[InstCombine] Signed saturation patterns
[llvm-core.git] / lib / Transforms / Instrumentation / DataFlowSanitizer.cpp
blobc0353cba0b2fd566261e4533d6839ca2716e36b5
1 //===- DataFlowSanitizer.cpp - dynamic data flow analysis -----------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 /// \file
10 /// This file is a part of DataFlowSanitizer, a generalised dynamic data flow
11 /// analysis.
12 ///
13 /// Unlike other Sanitizer tools, this tool is not designed to detect a specific
14 /// class of bugs on its own. Instead, it provides a generic dynamic data flow
15 /// analysis framework to be used by clients to help detect application-specific
16 /// issues within their own code.
17 ///
18 /// The analysis is based on automatic propagation of data flow labels (also
19 /// known as taint labels) through a program as it performs computation. Each
20 /// byte of application memory is backed by two bytes of shadow memory which
21 /// hold the label. On Linux/x86_64, memory is laid out as follows:
22 ///
23 /// +--------------------+ 0x800000000000 (top of memory)
24 /// | application memory |
25 /// +--------------------+ 0x700000008000 (kAppAddr)
26 /// | |
27 /// | unused |
28 /// | |
29 /// +--------------------+ 0x200200000000 (kUnusedAddr)
30 /// | union table |
31 /// +--------------------+ 0x200000000000 (kUnionTableAddr)
32 /// | shadow memory |
33 /// +--------------------+ 0x000000010000 (kShadowAddr)
34 /// | reserved by kernel |
35 /// +--------------------+ 0x000000000000
36 ///
37 /// To derive a shadow memory address from an application memory address,
38 /// bits 44-46 are cleared to bring the address into the range
39 /// [0x000000008000,0x100000000000). Then the address is shifted left by 1 to
40 /// account for the double byte representation of shadow labels and move the
41 /// address into the shadow memory range. See the function
42 /// DataFlowSanitizer::getShadowAddress below.
43 ///
44 /// For more information, please refer to the design document:
45 /// http://clang.llvm.org/docs/DataFlowSanitizerDesign.html
47 //===----------------------------------------------------------------------===//
49 #include "llvm/ADT/DenseMap.h"
50 #include "llvm/ADT/DenseSet.h"
51 #include "llvm/ADT/DepthFirstIterator.h"
52 #include "llvm/ADT/None.h"
53 #include "llvm/ADT/SmallPtrSet.h"
54 #include "llvm/ADT/SmallVector.h"
55 #include "llvm/ADT/StringExtras.h"
56 #include "llvm/ADT/StringRef.h"
57 #include "llvm/ADT/Triple.h"
58 #include "llvm/Transforms/Utils/Local.h"
59 #include "llvm/Analysis/ValueTracking.h"
60 #include "llvm/IR/Argument.h"
61 #include "llvm/IR/Attributes.h"
62 #include "llvm/IR/BasicBlock.h"
63 #include "llvm/IR/CallSite.h"
64 #include "llvm/IR/Constant.h"
65 #include "llvm/IR/Constants.h"
66 #include "llvm/IR/DataLayout.h"
67 #include "llvm/IR/DerivedTypes.h"
68 #include "llvm/IR/Dominators.h"
69 #include "llvm/IR/Function.h"
70 #include "llvm/IR/GlobalAlias.h"
71 #include "llvm/IR/GlobalValue.h"
72 #include "llvm/IR/GlobalVariable.h"
73 #include "llvm/IR/IRBuilder.h"
74 #include "llvm/IR/InlineAsm.h"
75 #include "llvm/IR/InstVisitor.h"
76 #include "llvm/IR/InstrTypes.h"
77 #include "llvm/IR/Instruction.h"
78 #include "llvm/IR/Instructions.h"
79 #include "llvm/IR/IntrinsicInst.h"
80 #include "llvm/IR/LLVMContext.h"
81 #include "llvm/IR/MDBuilder.h"
82 #include "llvm/IR/Module.h"
83 #include "llvm/IR/Type.h"
84 #include "llvm/IR/User.h"
85 #include "llvm/IR/Value.h"
86 #include "llvm/Pass.h"
87 #include "llvm/Support/Casting.h"
88 #include "llvm/Support/CommandLine.h"
89 #include "llvm/Support/ErrorHandling.h"
90 #include "llvm/Support/SpecialCaseList.h"
91 #include "llvm/Transforms/Instrumentation.h"
92 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
93 #include <algorithm>
94 #include <cassert>
95 #include <cstddef>
96 #include <cstdint>
97 #include <iterator>
98 #include <memory>
99 #include <set>
100 #include <string>
101 #include <utility>
102 #include <vector>
104 using namespace llvm;
106 // External symbol to be used when generating the shadow address for
107 // architectures with multiple VMAs. Instead of using a constant integer
108 // the runtime will set the external mask based on the VMA range.
109 static const char *const kDFSanExternShadowPtrMask = "__dfsan_shadow_ptr_mask";
111 // The -dfsan-preserve-alignment flag controls whether this pass assumes that
112 // alignment requirements provided by the input IR are correct. For example,
113 // if the input IR contains a load with alignment 8, this flag will cause
114 // the shadow load to have alignment 16. This flag is disabled by default as
115 // we have unfortunately encountered too much code (including Clang itself;
116 // see PR14291) which performs misaligned access.
117 static cl::opt<bool> ClPreserveAlignment(
118 "dfsan-preserve-alignment",
119 cl::desc("respect alignment requirements provided by input IR"), cl::Hidden,
120 cl::init(false));
122 // The ABI list files control how shadow parameters are passed. The pass treats
123 // every function labelled "uninstrumented" in the ABI list file as conforming
124 // to the "native" (i.e. unsanitized) ABI. Unless the ABI list contains
125 // additional annotations for those functions, a call to one of those functions
126 // will produce a warning message, as the labelling behaviour of the function is
127 // unknown. The other supported annotations are "functional" and "discard",
128 // which are described below under DataFlowSanitizer::WrapperKind.
129 static cl::list<std::string> ClABIListFiles(
130 "dfsan-abilist",
131 cl::desc("File listing native ABI functions and how the pass treats them"),
132 cl::Hidden);
134 // Controls whether the pass uses IA_Args or IA_TLS as the ABI for instrumented
135 // functions (see DataFlowSanitizer::InstrumentedABI below).
136 static cl::opt<bool> ClArgsABI(
137 "dfsan-args-abi",
138 cl::desc("Use the argument ABI rather than the TLS ABI"),
139 cl::Hidden);
141 // Controls whether the pass includes or ignores the labels of pointers in load
142 // instructions.
143 static cl::opt<bool> ClCombinePointerLabelsOnLoad(
144 "dfsan-combine-pointer-labels-on-load",
145 cl::desc("Combine the label of the pointer with the label of the data when "
146 "loading from memory."),
147 cl::Hidden, cl::init(true));
149 // Controls whether the pass includes or ignores the labels of pointers in
150 // stores instructions.
151 static cl::opt<bool> ClCombinePointerLabelsOnStore(
152 "dfsan-combine-pointer-labels-on-store",
153 cl::desc("Combine the label of the pointer with the label of the data when "
154 "storing in memory."),
155 cl::Hidden, cl::init(false));
157 static cl::opt<bool> ClDebugNonzeroLabels(
158 "dfsan-debug-nonzero-labels",
159 cl::desc("Insert calls to __dfsan_nonzero_label on observing a parameter, "
160 "load or return with a nonzero label"),
161 cl::Hidden);
163 static StringRef GetGlobalTypeString(const GlobalValue &G) {
164 // Types of GlobalVariables are always pointer types.
165 Type *GType = G.getValueType();
166 // For now we support blacklisting struct types only.
167 if (StructType *SGType = dyn_cast<StructType>(GType)) {
168 if (!SGType->isLiteral())
169 return SGType->getName();
171 return "<unknown type>";
174 namespace {
176 class DFSanABIList {
177 std::unique_ptr<SpecialCaseList> SCL;
179 public:
180 DFSanABIList() = default;
182 void set(std::unique_ptr<SpecialCaseList> List) { SCL = std::move(List); }
184 /// Returns whether either this function or its source file are listed in the
185 /// given category.
186 bool isIn(const Function &F, StringRef Category) const {
187 return isIn(*F.getParent(), Category) ||
188 SCL->inSection("dataflow", "fun", F.getName(), Category);
191 /// Returns whether this global alias is listed in the given category.
193 /// If GA aliases a function, the alias's name is matched as a function name
194 /// would be. Similarly, aliases of globals are matched like globals.
195 bool isIn(const GlobalAlias &GA, StringRef Category) const {
196 if (isIn(*GA.getParent(), Category))
197 return true;
199 if (isa<FunctionType>(GA.getValueType()))
200 return SCL->inSection("dataflow", "fun", GA.getName(), Category);
202 return SCL->inSection("dataflow", "global", GA.getName(), Category) ||
203 SCL->inSection("dataflow", "type", GetGlobalTypeString(GA),
204 Category);
207 /// Returns whether this module is listed in the given category.
208 bool isIn(const Module &M, StringRef Category) const {
209 return SCL->inSection("dataflow", "src", M.getModuleIdentifier(), Category);
213 /// TransformedFunction is used to express the result of transforming one
214 /// function type into another. This struct is immutable. It holds metadata
215 /// useful for updating calls of the old function to the new type.
216 struct TransformedFunction {
217 TransformedFunction(FunctionType* OriginalType,
218 FunctionType* TransformedType,
219 std::vector<unsigned> ArgumentIndexMapping)
220 : OriginalType(OriginalType),
221 TransformedType(TransformedType),
222 ArgumentIndexMapping(ArgumentIndexMapping) {}
224 // Disallow copies.
225 TransformedFunction(const TransformedFunction&) = delete;
226 TransformedFunction& operator=(const TransformedFunction&) = delete;
228 // Allow moves.
229 TransformedFunction(TransformedFunction&&) = default;
230 TransformedFunction& operator=(TransformedFunction&&) = default;
232 /// Type of the function before the transformation.
233 FunctionType *OriginalType;
235 /// Type of the function after the transformation.
236 FunctionType *TransformedType;
238 /// Transforming a function may change the position of arguments. This
239 /// member records the mapping from each argument's old position to its new
240 /// position. Argument positions are zero-indexed. If the transformation
241 /// from F to F' made the first argument of F into the third argument of F',
242 /// then ArgumentIndexMapping[0] will equal 2.
243 std::vector<unsigned> ArgumentIndexMapping;
246 /// Given function attributes from a call site for the original function,
247 /// return function attributes appropriate for a call to the transformed
248 /// function.
249 AttributeList TransformFunctionAttributes(
250 const TransformedFunction& TransformedFunction,
251 LLVMContext& Ctx, AttributeList CallSiteAttrs) {
253 // Construct a vector of AttributeSet for each function argument.
254 std::vector<llvm::AttributeSet> ArgumentAttributes(
255 TransformedFunction.TransformedType->getNumParams());
257 // Copy attributes from the parameter of the original function to the
258 // transformed version. 'ArgumentIndexMapping' holds the mapping from
259 // old argument position to new.
260 for (unsigned i=0, ie = TransformedFunction.ArgumentIndexMapping.size();
261 i < ie; ++i) {
262 unsigned TransformedIndex = TransformedFunction.ArgumentIndexMapping[i];
263 ArgumentAttributes[TransformedIndex] = CallSiteAttrs.getParamAttributes(i);
266 // Copy annotations on varargs arguments.
267 for (unsigned i = TransformedFunction.OriginalType->getNumParams(),
268 ie = CallSiteAttrs.getNumAttrSets(); i<ie; ++i) {
269 ArgumentAttributes.push_back(CallSiteAttrs.getParamAttributes(i));
272 return AttributeList::get(
273 Ctx,
274 CallSiteAttrs.getFnAttributes(),
275 CallSiteAttrs.getRetAttributes(),
276 llvm::makeArrayRef(ArgumentAttributes));
279 class DataFlowSanitizer : public ModulePass {
280 friend struct DFSanFunction;
281 friend class DFSanVisitor;
283 enum {
284 ShadowWidth = 16
287 /// Which ABI should be used for instrumented functions?
288 enum InstrumentedABI {
289 /// Argument and return value labels are passed through additional
290 /// arguments and by modifying the return type.
291 IA_Args,
293 /// Argument and return value labels are passed through TLS variables
294 /// __dfsan_arg_tls and __dfsan_retval_tls.
295 IA_TLS
298 /// How should calls to uninstrumented functions be handled?
299 enum WrapperKind {
300 /// This function is present in an uninstrumented form but we don't know
301 /// how it should be handled. Print a warning and call the function anyway.
302 /// Don't label the return value.
303 WK_Warning,
305 /// This function does not write to (user-accessible) memory, and its return
306 /// value is unlabelled.
307 WK_Discard,
309 /// This function does not write to (user-accessible) memory, and the label
310 /// of its return value is the union of the label of its arguments.
311 WK_Functional,
313 /// Instead of calling the function, a custom wrapper __dfsw_F is called,
314 /// where F is the name of the function. This function may wrap the
315 /// original function or provide its own implementation. This is similar to
316 /// the IA_Args ABI, except that IA_Args uses a struct return type to
317 /// pass the return value shadow in a register, while WK_Custom uses an
318 /// extra pointer argument to return the shadow. This allows the wrapped
319 /// form of the function type to be expressed in C.
320 WK_Custom
323 Module *Mod;
324 LLVMContext *Ctx;
325 IntegerType *ShadowTy;
326 PointerType *ShadowPtrTy;
327 IntegerType *IntptrTy;
328 ConstantInt *ZeroShadow;
329 ConstantInt *ShadowPtrMask;
330 ConstantInt *ShadowPtrMul;
331 Constant *ArgTLS;
332 Constant *RetvalTLS;
333 void *(*GetArgTLSPtr)();
334 void *(*GetRetvalTLSPtr)();
335 FunctionType *GetArgTLSTy;
336 FunctionType *GetRetvalTLSTy;
337 Constant *GetArgTLS;
338 Constant *GetRetvalTLS;
339 Constant *ExternalShadowMask;
340 FunctionType *DFSanUnionFnTy;
341 FunctionType *DFSanUnionLoadFnTy;
342 FunctionType *DFSanUnimplementedFnTy;
343 FunctionType *DFSanSetLabelFnTy;
344 FunctionType *DFSanNonzeroLabelFnTy;
345 FunctionType *DFSanVarargWrapperFnTy;
346 FunctionCallee DFSanUnionFn;
347 FunctionCallee DFSanCheckedUnionFn;
348 FunctionCallee DFSanUnionLoadFn;
349 FunctionCallee DFSanUnimplementedFn;
350 FunctionCallee DFSanSetLabelFn;
351 FunctionCallee DFSanNonzeroLabelFn;
352 FunctionCallee DFSanVarargWrapperFn;
353 MDNode *ColdCallWeights;
354 DFSanABIList ABIList;
355 DenseMap<Value *, Function *> UnwrappedFnMap;
356 AttrBuilder ReadOnlyNoneAttrs;
357 bool DFSanRuntimeShadowMask = false;
359 Value *getShadowAddress(Value *Addr, Instruction *Pos);
360 bool isInstrumented(const Function *F);
361 bool isInstrumented(const GlobalAlias *GA);
362 FunctionType *getArgsFunctionType(FunctionType *T);
363 FunctionType *getTrampolineFunctionType(FunctionType *T);
364 TransformedFunction getCustomFunctionType(FunctionType *T);
365 InstrumentedABI getInstrumentedABI();
366 WrapperKind getWrapperKind(Function *F);
367 void addGlobalNamePrefix(GlobalValue *GV);
368 Function *buildWrapperFunction(Function *F, StringRef NewFName,
369 GlobalValue::LinkageTypes NewFLink,
370 FunctionType *NewFT);
371 Constant *getOrBuildTrampolineFunction(FunctionType *FT, StringRef FName);
373 public:
374 static char ID;
376 DataFlowSanitizer(
377 const std::vector<std::string> &ABIListFiles = std::vector<std::string>(),
378 void *(*getArgTLS)() = nullptr, void *(*getRetValTLS)() = nullptr);
380 bool doInitialization(Module &M) override;
381 bool runOnModule(Module &M) override;
384 struct DFSanFunction {
385 DataFlowSanitizer &DFS;
386 Function *F;
387 DominatorTree DT;
388 DataFlowSanitizer::InstrumentedABI IA;
389 bool IsNativeABI;
390 Value *ArgTLSPtr = nullptr;
391 Value *RetvalTLSPtr = nullptr;
392 AllocaInst *LabelReturnAlloca = nullptr;
393 DenseMap<Value *, Value *> ValShadowMap;
394 DenseMap<AllocaInst *, AllocaInst *> AllocaShadowMap;
395 std::vector<std::pair<PHINode *, PHINode *>> PHIFixups;
396 DenseSet<Instruction *> SkipInsts;
397 std::vector<Value *> NonZeroChecks;
398 bool AvoidNewBlocks;
400 struct CachedCombinedShadow {
401 BasicBlock *Block;
402 Value *Shadow;
404 DenseMap<std::pair<Value *, Value *>, CachedCombinedShadow>
405 CachedCombinedShadows;
406 DenseMap<Value *, std::set<Value *>> ShadowElements;
408 DFSanFunction(DataFlowSanitizer &DFS, Function *F, bool IsNativeABI)
409 : DFS(DFS), F(F), IA(DFS.getInstrumentedABI()), IsNativeABI(IsNativeABI) {
410 DT.recalculate(*F);
411 // FIXME: Need to track down the register allocator issue which causes poor
412 // performance in pathological cases with large numbers of basic blocks.
413 AvoidNewBlocks = F->size() > 1000;
416 Value *getArgTLSPtr();
417 Value *getArgTLS(unsigned Index, Instruction *Pos);
418 Value *getRetvalTLS();
419 Value *getShadow(Value *V);
420 void setShadow(Instruction *I, Value *Shadow);
421 Value *combineShadows(Value *V1, Value *V2, Instruction *Pos);
422 Value *combineOperandShadows(Instruction *Inst);
423 Value *loadShadow(Value *ShadowAddr, uint64_t Size, uint64_t Align,
424 Instruction *Pos);
425 void storeShadow(Value *Addr, uint64_t Size, uint64_t Align, Value *Shadow,
426 Instruction *Pos);
429 class DFSanVisitor : public InstVisitor<DFSanVisitor> {
430 public:
431 DFSanFunction &DFSF;
433 DFSanVisitor(DFSanFunction &DFSF) : DFSF(DFSF) {}
435 const DataLayout &getDataLayout() const {
436 return DFSF.F->getParent()->getDataLayout();
439 void visitOperandShadowInst(Instruction &I);
440 void visitUnaryOperator(UnaryOperator &UO);
441 void visitBinaryOperator(BinaryOperator &BO);
442 void visitCastInst(CastInst &CI);
443 void visitCmpInst(CmpInst &CI);
444 void visitGetElementPtrInst(GetElementPtrInst &GEPI);
445 void visitLoadInst(LoadInst &LI);
446 void visitStoreInst(StoreInst &SI);
447 void visitReturnInst(ReturnInst &RI);
448 void visitCallSite(CallSite CS);
449 void visitPHINode(PHINode &PN);
450 void visitExtractElementInst(ExtractElementInst &I);
451 void visitInsertElementInst(InsertElementInst &I);
452 void visitShuffleVectorInst(ShuffleVectorInst &I);
453 void visitExtractValueInst(ExtractValueInst &I);
454 void visitInsertValueInst(InsertValueInst &I);
455 void visitAllocaInst(AllocaInst &I);
456 void visitSelectInst(SelectInst &I);
457 void visitMemSetInst(MemSetInst &I);
458 void visitMemTransferInst(MemTransferInst &I);
461 } // end anonymous namespace
463 char DataFlowSanitizer::ID;
465 INITIALIZE_PASS(DataFlowSanitizer, "dfsan",
466 "DataFlowSanitizer: dynamic data flow analysis.", false, false)
468 ModulePass *
469 llvm::createDataFlowSanitizerPass(const std::vector<std::string> &ABIListFiles,
470 void *(*getArgTLS)(),
471 void *(*getRetValTLS)()) {
472 return new DataFlowSanitizer(ABIListFiles, getArgTLS, getRetValTLS);
475 DataFlowSanitizer::DataFlowSanitizer(
476 const std::vector<std::string> &ABIListFiles, void *(*getArgTLS)(),
477 void *(*getRetValTLS)())
478 : ModulePass(ID), GetArgTLSPtr(getArgTLS), GetRetvalTLSPtr(getRetValTLS) {
479 std::vector<std::string> AllABIListFiles(std::move(ABIListFiles));
480 AllABIListFiles.insert(AllABIListFiles.end(), ClABIListFiles.begin(),
481 ClABIListFiles.end());
482 ABIList.set(SpecialCaseList::createOrDie(AllABIListFiles));
485 FunctionType *DataFlowSanitizer::getArgsFunctionType(FunctionType *T) {
486 SmallVector<Type *, 4> ArgTypes(T->param_begin(), T->param_end());
487 ArgTypes.append(T->getNumParams(), ShadowTy);
488 if (T->isVarArg())
489 ArgTypes.push_back(ShadowPtrTy);
490 Type *RetType = T->getReturnType();
491 if (!RetType->isVoidTy())
492 RetType = StructType::get(RetType, ShadowTy);
493 return FunctionType::get(RetType, ArgTypes, T->isVarArg());
496 FunctionType *DataFlowSanitizer::getTrampolineFunctionType(FunctionType *T) {
497 assert(!T->isVarArg());
498 SmallVector<Type *, 4> ArgTypes;
499 ArgTypes.push_back(T->getPointerTo());
500 ArgTypes.append(T->param_begin(), T->param_end());
501 ArgTypes.append(T->getNumParams(), ShadowTy);
502 Type *RetType = T->getReturnType();
503 if (!RetType->isVoidTy())
504 ArgTypes.push_back(ShadowPtrTy);
505 return FunctionType::get(T->getReturnType(), ArgTypes, false);
508 TransformedFunction DataFlowSanitizer::getCustomFunctionType(FunctionType *T) {
509 SmallVector<Type *, 4> ArgTypes;
511 // Some parameters of the custom function being constructed are
512 // parameters of T. Record the mapping from parameters of T to
513 // parameters of the custom function, so that parameter attributes
514 // at call sites can be updated.
515 std::vector<unsigned> ArgumentIndexMapping;
516 for (unsigned i = 0, ie = T->getNumParams(); i != ie; ++i) {
517 Type* param_type = T->getParamType(i);
518 FunctionType *FT;
519 if (isa<PointerType>(param_type) && (FT = dyn_cast<FunctionType>(
520 cast<PointerType>(param_type)->getElementType()))) {
521 ArgumentIndexMapping.push_back(ArgTypes.size());
522 ArgTypes.push_back(getTrampolineFunctionType(FT)->getPointerTo());
523 ArgTypes.push_back(Type::getInt8PtrTy(*Ctx));
524 } else {
525 ArgumentIndexMapping.push_back(ArgTypes.size());
526 ArgTypes.push_back(param_type);
529 for (unsigned i = 0, e = T->getNumParams(); i != e; ++i)
530 ArgTypes.push_back(ShadowTy);
531 if (T->isVarArg())
532 ArgTypes.push_back(ShadowPtrTy);
533 Type *RetType = T->getReturnType();
534 if (!RetType->isVoidTy())
535 ArgTypes.push_back(ShadowPtrTy);
536 return TransformedFunction(
537 T, FunctionType::get(T->getReturnType(), ArgTypes, T->isVarArg()),
538 ArgumentIndexMapping);
541 bool DataFlowSanitizer::doInitialization(Module &M) {
542 Triple TargetTriple(M.getTargetTriple());
543 bool IsX86_64 = TargetTriple.getArch() == Triple::x86_64;
544 bool IsMIPS64 = TargetTriple.isMIPS64();
545 bool IsAArch64 = TargetTriple.getArch() == Triple::aarch64 ||
546 TargetTriple.getArch() == Triple::aarch64_be;
548 const DataLayout &DL = M.getDataLayout();
550 Mod = &M;
551 Ctx = &M.getContext();
552 ShadowTy = IntegerType::get(*Ctx, ShadowWidth);
553 ShadowPtrTy = PointerType::getUnqual(ShadowTy);
554 IntptrTy = DL.getIntPtrType(*Ctx);
555 ZeroShadow = ConstantInt::getSigned(ShadowTy, 0);
556 ShadowPtrMul = ConstantInt::getSigned(IntptrTy, ShadowWidth / 8);
557 if (IsX86_64)
558 ShadowPtrMask = ConstantInt::getSigned(IntptrTy, ~0x700000000000LL);
559 else if (IsMIPS64)
560 ShadowPtrMask = ConstantInt::getSigned(IntptrTy, ~0xF000000000LL);
561 // AArch64 supports multiple VMAs and the shadow mask is set at runtime.
562 else if (IsAArch64)
563 DFSanRuntimeShadowMask = true;
564 else
565 report_fatal_error("unsupported triple");
567 Type *DFSanUnionArgs[2] = { ShadowTy, ShadowTy };
568 DFSanUnionFnTy =
569 FunctionType::get(ShadowTy, DFSanUnionArgs, /*isVarArg=*/ false);
570 Type *DFSanUnionLoadArgs[2] = { ShadowPtrTy, IntptrTy };
571 DFSanUnionLoadFnTy =
572 FunctionType::get(ShadowTy, DFSanUnionLoadArgs, /*isVarArg=*/ false);
573 DFSanUnimplementedFnTy = FunctionType::get(
574 Type::getVoidTy(*Ctx), Type::getInt8PtrTy(*Ctx), /*isVarArg=*/false);
575 Type *DFSanSetLabelArgs[3] = { ShadowTy, Type::getInt8PtrTy(*Ctx), IntptrTy };
576 DFSanSetLabelFnTy = FunctionType::get(Type::getVoidTy(*Ctx),
577 DFSanSetLabelArgs, /*isVarArg=*/false);
578 DFSanNonzeroLabelFnTy = FunctionType::get(
579 Type::getVoidTy(*Ctx), None, /*isVarArg=*/false);
580 DFSanVarargWrapperFnTy = FunctionType::get(
581 Type::getVoidTy(*Ctx), Type::getInt8PtrTy(*Ctx), /*isVarArg=*/false);
583 if (GetArgTLSPtr) {
584 Type *ArgTLSTy = ArrayType::get(ShadowTy, 64);
585 ArgTLS = nullptr;
586 GetArgTLSTy = FunctionType::get(PointerType::getUnqual(ArgTLSTy), false);
587 GetArgTLS = ConstantExpr::getIntToPtr(
588 ConstantInt::get(IntptrTy, uintptr_t(GetArgTLSPtr)),
589 PointerType::getUnqual(GetArgTLSTy));
591 if (GetRetvalTLSPtr) {
592 RetvalTLS = nullptr;
593 GetRetvalTLSTy = FunctionType::get(PointerType::getUnqual(ShadowTy), false);
594 GetRetvalTLS = ConstantExpr::getIntToPtr(
595 ConstantInt::get(IntptrTy, uintptr_t(GetRetvalTLSPtr)),
596 PointerType::getUnqual(GetRetvalTLSTy));
599 ColdCallWeights = MDBuilder(*Ctx).createBranchWeights(1, 1000);
600 return true;
603 bool DataFlowSanitizer::isInstrumented(const Function *F) {
604 return !ABIList.isIn(*F, "uninstrumented");
607 bool DataFlowSanitizer::isInstrumented(const GlobalAlias *GA) {
608 return !ABIList.isIn(*GA, "uninstrumented");
611 DataFlowSanitizer::InstrumentedABI DataFlowSanitizer::getInstrumentedABI() {
612 return ClArgsABI ? IA_Args : IA_TLS;
615 DataFlowSanitizer::WrapperKind DataFlowSanitizer::getWrapperKind(Function *F) {
616 if (ABIList.isIn(*F, "functional"))
617 return WK_Functional;
618 if (ABIList.isIn(*F, "discard"))
619 return WK_Discard;
620 if (ABIList.isIn(*F, "custom"))
621 return WK_Custom;
623 return WK_Warning;
626 void DataFlowSanitizer::addGlobalNamePrefix(GlobalValue *GV) {
627 std::string GVName = GV->getName(), Prefix = "dfs$";
628 GV->setName(Prefix + GVName);
630 // Try to change the name of the function in module inline asm. We only do
631 // this for specific asm directives, currently only ".symver", to try to avoid
632 // corrupting asm which happens to contain the symbol name as a substring.
633 // Note that the substitution for .symver assumes that the versioned symbol
634 // also has an instrumented name.
635 std::string Asm = GV->getParent()->getModuleInlineAsm();
636 std::string SearchStr = ".symver " + GVName + ",";
637 size_t Pos = Asm.find(SearchStr);
638 if (Pos != std::string::npos) {
639 Asm.replace(Pos, SearchStr.size(),
640 ".symver " + Prefix + GVName + "," + Prefix);
641 GV->getParent()->setModuleInlineAsm(Asm);
645 Function *
646 DataFlowSanitizer::buildWrapperFunction(Function *F, StringRef NewFName,
647 GlobalValue::LinkageTypes NewFLink,
648 FunctionType *NewFT) {
649 FunctionType *FT = F->getFunctionType();
650 Function *NewF = Function::Create(NewFT, NewFLink, F->getAddressSpace(),
651 NewFName, F->getParent());
652 NewF->copyAttributesFrom(F);
653 NewF->removeAttributes(
654 AttributeList::ReturnIndex,
655 AttributeFuncs::typeIncompatible(NewFT->getReturnType()));
657 BasicBlock *BB = BasicBlock::Create(*Ctx, "entry", NewF);
658 if (F->isVarArg()) {
659 NewF->removeAttributes(AttributeList::FunctionIndex,
660 AttrBuilder().addAttribute("split-stack"));
661 CallInst::Create(DFSanVarargWrapperFn,
662 IRBuilder<>(BB).CreateGlobalStringPtr(F->getName()), "",
663 BB);
664 new UnreachableInst(*Ctx, BB);
665 } else {
666 std::vector<Value *> Args;
667 unsigned n = FT->getNumParams();
668 for (Function::arg_iterator ai = NewF->arg_begin(); n != 0; ++ai, --n)
669 Args.push_back(&*ai);
670 CallInst *CI = CallInst::Create(F, Args, "", BB);
671 if (FT->getReturnType()->isVoidTy())
672 ReturnInst::Create(*Ctx, BB);
673 else
674 ReturnInst::Create(*Ctx, CI, BB);
677 return NewF;
680 Constant *DataFlowSanitizer::getOrBuildTrampolineFunction(FunctionType *FT,
681 StringRef FName) {
682 FunctionType *FTT = getTrampolineFunctionType(FT);
683 FunctionCallee C = Mod->getOrInsertFunction(FName, FTT);
684 Function *F = dyn_cast<Function>(C.getCallee());
685 if (F && F->isDeclaration()) {
686 F->setLinkage(GlobalValue::LinkOnceODRLinkage);
687 BasicBlock *BB = BasicBlock::Create(*Ctx, "entry", F);
688 std::vector<Value *> Args;
689 Function::arg_iterator AI = F->arg_begin(); ++AI;
690 for (unsigned N = FT->getNumParams(); N != 0; ++AI, --N)
691 Args.push_back(&*AI);
692 CallInst *CI = CallInst::Create(FT, &*F->arg_begin(), Args, "", BB);
693 ReturnInst *RI;
694 if (FT->getReturnType()->isVoidTy())
695 RI = ReturnInst::Create(*Ctx, BB);
696 else
697 RI = ReturnInst::Create(*Ctx, CI, BB);
699 DFSanFunction DFSF(*this, F, /*IsNativeABI=*/true);
700 Function::arg_iterator ValAI = F->arg_begin(), ShadowAI = AI; ++ValAI;
701 for (unsigned N = FT->getNumParams(); N != 0; ++ValAI, ++ShadowAI, --N)
702 DFSF.ValShadowMap[&*ValAI] = &*ShadowAI;
703 DFSanVisitor(DFSF).visitCallInst(*CI);
704 if (!FT->getReturnType()->isVoidTy())
705 new StoreInst(DFSF.getShadow(RI->getReturnValue()),
706 &*std::prev(F->arg_end()), RI);
709 return cast<Constant>(C.getCallee());
712 bool DataFlowSanitizer::runOnModule(Module &M) {
713 if (ABIList.isIn(M, "skip"))
714 return false;
716 if (!GetArgTLSPtr) {
717 Type *ArgTLSTy = ArrayType::get(ShadowTy, 64);
718 ArgTLS = Mod->getOrInsertGlobal("__dfsan_arg_tls", ArgTLSTy);
719 if (GlobalVariable *G = dyn_cast<GlobalVariable>(ArgTLS))
720 G->setThreadLocalMode(GlobalVariable::InitialExecTLSModel);
722 if (!GetRetvalTLSPtr) {
723 RetvalTLS = Mod->getOrInsertGlobal("__dfsan_retval_tls", ShadowTy);
724 if (GlobalVariable *G = dyn_cast<GlobalVariable>(RetvalTLS))
725 G->setThreadLocalMode(GlobalVariable::InitialExecTLSModel);
728 ExternalShadowMask =
729 Mod->getOrInsertGlobal(kDFSanExternShadowPtrMask, IntptrTy);
732 AttributeList AL;
733 AL = AL.addAttribute(M.getContext(), AttributeList::FunctionIndex,
734 Attribute::NoUnwind);
735 AL = AL.addAttribute(M.getContext(), AttributeList::FunctionIndex,
736 Attribute::ReadNone);
737 AL = AL.addAttribute(M.getContext(), AttributeList::ReturnIndex,
738 Attribute::ZExt);
739 AL = AL.addParamAttribute(M.getContext(), 0, Attribute::ZExt);
740 AL = AL.addParamAttribute(M.getContext(), 1, Attribute::ZExt);
741 DFSanUnionFn =
742 Mod->getOrInsertFunction("__dfsan_union", DFSanUnionFnTy, AL);
746 AttributeList AL;
747 AL = AL.addAttribute(M.getContext(), AttributeList::FunctionIndex,
748 Attribute::NoUnwind);
749 AL = AL.addAttribute(M.getContext(), AttributeList::FunctionIndex,
750 Attribute::ReadNone);
751 AL = AL.addAttribute(M.getContext(), AttributeList::ReturnIndex,
752 Attribute::ZExt);
753 AL = AL.addParamAttribute(M.getContext(), 0, Attribute::ZExt);
754 AL = AL.addParamAttribute(M.getContext(), 1, Attribute::ZExt);
755 DFSanCheckedUnionFn =
756 Mod->getOrInsertFunction("dfsan_union", DFSanUnionFnTy, AL);
759 AttributeList AL;
760 AL = AL.addAttribute(M.getContext(), AttributeList::FunctionIndex,
761 Attribute::NoUnwind);
762 AL = AL.addAttribute(M.getContext(), AttributeList::FunctionIndex,
763 Attribute::ReadOnly);
764 AL = AL.addAttribute(M.getContext(), AttributeList::ReturnIndex,
765 Attribute::ZExt);
766 DFSanUnionLoadFn =
767 Mod->getOrInsertFunction("__dfsan_union_load", DFSanUnionLoadFnTy, AL);
769 DFSanUnimplementedFn =
770 Mod->getOrInsertFunction("__dfsan_unimplemented", DFSanUnimplementedFnTy);
772 AttributeList AL;
773 AL = AL.addParamAttribute(M.getContext(), 0, Attribute::ZExt);
774 DFSanSetLabelFn =
775 Mod->getOrInsertFunction("__dfsan_set_label", DFSanSetLabelFnTy, AL);
777 DFSanNonzeroLabelFn =
778 Mod->getOrInsertFunction("__dfsan_nonzero_label", DFSanNonzeroLabelFnTy);
779 DFSanVarargWrapperFn = Mod->getOrInsertFunction("__dfsan_vararg_wrapper",
780 DFSanVarargWrapperFnTy);
782 std::vector<Function *> FnsToInstrument;
783 SmallPtrSet<Function *, 2> FnsWithNativeABI;
784 for (Function &i : M) {
785 if (!i.isIntrinsic() &&
786 &i != DFSanUnionFn.getCallee()->stripPointerCasts() &&
787 &i != DFSanCheckedUnionFn.getCallee()->stripPointerCasts() &&
788 &i != DFSanUnionLoadFn.getCallee()->stripPointerCasts() &&
789 &i != DFSanUnimplementedFn.getCallee()->stripPointerCasts() &&
790 &i != DFSanSetLabelFn.getCallee()->stripPointerCasts() &&
791 &i != DFSanNonzeroLabelFn.getCallee()->stripPointerCasts() &&
792 &i != DFSanVarargWrapperFn.getCallee()->stripPointerCasts())
793 FnsToInstrument.push_back(&i);
796 // Give function aliases prefixes when necessary, and build wrappers where the
797 // instrumentedness is inconsistent.
798 for (Module::alias_iterator i = M.alias_begin(), e = M.alias_end(); i != e;) {
799 GlobalAlias *GA = &*i;
800 ++i;
801 // Don't stop on weak. We assume people aren't playing games with the
802 // instrumentedness of overridden weak aliases.
803 if (auto F = dyn_cast<Function>(GA->getBaseObject())) {
804 bool GAInst = isInstrumented(GA), FInst = isInstrumented(F);
805 if (GAInst && FInst) {
806 addGlobalNamePrefix(GA);
807 } else if (GAInst != FInst) {
808 // Non-instrumented alias of an instrumented function, or vice versa.
809 // Replace the alias with a native-ABI wrapper of the aliasee. The pass
810 // below will take care of instrumenting it.
811 Function *NewF =
812 buildWrapperFunction(F, "", GA->getLinkage(), F->getFunctionType());
813 GA->replaceAllUsesWith(ConstantExpr::getBitCast(NewF, GA->getType()));
814 NewF->takeName(GA);
815 GA->eraseFromParent();
816 FnsToInstrument.push_back(NewF);
821 ReadOnlyNoneAttrs.addAttribute(Attribute::ReadOnly)
822 .addAttribute(Attribute::ReadNone);
824 // First, change the ABI of every function in the module. ABI-listed
825 // functions keep their original ABI and get a wrapper function.
826 for (std::vector<Function *>::iterator i = FnsToInstrument.begin(),
827 e = FnsToInstrument.end();
828 i != e; ++i) {
829 Function &F = **i;
830 FunctionType *FT = F.getFunctionType();
832 bool IsZeroArgsVoidRet = (FT->getNumParams() == 0 && !FT->isVarArg() &&
833 FT->getReturnType()->isVoidTy());
835 if (isInstrumented(&F)) {
836 // Instrumented functions get a 'dfs$' prefix. This allows us to more
837 // easily identify cases of mismatching ABIs.
838 if (getInstrumentedABI() == IA_Args && !IsZeroArgsVoidRet) {
839 FunctionType *NewFT = getArgsFunctionType(FT);
840 Function *NewF = Function::Create(NewFT, F.getLinkage(),
841 F.getAddressSpace(), "", &M);
842 NewF->copyAttributesFrom(&F);
843 NewF->removeAttributes(
844 AttributeList::ReturnIndex,
845 AttributeFuncs::typeIncompatible(NewFT->getReturnType()));
846 for (Function::arg_iterator FArg = F.arg_begin(),
847 NewFArg = NewF->arg_begin(),
848 FArgEnd = F.arg_end();
849 FArg != FArgEnd; ++FArg, ++NewFArg) {
850 FArg->replaceAllUsesWith(&*NewFArg);
852 NewF->getBasicBlockList().splice(NewF->begin(), F.getBasicBlockList());
854 for (Function::user_iterator UI = F.user_begin(), UE = F.user_end();
855 UI != UE;) {
856 BlockAddress *BA = dyn_cast<BlockAddress>(*UI);
857 ++UI;
858 if (BA) {
859 BA->replaceAllUsesWith(
860 BlockAddress::get(NewF, BA->getBasicBlock()));
861 delete BA;
864 F.replaceAllUsesWith(
865 ConstantExpr::getBitCast(NewF, PointerType::getUnqual(FT)));
866 NewF->takeName(&F);
867 F.eraseFromParent();
868 *i = NewF;
869 addGlobalNamePrefix(NewF);
870 } else {
871 addGlobalNamePrefix(&F);
873 } else if (!IsZeroArgsVoidRet || getWrapperKind(&F) == WK_Custom) {
874 // Build a wrapper function for F. The wrapper simply calls F, and is
875 // added to FnsToInstrument so that any instrumentation according to its
876 // WrapperKind is done in the second pass below.
877 FunctionType *NewFT = getInstrumentedABI() == IA_Args
878 ? getArgsFunctionType(FT)
879 : FT;
881 // If the function being wrapped has local linkage, then preserve the
882 // function's linkage in the wrapper function.
883 GlobalValue::LinkageTypes wrapperLinkage =
884 F.hasLocalLinkage()
885 ? F.getLinkage()
886 : GlobalValue::LinkOnceODRLinkage;
888 Function *NewF = buildWrapperFunction(
889 &F, std::string("dfsw$") + std::string(F.getName()),
890 wrapperLinkage, NewFT);
891 if (getInstrumentedABI() == IA_TLS)
892 NewF->removeAttributes(AttributeList::FunctionIndex, ReadOnlyNoneAttrs);
894 Value *WrappedFnCst =
895 ConstantExpr::getBitCast(NewF, PointerType::getUnqual(FT));
896 F.replaceAllUsesWith(WrappedFnCst);
898 UnwrappedFnMap[WrappedFnCst] = &F;
899 *i = NewF;
901 if (!F.isDeclaration()) {
902 // This function is probably defining an interposition of an
903 // uninstrumented function and hence needs to keep the original ABI.
904 // But any functions it may call need to use the instrumented ABI, so
905 // we instrument it in a mode which preserves the original ABI.
906 FnsWithNativeABI.insert(&F);
908 // This code needs to rebuild the iterators, as they may be invalidated
909 // by the push_back, taking care that the new range does not include
910 // any functions added by this code.
911 size_t N = i - FnsToInstrument.begin(),
912 Count = e - FnsToInstrument.begin();
913 FnsToInstrument.push_back(&F);
914 i = FnsToInstrument.begin() + N;
915 e = FnsToInstrument.begin() + Count;
917 // Hopefully, nobody will try to indirectly call a vararg
918 // function... yet.
919 } else if (FT->isVarArg()) {
920 UnwrappedFnMap[&F] = &F;
921 *i = nullptr;
925 for (Function *i : FnsToInstrument) {
926 if (!i || i->isDeclaration())
927 continue;
929 removeUnreachableBlocks(*i);
931 DFSanFunction DFSF(*this, i, FnsWithNativeABI.count(i));
933 // DFSanVisitor may create new basic blocks, which confuses df_iterator.
934 // Build a copy of the list before iterating over it.
935 SmallVector<BasicBlock *, 4> BBList(depth_first(&i->getEntryBlock()));
937 for (BasicBlock *i : BBList) {
938 Instruction *Inst = &i->front();
939 while (true) {
940 // DFSanVisitor may split the current basic block, changing the current
941 // instruction's next pointer and moving the next instruction to the
942 // tail block from which we should continue.
943 Instruction *Next = Inst->getNextNode();
944 // DFSanVisitor may delete Inst, so keep track of whether it was a
945 // terminator.
946 bool IsTerminator = Inst->isTerminator();
947 if (!DFSF.SkipInsts.count(Inst))
948 DFSanVisitor(DFSF).visit(Inst);
949 if (IsTerminator)
950 break;
951 Inst = Next;
955 // We will not necessarily be able to compute the shadow for every phi node
956 // until we have visited every block. Therefore, the code that handles phi
957 // nodes adds them to the PHIFixups list so that they can be properly
958 // handled here.
959 for (std::vector<std::pair<PHINode *, PHINode *>>::iterator
960 i = DFSF.PHIFixups.begin(),
961 e = DFSF.PHIFixups.end();
962 i != e; ++i) {
963 for (unsigned val = 0, n = i->first->getNumIncomingValues(); val != n;
964 ++val) {
965 i->second->setIncomingValue(
966 val, DFSF.getShadow(i->first->getIncomingValue(val)));
970 // -dfsan-debug-nonzero-labels will split the CFG in all kinds of crazy
971 // places (i.e. instructions in basic blocks we haven't even begun visiting
972 // yet). To make our life easier, do this work in a pass after the main
973 // instrumentation.
974 if (ClDebugNonzeroLabels) {
975 for (Value *V : DFSF.NonZeroChecks) {
976 Instruction *Pos;
977 if (Instruction *I = dyn_cast<Instruction>(V))
978 Pos = I->getNextNode();
979 else
980 Pos = &DFSF.F->getEntryBlock().front();
981 while (isa<PHINode>(Pos) || isa<AllocaInst>(Pos))
982 Pos = Pos->getNextNode();
983 IRBuilder<> IRB(Pos);
984 Value *Ne = IRB.CreateICmpNE(V, DFSF.DFS.ZeroShadow);
985 BranchInst *BI = cast<BranchInst>(SplitBlockAndInsertIfThen(
986 Ne, Pos, /*Unreachable=*/false, ColdCallWeights));
987 IRBuilder<> ThenIRB(BI);
988 ThenIRB.CreateCall(DFSF.DFS.DFSanNonzeroLabelFn, {});
993 return false;
996 Value *DFSanFunction::getArgTLSPtr() {
997 if (ArgTLSPtr)
998 return ArgTLSPtr;
999 if (DFS.ArgTLS)
1000 return ArgTLSPtr = DFS.ArgTLS;
1002 IRBuilder<> IRB(&F->getEntryBlock().front());
1003 return ArgTLSPtr = IRB.CreateCall(DFS.GetArgTLSTy, DFS.GetArgTLS, {});
1006 Value *DFSanFunction::getRetvalTLS() {
1007 if (RetvalTLSPtr)
1008 return RetvalTLSPtr;
1009 if (DFS.RetvalTLS)
1010 return RetvalTLSPtr = DFS.RetvalTLS;
1012 IRBuilder<> IRB(&F->getEntryBlock().front());
1013 return RetvalTLSPtr =
1014 IRB.CreateCall(DFS.GetRetvalTLSTy, DFS.GetRetvalTLS, {});
1017 Value *DFSanFunction::getArgTLS(unsigned Idx, Instruction *Pos) {
1018 IRBuilder<> IRB(Pos);
1019 return IRB.CreateConstGEP2_64(ArrayType::get(DFS.ShadowTy, 64),
1020 getArgTLSPtr(), 0, Idx);
1023 Value *DFSanFunction::getShadow(Value *V) {
1024 if (!isa<Argument>(V) && !isa<Instruction>(V))
1025 return DFS.ZeroShadow;
1026 Value *&Shadow = ValShadowMap[V];
1027 if (!Shadow) {
1028 if (Argument *A = dyn_cast<Argument>(V)) {
1029 if (IsNativeABI)
1030 return DFS.ZeroShadow;
1031 switch (IA) {
1032 case DataFlowSanitizer::IA_TLS: {
1033 Value *ArgTLSPtr = getArgTLSPtr();
1034 Instruction *ArgTLSPos =
1035 DFS.ArgTLS ? &*F->getEntryBlock().begin()
1036 : cast<Instruction>(ArgTLSPtr)->getNextNode();
1037 IRBuilder<> IRB(ArgTLSPos);
1038 Shadow =
1039 IRB.CreateLoad(DFS.ShadowTy, getArgTLS(A->getArgNo(), ArgTLSPos));
1040 break;
1042 case DataFlowSanitizer::IA_Args: {
1043 unsigned ArgIdx = A->getArgNo() + F->arg_size() / 2;
1044 Function::arg_iterator i = F->arg_begin();
1045 while (ArgIdx--)
1046 ++i;
1047 Shadow = &*i;
1048 assert(Shadow->getType() == DFS.ShadowTy);
1049 break;
1052 NonZeroChecks.push_back(Shadow);
1053 } else {
1054 Shadow = DFS.ZeroShadow;
1057 return Shadow;
1060 void DFSanFunction::setShadow(Instruction *I, Value *Shadow) {
1061 assert(!ValShadowMap.count(I));
1062 assert(Shadow->getType() == DFS.ShadowTy);
1063 ValShadowMap[I] = Shadow;
1066 Value *DataFlowSanitizer::getShadowAddress(Value *Addr, Instruction *Pos) {
1067 assert(Addr != RetvalTLS && "Reinstrumenting?");
1068 IRBuilder<> IRB(Pos);
1069 Value *ShadowPtrMaskValue;
1070 if (DFSanRuntimeShadowMask)
1071 ShadowPtrMaskValue = IRB.CreateLoad(IntptrTy, ExternalShadowMask);
1072 else
1073 ShadowPtrMaskValue = ShadowPtrMask;
1074 return IRB.CreateIntToPtr(
1075 IRB.CreateMul(
1076 IRB.CreateAnd(IRB.CreatePtrToInt(Addr, IntptrTy),
1077 IRB.CreatePtrToInt(ShadowPtrMaskValue, IntptrTy)),
1078 ShadowPtrMul),
1079 ShadowPtrTy);
1082 // Generates IR to compute the union of the two given shadows, inserting it
1083 // before Pos. Returns the computed union Value.
1084 Value *DFSanFunction::combineShadows(Value *V1, Value *V2, Instruction *Pos) {
1085 if (V1 == DFS.ZeroShadow)
1086 return V2;
1087 if (V2 == DFS.ZeroShadow)
1088 return V1;
1089 if (V1 == V2)
1090 return V1;
1092 auto V1Elems = ShadowElements.find(V1);
1093 auto V2Elems = ShadowElements.find(V2);
1094 if (V1Elems != ShadowElements.end() && V2Elems != ShadowElements.end()) {
1095 if (std::includes(V1Elems->second.begin(), V1Elems->second.end(),
1096 V2Elems->second.begin(), V2Elems->second.end())) {
1097 return V1;
1098 } else if (std::includes(V2Elems->second.begin(), V2Elems->second.end(),
1099 V1Elems->second.begin(), V1Elems->second.end())) {
1100 return V2;
1102 } else if (V1Elems != ShadowElements.end()) {
1103 if (V1Elems->second.count(V2))
1104 return V1;
1105 } else if (V2Elems != ShadowElements.end()) {
1106 if (V2Elems->second.count(V1))
1107 return V2;
1110 auto Key = std::make_pair(V1, V2);
1111 if (V1 > V2)
1112 std::swap(Key.first, Key.second);
1113 CachedCombinedShadow &CCS = CachedCombinedShadows[Key];
1114 if (CCS.Block && DT.dominates(CCS.Block, Pos->getParent()))
1115 return CCS.Shadow;
1117 IRBuilder<> IRB(Pos);
1118 if (AvoidNewBlocks) {
1119 CallInst *Call = IRB.CreateCall(DFS.DFSanCheckedUnionFn, {V1, V2});
1120 Call->addAttribute(AttributeList::ReturnIndex, Attribute::ZExt);
1121 Call->addParamAttr(0, Attribute::ZExt);
1122 Call->addParamAttr(1, Attribute::ZExt);
1124 CCS.Block = Pos->getParent();
1125 CCS.Shadow = Call;
1126 } else {
1127 BasicBlock *Head = Pos->getParent();
1128 Value *Ne = IRB.CreateICmpNE(V1, V2);
1129 BranchInst *BI = cast<BranchInst>(SplitBlockAndInsertIfThen(
1130 Ne, Pos, /*Unreachable=*/false, DFS.ColdCallWeights, &DT));
1131 IRBuilder<> ThenIRB(BI);
1132 CallInst *Call = ThenIRB.CreateCall(DFS.DFSanUnionFn, {V1, V2});
1133 Call->addAttribute(AttributeList::ReturnIndex, Attribute::ZExt);
1134 Call->addParamAttr(0, Attribute::ZExt);
1135 Call->addParamAttr(1, Attribute::ZExt);
1137 BasicBlock *Tail = BI->getSuccessor(0);
1138 PHINode *Phi = PHINode::Create(DFS.ShadowTy, 2, "", &Tail->front());
1139 Phi->addIncoming(Call, Call->getParent());
1140 Phi->addIncoming(V1, Head);
1142 CCS.Block = Tail;
1143 CCS.Shadow = Phi;
1146 std::set<Value *> UnionElems;
1147 if (V1Elems != ShadowElements.end()) {
1148 UnionElems = V1Elems->second;
1149 } else {
1150 UnionElems.insert(V1);
1152 if (V2Elems != ShadowElements.end()) {
1153 UnionElems.insert(V2Elems->second.begin(), V2Elems->second.end());
1154 } else {
1155 UnionElems.insert(V2);
1157 ShadowElements[CCS.Shadow] = std::move(UnionElems);
1159 return CCS.Shadow;
1162 // A convenience function which folds the shadows of each of the operands
1163 // of the provided instruction Inst, inserting the IR before Inst. Returns
1164 // the computed union Value.
1165 Value *DFSanFunction::combineOperandShadows(Instruction *Inst) {
1166 if (Inst->getNumOperands() == 0)
1167 return DFS.ZeroShadow;
1169 Value *Shadow = getShadow(Inst->getOperand(0));
1170 for (unsigned i = 1, n = Inst->getNumOperands(); i != n; ++i) {
1171 Shadow = combineShadows(Shadow, getShadow(Inst->getOperand(i)), Inst);
1173 return Shadow;
1176 void DFSanVisitor::visitOperandShadowInst(Instruction &I) {
1177 Value *CombinedShadow = DFSF.combineOperandShadows(&I);
1178 DFSF.setShadow(&I, CombinedShadow);
1181 // Generates IR to load shadow corresponding to bytes [Addr, Addr+Size), where
1182 // Addr has alignment Align, and take the union of each of those shadows.
1183 Value *DFSanFunction::loadShadow(Value *Addr, uint64_t Size, uint64_t Align,
1184 Instruction *Pos) {
1185 if (AllocaInst *AI = dyn_cast<AllocaInst>(Addr)) {
1186 const auto i = AllocaShadowMap.find(AI);
1187 if (i != AllocaShadowMap.end()) {
1188 IRBuilder<> IRB(Pos);
1189 return IRB.CreateLoad(DFS.ShadowTy, i->second);
1193 uint64_t ShadowAlign = Align * DFS.ShadowWidth / 8;
1194 SmallVector<const Value *, 2> Objs;
1195 GetUnderlyingObjects(Addr, Objs, Pos->getModule()->getDataLayout());
1196 bool AllConstants = true;
1197 for (const Value *Obj : Objs) {
1198 if (isa<Function>(Obj) || isa<BlockAddress>(Obj))
1199 continue;
1200 if (isa<GlobalVariable>(Obj) && cast<GlobalVariable>(Obj)->isConstant())
1201 continue;
1203 AllConstants = false;
1204 break;
1206 if (AllConstants)
1207 return DFS.ZeroShadow;
1209 Value *ShadowAddr = DFS.getShadowAddress(Addr, Pos);
1210 switch (Size) {
1211 case 0:
1212 return DFS.ZeroShadow;
1213 case 1: {
1214 LoadInst *LI = new LoadInst(DFS.ShadowTy, ShadowAddr, "", Pos);
1215 LI->setAlignment(MaybeAlign(ShadowAlign));
1216 return LI;
1218 case 2: {
1219 IRBuilder<> IRB(Pos);
1220 Value *ShadowAddr1 = IRB.CreateGEP(DFS.ShadowTy, ShadowAddr,
1221 ConstantInt::get(DFS.IntptrTy, 1));
1222 return combineShadows(
1223 IRB.CreateAlignedLoad(DFS.ShadowTy, ShadowAddr, ShadowAlign),
1224 IRB.CreateAlignedLoad(DFS.ShadowTy, ShadowAddr1, ShadowAlign), Pos);
1227 if (!AvoidNewBlocks && Size % (64 / DFS.ShadowWidth) == 0) {
1228 // Fast path for the common case where each byte has identical shadow: load
1229 // shadow 64 bits at a time, fall out to a __dfsan_union_load call if any
1230 // shadow is non-equal.
1231 BasicBlock *FallbackBB = BasicBlock::Create(*DFS.Ctx, "", F);
1232 IRBuilder<> FallbackIRB(FallbackBB);
1233 CallInst *FallbackCall = FallbackIRB.CreateCall(
1234 DFS.DFSanUnionLoadFn,
1235 {ShadowAddr, ConstantInt::get(DFS.IntptrTy, Size)});
1236 FallbackCall->addAttribute(AttributeList::ReturnIndex, Attribute::ZExt);
1238 // Compare each of the shadows stored in the loaded 64 bits to each other,
1239 // by computing (WideShadow rotl ShadowWidth) == WideShadow.
1240 IRBuilder<> IRB(Pos);
1241 Value *WideAddr =
1242 IRB.CreateBitCast(ShadowAddr, Type::getInt64PtrTy(*DFS.Ctx));
1243 Value *WideShadow =
1244 IRB.CreateAlignedLoad(IRB.getInt64Ty(), WideAddr, ShadowAlign);
1245 Value *TruncShadow = IRB.CreateTrunc(WideShadow, DFS.ShadowTy);
1246 Value *ShlShadow = IRB.CreateShl(WideShadow, DFS.ShadowWidth);
1247 Value *ShrShadow = IRB.CreateLShr(WideShadow, 64 - DFS.ShadowWidth);
1248 Value *RotShadow = IRB.CreateOr(ShlShadow, ShrShadow);
1249 Value *ShadowsEq = IRB.CreateICmpEQ(WideShadow, RotShadow);
1251 BasicBlock *Head = Pos->getParent();
1252 BasicBlock *Tail = Head->splitBasicBlock(Pos->getIterator());
1254 if (DomTreeNode *OldNode = DT.getNode(Head)) {
1255 std::vector<DomTreeNode *> Children(OldNode->begin(), OldNode->end());
1257 DomTreeNode *NewNode = DT.addNewBlock(Tail, Head);
1258 for (auto Child : Children)
1259 DT.changeImmediateDominator(Child, NewNode);
1262 // In the following code LastBr will refer to the previous basic block's
1263 // conditional branch instruction, whose true successor is fixed up to point
1264 // to the next block during the loop below or to the tail after the final
1265 // iteration.
1266 BranchInst *LastBr = BranchInst::Create(FallbackBB, FallbackBB, ShadowsEq);
1267 ReplaceInstWithInst(Head->getTerminator(), LastBr);
1268 DT.addNewBlock(FallbackBB, Head);
1270 for (uint64_t Ofs = 64 / DFS.ShadowWidth; Ofs != Size;
1271 Ofs += 64 / DFS.ShadowWidth) {
1272 BasicBlock *NextBB = BasicBlock::Create(*DFS.Ctx, "", F);
1273 DT.addNewBlock(NextBB, LastBr->getParent());
1274 IRBuilder<> NextIRB(NextBB);
1275 WideAddr = NextIRB.CreateGEP(Type::getInt64Ty(*DFS.Ctx), WideAddr,
1276 ConstantInt::get(DFS.IntptrTy, 1));
1277 Value *NextWideShadow = NextIRB.CreateAlignedLoad(NextIRB.getInt64Ty(),
1278 WideAddr, ShadowAlign);
1279 ShadowsEq = NextIRB.CreateICmpEQ(WideShadow, NextWideShadow);
1280 LastBr->setSuccessor(0, NextBB);
1281 LastBr = NextIRB.CreateCondBr(ShadowsEq, FallbackBB, FallbackBB);
1284 LastBr->setSuccessor(0, Tail);
1285 FallbackIRB.CreateBr(Tail);
1286 PHINode *Shadow = PHINode::Create(DFS.ShadowTy, 2, "", &Tail->front());
1287 Shadow->addIncoming(FallbackCall, FallbackBB);
1288 Shadow->addIncoming(TruncShadow, LastBr->getParent());
1289 return Shadow;
1292 IRBuilder<> IRB(Pos);
1293 CallInst *FallbackCall = IRB.CreateCall(
1294 DFS.DFSanUnionLoadFn, {ShadowAddr, ConstantInt::get(DFS.IntptrTy, Size)});
1295 FallbackCall->addAttribute(AttributeList::ReturnIndex, Attribute::ZExt);
1296 return FallbackCall;
1299 void DFSanVisitor::visitLoadInst(LoadInst &LI) {
1300 auto &DL = LI.getModule()->getDataLayout();
1301 uint64_t Size = DL.getTypeStoreSize(LI.getType());
1302 if (Size == 0) {
1303 DFSF.setShadow(&LI, DFSF.DFS.ZeroShadow);
1304 return;
1307 uint64_t Align;
1308 if (ClPreserveAlignment) {
1309 Align = LI.getAlignment();
1310 if (Align == 0)
1311 Align = DL.getABITypeAlignment(LI.getType());
1312 } else {
1313 Align = 1;
1315 IRBuilder<> IRB(&LI);
1316 Value *Shadow = DFSF.loadShadow(LI.getPointerOperand(), Size, Align, &LI);
1317 if (ClCombinePointerLabelsOnLoad) {
1318 Value *PtrShadow = DFSF.getShadow(LI.getPointerOperand());
1319 Shadow = DFSF.combineShadows(Shadow, PtrShadow, &LI);
1321 if (Shadow != DFSF.DFS.ZeroShadow)
1322 DFSF.NonZeroChecks.push_back(Shadow);
1324 DFSF.setShadow(&LI, Shadow);
1327 void DFSanFunction::storeShadow(Value *Addr, uint64_t Size, uint64_t Align,
1328 Value *Shadow, Instruction *Pos) {
1329 if (AllocaInst *AI = dyn_cast<AllocaInst>(Addr)) {
1330 const auto i = AllocaShadowMap.find(AI);
1331 if (i != AllocaShadowMap.end()) {
1332 IRBuilder<> IRB(Pos);
1333 IRB.CreateStore(Shadow, i->second);
1334 return;
1338 uint64_t ShadowAlign = Align * DFS.ShadowWidth / 8;
1339 IRBuilder<> IRB(Pos);
1340 Value *ShadowAddr = DFS.getShadowAddress(Addr, Pos);
1341 if (Shadow == DFS.ZeroShadow) {
1342 IntegerType *ShadowTy = IntegerType::get(*DFS.Ctx, Size * DFS.ShadowWidth);
1343 Value *ExtZeroShadow = ConstantInt::get(ShadowTy, 0);
1344 Value *ExtShadowAddr =
1345 IRB.CreateBitCast(ShadowAddr, PointerType::getUnqual(ShadowTy));
1346 IRB.CreateAlignedStore(ExtZeroShadow, ExtShadowAddr, ShadowAlign);
1347 return;
1350 const unsigned ShadowVecSize = 128 / DFS.ShadowWidth;
1351 uint64_t Offset = 0;
1352 if (Size >= ShadowVecSize) {
1353 VectorType *ShadowVecTy = VectorType::get(DFS.ShadowTy, ShadowVecSize);
1354 Value *ShadowVec = UndefValue::get(ShadowVecTy);
1355 for (unsigned i = 0; i != ShadowVecSize; ++i) {
1356 ShadowVec = IRB.CreateInsertElement(
1357 ShadowVec, Shadow, ConstantInt::get(Type::getInt32Ty(*DFS.Ctx), i));
1359 Value *ShadowVecAddr =
1360 IRB.CreateBitCast(ShadowAddr, PointerType::getUnqual(ShadowVecTy));
1361 do {
1362 Value *CurShadowVecAddr =
1363 IRB.CreateConstGEP1_32(ShadowVecTy, ShadowVecAddr, Offset);
1364 IRB.CreateAlignedStore(ShadowVec, CurShadowVecAddr, ShadowAlign);
1365 Size -= ShadowVecSize;
1366 ++Offset;
1367 } while (Size >= ShadowVecSize);
1368 Offset *= ShadowVecSize;
1370 while (Size > 0) {
1371 Value *CurShadowAddr =
1372 IRB.CreateConstGEP1_32(DFS.ShadowTy, ShadowAddr, Offset);
1373 IRB.CreateAlignedStore(Shadow, CurShadowAddr, ShadowAlign);
1374 --Size;
1375 ++Offset;
1379 void DFSanVisitor::visitStoreInst(StoreInst &SI) {
1380 auto &DL = SI.getModule()->getDataLayout();
1381 uint64_t Size = DL.getTypeStoreSize(SI.getValueOperand()->getType());
1382 if (Size == 0)
1383 return;
1385 uint64_t Align;
1386 if (ClPreserveAlignment) {
1387 Align = SI.getAlignment();
1388 if (Align == 0)
1389 Align = DL.getABITypeAlignment(SI.getValueOperand()->getType());
1390 } else {
1391 Align = 1;
1394 Value* Shadow = DFSF.getShadow(SI.getValueOperand());
1395 if (ClCombinePointerLabelsOnStore) {
1396 Value *PtrShadow = DFSF.getShadow(SI.getPointerOperand());
1397 Shadow = DFSF.combineShadows(Shadow, PtrShadow, &SI);
1399 DFSF.storeShadow(SI.getPointerOperand(), Size, Align, Shadow, &SI);
1402 void DFSanVisitor::visitUnaryOperator(UnaryOperator &UO) {
1403 visitOperandShadowInst(UO);
1406 void DFSanVisitor::visitBinaryOperator(BinaryOperator &BO) {
1407 visitOperandShadowInst(BO);
1410 void DFSanVisitor::visitCastInst(CastInst &CI) { visitOperandShadowInst(CI); }
1412 void DFSanVisitor::visitCmpInst(CmpInst &CI) { visitOperandShadowInst(CI); }
1414 void DFSanVisitor::visitGetElementPtrInst(GetElementPtrInst &GEPI) {
1415 visitOperandShadowInst(GEPI);
1418 void DFSanVisitor::visitExtractElementInst(ExtractElementInst &I) {
1419 visitOperandShadowInst(I);
1422 void DFSanVisitor::visitInsertElementInst(InsertElementInst &I) {
1423 visitOperandShadowInst(I);
1426 void DFSanVisitor::visitShuffleVectorInst(ShuffleVectorInst &I) {
1427 visitOperandShadowInst(I);
1430 void DFSanVisitor::visitExtractValueInst(ExtractValueInst &I) {
1431 visitOperandShadowInst(I);
1434 void DFSanVisitor::visitInsertValueInst(InsertValueInst &I) {
1435 visitOperandShadowInst(I);
1438 void DFSanVisitor::visitAllocaInst(AllocaInst &I) {
1439 bool AllLoadsStores = true;
1440 for (User *U : I.users()) {
1441 if (isa<LoadInst>(U))
1442 continue;
1444 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
1445 if (SI->getPointerOperand() == &I)
1446 continue;
1449 AllLoadsStores = false;
1450 break;
1452 if (AllLoadsStores) {
1453 IRBuilder<> IRB(&I);
1454 DFSF.AllocaShadowMap[&I] = IRB.CreateAlloca(DFSF.DFS.ShadowTy);
1456 DFSF.setShadow(&I, DFSF.DFS.ZeroShadow);
1459 void DFSanVisitor::visitSelectInst(SelectInst &I) {
1460 Value *CondShadow = DFSF.getShadow(I.getCondition());
1461 Value *TrueShadow = DFSF.getShadow(I.getTrueValue());
1462 Value *FalseShadow = DFSF.getShadow(I.getFalseValue());
1464 if (isa<VectorType>(I.getCondition()->getType())) {
1465 DFSF.setShadow(
1467 DFSF.combineShadows(
1468 CondShadow, DFSF.combineShadows(TrueShadow, FalseShadow, &I), &I));
1469 } else {
1470 Value *ShadowSel;
1471 if (TrueShadow == FalseShadow) {
1472 ShadowSel = TrueShadow;
1473 } else {
1474 ShadowSel =
1475 SelectInst::Create(I.getCondition(), TrueShadow, FalseShadow, "", &I);
1477 DFSF.setShadow(&I, DFSF.combineShadows(CondShadow, ShadowSel, &I));
1481 void DFSanVisitor::visitMemSetInst(MemSetInst &I) {
1482 IRBuilder<> IRB(&I);
1483 Value *ValShadow = DFSF.getShadow(I.getValue());
1484 IRB.CreateCall(DFSF.DFS.DFSanSetLabelFn,
1485 {ValShadow, IRB.CreateBitCast(I.getDest(), Type::getInt8PtrTy(
1486 *DFSF.DFS.Ctx)),
1487 IRB.CreateZExtOrTrunc(I.getLength(), DFSF.DFS.IntptrTy)});
1490 void DFSanVisitor::visitMemTransferInst(MemTransferInst &I) {
1491 IRBuilder<> IRB(&I);
1492 Value *DestShadow = DFSF.DFS.getShadowAddress(I.getDest(), &I);
1493 Value *SrcShadow = DFSF.DFS.getShadowAddress(I.getSource(), &I);
1494 Value *LenShadow = IRB.CreateMul(
1495 I.getLength(),
1496 ConstantInt::get(I.getLength()->getType(), DFSF.DFS.ShadowWidth / 8));
1497 Type *Int8Ptr = Type::getInt8PtrTy(*DFSF.DFS.Ctx);
1498 DestShadow = IRB.CreateBitCast(DestShadow, Int8Ptr);
1499 SrcShadow = IRB.CreateBitCast(SrcShadow, Int8Ptr);
1500 auto *MTI = cast<MemTransferInst>(
1501 IRB.CreateCall(I.getFunctionType(), I.getCalledValue(),
1502 {DestShadow, SrcShadow, LenShadow, I.getVolatileCst()}));
1503 if (ClPreserveAlignment) {
1504 MTI->setDestAlignment(I.getDestAlignment() * (DFSF.DFS.ShadowWidth / 8));
1505 MTI->setSourceAlignment(I.getSourceAlignment() * (DFSF.DFS.ShadowWidth / 8));
1506 } else {
1507 MTI->setDestAlignment(DFSF.DFS.ShadowWidth / 8);
1508 MTI->setSourceAlignment(DFSF.DFS.ShadowWidth / 8);
1512 void DFSanVisitor::visitReturnInst(ReturnInst &RI) {
1513 if (!DFSF.IsNativeABI && RI.getReturnValue()) {
1514 switch (DFSF.IA) {
1515 case DataFlowSanitizer::IA_TLS: {
1516 Value *S = DFSF.getShadow(RI.getReturnValue());
1517 IRBuilder<> IRB(&RI);
1518 IRB.CreateStore(S, DFSF.getRetvalTLS());
1519 break;
1521 case DataFlowSanitizer::IA_Args: {
1522 IRBuilder<> IRB(&RI);
1523 Type *RT = DFSF.F->getFunctionType()->getReturnType();
1524 Value *InsVal =
1525 IRB.CreateInsertValue(UndefValue::get(RT), RI.getReturnValue(), 0);
1526 Value *InsShadow =
1527 IRB.CreateInsertValue(InsVal, DFSF.getShadow(RI.getReturnValue()), 1);
1528 RI.setOperand(0, InsShadow);
1529 break;
1535 void DFSanVisitor::visitCallSite(CallSite CS) {
1536 Function *F = CS.getCalledFunction();
1537 if ((F && F->isIntrinsic()) || isa<InlineAsm>(CS.getCalledValue())) {
1538 visitOperandShadowInst(*CS.getInstruction());
1539 return;
1542 // Calls to this function are synthesized in wrappers, and we shouldn't
1543 // instrument them.
1544 if (F == DFSF.DFS.DFSanVarargWrapperFn.getCallee()->stripPointerCasts())
1545 return;
1547 IRBuilder<> IRB(CS.getInstruction());
1549 DenseMap<Value *, Function *>::iterator i =
1550 DFSF.DFS.UnwrappedFnMap.find(CS.getCalledValue());
1551 if (i != DFSF.DFS.UnwrappedFnMap.end()) {
1552 Function *F = i->second;
1553 switch (DFSF.DFS.getWrapperKind(F)) {
1554 case DataFlowSanitizer::WK_Warning:
1555 CS.setCalledFunction(F);
1556 IRB.CreateCall(DFSF.DFS.DFSanUnimplementedFn,
1557 IRB.CreateGlobalStringPtr(F->getName()));
1558 DFSF.setShadow(CS.getInstruction(), DFSF.DFS.ZeroShadow);
1559 return;
1560 case DataFlowSanitizer::WK_Discard:
1561 CS.setCalledFunction(F);
1562 DFSF.setShadow(CS.getInstruction(), DFSF.DFS.ZeroShadow);
1563 return;
1564 case DataFlowSanitizer::WK_Functional:
1565 CS.setCalledFunction(F);
1566 visitOperandShadowInst(*CS.getInstruction());
1567 return;
1568 case DataFlowSanitizer::WK_Custom:
1569 // Don't try to handle invokes of custom functions, it's too complicated.
1570 // Instead, invoke the dfsw$ wrapper, which will in turn call the __dfsw_
1571 // wrapper.
1572 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) {
1573 FunctionType *FT = F->getFunctionType();
1574 TransformedFunction CustomFn = DFSF.DFS.getCustomFunctionType(FT);
1575 std::string CustomFName = "__dfsw_";
1576 CustomFName += F->getName();
1577 FunctionCallee CustomF = DFSF.DFS.Mod->getOrInsertFunction(
1578 CustomFName, CustomFn.TransformedType);
1579 if (Function *CustomFn = dyn_cast<Function>(CustomF.getCallee())) {
1580 CustomFn->copyAttributesFrom(F);
1582 // Custom functions returning non-void will write to the return label.
1583 if (!FT->getReturnType()->isVoidTy()) {
1584 CustomFn->removeAttributes(AttributeList::FunctionIndex,
1585 DFSF.DFS.ReadOnlyNoneAttrs);
1589 std::vector<Value *> Args;
1591 CallSite::arg_iterator i = CS.arg_begin();
1592 for (unsigned n = FT->getNumParams(); n != 0; ++i, --n) {
1593 Type *T = (*i)->getType();
1594 FunctionType *ParamFT;
1595 if (isa<PointerType>(T) &&
1596 (ParamFT = dyn_cast<FunctionType>(
1597 cast<PointerType>(T)->getElementType()))) {
1598 std::string TName = "dfst";
1599 TName += utostr(FT->getNumParams() - n);
1600 TName += "$";
1601 TName += F->getName();
1602 Constant *T = DFSF.DFS.getOrBuildTrampolineFunction(ParamFT, TName);
1603 Args.push_back(T);
1604 Args.push_back(
1605 IRB.CreateBitCast(*i, Type::getInt8PtrTy(*DFSF.DFS.Ctx)));
1606 } else {
1607 Args.push_back(*i);
1611 i = CS.arg_begin();
1612 const unsigned ShadowArgStart = Args.size();
1613 for (unsigned n = FT->getNumParams(); n != 0; ++i, --n)
1614 Args.push_back(DFSF.getShadow(*i));
1616 if (FT->isVarArg()) {
1617 auto *LabelVATy = ArrayType::get(DFSF.DFS.ShadowTy,
1618 CS.arg_size() - FT->getNumParams());
1619 auto *LabelVAAlloca = new AllocaInst(
1620 LabelVATy, getDataLayout().getAllocaAddrSpace(),
1621 "labelva", &DFSF.F->getEntryBlock().front());
1623 for (unsigned n = 0; i != CS.arg_end(); ++i, ++n) {
1624 auto LabelVAPtr = IRB.CreateStructGEP(LabelVATy, LabelVAAlloca, n);
1625 IRB.CreateStore(DFSF.getShadow(*i), LabelVAPtr);
1628 Args.push_back(IRB.CreateStructGEP(LabelVATy, LabelVAAlloca, 0));
1631 if (!FT->getReturnType()->isVoidTy()) {
1632 if (!DFSF.LabelReturnAlloca) {
1633 DFSF.LabelReturnAlloca =
1634 new AllocaInst(DFSF.DFS.ShadowTy,
1635 getDataLayout().getAllocaAddrSpace(),
1636 "labelreturn", &DFSF.F->getEntryBlock().front());
1638 Args.push_back(DFSF.LabelReturnAlloca);
1641 for (i = CS.arg_begin() + FT->getNumParams(); i != CS.arg_end(); ++i)
1642 Args.push_back(*i);
1644 CallInst *CustomCI = IRB.CreateCall(CustomF, Args);
1645 CustomCI->setCallingConv(CI->getCallingConv());
1646 CustomCI->setAttributes(TransformFunctionAttributes(CustomFn,
1647 CI->getContext(), CI->getAttributes()));
1649 // Update the parameter attributes of the custom call instruction to
1650 // zero extend the shadow parameters. This is required for targets
1651 // which consider ShadowTy an illegal type.
1652 for (unsigned n = 0; n < FT->getNumParams(); n++) {
1653 const unsigned ArgNo = ShadowArgStart + n;
1654 if (CustomCI->getArgOperand(ArgNo)->getType() == DFSF.DFS.ShadowTy)
1655 CustomCI->addParamAttr(ArgNo, Attribute::ZExt);
1658 if (!FT->getReturnType()->isVoidTy()) {
1659 LoadInst *LabelLoad =
1660 IRB.CreateLoad(DFSF.DFS.ShadowTy, DFSF.LabelReturnAlloca);
1661 DFSF.setShadow(CustomCI, LabelLoad);
1664 CI->replaceAllUsesWith(CustomCI);
1665 CI->eraseFromParent();
1666 return;
1668 break;
1672 FunctionType *FT = cast<FunctionType>(
1673 CS.getCalledValue()->getType()->getPointerElementType());
1674 if (DFSF.DFS.getInstrumentedABI() == DataFlowSanitizer::IA_TLS) {
1675 for (unsigned i = 0, n = FT->getNumParams(); i != n; ++i) {
1676 IRB.CreateStore(DFSF.getShadow(CS.getArgument(i)),
1677 DFSF.getArgTLS(i, CS.getInstruction()));
1681 Instruction *Next = nullptr;
1682 if (!CS.getType()->isVoidTy()) {
1683 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
1684 if (II->getNormalDest()->getSinglePredecessor()) {
1685 Next = &II->getNormalDest()->front();
1686 } else {
1687 BasicBlock *NewBB =
1688 SplitEdge(II->getParent(), II->getNormalDest(), &DFSF.DT);
1689 Next = &NewBB->front();
1691 } else {
1692 assert(CS->getIterator() != CS->getParent()->end());
1693 Next = CS->getNextNode();
1696 if (DFSF.DFS.getInstrumentedABI() == DataFlowSanitizer::IA_TLS) {
1697 IRBuilder<> NextIRB(Next);
1698 LoadInst *LI = NextIRB.CreateLoad(DFSF.DFS.ShadowTy, DFSF.getRetvalTLS());
1699 DFSF.SkipInsts.insert(LI);
1700 DFSF.setShadow(CS.getInstruction(), LI);
1701 DFSF.NonZeroChecks.push_back(LI);
1705 // Do all instrumentation for IA_Args down here to defer tampering with the
1706 // CFG in a way that SplitEdge may be able to detect.
1707 if (DFSF.DFS.getInstrumentedABI() == DataFlowSanitizer::IA_Args) {
1708 FunctionType *NewFT = DFSF.DFS.getArgsFunctionType(FT);
1709 Value *Func =
1710 IRB.CreateBitCast(CS.getCalledValue(), PointerType::getUnqual(NewFT));
1711 std::vector<Value *> Args;
1713 CallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
1714 for (unsigned n = FT->getNumParams(); n != 0; ++i, --n)
1715 Args.push_back(*i);
1717 i = CS.arg_begin();
1718 for (unsigned n = FT->getNumParams(); n != 0; ++i, --n)
1719 Args.push_back(DFSF.getShadow(*i));
1721 if (FT->isVarArg()) {
1722 unsigned VarArgSize = CS.arg_size() - FT->getNumParams();
1723 ArrayType *VarArgArrayTy = ArrayType::get(DFSF.DFS.ShadowTy, VarArgSize);
1724 AllocaInst *VarArgShadow =
1725 new AllocaInst(VarArgArrayTy, getDataLayout().getAllocaAddrSpace(),
1726 "", &DFSF.F->getEntryBlock().front());
1727 Args.push_back(IRB.CreateConstGEP2_32(VarArgArrayTy, VarArgShadow, 0, 0));
1728 for (unsigned n = 0; i != e; ++i, ++n) {
1729 IRB.CreateStore(
1730 DFSF.getShadow(*i),
1731 IRB.CreateConstGEP2_32(VarArgArrayTy, VarArgShadow, 0, n));
1732 Args.push_back(*i);
1736 CallSite NewCS;
1737 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
1738 NewCS = IRB.CreateInvoke(NewFT, Func, II->getNormalDest(),
1739 II->getUnwindDest(), Args);
1740 } else {
1741 NewCS = IRB.CreateCall(NewFT, Func, Args);
1743 NewCS.setCallingConv(CS.getCallingConv());
1744 NewCS.setAttributes(CS.getAttributes().removeAttributes(
1745 *DFSF.DFS.Ctx, AttributeList::ReturnIndex,
1746 AttributeFuncs::typeIncompatible(NewCS.getInstruction()->getType())));
1748 if (Next) {
1749 ExtractValueInst *ExVal =
1750 ExtractValueInst::Create(NewCS.getInstruction(), 0, "", Next);
1751 DFSF.SkipInsts.insert(ExVal);
1752 ExtractValueInst *ExShadow =
1753 ExtractValueInst::Create(NewCS.getInstruction(), 1, "", Next);
1754 DFSF.SkipInsts.insert(ExShadow);
1755 DFSF.setShadow(ExVal, ExShadow);
1756 DFSF.NonZeroChecks.push_back(ExShadow);
1758 CS.getInstruction()->replaceAllUsesWith(ExVal);
1761 CS.getInstruction()->eraseFromParent();
1765 void DFSanVisitor::visitPHINode(PHINode &PN) {
1766 PHINode *ShadowPN =
1767 PHINode::Create(DFSF.DFS.ShadowTy, PN.getNumIncomingValues(), "", &PN);
1769 // Give the shadow phi node valid predecessors to fool SplitEdge into working.
1770 Value *UndefShadow = UndefValue::get(DFSF.DFS.ShadowTy);
1771 for (PHINode::block_iterator i = PN.block_begin(), e = PN.block_end(); i != e;
1772 ++i) {
1773 ShadowPN->addIncoming(UndefShadow, *i);
1776 DFSF.PHIFixups.push_back(std::make_pair(&PN, ShadowPN));
1777 DFSF.setShadow(&PN, ShadowPN);