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 if (dstTy->isIntegerTy() && srcTy->isIntegerTy())
1580 return IRB.CreateIntCast(V, dstTy, Signed);
1581 if (dstTy->isVectorTy() && srcTy->isVectorTy() &&
1582 dstTy->getVectorNumElements() == srcTy->getVectorNumElements())
1583 return IRB.CreateIntCast(V, dstTy, Signed);
1584 size_t srcSizeInBits = VectorOrPrimitiveTypeSizeInBits(srcTy);
1585 size_t dstSizeInBits = VectorOrPrimitiveTypeSizeInBits(dstTy);
1586 Value *V1 = IRB.CreateBitCast(V, Type::getIntNTy(*MS.C, srcSizeInBits));
1588 IRB.CreateIntCast(V1, Type::getIntNTy(*MS.C, dstSizeInBits), Signed);
1589 return IRB.CreateBitCast(V2, dstTy);
1590 // TODO: handle struct types.
1593 /// \brief Cast an application value to the type of its own shadow.
1594 Value *CreateAppToShadowCast(IRBuilder<> &IRB, Value *V) {
1595 Type *ShadowTy = getShadowTy(V);
1596 if (V->getType() == ShadowTy)
1598 if (V->getType()->isPtrOrPtrVectorTy())
1599 return IRB.CreatePtrToInt(V, ShadowTy);
1601 return IRB.CreateBitCast(V, ShadowTy);
1604 /// \brief Propagate shadow for arbitrary operation.
1605 void handleShadowOr(Instruction &I) {
1606 IRBuilder<> IRB(&I);
1607 ShadowAndOriginCombiner SC(this, IRB);
1608 for (Instruction::op_iterator OI = I.op_begin(); OI != I.op_end(); ++OI)
1613 // \brief Handle multiplication by constant.
1615 // Handle a special case of multiplication by constant that may have one or
1616 // more zeros in the lower bits. This makes corresponding number of lower bits
1617 // of the result zero as well. We model it by shifting the other operand
1618 // shadow left by the required number of bits. Effectively, we transform
1619 // (X * (A * 2**B)) to ((X << B) * A) and instrument (X << B) as (Sx << B).
1620 // We use multiplication by 2**N instead of shift to cover the case of
1621 // multiplication by 0, which may occur in some elements of a vector operand.
1622 void handleMulByConstant(BinaryOperator &I, Constant *ConstArg,
1624 Constant *ShadowMul;
1625 Type *Ty = ConstArg->getType();
1626 if (Ty->isVectorTy()) {
1627 unsigned NumElements = Ty->getVectorNumElements();
1628 Type *EltTy = Ty->getSequentialElementType();
1629 SmallVector<Constant *, 16> Elements;
1630 for (unsigned Idx = 0; Idx < NumElements; ++Idx) {
1631 if (ConstantInt *Elt =
1632 dyn_cast<ConstantInt>(ConstArg->getAggregateElement(Idx))) {
1633 const APInt &V = Elt->getValue();
1634 APInt V2 = APInt(V.getBitWidth(), 1) << V.countTrailingZeros();
1635 Elements.push_back(ConstantInt::get(EltTy, V2));
1637 Elements.push_back(ConstantInt::get(EltTy, 1));
1640 ShadowMul = ConstantVector::get(Elements);
1642 if (ConstantInt *Elt = dyn_cast<ConstantInt>(ConstArg)) {
1643 const APInt &V = Elt->getValue();
1644 APInt V2 = APInt(V.getBitWidth(), 1) << V.countTrailingZeros();
1645 ShadowMul = ConstantInt::get(Ty, V2);
1647 ShadowMul = ConstantInt::get(Ty, 1);
1651 IRBuilder<> IRB(&I);
1653 IRB.CreateMul(getShadow(OtherArg), ShadowMul, "msprop_mul_cst"));
1654 setOrigin(&I, getOrigin(OtherArg));
1657 void visitMul(BinaryOperator &I) {
1658 Constant *constOp0 = dyn_cast<Constant>(I.getOperand(0));
1659 Constant *constOp1 = dyn_cast<Constant>(I.getOperand(1));
1660 if (constOp0 && !constOp1)
1661 handleMulByConstant(I, constOp0, I.getOperand(1));
1662 else if (constOp1 && !constOp0)
1663 handleMulByConstant(I, constOp1, I.getOperand(0));
1668 void visitFAdd(BinaryOperator &I) { handleShadowOr(I); }
1669 void visitFSub(BinaryOperator &I) { handleShadowOr(I); }
1670 void visitFMul(BinaryOperator &I) { handleShadowOr(I); }
1671 void visitAdd(BinaryOperator &I) { handleShadowOr(I); }
1672 void visitSub(BinaryOperator &I) { handleShadowOr(I); }
1673 void visitXor(BinaryOperator &I) { handleShadowOr(I); }
1675 void handleDiv(Instruction &I) {
1676 IRBuilder<> IRB(&I);
1677 // Strict on the second argument.
1678 insertShadowCheck(I.getOperand(1), &I);
1679 setShadow(&I, getShadow(&I, 0));
1680 setOrigin(&I, getOrigin(&I, 0));
1683 void visitUDiv(BinaryOperator &I) { handleDiv(I); }
1684 void visitSDiv(BinaryOperator &I) { handleDiv(I); }
1685 void visitFDiv(BinaryOperator &I) { handleDiv(I); }
1686 void visitURem(BinaryOperator &I) { handleDiv(I); }
1687 void visitSRem(BinaryOperator &I) { handleDiv(I); }
1688 void visitFRem(BinaryOperator &I) { handleDiv(I); }
1690 /// \brief Instrument == and != comparisons.
1692 /// Sometimes the comparison result is known even if some of the bits of the
1693 /// arguments are not.
1694 void handleEqualityComparison(ICmpInst &I) {
1695 IRBuilder<> IRB(&I);
1696 Value *A = I.getOperand(0);
1697 Value *B = I.getOperand(1);
1698 Value *Sa = getShadow(A);
1699 Value *Sb = getShadow(B);
1701 // Get rid of pointers and vectors of pointers.
1702 // For ints (and vectors of ints), types of A and Sa match,
1703 // and this is a no-op.
1704 A = IRB.CreatePointerCast(A, Sa->getType());
1705 B = IRB.CreatePointerCast(B, Sb->getType());
1707 // A == B <==> (C = A^B) == 0
1708 // A != B <==> (C = A^B) != 0
1710 Value *C = IRB.CreateXor(A, B);
1711 Value *Sc = IRB.CreateOr(Sa, Sb);
1712 // Now dealing with i = (C == 0) comparison (or C != 0, does not matter now)
1713 // Result is defined if one of the following is true
1714 // * there is a defined 1 bit in C
1715 // * C is fully defined
1716 // Si = !(C & ~Sc) && Sc
1717 Value *Zero = Constant::getNullValue(Sc->getType());
1718 Value *MinusOne = Constant::getAllOnesValue(Sc->getType());
1720 IRB.CreateAnd(IRB.CreateICmpNE(Sc, Zero),
1722 IRB.CreateAnd(IRB.CreateXor(Sc, MinusOne), C), Zero));
1723 Si->setName("_msprop_icmp");
1725 setOriginForNaryOp(I);
1728 /// \brief Build the lowest possible value of V, taking into account V's
1729 /// uninitialized bits.
1730 Value *getLowestPossibleValue(IRBuilder<> &IRB, Value *A, Value *Sa,
1733 // Split shadow into sign bit and other bits.
1734 Value *SaOtherBits = IRB.CreateLShr(IRB.CreateShl(Sa, 1), 1);
1735 Value *SaSignBit = IRB.CreateXor(Sa, SaOtherBits);
1736 // Maximise the undefined shadow bit, minimize other undefined bits.
1738 IRB.CreateOr(IRB.CreateAnd(A, IRB.CreateNot(SaOtherBits)), SaSignBit);
1740 // Minimize undefined bits.
1741 return IRB.CreateAnd(A, IRB.CreateNot(Sa));
1745 /// \brief Build the highest possible value of V, taking into account V's
1746 /// uninitialized bits.
1747 Value *getHighestPossibleValue(IRBuilder<> &IRB, Value *A, Value *Sa,
1750 // Split shadow into sign bit and other bits.
1751 Value *SaOtherBits = IRB.CreateLShr(IRB.CreateShl(Sa, 1), 1);
1752 Value *SaSignBit = IRB.CreateXor(Sa, SaOtherBits);
1753 // Minimise the undefined shadow bit, maximise other undefined bits.
1755 IRB.CreateOr(IRB.CreateAnd(A, IRB.CreateNot(SaSignBit)), SaOtherBits);
1757 // Maximize undefined bits.
1758 return IRB.CreateOr(A, Sa);
1762 /// \brief Instrument relational comparisons.
1764 /// This function does exact shadow propagation for all relational
1765 /// comparisons of integers, pointers and vectors of those.
1766 /// FIXME: output seems suboptimal when one of the operands is a constant
1767 void handleRelationalComparisonExact(ICmpInst &I) {
1768 IRBuilder<> IRB(&I);
1769 Value *A = I.getOperand(0);
1770 Value *B = I.getOperand(1);
1771 Value *Sa = getShadow(A);
1772 Value *Sb = getShadow(B);
1774 // Get rid of pointers and vectors of pointers.
1775 // For ints (and vectors of ints), types of A and Sa match,
1776 // and this is a no-op.
1777 A = IRB.CreatePointerCast(A, Sa->getType());
1778 B = IRB.CreatePointerCast(B, Sb->getType());
1780 // Let [a0, a1] be the interval of possible values of A, taking into account
1781 // its undefined bits. Let [b0, b1] be the interval of possible values of B.
1782 // Then (A cmp B) is defined iff (a0 cmp b1) == (a1 cmp b0).
1783 bool IsSigned = I.isSigned();
1784 Value *S1 = IRB.CreateICmp(I.getPredicate(),
1785 getLowestPossibleValue(IRB, A, Sa, IsSigned),
1786 getHighestPossibleValue(IRB, B, Sb, IsSigned));
1787 Value *S2 = IRB.CreateICmp(I.getPredicate(),
1788 getHighestPossibleValue(IRB, A, Sa, IsSigned),
1789 getLowestPossibleValue(IRB, B, Sb, IsSigned));
1790 Value *Si = IRB.CreateXor(S1, S2);
1792 setOriginForNaryOp(I);
1795 /// \brief Instrument signed relational comparisons.
1797 /// Handle sign bit tests: x<0, x>=0, x<=-1, x>-1 by propagating the highest
1798 /// bit of the shadow. Everything else is delegated to handleShadowOr().
1799 void handleSignedRelationalComparison(ICmpInst &I) {
1801 Value *op = nullptr;
1802 CmpInst::Predicate pre;
1803 if ((constOp = dyn_cast<Constant>(I.getOperand(1)))) {
1804 op = I.getOperand(0);
1805 pre = I.getPredicate();
1806 } else if ((constOp = dyn_cast<Constant>(I.getOperand(0)))) {
1807 op = I.getOperand(1);
1808 pre = I.getSwappedPredicate();
1814 if ((constOp->isNullValue() &&
1815 (pre == CmpInst::ICMP_SLT || pre == CmpInst::ICMP_SGE)) ||
1816 (constOp->isAllOnesValue() &&
1817 (pre == CmpInst::ICMP_SGT || pre == CmpInst::ICMP_SLE))) {
1818 IRBuilder<> IRB(&I);
1819 Value *Shadow = IRB.CreateICmpSLT(getShadow(op), getCleanShadow(op),
1821 setShadow(&I, Shadow);
1822 setOrigin(&I, getOrigin(op));
1828 void visitICmpInst(ICmpInst &I) {
1829 if (!ClHandleICmp) {
1833 if (I.isEquality()) {
1834 handleEqualityComparison(I);
1838 assert(I.isRelational());
1839 if (ClHandleICmpExact) {
1840 handleRelationalComparisonExact(I);
1844 handleSignedRelationalComparison(I);
1848 assert(I.isUnsigned());
1849 if ((isa<Constant>(I.getOperand(0)) || isa<Constant>(I.getOperand(1)))) {
1850 handleRelationalComparisonExact(I);
1857 void visitFCmpInst(FCmpInst &I) {
1861 void handleShift(BinaryOperator &I) {
1862 IRBuilder<> IRB(&I);
1863 // If any of the S2 bits are poisoned, the whole thing is poisoned.
1864 // Otherwise perform the same shift on S1.
1865 Value *S1 = getShadow(&I, 0);
1866 Value *S2 = getShadow(&I, 1);
1867 Value *S2Conv = IRB.CreateSExt(IRB.CreateICmpNE(S2, getCleanShadow(S2)),
1869 Value *V2 = I.getOperand(1);
1870 Value *Shift = IRB.CreateBinOp(I.getOpcode(), S1, V2);
1871 setShadow(&I, IRB.CreateOr(Shift, S2Conv));
1872 setOriginForNaryOp(I);
1875 void visitShl(BinaryOperator &I) { handleShift(I); }
1876 void visitAShr(BinaryOperator &I) { handleShift(I); }
1877 void visitLShr(BinaryOperator &I) { handleShift(I); }
1879 /// \brief Instrument llvm.memmove
1881 /// At this point we don't know if llvm.memmove will be inlined or not.
1882 /// If we don't instrument it and it gets inlined,
1883 /// our interceptor will not kick in and we will lose the memmove.
1884 /// If we instrument the call here, but it does not get inlined,
1885 /// we will memove the shadow twice: which is bad in case
1886 /// of overlapping regions. So, we simply lower the intrinsic to a call.
1888 /// Similar situation exists for memcpy and memset.
1889 void visitMemMoveInst(MemMoveInst &I) {
1890 IRBuilder<> IRB(&I);
1893 {IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1894 IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()),
1895 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false)});
1896 I.eraseFromParent();
1899 // Similar to memmove: avoid copying shadow twice.
1900 // This is somewhat unfortunate as it may slowdown small constant memcpys.
1901 // FIXME: consider doing manual inline for small constant sizes and proper
1903 void visitMemCpyInst(MemCpyInst &I) {
1904 IRBuilder<> IRB(&I);
1907 {IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1908 IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()),
1909 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false)});
1910 I.eraseFromParent();
1914 void visitMemSetInst(MemSetInst &I) {
1915 IRBuilder<> IRB(&I);
1918 {IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1919 IRB.CreateIntCast(I.getArgOperand(1), IRB.getInt32Ty(), false),
1920 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false)});
1921 I.eraseFromParent();
1924 void visitVAStartInst(VAStartInst &I) {
1925 VAHelper->visitVAStartInst(I);
1928 void visitVACopyInst(VACopyInst &I) {
1929 VAHelper->visitVACopyInst(I);
1932 /// \brief Handle vector store-like intrinsics.
1934 /// Instrument intrinsics that look like a simple SIMD store: writes memory,
1935 /// has 1 pointer argument and 1 vector argument, returns void.
1936 bool handleVectorStoreIntrinsic(IntrinsicInst &I) {
1937 IRBuilder<> IRB(&I);
1938 Value* Addr = I.getArgOperand(0);
1939 Value *Shadow = getShadow(&I, 1);
1940 Value *ShadowPtr = getShadowPtr(Addr, Shadow->getType(), IRB);
1942 // We don't know the pointer alignment (could be unaligned SSE store!).
1943 // Have to assume to worst case.
1944 IRB.CreateAlignedStore(Shadow, ShadowPtr, 1);
1946 if (ClCheckAccessAddress)
1947 insertShadowCheck(Addr, &I);
1949 // FIXME: factor out common code from materializeStores
1950 if (MS.TrackOrigins)
1951 IRB.CreateStore(getOrigin(&I, 1), getOriginPtr(Addr, IRB, 1));
1955 /// \brief Handle vector load-like intrinsics.
1957 /// Instrument intrinsics that look like a simple SIMD load: reads memory,
1958 /// has 1 pointer argument, returns a vector.
1959 bool handleVectorLoadIntrinsic(IntrinsicInst &I) {
1960 IRBuilder<> IRB(&I);
1961 Value *Addr = I.getArgOperand(0);
1963 Type *ShadowTy = getShadowTy(&I);
1964 if (PropagateShadow) {
1965 Value *ShadowPtr = getShadowPtr(Addr, ShadowTy, IRB);
1966 // We don't know the pointer alignment (could be unaligned SSE load!).
1967 // Have to assume to worst case.
1968 setShadow(&I, IRB.CreateAlignedLoad(ShadowPtr, 1, "_msld"));
1970 setShadow(&I, getCleanShadow(&I));
1973 if (ClCheckAccessAddress)
1974 insertShadowCheck(Addr, &I);
1976 if (MS.TrackOrigins) {
1977 if (PropagateShadow)
1978 setOrigin(&I, IRB.CreateLoad(getOriginPtr(Addr, IRB, 1)));
1980 setOrigin(&I, getCleanOrigin());
1985 /// \brief Handle (SIMD arithmetic)-like intrinsics.
1987 /// Instrument intrinsics with any number of arguments of the same type,
1988 /// equal to the return type. The type should be simple (no aggregates or
1989 /// pointers; vectors are fine).
1990 /// Caller guarantees that this intrinsic does not access memory.
1991 bool maybeHandleSimpleNomemIntrinsic(IntrinsicInst &I) {
1992 Type *RetTy = I.getType();
1993 if (!(RetTy->isIntOrIntVectorTy() ||
1994 RetTy->isFPOrFPVectorTy() ||
1995 RetTy->isX86_MMXTy()))
1998 unsigned NumArgOperands = I.getNumArgOperands();
2000 for (unsigned i = 0; i < NumArgOperands; ++i) {
2001 Type *Ty = I.getArgOperand(i)->getType();
2006 IRBuilder<> IRB(&I);
2007 ShadowAndOriginCombiner SC(this, IRB);
2008 for (unsigned i = 0; i < NumArgOperands; ++i)
2009 SC.Add(I.getArgOperand(i));
2015 /// \brief Heuristically instrument unknown intrinsics.
2017 /// The main purpose of this code is to do something reasonable with all
2018 /// random intrinsics we might encounter, most importantly - SIMD intrinsics.
2019 /// We recognize several classes of intrinsics by their argument types and
2020 /// ModRefBehaviour and apply special intrumentation when we are reasonably
2021 /// sure that we know what the intrinsic does.
2023 /// We special-case intrinsics where this approach fails. See llvm.bswap
2024 /// handling as an example of that.
2025 bool handleUnknownIntrinsic(IntrinsicInst &I) {
2026 unsigned NumArgOperands = I.getNumArgOperands();
2027 if (NumArgOperands == 0)
2030 if (NumArgOperands == 2 &&
2031 I.getArgOperand(0)->getType()->isPointerTy() &&
2032 I.getArgOperand(1)->getType()->isVectorTy() &&
2033 I.getType()->isVoidTy() &&
2034 !I.onlyReadsMemory()) {
2035 // This looks like a vector store.
2036 return handleVectorStoreIntrinsic(I);
2039 if (NumArgOperands == 1 &&
2040 I.getArgOperand(0)->getType()->isPointerTy() &&
2041 I.getType()->isVectorTy() &&
2042 I.onlyReadsMemory()) {
2043 // This looks like a vector load.
2044 return handleVectorLoadIntrinsic(I);
2047 if (I.doesNotAccessMemory())
2048 if (maybeHandleSimpleNomemIntrinsic(I))
2051 // FIXME: detect and handle SSE maskstore/maskload
2055 void handleBswap(IntrinsicInst &I) {
2056 IRBuilder<> IRB(&I);
2057 Value *Op = I.getArgOperand(0);
2058 Type *OpType = Op->getType();
2059 Function *BswapFunc = Intrinsic::getDeclaration(
2060 F.getParent(), Intrinsic::bswap, makeArrayRef(&OpType, 1));
2061 setShadow(&I, IRB.CreateCall(BswapFunc, getShadow(Op)));
2062 setOrigin(&I, getOrigin(Op));
2065 // \brief Instrument vector convert instrinsic.
2067 // This function instruments intrinsics like cvtsi2ss:
2068 // %Out = int_xxx_cvtyyy(%ConvertOp)
2070 // %Out = int_xxx_cvtyyy(%CopyOp, %ConvertOp)
2071 // Intrinsic converts \p NumUsedElements elements of \p ConvertOp to the same
2072 // number \p Out elements, and (if has 2 arguments) copies the rest of the
2073 // elements from \p CopyOp.
2074 // In most cases conversion involves floating-point value which may trigger a
2075 // hardware exception when not fully initialized. For this reason we require
2076 // \p ConvertOp[0:NumUsedElements] to be fully initialized and trap otherwise.
2077 // We copy the shadow of \p CopyOp[NumUsedElements:] to \p
2078 // Out[NumUsedElements:]. This means that intrinsics without \p CopyOp always
2079 // return a fully initialized value.
2080 void handleVectorConvertIntrinsic(IntrinsicInst &I, int NumUsedElements) {
2081 IRBuilder<> IRB(&I);
2082 Value *CopyOp, *ConvertOp;
2084 switch (I.getNumArgOperands()) {
2086 assert(isa<ConstantInt>(I.getArgOperand(2)) && "Invalid rounding mode");
2088 CopyOp = I.getArgOperand(0);
2089 ConvertOp = I.getArgOperand(1);
2092 ConvertOp = I.getArgOperand(0);
2096 llvm_unreachable("Cvt intrinsic with unsupported number of arguments.");
2099 // The first *NumUsedElements* elements of ConvertOp are converted to the
2100 // same number of output elements. The rest of the output is copied from
2101 // CopyOp, or (if not available) filled with zeroes.
2102 // Combine shadow for elements of ConvertOp that are used in this operation,
2103 // and insert a check.
2104 // FIXME: consider propagating shadow of ConvertOp, at least in the case of
2105 // int->any conversion.
2106 Value *ConvertShadow = getShadow(ConvertOp);
2107 Value *AggShadow = nullptr;
2108 if (ConvertOp->getType()->isVectorTy()) {
2109 AggShadow = IRB.CreateExtractElement(
2110 ConvertShadow, ConstantInt::get(IRB.getInt32Ty(), 0));
2111 for (int i = 1; i < NumUsedElements; ++i) {
2112 Value *MoreShadow = IRB.CreateExtractElement(
2113 ConvertShadow, ConstantInt::get(IRB.getInt32Ty(), i));
2114 AggShadow = IRB.CreateOr(AggShadow, MoreShadow);
2117 AggShadow = ConvertShadow;
2119 assert(AggShadow->getType()->isIntegerTy());
2120 insertShadowCheck(AggShadow, getOrigin(ConvertOp), &I);
2122 // Build result shadow by zero-filling parts of CopyOp shadow that come from
2125 assert(CopyOp->getType() == I.getType());
2126 assert(CopyOp->getType()->isVectorTy());
2127 Value *ResultShadow = getShadow(CopyOp);
2128 Type *EltTy = ResultShadow->getType()->getVectorElementType();
2129 for (int i = 0; i < NumUsedElements; ++i) {
2130 ResultShadow = IRB.CreateInsertElement(
2131 ResultShadow, ConstantInt::getNullValue(EltTy),
2132 ConstantInt::get(IRB.getInt32Ty(), i));
2134 setShadow(&I, ResultShadow);
2135 setOrigin(&I, getOrigin(CopyOp));
2137 setShadow(&I, getCleanShadow(&I));
2138 setOrigin(&I, getCleanOrigin());
2142 // Given a scalar or vector, extract lower 64 bits (or less), and return all
2143 // zeroes if it is zero, and all ones otherwise.
2144 Value *Lower64ShadowExtend(IRBuilder<> &IRB, Value *S, Type *T) {
2145 if (S->getType()->isVectorTy())
2146 S = CreateShadowCast(IRB, S, IRB.getInt64Ty(), /* Signed */ true);
2147 assert(S->getType()->getPrimitiveSizeInBits() <= 64);
2148 Value *S2 = IRB.CreateICmpNE(S, getCleanShadow(S));
2149 return CreateShadowCast(IRB, S2, T, /* Signed */ true);
2152 // Given a vector, extract its first element, and return all
2153 // zeroes if it is zero, and all ones otherwise.
2154 Value *LowerElementShadowExtend(IRBuilder<> &IRB, Value *S, Type *T) {
2155 Value *S1 = IRB.CreateExtractElement(S, (uint64_t)0);
2156 Value *S2 = IRB.CreateICmpNE(S1, getCleanShadow(S1));
2157 return CreateShadowCast(IRB, S2, T, /* Signed */ true);
2160 Value *VariableShadowExtend(IRBuilder<> &IRB, Value *S) {
2161 Type *T = S->getType();
2162 assert(T->isVectorTy());
2163 Value *S2 = IRB.CreateICmpNE(S, getCleanShadow(S));
2164 return IRB.CreateSExt(S2, T);
2167 // \brief Instrument vector shift instrinsic.
2169 // This function instruments intrinsics like int_x86_avx2_psll_w.
2170 // Intrinsic shifts %In by %ShiftSize bits.
2171 // %ShiftSize may be a vector. In that case the lower 64 bits determine shift
2172 // size, and the rest is ignored. Behavior is defined even if shift size is
2173 // greater than register (or field) width.
2174 void handleVectorShiftIntrinsic(IntrinsicInst &I, bool Variable) {
2175 assert(I.getNumArgOperands() == 2);
2176 IRBuilder<> IRB(&I);
2177 // If any of the S2 bits are poisoned, the whole thing is poisoned.
2178 // Otherwise perform the same shift on S1.
2179 Value *S1 = getShadow(&I, 0);
2180 Value *S2 = getShadow(&I, 1);
2181 Value *S2Conv = Variable ? VariableShadowExtend(IRB, S2)
2182 : Lower64ShadowExtend(IRB, S2, getShadowTy(&I));
2183 Value *V1 = I.getOperand(0);
2184 Value *V2 = I.getOperand(1);
2185 Value *Shift = IRB.CreateCall(I.getCalledValue(),
2186 {IRB.CreateBitCast(S1, V1->getType()), V2});
2187 Shift = IRB.CreateBitCast(Shift, getShadowTy(&I));
2188 setShadow(&I, IRB.CreateOr(Shift, S2Conv));
2189 setOriginForNaryOp(I);
2192 // \brief Get an X86_MMX-sized vector type.
2193 Type *getMMXVectorTy(unsigned EltSizeInBits) {
2194 const unsigned X86_MMXSizeInBits = 64;
2195 return VectorType::get(IntegerType::get(*MS.C, EltSizeInBits),
2196 X86_MMXSizeInBits / EltSizeInBits);
2199 // \brief Returns a signed counterpart for an (un)signed-saturate-and-pack
2201 Intrinsic::ID getSignedPackIntrinsic(Intrinsic::ID id) {
2203 case llvm::Intrinsic::x86_sse2_packsswb_128:
2204 case llvm::Intrinsic::x86_sse2_packuswb_128:
2205 return llvm::Intrinsic::x86_sse2_packsswb_128;
2207 case llvm::Intrinsic::x86_sse2_packssdw_128:
2208 case llvm::Intrinsic::x86_sse41_packusdw:
2209 return llvm::Intrinsic::x86_sse2_packssdw_128;
2211 case llvm::Intrinsic::x86_avx2_packsswb:
2212 case llvm::Intrinsic::x86_avx2_packuswb:
2213 return llvm::Intrinsic::x86_avx2_packsswb;
2215 case llvm::Intrinsic::x86_avx2_packssdw:
2216 case llvm::Intrinsic::x86_avx2_packusdw:
2217 return llvm::Intrinsic::x86_avx2_packssdw;
2219 case llvm::Intrinsic::x86_mmx_packsswb:
2220 case llvm::Intrinsic::x86_mmx_packuswb:
2221 return llvm::Intrinsic::x86_mmx_packsswb;
2223 case llvm::Intrinsic::x86_mmx_packssdw:
2224 return llvm::Intrinsic::x86_mmx_packssdw;
2226 llvm_unreachable("unexpected intrinsic id");
2230 // \brief Instrument vector pack instrinsic.
2232 // This function instruments intrinsics like x86_mmx_packsswb, that
2233 // packs elements of 2 input vectors into half as many bits with saturation.
2234 // Shadow is propagated with the signed variant of the same intrinsic applied
2235 // to sext(Sa != zeroinitializer), sext(Sb != zeroinitializer).
2236 // EltSizeInBits is used only for x86mmx arguments.
2237 void handleVectorPackIntrinsic(IntrinsicInst &I, unsigned EltSizeInBits = 0) {
2238 assert(I.getNumArgOperands() == 2);
2239 bool isX86_MMX = I.getOperand(0)->getType()->isX86_MMXTy();
2240 IRBuilder<> IRB(&I);
2241 Value *S1 = getShadow(&I, 0);
2242 Value *S2 = getShadow(&I, 1);
2243 assert(isX86_MMX || S1->getType()->isVectorTy());
2245 // SExt and ICmpNE below must apply to individual elements of input vectors.
2246 // In case of x86mmx arguments, cast them to appropriate vector types and
2248 Type *T = isX86_MMX ? getMMXVectorTy(EltSizeInBits) : S1->getType();
2250 S1 = IRB.CreateBitCast(S1, T);
2251 S2 = IRB.CreateBitCast(S2, T);
2253 Value *S1_ext = IRB.CreateSExt(
2254 IRB.CreateICmpNE(S1, llvm::Constant::getNullValue(T)), T);
2255 Value *S2_ext = IRB.CreateSExt(
2256 IRB.CreateICmpNE(S2, llvm::Constant::getNullValue(T)), T);
2258 Type *X86_MMXTy = Type::getX86_MMXTy(*MS.C);
2259 S1_ext = IRB.CreateBitCast(S1_ext, X86_MMXTy);
2260 S2_ext = IRB.CreateBitCast(S2_ext, X86_MMXTy);
2263 Function *ShadowFn = Intrinsic::getDeclaration(
2264 F.getParent(), getSignedPackIntrinsic(I.getIntrinsicID()));
2267 IRB.CreateCall(ShadowFn, {S1_ext, S2_ext}, "_msprop_vector_pack");
2268 if (isX86_MMX) S = IRB.CreateBitCast(S, getShadowTy(&I));
2270 setOriginForNaryOp(I);
2273 // \brief Instrument sum-of-absolute-differencies intrinsic.
2274 void handleVectorSadIntrinsic(IntrinsicInst &I) {
2275 const unsigned SignificantBitsPerResultElement = 16;
2276 bool isX86_MMX = I.getOperand(0)->getType()->isX86_MMXTy();
2277 Type *ResTy = isX86_MMX ? IntegerType::get(*MS.C, 64) : I.getType();
2278 unsigned ZeroBitsPerResultElement =
2279 ResTy->getScalarSizeInBits() - SignificantBitsPerResultElement;
2281 IRBuilder<> IRB(&I);
2282 Value *S = IRB.CreateOr(getShadow(&I, 0), getShadow(&I, 1));
2283 S = IRB.CreateBitCast(S, ResTy);
2284 S = IRB.CreateSExt(IRB.CreateICmpNE(S, Constant::getNullValue(ResTy)),
2286 S = IRB.CreateLShr(S, ZeroBitsPerResultElement);
2287 S = IRB.CreateBitCast(S, getShadowTy(&I));
2289 setOriginForNaryOp(I);
2292 // \brief Instrument multiply-add intrinsic.
2293 void handleVectorPmaddIntrinsic(IntrinsicInst &I,
2294 unsigned EltSizeInBits = 0) {
2295 bool isX86_MMX = I.getOperand(0)->getType()->isX86_MMXTy();
2296 Type *ResTy = isX86_MMX ? getMMXVectorTy(EltSizeInBits * 2) : I.getType();
2297 IRBuilder<> IRB(&I);
2298 Value *S = IRB.CreateOr(getShadow(&I, 0), getShadow(&I, 1));
2299 S = IRB.CreateBitCast(S, ResTy);
2300 S = IRB.CreateSExt(IRB.CreateICmpNE(S, Constant::getNullValue(ResTy)),
2302 S = IRB.CreateBitCast(S, getShadowTy(&I));
2304 setOriginForNaryOp(I);
2307 // \brief Instrument compare-packed intrinsic.
2308 // Basically, an or followed by sext(icmp ne 0) to end up with all-zeros or
2310 void handleVectorComparePackedIntrinsic(IntrinsicInst &I) {
2311 IRBuilder<> IRB(&I);
2312 Type *ResTy = getShadowTy(&I);
2313 Value *S0 = IRB.CreateOr(getShadow(&I, 0), getShadow(&I, 1));
2314 Value *S = IRB.CreateSExt(
2315 IRB.CreateICmpNE(S0, Constant::getNullValue(ResTy)), ResTy);
2317 setOriginForNaryOp(I);
2320 // \brief Instrument compare-scalar intrinsic.
2321 // This handles both cmp* intrinsics which return the result in the first
2322 // element of a vector, and comi* which return the result as i32.
2323 void handleVectorCompareScalarIntrinsic(IntrinsicInst &I) {
2324 IRBuilder<> IRB(&I);
2325 Value *S0 = IRB.CreateOr(getShadow(&I, 0), getShadow(&I, 1));
2326 Value *S = LowerElementShadowExtend(IRB, S0, getShadowTy(&I));
2328 setOriginForNaryOp(I);
2331 void handleStmxcsr(IntrinsicInst &I) {
2332 IRBuilder<> IRB(&I);
2333 Value* Addr = I.getArgOperand(0);
2334 Type *Ty = IRB.getInt32Ty();
2335 Value *ShadowPtr = getShadowPtr(Addr, Ty, IRB);
2337 IRB.CreateStore(getCleanShadow(Ty),
2338 IRB.CreatePointerCast(ShadowPtr, Ty->getPointerTo()));
2340 if (ClCheckAccessAddress)
2341 insertShadowCheck(Addr, &I);
2344 void handleLdmxcsr(IntrinsicInst &I) {
2345 if (!InsertChecks) return;
2347 IRBuilder<> IRB(&I);
2348 Value *Addr = I.getArgOperand(0);
2349 Type *Ty = IRB.getInt32Ty();
2350 unsigned Alignment = 1;
2352 if (ClCheckAccessAddress)
2353 insertShadowCheck(Addr, &I);
2355 Value *Shadow = IRB.CreateAlignedLoad(getShadowPtr(Addr, Ty, IRB),
2356 Alignment, "_ldmxcsr");
2357 Value *Origin = MS.TrackOrigins
2358 ? IRB.CreateLoad(getOriginPtr(Addr, IRB, Alignment))
2360 insertShadowCheck(Shadow, Origin, &I);
2363 void visitIntrinsicInst(IntrinsicInst &I) {
2364 switch (I.getIntrinsicID()) {
2365 case llvm::Intrinsic::bswap:
2368 case llvm::Intrinsic::x86_sse_stmxcsr:
2371 case llvm::Intrinsic::x86_sse_ldmxcsr:
2374 case llvm::Intrinsic::x86_avx512_vcvtsd2usi64:
2375 case llvm::Intrinsic::x86_avx512_vcvtsd2usi32:
2376 case llvm::Intrinsic::x86_avx512_vcvtss2usi64:
2377 case llvm::Intrinsic::x86_avx512_vcvtss2usi32:
2378 case llvm::Intrinsic::x86_avx512_cvttss2usi64:
2379 case llvm::Intrinsic::x86_avx512_cvttss2usi:
2380 case llvm::Intrinsic::x86_avx512_cvttsd2usi64:
2381 case llvm::Intrinsic::x86_avx512_cvttsd2usi:
2382 case llvm::Intrinsic::x86_avx512_cvtusi2sd:
2383 case llvm::Intrinsic::x86_avx512_cvtusi2ss:
2384 case llvm::Intrinsic::x86_avx512_cvtusi642sd:
2385 case llvm::Intrinsic::x86_avx512_cvtusi642ss:
2386 case llvm::Intrinsic::x86_sse2_cvtsd2si64:
2387 case llvm::Intrinsic::x86_sse2_cvtsd2si:
2388 case llvm::Intrinsic::x86_sse2_cvtsd2ss:
2389 case llvm::Intrinsic::x86_sse2_cvtsi2sd:
2390 case llvm::Intrinsic::x86_sse2_cvtsi642sd:
2391 case llvm::Intrinsic::x86_sse2_cvtss2sd:
2392 case llvm::Intrinsic::x86_sse2_cvttsd2si64:
2393 case llvm::Intrinsic::x86_sse2_cvttsd2si:
2394 case llvm::Intrinsic::x86_sse_cvtsi2ss:
2395 case llvm::Intrinsic::x86_sse_cvtsi642ss:
2396 case llvm::Intrinsic::x86_sse_cvtss2si64:
2397 case llvm::Intrinsic::x86_sse_cvtss2si:
2398 case llvm::Intrinsic::x86_sse_cvttss2si64:
2399 case llvm::Intrinsic::x86_sse_cvttss2si:
2400 handleVectorConvertIntrinsic(I, 1);
2402 case llvm::Intrinsic::x86_sse_cvtps2pi:
2403 case llvm::Intrinsic::x86_sse_cvttps2pi:
2404 handleVectorConvertIntrinsic(I, 2);
2407 case llvm::Intrinsic::x86_avx512_psll_w_512:
2408 case llvm::Intrinsic::x86_avx512_psll_d_512:
2409 case llvm::Intrinsic::x86_avx512_psll_q_512:
2410 case llvm::Intrinsic::x86_avx512_pslli_w_512:
2411 case llvm::Intrinsic::x86_avx512_pslli_d_512:
2412 case llvm::Intrinsic::x86_avx512_pslli_q_512:
2413 case llvm::Intrinsic::x86_avx512_psrl_w_512:
2414 case llvm::Intrinsic::x86_avx512_psrl_d_512:
2415 case llvm::Intrinsic::x86_avx512_psrl_q_512:
2416 case llvm::Intrinsic::x86_avx512_psra_w_512:
2417 case llvm::Intrinsic::x86_avx512_psra_d_512:
2418 case llvm::Intrinsic::x86_avx512_psra_q_512:
2419 case llvm::Intrinsic::x86_avx512_psrli_w_512:
2420 case llvm::Intrinsic::x86_avx512_psrli_d_512:
2421 case llvm::Intrinsic::x86_avx512_psrli_q_512:
2422 case llvm::Intrinsic::x86_avx512_psrai_w_512:
2423 case llvm::Intrinsic::x86_avx512_psrai_d_512:
2424 case llvm::Intrinsic::x86_avx512_psrai_q_512:
2425 case llvm::Intrinsic::x86_avx512_psra_q_256:
2426 case llvm::Intrinsic::x86_avx512_psra_q_128:
2427 case llvm::Intrinsic::x86_avx512_psrai_q_256:
2428 case llvm::Intrinsic::x86_avx512_psrai_q_128:
2429 case llvm::Intrinsic::x86_avx2_psll_w:
2430 case llvm::Intrinsic::x86_avx2_psll_d:
2431 case llvm::Intrinsic::x86_avx2_psll_q:
2432 case llvm::Intrinsic::x86_avx2_pslli_w:
2433 case llvm::Intrinsic::x86_avx2_pslli_d:
2434 case llvm::Intrinsic::x86_avx2_pslli_q:
2435 case llvm::Intrinsic::x86_avx2_psrl_w:
2436 case llvm::Intrinsic::x86_avx2_psrl_d:
2437 case llvm::Intrinsic::x86_avx2_psrl_q:
2438 case llvm::Intrinsic::x86_avx2_psra_w:
2439 case llvm::Intrinsic::x86_avx2_psra_d:
2440 case llvm::Intrinsic::x86_avx2_psrli_w:
2441 case llvm::Intrinsic::x86_avx2_psrli_d:
2442 case llvm::Intrinsic::x86_avx2_psrli_q:
2443 case llvm::Intrinsic::x86_avx2_psrai_w:
2444 case llvm::Intrinsic::x86_avx2_psrai_d:
2445 case llvm::Intrinsic::x86_sse2_psll_w:
2446 case llvm::Intrinsic::x86_sse2_psll_d:
2447 case llvm::Intrinsic::x86_sse2_psll_q:
2448 case llvm::Intrinsic::x86_sse2_pslli_w:
2449 case llvm::Intrinsic::x86_sse2_pslli_d:
2450 case llvm::Intrinsic::x86_sse2_pslli_q:
2451 case llvm::Intrinsic::x86_sse2_psrl_w:
2452 case llvm::Intrinsic::x86_sse2_psrl_d:
2453 case llvm::Intrinsic::x86_sse2_psrl_q:
2454 case llvm::Intrinsic::x86_sse2_psra_w:
2455 case llvm::Intrinsic::x86_sse2_psra_d:
2456 case llvm::Intrinsic::x86_sse2_psrli_w:
2457 case llvm::Intrinsic::x86_sse2_psrli_d:
2458 case llvm::Intrinsic::x86_sse2_psrli_q:
2459 case llvm::Intrinsic::x86_sse2_psrai_w:
2460 case llvm::Intrinsic::x86_sse2_psrai_d:
2461 case llvm::Intrinsic::x86_mmx_psll_w:
2462 case llvm::Intrinsic::x86_mmx_psll_d:
2463 case llvm::Intrinsic::x86_mmx_psll_q:
2464 case llvm::Intrinsic::x86_mmx_pslli_w:
2465 case llvm::Intrinsic::x86_mmx_pslli_d:
2466 case llvm::Intrinsic::x86_mmx_pslli_q:
2467 case llvm::Intrinsic::x86_mmx_psrl_w:
2468 case llvm::Intrinsic::x86_mmx_psrl_d:
2469 case llvm::Intrinsic::x86_mmx_psrl_q:
2470 case llvm::Intrinsic::x86_mmx_psra_w:
2471 case llvm::Intrinsic::x86_mmx_psra_d:
2472 case llvm::Intrinsic::x86_mmx_psrli_w:
2473 case llvm::Intrinsic::x86_mmx_psrli_d:
2474 case llvm::Intrinsic::x86_mmx_psrli_q:
2475 case llvm::Intrinsic::x86_mmx_psrai_w:
2476 case llvm::Intrinsic::x86_mmx_psrai_d:
2477 handleVectorShiftIntrinsic(I, /* Variable */ false);
2479 case llvm::Intrinsic::x86_avx2_psllv_d:
2480 case llvm::Intrinsic::x86_avx2_psllv_d_256:
2481 case llvm::Intrinsic::x86_avx512_psllv_d_512:
2482 case llvm::Intrinsic::x86_avx2_psllv_q:
2483 case llvm::Intrinsic::x86_avx2_psllv_q_256:
2484 case llvm::Intrinsic::x86_avx512_psllv_q_512:
2485 case llvm::Intrinsic::x86_avx2_psrlv_d:
2486 case llvm::Intrinsic::x86_avx2_psrlv_d_256:
2487 case llvm::Intrinsic::x86_avx512_psrlv_d_512:
2488 case llvm::Intrinsic::x86_avx2_psrlv_q:
2489 case llvm::Intrinsic::x86_avx2_psrlv_q_256:
2490 case llvm::Intrinsic::x86_avx512_psrlv_q_512:
2491 case llvm::Intrinsic::x86_avx2_psrav_d:
2492 case llvm::Intrinsic::x86_avx2_psrav_d_256:
2493 case llvm::Intrinsic::x86_avx512_psrav_d_512:
2494 case llvm::Intrinsic::x86_avx512_psrav_q_128:
2495 case llvm::Intrinsic::x86_avx512_psrav_q_256:
2496 case llvm::Intrinsic::x86_avx512_psrav_q_512:
2497 handleVectorShiftIntrinsic(I, /* Variable */ true);
2500 case llvm::Intrinsic::x86_sse2_packsswb_128:
2501 case llvm::Intrinsic::x86_sse2_packssdw_128:
2502 case llvm::Intrinsic::x86_sse2_packuswb_128:
2503 case llvm::Intrinsic::x86_sse41_packusdw:
2504 case llvm::Intrinsic::x86_avx2_packsswb:
2505 case llvm::Intrinsic::x86_avx2_packssdw:
2506 case llvm::Intrinsic::x86_avx2_packuswb:
2507 case llvm::Intrinsic::x86_avx2_packusdw:
2508 handleVectorPackIntrinsic(I);
2511 case llvm::Intrinsic::x86_mmx_packsswb:
2512 case llvm::Intrinsic::x86_mmx_packuswb:
2513 handleVectorPackIntrinsic(I, 16);
2516 case llvm::Intrinsic::x86_mmx_packssdw:
2517 handleVectorPackIntrinsic(I, 32);
2520 case llvm::Intrinsic::x86_mmx_psad_bw:
2521 case llvm::Intrinsic::x86_sse2_psad_bw:
2522 case llvm::Intrinsic::x86_avx2_psad_bw:
2523 handleVectorSadIntrinsic(I);
2526 case llvm::Intrinsic::x86_sse2_pmadd_wd:
2527 case llvm::Intrinsic::x86_avx2_pmadd_wd:
2528 case llvm::Intrinsic::x86_ssse3_pmadd_ub_sw_128:
2529 case llvm::Intrinsic::x86_avx2_pmadd_ub_sw:
2530 handleVectorPmaddIntrinsic(I);
2533 case llvm::Intrinsic::x86_ssse3_pmadd_ub_sw:
2534 handleVectorPmaddIntrinsic(I, 8);
2537 case llvm::Intrinsic::x86_mmx_pmadd_wd:
2538 handleVectorPmaddIntrinsic(I, 16);
2541 case llvm::Intrinsic::x86_sse_cmp_ss:
2542 case llvm::Intrinsic::x86_sse2_cmp_sd:
2543 case llvm::Intrinsic::x86_sse_comieq_ss:
2544 case llvm::Intrinsic::x86_sse_comilt_ss:
2545 case llvm::Intrinsic::x86_sse_comile_ss:
2546 case llvm::Intrinsic::x86_sse_comigt_ss:
2547 case llvm::Intrinsic::x86_sse_comige_ss:
2548 case llvm::Intrinsic::x86_sse_comineq_ss:
2549 case llvm::Intrinsic::x86_sse_ucomieq_ss:
2550 case llvm::Intrinsic::x86_sse_ucomilt_ss:
2551 case llvm::Intrinsic::x86_sse_ucomile_ss:
2552 case llvm::Intrinsic::x86_sse_ucomigt_ss:
2553 case llvm::Intrinsic::x86_sse_ucomige_ss:
2554 case llvm::Intrinsic::x86_sse_ucomineq_ss:
2555 case llvm::Intrinsic::x86_sse2_comieq_sd:
2556 case llvm::Intrinsic::x86_sse2_comilt_sd:
2557 case llvm::Intrinsic::x86_sse2_comile_sd:
2558 case llvm::Intrinsic::x86_sse2_comigt_sd:
2559 case llvm::Intrinsic::x86_sse2_comige_sd:
2560 case llvm::Intrinsic::x86_sse2_comineq_sd:
2561 case llvm::Intrinsic::x86_sse2_ucomieq_sd:
2562 case llvm::Intrinsic::x86_sse2_ucomilt_sd:
2563 case llvm::Intrinsic::x86_sse2_ucomile_sd:
2564 case llvm::Intrinsic::x86_sse2_ucomigt_sd:
2565 case llvm::Intrinsic::x86_sse2_ucomige_sd:
2566 case llvm::Intrinsic::x86_sse2_ucomineq_sd:
2567 handleVectorCompareScalarIntrinsic(I);
2570 case llvm::Intrinsic::x86_sse_cmp_ps:
2571 case llvm::Intrinsic::x86_sse2_cmp_pd:
2572 // FIXME: For x86_avx_cmp_pd_256 and x86_avx_cmp_ps_256 this function
2573 // generates reasonably looking IR that fails in the backend with "Do not
2574 // know how to split the result of this operator!".
2575 handleVectorComparePackedIntrinsic(I);
2579 if (!handleUnknownIntrinsic(I))
2580 visitInstruction(I);
2585 void visitCallSite(CallSite CS) {
2586 Instruction &I = *CS.getInstruction();
2587 assert((CS.isCall() || CS.isInvoke()) && "Unknown type of CallSite");
2589 CallInst *Call = cast<CallInst>(&I);
2591 // For inline asm, do the usual thing: check argument shadow and mark all
2592 // outputs as clean. Note that any side effects of the inline asm that are
2593 // not immediately visible in its constraints are not handled.
2594 if (Call->isInlineAsm()) {
2595 visitInstruction(I);
2599 assert(!isa<IntrinsicInst>(&I) && "intrinsics are handled elsewhere");
2601 // We are going to insert code that relies on the fact that the callee
2602 // will become a non-readonly function after it is instrumented by us. To
2603 // prevent this code from being optimized out, mark that function
2604 // non-readonly in advance.
2605 if (Function *Func = Call->getCalledFunction()) {
2606 // Clear out readonly/readnone attributes.
2608 B.addAttribute(Attribute::ReadOnly)
2609 .addAttribute(Attribute::ReadNone);
2610 Func->removeAttributes(AttributeList::FunctionIndex,
2611 AttributeList::get(Func->getContext(),
2612 AttributeList::FunctionIndex,
2616 maybeMarkSanitizerLibraryCallNoBuiltin(Call, TLI);
2618 IRBuilder<> IRB(&I);
2620 unsigned ArgOffset = 0;
2621 DEBUG(dbgs() << " CallSite: " << I << "\n");
2622 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
2623 ArgIt != End; ++ArgIt) {
2625 unsigned i = ArgIt - CS.arg_begin();
2626 if (!A->getType()->isSized()) {
2627 DEBUG(dbgs() << "Arg " << i << " is not sized: " << I << "\n");
2631 Value *Store = nullptr;
2632 // Compute the Shadow for arg even if it is ByVal, because
2633 // in that case getShadow() will copy the actual arg shadow to
2634 // __msan_param_tls.
2635 Value *ArgShadow = getShadow(A);
2636 Value *ArgShadowBase = getShadowPtrForArgument(A, IRB, ArgOffset);
2637 DEBUG(dbgs() << " Arg#" << i << ": " << *A <<
2638 " Shadow: " << *ArgShadow << "\n");
2639 bool ArgIsInitialized = false;
2640 const DataLayout &DL = F.getParent()->getDataLayout();
2641 if (CS.paramHasAttr(i, Attribute::ByVal)) {
2642 assert(A->getType()->isPointerTy() &&
2643 "ByVal argument is not a pointer!");
2644 Size = DL.getTypeAllocSize(A->getType()->getPointerElementType());
2645 if (ArgOffset + Size > kParamTLSSize) break;
2646 unsigned ParamAlignment = CS.getParamAlignment(i + 1);
2647 unsigned Alignment = std::min(ParamAlignment, kShadowTLSAlignment);
2648 Store = IRB.CreateMemCpy(ArgShadowBase,
2649 getShadowPtr(A, Type::getInt8Ty(*MS.C), IRB),
2652 Size = DL.getTypeAllocSize(A->getType());
2653 if (ArgOffset + Size > kParamTLSSize) break;
2654 Store = IRB.CreateAlignedStore(ArgShadow, ArgShadowBase,
2655 kShadowTLSAlignment);
2656 Constant *Cst = dyn_cast<Constant>(ArgShadow);
2657 if (Cst && Cst->isNullValue()) ArgIsInitialized = true;
2659 if (MS.TrackOrigins && !ArgIsInitialized)
2660 IRB.CreateStore(getOrigin(A),
2661 getOriginPtrForArgument(A, IRB, ArgOffset));
2663 assert(Size != 0 && Store != nullptr);
2664 DEBUG(dbgs() << " Param:" << *Store << "\n");
2665 ArgOffset += alignTo(Size, 8);
2667 DEBUG(dbgs() << " done with call args\n");
2670 cast<FunctionType>(CS.getCalledValue()->getType()->getContainedType(0));
2671 if (FT->isVarArg()) {
2672 VAHelper->visitCallSite(CS, IRB);
2675 // Now, get the shadow for the RetVal.
2676 if (!I.getType()->isSized()) return;
2677 // Don't emit the epilogue for musttail call returns.
2678 if (CS.isCall() && cast<CallInst>(&I)->isMustTailCall()) return;
2679 IRBuilder<> IRBBefore(&I);
2680 // Until we have full dynamic coverage, make sure the retval shadow is 0.
2681 Value *Base = getShadowPtrForRetval(&I, IRBBefore);
2682 IRBBefore.CreateAlignedStore(getCleanShadow(&I), Base, kShadowTLSAlignment);
2683 BasicBlock::iterator NextInsn;
2685 NextInsn = ++I.getIterator();
2686 assert(NextInsn != I.getParent()->end());
2688 BasicBlock *NormalDest = cast<InvokeInst>(&I)->getNormalDest();
2689 if (!NormalDest->getSinglePredecessor()) {
2690 // FIXME: this case is tricky, so we are just conservative here.
2691 // Perhaps we need to split the edge between this BB and NormalDest,
2692 // but a naive attempt to use SplitEdge leads to a crash.
2693 setShadow(&I, getCleanShadow(&I));
2694 setOrigin(&I, getCleanOrigin());
2697 NextInsn = NormalDest->getFirstInsertionPt();
2698 assert(NextInsn != NormalDest->end() &&
2699 "Could not find insertion point for retval shadow load");
2701 IRBuilder<> IRBAfter(&*NextInsn);
2702 Value *RetvalShadow =
2703 IRBAfter.CreateAlignedLoad(getShadowPtrForRetval(&I, IRBAfter),
2704 kShadowTLSAlignment, "_msret");
2705 setShadow(&I, RetvalShadow);
2706 if (MS.TrackOrigins)
2707 setOrigin(&I, IRBAfter.CreateLoad(getOriginPtrForRetval(IRBAfter)));
2710 bool isAMustTailRetVal(Value *RetVal) {
2711 if (auto *I = dyn_cast<BitCastInst>(RetVal)) {
2712 RetVal = I->getOperand(0);
2714 if (auto *I = dyn_cast<CallInst>(RetVal)) {
2715 return I->isMustTailCall();
2720 void visitReturnInst(ReturnInst &I) {
2721 IRBuilder<> IRB(&I);
2722 Value *RetVal = I.getReturnValue();
2723 if (!RetVal) return;
2724 // Don't emit the epilogue for musttail call returns.
2725 if (isAMustTailRetVal(RetVal)) return;
2726 Value *ShadowPtr = getShadowPtrForRetval(RetVal, IRB);
2727 if (CheckReturnValue) {
2728 insertShadowCheck(RetVal, &I);
2729 Value *Shadow = getCleanShadow(RetVal);
2730 IRB.CreateAlignedStore(Shadow, ShadowPtr, kShadowTLSAlignment);
2732 Value *Shadow = getShadow(RetVal);
2733 IRB.CreateAlignedStore(Shadow, ShadowPtr, kShadowTLSAlignment);
2734 if (MS.TrackOrigins)
2735 IRB.CreateStore(getOrigin(RetVal), getOriginPtrForRetval(IRB));
2739 void visitPHINode(PHINode &I) {
2740 IRBuilder<> IRB(&I);
2741 if (!PropagateShadow) {
2742 setShadow(&I, getCleanShadow(&I));
2743 setOrigin(&I, getCleanOrigin());
2747 ShadowPHINodes.push_back(&I);
2748 setShadow(&I, IRB.CreatePHI(getShadowTy(&I), I.getNumIncomingValues(),
2750 if (MS.TrackOrigins)
2751 setOrigin(&I, IRB.CreatePHI(MS.OriginTy, I.getNumIncomingValues(),
2755 void visitAllocaInst(AllocaInst &I) {
2756 setShadow(&I, getCleanShadow(&I));
2757 setOrigin(&I, getCleanOrigin());
2758 IRBuilder<> IRB(I.getNextNode());
2759 const DataLayout &DL = F.getParent()->getDataLayout();
2760 uint64_t TypeSize = DL.getTypeAllocSize(I.getAllocatedType());
2761 Value *Len = ConstantInt::get(MS.IntptrTy, TypeSize);
2762 if (I.isArrayAllocation())
2763 Len = IRB.CreateMul(Len, I.getArraySize());
2764 if (PoisonStack && ClPoisonStackWithCall) {
2765 IRB.CreateCall(MS.MsanPoisonStackFn,
2766 {IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()), Len});
2768 Value *ShadowBase = getShadowPtr(&I, Type::getInt8PtrTy(*MS.C), IRB);
2769 Value *PoisonValue = IRB.getInt8(PoisonStack ? ClPoisonStackPattern : 0);
2770 IRB.CreateMemSet(ShadowBase, PoisonValue, Len, I.getAlignment());
2773 if (PoisonStack && MS.TrackOrigins) {
2774 SmallString<2048> StackDescriptionStorage;
2775 raw_svector_ostream StackDescription(StackDescriptionStorage);
2776 // We create a string with a description of the stack allocation and
2777 // pass it into __msan_set_alloca_origin.
2778 // It will be printed by the run-time if stack-originated UMR is found.
2779 // The first 4 bytes of the string are set to '----' and will be replaced
2780 // by __msan_va_arg_overflow_size_tls at the first call.
2781 StackDescription << "----" << I.getName() << "@" << F.getName();
2783 createPrivateNonConstGlobalForString(*F.getParent(),
2784 StackDescription.str());
2786 IRB.CreateCall(MS.MsanSetAllocaOrigin4Fn,
2787 {IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()), Len,
2788 IRB.CreatePointerCast(Descr, IRB.getInt8PtrTy()),
2789 IRB.CreatePointerCast(&F, MS.IntptrTy)});
2793 void visitSelectInst(SelectInst& I) {
2794 IRBuilder<> IRB(&I);
2795 // a = select b, c, d
2796 Value *B = I.getCondition();
2797 Value *C = I.getTrueValue();
2798 Value *D = I.getFalseValue();
2799 Value *Sb = getShadow(B);
2800 Value *Sc = getShadow(C);
2801 Value *Sd = getShadow(D);
2803 // Result shadow if condition shadow is 0.
2804 Value *Sa0 = IRB.CreateSelect(B, Sc, Sd);
2806 if (I.getType()->isAggregateType()) {
2807 // To avoid "sign extending" i1 to an arbitrary aggregate type, we just do
2808 // an extra "select". This results in much more compact IR.
2809 // Sa = select Sb, poisoned, (select b, Sc, Sd)
2810 Sa1 = getPoisonedShadow(getShadowTy(I.getType()));
2812 // Sa = select Sb, [ (c^d) | Sc | Sd ], [ b ? Sc : Sd ]
2813 // If Sb (condition is poisoned), look for bits in c and d that are equal
2814 // and both unpoisoned.
2815 // If !Sb (condition is unpoisoned), simply pick one of Sc and Sd.
2817 // Cast arguments to shadow-compatible type.
2818 C = CreateAppToShadowCast(IRB, C);
2819 D = CreateAppToShadowCast(IRB, D);
2821 // Result shadow if condition shadow is 1.
2822 Sa1 = IRB.CreateOr(IRB.CreateXor(C, D), IRB.CreateOr(Sc, Sd));
2824 Value *Sa = IRB.CreateSelect(Sb, Sa1, Sa0, "_msprop_select");
2826 if (MS.TrackOrigins) {
2827 // Origins are always i32, so any vector conditions must be flattened.
2828 // FIXME: consider tracking vector origins for app vectors?
2829 if (B->getType()->isVectorTy()) {
2830 Type *FlatTy = getShadowTyNoVec(B->getType());
2831 B = IRB.CreateICmpNE(IRB.CreateBitCast(B, FlatTy),
2832 ConstantInt::getNullValue(FlatTy));
2833 Sb = IRB.CreateICmpNE(IRB.CreateBitCast(Sb, FlatTy),
2834 ConstantInt::getNullValue(FlatTy));
2836 // a = select b, c, d
2837 // Oa = Sb ? Ob : (b ? Oc : Od)
2839 &I, IRB.CreateSelect(Sb, getOrigin(I.getCondition()),
2840 IRB.CreateSelect(B, getOrigin(I.getTrueValue()),
2841 getOrigin(I.getFalseValue()))));
2845 void visitLandingPadInst(LandingPadInst &I) {
2847 // See http://code.google.com/p/memory-sanitizer/issues/detail?id=1
2848 setShadow(&I, getCleanShadow(&I));
2849 setOrigin(&I, getCleanOrigin());
2852 void visitCatchSwitchInst(CatchSwitchInst &I) {
2853 setShadow(&I, getCleanShadow(&I));
2854 setOrigin(&I, getCleanOrigin());
2857 void visitFuncletPadInst(FuncletPadInst &I) {
2858 setShadow(&I, getCleanShadow(&I));
2859 setOrigin(&I, getCleanOrigin());
2862 void visitGetElementPtrInst(GetElementPtrInst &I) {
2866 void visitExtractValueInst(ExtractValueInst &I) {
2867 IRBuilder<> IRB(&I);
2868 Value *Agg = I.getAggregateOperand();
2869 DEBUG(dbgs() << "ExtractValue: " << I << "\n");
2870 Value *AggShadow = getShadow(Agg);
2871 DEBUG(dbgs() << " AggShadow: " << *AggShadow << "\n");
2872 Value *ResShadow = IRB.CreateExtractValue(AggShadow, I.getIndices());
2873 DEBUG(dbgs() << " ResShadow: " << *ResShadow << "\n");
2874 setShadow(&I, ResShadow);
2875 setOriginForNaryOp(I);
2878 void visitInsertValueInst(InsertValueInst &I) {
2879 IRBuilder<> IRB(&I);
2880 DEBUG(dbgs() << "InsertValue: " << I << "\n");
2881 Value *AggShadow = getShadow(I.getAggregateOperand());
2882 Value *InsShadow = getShadow(I.getInsertedValueOperand());
2883 DEBUG(dbgs() << " AggShadow: " << *AggShadow << "\n");
2884 DEBUG(dbgs() << " InsShadow: " << *InsShadow << "\n");
2885 Value *Res = IRB.CreateInsertValue(AggShadow, InsShadow, I.getIndices());
2886 DEBUG(dbgs() << " Res: " << *Res << "\n");
2888 setOriginForNaryOp(I);
2891 void dumpInst(Instruction &I) {
2892 if (CallInst *CI = dyn_cast<CallInst>(&I)) {
2893 errs() << "ZZZ call " << CI->getCalledFunction()->getName() << "\n";
2895 errs() << "ZZZ " << I.getOpcodeName() << "\n";
2897 errs() << "QQQ " << I << "\n";
2900 void visitResumeInst(ResumeInst &I) {
2901 DEBUG(dbgs() << "Resume: " << I << "\n");
2902 // Nothing to do here.
2905 void visitCleanupReturnInst(CleanupReturnInst &CRI) {
2906 DEBUG(dbgs() << "CleanupReturn: " << CRI << "\n");
2907 // Nothing to do here.
2910 void visitCatchReturnInst(CatchReturnInst &CRI) {
2911 DEBUG(dbgs() << "CatchReturn: " << CRI << "\n");
2912 // Nothing to do here.
2915 void visitInstruction(Instruction &I) {
2916 // Everything else: stop propagating and check for poisoned shadow.
2917 if (ClDumpStrictInstructions)
2919 DEBUG(dbgs() << "DEFAULT: " << I << "\n");
2920 for (size_t i = 0, n = I.getNumOperands(); i < n; i++)
2921 insertShadowCheck(I.getOperand(i), &I);
2922 setShadow(&I, getCleanShadow(&I));
2923 setOrigin(&I, getCleanOrigin());
2927 /// \brief AMD64-specific implementation of VarArgHelper.
2928 struct VarArgAMD64Helper : public VarArgHelper {
2929 // An unfortunate workaround for asymmetric lowering of va_arg stuff.
2930 // See a comment in visitCallSite for more details.
2931 static const unsigned AMD64GpEndOffset = 48; // AMD64 ABI Draft 0.99.6 p3.5.7
2932 static const unsigned AMD64FpEndOffset = 176;
2935 MemorySanitizer &MS;
2936 MemorySanitizerVisitor &MSV;
2937 Value *VAArgTLSCopy;
2938 Value *VAArgOverflowSize;
2940 SmallVector<CallInst*, 16> VAStartInstrumentationList;
2942 VarArgAMD64Helper(Function &F, MemorySanitizer &MS,
2943 MemorySanitizerVisitor &MSV)
2944 : F(F), MS(MS), MSV(MSV), VAArgTLSCopy(nullptr),
2945 VAArgOverflowSize(nullptr) {}
2947 enum ArgKind { AK_GeneralPurpose, AK_FloatingPoint, AK_Memory };
2949 ArgKind classifyArgument(Value* arg) {
2950 // A very rough approximation of X86_64 argument classification rules.
2951 Type *T = arg->getType();
2952 if (T->isFPOrFPVectorTy() || T->isX86_MMXTy())
2953 return AK_FloatingPoint;
2954 if (T->isIntegerTy() && T->getPrimitiveSizeInBits() <= 64)
2955 return AK_GeneralPurpose;
2956 if (T->isPointerTy())
2957 return AK_GeneralPurpose;
2961 // For VarArg functions, store the argument shadow in an ABI-specific format
2962 // that corresponds to va_list layout.
2963 // We do this because Clang lowers va_arg in the frontend, and this pass
2964 // only sees the low level code that deals with va_list internals.
2965 // A much easier alternative (provided that Clang emits va_arg instructions)
2966 // would have been to associate each live instance of va_list with a copy of
2967 // MSanParamTLS, and extract shadow on va_arg() call in the argument list
2969 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) override {
2970 unsigned GpOffset = 0;
2971 unsigned FpOffset = AMD64GpEndOffset;
2972 unsigned OverflowOffset = AMD64FpEndOffset;
2973 const DataLayout &DL = F.getParent()->getDataLayout();
2974 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
2975 ArgIt != End; ++ArgIt) {
2977 unsigned ArgNo = CS.getArgumentNo(ArgIt);
2978 bool IsFixed = ArgNo < CS.getFunctionType()->getNumParams();
2979 bool IsByVal = CS.paramHasAttr(ArgNo, Attribute::ByVal);
2981 // ByVal arguments always go to the overflow area.
2982 // Fixed arguments passed through the overflow area will be stepped
2983 // over by va_start, so don't count them towards the offset.
2986 assert(A->getType()->isPointerTy());
2987 Type *RealTy = A->getType()->getPointerElementType();
2988 uint64_t ArgSize = DL.getTypeAllocSize(RealTy);
2989 Value *Base = getShadowPtrForVAArgument(RealTy, IRB, OverflowOffset);
2990 OverflowOffset += alignTo(ArgSize, 8);
2991 IRB.CreateMemCpy(Base, MSV.getShadowPtr(A, IRB.getInt8Ty(), IRB),
2992 ArgSize, kShadowTLSAlignment);
2994 ArgKind AK = classifyArgument(A);
2995 if (AK == AK_GeneralPurpose && GpOffset >= AMD64GpEndOffset)
2997 if (AK == AK_FloatingPoint && FpOffset >= AMD64FpEndOffset)
3001 case AK_GeneralPurpose:
3002 Base = getShadowPtrForVAArgument(A->getType(), IRB, GpOffset);
3005 case AK_FloatingPoint:
3006 Base = getShadowPtrForVAArgument(A->getType(), IRB, FpOffset);
3012 uint64_t ArgSize = DL.getTypeAllocSize(A->getType());
3013 Base = getShadowPtrForVAArgument(A->getType(), IRB, OverflowOffset);
3014 OverflowOffset += alignTo(ArgSize, 8);
3016 // Take fixed arguments into account for GpOffset and FpOffset,
3017 // but don't actually store shadows for them.
3020 IRB.CreateAlignedStore(MSV.getShadow(A), Base, kShadowTLSAlignment);
3023 Constant *OverflowSize =
3024 ConstantInt::get(IRB.getInt64Ty(), OverflowOffset - AMD64FpEndOffset);
3025 IRB.CreateStore(OverflowSize, MS.VAArgOverflowSizeTLS);
3028 /// \brief Compute the shadow address for a given va_arg.
3029 Value *getShadowPtrForVAArgument(Type *Ty, IRBuilder<> &IRB,
3031 Value *Base = IRB.CreatePointerCast(MS.VAArgTLS, MS.IntptrTy);
3032 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
3033 return IRB.CreateIntToPtr(Base, PointerType::get(MSV.getShadowTy(Ty), 0),
3037 void visitVAStartInst(VAStartInst &I) override {
3038 if (F.getCallingConv() == CallingConv::X86_64_Win64)
3040 IRBuilder<> IRB(&I);
3041 VAStartInstrumentationList.push_back(&I);
3042 Value *VAListTag = I.getArgOperand(0);
3043 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
3045 // Unpoison the whole __va_list_tag.
3046 // FIXME: magic ABI constants.
3047 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
3048 /* size */24, /* alignment */8, false);
3051 void visitVACopyInst(VACopyInst &I) override {
3052 if (F.getCallingConv() == CallingConv::X86_64_Win64)
3054 IRBuilder<> IRB(&I);
3055 Value *VAListTag = I.getArgOperand(0);
3056 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
3058 // Unpoison the whole __va_list_tag.
3059 // FIXME: magic ABI constants.
3060 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
3061 /* size */24, /* alignment */8, false);
3064 void finalizeInstrumentation() override {
3065 assert(!VAArgOverflowSize && !VAArgTLSCopy &&
3066 "finalizeInstrumentation called twice");
3067 if (!VAStartInstrumentationList.empty()) {
3068 // If there is a va_start in this function, make a backup copy of
3069 // va_arg_tls somewhere in the function entry block.
3070 IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
3071 VAArgOverflowSize = IRB.CreateLoad(MS.VAArgOverflowSizeTLS);
3073 IRB.CreateAdd(ConstantInt::get(MS.IntptrTy, AMD64FpEndOffset),
3075 VAArgTLSCopy = IRB.CreateAlloca(Type::getInt8Ty(*MS.C), CopySize);
3076 IRB.CreateMemCpy(VAArgTLSCopy, MS.VAArgTLS, CopySize, 8);
3079 // Instrument va_start.
3080 // Copy va_list shadow from the backup copy of the TLS contents.
3081 for (size_t i = 0, n = VAStartInstrumentationList.size(); i < n; i++) {
3082 CallInst *OrigInst = VAStartInstrumentationList[i];
3083 IRBuilder<> IRB(OrigInst->getNextNode());
3084 Value *VAListTag = OrigInst->getArgOperand(0);
3086 Value *RegSaveAreaPtrPtr =
3088 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
3089 ConstantInt::get(MS.IntptrTy, 16)),
3090 Type::getInt64PtrTy(*MS.C));
3091 Value *RegSaveAreaPtr = IRB.CreateLoad(RegSaveAreaPtrPtr);
3092 Value *RegSaveAreaShadowPtr =
3093 MSV.getShadowPtr(RegSaveAreaPtr, IRB.getInt8Ty(), IRB);
3094 IRB.CreateMemCpy(RegSaveAreaShadowPtr, VAArgTLSCopy,
3095 AMD64FpEndOffset, 16);
3097 Value *OverflowArgAreaPtrPtr =
3099 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
3100 ConstantInt::get(MS.IntptrTy, 8)),
3101 Type::getInt64PtrTy(*MS.C));
3102 Value *OverflowArgAreaPtr = IRB.CreateLoad(OverflowArgAreaPtrPtr);
3103 Value *OverflowArgAreaShadowPtr =
3104 MSV.getShadowPtr(OverflowArgAreaPtr, IRB.getInt8Ty(), IRB);
3105 Value *SrcPtr = IRB.CreateConstGEP1_32(IRB.getInt8Ty(), VAArgTLSCopy,
3107 IRB.CreateMemCpy(OverflowArgAreaShadowPtr, SrcPtr, VAArgOverflowSize, 16);
3112 /// \brief MIPS64-specific implementation of VarArgHelper.
3113 struct VarArgMIPS64Helper : public VarArgHelper {
3115 MemorySanitizer &MS;
3116 MemorySanitizerVisitor &MSV;
3117 Value *VAArgTLSCopy;
3120 SmallVector<CallInst*, 16> VAStartInstrumentationList;
3122 VarArgMIPS64Helper(Function &F, MemorySanitizer &MS,
3123 MemorySanitizerVisitor &MSV)
3124 : F(F), MS(MS), MSV(MSV), VAArgTLSCopy(nullptr),
3125 VAArgSize(nullptr) {}
3127 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) override {
3128 unsigned VAArgOffset = 0;
3129 const DataLayout &DL = F.getParent()->getDataLayout();
3130 for (CallSite::arg_iterator ArgIt = CS.arg_begin() +
3131 CS.getFunctionType()->getNumParams(), End = CS.arg_end();
3132 ArgIt != End; ++ArgIt) {
3133 llvm::Triple TargetTriple(F.getParent()->getTargetTriple());
3136 uint64_t ArgSize = DL.getTypeAllocSize(A->getType());
3137 if (TargetTriple.getArch() == llvm::Triple::mips64) {
3138 // Adjusting the shadow for argument with size < 8 to match the placement
3139 // of bits in big endian system
3141 VAArgOffset += (8 - ArgSize);
3143 Base = getShadowPtrForVAArgument(A->getType(), IRB, VAArgOffset);
3144 VAArgOffset += ArgSize;
3145 VAArgOffset = alignTo(VAArgOffset, 8);
3146 IRB.CreateAlignedStore(MSV.getShadow(A), Base, kShadowTLSAlignment);
3149 Constant *TotalVAArgSize = ConstantInt::get(IRB.getInt64Ty(), VAArgOffset);
3150 // Here using VAArgOverflowSizeTLS as VAArgSizeTLS to avoid creation of
3151 // a new class member i.e. it is the total size of all VarArgs.
3152 IRB.CreateStore(TotalVAArgSize, MS.VAArgOverflowSizeTLS);
3155 /// \brief Compute the shadow address for a given va_arg.
3156 Value *getShadowPtrForVAArgument(Type *Ty, IRBuilder<> &IRB,
3158 Value *Base = IRB.CreatePointerCast(MS.VAArgTLS, MS.IntptrTy);
3159 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
3160 return IRB.CreateIntToPtr(Base, PointerType::get(MSV.getShadowTy(Ty), 0),
3164 void visitVAStartInst(VAStartInst &I) override {
3165 IRBuilder<> IRB(&I);
3166 VAStartInstrumentationList.push_back(&I);
3167 Value *VAListTag = I.getArgOperand(0);
3168 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
3169 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
3170 /* size */8, /* alignment */8, false);
3173 void visitVACopyInst(VACopyInst &I) override {
3174 IRBuilder<> IRB(&I);
3175 Value *VAListTag = I.getArgOperand(0);
3176 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
3177 // Unpoison the whole __va_list_tag.
3178 // FIXME: magic ABI constants.
3179 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
3180 /* size */8, /* alignment */8, false);
3183 void finalizeInstrumentation() override {
3184 assert(!VAArgSize && !VAArgTLSCopy &&
3185 "finalizeInstrumentation called twice");
3186 IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
3187 VAArgSize = IRB.CreateLoad(MS.VAArgOverflowSizeTLS);
3188 Value *CopySize = IRB.CreateAdd(ConstantInt::get(MS.IntptrTy, 0),
3191 if (!VAStartInstrumentationList.empty()) {
3192 // If there is a va_start in this function, make a backup copy of
3193 // va_arg_tls somewhere in the function entry block.
3194 VAArgTLSCopy = IRB.CreateAlloca(Type::getInt8Ty(*MS.C), CopySize);
3195 IRB.CreateMemCpy(VAArgTLSCopy, MS.VAArgTLS, CopySize, 8);
3198 // Instrument va_start.
3199 // Copy va_list shadow from the backup copy of the TLS contents.
3200 for (size_t i = 0, n = VAStartInstrumentationList.size(); i < n; i++) {
3201 CallInst *OrigInst = VAStartInstrumentationList[i];
3202 IRBuilder<> IRB(OrigInst->getNextNode());
3203 Value *VAListTag = OrigInst->getArgOperand(0);
3204 Value *RegSaveAreaPtrPtr =
3205 IRB.CreateIntToPtr(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
3206 Type::getInt64PtrTy(*MS.C));
3207 Value *RegSaveAreaPtr = IRB.CreateLoad(RegSaveAreaPtrPtr);
3208 Value *RegSaveAreaShadowPtr =
3209 MSV.getShadowPtr(RegSaveAreaPtr, IRB.getInt8Ty(), IRB);
3210 IRB.CreateMemCpy(RegSaveAreaShadowPtr, VAArgTLSCopy, CopySize, 8);
3216 /// \brief AArch64-specific implementation of VarArgHelper.
3217 struct VarArgAArch64Helper : public VarArgHelper {
3218 static const unsigned kAArch64GrArgSize = 64;
3219 static const unsigned kAArch64VrArgSize = 128;
3221 static const unsigned AArch64GrBegOffset = 0;
3222 static const unsigned AArch64GrEndOffset = kAArch64GrArgSize;
3223 // Make VR space aligned to 16 bytes.
3224 static const unsigned AArch64VrBegOffset = AArch64GrEndOffset;
3225 static const unsigned AArch64VrEndOffset = AArch64VrBegOffset
3226 + kAArch64VrArgSize;
3227 static const unsigned AArch64VAEndOffset = AArch64VrEndOffset;
3230 MemorySanitizer &MS;
3231 MemorySanitizerVisitor &MSV;
3232 Value *VAArgTLSCopy;
3233 Value *VAArgOverflowSize;
3235 SmallVector<CallInst*, 16> VAStartInstrumentationList;
3237 VarArgAArch64Helper(Function &F, MemorySanitizer &MS,
3238 MemorySanitizerVisitor &MSV)
3239 : F(F), MS(MS), MSV(MSV), VAArgTLSCopy(nullptr),
3240 VAArgOverflowSize(nullptr) {}
3242 enum ArgKind { AK_GeneralPurpose, AK_FloatingPoint, AK_Memory };
3244 ArgKind classifyArgument(Value* arg) {
3245 Type *T = arg->getType();
3246 if (T->isFPOrFPVectorTy())
3247 return AK_FloatingPoint;
3248 if ((T->isIntegerTy() && T->getPrimitiveSizeInBits() <= 64)
3249 || (T->isPointerTy()))
3250 return AK_GeneralPurpose;
3254 // The instrumentation stores the argument shadow in a non ABI-specific
3255 // format because it does not know which argument is named (since Clang,
3256 // like x86_64 case, lowers the va_args in the frontend and this pass only
3257 // sees the low level code that deals with va_list internals).
3258 // The first seven GR registers are saved in the first 56 bytes of the
3259 // va_arg tls arra, followers by the first 8 FP/SIMD registers, and then
3260 // the remaining arguments.
3261 // Using constant offset within the va_arg TLS array allows fast copy
3262 // in the finalize instrumentation.
3263 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) override {
3264 unsigned GrOffset = AArch64GrBegOffset;
3265 unsigned VrOffset = AArch64VrBegOffset;
3266 unsigned OverflowOffset = AArch64VAEndOffset;
3268 const DataLayout &DL = F.getParent()->getDataLayout();
3269 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
3270 ArgIt != End; ++ArgIt) {
3272 unsigned ArgNo = CS.getArgumentNo(ArgIt);
3273 bool IsFixed = ArgNo < CS.getFunctionType()->getNumParams();
3274 ArgKind AK = classifyArgument(A);
3275 if (AK == AK_GeneralPurpose && GrOffset >= AArch64GrEndOffset)
3277 if (AK == AK_FloatingPoint && VrOffset >= AArch64VrEndOffset)
3281 case AK_GeneralPurpose:
3282 Base = getShadowPtrForVAArgument(A->getType(), IRB, GrOffset);
3285 case AK_FloatingPoint:
3286 Base = getShadowPtrForVAArgument(A->getType(), IRB, VrOffset);
3290 // Don't count fixed arguments in the overflow area - va_start will
3291 // skip right over them.
3294 uint64_t ArgSize = DL.getTypeAllocSize(A->getType());
3295 Base = getShadowPtrForVAArgument(A->getType(), IRB, OverflowOffset);
3296 OverflowOffset += alignTo(ArgSize, 8);
3299 // Count Gp/Vr fixed arguments to their respective offsets, but don't
3300 // bother to actually store a shadow.
3303 IRB.CreateAlignedStore(MSV.getShadow(A), Base, kShadowTLSAlignment);
3305 Constant *OverflowSize =
3306 ConstantInt::get(IRB.getInt64Ty(), OverflowOffset - AArch64VAEndOffset);
3307 IRB.CreateStore(OverflowSize, MS.VAArgOverflowSizeTLS);
3310 /// Compute the shadow address for a given va_arg.
3311 Value *getShadowPtrForVAArgument(Type *Ty, IRBuilder<> &IRB,
3313 Value *Base = IRB.CreatePointerCast(MS.VAArgTLS, MS.IntptrTy);
3314 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
3315 return IRB.CreateIntToPtr(Base, PointerType::get(MSV.getShadowTy(Ty), 0),
3319 void visitVAStartInst(VAStartInst &I) override {
3320 IRBuilder<> IRB(&I);
3321 VAStartInstrumentationList.push_back(&I);
3322 Value *VAListTag = I.getArgOperand(0);
3323 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
3324 // Unpoison the whole __va_list_tag.
3325 // FIXME: magic ABI constants (size of va_list).
3326 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
3327 /* size */32, /* alignment */8, false);
3330 void visitVACopyInst(VACopyInst &I) override {
3331 IRBuilder<> IRB(&I);
3332 Value *VAListTag = I.getArgOperand(0);
3333 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
3334 // Unpoison the whole __va_list_tag.
3335 // FIXME: magic ABI constants (size of va_list).
3336 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
3337 /* size */32, /* alignment */8, false);
3340 // Retrieve a va_list field of 'void*' size.
3341 Value* getVAField64(IRBuilder<> &IRB, Value *VAListTag, int offset) {
3342 Value *SaveAreaPtrPtr =
3344 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
3345 ConstantInt::get(MS.IntptrTy, offset)),
3346 Type::getInt64PtrTy(*MS.C));
3347 return IRB.CreateLoad(SaveAreaPtrPtr);
3350 // Retrieve a va_list field of 'int' size.
3351 Value* getVAField32(IRBuilder<> &IRB, Value *VAListTag, int offset) {
3352 Value *SaveAreaPtr =
3354 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
3355 ConstantInt::get(MS.IntptrTy, offset)),
3356 Type::getInt32PtrTy(*MS.C));
3357 Value *SaveArea32 = IRB.CreateLoad(SaveAreaPtr);
3358 return IRB.CreateSExt(SaveArea32, MS.IntptrTy);
3361 void finalizeInstrumentation() override {
3362 assert(!VAArgOverflowSize && !VAArgTLSCopy &&
3363 "finalizeInstrumentation called twice");
3364 if (!VAStartInstrumentationList.empty()) {
3365 // If there is a va_start in this function, make a backup copy of
3366 // va_arg_tls somewhere in the function entry block.
3367 IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
3368 VAArgOverflowSize = IRB.CreateLoad(MS.VAArgOverflowSizeTLS);
3370 IRB.CreateAdd(ConstantInt::get(MS.IntptrTy, AArch64VAEndOffset),
3372 VAArgTLSCopy = IRB.CreateAlloca(Type::getInt8Ty(*MS.C), CopySize);
3373 IRB.CreateMemCpy(VAArgTLSCopy, MS.VAArgTLS, CopySize, 8);
3376 Value *GrArgSize = ConstantInt::get(MS.IntptrTy, kAArch64GrArgSize);
3377 Value *VrArgSize = ConstantInt::get(MS.IntptrTy, kAArch64VrArgSize);
3379 // Instrument va_start, copy va_list shadow from the backup copy of
3380 // the TLS contents.
3381 for (size_t i = 0, n = VAStartInstrumentationList.size(); i < n; i++) {
3382 CallInst *OrigInst = VAStartInstrumentationList[i];
3383 IRBuilder<> IRB(OrigInst->getNextNode());
3385 Value *VAListTag = OrigInst->getArgOperand(0);
3387 // The variadic ABI for AArch64 creates two areas to save the incoming
3388 // argument registers (one for 64-bit general register xn-x7 and another
3389 // for 128-bit FP/SIMD vn-v7).
3390 // We need then to propagate the shadow arguments on both regions
3391 // 'va::__gr_top + va::__gr_offs' and 'va::__vr_top + va::__vr_offs'.
3392 // The remaning arguments are saved on shadow for 'va::stack'.
3393 // One caveat is it requires only to propagate the non-named arguments,
3394 // however on the call site instrumentation 'all' the arguments are
3395 // saved. So to copy the shadow values from the va_arg TLS array
3396 // we need to adjust the offset for both GR and VR fields based on
3397 // the __{gr,vr}_offs value (since they are stores based on incoming
3398 // named arguments).
3400 // Read the stack pointer from the va_list.
3401 Value *StackSaveAreaPtr = getVAField64(IRB, VAListTag, 0);
3403 // Read both the __gr_top and __gr_off and add them up.
3404 Value *GrTopSaveAreaPtr = getVAField64(IRB, VAListTag, 8);
3405 Value *GrOffSaveArea = getVAField32(IRB, VAListTag, 24);
3407 Value *GrRegSaveAreaPtr = IRB.CreateAdd(GrTopSaveAreaPtr, GrOffSaveArea);
3409 // Read both the __vr_top and __vr_off and add them up.
3410 Value *VrTopSaveAreaPtr = getVAField64(IRB, VAListTag, 16);
3411 Value *VrOffSaveArea = getVAField32(IRB, VAListTag, 28);
3413 Value *VrRegSaveAreaPtr = IRB.CreateAdd(VrTopSaveAreaPtr, VrOffSaveArea);
3415 // It does not know how many named arguments is being used and, on the
3416 // callsite all the arguments were saved. Since __gr_off is defined as
3417 // '0 - ((8 - named_gr) * 8)', the idea is to just propagate the variadic
3418 // argument by ignoring the bytes of shadow from named arguments.
3419 Value *GrRegSaveAreaShadowPtrOff =
3420 IRB.CreateAdd(GrArgSize, GrOffSaveArea);
3422 Value *GrRegSaveAreaShadowPtr =
3423 MSV.getShadowPtr(GrRegSaveAreaPtr, IRB.getInt8Ty(), IRB);
3425 Value *GrSrcPtr = IRB.CreateInBoundsGEP(IRB.getInt8Ty(), VAArgTLSCopy,
3426 GrRegSaveAreaShadowPtrOff);
3427 Value *GrCopySize = IRB.CreateSub(GrArgSize, GrRegSaveAreaShadowPtrOff);
3429 IRB.CreateMemCpy(GrRegSaveAreaShadowPtr, GrSrcPtr, GrCopySize, 8);
3431 // Again, but for FP/SIMD values.
3432 Value *VrRegSaveAreaShadowPtrOff =
3433 IRB.CreateAdd(VrArgSize, VrOffSaveArea);
3435 Value *VrRegSaveAreaShadowPtr =
3436 MSV.getShadowPtr(VrRegSaveAreaPtr, IRB.getInt8Ty(), IRB);
3438 Value *VrSrcPtr = IRB.CreateInBoundsGEP(
3440 IRB.CreateInBoundsGEP(IRB.getInt8Ty(), VAArgTLSCopy,
3441 IRB.getInt32(AArch64VrBegOffset)),
3442 VrRegSaveAreaShadowPtrOff);
3443 Value *VrCopySize = IRB.CreateSub(VrArgSize, VrRegSaveAreaShadowPtrOff);
3445 IRB.CreateMemCpy(VrRegSaveAreaShadowPtr, VrSrcPtr, VrCopySize, 8);
3447 // And finally for remaining arguments.
3448 Value *StackSaveAreaShadowPtr =
3449 MSV.getShadowPtr(StackSaveAreaPtr, IRB.getInt8Ty(), IRB);
3451 Value *StackSrcPtr =
3452 IRB.CreateInBoundsGEP(IRB.getInt8Ty(), VAArgTLSCopy,
3453 IRB.getInt32(AArch64VAEndOffset));
3455 IRB.CreateMemCpy(StackSaveAreaShadowPtr, StackSrcPtr,
3456 VAArgOverflowSize, 16);
3461 /// \brief PowerPC64-specific implementation of VarArgHelper.
3462 struct VarArgPowerPC64Helper : public VarArgHelper {
3464 MemorySanitizer &MS;
3465 MemorySanitizerVisitor &MSV;
3466 Value *VAArgTLSCopy;
3469 SmallVector<CallInst*, 16> VAStartInstrumentationList;
3471 VarArgPowerPC64Helper(Function &F, MemorySanitizer &MS,
3472 MemorySanitizerVisitor &MSV)
3473 : F(F), MS(MS), MSV(MSV), VAArgTLSCopy(nullptr),
3474 VAArgSize(nullptr) {}
3476 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) override {
3477 // For PowerPC, we need to deal with alignment of stack arguments -
3478 // they are mostly aligned to 8 bytes, but vectors and i128 arrays
3479 // are aligned to 16 bytes, byvals can be aligned to 8 or 16 bytes,
3480 // and QPX vectors are aligned to 32 bytes. For that reason, we
3481 // compute current offset from stack pointer (which is always properly
3482 // aligned), and offset for the first vararg, then subtract them.
3484 llvm::Triple TargetTriple(F.getParent()->getTargetTriple());
3485 // Parameter save area starts at 48 bytes from frame pointer for ABIv1,
3486 // and 32 bytes for ABIv2. This is usually determined by target
3487 // endianness, but in theory could be overriden by function attribute.
3488 // For simplicity, we ignore it here (it'd only matter for QPX vectors).
3489 if (TargetTriple.getArch() == llvm::Triple::ppc64)
3493 unsigned VAArgOffset = VAArgBase;
3494 const DataLayout &DL = F.getParent()->getDataLayout();
3495 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
3496 ArgIt != End; ++ArgIt) {
3498 unsigned ArgNo = CS.getArgumentNo(ArgIt);
3499 bool IsFixed = ArgNo < CS.getFunctionType()->getNumParams();
3500 bool IsByVal = CS.paramHasAttr(ArgNo, Attribute::ByVal);
3502 assert(A->getType()->isPointerTy());
3503 Type *RealTy = A->getType()->getPointerElementType();
3504 uint64_t ArgSize = DL.getTypeAllocSize(RealTy);
3505 uint64_t ArgAlign = CS.getParamAlignment(ArgNo + 1);
3508 VAArgOffset = alignTo(VAArgOffset, ArgAlign);
3510 Value *Base = getShadowPtrForVAArgument(RealTy, IRB,
3511 VAArgOffset - VAArgBase);
3512 IRB.CreateMemCpy(Base, MSV.getShadowPtr(A, IRB.getInt8Ty(), IRB),
3513 ArgSize, kShadowTLSAlignment);
3515 VAArgOffset += alignTo(ArgSize, 8);
3518 uint64_t ArgSize = DL.getTypeAllocSize(A->getType());
3519 uint64_t ArgAlign = 8;
3520 if (A->getType()->isArrayTy()) {
3521 // Arrays are aligned to element size, except for long double
3522 // arrays, which are aligned to 8 bytes.
3523 Type *ElementTy = A->getType()->getArrayElementType();
3524 if (!ElementTy->isPPC_FP128Ty())
3525 ArgAlign = DL.getTypeAllocSize(ElementTy);
3526 } else if (A->getType()->isVectorTy()) {
3527 // Vectors are naturally aligned.
3528 ArgAlign = DL.getTypeAllocSize(A->getType());
3532 VAArgOffset = alignTo(VAArgOffset, ArgAlign);
3533 if (DL.isBigEndian()) {
3534 // Adjusting the shadow for argument with size < 8 to match the placement
3535 // of bits in big endian system
3537 VAArgOffset += (8 - ArgSize);
3540 Base = getShadowPtrForVAArgument(A->getType(), IRB,
3541 VAArgOffset - VAArgBase);
3542 IRB.CreateAlignedStore(MSV.getShadow(A), Base, kShadowTLSAlignment);
3544 VAArgOffset += ArgSize;
3545 VAArgOffset = alignTo(VAArgOffset, 8);
3548 VAArgBase = VAArgOffset;
3551 Constant *TotalVAArgSize = ConstantInt::get(IRB.getInt64Ty(),
3552 VAArgOffset - VAArgBase);
3553 // Here using VAArgOverflowSizeTLS as VAArgSizeTLS to avoid creation of
3554 // a new class member i.e. it is the total size of all VarArgs.
3555 IRB.CreateStore(TotalVAArgSize, MS.VAArgOverflowSizeTLS);
3558 /// \brief Compute the shadow address for a given va_arg.
3559 Value *getShadowPtrForVAArgument(Type *Ty, IRBuilder<> &IRB,
3561 Value *Base = IRB.CreatePointerCast(MS.VAArgTLS, MS.IntptrTy);
3562 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
3563 return IRB.CreateIntToPtr(Base, PointerType::get(MSV.getShadowTy(Ty), 0),
3567 void visitVAStartInst(VAStartInst &I) override {
3568 IRBuilder<> IRB(&I);
3569 VAStartInstrumentationList.push_back(&I);
3570 Value *VAListTag = I.getArgOperand(0);
3571 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
3572 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
3573 /* size */8, /* alignment */8, false);
3576 void visitVACopyInst(VACopyInst &I) override {
3577 IRBuilder<> IRB(&I);
3578 Value *VAListTag = I.getArgOperand(0);
3579 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
3580 // Unpoison the whole __va_list_tag.
3581 // FIXME: magic ABI constants.
3582 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
3583 /* size */8, /* alignment */8, false);
3586 void finalizeInstrumentation() override {
3587 assert(!VAArgSize && !VAArgTLSCopy &&
3588 "finalizeInstrumentation called twice");
3589 IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
3590 VAArgSize = IRB.CreateLoad(MS.VAArgOverflowSizeTLS);
3591 Value *CopySize = IRB.CreateAdd(ConstantInt::get(MS.IntptrTy, 0),
3594 if (!VAStartInstrumentationList.empty()) {
3595 // If there is a va_start in this function, make a backup copy of
3596 // va_arg_tls somewhere in the function entry block.
3597 VAArgTLSCopy = IRB.CreateAlloca(Type::getInt8Ty(*MS.C), CopySize);
3598 IRB.CreateMemCpy(VAArgTLSCopy, MS.VAArgTLS, CopySize, 8);
3601 // Instrument va_start.
3602 // Copy va_list shadow from the backup copy of the TLS contents.
3603 for (size_t i = 0, n = VAStartInstrumentationList.size(); i < n; i++) {
3604 CallInst *OrigInst = VAStartInstrumentationList[i];
3605 IRBuilder<> IRB(OrigInst->getNextNode());
3606 Value *VAListTag = OrigInst->getArgOperand(0);
3607 Value *RegSaveAreaPtrPtr =
3608 IRB.CreateIntToPtr(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
3609 Type::getInt64PtrTy(*MS.C));
3610 Value *RegSaveAreaPtr = IRB.CreateLoad(RegSaveAreaPtrPtr);
3611 Value *RegSaveAreaShadowPtr =
3612 MSV.getShadowPtr(RegSaveAreaPtr, IRB.getInt8Ty(), IRB);
3613 IRB.CreateMemCpy(RegSaveAreaShadowPtr, VAArgTLSCopy, CopySize, 8);
3618 /// \brief A no-op implementation of VarArgHelper.
3619 struct VarArgNoOpHelper : public VarArgHelper {
3620 VarArgNoOpHelper(Function &F, MemorySanitizer &MS,
3621 MemorySanitizerVisitor &MSV) {}
3623 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) override {}
3625 void visitVAStartInst(VAStartInst &I) override {}
3627 void visitVACopyInst(VACopyInst &I) override {}
3629 void finalizeInstrumentation() override {}
3632 VarArgHelper *CreateVarArgHelper(Function &Func, MemorySanitizer &Msan,
3633 MemorySanitizerVisitor &Visitor) {
3634 // VarArg handling is only implemented on AMD64. False positives are possible
3635 // on other platforms.
3636 llvm::Triple TargetTriple(Func.getParent()->getTargetTriple());
3637 if (TargetTriple.getArch() == llvm::Triple::x86_64)
3638 return new VarArgAMD64Helper(Func, Msan, Visitor);
3639 else if (TargetTriple.getArch() == llvm::Triple::mips64 ||
3640 TargetTriple.getArch() == llvm::Triple::mips64el)
3641 return new VarArgMIPS64Helper(Func, Msan, Visitor);
3642 else if (TargetTriple.getArch() == llvm::Triple::aarch64)
3643 return new VarArgAArch64Helper(Func, Msan, Visitor);
3644 else if (TargetTriple.getArch() == llvm::Triple::ppc64 ||
3645 TargetTriple.getArch() == llvm::Triple::ppc64le)
3646 return new VarArgPowerPC64Helper(Func, Msan, Visitor);
3648 return new VarArgNoOpHelper(Func, Msan, Visitor);
3651 } // anonymous namespace
3653 bool MemorySanitizer::runOnFunction(Function &F) {
3654 if (&F == MsanCtorFunction)
3656 MemorySanitizerVisitor Visitor(F, *this);
3658 // Clear out readonly/readnone attributes.
3660 B.addAttribute(Attribute::ReadOnly)
3661 .addAttribute(Attribute::ReadNone);
3663 AttributeList::FunctionIndex,
3664 AttributeList::get(F.getContext(), AttributeList::FunctionIndex, B));
3666 return Visitor.runOnFunction();