(&I, getCleanOrigin());
2759 IRBuilder<> IRB(I.getNextNode());
2760 const DataLayout &DL = F.getParent()->getDataLayout();
2761 uint64_t TypeSize = DL.getTypeAllocSize(I.getAllocatedType());
2762 Value *Len = ConstantInt::get(MS.IntptrTy, TypeSize);
2763 if (I.isArrayAllocation())
2764 Len = IRB.CreateMul(Len, I.getArraySize());
2765 if (PoisonStack && ClPoisonStackWithCall) {
2766 IRB.CreateCall(MS.MsanPoisonStackFn,
2767 {IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()), Len});
2769 Value *ShadowBase = getShadowPtr(&I, Type::getInt8PtrTy(*MS.C), IRB);
2770 Value *PoisonValue = IRB.getInt8(PoisonStack ? ClPoisonStackPattern : 0);
2771 IRB.CreateMemSet(ShadowBase, PoisonValue, Len, I.getAlignment());
2774 if (PoisonStack && MS.TrackOrigins) {
2775 SmallString<2048> StackDescriptionStorage;
2776 raw_svector_ostream StackDescription(StackDescriptionStorage);
2777 // We create a string with a description of the stack allocation and
2778 // pass it into __msan_set_alloca_origin.
2779 // It will be printed by the run-time if stack-originated UMR is found.
2780 // The first 4 bytes of the string are set to '----' and will be replaced
2781 // by __msan_va_arg_overflow_size_tls at the first call.
2782 StackDescription << "----" << I.getName() << "@" << F.getName();
2784 createPrivateNonConstGlobalForString(*F.getParent(),
2785 StackDescription.str());
2787 IRB.CreateCall(MS.MsanSetAllocaOrigin4Fn,
2788 {IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()), Len,
2789 IRB.CreatePointerCast(Descr, IRB.getInt8PtrTy()),
2790 IRB.CreatePointerCast(&F, MS.IntptrTy)});
2794 void visitSelectInst(SelectInst& I) {
2795 IRBuilder<> IRB(&I);
2796 // a = select b, c, d
2797 Value *B = I.getCondition();
2798 Value *C = I.getTrueValue();
2799 Value *D = I.getFalseValue();
2800 Value *Sb = getShadow(B);
2801 Value *Sc = getShadow(C);
2802 Value *Sd = getShadow(D);
2804 // Result shadow if condition shadow is 0.
2805 Value *Sa0 = IRB.CreateSelect(B, Sc, Sd);
2807 if (I.getType()->isAggregateType()) {
2808 // To avoid "sign extending" i1 to an arbitrary aggregate type, we just do
2809 // an extra "select". This results in much more compact IR.
2810 // Sa = select Sb, poisoned, (select b, Sc, Sd)
2811 Sa1 = getPoisonedShadow(getShadowTy(I.getType()));
2813 // Sa = select Sb, [ (c^d) | Sc | Sd ], [ b ? Sc : Sd ]
2814 // If Sb (condition is poisoned), look for bits in c and d that are equal
2815 // and both unpoisoned.
2816 // If !Sb (condition is unpoisoned), simply pick one of Sc and Sd.
2818 // Cast arguments to shadow-compatible type.
2819 C = CreateAppToShadowCast(IRB, C);
2820 D = CreateAppToShadowCast(IRB, D);
2822 // Result shadow if condition shadow is 1.
2823 Sa1 = IRB.CreateOr(IRB.CreateXor(C, D), IRB.CreateOr(Sc, Sd));
2825 Value *Sa = IRB.CreateSelect(Sb, Sa1, Sa0, "_msprop_select");
2827 if (MS.TrackOrigins) {
2828 // Origins are always i32, so any vector conditions must be flattened.
2829 // FIXME: consider tracking vector origins for app vectors?
2830 if (B->getType()->isVectorTy()) {
2831 Type *FlatTy = getShadowTyNoVec(B->getType());
2832 B = IRB.CreateICmpNE(IRB.CreateBitCast(B, FlatTy),
2833 ConstantInt::getNullValue(FlatTy));
2834 Sb = IRB.CreateICmpNE(IRB.CreateBitCast(Sb, FlatTy),
2835 ConstantInt::getNullValue(FlatTy));
2837 // a = select b, c, d
2838 // Oa = Sb ? Ob : (b ? Oc : Od)
2840 &I, IRB.CreateSelect(Sb, getOrigin(I.getCondition()),
2841 IRB.CreateSelect(B, getOrigin(I.getTrueValue()),
2842 getOrigin(I.getFalseValue()))));
2846 void visitLandingPadInst(LandingPadInst &I) {
2848 // See http://code.google.com/p/memory-sanitizer/issues/detail?id=1
2849 setShadow(&I, getCleanShadow(&I));
2850 setOrigin(&I, getCleanOrigin());
2853 void visitCatchSwitchInst(CatchSwitchInst &I) {
2854 setShadow(&I, getCleanShadow(&I));
2855 setOrigin(&I, getCleanOrigin());
2858 void visitFuncletPadInst(FuncletPadInst &I) {
2859 setShadow(&I, getCleanShadow(&I));
2860 setOrigin(&I, getCleanOrigin());
2863 void visitGetElementPtrInst(GetElementPtrInst &I) {
2867 void visitExtractValueInst(ExtractValueInst &I) {
2868 IRBuilder<> IRB(&I);
2869 Value *Agg = I.getAggregateOperand();
2870 DEBUG(dbgs() << "ExtractValue: " << I << "\n");
2871 Value *AggShadow = getShadow(Agg);
2872 DEBUG(dbgs() << " AggShadow: " << *AggShadow << "\n");
2873 Value *ResShadow = IRB.CreateExtractValue(AggShadow, I.getIndices());
2874 DEBUG(dbgs() << " ResShadow: " << *ResShadow << "\n");
2875 setShadow(&I, ResShadow);
2876 setOriginForNaryOp(I);
2879 void visitInsertValueInst(InsertValueInst &I) {
2880 IRBuilder<> IRB(&I);
2881 DEBUG(dbgs() << "InsertValue: " << I << "\n");
2882 Value *AggShadow = getShadow(I.getAggregateOperand());
2883 Value *InsShadow = getShadow(I.getInsertedValueOperand());
2884 DEBUG(dbgs() << " AggShadow: " << *AggShadow << "\n");
2885 DEBUG(dbgs() << " InsShadow: " << *InsShadow << "\n");
2886 Value *Res = IRB.CreateInsertValue(AggShadow, InsShadow, I.getIndices());
2887 DEBUG(dbgs() << " Res: " << *Res << "\n");
2889 setOriginForNaryOp(I);
2892 void dumpInst(Instruction &I) {
2893 if (CallInst *CI = dyn_cast<CallInst>(&I)) {
2894 errs() << "ZZZ call " << CI->getCalledFunction()->getName() << "\n";
2896 errs() << "ZZZ " << I.getOpcodeName() << "\n";
2898 errs() << "QQQ " << I << "\n";
2901 void visitResumeInst(ResumeInst &I) {
2902 DEBUG(dbgs() << "Resume: " << I << "\n");
2903 // Nothing to do here.
2906 void visitCleanupReturnInst(CleanupReturnInst &CRI) {
2907 DEBUG(dbgs() << "CleanupReturn: " << CRI << "\n");
2908 // Nothing to do here.
2911 void visitCatchReturnInst(CatchReturnInst &CRI) {
2912 DEBUG(dbgs() << "CatchReturn: " << CRI << "\n");
2913 // Nothing to do here.
2916 void visitInstruction(Instruction &I) {
2917 // Everything else: stop propagating and check for poisoned shadow.
2918 if (ClDumpStrictInstructions)
2920 DEBUG(dbgs() << "DEFAULT: " << I << "\n");
2921 for (size_t i = 0, n = I.getNumOperands(); i < n; i++) {
2922 Value *Operand = I.getOperand(i);
2923 if (Operand->getType()->isSized())
2924 insertShadowCheck(Operand, &I);
2926 setShadow(&I, getCleanShadow(&I));
2927 setOrigin(&I, getCleanOrigin());
2931 /// \brief AMD64-specific implementation of VarArgHelper.
2932 struct VarArgAMD64Helper : public VarArgHelper {
2933 // An unfortunate workaround for asymmetric lowering of va_arg stuff.
2934 // See a comment in visitCallSite for more details.
2935 static const unsigned AMD64GpEndOffset = 48; // AMD64 ABI Draft 0.99.6 p3.5.7
2936 static const unsigned AMD64FpEndOffset = 176;
2939 MemorySanitizer &MS;
2940 MemorySanitizerVisitor &MSV;
2941 Value *VAArgTLSCopy;
2942 Value *VAArgOverflowSize;
2944 SmallVector<CallInst*, 16> VAStartInstrumentationList;
2946 VarArgAMD64Helper(Function &F, MemorySanitizer &MS,
2947 MemorySanitizerVisitor &MSV)
2948 : F(F), MS(MS), MSV(MSV), VAArgTLSCopy(nullptr),
2949 VAArgOverflowSize(nullptr) {}
2951 enum ArgKind { AK_GeneralPurpose, AK_FloatingPoint, AK_Memory };
2953 ArgKind classifyArgument(Value* arg) {
2954 // A very rough approximation of X86_64 argument classification rules.
2955 Type *T = arg->getType();
2956 if (T->isFPOrFPVectorTy() || T->isX86_MMXTy())
2957 return AK_FloatingPoint;
2958 if (T->isIntegerTy() && T->getPrimitiveSizeInBits() <= 64)
2959 return AK_GeneralPurpose;
2960 if (T->isPointerTy())
2961 return AK_GeneralPurpose;
2965 // For VarArg functions, store the argument shadow in an ABI-specific format
2966 // that corresponds to va_list layout.
2967 // We do this because Clang lowers va_arg in the frontend, and this pass
2968 // only sees the low level code that deals with va_list internals.
2969 // A much easier alternative (provided that Clang emits va_arg instructions)
2970 // would have been to associate each live instance of va_list with a copy of
2971 // MSanParamTLS, and extract shadow on va_arg() call in the argument list
2973 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) override {
2974 unsigned GpOffset = 0;
2975 unsigned FpOffset = AMD64GpEndOffset;
2976 unsigned OverflowOffset = AMD64FpEndOffset;
2977 const DataLayout &DL = F.getParent()->getDataLayout();
2978 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
2979 ArgIt != End; ++ArgIt) {
2981 unsigned ArgNo = CS.getArgumentNo(ArgIt);
2982 bool IsFixed = ArgNo < CS.getFunctionType()->getNumParams();
2983 bool IsByVal = CS.paramHasAttr(ArgNo, Attribute::ByVal);
2985 // ByVal arguments always go to the overflow area.
2986 // Fixed arguments passed through the overflow area will be stepped
2987 // over by va_start, so don't count them towards the offset.
2990 assert(A->getType()->isPointerTy());
2991 Type *RealTy = A->getType()->getPointerElementType();
2992 uint64_t ArgSize = DL.getTypeAllocSize(RealTy);
2993 Value *Base = getShadowPtrForVAArgument(RealTy, IRB, OverflowOffset);
2994 OverflowOffset += alignTo(ArgSize, 8);
2995 IRB.CreateMemCpy(Base, MSV.getShadowPtr(A, IRB.getInt8Ty(), IRB),
2996 ArgSize, kShadowTLSAlignment);
2998 ArgKind AK = classifyArgument(A);
2999 if (AK == AK_GeneralPurpose && GpOffset >= AMD64GpEndOffset)
3001 if (AK == AK_FloatingPoint && FpOffset >= AMD64FpEndOffset)
3005 case AK_GeneralPurpose:
3006 Base = getShadowPtrForVAArgument(A->getType(), IRB, GpOffset);
3009 case AK_FloatingPoint:
3010 Base = getShadowPtrForVAArgument(A->getType(), IRB, FpOffset);
3016 uint64_t ArgSize = DL.getTypeAllocSize(A->getType());
3017 Base = getShadowPtrForVAArgument(A->getType(), IRB, OverflowOffset);
3018 OverflowOffset += alignTo(ArgSize, 8);
3020 // Take fixed arguments into account for GpOffset and FpOffset,
3021 // but don't actually store shadows for them.
3024 IRB.CreateAlignedStore(MSV.getShadow(A), Base, kShadowTLSAlignment);
3027 Constant *OverflowSize =
3028 ConstantInt::get(IRB.getInt64Ty(), OverflowOffset - AMD64FpEndOffset);
3029 IRB.CreateStore(OverflowSize, MS.VAArgOverflowSizeTLS);
3032 /// \brief Compute the shadow address for a given va_arg.
3033 Value *getShadowPtrForVAArgument(Type *Ty, IRBuilder<> &IRB,
3035 Value *Base = IRB.CreatePointerCast(MS.VAArgTLS, MS.IntptrTy);
3036 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
3037 return IRB.CreateIntToPtr(Base, PointerType::get(MSV.getShadowTy(Ty), 0),
3041 void visitVAStartInst(VAStartInst &I) override {
3042 if (F.getCallingConv() == CallingConv::Win64)
3044 IRBuilder<> IRB(&I);
3045 VAStartInstrumentationList.push_back(&I);
3046 Value *VAListTag = I.getArgOperand(0);
3047 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
3049 // Unpoison the whole __va_list_tag.
3050 // FIXME: magic ABI constants.
3051 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
3052 /* size */24, /* alignment */8, false);
3055 void visitVACopyInst(VACopyInst &I) override {
3056 if (F.getCallingConv() == CallingConv::Win64)
3058 IRBuilder<> IRB(&I);
3059 Value *VAListTag = I.getArgOperand(0);
3060 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
3062 // Unpoison the whole __va_list_tag.
3063 // FIXME: magic ABI constants.
3064 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
3065 /* size */24, /* alignment */8, false);
3068 void finalizeInstrumentation() override {
3069 assert(!VAArgOverflowSize && !VAArgTLSCopy &&
3070 "finalizeInstrumentation called twice");
3071 if (!VAStartInstrumentationList.empty()) {
3072 // If there is a va_start in this function, make a backup copy of
3073 // va_arg_tls somewhere in the function entry block.
3074 IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
3075 VAArgOverflowSize = IRB.CreateLoad(MS.VAArgOverflowSizeTLS);
3077 IRB.CreateAdd(ConstantInt::get(MS.IntptrTy, AMD64FpEndOffset),
3079 VAArgTLSCopy = IRB.CreateAlloca(Type::getInt8Ty(*MS.C), CopySize);
3080 IRB.CreateMemCpy(VAArgTLSCopy, MS.VAArgTLS, CopySize, 8);
3083 // Instrument va_start.
3084 // Copy va_list shadow from the backup copy of the TLS contents.
3085 for (size_t i = 0, n = VAStartInstrumentationList.size(); i < n; i++) {
3086 CallInst *OrigInst = VAStartInstrumentationList[i];
3087 IRBuilder<> IRB(OrigInst->getNextNode());
3088 Value *VAListTag = OrigInst->getArgOperand(0);
3090 Value *RegSaveAreaPtrPtr =
3092 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
3093 ConstantInt::get(MS.IntptrTy, 16)),
3094 Type::getInt64PtrTy(*MS.C));
3095 Value *RegSaveAreaPtr = IRB.CreateLoad(RegSaveAreaPtrPtr);
3096 Value *RegSaveAreaShadowPtr =
3097 MSV.getShadowPtr(RegSaveAreaPtr, IRB.getInt8Ty(), IRB);
3098 IRB.CreateMemCpy(RegSaveAreaShadowPtr, VAArgTLSCopy,
3099 AMD64FpEndOffset, 16);
3101 Value *OverflowArgAreaPtrPtr =
3103 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
3104 ConstantInt::get(MS.IntptrTy, 8)),
3105 Type::getInt64PtrTy(*MS.C));
3106 Value *OverflowArgAreaPtr = IRB.CreateLoad(OverflowArgAreaPtrPtr);
3107 Value *OverflowArgAreaShadowPtr =
3108 MSV.getShadowPtr(OverflowArgAreaPtr, IRB.getInt8Ty(), IRB);
3109 Value *SrcPtr = IRB.CreateConstGEP1_32(IRB.getInt8Ty(), VAArgTLSCopy,
3111 IRB.CreateMemCpy(OverflowArgAreaShadowPtr, SrcPtr, VAArgOverflowSize, 16);
3116 /// \brief MIPS64-specific implementation of VarArgHelper.
3117 struct VarArgMIPS64Helper : public VarArgHelper {
3119 MemorySanitizer &MS;
3120 MemorySanitizerVisitor &MSV;
3121 Value *VAArgTLSCopy;
3124 SmallVector<CallInst*, 16> VAStartInstrumentationList;
3126 VarArgMIPS64Helper(Function &F, MemorySanitizer &MS,
3127 MemorySanitizerVisitor &MSV)
3128 : F(F), MS(MS), MSV(MSV), VAArgTLSCopy(nullptr),
3129 VAArgSize(nullptr) {}
3131 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) override {
3132 unsigned VAArgOffset = 0;
3133 const DataLayout &DL = F.getParent()->getDataLayout();
3134 for (CallSite::arg_iterator ArgIt = CS.arg_begin() +
3135 CS.getFunctionType()->getNumParams(), End = CS.arg_end();
3136 ArgIt != End; ++ArgIt) {
3137 llvm::Triple TargetTriple(F.getParent()->getTargetTriple());
3140 uint64_t ArgSize = DL.getTypeAllocSize(A->getType());
3141 if (TargetTriple.getArch() == llvm::Triple::mips64) {
3142 // Adjusting the shadow for argument with size < 8 to match the placement
3143 // of bits in big endian system
3145 VAArgOffset += (8 - ArgSize);
3147 Base = getShadowPtrForVAArgument(A->getType(), IRB, VAArgOffset);
3148 VAArgOffset += ArgSize;
3149 VAArgOffset = alignTo(VAArgOffset, 8);
3150 IRB.CreateAlignedStore(MSV.getShadow(A), Base, kShadowTLSAlignment);
3153 Constant *TotalVAArgSize = ConstantInt::get(IRB.getInt64Ty(), VAArgOffset);
3154 // Here using VAArgOverflowSizeTLS as VAArgSizeTLS to avoid creation of
3155 // a new class member i.e. it is the total size of all VarArgs.
3156 IRB.CreateStore(TotalVAArgSize, MS.VAArgOverflowSizeTLS);
3159 /// \brief Compute the shadow address for a given va_arg.
3160 Value *getShadowPtrForVAArgument(Type *Ty, IRBuilder<> &IRB,
3162 Value *Base = IRB.CreatePointerCast(MS.VAArgTLS, MS.IntptrTy);
3163 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
3164 return IRB.CreateIntToPtr(Base, PointerType::get(MSV.getShadowTy(Ty), 0),
3168 void visitVAStartInst(VAStartInst &I) override {
3169 IRBuilder<> IRB(&I);
3170 VAStartInstrumentationList.push_back(&I);
3171 Value *VAListTag = I.getArgOperand(0);
3172 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
3173 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
3174 /* size */8, /* alignment */8, false);
3177 void visitVACopyInst(VACopyInst &I) override {
3178 IRBuilder<> IRB(&I);
3179 Value *VAListTag = I.getArgOperand(0);
3180 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
3181 // Unpoison the whole __va_list_tag.
3182 // FIXME: magic ABI constants.
3183 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
3184 /* size */8, /* alignment */8, false);
3187 void finalizeInstrumentation() override {
3188 assert(!VAArgSize && !VAArgTLSCopy &&
3189 "finalizeInstrumentation called twice");
3190 IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
3191 VAArgSize = IRB.CreateLoad(MS.VAArgOverflowSizeTLS);
3192 Value *CopySize = IRB.CreateAdd(ConstantInt::get(MS.IntptrTy, 0),
3195 if (!VAStartInstrumentationList.empty()) {
3196 // If there is a va_start in this function, make a backup copy of
3197 // va_arg_tls somewhere in the function entry block.
3198 VAArgTLSCopy = IRB.CreateAlloca(Type::getInt8Ty(*MS.C), CopySize);
3199 IRB.CreateMemCpy(VAArgTLSCopy, MS.VAArgTLS, CopySize, 8);
3202 // Instrument va_start.
3203 // Copy va_list shadow from the backup copy of the TLS contents.
3204 for (size_t i = 0, n = VAStartInstrumentationList.size(); i < n; i++) {
3205 CallInst *OrigInst = VAStartInstrumentationList[i];
3206 IRBuilder<> IRB(OrigInst->getNextNode());
3207 Value *VAListTag = OrigInst->getArgOperand(0);
3208 Value *RegSaveAreaPtrPtr =
3209 IRB.CreateIntToPtr(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
3210 Type::getInt64PtrTy(*MS.C));
3211 Value *RegSaveAreaPtr = IRB.CreateLoad(RegSaveAreaPtrPtr);
3212 Value *RegSaveAreaShadowPtr =
3213 MSV.getShadowPtr(RegSaveAreaPtr, IRB.getInt8Ty(), IRB);
3214 IRB.CreateMemCpy(RegSaveAreaShadowPtr, VAArgTLSCopy, CopySize, 8);
3220 /// \brief AArch64-specific implementation of VarArgHelper.
3221 struct VarArgAArch64Helper : public VarArgHelper {
3222 static const unsigned kAArch64GrArgSize = 64;
3223 static const unsigned kAArch64VrArgSize = 128;
3225 static const unsigned AArch64GrBegOffset = 0;
3226 static const unsigned AArch64GrEndOffset = kAArch64GrArgSize;
3227 // Make VR space aligned to 16 bytes.
3228 static const unsigned AArch64VrBegOffset = AArch64GrEndOffset;
3229 static const unsigned AArch64VrEndOffset = AArch64VrBegOffset
3230 + kAArch64VrArgSize;
3231 static const unsigned AArch64VAEndOffset = AArch64VrEndOffset;
3234 MemorySanitizer &MS;
3235 MemorySanitizerVisitor &MSV;
3236 Value *VAArgTLSCopy;
3237 Value *VAArgOverflowSize;
3239 SmallVector<CallInst*, 16> VAStartInstrumentationList;
3241 VarArgAArch64Helper(Function &F, MemorySanitizer &MS,
3242 MemorySanitizerVisitor &MSV)
3243 : F(F), MS(MS), MSV(MSV), VAArgTLSCopy(nullptr),
3244 VAArgOverflowSize(nullptr) {}
3246 enum ArgKind { AK_GeneralPurpose, AK_FloatingPoint, AK_Memory };
3248 ArgKind classifyArgument(Value* arg) {
3249 Type *T = arg->getType();
3250 if (T->isFPOrFPVectorTy())
3251 return AK_FloatingPoint;
3252 if ((T->isIntegerTy() && T->getPrimitiveSizeInBits() <= 64)
3253 || (T->isPointerTy()))
3254 return AK_GeneralPurpose;
3258 // The instrumentation stores the argument shadow in a non ABI-specific
3259 // format because it does not know which argument is named (since Clang,
3260 // like x86_64 case, lowers the va_args in the frontend and this pass only
3261 // sees the low level code that deals with va_list internals).
3262 // The first seven GR registers are saved in the first 56 bytes of the
3263 // va_arg tls arra, followers by the first 8 FP/SIMD registers, and then
3264 // the remaining arguments.
3265 // Using constant offset within the va_arg TLS array allows fast copy
3266 // in the finalize instrumentation.
3267 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) override {
3268 unsigned GrOffset = AArch64GrBegOffset;
3269 unsigned VrOffset = AArch64VrBegOffset;
3270 unsigned OverflowOffset = AArch64VAEndOffset;
3272 const DataLayout &DL = F.getParent()->getDataLayout();
3273 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
3274 ArgIt != End; ++ArgIt) {
3276 unsigned ArgNo = CS.getArgumentNo(ArgIt);
3277 bool IsFixed = ArgNo < CS.getFunctionType()->getNumParams();
3278 ArgKind AK = classifyArgument(A);
3279 if (AK == AK_GeneralPurpose && GrOffset >= AArch64GrEndOffset)
3281 if (AK == AK_FloatingPoint && VrOffset >= AArch64VrEndOffset)
3285 case AK_GeneralPurpose:
3286 Base = getShadowPtrForVAArgument(A->getType(), IRB, GrOffset);
3289 case AK_FloatingPoint:
3290 Base = getShadowPtrForVAArgument(A->getType(), IRB, VrOffset);
3294 // Don't count fixed arguments in the overflow area - va_start will
3295 // skip right over them.
3298 uint64_t ArgSize = DL.getTypeAllocSize(A->getType());
3299 Base = getShadowPtrForVAArgument(A->getType(), IRB, OverflowOffset);
3300 OverflowOffset += alignTo(ArgSize, 8);
3303 // Count Gp/Vr fixed arguments to their respective offsets, but don't
3304 // bother to actually store a shadow.
3307 IRB.CreateAlignedStore(MSV.getShadow(A), Base, kShadowTLSAlignment);
3309 Constant *OverflowSize =
3310 ConstantInt::get(IRB.getInt64Ty(), OverflowOffset - AArch64VAEndOffset);
3311 IRB.CreateStore(OverflowSize, MS.VAArgOverflowSizeTLS);
3314 /// Compute the shadow address for a given va_arg.
3315 Value *getShadowPtrForVAArgument(Type *Ty, IRBuilder<> &IRB,
3317 Value *Base = IRB.CreatePointerCast(MS.VAArgTLS, MS.IntptrTy);
3318 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
3319 return IRB.CreateIntToPtr(Base, PointerType::get(MSV.getShadowTy(Ty), 0),
3323 void visitVAStartInst(VAStartInst &I) override {
3324 IRBuilder<> IRB(&I);
3325 VAStartInstrumentationList.push_back(&I);
3326 Value *VAListTag = I.getArgOperand(0);
3327 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
3328 // Unpoison the whole __va_list_tag.
3329 // FIXME: magic ABI constants (size of va_list).
3330 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
3331 /* size */32, /* alignment */8, false);
3334 void visitVACopyInst(VACopyInst &I) override {
3335 IRBuilder<> IRB(&I);
3336 Value *VAListTag = I.getArgOperand(0);
3337 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
3338 // Unpoison the whole __va_list_tag.
3339 // FIXME: magic ABI constants (size of va_list).
3340 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
3341 /* size */32, /* alignment */8, false);
3344 // Retrieve a va_list field of 'void*' size.
3345 Value* getVAField64(IRBuilder<> &IRB, Value *VAListTag, int offset) {
3346 Value *SaveAreaPtrPtr =
3348 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
3349 ConstantInt::get(MS.IntptrTy, offset)),
3350 Type::getInt64PtrTy(*MS.C));
3351 return IRB.CreateLoad(SaveAreaPtrPtr);
3354 // Retrieve a va_list field of 'int' size.
3355 Value* getVAField32(IRBuilder<> &IRB, Value *VAListTag, int offset) {
3356 Value *SaveAreaPtr =
3358 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
3359 ConstantInt::get(MS.IntptrTy, offset)),
3360 Type::getInt32PtrTy(*MS.C));
3361 Value *SaveArea32 = IRB.CreateLoad(SaveAreaPtr);
3362 return IRB.CreateSExt(SaveArea32, MS.IntptrTy);
3365 void finalizeInstrumentation() override {
3366 assert(!VAArgOverflowSize && !VAArgTLSCopy &&
3367 "finalizeInstrumentation called twice");
3368 if (!VAStartInstrumentationList.empty()) {
3369 // If there is a va_start in this function, make a backup copy of
3370 // va_arg_tls somewhere in the function entry block.
3371 IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
3372 VAArgOverflowSize = IRB.CreateLoad(MS.VAArgOverflowSizeTLS);
3374 IRB.CreateAdd(ConstantInt::get(MS.IntptrTy, AArch64VAEndOffset),
3376 VAArgTLSCopy = IRB.CreateAlloca(Type::getInt8Ty(*MS.C), CopySize);
3377 IRB.CreateMemCpy(VAArgTLSCopy, MS.VAArgTLS, CopySize, 8);
3380 Value *GrArgSize = ConstantInt::get(MS.IntptrTy, kAArch64GrArgSize);
3381 Value *VrArgSize = ConstantInt::get(MS.IntptrTy, kAArch64VrArgSize);
3383 // Instrument va_start, copy va_list shadow from the backup copy of
3384 // the TLS contents.
3385 for (size_t i = 0, n = VAStartInstrumentationList.size(); i < n; i++) {
3386 CallInst *OrigInst = VAStartInstrumentationList[i];
3387 IRBuilder<> IRB(OrigInst->getNextNode());
3389 Value *VAListTag = OrigInst->getArgOperand(0);
3391 // The variadic ABI for AArch64 creates two areas to save the incoming
3392 // argument registers (one for 64-bit general register xn-x7 and another
3393 // for 128-bit FP/SIMD vn-v7).
3394 // We need then to propagate the shadow arguments on both regions
3395 // 'va::__gr_top + va::__gr_offs' and 'va::__vr_top + va::__vr_offs'.
3396 // The remaning arguments are saved on shadow for 'va::stack'.
3397 // One caveat is it requires only to propagate the non-named arguments,
3398 // however on the call site instrumentation 'all' the arguments are
3399 // saved. So to copy the shadow values from the va_arg TLS array
3400 // we need to adjust the offset for both GR and VR fields based on
3401 // the __{gr,vr}_offs value (since they are stores based on incoming
3402 // named arguments).
3404 // Read the stack pointer from the va_list.
3405 Value *StackSaveAreaPtr = getVAField64(IRB, VAListTag, 0);
3407 // Read both the __gr_top and __gr_off and add them up.
3408 Value *GrTopSaveAreaPtr = getVAField64(IRB, VAListTag, 8);
3409 Value *GrOffSaveArea = getVAField32(IRB, VAListTag, 24);
3411 Value *GrRegSaveAreaPtr = IRB.CreateAdd(GrTopSaveAreaPtr, GrOffSaveArea);
3413 // Read both the __vr_top and __vr_off and add them up.
3414 Value *VrTopSaveAreaPtr = getVAField64(IRB, VAListTag, 16);
3415 Value *VrOffSaveArea = getVAField32(IRB, VAListTag, 28);
3417 Value *VrRegSaveAreaPtr = IRB.CreateAdd(VrTopSaveAreaPtr, VrOffSaveArea);
3419 // It does not know how many named arguments is being used and, on the
3420 // callsite all the arguments were saved. Since __gr_off is defined as
3421 // '0 - ((8 - named_gr) * 8)', the idea is to just propagate the variadic
3422 // argument by ignoring the bytes of shadow from named arguments.
3423 Value *GrRegSaveAreaShadowPtrOff =
3424 IRB.CreateAdd(GrArgSize, GrOffSaveArea);
3426 Value *GrRegSaveAreaShadowPtr =
3427 MSV.getShadowPtr(GrRegSaveAreaPtr, IRB.getInt8Ty(), IRB);
3429 Value *GrSrcPtr = IRB.CreateInBoundsGEP(IRB.getInt8Ty(), VAArgTLSCopy,
3430 GrRegSaveAreaShadowPtrOff);
3431 Value *GrCopySize = IRB.CreateSub(GrArgSize, GrRegSaveAreaShadowPtrOff);
3433 IRB.CreateMemCpy(GrRegSaveAreaShadowPtr, GrSrcPtr, GrCopySize, 8);
3435 // Again, but for FP/SIMD values.
3436 Value *VrRegSaveAreaShadowPtrOff =
3437 IRB.CreateAdd(VrArgSize, VrOffSaveArea);
3439 Value *VrRegSaveAreaShadowPtr =
3440 MSV.getShadowPtr(VrRegSaveAreaPtr, IRB.getInt8Ty(), IRB);
3442 Value *VrSrcPtr = IRB.CreateInBoundsGEP(
3444 IRB.CreateInBoundsGEP(IRB.getInt8Ty(), VAArgTLSCopy,
3445 IRB.getInt32(AArch64VrBegOffset)),
3446 VrRegSaveAreaShadowPtrOff);
3447 Value *VrCopySize = IRB.CreateSub(VrArgSize, VrRegSaveAreaShadowPtrOff);
3449 IRB.CreateMemCpy(VrRegSaveAreaShadowPtr, VrSrcPtr, VrCopySize, 8);
3451 // And finally for remaining arguments.
3452 Value *StackSaveAreaShadowPtr =
3453 MSV.getShadowPtr(StackSaveAreaPtr, IRB.getInt8Ty(), IRB);
3455 Value *StackSrcPtr =
3456 IRB.CreateInBoundsGEP(IRB.getInt8Ty(), VAArgTLSCopy,
3457 IRB.getInt32(AArch64VAEndOffset));
3459 IRB.CreateMemCpy(StackSaveAreaShadowPtr, StackSrcPtr,
3460 VAArgOverflowSize, 16);
3465 /// \brief PowerPC64-specific implementation of VarArgHelper.
3466 struct VarArgPowerPC64Helper : public VarArgHelper {
3468 MemorySanitizer &MS;
3469 MemorySanitizerVisitor &MSV;
3470 Value *VAArgTLSCopy;
3473 SmallVector<CallInst*, 16> VAStartInstrumentationList;
3475 VarArgPowerPC64Helper(Function &F, MemorySanitizer &MS,
3476 MemorySanitizerVisitor &MSV)
3477 : F(F), MS(MS), MSV(MSV), VAArgTLSCopy(nullptr),
3478 VAArgSize(nullptr) {}
3480 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) override {
3481 // For PowerPC, we need to deal with alignment of stack arguments -
3482 // they are mostly aligned to 8 bytes, but vectors and i128 arrays
3483 // are aligned to 16 bytes, byvals can be aligned to 8 or 16 bytes,
3484 // and QPX vectors are aligned to 32 bytes. For that reason, we
3485 // compute current offset from stack pointer (which is always properly
3486 // aligned), and offset for the first vararg, then subtract them.
3488 llvm::Triple TargetTriple(F.getParent()->getTargetTriple());
3489 // Parameter save area starts at 48 bytes from frame pointer for ABIv1,
3490 // and 32 bytes for ABIv2. This is usually determined by target
3491 // endianness, but in theory could be overriden by function attribute.
3492 // For simplicity, we ignore it here (it'd only matter for QPX vectors).
3493 if (TargetTriple.getArch() == llvm::Triple::ppc64)
3497 unsigned VAArgOffset = VAArgBase;
3498 const DataLayout &DL = F.getParent()->getDataLayout();
3499 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
3500 ArgIt != End; ++ArgIt) {
3502 unsigned ArgNo = CS.getArgumentNo(ArgIt);
3503 bool IsFixed = ArgNo < CS.getFunctionType()->getNumParams();
3504 bool IsByVal = CS.paramHasAttr(ArgNo, Attribute::ByVal);
3506 assert(A->getType()->isPointerTy());
3507 Type *RealTy = A->getType()->getPointerElementType();
3508 uint64_t ArgSize = DL.getTypeAllocSize(RealTy);
3509 uint64_t ArgAlign = CS.getParamAlignment(ArgNo);
3512 VAArgOffset = alignTo(VAArgOffset, ArgAlign);
3514 Value *Base = getShadowPtrForVAArgument(RealTy, IRB,
3515 VAArgOffset - VAArgBase);
3516 IRB.CreateMemCpy(Base, MSV.getShadowPtr(A, IRB.getInt8Ty(), IRB),
3517 ArgSize, kShadowTLSAlignment);
3519 VAArgOffset += alignTo(ArgSize, 8);
3522 uint64_t ArgSize = DL.getTypeAllocSize(A->getType());
3523 uint64_t ArgAlign = 8;
3524 if (A->getType()->isArrayTy()) {
3525 // Arrays are aligned to element size, except for long double
3526 // arrays, which are aligned to 8 bytes.
3527 Type *ElementTy = A->getType()->getArrayElementType();
3528 if (!ElementTy->isPPC_FP128Ty())
3529 ArgAlign = DL.getTypeAllocSize(ElementTy);
3530 } else if (A->getType()->isVectorTy()) {
3531 // Vectors are naturally aligned.
3532 ArgAlign = DL.getTypeAllocSize(A->getType());
3536 VAArgOffset = alignTo(VAArgOffset, ArgAlign);
3537 if (DL.isBigEndian()) {
3538 // Adjusting the shadow for argument with size < 8 to match the placement
3539 // of bits in big endian system
3541 VAArgOffset += (8 - ArgSize);
3544 Base = getShadowPtrForVAArgument(A->getType(), IRB,
3545 VAArgOffset - VAArgBase);
3546 IRB.CreateAlignedStore(MSV.getShadow(A), Base, kShadowTLSAlignment);
3548 VAArgOffset += ArgSize;
3549 VAArgOffset = alignTo(VAArgOffset, 8);
3552 VAArgBase = VAArgOffset;
3555 Constant *TotalVAArgSize = ConstantInt::get(IRB.getInt64Ty(),
3556 VAArgOffset - VAArgBase);
3557 // Here using VAArgOverflowSizeTLS as VAArgSizeTLS to avoid creation of
3558 // a new class member i.e. it is the total size of all VarArgs.
3559 IRB.CreateStore(TotalVAArgSize, MS.VAArgOverflowSizeTLS);
3562 /// \brief Compute the shadow address for a given va_arg.
3563 Value *getShadowPtrForVAArgument(Type *Ty, IRBuilder<> &IRB,
3565 Value *Base = IRB.CreatePointerCast(MS.VAArgTLS, MS.IntptrTy);
3566 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
3567 return IRB.CreateIntToPtr(Base, PointerType::get(MSV.getShadowTy(Ty), 0),
3571 void visitVAStartInst(VAStartInst &I) override {
3572 IRBuilder<> IRB(&I);
3573 VAStartInstrumentationList.push_back(&I);
3574 Value *VAListTag = I.getArgOperand(0);
3575 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
3576 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
3577 /* size */8, /* alignment */8, false);
3580 void visitVACopyInst(VACopyInst &I) override {
3581 IRBuilder<> IRB(&I);
3582 Value *VAListTag = I.getArgOperand(0);
3583 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
3584 // Unpoison the whole __va_list_tag.
3585 // FIXME: magic ABI constants.
3586 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
3587 /* size */8, /* alignment */8, false);
3590 void finalizeInstrumentation() override {
3591 assert(!VAArgSize && !VAArgTLSCopy &&
3592 "finalizeInstrumentation called twice");
3593 IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
3594 VAArgSize = IRB.CreateLoad(MS.VAArgOverflowSizeTLS);
3595 Value *CopySize = IRB.CreateAdd(ConstantInt::get(MS.IntptrTy, 0),
3598 if (!VAStartInstrumentationList.empty()) {
3599 // If there is a va_start in this function, make a backup copy of
3600 // va_arg_tls somewhere in the function entry block.
3601 VAArgTLSCopy = IRB.CreateAlloca(Type::getInt8Ty(*MS.C), CopySize);
3602 IRB.CreateMemCpy(VAArgTLSCopy, MS.VAArgTLS, CopySize, 8);
3605 // Instrument va_start.
3606 // Copy va_list shadow from the backup copy of the TLS contents.
3607 for (size_t i = 0, n = VAStartInstrumentationList.size(); i < n; i++) {
3608 CallInst *OrigInst = VAStartInstrumentationList[i];
3609 IRBuilder<> IRB(OrigInst->getNextNode());
3610 Value *VAListTag = OrigInst->getArgOperand(0);
3611 Value *RegSaveAreaPtrPtr =
3612 IRB.CreateIntToPtr(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
3613 Type::getInt64PtrTy(*MS.C));
3614 Value *RegSaveAreaPtr = IRB.CreateLoad(RegSaveAreaPtrPtr);
3615 Value *RegSaveAreaShadowPtr =
3616 MSV.getShadowPtr(RegSaveAreaPtr, IRB.getInt8Ty(), IRB);
3617 IRB.CreateMemCpy(RegSaveAreaShadowPtr, VAArgTLSCopy, CopySize, 8);
3622 /// \brief A no-op implementation of VarArgHelper.
3623 struct VarArgNoOpHelper : public VarArgHelper {
3624 VarArgNoOpHelper(Function &F, MemorySanitizer &MS,
3625 MemorySanitizerVisitor &MSV) {}
3627 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) override {}
3629 void visitVAStartInst(VAStartInst &I) override {}
3631 void visitVACopyInst(VACopyInst &I) override {}
3633 void finalizeInstrumentation() override {}
3636 VarArgHelper *CreateVarArgHelper(Function &Func, MemorySanitizer &Msan,
3637 MemorySanitizerVisitor &Visitor) {
3638 // VarArg handling is only implemented on AMD64. False positives are possible
3639 // on other platforms.
3640 llvm::Triple TargetTriple(Func.getParent()->getTargetTriple());
3641 if (TargetTriple.getArch() == llvm::Triple::x86_64)
3642 return new VarArgAMD64Helper(Func, Msan, Visitor);
3643 else if (TargetTriple.getArch() == llvm::Triple::mips64 ||
3644 TargetTriple.getArch() == llvm::Triple::mips64el)
3645 return new VarArgMIPS64Helper(Func, Msan, Visitor);
3646 else if (TargetTriple.getArch() == llvm::Triple::aarch64)
3647 return new VarArgAArch64Helper(Func, Msan, Visitor);
3648 else if (TargetTriple.getArch() == llvm::Triple::ppc64 ||
3649 TargetTriple.getArch() == llvm::Triple::ppc64le)
3650 return new VarArgPowerPC64Helper(Func, Msan, Visitor);
3652 return new VarArgNoOpHelper(Func, Msan, Visitor);
3655 } // anonymous namespace
3657 bool MemorySanitizer::runOnFunction(Function &F) {
3658 if (&F == MsanCtorFunction)
3660 MemorySanitizerVisitor Visitor(F, *this);
3662 // Clear out readonly/readnone attributes.
3664 B.addAttribute(Attribute::ReadOnly)
3665 .addAttribute(Attribute::ReadNone);
3666 F.removeAttributes(AttributeList::FunctionIndex, B);
3668 return Visitor.runOnFunction();