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/Transforms/Instrumentation.h"
95 #include "llvm/ADT/DepthFirstIterator.h"
96 #include "llvm/ADT/SmallString.h"
97 #include "llvm/ADT/SmallVector.h"
98 #include "llvm/ADT/StringExtras.h"
99 #include "llvm/ADT/Triple.h"
100 #include "llvm/IR/DataLayout.h"
101 #include "llvm/IR/Function.h"
102 #include "llvm/IR/IRBuilder.h"
103 #include "llvm/IR/InlineAsm.h"
104 #include "llvm/IR/InstVisitor.h"
105 #include "llvm/IR/IntrinsicInst.h"
106 #include "llvm/IR/LLVMContext.h"
107 #include "llvm/IR/MDBuilder.h"
108 #include "llvm/IR/Module.h"
109 #include "llvm/IR/Type.h"
110 #include "llvm/IR/ValueMap.h"
111 #include "llvm/Support/CommandLine.h"
112 #include "llvm/Support/Compiler.h"
113 #include "llvm/Support/Debug.h"
114 #include "llvm/Support/raw_ostream.h"
115 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
116 #include "llvm/Transforms/Utils/Local.h"
117 #include "llvm/Transforms/Utils/ModuleUtils.h"
119 using namespace llvm;
121 #define DEBUG_TYPE "msan"
123 // VMA size definition for architecture that support multiple sizes.
124 // AArch64 has 3 VMA sizes: 39, 42 and 48.
125 #ifndef SANITIZER_AARCH64_VMA
126 # define SANITIZER_AARCH64_VMA 39
128 # if SANITIZER_AARCH64_VMA != 39 && SANITIZER_AARCH64_VMA != 42
129 # error "invalid SANITIZER_AARCH64_VMA size"
133 static const unsigned kOriginSize = 4;
134 static const unsigned kMinOriginAlignment = 4;
135 static const unsigned kShadowTLSAlignment = 8;
137 // These constants must be kept in sync with the ones in msan.h.
138 static const unsigned kParamTLSSize = 800;
139 static const unsigned kRetvalTLSSize = 800;
141 // Accesses sizes are powers of two: 1, 2, 4, 8.
142 static const size_t kNumberOfAccessSizes = 4;
144 /// \brief Track origins of uninitialized values.
146 /// Adds a section to MemorySanitizer report that points to the allocation
147 /// (stack or heap) the uninitialized bits came from originally.
148 static cl::opt<int> ClTrackOrigins("msan-track-origins",
149 cl::desc("Track origins (allocation sites) of poisoned memory"),
150 cl::Hidden, cl::init(0));
151 static cl::opt<bool> ClKeepGoing("msan-keep-going",
152 cl::desc("keep going after reporting a UMR"),
153 cl::Hidden, cl::init(false));
154 static cl::opt<bool> ClPoisonStack("msan-poison-stack",
155 cl::desc("poison uninitialized stack variables"),
156 cl::Hidden, cl::init(true));
157 static cl::opt<bool> ClPoisonStackWithCall("msan-poison-stack-with-call",
158 cl::desc("poison uninitialized stack variables with a call"),
159 cl::Hidden, cl::init(false));
160 static cl::opt<int> ClPoisonStackPattern("msan-poison-stack-pattern",
161 cl::desc("poison uninitialized stack variables with the given patter"),
162 cl::Hidden, cl::init(0xff));
163 static cl::opt<bool> ClPoisonUndef("msan-poison-undef",
164 cl::desc("poison undef temps"),
165 cl::Hidden, cl::init(true));
167 static cl::opt<bool> ClHandleICmp("msan-handle-icmp",
168 cl::desc("propagate shadow through ICmpEQ and ICmpNE"),
169 cl::Hidden, cl::init(true));
171 static cl::opt<bool> ClHandleICmpExact("msan-handle-icmp-exact",
172 cl::desc("exact handling of relational integer ICmp"),
173 cl::Hidden, cl::init(false));
175 // This flag controls whether we check the shadow of the address
176 // operand of load or store. Such bugs are very rare, since load from
177 // a garbage address typically results in SEGV, but still happen
178 // (e.g. only lower bits of address are garbage, or the access happens
179 // early at program startup where malloc-ed memory is more likely to
180 // be zeroed. As of 2012-08-28 this flag adds 20% slowdown.
181 static cl::opt<bool> ClCheckAccessAddress("msan-check-access-address",
182 cl::desc("report accesses through a pointer which has poisoned shadow"),
183 cl::Hidden, cl::init(true));
185 static cl::opt<bool> ClDumpStrictInstructions("msan-dump-strict-instructions",
186 cl::desc("print out instructions with default strict semantics"),
187 cl::Hidden, cl::init(false));
189 static cl::opt<int> ClInstrumentationWithCallThreshold(
190 "msan-instrumentation-with-call-threshold",
192 "If the function being instrumented requires more than "
193 "this number of checks and origin stores, use callbacks instead of "
194 "inline checks (-1 means never use callbacks)."),
195 cl::Hidden, cl::init(3500));
197 // This is an experiment to enable handling of cases where shadow is a non-zero
198 // compile-time constant. For some unexplainable reason they were silently
199 // ignored in the instrumentation.
200 static cl::opt<bool> ClCheckConstantShadow("msan-check-constant-shadow",
201 cl::desc("Insert checks for constant shadow values"),
202 cl::Hidden, cl::init(false));
204 static const char *const kMsanModuleCtorName = "msan.module_ctor";
205 static const char *const kMsanInitName = "__msan_init";
209 // Memory map parameters used in application-to-shadow address calculation.
210 // Offset = (Addr & ~AndMask) ^ XorMask
211 // Shadow = ShadowBase + Offset
212 // Origin = OriginBase + Offset
213 struct MemoryMapParams {
220 struct PlatformMemoryMapParams {
221 const MemoryMapParams *bits32;
222 const MemoryMapParams *bits64;
226 static const MemoryMapParams Linux_I386_MemoryMapParams = {
227 0x000080000000, // AndMask
228 0, // XorMask (not used)
229 0, // ShadowBase (not used)
230 0x000040000000, // OriginBase
234 static const MemoryMapParams Linux_X86_64_MemoryMapParams = {
235 0x400000000000, // AndMask
236 0, // XorMask (not used)
237 0, // ShadowBase (not used)
238 0x200000000000, // OriginBase
242 static const MemoryMapParams Linux_MIPS64_MemoryMapParams = {
243 0x004000000000, // AndMask
244 0, // XorMask (not used)
245 0, // ShadowBase (not used)
246 0x002000000000, // OriginBase
250 static const MemoryMapParams Linux_PowerPC64_MemoryMapParams = {
251 0x200000000000, // AndMask
252 0x100000000000, // XorMask
253 0x080000000000, // ShadowBase
254 0x1C0000000000, // OriginBase
258 static const MemoryMapParams Linux_AArch64_MemoryMapParams = {
259 #if SANITIZER_AARCH64_VMA == 39
260 0x007C00000000, // AndMask
261 0x000100000000, // XorMask
262 0x004000000000, // ShadowBase
263 0x004300000000, // OriginBase
264 #elif SANITIZER_AARCH64_VMA == 42
265 0x03E000000000, // AndMask
266 0x001000000000, // XorMask
267 0x010000000000, // ShadowBase
268 0x012000000000, // OriginBase
273 static const MemoryMapParams FreeBSD_I386_MemoryMapParams = {
274 0x000180000000, // AndMask
275 0x000040000000, // XorMask
276 0x000020000000, // ShadowBase
277 0x000700000000, // OriginBase
281 static const MemoryMapParams FreeBSD_X86_64_MemoryMapParams = {
282 0xc00000000000, // AndMask
283 0x200000000000, // XorMask
284 0x100000000000, // ShadowBase
285 0x380000000000, // OriginBase
288 static const PlatformMemoryMapParams Linux_X86_MemoryMapParams = {
289 &Linux_I386_MemoryMapParams,
290 &Linux_X86_64_MemoryMapParams,
293 static const PlatformMemoryMapParams Linux_MIPS_MemoryMapParams = {
295 &Linux_MIPS64_MemoryMapParams,
298 static const PlatformMemoryMapParams Linux_PowerPC_MemoryMapParams = {
300 &Linux_PowerPC64_MemoryMapParams,
303 static const PlatformMemoryMapParams Linux_ARM_MemoryMapParams = {
305 &Linux_AArch64_MemoryMapParams,
308 static const PlatformMemoryMapParams FreeBSD_X86_MemoryMapParams = {
309 &FreeBSD_I386_MemoryMapParams,
310 &FreeBSD_X86_64_MemoryMapParams,
313 /// \brief An instrumentation pass implementing detection of uninitialized
316 /// MemorySanitizer: instrument the code in module to find
317 /// uninitialized reads.
318 class MemorySanitizer : public FunctionPass {
320 MemorySanitizer(int TrackOrigins = 0)
322 TrackOrigins(std::max(TrackOrigins, (int)ClTrackOrigins)),
323 WarningFn(nullptr) {}
324 const char *getPassName() const override { return "MemorySanitizer"; }
325 bool runOnFunction(Function &F) override;
326 bool doInitialization(Module &M) override;
327 static char ID; // Pass identification, replacement for typeid.
330 void initializeCallbacks(Module &M);
332 /// \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;
389 char MemorySanitizer::ID = 0;
390 INITIALIZE_PASS(MemorySanitizer, "msan",
391 "MemorySanitizer: detects uninitialized reads.",
394 FunctionPass *llvm::createMemorySanitizerPass(int TrackOrigins) {
395 return new MemorySanitizer(TrackOrigins);
398 /// \brief Create a non-const global initialized with the given string.
400 /// Creates a writable global for Str so that we can pass it to the
401 /// run-time lib. Runtime uses first 4 bytes of the string to store the
402 /// frame ID, so the string needs to be mutable.
403 static GlobalVariable *createPrivateNonConstGlobalForString(Module &M,
405 Constant *StrConst = ConstantDataArray::getString(M.getContext(), Str);
406 return new GlobalVariable(M, StrConst->getType(), /*isConstant=*/false,
407 GlobalValue::PrivateLinkage, StrConst, "");
411 /// \brief Insert extern declaration of runtime-provided functions and globals.
412 void MemorySanitizer::initializeCallbacks(Module &M) {
413 // Only do this once.
418 // Create the callback.
419 // FIXME: this function should have "Cold" calling conv,
420 // which is not yet implemented.
421 StringRef WarningFnName = ClKeepGoing ? "__msan_warning"
422 : "__msan_warning_noreturn";
423 WarningFn = M.getOrInsertFunction(WarningFnName, IRB.getVoidTy(), nullptr);
425 for (size_t AccessSizeIndex = 0; AccessSizeIndex < kNumberOfAccessSizes;
427 unsigned AccessSize = 1 << AccessSizeIndex;
428 std::string FunctionName = "__msan_maybe_warning_" + itostr(AccessSize);
429 MaybeWarningFn[AccessSizeIndex] = M.getOrInsertFunction(
430 FunctionName, IRB.getVoidTy(), IRB.getIntNTy(AccessSize * 8),
431 IRB.getInt32Ty(), nullptr);
433 FunctionName = "__msan_maybe_store_origin_" + itostr(AccessSize);
434 MaybeStoreOriginFn[AccessSizeIndex] = M.getOrInsertFunction(
435 FunctionName, IRB.getVoidTy(), IRB.getIntNTy(AccessSize * 8),
436 IRB.getInt8PtrTy(), IRB.getInt32Ty(), nullptr);
439 MsanSetAllocaOrigin4Fn = M.getOrInsertFunction(
440 "__msan_set_alloca_origin4", IRB.getVoidTy(), IRB.getInt8PtrTy(), IntptrTy,
441 IRB.getInt8PtrTy(), IntptrTy, nullptr);
443 M.getOrInsertFunction("__msan_poison_stack", IRB.getVoidTy(),
444 IRB.getInt8PtrTy(), IntptrTy, nullptr);
445 MsanChainOriginFn = M.getOrInsertFunction(
446 "__msan_chain_origin", IRB.getInt32Ty(), IRB.getInt32Ty(), nullptr);
447 MemmoveFn = M.getOrInsertFunction(
448 "__msan_memmove", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
449 IRB.getInt8PtrTy(), IntptrTy, nullptr);
450 MemcpyFn = M.getOrInsertFunction(
451 "__msan_memcpy", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
453 MemsetFn = M.getOrInsertFunction(
454 "__msan_memset", IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), IRB.getInt32Ty(),
458 RetvalTLS = new GlobalVariable(
459 M, ArrayType::get(IRB.getInt64Ty(), kRetvalTLSSize / 8), false,
460 GlobalVariable::ExternalLinkage, nullptr, "__msan_retval_tls", nullptr,
461 GlobalVariable::InitialExecTLSModel);
462 RetvalOriginTLS = new GlobalVariable(
463 M, OriginTy, false, GlobalVariable::ExternalLinkage, nullptr,
464 "__msan_retval_origin_tls", nullptr, GlobalVariable::InitialExecTLSModel);
466 ParamTLS = new GlobalVariable(
467 M, ArrayType::get(IRB.getInt64Ty(), kParamTLSSize / 8), false,
468 GlobalVariable::ExternalLinkage, nullptr, "__msan_param_tls", nullptr,
469 GlobalVariable::InitialExecTLSModel);
470 ParamOriginTLS = new GlobalVariable(
471 M, ArrayType::get(OriginTy, kParamTLSSize / 4), false,
472 GlobalVariable::ExternalLinkage, nullptr, "__msan_param_origin_tls",
473 nullptr, GlobalVariable::InitialExecTLSModel);
475 VAArgTLS = new GlobalVariable(
476 M, ArrayType::get(IRB.getInt64Ty(), kParamTLSSize / 8), false,
477 GlobalVariable::ExternalLinkage, nullptr, "__msan_va_arg_tls", nullptr,
478 GlobalVariable::InitialExecTLSModel);
479 VAArgOverflowSizeTLS = new GlobalVariable(
480 M, IRB.getInt64Ty(), false, GlobalVariable::ExternalLinkage, nullptr,
481 "__msan_va_arg_overflow_size_tls", nullptr,
482 GlobalVariable::InitialExecTLSModel);
483 OriginTLS = new GlobalVariable(
484 M, IRB.getInt32Ty(), false, GlobalVariable::ExternalLinkage, nullptr,
485 "__msan_origin_tls", nullptr, GlobalVariable::InitialExecTLSModel);
487 // We insert an empty inline asm after __msan_report* to avoid callback merge.
488 EmptyAsm = InlineAsm::get(FunctionType::get(IRB.getVoidTy(), false),
489 StringRef(""), StringRef(""),
490 /*hasSideEffects=*/true);
493 /// \brief Module-level initialization.
495 /// inserts a call to __msan_init to the module's constructor list.
496 bool MemorySanitizer::doInitialization(Module &M) {
497 auto &DL = M.getDataLayout();
499 Triple TargetTriple(M.getTargetTriple());
500 switch (TargetTriple.getOS()) {
501 case Triple::FreeBSD:
502 switch (TargetTriple.getArch()) {
504 MapParams = FreeBSD_X86_MemoryMapParams.bits64;
507 MapParams = FreeBSD_X86_MemoryMapParams.bits32;
510 report_fatal_error("unsupported architecture");
514 switch (TargetTriple.getArch()) {
516 MapParams = Linux_X86_MemoryMapParams.bits64;
519 MapParams = Linux_X86_MemoryMapParams.bits32;
522 case Triple::mips64el:
523 MapParams = Linux_MIPS_MemoryMapParams.bits64;
526 case Triple::ppc64le:
527 MapParams = Linux_PowerPC_MemoryMapParams.bits64;
529 case Triple::aarch64:
530 case Triple::aarch64_be:
531 MapParams = Linux_ARM_MemoryMapParams.bits64;
534 report_fatal_error("unsupported architecture");
538 report_fatal_error("unsupported operating system");
541 C = &(M.getContext());
543 IntptrTy = IRB.getIntPtrTy(DL);
544 OriginTy = IRB.getInt32Ty();
546 ColdCallWeights = MDBuilder(*C).createBranchWeights(1, 1000);
547 OriginStoreWeights = MDBuilder(*C).createBranchWeights(1, 1000);
549 std::tie(MsanCtorFunction, std::ignore) =
550 createSanitizerCtorAndInitFunctions(M, kMsanModuleCtorName, kMsanInitName,
554 appendToGlobalCtors(M, MsanCtorFunction, 0);
557 new GlobalVariable(M, IRB.getInt32Ty(), true, GlobalValue::WeakODRLinkage,
558 IRB.getInt32(TrackOrigins), "__msan_track_origins");
561 new GlobalVariable(M, IRB.getInt32Ty(), true, GlobalValue::WeakODRLinkage,
562 IRB.getInt32(ClKeepGoing), "__msan_keep_going");
569 /// \brief A helper class that handles instrumentation of VarArg
570 /// functions on a particular platform.
572 /// Implementations are expected to insert the instrumentation
573 /// necessary to propagate argument shadow through VarArg function
574 /// calls. Visit* methods are called during an InstVisitor pass over
575 /// the function, and should avoid creating new basic blocks. A new
576 /// instance of this class is created for each instrumented function.
577 struct VarArgHelper {
578 /// \brief Visit a CallSite.
579 virtual void visitCallSite(CallSite &CS, IRBuilder<> &IRB) = 0;
581 /// \brief Visit a va_start call.
582 virtual void visitVAStartInst(VAStartInst &I) = 0;
584 /// \brief Visit a va_copy call.
585 virtual void visitVACopyInst(VACopyInst &I) = 0;
587 /// \brief Finalize function instrumentation.
589 /// This method is called after visiting all interesting (see above)
590 /// instructions in a function.
591 virtual void finalizeInstrumentation() = 0;
593 virtual ~VarArgHelper() {}
596 struct MemorySanitizerVisitor;
599 CreateVarArgHelper(Function &Func, MemorySanitizer &Msan,
600 MemorySanitizerVisitor &Visitor);
602 unsigned TypeSizeToSizeIndex(unsigned TypeSize) {
603 if (TypeSize <= 8) return 0;
604 return Log2_32_Ceil(TypeSize / 8);
607 /// This class does all the work for a given function. Store and Load
608 /// instructions store and load corresponding shadow and origin
609 /// values. Most instructions propagate shadow from arguments to their
610 /// return values. Certain instructions (most importantly, BranchInst)
611 /// test their argument shadow and print reports (with a runtime call) if it's
613 struct MemorySanitizerVisitor : public InstVisitor<MemorySanitizerVisitor> {
616 SmallVector<PHINode *, 16> ShadowPHINodes, OriginPHINodes;
617 ValueMap<Value*, Value*> ShadowMap, OriginMap;
618 std::unique_ptr<VarArgHelper> VAHelper;
620 // The following flags disable parts of MSan instrumentation based on
621 // blacklist contents and command-line options.
623 bool PropagateShadow;
626 bool CheckReturnValue;
628 struct ShadowOriginAndInsertPoint {
631 Instruction *OrigIns;
632 ShadowOriginAndInsertPoint(Value *S, Value *O, Instruction *I)
633 : Shadow(S), Origin(O), OrigIns(I) { }
635 SmallVector<ShadowOriginAndInsertPoint, 16> InstrumentationList;
636 SmallVector<Instruction*, 16> StoreList;
638 MemorySanitizerVisitor(Function &F, MemorySanitizer &MS)
639 : F(F), MS(MS), VAHelper(CreateVarArgHelper(F, MS, *this)) {
640 bool SanitizeFunction = F.hasFnAttribute(Attribute::SanitizeMemory);
641 InsertChecks = SanitizeFunction;
642 PropagateShadow = SanitizeFunction;
643 PoisonStack = SanitizeFunction && ClPoisonStack;
644 PoisonUndef = SanitizeFunction && ClPoisonUndef;
645 // FIXME: Consider using SpecialCaseList to specify a list of functions that
646 // must always return fully initialized values. For now, we hardcode "main".
647 CheckReturnValue = SanitizeFunction && (F.getName() == "main");
649 DEBUG(if (!InsertChecks)
650 dbgs() << "MemorySanitizer is not inserting checks into '"
651 << F.getName() << "'\n");
654 Value *updateOrigin(Value *V, IRBuilder<> &IRB) {
655 if (MS.TrackOrigins <= 1) return V;
656 return IRB.CreateCall(MS.MsanChainOriginFn, V);
659 Value *originToIntptr(IRBuilder<> &IRB, Value *Origin) {
660 const DataLayout &DL = F.getParent()->getDataLayout();
661 unsigned IntptrSize = DL.getTypeStoreSize(MS.IntptrTy);
662 if (IntptrSize == kOriginSize) return Origin;
663 assert(IntptrSize == kOriginSize * 2);
664 Origin = IRB.CreateIntCast(Origin, MS.IntptrTy, /* isSigned */ false);
665 return IRB.CreateOr(Origin, IRB.CreateShl(Origin, kOriginSize * 8));
668 /// \brief Fill memory range with the given origin value.
669 void paintOrigin(IRBuilder<> &IRB, Value *Origin, Value *OriginPtr,
670 unsigned Size, unsigned Alignment) {
671 const DataLayout &DL = F.getParent()->getDataLayout();
672 unsigned IntptrAlignment = DL.getABITypeAlignment(MS.IntptrTy);
673 unsigned IntptrSize = DL.getTypeStoreSize(MS.IntptrTy);
674 assert(IntptrAlignment >= kMinOriginAlignment);
675 assert(IntptrSize >= kOriginSize);
678 unsigned CurrentAlignment = Alignment;
679 if (Alignment >= IntptrAlignment && IntptrSize > kOriginSize) {
680 Value *IntptrOrigin = originToIntptr(IRB, Origin);
681 Value *IntptrOriginPtr =
682 IRB.CreatePointerCast(OriginPtr, PointerType::get(MS.IntptrTy, 0));
683 for (unsigned i = 0; i < Size / IntptrSize; ++i) {
684 Value *Ptr = i ? IRB.CreateConstGEP1_32(MS.IntptrTy, IntptrOriginPtr, i)
686 IRB.CreateAlignedStore(IntptrOrigin, Ptr, CurrentAlignment);
687 Ofs += IntptrSize / kOriginSize;
688 CurrentAlignment = IntptrAlignment;
692 for (unsigned i = Ofs; i < (Size + kOriginSize - 1) / kOriginSize; ++i) {
694 i ? IRB.CreateConstGEP1_32(nullptr, OriginPtr, i) : OriginPtr;
695 IRB.CreateAlignedStore(Origin, GEP, CurrentAlignment);
696 CurrentAlignment = kMinOriginAlignment;
700 void storeOrigin(IRBuilder<> &IRB, Value *Addr, Value *Shadow, Value *Origin,
701 unsigned Alignment, bool AsCall) {
702 const DataLayout &DL = F.getParent()->getDataLayout();
703 unsigned OriginAlignment = std::max(kMinOriginAlignment, Alignment);
704 unsigned StoreSize = DL.getTypeStoreSize(Shadow->getType());
705 if (isa<StructType>(Shadow->getType())) {
706 paintOrigin(IRB, updateOrigin(Origin, IRB),
707 getOriginPtr(Addr, IRB, Alignment), StoreSize,
710 Value *ConvertedShadow = convertToShadowTyNoVec(Shadow, IRB);
711 Constant *ConstantShadow = dyn_cast_or_null<Constant>(ConvertedShadow);
712 if (ConstantShadow) {
713 if (ClCheckConstantShadow && !ConstantShadow->isZeroValue())
714 paintOrigin(IRB, updateOrigin(Origin, IRB),
715 getOriginPtr(Addr, IRB, Alignment), StoreSize,
720 unsigned TypeSizeInBits =
721 DL.getTypeSizeInBits(ConvertedShadow->getType());
722 unsigned SizeIndex = TypeSizeToSizeIndex(TypeSizeInBits);
723 if (AsCall && SizeIndex < kNumberOfAccessSizes) {
724 Value *Fn = MS.MaybeStoreOriginFn[SizeIndex];
725 Value *ConvertedShadow2 = IRB.CreateZExt(
726 ConvertedShadow, IRB.getIntNTy(8 * (1 << SizeIndex)));
727 IRB.CreateCall(Fn, {ConvertedShadow2,
728 IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()),
731 Value *Cmp = IRB.CreateICmpNE(
732 ConvertedShadow, getCleanShadow(ConvertedShadow), "_mscmp");
733 Instruction *CheckTerm = SplitBlockAndInsertIfThen(
734 Cmp, IRB.GetInsertPoint(), false, MS.OriginStoreWeights);
735 IRBuilder<> IRBNew(CheckTerm);
736 paintOrigin(IRBNew, updateOrigin(Origin, IRBNew),
737 getOriginPtr(Addr, IRBNew, Alignment), StoreSize,
743 void materializeStores(bool InstrumentWithCalls) {
744 for (auto Inst : StoreList) {
745 StoreInst &SI = *dyn_cast<StoreInst>(Inst);
747 IRBuilder<> IRB(&SI);
748 Value *Val = SI.getValueOperand();
749 Value *Addr = SI.getPointerOperand();
750 Value *Shadow = SI.isAtomic() ? getCleanShadow(Val) : getShadow(Val);
751 Value *ShadowPtr = getShadowPtr(Addr, Shadow->getType(), IRB);
754 IRB.CreateAlignedStore(Shadow, ShadowPtr, SI.getAlignment());
755 DEBUG(dbgs() << " STORE: " << *NewSI << "\n");
758 if (ClCheckAccessAddress) insertShadowCheck(Addr, &SI);
760 if (SI.isAtomic()) SI.setOrdering(addReleaseOrdering(SI.getOrdering()));
762 if (MS.TrackOrigins && !SI.isAtomic())
763 storeOrigin(IRB, Addr, Shadow, getOrigin(Val), SI.getAlignment(),
764 InstrumentWithCalls);
768 void materializeOneCheck(Instruction *OrigIns, Value *Shadow, Value *Origin,
770 IRBuilder<> IRB(OrigIns);
771 DEBUG(dbgs() << " SHAD0 : " << *Shadow << "\n");
772 Value *ConvertedShadow = convertToShadowTyNoVec(Shadow, IRB);
773 DEBUG(dbgs() << " SHAD1 : " << *ConvertedShadow << "\n");
775 Constant *ConstantShadow = dyn_cast_or_null<Constant>(ConvertedShadow);
776 if (ConstantShadow) {
777 if (ClCheckConstantShadow && !ConstantShadow->isZeroValue()) {
778 if (MS.TrackOrigins) {
779 IRB.CreateStore(Origin ? (Value *)Origin : (Value *)IRB.getInt32(0),
782 IRB.CreateCall(MS.WarningFn, {});
783 IRB.CreateCall(MS.EmptyAsm, {});
784 // FIXME: Insert UnreachableInst if !ClKeepGoing?
785 // This may invalidate some of the following checks and needs to be done
791 const DataLayout &DL = OrigIns->getModule()->getDataLayout();
793 unsigned TypeSizeInBits = DL.getTypeSizeInBits(ConvertedShadow->getType());
794 unsigned SizeIndex = TypeSizeToSizeIndex(TypeSizeInBits);
795 if (AsCall && SizeIndex < kNumberOfAccessSizes) {
796 Value *Fn = MS.MaybeWarningFn[SizeIndex];
797 Value *ConvertedShadow2 =
798 IRB.CreateZExt(ConvertedShadow, IRB.getIntNTy(8 * (1 << SizeIndex)));
799 IRB.CreateCall(Fn, {ConvertedShadow2, MS.TrackOrigins && Origin
801 : (Value *)IRB.getInt32(0)});
803 Value *Cmp = IRB.CreateICmpNE(ConvertedShadow,
804 getCleanShadow(ConvertedShadow), "_mscmp");
805 Instruction *CheckTerm = SplitBlockAndInsertIfThen(
807 /* Unreachable */ !ClKeepGoing, MS.ColdCallWeights);
809 IRB.SetInsertPoint(CheckTerm);
810 if (MS.TrackOrigins) {
811 IRB.CreateStore(Origin ? (Value *)Origin : (Value *)IRB.getInt32(0),
814 IRB.CreateCall(MS.WarningFn, {});
815 IRB.CreateCall(MS.EmptyAsm, {});
816 DEBUG(dbgs() << " CHECK: " << *Cmp << "\n");
820 void materializeChecks(bool InstrumentWithCalls) {
821 for (const auto &ShadowData : InstrumentationList) {
822 Instruction *OrigIns = ShadowData.OrigIns;
823 Value *Shadow = ShadowData.Shadow;
824 Value *Origin = ShadowData.Origin;
825 materializeOneCheck(OrigIns, Shadow, Origin, InstrumentWithCalls);
827 DEBUG(dbgs() << "DONE:\n" << F);
830 /// \brief Add MemorySanitizer instrumentation to a function.
831 bool runOnFunction() {
832 MS.initializeCallbacks(*F.getParent());
834 // In the presence of unreachable blocks, we may see Phi nodes with
835 // incoming nodes from such blocks. Since InstVisitor skips unreachable
836 // blocks, such nodes will not have any shadow value associated with them.
837 // It's easier to remove unreachable blocks than deal with missing shadow.
838 removeUnreachableBlocks(F);
840 // Iterate all BBs in depth-first order and create shadow instructions
841 // for all instructions (where applicable).
842 // For PHI nodes we create dummy shadow PHIs which will be finalized later.
843 for (BasicBlock *BB : depth_first(&F.getEntryBlock()))
847 // Finalize PHI nodes.
848 for (PHINode *PN : ShadowPHINodes) {
849 PHINode *PNS = cast<PHINode>(getShadow(PN));
850 PHINode *PNO = MS.TrackOrigins ? cast<PHINode>(getOrigin(PN)) : nullptr;
851 size_t NumValues = PN->getNumIncomingValues();
852 for (size_t v = 0; v < NumValues; v++) {
853 PNS->addIncoming(getShadow(PN, v), PN->getIncomingBlock(v));
854 if (PNO) PNO->addIncoming(getOrigin(PN, v), PN->getIncomingBlock(v));
858 VAHelper->finalizeInstrumentation();
860 bool InstrumentWithCalls = ClInstrumentationWithCallThreshold >= 0 &&
861 InstrumentationList.size() + StoreList.size() >
862 (unsigned)ClInstrumentationWithCallThreshold;
864 // Delayed instrumentation of StoreInst.
865 // This may add new checks to be inserted later.
866 materializeStores(InstrumentWithCalls);
868 // Insert shadow value checks.
869 materializeChecks(InstrumentWithCalls);
874 /// \brief Compute the shadow type that corresponds to a given Value.
875 Type *getShadowTy(Value *V) {
876 return getShadowTy(V->getType());
879 /// \brief Compute the shadow type that corresponds to a given Type.
880 Type *getShadowTy(Type *OrigTy) {
881 if (!OrigTy->isSized()) {
884 // For integer type, shadow is the same as the original type.
885 // This may return weird-sized types like i1.
886 if (IntegerType *IT = dyn_cast<IntegerType>(OrigTy))
888 const DataLayout &DL = F.getParent()->getDataLayout();
889 if (VectorType *VT = dyn_cast<VectorType>(OrigTy)) {
890 uint32_t EltSize = DL.getTypeSizeInBits(VT->getElementType());
891 return VectorType::get(IntegerType::get(*MS.C, EltSize),
892 VT->getNumElements());
894 if (ArrayType *AT = dyn_cast<ArrayType>(OrigTy)) {
895 return ArrayType::get(getShadowTy(AT->getElementType()),
896 AT->getNumElements());
898 if (StructType *ST = dyn_cast<StructType>(OrigTy)) {
899 SmallVector<Type*, 4> Elements;
900 for (unsigned i = 0, n = ST->getNumElements(); i < n; i++)
901 Elements.push_back(getShadowTy(ST->getElementType(i)));
902 StructType *Res = StructType::get(*MS.C, Elements, ST->isPacked());
903 DEBUG(dbgs() << "getShadowTy: " << *ST << " ===> " << *Res << "\n");
906 uint32_t TypeSize = DL.getTypeSizeInBits(OrigTy);
907 return IntegerType::get(*MS.C, TypeSize);
910 /// \brief Flatten a vector type.
911 Type *getShadowTyNoVec(Type *ty) {
912 if (VectorType *vt = dyn_cast<VectorType>(ty))
913 return IntegerType::get(*MS.C, vt->getBitWidth());
917 /// \brief Convert a shadow value to it's flattened variant.
918 Value *convertToShadowTyNoVec(Value *V, IRBuilder<> &IRB) {
919 Type *Ty = V->getType();
920 Type *NoVecTy = getShadowTyNoVec(Ty);
921 if (Ty == NoVecTy) return V;
922 return IRB.CreateBitCast(V, NoVecTy);
925 /// \brief Compute the integer shadow offset that corresponds to a given
926 /// application address.
928 /// Offset = (Addr & ~AndMask) ^ XorMask
929 Value *getShadowPtrOffset(Value *Addr, IRBuilder<> &IRB) {
930 uint64_t AndMask = MS.MapParams->AndMask;
931 assert(AndMask != 0 && "AndMask shall be specified");
933 IRB.CreateAnd(IRB.CreatePointerCast(Addr, MS.IntptrTy),
934 ConstantInt::get(MS.IntptrTy, ~AndMask));
936 uint64_t XorMask = MS.MapParams->XorMask;
938 OffsetLong = IRB.CreateXor(OffsetLong,
939 ConstantInt::get(MS.IntptrTy, XorMask));
943 /// \brief Compute the shadow address that corresponds to a given application
946 /// Shadow = ShadowBase + Offset
947 Value *getShadowPtr(Value *Addr, Type *ShadowTy,
949 Value *ShadowLong = getShadowPtrOffset(Addr, IRB);
950 uint64_t ShadowBase = MS.MapParams->ShadowBase;
953 IRB.CreateAdd(ShadowLong,
954 ConstantInt::get(MS.IntptrTy, ShadowBase));
955 return IRB.CreateIntToPtr(ShadowLong, PointerType::get(ShadowTy, 0));
958 /// \brief Compute the origin address that corresponds to a given application
961 /// OriginAddr = (OriginBase + Offset) & ~3ULL
962 Value *getOriginPtr(Value *Addr, IRBuilder<> &IRB, unsigned Alignment) {
963 Value *OriginLong = getShadowPtrOffset(Addr, IRB);
964 uint64_t OriginBase = MS.MapParams->OriginBase;
967 IRB.CreateAdd(OriginLong,
968 ConstantInt::get(MS.IntptrTy, OriginBase));
969 if (Alignment < kMinOriginAlignment) {
970 uint64_t Mask = kMinOriginAlignment - 1;
971 OriginLong = IRB.CreateAnd(OriginLong,
972 ConstantInt::get(MS.IntptrTy, ~Mask));
974 return IRB.CreateIntToPtr(OriginLong,
975 PointerType::get(IRB.getInt32Ty(), 0));
978 /// \brief Compute the shadow address for a given function argument.
980 /// Shadow = ParamTLS+ArgOffset.
981 Value *getShadowPtrForArgument(Value *A, IRBuilder<> &IRB,
983 Value *Base = IRB.CreatePointerCast(MS.ParamTLS, MS.IntptrTy);
984 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
985 return IRB.CreateIntToPtr(Base, PointerType::get(getShadowTy(A), 0),
989 /// \brief Compute the origin address for a given function argument.
990 Value *getOriginPtrForArgument(Value *A, IRBuilder<> &IRB,
992 if (!MS.TrackOrigins) return nullptr;
993 Value *Base = IRB.CreatePointerCast(MS.ParamOriginTLS, MS.IntptrTy);
994 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
995 return IRB.CreateIntToPtr(Base, PointerType::get(MS.OriginTy, 0),
999 /// \brief Compute the shadow address for a retval.
1000 Value *getShadowPtrForRetval(Value *A, IRBuilder<> &IRB) {
1001 Value *Base = IRB.CreatePointerCast(MS.RetvalTLS, MS.IntptrTy);
1002 return IRB.CreateIntToPtr(Base, PointerType::get(getShadowTy(A), 0),
1006 /// \brief Compute the origin address for a retval.
1007 Value *getOriginPtrForRetval(IRBuilder<> &IRB) {
1008 // We keep a single origin for the entire retval. Might be too optimistic.
1009 return MS.RetvalOriginTLS;
1012 /// \brief Set SV to be the shadow value for V.
1013 void setShadow(Value *V, Value *SV) {
1014 assert(!ShadowMap.count(V) && "Values may only have one shadow");
1015 ShadowMap[V] = PropagateShadow ? SV : getCleanShadow(V);
1018 /// \brief Set Origin to be the origin value for V.
1019 void setOrigin(Value *V, Value *Origin) {
1020 if (!MS.TrackOrigins) return;
1021 assert(!OriginMap.count(V) && "Values may only have one origin");
1022 DEBUG(dbgs() << "ORIGIN: " << *V << " ==> " << *Origin << "\n");
1023 OriginMap[V] = Origin;
1026 /// \brief Create a clean shadow value for a given value.
1028 /// Clean shadow (all zeroes) means all bits of the value are defined
1030 Constant *getCleanShadow(Value *V) {
1031 Type *ShadowTy = getShadowTy(V);
1034 return Constant::getNullValue(ShadowTy);
1037 /// \brief Create a dirty shadow of a given shadow type.
1038 Constant *getPoisonedShadow(Type *ShadowTy) {
1040 if (isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy))
1041 return Constant::getAllOnesValue(ShadowTy);
1042 if (ArrayType *AT = dyn_cast<ArrayType>(ShadowTy)) {
1043 SmallVector<Constant *, 4> Vals(AT->getNumElements(),
1044 getPoisonedShadow(AT->getElementType()));
1045 return ConstantArray::get(AT, Vals);
1047 if (StructType *ST = dyn_cast<StructType>(ShadowTy)) {
1048 SmallVector<Constant *, 4> Vals;
1049 for (unsigned i = 0, n = ST->getNumElements(); i < n; i++)
1050 Vals.push_back(getPoisonedShadow(ST->getElementType(i)));
1051 return ConstantStruct::get(ST, Vals);
1053 llvm_unreachable("Unexpected shadow type");
1056 /// \brief Create a dirty shadow for a given value.
1057 Constant *getPoisonedShadow(Value *V) {
1058 Type *ShadowTy = getShadowTy(V);
1061 return getPoisonedShadow(ShadowTy);
1064 /// \brief Create a clean (zero) origin.
1065 Value *getCleanOrigin() {
1066 return Constant::getNullValue(MS.OriginTy);
1069 /// \brief Get the shadow value for a given Value.
1071 /// This function either returns the value set earlier with setShadow,
1072 /// or extracts if from ParamTLS (for function arguments).
1073 Value *getShadow(Value *V) {
1074 if (!PropagateShadow) return getCleanShadow(V);
1075 if (Instruction *I = dyn_cast<Instruction>(V)) {
1076 // For instructions the shadow is already stored in the map.
1077 Value *Shadow = ShadowMap[V];
1079 DEBUG(dbgs() << "No shadow: " << *V << "\n" << *(I->getParent()));
1081 assert(Shadow && "No shadow for a value");
1085 if (UndefValue *U = dyn_cast<UndefValue>(V)) {
1086 Value *AllOnes = PoisonUndef ? getPoisonedShadow(V) : getCleanShadow(V);
1087 DEBUG(dbgs() << "Undef: " << *U << " ==> " << *AllOnes << "\n");
1091 if (Argument *A = dyn_cast<Argument>(V)) {
1092 // For arguments we compute the shadow on demand and store it in the map.
1093 Value **ShadowPtr = &ShadowMap[V];
1096 Function *F = A->getParent();
1097 IRBuilder<> EntryIRB(F->getEntryBlock().getFirstNonPHI());
1098 unsigned ArgOffset = 0;
1099 const DataLayout &DL = F->getParent()->getDataLayout();
1100 for (auto &FArg : F->args()) {
1101 if (!FArg.getType()->isSized()) {
1102 DEBUG(dbgs() << "Arg is not sized\n");
1107 ? DL.getTypeAllocSize(FArg.getType()->getPointerElementType())
1108 : DL.getTypeAllocSize(FArg.getType());
1110 bool Overflow = ArgOffset + Size > kParamTLSSize;
1111 Value *Base = getShadowPtrForArgument(&FArg, EntryIRB, ArgOffset);
1112 if (FArg.hasByValAttr()) {
1113 // ByVal pointer itself has clean shadow. We copy the actual
1114 // argument shadow to the underlying memory.
1115 // Figure out maximal valid memcpy alignment.
1116 unsigned ArgAlign = FArg.getParamAlignment();
1117 if (ArgAlign == 0) {
1118 Type *EltType = A->getType()->getPointerElementType();
1119 ArgAlign = DL.getABITypeAlignment(EltType);
1122 // ParamTLS overflow.
1123 EntryIRB.CreateMemSet(
1124 getShadowPtr(V, EntryIRB.getInt8Ty(), EntryIRB),
1125 Constant::getNullValue(EntryIRB.getInt8Ty()), Size, ArgAlign);
1127 unsigned CopyAlign = std::min(ArgAlign, kShadowTLSAlignment);
1128 Value *Cpy = EntryIRB.CreateMemCpy(
1129 getShadowPtr(V, EntryIRB.getInt8Ty(), EntryIRB), Base, Size,
1131 DEBUG(dbgs() << " ByValCpy: " << *Cpy << "\n");
1134 *ShadowPtr = getCleanShadow(V);
1137 // ParamTLS overflow.
1138 *ShadowPtr = getCleanShadow(V);
1141 EntryIRB.CreateAlignedLoad(Base, kShadowTLSAlignment);
1144 DEBUG(dbgs() << " ARG: " << FArg << " ==> " <<
1145 **ShadowPtr << "\n");
1146 if (MS.TrackOrigins && !Overflow) {
1148 getOriginPtrForArgument(&FArg, EntryIRB, ArgOffset);
1149 setOrigin(A, EntryIRB.CreateLoad(OriginPtr));
1151 setOrigin(A, getCleanOrigin());
1154 ArgOffset += RoundUpToAlignment(Size, kShadowTLSAlignment);
1156 assert(*ShadowPtr && "Could not find shadow for an argument");
1159 // For everything else the shadow is zero.
1160 return getCleanShadow(V);
1163 /// \brief Get the shadow for i-th argument of the instruction I.
1164 Value *getShadow(Instruction *I, int i) {
1165 return getShadow(I->getOperand(i));
1168 /// \brief Get the origin for a value.
1169 Value *getOrigin(Value *V) {
1170 if (!MS.TrackOrigins) return nullptr;
1171 if (!PropagateShadow) return getCleanOrigin();
1172 if (isa<Constant>(V)) return getCleanOrigin();
1173 assert((isa<Instruction>(V) || isa<Argument>(V)) &&
1174 "Unexpected value type in getOrigin()");
1175 Value *Origin = OriginMap[V];
1176 assert(Origin && "Missing origin");
1180 /// \brief Get the origin for i-th argument of the instruction I.
1181 Value *getOrigin(Instruction *I, int i) {
1182 return getOrigin(I->getOperand(i));
1185 /// \brief Remember the place where a shadow check should be inserted.
1187 /// This location will be later instrumented with a check that will print a
1188 /// UMR warning in runtime if the shadow value is not 0.
1189 void insertShadowCheck(Value *Shadow, Value *Origin, Instruction *OrigIns) {
1191 if (!InsertChecks) return;
1193 Type *ShadowTy = Shadow->getType();
1194 assert((isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy)) &&
1195 "Can only insert checks for integer and vector shadow types");
1197 InstrumentationList.push_back(
1198 ShadowOriginAndInsertPoint(Shadow, Origin, OrigIns));
1201 /// \brief Remember the place where a shadow check should be inserted.
1203 /// This location will be later instrumented with a check that will print a
1204 /// UMR warning in runtime if the value is not fully defined.
1205 void insertShadowCheck(Value *Val, Instruction *OrigIns) {
1207 Value *Shadow, *Origin;
1208 if (ClCheckConstantShadow) {
1209 Shadow = getShadow(Val);
1210 if (!Shadow) return;
1211 Origin = getOrigin(Val);
1213 Shadow = dyn_cast_or_null<Instruction>(getShadow(Val));
1214 if (!Shadow) return;
1215 Origin = dyn_cast_or_null<Instruction>(getOrigin(Val));
1217 insertShadowCheck(Shadow, Origin, OrigIns);
1220 AtomicOrdering addReleaseOrdering(AtomicOrdering a) {
1229 case AcquireRelease:
1230 return AcquireRelease;
1231 case SequentiallyConsistent:
1232 return SequentiallyConsistent;
1234 llvm_unreachable("Unknown ordering");
1237 AtomicOrdering addAcquireOrdering(AtomicOrdering a) {
1246 case AcquireRelease:
1247 return AcquireRelease;
1248 case SequentiallyConsistent:
1249 return SequentiallyConsistent;
1251 llvm_unreachable("Unknown ordering");
1254 // ------------------- Visitors.
1256 /// \brief Instrument LoadInst
1258 /// Loads the corresponding shadow and (optionally) origin.
1259 /// Optionally, checks that the load address is fully defined.
1260 void visitLoadInst(LoadInst &I) {
1261 assert(I.getType()->isSized() && "Load type must have size");
1262 IRBuilder<> IRB(I.getNextNode());
1263 Type *ShadowTy = getShadowTy(&I);
1264 Value *Addr = I.getPointerOperand();
1265 if (PropagateShadow && !I.getMetadata("nosanitize")) {
1266 Value *ShadowPtr = getShadowPtr(Addr, ShadowTy, IRB);
1268 IRB.CreateAlignedLoad(ShadowPtr, I.getAlignment(), "_msld"));
1270 setShadow(&I, getCleanShadow(&I));
1273 if (ClCheckAccessAddress)
1274 insertShadowCheck(I.getPointerOperand(), &I);
1277 I.setOrdering(addAcquireOrdering(I.getOrdering()));
1279 if (MS.TrackOrigins) {
1280 if (PropagateShadow) {
1281 unsigned Alignment = I.getAlignment();
1282 unsigned OriginAlignment = std::max(kMinOriginAlignment, Alignment);
1283 setOrigin(&I, IRB.CreateAlignedLoad(getOriginPtr(Addr, IRB, Alignment),
1286 setOrigin(&I, getCleanOrigin());
1291 /// \brief Instrument StoreInst
1293 /// Stores the corresponding shadow and (optionally) origin.
1294 /// Optionally, checks that the store address is fully defined.
1295 void visitStoreInst(StoreInst &I) {
1296 StoreList.push_back(&I);
1299 void handleCASOrRMW(Instruction &I) {
1300 assert(isa<AtomicRMWInst>(I) || isa<AtomicCmpXchgInst>(I));
1302 IRBuilder<> IRB(&I);
1303 Value *Addr = I.getOperand(0);
1304 Value *ShadowPtr = getShadowPtr(Addr, I.getType(), IRB);
1306 if (ClCheckAccessAddress)
1307 insertShadowCheck(Addr, &I);
1309 // Only test the conditional argument of cmpxchg instruction.
1310 // The other argument can potentially be uninitialized, but we can not
1311 // detect this situation reliably without possible false positives.
1312 if (isa<AtomicCmpXchgInst>(I))
1313 insertShadowCheck(I.getOperand(1), &I);
1315 IRB.CreateStore(getCleanShadow(&I), ShadowPtr);
1317 setShadow(&I, getCleanShadow(&I));
1318 setOrigin(&I, getCleanOrigin());
1321 void visitAtomicRMWInst(AtomicRMWInst &I) {
1323 I.setOrdering(addReleaseOrdering(I.getOrdering()));
1326 void visitAtomicCmpXchgInst(AtomicCmpXchgInst &I) {
1328 I.setSuccessOrdering(addReleaseOrdering(I.getSuccessOrdering()));
1331 // Vector manipulation.
1332 void visitExtractElementInst(ExtractElementInst &I) {
1333 insertShadowCheck(I.getOperand(1), &I);
1334 IRBuilder<> IRB(&I);
1335 setShadow(&I, IRB.CreateExtractElement(getShadow(&I, 0), I.getOperand(1),
1337 setOrigin(&I, getOrigin(&I, 0));
1340 void visitInsertElementInst(InsertElementInst &I) {
1341 insertShadowCheck(I.getOperand(2), &I);
1342 IRBuilder<> IRB(&I);
1343 setShadow(&I, IRB.CreateInsertElement(getShadow(&I, 0), getShadow(&I, 1),
1344 I.getOperand(2), "_msprop"));
1345 setOriginForNaryOp(I);
1348 void visitShuffleVectorInst(ShuffleVectorInst &I) {
1349 insertShadowCheck(I.getOperand(2), &I);
1350 IRBuilder<> IRB(&I);
1351 setShadow(&I, IRB.CreateShuffleVector(getShadow(&I, 0), getShadow(&I, 1),
1352 I.getOperand(2), "_msprop"));
1353 setOriginForNaryOp(I);
1357 void visitSExtInst(SExtInst &I) {
1358 IRBuilder<> IRB(&I);
1359 setShadow(&I, IRB.CreateSExt(getShadow(&I, 0), I.getType(), "_msprop"));
1360 setOrigin(&I, getOrigin(&I, 0));
1363 void visitZExtInst(ZExtInst &I) {
1364 IRBuilder<> IRB(&I);
1365 setShadow(&I, IRB.CreateZExt(getShadow(&I, 0), I.getType(), "_msprop"));
1366 setOrigin(&I, getOrigin(&I, 0));
1369 void visitTruncInst(TruncInst &I) {
1370 IRBuilder<> IRB(&I);
1371 setShadow(&I, IRB.CreateTrunc(getShadow(&I, 0), I.getType(), "_msprop"));
1372 setOrigin(&I, getOrigin(&I, 0));
1375 void visitBitCastInst(BitCastInst &I) {
1376 // Special case: if this is the bitcast (there is exactly 1 allowed) between
1377 // a musttail call and a ret, don't instrument. New instructions are not
1378 // allowed after a musttail call.
1379 if (auto *CI = dyn_cast<CallInst>(I.getOperand(0)))
1380 if (CI->isMustTailCall())
1382 IRBuilder<> IRB(&I);
1383 setShadow(&I, IRB.CreateBitCast(getShadow(&I, 0), getShadowTy(&I)));
1384 setOrigin(&I, getOrigin(&I, 0));
1387 void visitPtrToIntInst(PtrToIntInst &I) {
1388 IRBuilder<> IRB(&I);
1389 setShadow(&I, IRB.CreateIntCast(getShadow(&I, 0), getShadowTy(&I), false,
1390 "_msprop_ptrtoint"));
1391 setOrigin(&I, getOrigin(&I, 0));
1394 void visitIntToPtrInst(IntToPtrInst &I) {
1395 IRBuilder<> IRB(&I);
1396 setShadow(&I, IRB.CreateIntCast(getShadow(&I, 0), getShadowTy(&I), false,
1397 "_msprop_inttoptr"));
1398 setOrigin(&I, getOrigin(&I, 0));
1401 void visitFPToSIInst(CastInst& I) { handleShadowOr(I); }
1402 void visitFPToUIInst(CastInst& I) { handleShadowOr(I); }
1403 void visitSIToFPInst(CastInst& I) { handleShadowOr(I); }
1404 void visitUIToFPInst(CastInst& I) { handleShadowOr(I); }
1405 void visitFPExtInst(CastInst& I) { handleShadowOr(I); }
1406 void visitFPTruncInst(CastInst& I) { handleShadowOr(I); }
1408 /// \brief Propagate shadow for bitwise AND.
1410 /// This code is exact, i.e. if, for example, a bit in the left argument
1411 /// is defined and 0, then neither the value not definedness of the
1412 /// corresponding bit in B don't affect the resulting shadow.
1413 void visitAnd(BinaryOperator &I) {
1414 IRBuilder<> IRB(&I);
1415 // "And" of 0 and a poisoned value results in unpoisoned value.
1416 // 1&1 => 1; 0&1 => 0; p&1 => p;
1417 // 1&0 => 0; 0&0 => 0; p&0 => 0;
1418 // 1&p => p; 0&p => 0; p&p => p;
1419 // S = (S1 & S2) | (V1 & S2) | (S1 & V2)
1420 Value *S1 = getShadow(&I, 0);
1421 Value *S2 = getShadow(&I, 1);
1422 Value *V1 = I.getOperand(0);
1423 Value *V2 = I.getOperand(1);
1424 if (V1->getType() != S1->getType()) {
1425 V1 = IRB.CreateIntCast(V1, S1->getType(), false);
1426 V2 = IRB.CreateIntCast(V2, S2->getType(), false);
1428 Value *S1S2 = IRB.CreateAnd(S1, S2);
1429 Value *V1S2 = IRB.CreateAnd(V1, S2);
1430 Value *S1V2 = IRB.CreateAnd(S1, V2);
1431 setShadow(&I, IRB.CreateOr(S1S2, IRB.CreateOr(V1S2, S1V2)));
1432 setOriginForNaryOp(I);
1435 void visitOr(BinaryOperator &I) {
1436 IRBuilder<> IRB(&I);
1437 // "Or" of 1 and a poisoned value results in unpoisoned value.
1438 // 1|1 => 1; 0|1 => 1; p|1 => 1;
1439 // 1|0 => 1; 0|0 => 0; p|0 => p;
1440 // 1|p => 1; 0|p => p; p|p => p;
1441 // S = (S1 & S2) | (~V1 & S2) | (S1 & ~V2)
1442 Value *S1 = getShadow(&I, 0);
1443 Value *S2 = getShadow(&I, 1);
1444 Value *V1 = IRB.CreateNot(I.getOperand(0));
1445 Value *V2 = IRB.CreateNot(I.getOperand(1));
1446 if (V1->getType() != S1->getType()) {
1447 V1 = IRB.CreateIntCast(V1, S1->getType(), false);
1448 V2 = IRB.CreateIntCast(V2, S2->getType(), false);
1450 Value *S1S2 = IRB.CreateAnd(S1, S2);
1451 Value *V1S2 = IRB.CreateAnd(V1, S2);
1452 Value *S1V2 = IRB.CreateAnd(S1, V2);
1453 setShadow(&I, IRB.CreateOr(S1S2, IRB.CreateOr(V1S2, S1V2)));
1454 setOriginForNaryOp(I);
1457 /// \brief Default propagation of shadow and/or origin.
1459 /// This class implements the general case of shadow propagation, used in all
1460 /// cases where we don't know and/or don't care about what the operation
1461 /// actually does. It converts all input shadow values to a common type
1462 /// (extending or truncating as necessary), and bitwise OR's them.
1464 /// This is much cheaper than inserting checks (i.e. requiring inputs to be
1465 /// fully initialized), and less prone to false positives.
1467 /// This class also implements the general case of origin propagation. For a
1468 /// Nary operation, result origin is set to the origin of an argument that is
1469 /// not entirely initialized. If there is more than one such arguments, the
1470 /// rightmost of them is picked. It does not matter which one is picked if all
1471 /// arguments are initialized.
1472 template <bool CombineShadow>
1477 MemorySanitizerVisitor *MSV;
1480 Combiner(MemorySanitizerVisitor *MSV, IRBuilder<> &IRB) :
1481 Shadow(nullptr), Origin(nullptr), IRB(IRB), MSV(MSV) {}
1483 /// \brief Add a pair of shadow and origin values to the mix.
1484 Combiner &Add(Value *OpShadow, Value *OpOrigin) {
1485 if (CombineShadow) {
1490 OpShadow = MSV->CreateShadowCast(IRB, OpShadow, Shadow->getType());
1491 Shadow = IRB.CreateOr(Shadow, OpShadow, "_msprop");
1495 if (MSV->MS.TrackOrigins) {
1500 Constant *ConstOrigin = dyn_cast<Constant>(OpOrigin);
1501 // No point in adding something that might result in 0 origin value.
1502 if (!ConstOrigin || !ConstOrigin->isNullValue()) {
1503 Value *FlatShadow = MSV->convertToShadowTyNoVec(OpShadow, IRB);
1505 IRB.CreateICmpNE(FlatShadow, MSV->getCleanShadow(FlatShadow));
1506 Origin = IRB.CreateSelect(Cond, OpOrigin, Origin);
1513 /// \brief Add an application value to the mix.
1514 Combiner &Add(Value *V) {
1515 Value *OpShadow = MSV->getShadow(V);
1516 Value *OpOrigin = MSV->MS.TrackOrigins ? MSV->getOrigin(V) : nullptr;
1517 return Add(OpShadow, OpOrigin);
1520 /// \brief Set the current combined values as the given instruction's shadow
1522 void Done(Instruction *I) {
1523 if (CombineShadow) {
1525 Shadow = MSV->CreateShadowCast(IRB, Shadow, MSV->getShadowTy(I));
1526 MSV->setShadow(I, Shadow);
1528 if (MSV->MS.TrackOrigins) {
1530 MSV->setOrigin(I, Origin);
1535 typedef Combiner<true> ShadowAndOriginCombiner;
1536 typedef Combiner<false> OriginCombiner;
1538 /// \brief Propagate origin for arbitrary operation.
1539 void setOriginForNaryOp(Instruction &I) {
1540 if (!MS.TrackOrigins) return;
1541 IRBuilder<> IRB(&I);
1542 OriginCombiner OC(this, IRB);
1543 for (Instruction::op_iterator OI = I.op_begin(); OI != I.op_end(); ++OI)
1548 size_t VectorOrPrimitiveTypeSizeInBits(Type *Ty) {
1549 assert(!(Ty->isVectorTy() && Ty->getScalarType()->isPointerTy()) &&
1550 "Vector of pointers is not a valid shadow type");
1551 return Ty->isVectorTy() ?
1552 Ty->getVectorNumElements() * Ty->getScalarSizeInBits() :
1553 Ty->getPrimitiveSizeInBits();
1556 /// \brief Cast between two shadow types, extending or truncating as
1558 Value *CreateShadowCast(IRBuilder<> &IRB, Value *V, Type *dstTy,
1559 bool Signed = false) {
1560 Type *srcTy = V->getType();
1561 if (dstTy->isIntegerTy() && srcTy->isIntegerTy())
1562 return IRB.CreateIntCast(V, dstTy, Signed);
1563 if (dstTy->isVectorTy() && srcTy->isVectorTy() &&
1564 dstTy->getVectorNumElements() == srcTy->getVectorNumElements())
1565 return IRB.CreateIntCast(V, dstTy, Signed);
1566 size_t srcSizeInBits = VectorOrPrimitiveTypeSizeInBits(srcTy);
1567 size_t dstSizeInBits = VectorOrPrimitiveTypeSizeInBits(dstTy);
1568 Value *V1 = IRB.CreateBitCast(V, Type::getIntNTy(*MS.C, srcSizeInBits));
1570 IRB.CreateIntCast(V1, Type::getIntNTy(*MS.C, dstSizeInBits), Signed);
1571 return IRB.CreateBitCast(V2, dstTy);
1572 // TODO: handle struct types.
1575 /// \brief Cast an application value to the type of its own shadow.
1576 Value *CreateAppToShadowCast(IRBuilder<> &IRB, Value *V) {
1577 Type *ShadowTy = getShadowTy(V);
1578 if (V->getType() == ShadowTy)
1580 if (V->getType()->isPtrOrPtrVectorTy())
1581 return IRB.CreatePtrToInt(V, ShadowTy);
1583 return IRB.CreateBitCast(V, ShadowTy);
1586 /// \brief Propagate shadow for arbitrary operation.
1587 void handleShadowOr(Instruction &I) {
1588 IRBuilder<> IRB(&I);
1589 ShadowAndOriginCombiner SC(this, IRB);
1590 for (Instruction::op_iterator OI = I.op_begin(); OI != I.op_end(); ++OI)
1595 // \brief Handle multiplication by constant.
1597 // Handle a special case of multiplication by constant that may have one or
1598 // more zeros in the lower bits. This makes corresponding number of lower bits
1599 // of the result zero as well. We model it by shifting the other operand
1600 // shadow left by the required number of bits. Effectively, we transform
1601 // (X * (A * 2**B)) to ((X << B) * A) and instrument (X << B) as (Sx << B).
1602 // We use multiplication by 2**N instead of shift to cover the case of
1603 // multiplication by 0, which may occur in some elements of a vector operand.
1604 void handleMulByConstant(BinaryOperator &I, Constant *ConstArg,
1606 Constant *ShadowMul;
1607 Type *Ty = ConstArg->getType();
1608 if (Ty->isVectorTy()) {
1609 unsigned NumElements = Ty->getVectorNumElements();
1610 Type *EltTy = Ty->getSequentialElementType();
1611 SmallVector<Constant *, 16> Elements;
1612 for (unsigned Idx = 0; Idx < NumElements; ++Idx) {
1614 dyn_cast<ConstantInt>(ConstArg->getAggregateElement(Idx));
1615 APInt V = Elt->getValue();
1616 APInt V2 = APInt(V.getBitWidth(), 1) << V.countTrailingZeros();
1617 Elements.push_back(ConstantInt::get(EltTy, V2));
1619 ShadowMul = ConstantVector::get(Elements);
1621 ConstantInt *Elt = dyn_cast<ConstantInt>(ConstArg);
1622 APInt V = Elt->getValue();
1623 APInt V2 = APInt(V.getBitWidth(), 1) << V.countTrailingZeros();
1624 ShadowMul = ConstantInt::get(Elt->getType(), V2);
1627 IRBuilder<> IRB(&I);
1629 IRB.CreateMul(getShadow(OtherArg), ShadowMul, "msprop_mul_cst"));
1630 setOrigin(&I, getOrigin(OtherArg));
1633 void visitMul(BinaryOperator &I) {
1634 Constant *constOp0 = dyn_cast<Constant>(I.getOperand(0));
1635 Constant *constOp1 = dyn_cast<Constant>(I.getOperand(1));
1636 if (constOp0 && !constOp1)
1637 handleMulByConstant(I, constOp0, I.getOperand(1));
1638 else if (constOp1 && !constOp0)
1639 handleMulByConstant(I, constOp1, I.getOperand(0));
1644 void visitFAdd(BinaryOperator &I) { handleShadowOr(I); }
1645 void visitFSub(BinaryOperator &I) { handleShadowOr(I); }
1646 void visitFMul(BinaryOperator &I) { handleShadowOr(I); }
1647 void visitAdd(BinaryOperator &I) { handleShadowOr(I); }
1648 void visitSub(BinaryOperator &I) { handleShadowOr(I); }
1649 void visitXor(BinaryOperator &I) { handleShadowOr(I); }
1651 void handleDiv(Instruction &I) {
1652 IRBuilder<> IRB(&I);
1653 // Strict on the second argument.
1654 insertShadowCheck(I.getOperand(1), &I);
1655 setShadow(&I, getShadow(&I, 0));
1656 setOrigin(&I, getOrigin(&I, 0));
1659 void visitUDiv(BinaryOperator &I) { handleDiv(I); }
1660 void visitSDiv(BinaryOperator &I) { handleDiv(I); }
1661 void visitFDiv(BinaryOperator &I) { handleDiv(I); }
1662 void visitURem(BinaryOperator &I) { handleDiv(I); }
1663 void visitSRem(BinaryOperator &I) { handleDiv(I); }
1664 void visitFRem(BinaryOperator &I) { handleDiv(I); }
1666 /// \brief Instrument == and != comparisons.
1668 /// Sometimes the comparison result is known even if some of the bits of the
1669 /// arguments are not.
1670 void handleEqualityComparison(ICmpInst &I) {
1671 IRBuilder<> IRB(&I);
1672 Value *A = I.getOperand(0);
1673 Value *B = I.getOperand(1);
1674 Value *Sa = getShadow(A);
1675 Value *Sb = getShadow(B);
1677 // Get rid of pointers and vectors of pointers.
1678 // For ints (and vectors of ints), types of A and Sa match,
1679 // and this is a no-op.
1680 A = IRB.CreatePointerCast(A, Sa->getType());
1681 B = IRB.CreatePointerCast(B, Sb->getType());
1683 // A == B <==> (C = A^B) == 0
1684 // A != B <==> (C = A^B) != 0
1686 Value *C = IRB.CreateXor(A, B);
1687 Value *Sc = IRB.CreateOr(Sa, Sb);
1688 // Now dealing with i = (C == 0) comparison (or C != 0, does not matter now)
1689 // Result is defined if one of the following is true
1690 // * there is a defined 1 bit in C
1691 // * C is fully defined
1692 // Si = !(C & ~Sc) && Sc
1693 Value *Zero = Constant::getNullValue(Sc->getType());
1694 Value *MinusOne = Constant::getAllOnesValue(Sc->getType());
1696 IRB.CreateAnd(IRB.CreateICmpNE(Sc, Zero),
1698 IRB.CreateAnd(IRB.CreateXor(Sc, MinusOne), C), Zero));
1699 Si->setName("_msprop_icmp");
1701 setOriginForNaryOp(I);
1704 /// \brief Build the lowest possible value of V, taking into account V's
1705 /// uninitialized bits.
1706 Value *getLowestPossibleValue(IRBuilder<> &IRB, Value *A, Value *Sa,
1709 // Split shadow into sign bit and other bits.
1710 Value *SaOtherBits = IRB.CreateLShr(IRB.CreateShl(Sa, 1), 1);
1711 Value *SaSignBit = IRB.CreateXor(Sa, SaOtherBits);
1712 // Maximise the undefined shadow bit, minimize other undefined bits.
1714 IRB.CreateOr(IRB.CreateAnd(A, IRB.CreateNot(SaOtherBits)), SaSignBit);
1716 // Minimize undefined bits.
1717 return IRB.CreateAnd(A, IRB.CreateNot(Sa));
1721 /// \brief Build the highest possible value of V, taking into account V's
1722 /// uninitialized bits.
1723 Value *getHighestPossibleValue(IRBuilder<> &IRB, Value *A, Value *Sa,
1726 // Split shadow into sign bit and other bits.
1727 Value *SaOtherBits = IRB.CreateLShr(IRB.CreateShl(Sa, 1), 1);
1728 Value *SaSignBit = IRB.CreateXor(Sa, SaOtherBits);
1729 // Minimise the undefined shadow bit, maximise other undefined bits.
1731 IRB.CreateOr(IRB.CreateAnd(A, IRB.CreateNot(SaSignBit)), SaOtherBits);
1733 // Maximize undefined bits.
1734 return IRB.CreateOr(A, Sa);
1738 /// \brief Instrument relational comparisons.
1740 /// This function does exact shadow propagation for all relational
1741 /// comparisons of integers, pointers and vectors of those.
1742 /// FIXME: output seems suboptimal when one of the operands is a constant
1743 void handleRelationalComparisonExact(ICmpInst &I) {
1744 IRBuilder<> IRB(&I);
1745 Value *A = I.getOperand(0);
1746 Value *B = I.getOperand(1);
1747 Value *Sa = getShadow(A);
1748 Value *Sb = getShadow(B);
1750 // Get rid of pointers and vectors of pointers.
1751 // For ints (and vectors of ints), types of A and Sa match,
1752 // and this is a no-op.
1753 A = IRB.CreatePointerCast(A, Sa->getType());
1754 B = IRB.CreatePointerCast(B, Sb->getType());
1756 // Let [a0, a1] be the interval of possible values of A, taking into account
1757 // its undefined bits. Let [b0, b1] be the interval of possible values of B.
1758 // Then (A cmp B) is defined iff (a0 cmp b1) == (a1 cmp b0).
1759 bool IsSigned = I.isSigned();
1760 Value *S1 = IRB.CreateICmp(I.getPredicate(),
1761 getLowestPossibleValue(IRB, A, Sa, IsSigned),
1762 getHighestPossibleValue(IRB, B, Sb, IsSigned));
1763 Value *S2 = IRB.CreateICmp(I.getPredicate(),
1764 getHighestPossibleValue(IRB, A, Sa, IsSigned),
1765 getLowestPossibleValue(IRB, B, Sb, IsSigned));
1766 Value *Si = IRB.CreateXor(S1, S2);
1768 setOriginForNaryOp(I);
1771 /// \brief Instrument signed relational comparisons.
1773 /// Handle sign bit tests: x<0, x>=0, x<=-1, x>-1 by propagating the highest
1774 /// bit of the shadow. Everything else is delegated to handleShadowOr().
1775 void handleSignedRelationalComparison(ICmpInst &I) {
1777 Value *op = nullptr;
1778 CmpInst::Predicate pre;
1779 if ((constOp = dyn_cast<Constant>(I.getOperand(1)))) {
1780 op = I.getOperand(0);
1781 pre = I.getPredicate();
1782 } else if ((constOp = dyn_cast<Constant>(I.getOperand(0)))) {
1783 op = I.getOperand(1);
1784 pre = I.getSwappedPredicate();
1790 if ((constOp->isNullValue() &&
1791 (pre == CmpInst::ICMP_SLT || pre == CmpInst::ICMP_SGE)) ||
1792 (constOp->isAllOnesValue() &&
1793 (pre == CmpInst::ICMP_SGT || pre == CmpInst::ICMP_SLE))) {
1794 IRBuilder<> IRB(&I);
1795 Value *Shadow = IRB.CreateICmpSLT(getShadow(op), getCleanShadow(op),
1797 setShadow(&I, Shadow);
1798 setOrigin(&I, getOrigin(op));
1804 void visitICmpInst(ICmpInst &I) {
1805 if (!ClHandleICmp) {
1809 if (I.isEquality()) {
1810 handleEqualityComparison(I);
1814 assert(I.isRelational());
1815 if (ClHandleICmpExact) {
1816 handleRelationalComparisonExact(I);
1820 handleSignedRelationalComparison(I);
1824 assert(I.isUnsigned());
1825 if ((isa<Constant>(I.getOperand(0)) || isa<Constant>(I.getOperand(1)))) {
1826 handleRelationalComparisonExact(I);
1833 void visitFCmpInst(FCmpInst &I) {
1837 void handleShift(BinaryOperator &I) {
1838 IRBuilder<> IRB(&I);
1839 // If any of the S2 bits are poisoned, the whole thing is poisoned.
1840 // Otherwise perform the same shift on S1.
1841 Value *S1 = getShadow(&I, 0);
1842 Value *S2 = getShadow(&I, 1);
1843 Value *S2Conv = IRB.CreateSExt(IRB.CreateICmpNE(S2, getCleanShadow(S2)),
1845 Value *V2 = I.getOperand(1);
1846 Value *Shift = IRB.CreateBinOp(I.getOpcode(), S1, V2);
1847 setShadow(&I, IRB.CreateOr(Shift, S2Conv));
1848 setOriginForNaryOp(I);
1851 void visitShl(BinaryOperator &I) { handleShift(I); }
1852 void visitAShr(BinaryOperator &I) { handleShift(I); }
1853 void visitLShr(BinaryOperator &I) { handleShift(I); }
1855 /// \brief Instrument llvm.memmove
1857 /// At this point we don't know if llvm.memmove will be inlined or not.
1858 /// If we don't instrument it and it gets inlined,
1859 /// our interceptor will not kick in and we will lose the memmove.
1860 /// If we instrument the call here, but it does not get inlined,
1861 /// we will memove the shadow twice: which is bad in case
1862 /// of overlapping regions. So, we simply lower the intrinsic to a call.
1864 /// Similar situation exists for memcpy and memset.
1865 void visitMemMoveInst(MemMoveInst &I) {
1866 IRBuilder<> IRB(&I);
1869 {IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1870 IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()),
1871 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false)});
1872 I.eraseFromParent();
1875 // Similar to memmove: avoid copying shadow twice.
1876 // This is somewhat unfortunate as it may slowdown small constant memcpys.
1877 // FIXME: consider doing manual inline for small constant sizes and proper
1879 void visitMemCpyInst(MemCpyInst &I) {
1880 IRBuilder<> IRB(&I);
1883 {IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1884 IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()),
1885 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false)});
1886 I.eraseFromParent();
1890 void visitMemSetInst(MemSetInst &I) {
1891 IRBuilder<> IRB(&I);
1894 {IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
1895 IRB.CreateIntCast(I.getArgOperand(1), IRB.getInt32Ty(), false),
1896 IRB.CreateIntCast(I.getArgOperand(2), MS.IntptrTy, false)});
1897 I.eraseFromParent();
1900 void visitVAStartInst(VAStartInst &I) {
1901 VAHelper->visitVAStartInst(I);
1904 void visitVACopyInst(VACopyInst &I) {
1905 VAHelper->visitVACopyInst(I);
1908 enum IntrinsicKind {
1909 IK_DoesNotAccessMemory,
1914 static IntrinsicKind getIntrinsicKind(Intrinsic::ID iid) {
1915 const int FMRB_DoesNotAccessMemory = IK_DoesNotAccessMemory;
1916 const int FMRB_OnlyReadsArgumentPointees = IK_OnlyReadsMemory;
1917 const int FMRB_OnlyReadsMemory = IK_OnlyReadsMemory;
1918 const int FMRB_OnlyAccessesArgumentPointees = IK_WritesMemory;
1919 const int FMRB_UnknownModRefBehavior = IK_WritesMemory;
1920 #define GET_INTRINSIC_MODREF_BEHAVIOR
1921 #define FunctionModRefBehavior IntrinsicKind
1922 #include "llvm/IR/Intrinsics.gen"
1923 #undef FunctionModRefBehavior
1924 #undef GET_INTRINSIC_MODREF_BEHAVIOR
1927 /// \brief Handle vector store-like intrinsics.
1929 /// Instrument intrinsics that look like a simple SIMD store: writes memory,
1930 /// has 1 pointer argument and 1 vector argument, returns void.
1931 bool handleVectorStoreIntrinsic(IntrinsicInst &I) {
1932 IRBuilder<> IRB(&I);
1933 Value* Addr = I.getArgOperand(0);
1934 Value *Shadow = getShadow(&I, 1);
1935 Value *ShadowPtr = getShadowPtr(Addr, Shadow->getType(), IRB);
1937 // We don't know the pointer alignment (could be unaligned SSE store!).
1938 // Have to assume to worst case.
1939 IRB.CreateAlignedStore(Shadow, ShadowPtr, 1);
1941 if (ClCheckAccessAddress)
1942 insertShadowCheck(Addr, &I);
1944 // FIXME: use ClStoreCleanOrigin
1945 // FIXME: factor out common code from materializeStores
1946 if (MS.TrackOrigins)
1947 IRB.CreateStore(getOrigin(&I, 1), getOriginPtr(Addr, IRB, 1));
1951 /// \brief Handle vector load-like intrinsics.
1953 /// Instrument intrinsics that look like a simple SIMD load: reads memory,
1954 /// has 1 pointer argument, returns a vector.
1955 bool handleVectorLoadIntrinsic(IntrinsicInst &I) {
1956 IRBuilder<> IRB(&I);
1957 Value *Addr = I.getArgOperand(0);
1959 Type *ShadowTy = getShadowTy(&I);
1960 if (PropagateShadow) {
1961 Value *ShadowPtr = getShadowPtr(Addr, ShadowTy, IRB);
1962 // We don't know the pointer alignment (could be unaligned SSE load!).
1963 // Have to assume to worst case.
1964 setShadow(&I, IRB.CreateAlignedLoad(ShadowPtr, 1, "_msld"));
1966 setShadow(&I, getCleanShadow(&I));
1969 if (ClCheckAccessAddress)
1970 insertShadowCheck(Addr, &I);
1972 if (MS.TrackOrigins) {
1973 if (PropagateShadow)
1974 setOrigin(&I, IRB.CreateLoad(getOriginPtr(Addr, IRB, 1)));
1976 setOrigin(&I, getCleanOrigin());
1981 /// \brief Handle (SIMD arithmetic)-like intrinsics.
1983 /// Instrument intrinsics with any number of arguments of the same type,
1984 /// equal to the return type. The type should be simple (no aggregates or
1985 /// pointers; vectors are fine).
1986 /// Caller guarantees that this intrinsic does not access memory.
1987 bool maybeHandleSimpleNomemIntrinsic(IntrinsicInst &I) {
1988 Type *RetTy = I.getType();
1989 if (!(RetTy->isIntOrIntVectorTy() ||
1990 RetTy->isFPOrFPVectorTy() ||
1991 RetTy->isX86_MMXTy()))
1994 unsigned NumArgOperands = I.getNumArgOperands();
1996 for (unsigned i = 0; i < NumArgOperands; ++i) {
1997 Type *Ty = I.getArgOperand(i)->getType();
2002 IRBuilder<> IRB(&I);
2003 ShadowAndOriginCombiner SC(this, IRB);
2004 for (unsigned i = 0; i < NumArgOperands; ++i)
2005 SC.Add(I.getArgOperand(i));
2011 /// \brief Heuristically instrument unknown intrinsics.
2013 /// The main purpose of this code is to do something reasonable with all
2014 /// random intrinsics we might encounter, most importantly - SIMD intrinsics.
2015 /// We recognize several classes of intrinsics by their argument types and
2016 /// ModRefBehaviour and apply special intrumentation when we are reasonably
2017 /// sure that we know what the intrinsic does.
2019 /// We special-case intrinsics where this approach fails. See llvm.bswap
2020 /// handling as an example of that.
2021 bool handleUnknownIntrinsic(IntrinsicInst &I) {
2022 unsigned NumArgOperands = I.getNumArgOperands();
2023 if (NumArgOperands == 0)
2026 Intrinsic::ID iid = I.getIntrinsicID();
2027 IntrinsicKind IK = getIntrinsicKind(iid);
2028 bool OnlyReadsMemory = IK == IK_OnlyReadsMemory;
2029 bool WritesMemory = IK == IK_WritesMemory;
2030 assert(!(OnlyReadsMemory && WritesMemory));
2032 if (NumArgOperands == 2 &&
2033 I.getArgOperand(0)->getType()->isPointerTy() &&
2034 I.getArgOperand(1)->getType()->isVectorTy() &&
2035 I.getType()->isVoidTy() &&
2037 // This looks like a vector store.
2038 return handleVectorStoreIntrinsic(I);
2041 if (NumArgOperands == 1 &&
2042 I.getArgOperand(0)->getType()->isPointerTy() &&
2043 I.getType()->isVectorTy() &&
2045 // This looks like a vector load.
2046 return handleVectorLoadIntrinsic(I);
2049 if (!OnlyReadsMemory && !WritesMemory)
2050 if (maybeHandleSimpleNomemIntrinsic(I))
2053 // FIXME: detect and handle SSE maskstore/maskload
2057 void handleBswap(IntrinsicInst &I) {
2058 IRBuilder<> IRB(&I);
2059 Value *Op = I.getArgOperand(0);
2060 Type *OpType = Op->getType();
2061 Function *BswapFunc = Intrinsic::getDeclaration(
2062 F.getParent(), Intrinsic::bswap, makeArrayRef(&OpType, 1));
2063 setShadow(&I, IRB.CreateCall(BswapFunc, getShadow(Op)));
2064 setOrigin(&I, getOrigin(Op));
2067 // \brief Instrument vector convert instrinsic.
2069 // This function instruments intrinsics like cvtsi2ss:
2070 // %Out = int_xxx_cvtyyy(%ConvertOp)
2072 // %Out = int_xxx_cvtyyy(%CopyOp, %ConvertOp)
2073 // Intrinsic converts \p NumUsedElements elements of \p ConvertOp to the same
2074 // number \p Out elements, and (if has 2 arguments) copies the rest of the
2075 // elements from \p CopyOp.
2076 // In most cases conversion involves floating-point value which may trigger a
2077 // hardware exception when not fully initialized. For this reason we require
2078 // \p ConvertOp[0:NumUsedElements] to be fully initialized and trap otherwise.
2079 // We copy the shadow of \p CopyOp[NumUsedElements:] to \p
2080 // Out[NumUsedElements:]. This means that intrinsics without \p CopyOp always
2081 // return a fully initialized value.
2082 void handleVectorConvertIntrinsic(IntrinsicInst &I, int NumUsedElements) {
2083 IRBuilder<> IRB(&I);
2084 Value *CopyOp, *ConvertOp;
2086 switch (I.getNumArgOperands()) {
2088 assert(isa<ConstantInt>(I.getArgOperand(2)) && "Invalid rounding mode");
2090 CopyOp = I.getArgOperand(0);
2091 ConvertOp = I.getArgOperand(1);
2094 ConvertOp = I.getArgOperand(0);
2098 llvm_unreachable("Cvt intrinsic with unsupported number of arguments.");
2101 // The first *NumUsedElements* elements of ConvertOp are converted to the
2102 // same number of output elements. The rest of the output is copied from
2103 // CopyOp, or (if not available) filled with zeroes.
2104 // Combine shadow for elements of ConvertOp that are used in this operation,
2105 // and insert a check.
2106 // FIXME: consider propagating shadow of ConvertOp, at least in the case of
2107 // int->any conversion.
2108 Value *ConvertShadow = getShadow(ConvertOp);
2109 Value *AggShadow = nullptr;
2110 if (ConvertOp->getType()->isVectorTy()) {
2111 AggShadow = IRB.CreateExtractElement(
2112 ConvertShadow, ConstantInt::get(IRB.getInt32Ty(), 0));
2113 for (int i = 1; i < NumUsedElements; ++i) {
2114 Value *MoreShadow = IRB.CreateExtractElement(
2115 ConvertShadow, ConstantInt::get(IRB.getInt32Ty(), i));
2116 AggShadow = IRB.CreateOr(AggShadow, MoreShadow);
2119 AggShadow = ConvertShadow;
2121 assert(AggShadow->getType()->isIntegerTy());
2122 insertShadowCheck(AggShadow, getOrigin(ConvertOp), &I);
2124 // Build result shadow by zero-filling parts of CopyOp shadow that come from
2127 assert(CopyOp->getType() == I.getType());
2128 assert(CopyOp->getType()->isVectorTy());
2129 Value *ResultShadow = getShadow(CopyOp);
2130 Type *EltTy = ResultShadow->getType()->getVectorElementType();
2131 for (int i = 0; i < NumUsedElements; ++i) {
2132 ResultShadow = IRB.CreateInsertElement(
2133 ResultShadow, ConstantInt::getNullValue(EltTy),
2134 ConstantInt::get(IRB.getInt32Ty(), i));
2136 setShadow(&I, ResultShadow);
2137 setOrigin(&I, getOrigin(CopyOp));
2139 setShadow(&I, getCleanShadow(&I));
2140 setOrigin(&I, getCleanOrigin());
2144 // Given a scalar or vector, extract lower 64 bits (or less), and return all
2145 // zeroes if it is zero, and all ones otherwise.
2146 Value *Lower64ShadowExtend(IRBuilder<> &IRB, Value *S, Type *T) {
2147 if (S->getType()->isVectorTy())
2148 S = CreateShadowCast(IRB, S, IRB.getInt64Ty(), /* Signed */ true);
2149 assert(S->getType()->getPrimitiveSizeInBits() <= 64);
2150 Value *S2 = IRB.CreateICmpNE(S, getCleanShadow(S));
2151 return CreateShadowCast(IRB, S2, T, /* Signed */ true);
2154 Value *VariableShadowExtend(IRBuilder<> &IRB, Value *S) {
2155 Type *T = S->getType();
2156 assert(T->isVectorTy());
2157 Value *S2 = IRB.CreateICmpNE(S, getCleanShadow(S));
2158 return IRB.CreateSExt(S2, T);
2161 // \brief Instrument vector shift instrinsic.
2163 // This function instruments intrinsics like int_x86_avx2_psll_w.
2164 // Intrinsic shifts %In by %ShiftSize bits.
2165 // %ShiftSize may be a vector. In that case the lower 64 bits determine shift
2166 // size, and the rest is ignored. Behavior is defined even if shift size is
2167 // greater than register (or field) width.
2168 void handleVectorShiftIntrinsic(IntrinsicInst &I, bool Variable) {
2169 assert(I.getNumArgOperands() == 2);
2170 IRBuilder<> IRB(&I);
2171 // If any of the S2 bits are poisoned, the whole thing is poisoned.
2172 // Otherwise perform the same shift on S1.
2173 Value *S1 = getShadow(&I, 0);
2174 Value *S2 = getShadow(&I, 1);
2175 Value *S2Conv = Variable ? VariableShadowExtend(IRB, S2)
2176 : Lower64ShadowExtend(IRB, S2, getShadowTy(&I));
2177 Value *V1 = I.getOperand(0);
2178 Value *V2 = I.getOperand(1);
2179 Value *Shift = IRB.CreateCall(I.getCalledValue(),
2180 {IRB.CreateBitCast(S1, V1->getType()), V2});
2181 Shift = IRB.CreateBitCast(Shift, getShadowTy(&I));
2182 setShadow(&I, IRB.CreateOr(Shift, S2Conv));
2183 setOriginForNaryOp(I);
2186 // \brief Get an X86_MMX-sized vector type.
2187 Type *getMMXVectorTy(unsigned EltSizeInBits) {
2188 const unsigned X86_MMXSizeInBits = 64;
2189 return VectorType::get(IntegerType::get(*MS.C, EltSizeInBits),
2190 X86_MMXSizeInBits / EltSizeInBits);
2193 // \brief Returns a signed counterpart for an (un)signed-saturate-and-pack
2195 Intrinsic::ID getSignedPackIntrinsic(Intrinsic::ID id) {
2197 case llvm::Intrinsic::x86_sse2_packsswb_128:
2198 case llvm::Intrinsic::x86_sse2_packuswb_128:
2199 return llvm::Intrinsic::x86_sse2_packsswb_128;
2201 case llvm::Intrinsic::x86_sse2_packssdw_128:
2202 case llvm::Intrinsic::x86_sse41_packusdw:
2203 return llvm::Intrinsic::x86_sse2_packssdw_128;
2205 case llvm::Intrinsic::x86_avx2_packsswb:
2206 case llvm::Intrinsic::x86_avx2_packuswb:
2207 return llvm::Intrinsic::x86_avx2_packsswb;
2209 case llvm::Intrinsic::x86_avx2_packssdw:
2210 case llvm::Intrinsic::x86_avx2_packusdw:
2211 return llvm::Intrinsic::x86_avx2_packssdw;
2213 case llvm::Intrinsic::x86_mmx_packsswb:
2214 case llvm::Intrinsic::x86_mmx_packuswb:
2215 return llvm::Intrinsic::x86_mmx_packsswb;
2217 case llvm::Intrinsic::x86_mmx_packssdw:
2218 return llvm::Intrinsic::x86_mmx_packssdw;
2220 llvm_unreachable("unexpected intrinsic id");
2224 // \brief Instrument vector pack instrinsic.
2226 // This function instruments intrinsics like x86_mmx_packsswb, that
2227 // packs elements of 2 input vectors into half as many bits with saturation.
2228 // Shadow is propagated with the signed variant of the same intrinsic applied
2229 // to sext(Sa != zeroinitializer), sext(Sb != zeroinitializer).
2230 // EltSizeInBits is used only for x86mmx arguments.
2231 void handleVectorPackIntrinsic(IntrinsicInst &I, unsigned EltSizeInBits = 0) {
2232 assert(I.getNumArgOperands() == 2);
2233 bool isX86_MMX = I.getOperand(0)->getType()->isX86_MMXTy();
2234 IRBuilder<> IRB(&I);
2235 Value *S1 = getShadow(&I, 0);
2236 Value *S2 = getShadow(&I, 1);
2237 assert(isX86_MMX || S1->getType()->isVectorTy());
2239 // SExt and ICmpNE below must apply to individual elements of input vectors.
2240 // In case of x86mmx arguments, cast them to appropriate vector types and
2242 Type *T = isX86_MMX ? getMMXVectorTy(EltSizeInBits) : S1->getType();
2244 S1 = IRB.CreateBitCast(S1, T);
2245 S2 = IRB.CreateBitCast(S2, T);
2247 Value *S1_ext = IRB.CreateSExt(
2248 IRB.CreateICmpNE(S1, llvm::Constant::getNullValue(T)), T);
2249 Value *S2_ext = IRB.CreateSExt(
2250 IRB.CreateICmpNE(S2, llvm::Constant::getNullValue(T)), T);
2252 Type *X86_MMXTy = Type::getX86_MMXTy(*MS.C);
2253 S1_ext = IRB.CreateBitCast(S1_ext, X86_MMXTy);
2254 S2_ext = IRB.CreateBitCast(S2_ext, X86_MMXTy);
2257 Function *ShadowFn = Intrinsic::getDeclaration(
2258 F.getParent(), getSignedPackIntrinsic(I.getIntrinsicID()));
2261 IRB.CreateCall(ShadowFn, {S1_ext, S2_ext}, "_msprop_vector_pack");
2262 if (isX86_MMX) S = IRB.CreateBitCast(S, getShadowTy(&I));
2264 setOriginForNaryOp(I);
2267 // \brief Instrument sum-of-absolute-differencies intrinsic.
2268 void handleVectorSadIntrinsic(IntrinsicInst &I) {
2269 const unsigned SignificantBitsPerResultElement = 16;
2270 bool isX86_MMX = I.getOperand(0)->getType()->isX86_MMXTy();
2271 Type *ResTy = isX86_MMX ? IntegerType::get(*MS.C, 64) : I.getType();
2272 unsigned ZeroBitsPerResultElement =
2273 ResTy->getScalarSizeInBits() - SignificantBitsPerResultElement;
2275 IRBuilder<> IRB(&I);
2276 Value *S = IRB.CreateOr(getShadow(&I, 0), getShadow(&I, 1));
2277 S = IRB.CreateBitCast(S, ResTy);
2278 S = IRB.CreateSExt(IRB.CreateICmpNE(S, Constant::getNullValue(ResTy)),
2280 S = IRB.CreateLShr(S, ZeroBitsPerResultElement);
2281 S = IRB.CreateBitCast(S, getShadowTy(&I));
2283 setOriginForNaryOp(I);
2286 // \brief Instrument multiply-add intrinsic.
2287 void handleVectorPmaddIntrinsic(IntrinsicInst &I,
2288 unsigned EltSizeInBits = 0) {
2289 bool isX86_MMX = I.getOperand(0)->getType()->isX86_MMXTy();
2290 Type *ResTy = isX86_MMX ? getMMXVectorTy(EltSizeInBits * 2) : I.getType();
2291 IRBuilder<> IRB(&I);
2292 Value *S = IRB.CreateOr(getShadow(&I, 0), getShadow(&I, 1));
2293 S = IRB.CreateBitCast(S, ResTy);
2294 S = IRB.CreateSExt(IRB.CreateICmpNE(S, Constant::getNullValue(ResTy)),
2296 S = IRB.CreateBitCast(S, getShadowTy(&I));
2298 setOriginForNaryOp(I);
2301 void visitIntrinsicInst(IntrinsicInst &I) {
2302 switch (I.getIntrinsicID()) {
2303 case llvm::Intrinsic::bswap:
2306 case llvm::Intrinsic::x86_avx512_cvtsd2usi64:
2307 case llvm::Intrinsic::x86_avx512_cvtsd2usi:
2308 case llvm::Intrinsic::x86_avx512_cvtss2usi64:
2309 case llvm::Intrinsic::x86_avx512_cvtss2usi:
2310 case llvm::Intrinsic::x86_avx512_cvttss2usi64:
2311 case llvm::Intrinsic::x86_avx512_cvttss2usi:
2312 case llvm::Intrinsic::x86_avx512_cvttsd2usi64:
2313 case llvm::Intrinsic::x86_avx512_cvttsd2usi:
2314 case llvm::Intrinsic::x86_avx512_cvtusi2sd:
2315 case llvm::Intrinsic::x86_avx512_cvtusi2ss:
2316 case llvm::Intrinsic::x86_avx512_cvtusi642sd:
2317 case llvm::Intrinsic::x86_avx512_cvtusi642ss:
2318 case llvm::Intrinsic::x86_sse2_cvtsd2si64:
2319 case llvm::Intrinsic::x86_sse2_cvtsd2si:
2320 case llvm::Intrinsic::x86_sse2_cvtsd2ss:
2321 case llvm::Intrinsic::x86_sse2_cvtsi2sd:
2322 case llvm::Intrinsic::x86_sse2_cvtsi642sd:
2323 case llvm::Intrinsic::x86_sse2_cvtss2sd:
2324 case llvm::Intrinsic::x86_sse2_cvttsd2si64:
2325 case llvm::Intrinsic::x86_sse2_cvttsd2si:
2326 case llvm::Intrinsic::x86_sse_cvtsi2ss:
2327 case llvm::Intrinsic::x86_sse_cvtsi642ss:
2328 case llvm::Intrinsic::x86_sse_cvtss2si64:
2329 case llvm::Intrinsic::x86_sse_cvtss2si:
2330 case llvm::Intrinsic::x86_sse_cvttss2si64:
2331 case llvm::Intrinsic::x86_sse_cvttss2si:
2332 handleVectorConvertIntrinsic(I, 1);
2334 case llvm::Intrinsic::x86_sse2_cvtdq2pd:
2335 case llvm::Intrinsic::x86_sse2_cvtps2pd:
2336 case llvm::Intrinsic::x86_sse_cvtps2pi:
2337 case llvm::Intrinsic::x86_sse_cvttps2pi:
2338 handleVectorConvertIntrinsic(I, 2);
2340 case llvm::Intrinsic::x86_avx2_psll_w:
2341 case llvm::Intrinsic::x86_avx2_psll_d:
2342 case llvm::Intrinsic::x86_avx2_psll_q:
2343 case llvm::Intrinsic::x86_avx2_pslli_w:
2344 case llvm::Intrinsic::x86_avx2_pslli_d:
2345 case llvm::Intrinsic::x86_avx2_pslli_q:
2346 case llvm::Intrinsic::x86_avx2_psrl_w:
2347 case llvm::Intrinsic::x86_avx2_psrl_d:
2348 case llvm::Intrinsic::x86_avx2_psrl_q:
2349 case llvm::Intrinsic::x86_avx2_psra_w:
2350 case llvm::Intrinsic::x86_avx2_psra_d:
2351 case llvm::Intrinsic::x86_avx2_psrli_w:
2352 case llvm::Intrinsic::x86_avx2_psrli_d:
2353 case llvm::Intrinsic::x86_avx2_psrli_q:
2354 case llvm::Intrinsic::x86_avx2_psrai_w:
2355 case llvm::Intrinsic::x86_avx2_psrai_d:
2356 case llvm::Intrinsic::x86_sse2_psll_w:
2357 case llvm::Intrinsic::x86_sse2_psll_d:
2358 case llvm::Intrinsic::x86_sse2_psll_q:
2359 case llvm::Intrinsic::x86_sse2_pslli_w:
2360 case llvm::Intrinsic::x86_sse2_pslli_d:
2361 case llvm::Intrinsic::x86_sse2_pslli_q:
2362 case llvm::Intrinsic::x86_sse2_psrl_w:
2363 case llvm::Intrinsic::x86_sse2_psrl_d:
2364 case llvm::Intrinsic::x86_sse2_psrl_q:
2365 case llvm::Intrinsic::x86_sse2_psra_w:
2366 case llvm::Intrinsic::x86_sse2_psra_d:
2367 case llvm::Intrinsic::x86_sse2_psrli_w:
2368 case llvm::Intrinsic::x86_sse2_psrli_d:
2369 case llvm::Intrinsic::x86_sse2_psrli_q:
2370 case llvm::Intrinsic::x86_sse2_psrai_w:
2371 case llvm::Intrinsic::x86_sse2_psrai_d:
2372 case llvm::Intrinsic::x86_mmx_psll_w:
2373 case llvm::Intrinsic::x86_mmx_psll_d:
2374 case llvm::Intrinsic::x86_mmx_psll_q:
2375 case llvm::Intrinsic::x86_mmx_pslli_w:
2376 case llvm::Intrinsic::x86_mmx_pslli_d:
2377 case llvm::Intrinsic::x86_mmx_pslli_q:
2378 case llvm::Intrinsic::x86_mmx_psrl_w:
2379 case llvm::Intrinsic::x86_mmx_psrl_d:
2380 case llvm::Intrinsic::x86_mmx_psrl_q:
2381 case llvm::Intrinsic::x86_mmx_psra_w:
2382 case llvm::Intrinsic::x86_mmx_psra_d:
2383 case llvm::Intrinsic::x86_mmx_psrli_w:
2384 case llvm::Intrinsic::x86_mmx_psrli_d:
2385 case llvm::Intrinsic::x86_mmx_psrli_q:
2386 case llvm::Intrinsic::x86_mmx_psrai_w:
2387 case llvm::Intrinsic::x86_mmx_psrai_d:
2388 handleVectorShiftIntrinsic(I, /* Variable */ false);
2390 case llvm::Intrinsic::x86_avx2_psllv_d:
2391 case llvm::Intrinsic::x86_avx2_psllv_d_256:
2392 case llvm::Intrinsic::x86_avx2_psllv_q:
2393 case llvm::Intrinsic::x86_avx2_psllv_q_256:
2394 case llvm::Intrinsic::x86_avx2_psrlv_d:
2395 case llvm::Intrinsic::x86_avx2_psrlv_d_256:
2396 case llvm::Intrinsic::x86_avx2_psrlv_q:
2397 case llvm::Intrinsic::x86_avx2_psrlv_q_256:
2398 case llvm::Intrinsic::x86_avx2_psrav_d:
2399 case llvm::Intrinsic::x86_avx2_psrav_d_256:
2400 handleVectorShiftIntrinsic(I, /* Variable */ true);
2403 case llvm::Intrinsic::x86_sse2_packsswb_128:
2404 case llvm::Intrinsic::x86_sse2_packssdw_128:
2405 case llvm::Intrinsic::x86_sse2_packuswb_128:
2406 case llvm::Intrinsic::x86_sse41_packusdw:
2407 case llvm::Intrinsic::x86_avx2_packsswb:
2408 case llvm::Intrinsic::x86_avx2_packssdw:
2409 case llvm::Intrinsic::x86_avx2_packuswb:
2410 case llvm::Intrinsic::x86_avx2_packusdw:
2411 handleVectorPackIntrinsic(I);
2414 case llvm::Intrinsic::x86_mmx_packsswb:
2415 case llvm::Intrinsic::x86_mmx_packuswb:
2416 handleVectorPackIntrinsic(I, 16);
2419 case llvm::Intrinsic::x86_mmx_packssdw:
2420 handleVectorPackIntrinsic(I, 32);
2423 case llvm::Intrinsic::x86_mmx_psad_bw:
2424 case llvm::Intrinsic::x86_sse2_psad_bw:
2425 case llvm::Intrinsic::x86_avx2_psad_bw:
2426 handleVectorSadIntrinsic(I);
2429 case llvm::Intrinsic::x86_sse2_pmadd_wd:
2430 case llvm::Intrinsic::x86_avx2_pmadd_wd:
2431 case llvm::Intrinsic::x86_ssse3_pmadd_ub_sw_128:
2432 case llvm::Intrinsic::x86_avx2_pmadd_ub_sw:
2433 handleVectorPmaddIntrinsic(I);
2436 case llvm::Intrinsic::x86_ssse3_pmadd_ub_sw:
2437 handleVectorPmaddIntrinsic(I, 8);
2440 case llvm::Intrinsic::x86_mmx_pmadd_wd:
2441 handleVectorPmaddIntrinsic(I, 16);
2445 if (!handleUnknownIntrinsic(I))
2446 visitInstruction(I);
2451 void visitCallSite(CallSite CS) {
2452 Instruction &I = *CS.getInstruction();
2453 assert((CS.isCall() || CS.isInvoke()) && "Unknown type of CallSite");
2455 CallInst *Call = cast<CallInst>(&I);
2457 // For inline asm, do the usual thing: check argument shadow and mark all
2458 // outputs as clean. Note that any side effects of the inline asm that are
2459 // not immediately visible in its constraints are not handled.
2460 if (Call->isInlineAsm()) {
2461 visitInstruction(I);
2465 assert(!isa<IntrinsicInst>(&I) && "intrinsics are handled elsewhere");
2467 // We are going to insert code that relies on the fact that the callee
2468 // will become a non-readonly function after it is instrumented by us. To
2469 // prevent this code from being optimized out, mark that function
2470 // non-readonly in advance.
2471 if (Function *Func = Call->getCalledFunction()) {
2472 // Clear out readonly/readnone attributes.
2474 B.addAttribute(Attribute::ReadOnly)
2475 .addAttribute(Attribute::ReadNone);
2476 Func->removeAttributes(AttributeSet::FunctionIndex,
2477 AttributeSet::get(Func->getContext(),
2478 AttributeSet::FunctionIndex,
2482 IRBuilder<> IRB(&I);
2484 unsigned ArgOffset = 0;
2485 DEBUG(dbgs() << " CallSite: " << I << "\n");
2486 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
2487 ArgIt != End; ++ArgIt) {
2489 unsigned i = ArgIt - CS.arg_begin();
2490 if (!A->getType()->isSized()) {
2491 DEBUG(dbgs() << "Arg " << i << " is not sized: " << I << "\n");
2495 Value *Store = nullptr;
2496 // Compute the Shadow for arg even if it is ByVal, because
2497 // in that case getShadow() will copy the actual arg shadow to
2498 // __msan_param_tls.
2499 Value *ArgShadow = getShadow(A);
2500 Value *ArgShadowBase = getShadowPtrForArgument(A, IRB, ArgOffset);
2501 DEBUG(dbgs() << " Arg#" << i << ": " << *A <<
2502 " Shadow: " << *ArgShadow << "\n");
2503 bool ArgIsInitialized = false;
2504 const DataLayout &DL = F.getParent()->getDataLayout();
2505 if (CS.paramHasAttr(i + 1, Attribute::ByVal)) {
2506 assert(A->getType()->isPointerTy() &&
2507 "ByVal argument is not a pointer!");
2508 Size = DL.getTypeAllocSize(A->getType()->getPointerElementType());
2509 if (ArgOffset + Size > kParamTLSSize) break;
2510 unsigned ParamAlignment = CS.getParamAlignment(i + 1);
2511 unsigned Alignment = std::min(ParamAlignment, kShadowTLSAlignment);
2512 Store = IRB.CreateMemCpy(ArgShadowBase,
2513 getShadowPtr(A, Type::getInt8Ty(*MS.C), IRB),
2516 Size = DL.getTypeAllocSize(A->getType());
2517 if (ArgOffset + Size > kParamTLSSize) break;
2518 Store = IRB.CreateAlignedStore(ArgShadow, ArgShadowBase,
2519 kShadowTLSAlignment);
2520 Constant *Cst = dyn_cast<Constant>(ArgShadow);
2521 if (Cst && Cst->isNullValue()) ArgIsInitialized = true;
2523 if (MS.TrackOrigins && !ArgIsInitialized)
2524 IRB.CreateStore(getOrigin(A),
2525 getOriginPtrForArgument(A, IRB, ArgOffset));
2527 assert(Size != 0 && Store != nullptr);
2528 DEBUG(dbgs() << " Param:" << *Store << "\n");
2529 ArgOffset += RoundUpToAlignment(Size, 8);
2531 DEBUG(dbgs() << " done with call args\n");
2534 cast<FunctionType>(CS.getCalledValue()->getType()->getContainedType(0));
2535 if (FT->isVarArg()) {
2536 VAHelper->visitCallSite(CS, IRB);
2539 // Now, get the shadow for the RetVal.
2540 if (!I.getType()->isSized()) return;
2541 // Don't emit the epilogue for musttail call returns.
2542 if (CS.isCall() && cast<CallInst>(&I)->isMustTailCall()) return;
2543 IRBuilder<> IRBBefore(&I);
2544 // Until we have full dynamic coverage, make sure the retval shadow is 0.
2545 Value *Base = getShadowPtrForRetval(&I, IRBBefore);
2546 IRBBefore.CreateAlignedStore(getCleanShadow(&I), Base, kShadowTLSAlignment);
2547 Instruction *NextInsn = nullptr;
2549 NextInsn = I.getNextNode();
2551 BasicBlock *NormalDest = cast<InvokeInst>(&I)->getNormalDest();
2552 if (!NormalDest->getSinglePredecessor()) {
2553 // FIXME: this case is tricky, so we are just conservative here.
2554 // Perhaps we need to split the edge between this BB and NormalDest,
2555 // but a naive attempt to use SplitEdge leads to a crash.
2556 setShadow(&I, getCleanShadow(&I));
2557 setOrigin(&I, getCleanOrigin());
2560 NextInsn = NormalDest->getFirstInsertionPt();
2562 "Could not find insertion point for retval shadow load");
2564 IRBuilder<> IRBAfter(NextInsn);
2565 Value *RetvalShadow =
2566 IRBAfter.CreateAlignedLoad(getShadowPtrForRetval(&I, IRBAfter),
2567 kShadowTLSAlignment, "_msret");
2568 setShadow(&I, RetvalShadow);
2569 if (MS.TrackOrigins)
2570 setOrigin(&I, IRBAfter.CreateLoad(getOriginPtrForRetval(IRBAfter)));
2573 bool isAMustTailRetVal(Value *RetVal) {
2574 if (auto *I = dyn_cast<BitCastInst>(RetVal)) {
2575 RetVal = I->getOperand(0);
2577 if (auto *I = dyn_cast<CallInst>(RetVal)) {
2578 return I->isMustTailCall();
2583 void visitReturnInst(ReturnInst &I) {
2584 IRBuilder<> IRB(&I);
2585 Value *RetVal = I.getReturnValue();
2586 if (!RetVal) return;
2587 // Don't emit the epilogue for musttail call returns.
2588 if (isAMustTailRetVal(RetVal)) return;
2589 Value *ShadowPtr = getShadowPtrForRetval(RetVal, IRB);
2590 if (CheckReturnValue) {
2591 insertShadowCheck(RetVal, &I);
2592 Value *Shadow = getCleanShadow(RetVal);
2593 IRB.CreateAlignedStore(Shadow, ShadowPtr, kShadowTLSAlignment);
2595 Value *Shadow = getShadow(RetVal);
2596 IRB.CreateAlignedStore(Shadow, ShadowPtr, kShadowTLSAlignment);
2597 // FIXME: make it conditional if ClStoreCleanOrigin==0
2598 if (MS.TrackOrigins)
2599 IRB.CreateStore(getOrigin(RetVal), getOriginPtrForRetval(IRB));
2603 void visitPHINode(PHINode &I) {
2604 IRBuilder<> IRB(&I);
2605 if (!PropagateShadow) {
2606 setShadow(&I, getCleanShadow(&I));
2607 setOrigin(&I, getCleanOrigin());
2611 ShadowPHINodes.push_back(&I);
2612 setShadow(&I, IRB.CreatePHI(getShadowTy(&I), I.getNumIncomingValues(),
2614 if (MS.TrackOrigins)
2615 setOrigin(&I, IRB.CreatePHI(MS.OriginTy, I.getNumIncomingValues(),
2619 void visitAllocaInst(AllocaInst &I) {
2620 setShadow(&I, getCleanShadow(&I));
2621 setOrigin(&I, getCleanOrigin());
2622 IRBuilder<> IRB(I.getNextNode());
2623 const DataLayout &DL = F.getParent()->getDataLayout();
2624 uint64_t Size = DL.getTypeAllocSize(I.getAllocatedType());
2625 if (PoisonStack && ClPoisonStackWithCall) {
2626 IRB.CreateCall(MS.MsanPoisonStackFn,
2627 {IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()),
2628 ConstantInt::get(MS.IntptrTy, Size)});
2630 Value *ShadowBase = getShadowPtr(&I, Type::getInt8PtrTy(*MS.C), IRB);
2631 Value *PoisonValue = IRB.getInt8(PoisonStack ? ClPoisonStackPattern : 0);
2632 IRB.CreateMemSet(ShadowBase, PoisonValue, Size, I.getAlignment());
2635 if (PoisonStack && MS.TrackOrigins) {
2636 SmallString<2048> StackDescriptionStorage;
2637 raw_svector_ostream StackDescription(StackDescriptionStorage);
2638 // We create a string with a description of the stack allocation and
2639 // pass it into __msan_set_alloca_origin.
2640 // It will be printed by the run-time if stack-originated UMR is found.
2641 // The first 4 bytes of the string are set to '----' and will be replaced
2642 // by __msan_va_arg_overflow_size_tls at the first call.
2643 StackDescription << "----" << I.getName() << "@" << F.getName();
2645 createPrivateNonConstGlobalForString(*F.getParent(),
2646 StackDescription.str());
2648 IRB.CreateCall(MS.MsanSetAllocaOrigin4Fn,
2649 {IRB.CreatePointerCast(&I, IRB.getInt8PtrTy()),
2650 ConstantInt::get(MS.IntptrTy, Size),
2651 IRB.CreatePointerCast(Descr, IRB.getInt8PtrTy()),
2652 IRB.CreatePointerCast(&F, MS.IntptrTy)});
2656 void visitSelectInst(SelectInst& I) {
2657 IRBuilder<> IRB(&I);
2658 // a = select b, c, d
2659 Value *B = I.getCondition();
2660 Value *C = I.getTrueValue();
2661 Value *D = I.getFalseValue();
2662 Value *Sb = getShadow(B);
2663 Value *Sc = getShadow(C);
2664 Value *Sd = getShadow(D);
2666 // Result shadow if condition shadow is 0.
2667 Value *Sa0 = IRB.CreateSelect(B, Sc, Sd);
2669 if (I.getType()->isAggregateType()) {
2670 // To avoid "sign extending" i1 to an arbitrary aggregate type, we just do
2671 // an extra "select". This results in much more compact IR.
2672 // Sa = select Sb, poisoned, (select b, Sc, Sd)
2673 Sa1 = getPoisonedShadow(getShadowTy(I.getType()));
2675 // Sa = select Sb, [ (c^d) | Sc | Sd ], [ b ? Sc : Sd ]
2676 // If Sb (condition is poisoned), look for bits in c and d that are equal
2677 // and both unpoisoned.
2678 // If !Sb (condition is unpoisoned), simply pick one of Sc and Sd.
2680 // Cast arguments to shadow-compatible type.
2681 C = CreateAppToShadowCast(IRB, C);
2682 D = CreateAppToShadowCast(IRB, D);
2684 // Result shadow if condition shadow is 1.
2685 Sa1 = IRB.CreateOr(IRB.CreateXor(C, D), IRB.CreateOr(Sc, Sd));
2687 Value *Sa = IRB.CreateSelect(Sb, Sa1, Sa0, "_msprop_select");
2689 if (MS.TrackOrigins) {
2690 // Origins are always i32, so any vector conditions must be flattened.
2691 // FIXME: consider tracking vector origins for app vectors?
2692 if (B->getType()->isVectorTy()) {
2693 Type *FlatTy = getShadowTyNoVec(B->getType());
2694 B = IRB.CreateICmpNE(IRB.CreateBitCast(B, FlatTy),
2695 ConstantInt::getNullValue(FlatTy));
2696 Sb = IRB.CreateICmpNE(IRB.CreateBitCast(Sb, FlatTy),
2697 ConstantInt::getNullValue(FlatTy));
2699 // a = select b, c, d
2700 // Oa = Sb ? Ob : (b ? Oc : Od)
2702 &I, IRB.CreateSelect(Sb, getOrigin(I.getCondition()),
2703 IRB.CreateSelect(B, getOrigin(I.getTrueValue()),
2704 getOrigin(I.getFalseValue()))));
2708 void visitLandingPadInst(LandingPadInst &I) {
2710 // See http://code.google.com/p/memory-sanitizer/issues/detail?id=1
2711 setShadow(&I, getCleanShadow(&I));
2712 setOrigin(&I, getCleanOrigin());
2715 void visitCleanupPadInst(CleanupPadInst &I) {
2716 setShadow(&I, getCleanShadow(&I));
2717 setOrigin(&I, getCleanOrigin());
2720 void visitCatchPad(CatchPadInst &I) {
2721 setShadow(&I, getCleanShadow(&I));
2722 setOrigin(&I, getCleanOrigin());
2725 void visitTerminatePad(TerminatePadInst &I) {
2726 DEBUG(dbgs() << "TerminatePad: " << I << "\n");
2727 // Nothing to do here.
2730 void visitCatchEndPadInst(CatchEndPadInst &I) {
2731 DEBUG(dbgs() << "CatchEndPad: " << I << "\n");
2732 // Nothing to do here.
2735 void visitCleanupEndPadInst(CleanupEndPadInst &I) {
2736 DEBUG(dbgs() << "CleanupEndPad: " << I << "\n");
2737 // Nothing to do here.
2740 void visitGetElementPtrInst(GetElementPtrInst &I) {
2744 void visitExtractValueInst(ExtractValueInst &I) {
2745 IRBuilder<> IRB(&I);
2746 Value *Agg = I.getAggregateOperand();
2747 DEBUG(dbgs() << "ExtractValue: " << I << "\n");
2748 Value *AggShadow = getShadow(Agg);
2749 DEBUG(dbgs() << " AggShadow: " << *AggShadow << "\n");
2750 Value *ResShadow = IRB.CreateExtractValue(AggShadow, I.getIndices());
2751 DEBUG(dbgs() << " ResShadow: " << *ResShadow << "\n");
2752 setShadow(&I, ResShadow);
2753 setOriginForNaryOp(I);
2756 void visitInsertValueInst(InsertValueInst &I) {
2757 IRBuilder<> IRB(&I);
2758 DEBUG(dbgs() << "InsertValue: " << I << "\n");
2759 Value *AggShadow = getShadow(I.getAggregateOperand());
2760 Value *InsShadow = getShadow(I.getInsertedValueOperand());
2761 DEBUG(dbgs() << " AggShadow: " << *AggShadow << "\n");
2762 DEBUG(dbgs() << " InsShadow: " << *InsShadow << "\n");
2763 Value *Res = IRB.CreateInsertValue(AggShadow, InsShadow, I.getIndices());
2764 DEBUG(dbgs() << " Res: " << *Res << "\n");
2766 setOriginForNaryOp(I);
2769 void dumpInst(Instruction &I) {
2770 if (CallInst *CI = dyn_cast<CallInst>(&I)) {
2771 errs() << "ZZZ call " << CI->getCalledFunction()->getName() << "\n";
2773 errs() << "ZZZ " << I.getOpcodeName() << "\n";
2775 errs() << "QQQ " << I << "\n";
2778 void visitResumeInst(ResumeInst &I) {
2779 DEBUG(dbgs() << "Resume: " << I << "\n");
2780 // Nothing to do here.
2783 void visitCleanupReturnInst(CleanupReturnInst &CRI) {
2784 DEBUG(dbgs() << "CleanupReturn: " << CRI << "\n");
2785 // Nothing to do here.
2788 void visitCatchReturnInst(CatchReturnInst &CRI) {
2789 DEBUG(dbgs() << "CatchReturn: " << CRI << "\n");
2790 // Nothing to do here.
2793 void visitInstruction(Instruction &I) {
2794 // Everything else: stop propagating and check for poisoned shadow.
2795 if (ClDumpStrictInstructions)
2797 DEBUG(dbgs() << "DEFAULT: " << I << "\n");
2798 for (size_t i = 0, n = I.getNumOperands(); i < n; i++)
2799 insertShadowCheck(I.getOperand(i), &I);
2800 setShadow(&I, getCleanShadow(&I));
2801 setOrigin(&I, getCleanOrigin());
2805 /// \brief AMD64-specific implementation of VarArgHelper.
2806 struct VarArgAMD64Helper : public VarArgHelper {
2807 // An unfortunate workaround for asymmetric lowering of va_arg stuff.
2808 // See a comment in visitCallSite for more details.
2809 static const unsigned AMD64GpEndOffset = 48; // AMD64 ABI Draft 0.99.6 p3.5.7
2810 static const unsigned AMD64FpEndOffset = 176;
2813 MemorySanitizer &MS;
2814 MemorySanitizerVisitor &MSV;
2815 Value *VAArgTLSCopy;
2816 Value *VAArgOverflowSize;
2818 SmallVector<CallInst*, 16> VAStartInstrumentationList;
2820 VarArgAMD64Helper(Function &F, MemorySanitizer &MS,
2821 MemorySanitizerVisitor &MSV)
2822 : F(F), MS(MS), MSV(MSV), VAArgTLSCopy(nullptr),
2823 VAArgOverflowSize(nullptr) {}
2825 enum ArgKind { AK_GeneralPurpose, AK_FloatingPoint, AK_Memory };
2827 ArgKind classifyArgument(Value* arg) {
2828 // A very rough approximation of X86_64 argument classification rules.
2829 Type *T = arg->getType();
2830 if (T->isFPOrFPVectorTy() || T->isX86_MMXTy())
2831 return AK_FloatingPoint;
2832 if (T->isIntegerTy() && T->getPrimitiveSizeInBits() <= 64)
2833 return AK_GeneralPurpose;
2834 if (T->isPointerTy())
2835 return AK_GeneralPurpose;
2839 // For VarArg functions, store the argument shadow in an ABI-specific format
2840 // that corresponds to va_list layout.
2841 // We do this because Clang lowers va_arg in the frontend, and this pass
2842 // only sees the low level code that deals with va_list internals.
2843 // A much easier alternative (provided that Clang emits va_arg instructions)
2844 // would have been to associate each live instance of va_list with a copy of
2845 // MSanParamTLS, and extract shadow on va_arg() call in the argument list
2847 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) override {
2848 unsigned GpOffset = 0;
2849 unsigned FpOffset = AMD64GpEndOffset;
2850 unsigned OverflowOffset = AMD64FpEndOffset;
2851 const DataLayout &DL = F.getParent()->getDataLayout();
2852 for (CallSite::arg_iterator ArgIt = CS.arg_begin(), End = CS.arg_end();
2853 ArgIt != End; ++ArgIt) {
2855 unsigned ArgNo = CS.getArgumentNo(ArgIt);
2856 bool IsByVal = CS.paramHasAttr(ArgNo + 1, Attribute::ByVal);
2858 // ByVal arguments always go to the overflow area.
2859 assert(A->getType()->isPointerTy());
2860 Type *RealTy = A->getType()->getPointerElementType();
2861 uint64_t ArgSize = DL.getTypeAllocSize(RealTy);
2862 Value *Base = getShadowPtrForVAArgument(RealTy, IRB, OverflowOffset);
2863 OverflowOffset += RoundUpToAlignment(ArgSize, 8);
2864 IRB.CreateMemCpy(Base, MSV.getShadowPtr(A, IRB.getInt8Ty(), IRB),
2865 ArgSize, kShadowTLSAlignment);
2867 ArgKind AK = classifyArgument(A);
2868 if (AK == AK_GeneralPurpose && GpOffset >= AMD64GpEndOffset)
2870 if (AK == AK_FloatingPoint && FpOffset >= AMD64FpEndOffset)
2874 case AK_GeneralPurpose:
2875 Base = getShadowPtrForVAArgument(A->getType(), IRB, GpOffset);
2878 case AK_FloatingPoint:
2879 Base = getShadowPtrForVAArgument(A->getType(), IRB, FpOffset);
2883 uint64_t ArgSize = DL.getTypeAllocSize(A->getType());
2884 Base = getShadowPtrForVAArgument(A->getType(), IRB, OverflowOffset);
2885 OverflowOffset += RoundUpToAlignment(ArgSize, 8);
2887 IRB.CreateAlignedStore(MSV.getShadow(A), Base, kShadowTLSAlignment);
2890 Constant *OverflowSize =
2891 ConstantInt::get(IRB.getInt64Ty(), OverflowOffset - AMD64FpEndOffset);
2892 IRB.CreateStore(OverflowSize, MS.VAArgOverflowSizeTLS);
2895 /// \brief Compute the shadow address for a given va_arg.
2896 Value *getShadowPtrForVAArgument(Type *Ty, IRBuilder<> &IRB,
2898 Value *Base = IRB.CreatePointerCast(MS.VAArgTLS, MS.IntptrTy);
2899 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
2900 return IRB.CreateIntToPtr(Base, PointerType::get(MSV.getShadowTy(Ty), 0),
2904 void visitVAStartInst(VAStartInst &I) override {
2905 if (F.getCallingConv() == CallingConv::X86_64_Win64)
2907 IRBuilder<> IRB(&I);
2908 VAStartInstrumentationList.push_back(&I);
2909 Value *VAListTag = I.getArgOperand(0);
2910 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
2912 // Unpoison the whole __va_list_tag.
2913 // FIXME: magic ABI constants.
2914 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
2915 /* size */24, /* alignment */8, false);
2918 void visitVACopyInst(VACopyInst &I) override {
2919 if (F.getCallingConv() == CallingConv::X86_64_Win64)
2921 IRBuilder<> IRB(&I);
2922 Value *VAListTag = I.getArgOperand(0);
2923 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
2925 // Unpoison the whole __va_list_tag.
2926 // FIXME: magic ABI constants.
2927 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
2928 /* size */24, /* alignment */8, false);
2931 void finalizeInstrumentation() override {
2932 assert(!VAArgOverflowSize && !VAArgTLSCopy &&
2933 "finalizeInstrumentation called twice");
2934 if (!VAStartInstrumentationList.empty()) {
2935 // If there is a va_start in this function, make a backup copy of
2936 // va_arg_tls somewhere in the function entry block.
2937 IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
2938 VAArgOverflowSize = IRB.CreateLoad(MS.VAArgOverflowSizeTLS);
2940 IRB.CreateAdd(ConstantInt::get(MS.IntptrTy, AMD64FpEndOffset),
2942 VAArgTLSCopy = IRB.CreateAlloca(Type::getInt8Ty(*MS.C), CopySize);
2943 IRB.CreateMemCpy(VAArgTLSCopy, MS.VAArgTLS, CopySize, 8);
2946 // Instrument va_start.
2947 // Copy va_list shadow from the backup copy of the TLS contents.
2948 for (size_t i = 0, n = VAStartInstrumentationList.size(); i < n; i++) {
2949 CallInst *OrigInst = VAStartInstrumentationList[i];
2950 IRBuilder<> IRB(OrigInst->getNextNode());
2951 Value *VAListTag = OrigInst->getArgOperand(0);
2953 Value *RegSaveAreaPtrPtr =
2955 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
2956 ConstantInt::get(MS.IntptrTy, 16)),
2957 Type::getInt64PtrTy(*MS.C));
2958 Value *RegSaveAreaPtr = IRB.CreateLoad(RegSaveAreaPtrPtr);
2959 Value *RegSaveAreaShadowPtr =
2960 MSV.getShadowPtr(RegSaveAreaPtr, IRB.getInt8Ty(), IRB);
2961 IRB.CreateMemCpy(RegSaveAreaShadowPtr, VAArgTLSCopy,
2962 AMD64FpEndOffset, 16);
2964 Value *OverflowArgAreaPtrPtr =
2966 IRB.CreateAdd(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
2967 ConstantInt::get(MS.IntptrTy, 8)),
2968 Type::getInt64PtrTy(*MS.C));
2969 Value *OverflowArgAreaPtr = IRB.CreateLoad(OverflowArgAreaPtrPtr);
2970 Value *OverflowArgAreaShadowPtr =
2971 MSV.getShadowPtr(OverflowArgAreaPtr, IRB.getInt8Ty(), IRB);
2972 Value *SrcPtr = IRB.CreateConstGEP1_32(IRB.getInt8Ty(), VAArgTLSCopy,
2974 IRB.CreateMemCpy(OverflowArgAreaShadowPtr, SrcPtr, VAArgOverflowSize, 16);
2979 /// \brief MIPS64-specific implementation of VarArgHelper.
2980 struct VarArgMIPS64Helper : public VarArgHelper {
2982 MemorySanitizer &MS;
2983 MemorySanitizerVisitor &MSV;
2984 Value *VAArgTLSCopy;
2987 SmallVector<CallInst*, 16> VAStartInstrumentationList;
2989 VarArgMIPS64Helper(Function &F, MemorySanitizer &MS,
2990 MemorySanitizerVisitor &MSV)
2991 : F(F), MS(MS), MSV(MSV), VAArgTLSCopy(nullptr),
2992 VAArgSize(nullptr) {}
2994 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) override {
2995 unsigned VAArgOffset = 0;
2996 const DataLayout &DL = F.getParent()->getDataLayout();
2997 for (CallSite::arg_iterator ArgIt = CS.arg_begin() + 1, End = CS.arg_end();
2998 ArgIt != End; ++ArgIt) {
3001 uint64_t ArgSize = DL.getTypeAllocSize(A->getType());
3002 #if defined(__MIPSEB__) || defined(MIPSEB)
3003 // Adjusting the shadow for argument with size < 8 to match the placement
3004 // of bits in big endian system
3006 VAArgOffset += (8 - ArgSize);
3008 Base = getShadowPtrForVAArgument(A->getType(), IRB, VAArgOffset);
3009 VAArgOffset += ArgSize;
3010 VAArgOffset = RoundUpToAlignment(VAArgOffset, 8);
3011 IRB.CreateAlignedStore(MSV.getShadow(A), Base, kShadowTLSAlignment);
3014 Constant *TotalVAArgSize = ConstantInt::get(IRB.getInt64Ty(), VAArgOffset);
3015 // Here using VAArgOverflowSizeTLS as VAArgSizeTLS to avoid creation of
3016 // a new class member i.e. it is the total size of all VarArgs.
3017 IRB.CreateStore(TotalVAArgSize, MS.VAArgOverflowSizeTLS);
3020 /// \brief Compute the shadow address for a given va_arg.
3021 Value *getShadowPtrForVAArgument(Type *Ty, IRBuilder<> &IRB,
3023 Value *Base = IRB.CreatePointerCast(MS.VAArgTLS, MS.IntptrTy);
3024 Base = IRB.CreateAdd(Base, ConstantInt::get(MS.IntptrTy, ArgOffset));
3025 return IRB.CreateIntToPtr(Base, PointerType::get(MSV.getShadowTy(Ty), 0),
3029 void visitVAStartInst(VAStartInst &I) override {
3030 IRBuilder<> IRB(&I);
3031 VAStartInstrumentationList.push_back(&I);
3032 Value *VAListTag = I.getArgOperand(0);
3033 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
3034 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
3035 /* size */8, /* alignment */8, false);
3038 void visitVACopyInst(VACopyInst &I) override {
3039 IRBuilder<> IRB(&I);
3040 Value *VAListTag = I.getArgOperand(0);
3041 Value *ShadowPtr = MSV.getShadowPtr(VAListTag, IRB.getInt8Ty(), IRB);
3042 // Unpoison the whole __va_list_tag.
3043 // FIXME: magic ABI constants.
3044 IRB.CreateMemSet(ShadowPtr, Constant::getNullValue(IRB.getInt8Ty()),
3045 /* size */8, /* alignment */8, false);
3048 void finalizeInstrumentation() override {
3049 assert(!VAArgSize && !VAArgTLSCopy &&
3050 "finalizeInstrumentation called twice");
3051 IRBuilder<> IRB(F.getEntryBlock().getFirstNonPHI());
3052 VAArgSize = IRB.CreateLoad(MS.VAArgOverflowSizeTLS);
3053 Value *CopySize = IRB.CreateAdd(ConstantInt::get(MS.IntptrTy, 0),
3056 if (!VAStartInstrumentationList.empty()) {
3057 // If there is a va_start in this function, make a backup copy of
3058 // va_arg_tls somewhere in the function entry block.
3059 VAArgTLSCopy = IRB.CreateAlloca(Type::getInt8Ty(*MS.C), CopySize);
3060 IRB.CreateMemCpy(VAArgTLSCopy, MS.VAArgTLS, CopySize, 8);
3063 // Instrument va_start.
3064 // Copy va_list shadow from the backup copy of the TLS contents.
3065 for (size_t i = 0, n = VAStartInstrumentationList.size(); i < n; i++) {
3066 CallInst *OrigInst = VAStartInstrumentationList[i];
3067 IRBuilder<> IRB(OrigInst->getNextNode());
3068 Value *VAListTag = OrigInst->getArgOperand(0);
3069 Value *RegSaveAreaPtrPtr =
3070 IRB.CreateIntToPtr(IRB.CreatePtrToInt(VAListTag, MS.IntptrTy),
3071 Type::getInt64PtrTy(*MS.C));
3072 Value *RegSaveAreaPtr = IRB.CreateLoad(RegSaveAreaPtrPtr);
3073 Value *RegSaveAreaShadowPtr =
3074 MSV.getShadowPtr(RegSaveAreaPtr, IRB.getInt8Ty(), IRB);
3075 IRB.CreateMemCpy(RegSaveAreaShadowPtr, VAArgTLSCopy, CopySize, 8);
3080 /// \brief A no-op implementation of VarArgHelper.
3081 struct VarArgNoOpHelper : public VarArgHelper {
3082 VarArgNoOpHelper(Function &F, MemorySanitizer &MS,
3083 MemorySanitizerVisitor &MSV) {}
3085 void visitCallSite(CallSite &CS, IRBuilder<> &IRB) override {}
3087 void visitVAStartInst(VAStartInst &I) override {}
3089 void visitVACopyInst(VACopyInst &I) override {}
3091 void finalizeInstrumentation() override {}
3094 VarArgHelper *CreateVarArgHelper(Function &Func, MemorySanitizer &Msan,
3095 MemorySanitizerVisitor &Visitor) {
3096 // VarArg handling is only implemented on AMD64. False positives are possible
3097 // on other platforms.
3098 llvm::Triple TargetTriple(Func.getParent()->getTargetTriple());
3099 if (TargetTriple.getArch() == llvm::Triple::x86_64)
3100 return new VarArgAMD64Helper(Func, Msan, Visitor);
3101 else if (TargetTriple.getArch() == llvm::Triple::mips64 ||
3102 TargetTriple.getArch() == llvm::Triple::mips64el)
3103 return new VarArgMIPS64Helper(Func, Msan, Visitor);
3105 return new VarArgNoOpHelper(Func, Msan, Visitor);
3110 bool MemorySanitizer::runOnFunction(Function &F) {
3111 if (&F == MsanCtorFunction)
3113 MemorySanitizerVisitor Visitor(F, *this);
3115 // Clear out readonly/readnone attributes.
3117 B.addAttribute(Attribute::ReadOnly)
3118 .addAttribute(Attribute::ReadNone);
3119 F.removeAttributes(AttributeSet::FunctionIndex,
3120 AttributeSet::get(F.getContext(),
3121 AttributeSet::FunctionIndex, B));
3123 return Visitor.runOnFunction();