Bug Summary

File:src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Support/Alignment.h
Warning:line 85, column 47
The result of the left shift is undefined due to shifting by '255', which is greater or equal to the width of type 'uint64_t'

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -cc1 -triple amd64-unknown-openbsd7.0 -analyze -disable-free -disable-llvm-verifier -discard-value-names -main-file-name DataFlowSanitizer.cpp -analyzer-store=region -analyzer-opt-analyze-nested-blocks -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=cplusplus -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -setup-static-analyzer -mrelocation-model static -mframe-pointer=all -relaxed-aliasing -fno-rounding-math -mconstructor-aliases -munwind-tables -target-cpu x86-64 -tune-cpu generic -debugger-tuning=gdb -fcoverage-compilation-dir=/usr/src/gnu/usr.bin/clang/libLLVM/obj -resource-dir /usr/local/lib/clang/13.0.0 -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Transforms -I /usr/src/gnu/usr.bin/clang/libLLVM/obj/../include/llvm/AMDGPU -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/lib/Target/AMDGPU -I /usr/src/gnu/usr.bin/clang/libLLVM/obj/../include/llvm/AMDGPU -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/lib/Target/AMDGPU -I /usr/src/gnu/usr.bin/clang/libLLVM/obj/../include/llvm/AMDGPU -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/lib/Target/AMDGPU -I /usr/src/gnu/usr.bin/clang/libLLVM/obj/../include/llvm/AMDGPU -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/lib/Target/AMDGPU -I /usr/src/gnu/usr.bin/clang/libLLVM/obj/../include/llvm/AMDGPU -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/lib/Target/AMDGPU -I /usr/src/gnu/usr.bin/clang/libLLVM/obj/../include/llvm/AMDGPU -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/lib/Target/AMDGPU -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Analysis -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/ASMParser -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/BinaryFormat -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Bitcode -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Bitcode -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Bitstream -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Transforms -I /include/llvm/CodeGen -I /include/llvm/CodeGen/PBQP -I /usr/src/gnu/usr.bin/clang/libLLVM/obj/../include/llvm/IR -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/IR -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Transforms -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Transforms/Coroutines -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/ProfileData/Coverage -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/DebugInfo/CodeView -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/DebugInfo/DWARF -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/DebugInfo -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/DebugInfo/MSF -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/DebugInfo/PDB -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Demangle -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/ExecutionEngine -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/ExecutionEngine/JITLink -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/ExecutionEngine/Orc -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Frontend -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Frontend/OpenACC -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Frontend -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Frontend/OpenMP -I /include/llvm/CodeGen/GlobalISel -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/IRReader -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Transforms -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Transforms/InstCombine -I /usr/src/gnu/usr.bin/clang/libLLVM/obj/../include/llvm/Transforms/InstCombine -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Transforms -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/LTO -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Linker -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/MC -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/MC/MCParser -I /include/llvm/CodeGen/MIRParser -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Transforms -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Object -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Option -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Passes -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/ -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/ProfileData -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Transforms -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Transforms/Scalar -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/ADT -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Support -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/DebugInfo/Symbolize -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Target -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Transforms -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Transforms/Utils -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Transforms -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Transforms/Vectorize -I /usr/src/gnu/usr.bin/clang/libLLVM/obj/../include/llvm/X86 -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/lib/Target/X86 -I /usr/src/gnu/usr.bin/clang/libLLVM/obj/../include/llvm/X86 -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/lib/Target/X86 -I /usr/src/gnu/usr.bin/clang/libLLVM/obj/../include/llvm/X86 -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/lib/Target/X86 -I /usr/src/gnu/usr.bin/clang/libLLVM/obj/../include/llvm/X86 -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/lib/Target/X86 -I /usr/src/gnu/usr.bin/clang/libLLVM/obj/../include/llvm/X86 -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/lib/Target/X86 -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Transforms -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Transforms/IPO -I /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include -I /usr/src/gnu/usr.bin/clang/libLLVM/../include -I /usr/src/gnu/usr.bin/clang/libLLVM/obj -I /usr/src/gnu/usr.bin/clang/libLLVM/obj/../include -D NDEBUG -D __STDC_LIMIT_MACROS -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D LLVM_PREFIX="/usr" -internal-isystem /usr/include/c++/v1 -internal-isystem /usr/local/lib/clang/13.0.0/include -internal-externc-isystem /usr/include -O2 -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-comment -std=c++14 -fdeprecated-macro -fdebug-compilation-dir=/usr/src/gnu/usr.bin/clang/libLLVM/obj -ferror-limit 19 -fvisibility-inlines-hidden -fwrapv -stack-protector 2 -fno-rtti -fgnuc-version=4.2.1 -vectorize-loops -vectorize-slp -fno-builtin-malloc -fno-builtin-calloc -fno-builtin-realloc -fno-builtin-valloc -fno-builtin-free -fno-builtin-strdup -fno-builtin-strndup -analyzer-output=html -faddrsig -D__GCC_HAVE_DWARF2_CFI_ASM=1 -o /home/ben/Projects/vmm/scan-build/2022-01-12-194120-40624-1 -x c++ /usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/lib/Transforms/Instrumentation/DataFlowSanitizer.cpp

/usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/lib/Transforms/Instrumentation/DataFlowSanitizer.cpp

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.
20///
21/// Each byte of application memory is backed by a shadow memory byte. The
22/// shadow byte can represent up to 8 labels. On Linux/x86_64, memory is then
23/// laid out as follows:
24///
25/// +--------------------+ 0x800000000000 (top of memory)
26/// | application 3 |
27/// +--------------------+ 0x700000000000
28/// | invalid |
29/// +--------------------+ 0x610000000000
30/// | origin 1 |
31/// +--------------------+ 0x600000000000
32/// | application 2 |
33/// +--------------------+ 0x510000000000
34/// | shadow 1 |
35/// +--------------------+ 0x500000000000
36/// | invalid |
37/// +--------------------+ 0x400000000000
38/// | origin 3 |
39/// +--------------------+ 0x300000000000
40/// | shadow 3 |
41/// +--------------------+ 0x200000000000
42/// | origin 2 |
43/// +--------------------+ 0x110000000000
44/// | invalid |
45/// +--------------------+ 0x100000000000
46/// | shadow 2 |
47/// +--------------------+ 0x010000000000
48/// | application 1 |
49/// +--------------------+ 0x000000000000
50///
51/// MEM_TO_SHADOW(mem) = mem ^ 0x500000000000
52/// SHADOW_TO_ORIGIN(shadow) = shadow + 0x100000000000
53///
54/// For more information, please refer to the design document:
55/// http://clang.llvm.org/docs/DataFlowSanitizerDesign.html
56//
57//===----------------------------------------------------------------------===//
58
59#include "llvm/Transforms/Instrumentation/DataFlowSanitizer.h"
60#include "llvm/ADT/DenseMap.h"
61#include "llvm/ADT/DenseSet.h"
62#include "llvm/ADT/DepthFirstIterator.h"
63#include "llvm/ADT/None.h"
64#include "llvm/ADT/SmallPtrSet.h"
65#include "llvm/ADT/SmallVector.h"
66#include "llvm/ADT/StringExtras.h"
67#include "llvm/ADT/StringRef.h"
68#include "llvm/ADT/Triple.h"
69#include "llvm/ADT/iterator.h"
70#include "llvm/Analysis/ValueTracking.h"
71#include "llvm/IR/Argument.h"
72#include "llvm/IR/Attributes.h"
73#include "llvm/IR/BasicBlock.h"
74#include "llvm/IR/Constant.h"
75#include "llvm/IR/Constants.h"
76#include "llvm/IR/DataLayout.h"
77#include "llvm/IR/DerivedTypes.h"
78#include "llvm/IR/Dominators.h"
79#include "llvm/IR/Function.h"
80#include "llvm/IR/GlobalAlias.h"
81#include "llvm/IR/GlobalValue.h"
82#include "llvm/IR/GlobalVariable.h"
83#include "llvm/IR/IRBuilder.h"
84#include "llvm/IR/InlineAsm.h"
85#include "llvm/IR/InstVisitor.h"
86#include "llvm/IR/InstrTypes.h"
87#include "llvm/IR/Instruction.h"
88#include "llvm/IR/Instructions.h"
89#include "llvm/IR/IntrinsicInst.h"
90#include "llvm/IR/LLVMContext.h"
91#include "llvm/IR/MDBuilder.h"
92#include "llvm/IR/Module.h"
93#include "llvm/IR/PassManager.h"
94#include "llvm/IR/Type.h"
95#include "llvm/IR/User.h"
96#include "llvm/IR/Value.h"
97#include "llvm/InitializePasses.h"
98#include "llvm/Pass.h"
99#include "llvm/Support/Alignment.h"
100#include "llvm/Support/Casting.h"
101#include "llvm/Support/CommandLine.h"
102#include "llvm/Support/ErrorHandling.h"
103#include "llvm/Support/SpecialCaseList.h"
104#include "llvm/Support/VirtualFileSystem.h"
105#include "llvm/Transforms/Instrumentation.h"
106#include "llvm/Transforms/Utils/BasicBlockUtils.h"
107#include "llvm/Transforms/Utils/Local.h"
108#include <algorithm>
109#include <cassert>
110#include <cstddef>
111#include <cstdint>
112#include <iterator>
113#include <memory>
114#include <set>
115#include <string>
116#include <utility>
117#include <vector>
118
119using namespace llvm;
120
121// This must be consistent with ShadowWidthBits.
122static const Align ShadowTLSAlignment = Align(2);
123
124static const Align MinOriginAlignment = Align(4);
125
126// The size of TLS variables. These constants must be kept in sync with the ones
127// in dfsan.cpp.
128static const unsigned ArgTLSSize = 800;
129static const unsigned RetvalTLSSize = 800;
130
131// The -dfsan-preserve-alignment flag controls whether this pass assumes that
132// alignment requirements provided by the input IR are correct. For example,
133// if the input IR contains a load with alignment 8, this flag will cause
134// the shadow load to have alignment 16. This flag is disabled by default as
135// we have unfortunately encountered too much code (including Clang itself;
136// see PR14291) which performs misaligned access.
137static cl::opt<bool> ClPreserveAlignment(
138 "dfsan-preserve-alignment",
139 cl::desc("respect alignment requirements provided by input IR"), cl::Hidden,
140 cl::init(false));
141
142// The ABI list files control how shadow parameters are passed. The pass treats
143// every function labelled "uninstrumented" in the ABI list file as conforming
144// to the "native" (i.e. unsanitized) ABI. Unless the ABI list contains
145// additional annotations for those functions, a call to one of those functions
146// will produce a warning message, as the labelling behaviour of the function is
147// unknown. The other supported annotations are "functional" and "discard",
148// which are described below under DataFlowSanitizer::WrapperKind.
149static cl::list<std::string> ClABIListFiles(
150 "dfsan-abilist",
151 cl::desc("File listing native ABI functions and how the pass treats them"),
152 cl::Hidden);
153
154// Controls whether the pass uses IA_Args or IA_TLS as the ABI for instrumented
155// functions (see DataFlowSanitizer::InstrumentedABI below).
156static cl::opt<bool>
157 ClArgsABI("dfsan-args-abi",
158 cl::desc("Use the argument ABI rather than the TLS ABI"),
159 cl::Hidden);
160
161// Controls whether the pass includes or ignores the labels of pointers in load
162// instructions.
163static cl::opt<bool> ClCombinePointerLabelsOnLoad(
164 "dfsan-combine-pointer-labels-on-load",
165 cl::desc("Combine the label of the pointer with the label of the data when "
166 "loading from memory."),
167 cl::Hidden, cl::init(true));
168
169// Controls whether the pass includes or ignores the labels of pointers in
170// stores instructions.
171static cl::opt<bool> ClCombinePointerLabelsOnStore(
172 "dfsan-combine-pointer-labels-on-store",
173 cl::desc("Combine the label of the pointer with the label of the data when "
174 "storing in memory."),
175 cl::Hidden, cl::init(false));
176
177// Controls whether the pass propagates labels of offsets in GEP instructions.
178static cl::opt<bool> ClCombineOffsetLabelsOnGEP(
179 "dfsan-combine-offset-labels-on-gep",
180 cl::desc(
181 "Combine the label of the offset with the label of the pointer when "
182 "doing pointer arithmetic."),
183 cl::Hidden, cl::init(true));
184
185static cl::opt<bool> ClDebugNonzeroLabels(
186 "dfsan-debug-nonzero-labels",
187 cl::desc("Insert calls to __dfsan_nonzero_label on observing a parameter, "
188 "load or return with a nonzero label"),
189 cl::Hidden);
190
191// Experimental feature that inserts callbacks for certain data events.
192// Currently callbacks are only inserted for loads, stores, memory transfers
193// (i.e. memcpy and memmove), and comparisons.
194//
195// If this flag is set to true, the user must provide definitions for the
196// following callback functions:
197// void __dfsan_load_callback(dfsan_label Label, void* addr);
198// void __dfsan_store_callback(dfsan_label Label, void* addr);
199// void __dfsan_mem_transfer_callback(dfsan_label *Start, size_t Len);
200// void __dfsan_cmp_callback(dfsan_label CombinedLabel);
201static cl::opt<bool> ClEventCallbacks(
202 "dfsan-event-callbacks",
203 cl::desc("Insert calls to __dfsan_*_callback functions on data events."),
204 cl::Hidden, cl::init(false));
205
206// Controls whether the pass tracks the control flow of select instructions.
207static cl::opt<bool> ClTrackSelectControlFlow(
208 "dfsan-track-select-control-flow",
209 cl::desc("Propagate labels from condition values of select instructions "
210 "to results."),
211 cl::Hidden, cl::init(true));
212
213// TODO: This default value follows MSan. DFSan may use a different value.
214static cl::opt<int> ClInstrumentWithCallThreshold(
215 "dfsan-instrument-with-call-threshold",
216 cl::desc("If the function being instrumented requires more than "
217 "this number of origin stores, use callbacks instead of "
218 "inline checks (-1 means never use callbacks)."),
219 cl::Hidden, cl::init(3500));
220
221// Controls how to track origins.
222// * 0: do not track origins.
223// * 1: track origins at memory store operations.
224// * 2: track origins at memory load and store operations.
225// TODO: track callsites.
226static cl::opt<int> ClTrackOrigins("dfsan-track-origins",
227 cl::desc("Track origins of labels"),
228 cl::Hidden, cl::init(0));
229
230static StringRef getGlobalTypeString(const GlobalValue &G) {
231 // Types of GlobalVariables are always pointer types.
232 Type *GType = G.getValueType();
233 // For now we support excluding struct types only.
234 if (StructType *SGType = dyn_cast<StructType>(GType)) {
235 if (!SGType->isLiteral())
236 return SGType->getName();
237 }
238 return "<unknown type>";
239}
240
241namespace {
242
243// Memory map parameters used in application-to-shadow address calculation.
244// Offset = (Addr & ~AndMask) ^ XorMask
245// Shadow = ShadowBase + Offset
246// Origin = (OriginBase + Offset) & ~3ULL
247struct MemoryMapParams {
248 uint64_t AndMask;
249 uint64_t XorMask;
250 uint64_t ShadowBase;
251 uint64_t OriginBase;
252};
253
254} // end anonymous namespace
255
256// x86_64 Linux
257// NOLINTNEXTLINE(readability-identifier-naming)
258static const MemoryMapParams Linux_X86_64_MemoryMapParams = {
259 0, // AndMask (not used)
260 0x500000000000, // XorMask
261 0, // ShadowBase (not used)
262 0x100000000000, // OriginBase
263};
264
265namespace {
266
267class DFSanABIList {
268 std::unique_ptr<SpecialCaseList> SCL;
269
270public:
271 DFSanABIList() = default;
272
273 void set(std::unique_ptr<SpecialCaseList> List) { SCL = std::move(List); }
274
275 /// Returns whether either this function or its source file are listed in the
276 /// given category.
277 bool isIn(const Function &F, StringRef Category) const {
278 return isIn(*F.getParent(), Category) ||
279 SCL->inSection("dataflow", "fun", F.getName(), Category);
280 }
281
282 /// Returns whether this global alias is listed in the given category.
283 ///
284 /// If GA aliases a function, the alias's name is matched as a function name
285 /// would be. Similarly, aliases of globals are matched like globals.
286 bool isIn(const GlobalAlias &GA, StringRef Category) const {
287 if (isIn(*GA.getParent(), Category))
288 return true;
289
290 if (isa<FunctionType>(GA.getValueType()))
291 return SCL->inSection("dataflow", "fun", GA.getName(), Category);
292
293 return SCL->inSection("dataflow", "global", GA.getName(), Category) ||
294 SCL->inSection("dataflow", "type", getGlobalTypeString(GA),
295 Category);
296 }
297
298 /// Returns whether this module is listed in the given category.
299 bool isIn(const Module &M, StringRef Category) const {
300 return SCL->inSection("dataflow", "src", M.getModuleIdentifier(), Category);
301 }
302};
303
304/// TransformedFunction is used to express the result of transforming one
305/// function type into another. This struct is immutable. It holds metadata
306/// useful for updating calls of the old function to the new type.
307struct TransformedFunction {
308 TransformedFunction(FunctionType *OriginalType, FunctionType *TransformedType,
309 std::vector<unsigned> ArgumentIndexMapping)
310 : OriginalType(OriginalType), TransformedType(TransformedType),
311 ArgumentIndexMapping(ArgumentIndexMapping) {}
312
313 // Disallow copies.
314 TransformedFunction(const TransformedFunction &) = delete;
315 TransformedFunction &operator=(const TransformedFunction &) = delete;
316
317 // Allow moves.
318 TransformedFunction(TransformedFunction &&) = default;
319 TransformedFunction &operator=(TransformedFunction &&) = default;
320
321 /// Type of the function before the transformation.
322 FunctionType *OriginalType;
323
324 /// Type of the function after the transformation.
325 FunctionType *TransformedType;
326
327 /// Transforming a function may change the position of arguments. This
328 /// member records the mapping from each argument's old position to its new
329 /// position. Argument positions are zero-indexed. If the transformation
330 /// from F to F' made the first argument of F into the third argument of F',
331 /// then ArgumentIndexMapping[0] will equal 2.
332 std::vector<unsigned> ArgumentIndexMapping;
333};
334
335/// Given function attributes from a call site for the original function,
336/// return function attributes appropriate for a call to the transformed
337/// function.
338AttributeList
339transformFunctionAttributes(const TransformedFunction &TransformedFunction,
340 LLVMContext &Ctx, AttributeList CallSiteAttrs) {
341
342 // Construct a vector of AttributeSet for each function argument.
343 std::vector<llvm::AttributeSet> ArgumentAttributes(
344 TransformedFunction.TransformedType->getNumParams());
345
346 // Copy attributes from the parameter of the original function to the
347 // transformed version. 'ArgumentIndexMapping' holds the mapping from
348 // old argument position to new.
349 for (unsigned I = 0, IE = TransformedFunction.ArgumentIndexMapping.size();
350 I < IE; ++I) {
351 unsigned TransformedIndex = TransformedFunction.ArgumentIndexMapping[I];
352 ArgumentAttributes[TransformedIndex] = CallSiteAttrs.getParamAttributes(I);
353 }
354
355 // Copy annotations on varargs arguments.
356 for (unsigned I = TransformedFunction.OriginalType->getNumParams(),
357 IE = CallSiteAttrs.getNumAttrSets();
358 I < IE; ++I) {
359 ArgumentAttributes.push_back(CallSiteAttrs.getParamAttributes(I));
360 }
361
362 return AttributeList::get(Ctx, CallSiteAttrs.getFnAttributes(),
363 CallSiteAttrs.getRetAttributes(),
364 llvm::makeArrayRef(ArgumentAttributes));
365}
366
367class DataFlowSanitizer {
368 friend struct DFSanFunction;
369 friend class DFSanVisitor;
370
371 enum { ShadowWidthBits = 8, ShadowWidthBytes = ShadowWidthBits / 8 };
372
373 enum { OriginWidthBits = 32, OriginWidthBytes = OriginWidthBits / 8 };
374
375 /// Which ABI should be used for instrumented functions?
376 enum InstrumentedABI {
377 /// Argument and return value labels are passed through additional
378 /// arguments and by modifying the return type.
379 IA_Args,
380
381 /// Argument and return value labels are passed through TLS variables
382 /// __dfsan_arg_tls and __dfsan_retval_tls.
383 IA_TLS
384 };
385
386 /// How should calls to uninstrumented functions be handled?
387 enum WrapperKind {
388 /// This function is present in an uninstrumented form but we don't know
389 /// how it should be handled. Print a warning and call the function anyway.
390 /// Don't label the return value.
391 WK_Warning,
392
393 /// This function does not write to (user-accessible) memory, and its return
394 /// value is unlabelled.
395 WK_Discard,
396
397 /// This function does not write to (user-accessible) memory, and the label
398 /// of its return value is the union of the label of its arguments.
399 WK_Functional,
400
401 /// Instead of calling the function, a custom wrapper __dfsw_F is called,
402 /// where F is the name of the function. This function may wrap the
403 /// original function or provide its own implementation. This is similar to
404 /// the IA_Args ABI, except that IA_Args uses a struct return type to
405 /// pass the return value shadow in a register, while WK_Custom uses an
406 /// extra pointer argument to return the shadow. This allows the wrapped
407 /// form of the function type to be expressed in C.
408 WK_Custom
409 };
410
411 Module *Mod;
412 LLVMContext *Ctx;
413 Type *Int8Ptr;
414 IntegerType *OriginTy;
415 PointerType *OriginPtrTy;
416 ConstantInt *ZeroOrigin;
417 /// The shadow type for all primitive types and vector types.
418 IntegerType *PrimitiveShadowTy;
419 PointerType *PrimitiveShadowPtrTy;
420 IntegerType *IntptrTy;
421 ConstantInt *ZeroPrimitiveShadow;
422 Constant *ArgTLS;
423 ArrayType *ArgOriginTLSTy;
424 Constant *ArgOriginTLS;
425 Constant *RetvalTLS;
426 Constant *RetvalOriginTLS;
427 FunctionType *DFSanUnionLoadFnTy;
428 FunctionType *DFSanLoadLabelAndOriginFnTy;
429 FunctionType *DFSanUnimplementedFnTy;
430 FunctionType *DFSanSetLabelFnTy;
431 FunctionType *DFSanNonzeroLabelFnTy;
432 FunctionType *DFSanVarargWrapperFnTy;
433 FunctionType *DFSanCmpCallbackFnTy;
434 FunctionType *DFSanLoadStoreCallbackFnTy;
435 FunctionType *DFSanMemTransferCallbackFnTy;
436 FunctionType *DFSanChainOriginFnTy;
437 FunctionType *DFSanChainOriginIfTaintedFnTy;
438 FunctionType *DFSanMemOriginTransferFnTy;
439 FunctionType *DFSanMaybeStoreOriginFnTy;
440 FunctionCallee DFSanUnionLoadFn;
441 FunctionCallee DFSanLoadLabelAndOriginFn;
442 FunctionCallee DFSanUnimplementedFn;
443 FunctionCallee DFSanSetLabelFn;
444 FunctionCallee DFSanNonzeroLabelFn;
445 FunctionCallee DFSanVarargWrapperFn;
446 FunctionCallee DFSanLoadCallbackFn;
447 FunctionCallee DFSanStoreCallbackFn;
448 FunctionCallee DFSanMemTransferCallbackFn;
449 FunctionCallee DFSanCmpCallbackFn;
450 FunctionCallee DFSanChainOriginFn;
451 FunctionCallee DFSanChainOriginIfTaintedFn;
452 FunctionCallee DFSanMemOriginTransferFn;
453 FunctionCallee DFSanMaybeStoreOriginFn;
454 SmallPtrSet<Value *, 16> DFSanRuntimeFunctions;
455 MDNode *ColdCallWeights;
456 MDNode *OriginStoreWeights;
457 DFSanABIList ABIList;
458 DenseMap<Value *, Function *> UnwrappedFnMap;
459 AttrBuilder ReadOnlyNoneAttrs;
460
461 /// Memory map parameters used in calculation mapping application addresses
462 /// to shadow addresses and origin addresses.
463 const MemoryMapParams *MapParams;
464
465 Value *getShadowOffset(Value *Addr, IRBuilder<> &IRB);
466 Value *getShadowAddress(Value *Addr, Instruction *Pos);
467 Value *getShadowAddress(Value *Addr, Instruction *Pos, Value *ShadowOffset);
468 std::pair<Value *, Value *>
469 getShadowOriginAddress(Value *Addr, Align InstAlignment, Instruction *Pos);
470 bool isInstrumented(const Function *F);
471 bool isInstrumented(const GlobalAlias *GA);
472 FunctionType *getArgsFunctionType(FunctionType *T);
473 FunctionType *getTrampolineFunctionType(FunctionType *T);
474 TransformedFunction getCustomFunctionType(FunctionType *T);
475 InstrumentedABI getInstrumentedABI();
476 WrapperKind getWrapperKind(Function *F);
477 void addGlobalNameSuffix(GlobalValue *GV);
478 Function *buildWrapperFunction(Function *F, StringRef NewFName,
479 GlobalValue::LinkageTypes NewFLink,
480 FunctionType *NewFT);
481 Constant *getOrBuildTrampolineFunction(FunctionType *FT, StringRef FName);
482 void initializeCallbackFunctions(Module &M);
483 void initializeRuntimeFunctions(Module &M);
484 void injectMetadataGlobals(Module &M);
485 bool initializeModule(Module &M);
486
487 /// Advances \p OriginAddr to point to the next 32-bit origin and then loads
488 /// from it. Returns the origin's loaded value.
489 Value *loadNextOrigin(Instruction *Pos, Align OriginAlign,
490 Value **OriginAddr);
491
492 /// Returns whether the given load byte size is amenable to inlined
493 /// optimization patterns.
494 bool hasLoadSizeForFastPath(uint64_t Size);
495
496 /// Returns whether the pass tracks origins. Supports only TLS ABI mode.
497 bool shouldTrackOrigins();
498
499 /// Returns whether the pass tracks labels for struct fields and array
500 /// indices. Supports only TLS ABI mode.
501 bool shouldTrackFieldsAndIndices();
502
503 /// Returns a zero constant with the shadow type of OrigTy.
504 ///
505 /// getZeroShadow({T1,T2,...}) = {getZeroShadow(T1),getZeroShadow(T2,...}
506 /// getZeroShadow([n x T]) = [n x getZeroShadow(T)]
507 /// getZeroShadow(other type) = i16(0)
508 ///
509 /// Note that a zero shadow is always i16(0) when shouldTrackFieldsAndIndices
510 /// returns false.
511 Constant *getZeroShadow(Type *OrigTy);
512 /// Returns a zero constant with the shadow type of V's type.
513 Constant *getZeroShadow(Value *V);
514
515 /// Checks if V is a zero shadow.
516 bool isZeroShadow(Value *V);
517
518 /// Returns the shadow type of OrigTy.
519 ///
520 /// getShadowTy({T1,T2,...}) = {getShadowTy(T1),getShadowTy(T2),...}
521 /// getShadowTy([n x T]) = [n x getShadowTy(T)]
522 /// getShadowTy(other type) = i16
523 ///
524 /// Note that a shadow type is always i16 when shouldTrackFieldsAndIndices
525 /// returns false.
526 Type *getShadowTy(Type *OrigTy);
527 /// Returns the shadow type of of V's type.
528 Type *getShadowTy(Value *V);
529
530 const uint64_t NumOfElementsInArgOrgTLS = ArgTLSSize / OriginWidthBytes;
531
532public:
533 DataFlowSanitizer(const std::vector<std::string> &ABIListFiles);
534
535 bool runImpl(Module &M);
536};
537
538struct DFSanFunction {
539 DataFlowSanitizer &DFS;
540 Function *F;
541 DominatorTree DT;
542 DataFlowSanitizer::InstrumentedABI IA;
543 bool IsNativeABI;
544 AllocaInst *LabelReturnAlloca = nullptr;
545 AllocaInst *OriginReturnAlloca = nullptr;
546 DenseMap<Value *, Value *> ValShadowMap;
547 DenseMap<Value *, Value *> ValOriginMap;
548 DenseMap<AllocaInst *, AllocaInst *> AllocaShadowMap;
549 DenseMap<AllocaInst *, AllocaInst *> AllocaOriginMap;
550
551 struct PHIFixupElement {
552 PHINode *Phi;
553 PHINode *ShadowPhi;
554 PHINode *OriginPhi;
555 };
556 std::vector<PHIFixupElement> PHIFixups;
557
558 DenseSet<Instruction *> SkipInsts;
559 std::vector<Value *> NonZeroChecks;
560
561 struct CachedShadow {
562 BasicBlock *Block; // The block where Shadow is defined.
563 Value *Shadow;
564 };
565 /// Maps a value to its latest shadow value in terms of domination tree.
566 DenseMap<std::pair<Value *, Value *>, CachedShadow> CachedShadows;
567 /// Maps a value to its latest collapsed shadow value it was converted to in
568 /// terms of domination tree. When ClDebugNonzeroLabels is on, this cache is
569 /// used at a post process where CFG blocks are split. So it does not cache
570 /// BasicBlock like CachedShadows, but uses domination between values.
571 DenseMap<Value *, Value *> CachedCollapsedShadows;
572 DenseMap<Value *, std::set<Value *>> ShadowElements;
573
574 DFSanFunction(DataFlowSanitizer &DFS, Function *F, bool IsNativeABI)
575 : DFS(DFS), F(F), IA(DFS.getInstrumentedABI()), IsNativeABI(IsNativeABI) {
576 DT.recalculate(*F);
577 }
578
579 /// Computes the shadow address for a given function argument.
580 ///
581 /// Shadow = ArgTLS+ArgOffset.
582 Value *getArgTLS(Type *T, unsigned ArgOffset, IRBuilder<> &IRB);
583
584 /// Computes the shadow address for a return value.
585 Value *getRetvalTLS(Type *T, IRBuilder<> &IRB);
586
587 /// Computes the origin address for a given function argument.
588 ///
589 /// Origin = ArgOriginTLS[ArgNo].
590 Value *getArgOriginTLS(unsigned ArgNo, IRBuilder<> &IRB);
591
592 /// Computes the origin address for a return value.
593 Value *getRetvalOriginTLS();
594
595 Value *getOrigin(Value *V);
596 void setOrigin(Instruction *I, Value *Origin);
597 /// Generates IR to compute the origin of the last operand with a taint label.
598 Value *combineOperandOrigins(Instruction *Inst);
599 /// Before the instruction Pos, generates IR to compute the last origin with a
600 /// taint label. Labels and origins are from vectors Shadows and Origins
601 /// correspondingly. The generated IR is like
602 /// Sn-1 != Zero ? On-1: ... S2 != Zero ? O2: S1 != Zero ? O1: O0
603 /// When Zero is nullptr, it uses ZeroPrimitiveShadow. Otherwise it can be
604 /// zeros with other bitwidths.
605 Value *combineOrigins(const std::vector<Value *> &Shadows,
606 const std::vector<Value *> &Origins, Instruction *Pos,
607 ConstantInt *Zero = nullptr);
608
609 Value *getShadow(Value *V);
610 void setShadow(Instruction *I, Value *Shadow);
611 /// Generates IR to compute the union of the two given shadows, inserting it
612 /// before Pos. The combined value is with primitive type.
613 Value *combineShadows(Value *V1, Value *V2, Instruction *Pos);
614 /// Combines the shadow values of V1 and V2, then converts the combined value
615 /// with primitive type into a shadow value with the original type T.
616 Value *combineShadowsThenConvert(Type *T, Value *V1, Value *V2,
617 Instruction *Pos);
618 Value *combineOperandShadows(Instruction *Inst);
619
620 /// Generates IR to load shadow and origin corresponding to bytes [\p
621 /// Addr, \p Addr + \p Size), where addr has alignment \p
622 /// InstAlignment, and take the union of each of those shadows. The returned
623 /// shadow always has primitive type.
624 ///
625 /// When tracking loads is enabled, the returned origin is a chain at the
626 /// current stack if the returned shadow is tainted.
627 std::pair<Value *, Value *> loadShadowOrigin(Value *Addr, uint64_t Size,
628 Align InstAlignment,
629 Instruction *Pos);
630
631 void storePrimitiveShadowOrigin(Value *Addr, uint64_t Size,
632 Align InstAlignment, Value *PrimitiveShadow,
633 Value *Origin, Instruction *Pos);
634 /// Applies PrimitiveShadow to all primitive subtypes of T, returning
635 /// the expanded shadow value.
636 ///
637 /// EFP({T1,T2, ...}, PS) = {EFP(T1,PS),EFP(T2,PS),...}
638 /// EFP([n x T], PS) = [n x EFP(T,PS)]
639 /// EFP(other types, PS) = PS
640 Value *expandFromPrimitiveShadow(Type *T, Value *PrimitiveShadow,
641 Instruction *Pos);
642 /// Collapses Shadow into a single primitive shadow value, unioning all
643 /// primitive shadow values in the process. Returns the final primitive
644 /// shadow value.
645 ///
646 /// CTP({V1,V2, ...}) = UNION(CFP(V1,PS),CFP(V2,PS),...)
647 /// CTP([V1,V2,...]) = UNION(CFP(V1,PS),CFP(V2,PS),...)
648 /// CTP(other types, PS) = PS
649 Value *collapseToPrimitiveShadow(Value *Shadow, Instruction *Pos);
650
651 void storeZeroPrimitiveShadow(Value *Addr, uint64_t Size, Align ShadowAlign,
652 Instruction *Pos);
653
654 Align getShadowAlign(Align InstAlignment);
655
656private:
657 /// Collapses the shadow with aggregate type into a single primitive shadow
658 /// value.
659 template <class AggregateType>
660 Value *collapseAggregateShadow(AggregateType *AT, Value *Shadow,
661 IRBuilder<> &IRB);
662
663 Value *collapseToPrimitiveShadow(Value *Shadow, IRBuilder<> &IRB);
664
665 /// Returns the shadow value of an argument A.
666 Value *getShadowForTLSArgument(Argument *A);
667
668 /// The fast path of loading shadows.
669 std::pair<Value *, Value *>
670 loadShadowFast(Value *ShadowAddr, Value *OriginAddr, uint64_t Size,
671 Align ShadowAlign, Align OriginAlign, Value *FirstOrigin,
672 Instruction *Pos);
673
674 Align getOriginAlign(Align InstAlignment);
675
676 /// Because 4 contiguous bytes share one 4-byte origin, the most accurate load
677 /// is __dfsan_load_label_and_origin. This function returns the union of all
678 /// labels and the origin of the first taint label. However this is an
679 /// additional call with many instructions. To ensure common cases are fast,
680 /// checks if it is possible to load labels and origins without using the
681 /// callback function.
682 ///
683 /// When enabling tracking load instructions, we always use
684 /// __dfsan_load_label_and_origin to reduce code size.
685 bool useCallbackLoadLabelAndOrigin(uint64_t Size, Align InstAlignment);
686
687 /// Returns a chain at the current stack with previous origin V.
688 Value *updateOrigin(Value *V, IRBuilder<> &IRB);
689
690 /// Returns a chain at the current stack with previous origin V if Shadow is
691 /// tainted.
692 Value *updateOriginIfTainted(Value *Shadow, Value *Origin, IRBuilder<> &IRB);
693
694 /// Creates an Intptr = Origin | Origin << 32 if Intptr's size is 64. Returns
695 /// Origin otherwise.
696 Value *originToIntptr(IRBuilder<> &IRB, Value *Origin);
697
698 /// Stores Origin into the address range [StoreOriginAddr, StoreOriginAddr +
699 /// Size).
700 void paintOrigin(IRBuilder<> &IRB, Value *Origin, Value *StoreOriginAddr,
701 uint64_t StoreOriginSize, Align Alignment);
702
703 /// Stores Origin in terms of its Shadow value.
704 /// * Do not write origins for zero shadows because we do not trace origins
705 /// for untainted sinks.
706 /// * Use __dfsan_maybe_store_origin if there are too many origin store
707 /// instrumentations.
708 void storeOrigin(Instruction *Pos, Value *Addr, uint64_t Size, Value *Shadow,
709 Value *Origin, Value *StoreOriginAddr, Align InstAlignment);
710
711 /// Convert a scalar value to an i1 by comparing with 0.
712 Value *convertToBool(Value *V, IRBuilder<> &IRB, const Twine &Name = "");
713
714 bool shouldInstrumentWithCall();
715
716 /// Generates IR to load shadow and origin corresponding to bytes [\p
717 /// Addr, \p Addr + \p Size), where addr has alignment \p
718 /// InstAlignment, and take the union of each of those shadows. The returned
719 /// shadow always has primitive type.
720 std::pair<Value *, Value *>
721 loadShadowOriginSansLoadTracking(Value *Addr, uint64_t Size,
722 Align InstAlignment, Instruction *Pos);
723 int NumOriginStores = 0;
724};
725
726class DFSanVisitor : public InstVisitor<DFSanVisitor> {
727public:
728 DFSanFunction &DFSF;
729
730 DFSanVisitor(DFSanFunction &DFSF) : DFSF(DFSF) {}
731
732 const DataLayout &getDataLayout() const {
733 return DFSF.F->getParent()->getDataLayout();
734 }
735
736 // Combines shadow values and origins for all of I's operands.
737 void visitInstOperands(Instruction &I);
738
739 void visitUnaryOperator(UnaryOperator &UO);
740 void visitBinaryOperator(BinaryOperator &BO);
741 void visitBitCastInst(BitCastInst &BCI);
742 void visitCastInst(CastInst &CI);
743 void visitCmpInst(CmpInst &CI);
744 void visitLandingPadInst(LandingPadInst &LPI);
745 void visitGetElementPtrInst(GetElementPtrInst &GEPI);
746 void visitLoadInst(LoadInst &LI);
747 void visitStoreInst(StoreInst &SI);
748 void visitAtomicRMWInst(AtomicRMWInst &I);
749 void visitAtomicCmpXchgInst(AtomicCmpXchgInst &I);
750 void visitReturnInst(ReturnInst &RI);
751 void visitCallBase(CallBase &CB);
752 void visitPHINode(PHINode &PN);
753 void visitExtractElementInst(ExtractElementInst &I);
754 void visitInsertElementInst(InsertElementInst &I);
755 void visitShuffleVectorInst(ShuffleVectorInst &I);
756 void visitExtractValueInst(ExtractValueInst &I);
757 void visitInsertValueInst(InsertValueInst &I);
758 void visitAllocaInst(AllocaInst &I);
759 void visitSelectInst(SelectInst &I);
760 void visitMemSetInst(MemSetInst &I);
761 void visitMemTransferInst(MemTransferInst &I);
762
763private:
764 void visitCASOrRMW(Align InstAlignment, Instruction &I);
765
766 // Returns false when this is an invoke of a custom function.
767 bool visitWrappedCallBase(Function &F, CallBase &CB);
768
769 // Combines origins for all of I's operands.
770 void visitInstOperandOrigins(Instruction &I);
771
772 void addShadowArguments(Function &F, CallBase &CB, std::vector<Value *> &Args,
773 IRBuilder<> &IRB);
774
775 void addOriginArguments(Function &F, CallBase &CB, std::vector<Value *> &Args,
776 IRBuilder<> &IRB);
777};
778
779} // end anonymous namespace
780
781DataFlowSanitizer::DataFlowSanitizer(
782 const std::vector<std::string> &ABIListFiles) {
783 std::vector<std::string> AllABIListFiles(std::move(ABIListFiles));
784 llvm::append_range(AllABIListFiles, ClABIListFiles);
785 // FIXME: should we propagate vfs::FileSystem to this constructor?
786 ABIList.set(
787 SpecialCaseList::createOrDie(AllABIListFiles, *vfs::getRealFileSystem()));
788}
789
790FunctionType *DataFlowSanitizer::getArgsFunctionType(FunctionType *T) {
791 SmallVector<Type *, 4> ArgTypes(T->param_begin(), T->param_end());
792 ArgTypes.append(T->getNumParams(), PrimitiveShadowTy);
793 if (T->isVarArg())
794 ArgTypes.push_back(PrimitiveShadowPtrTy);
795 Type *RetType = T->getReturnType();
796 if (!RetType->isVoidTy())
797 RetType = StructType::get(RetType, PrimitiveShadowTy);
798 return FunctionType::get(RetType, ArgTypes, T->isVarArg());
799}
800
801FunctionType *DataFlowSanitizer::getTrampolineFunctionType(FunctionType *T) {
802 assert(!T->isVarArg())((void)0);
803 SmallVector<Type *, 4> ArgTypes;
804 ArgTypes.push_back(T->getPointerTo());
805 ArgTypes.append(T->param_begin(), T->param_end());
806 ArgTypes.append(T->getNumParams(), PrimitiveShadowTy);
807 Type *RetType = T->getReturnType();
808 if (!RetType->isVoidTy())
809 ArgTypes.push_back(PrimitiveShadowPtrTy);
810
811 if (shouldTrackOrigins()) {
812 ArgTypes.append(T->getNumParams(), OriginTy);
813 if (!RetType->isVoidTy())
814 ArgTypes.push_back(OriginPtrTy);
815 }
816
817 return FunctionType::get(T->getReturnType(), ArgTypes, false);
818}
819
820TransformedFunction DataFlowSanitizer::getCustomFunctionType(FunctionType *T) {
821 SmallVector<Type *, 4> ArgTypes;
822
823 // Some parameters of the custom function being constructed are
824 // parameters of T. Record the mapping from parameters of T to
825 // parameters of the custom function, so that parameter attributes
826 // at call sites can be updated.
827 std::vector<unsigned> ArgumentIndexMapping;
828 for (unsigned I = 0, E = T->getNumParams(); I != E; ++I) {
829 Type *ParamType = T->getParamType(I);
830 FunctionType *FT;
831 if (isa<PointerType>(ParamType) &&
832 (FT = dyn_cast<FunctionType>(ParamType->getPointerElementType()))) {
833 ArgumentIndexMapping.push_back(ArgTypes.size());
834 ArgTypes.push_back(getTrampolineFunctionType(FT)->getPointerTo());
835 ArgTypes.push_back(Type::getInt8PtrTy(*Ctx));
836 } else {
837 ArgumentIndexMapping.push_back(ArgTypes.size());
838 ArgTypes.push_back(ParamType);
839 }
840 }
841 for (unsigned I = 0, E = T->getNumParams(); I != E; ++I)
842 ArgTypes.push_back(PrimitiveShadowTy);
843 if (T->isVarArg())
844 ArgTypes.push_back(PrimitiveShadowPtrTy);
845 Type *RetType = T->getReturnType();
846 if (!RetType->isVoidTy())
847 ArgTypes.push_back(PrimitiveShadowPtrTy);
848
849 if (shouldTrackOrigins()) {
850 for (unsigned I = 0, E = T->getNumParams(); I != E; ++I)
851 ArgTypes.push_back(OriginTy);
852 if (T->isVarArg())
853 ArgTypes.push_back(OriginPtrTy);
854 if (!RetType->isVoidTy())
855 ArgTypes.push_back(OriginPtrTy);
856 }
857
858 return TransformedFunction(
859 T, FunctionType::get(T->getReturnType(), ArgTypes, T->isVarArg()),
860 ArgumentIndexMapping);
861}
862
863bool DataFlowSanitizer::isZeroShadow(Value *V) {
864 if (!shouldTrackFieldsAndIndices())
865 return ZeroPrimitiveShadow == V;
866
867 Type *T = V->getType();
868 if (!isa<ArrayType>(T) && !isa<StructType>(T)) {
869 if (const ConstantInt *CI = dyn_cast<ConstantInt>(V))
870 return CI->isZero();
871 return false;
872 }
873
874 return isa<ConstantAggregateZero>(V);
875}
876
877bool DataFlowSanitizer::hasLoadSizeForFastPath(uint64_t Size) {
878 uint64_t ShadowSize = Size * ShadowWidthBytes;
879 return ShadowSize % 8 == 0 || ShadowSize == 4;
880}
881
882bool DataFlowSanitizer::shouldTrackOrigins() {
883 static const bool ShouldTrackOrigins =
884 ClTrackOrigins && getInstrumentedABI() == DataFlowSanitizer::IA_TLS;
885 return ShouldTrackOrigins;
886}
887
888bool DataFlowSanitizer::shouldTrackFieldsAndIndices() {
889 return getInstrumentedABI() == DataFlowSanitizer::IA_TLS;
890}
891
892Constant *DataFlowSanitizer::getZeroShadow(Type *OrigTy) {
893 if (!shouldTrackFieldsAndIndices())
894 return ZeroPrimitiveShadow;
895
896 if (!isa<ArrayType>(OrigTy) && !isa<StructType>(OrigTy))
897 return ZeroPrimitiveShadow;
898 Type *ShadowTy = getShadowTy(OrigTy);
899 return ConstantAggregateZero::get(ShadowTy);
900}
901
902Constant *DataFlowSanitizer::getZeroShadow(Value *V) {
903 return getZeroShadow(V->getType());
904}
905
906static Value *expandFromPrimitiveShadowRecursive(
907 Value *Shadow, SmallVector<unsigned, 4> &Indices, Type *SubShadowTy,
908 Value *PrimitiveShadow, IRBuilder<> &IRB) {
909 if (!isa<ArrayType>(SubShadowTy) && !isa<StructType>(SubShadowTy))
910 return IRB.CreateInsertValue(Shadow, PrimitiveShadow, Indices);
911
912 if (ArrayType *AT = dyn_cast<ArrayType>(SubShadowTy)) {
913 for (unsigned Idx = 0; Idx < AT->getNumElements(); Idx++) {
914 Indices.push_back(Idx);
915 Shadow = expandFromPrimitiveShadowRecursive(
916 Shadow, Indices, AT->getElementType(), PrimitiveShadow, IRB);
917 Indices.pop_back();
918 }
919 return Shadow;
920 }
921
922 if (StructType *ST = dyn_cast<StructType>(SubShadowTy)) {
923 for (unsigned Idx = 0; Idx < ST->getNumElements(); Idx++) {
924 Indices.push_back(Idx);
925 Shadow = expandFromPrimitiveShadowRecursive(
926 Shadow, Indices, ST->getElementType(Idx), PrimitiveShadow, IRB);
927 Indices.pop_back();
928 }
929 return Shadow;
930 }
931 llvm_unreachable("Unexpected shadow type")__builtin_unreachable();
932}
933
934bool DFSanFunction::shouldInstrumentWithCall() {
935 return ClInstrumentWithCallThreshold >= 0 &&
936 NumOriginStores >= ClInstrumentWithCallThreshold;
937}
938
939Value *DFSanFunction::expandFromPrimitiveShadow(Type *T, Value *PrimitiveShadow,
940 Instruction *Pos) {
941 Type *ShadowTy = DFS.getShadowTy(T);
942
943 if (!isa<ArrayType>(ShadowTy) && !isa<StructType>(ShadowTy))
944 return PrimitiveShadow;
945
946 if (DFS.isZeroShadow(PrimitiveShadow))
947 return DFS.getZeroShadow(ShadowTy);
948
949 IRBuilder<> IRB(Pos);
950 SmallVector<unsigned, 4> Indices;
951 Value *Shadow = UndefValue::get(ShadowTy);
952 Shadow = expandFromPrimitiveShadowRecursive(Shadow, Indices, ShadowTy,
953 PrimitiveShadow, IRB);
954
955 // Caches the primitive shadow value that built the shadow value.
956 CachedCollapsedShadows[Shadow] = PrimitiveShadow;
957 return Shadow;
958}
959
960template <class AggregateType>
961Value *DFSanFunction::collapseAggregateShadow(AggregateType *AT, Value *Shadow,
962 IRBuilder<> &IRB) {
963 if (!AT->getNumElements())
964 return DFS.ZeroPrimitiveShadow;
965
966 Value *FirstItem = IRB.CreateExtractValue(Shadow, 0);
967 Value *Aggregator = collapseToPrimitiveShadow(FirstItem, IRB);
968
969 for (unsigned Idx = 1; Idx < AT->getNumElements(); Idx++) {
970 Value *ShadowItem = IRB.CreateExtractValue(Shadow, Idx);
971 Value *ShadowInner = collapseToPrimitiveShadow(ShadowItem, IRB);
972 Aggregator = IRB.CreateOr(Aggregator, ShadowInner);
973 }
974 return Aggregator;
975}
976
977Value *DFSanFunction::collapseToPrimitiveShadow(Value *Shadow,
978 IRBuilder<> &IRB) {
979 Type *ShadowTy = Shadow->getType();
980 if (!isa<ArrayType>(ShadowTy) && !isa<StructType>(ShadowTy))
981 return Shadow;
982 if (ArrayType *AT = dyn_cast<ArrayType>(ShadowTy))
983 return collapseAggregateShadow<>(AT, Shadow, IRB);
984 if (StructType *ST = dyn_cast<StructType>(ShadowTy))
985 return collapseAggregateShadow<>(ST, Shadow, IRB);
986 llvm_unreachable("Unexpected shadow type")__builtin_unreachable();
987}
988
989Value *DFSanFunction::collapseToPrimitiveShadow(Value *Shadow,
990 Instruction *Pos) {
991 Type *ShadowTy = Shadow->getType();
992 if (!isa<ArrayType>(ShadowTy) && !isa<StructType>(ShadowTy))
993 return Shadow;
994
995 assert(DFS.shouldTrackFieldsAndIndices())((void)0);
996
997 // Checks if the cached collapsed shadow value dominates Pos.
998 Value *&CS = CachedCollapsedShadows[Shadow];
999 if (CS && DT.dominates(CS, Pos))
1000 return CS;
1001
1002 IRBuilder<> IRB(Pos);
1003 Value *PrimitiveShadow = collapseToPrimitiveShadow(Shadow, IRB);
1004 // Caches the converted primitive shadow value.
1005 CS = PrimitiveShadow;
1006 return PrimitiveShadow;
1007}
1008
1009Type *DataFlowSanitizer::getShadowTy(Type *OrigTy) {
1010 if (!shouldTrackFieldsAndIndices())
1011 return PrimitiveShadowTy;
1012
1013 if (!OrigTy->isSized())
1014 return PrimitiveShadowTy;
1015 if (isa<IntegerType>(OrigTy))
1016 return PrimitiveShadowTy;
1017 if (isa<VectorType>(OrigTy))
1018 return PrimitiveShadowTy;
1019 if (ArrayType *AT = dyn_cast<ArrayType>(OrigTy))
1020 return ArrayType::get(getShadowTy(AT->getElementType()),
1021 AT->getNumElements());
1022 if (StructType *ST = dyn_cast<StructType>(OrigTy)) {
1023 SmallVector<Type *, 4> Elements;
1024 for (unsigned I = 0, N = ST->getNumElements(); I < N; ++I)
1025 Elements.push_back(getShadowTy(ST->getElementType(I)));
1026 return StructType::get(*Ctx, Elements);
1027 }
1028 return PrimitiveShadowTy;
1029}
1030
1031Type *DataFlowSanitizer::getShadowTy(Value *V) {
1032 return getShadowTy(V->getType());
1033}
1034
1035bool DataFlowSanitizer::initializeModule(Module &M) {
1036 Triple TargetTriple(M.getTargetTriple());
1037 const DataLayout &DL = M.getDataLayout();
1038
1039 if (TargetTriple.getOS() != Triple::Linux)
1040 report_fatal_error("unsupported operating system");
1041 if (TargetTriple.getArch() != Triple::x86_64)
1042 report_fatal_error("unsupported architecture");
1043 MapParams = &Linux_X86_64_MemoryMapParams;
1044
1045 Mod = &M;
1046 Ctx = &M.getContext();
1047 Int8Ptr = Type::getInt8PtrTy(*Ctx);
1048 OriginTy = IntegerType::get(*Ctx, OriginWidthBits);
1049 OriginPtrTy = PointerType::getUnqual(OriginTy);
1050 PrimitiveShadowTy = IntegerType::get(*Ctx, ShadowWidthBits);
1051 PrimitiveShadowPtrTy = PointerType::getUnqual(PrimitiveShadowTy);
1052 IntptrTy = DL.getIntPtrType(*Ctx);
1053 ZeroPrimitiveShadow = ConstantInt::getSigned(PrimitiveShadowTy, 0);
1054 ZeroOrigin = ConstantInt::getSigned(OriginTy, 0);
1055
1056 Type *DFSanUnionLoadArgs[2] = {PrimitiveShadowPtrTy, IntptrTy};
1057 DFSanUnionLoadFnTy = FunctionType::get(PrimitiveShadowTy, DFSanUnionLoadArgs,
1058 /*isVarArg=*/false);
1059 Type *DFSanLoadLabelAndOriginArgs[2] = {Int8Ptr, IntptrTy};
1060 DFSanLoadLabelAndOriginFnTy =
1061 FunctionType::get(IntegerType::get(*Ctx, 64), DFSanLoadLabelAndOriginArgs,
1062 /*isVarArg=*/false);
1063 DFSanUnimplementedFnTy = FunctionType::get(
1064 Type::getVoidTy(*Ctx), Type::getInt8PtrTy(*Ctx), /*isVarArg=*/false);
1065 Type *DFSanSetLabelArgs[4] = {PrimitiveShadowTy, OriginTy,
1066 Type::getInt8PtrTy(*Ctx), IntptrTy};
1067 DFSanSetLabelFnTy = FunctionType::get(Type::getVoidTy(*Ctx),
1068 DFSanSetLabelArgs, /*isVarArg=*/false);
1069 DFSanNonzeroLabelFnTy =
1070 FunctionType::get(Type::getVoidTy(*Ctx), None, /*isVarArg=*/false);
1071 DFSanVarargWrapperFnTy = FunctionType::get(
1072 Type::getVoidTy(*Ctx), Type::getInt8PtrTy(*Ctx), /*isVarArg=*/false);
1073 DFSanCmpCallbackFnTy =
1074 FunctionType::get(Type::getVoidTy(*Ctx), PrimitiveShadowTy,
1075 /*isVarArg=*/false);
1076 DFSanChainOriginFnTy =
1077 FunctionType::get(OriginTy, OriginTy, /*isVarArg=*/false);
1078 Type *DFSanChainOriginIfTaintedArgs[2] = {PrimitiveShadowTy, OriginTy};
1079 DFSanChainOriginIfTaintedFnTy = FunctionType::get(
1080 OriginTy, DFSanChainOriginIfTaintedArgs, /*isVarArg=*/false);
1081 Type *DFSanMaybeStoreOriginArgs[4] = {IntegerType::get(*Ctx, ShadowWidthBits),
1082 Int8Ptr, IntptrTy, OriginTy};
1083 DFSanMaybeStoreOriginFnTy = FunctionType::get(
1084 Type::getVoidTy(*Ctx), DFSanMaybeStoreOriginArgs, /*isVarArg=*/false);
1085 Type *DFSanMemOriginTransferArgs[3] = {Int8Ptr, Int8Ptr, IntptrTy};
1086 DFSanMemOriginTransferFnTy = FunctionType::get(
1087 Type::getVoidTy(*Ctx), DFSanMemOriginTransferArgs, /*isVarArg=*/false);
1088 Type *DFSanLoadStoreCallbackArgs[2] = {PrimitiveShadowTy, Int8Ptr};
1089 DFSanLoadStoreCallbackFnTy =
1090 FunctionType::get(Type::getVoidTy(*Ctx), DFSanLoadStoreCallbackArgs,
1091 /*isVarArg=*/false);
1092 Type *DFSanMemTransferCallbackArgs[2] = {PrimitiveShadowPtrTy, IntptrTy};
1093 DFSanMemTransferCallbackFnTy =
1094 FunctionType::get(Type::getVoidTy(*Ctx), DFSanMemTransferCallbackArgs,
1095 /*isVarArg=*/false);
1096
1097 ColdCallWeights = MDBuilder(*Ctx).createBranchWeights(1, 1000);
1098 OriginStoreWeights = MDBuilder(*Ctx).createBranchWeights(1, 1000);
1099 return true;
1100}
1101
1102bool DataFlowSanitizer::isInstrumented(const Function *F) {
1103 return !ABIList.isIn(*F, "uninstrumented");
1104}
1105
1106bool DataFlowSanitizer::isInstrumented(const GlobalAlias *GA) {
1107 return !ABIList.isIn(*GA, "uninstrumented");
1108}
1109
1110DataFlowSanitizer::InstrumentedABI DataFlowSanitizer::getInstrumentedABI() {
1111 return ClArgsABI ? IA_Args : IA_TLS;
1112}
1113
1114DataFlowSanitizer::WrapperKind DataFlowSanitizer::getWrapperKind(Function *F) {
1115 if (ABIList.isIn(*F, "functional"))
1116 return WK_Functional;
1117 if (ABIList.isIn(*F, "discard"))
1118 return WK_Discard;
1119 if (ABIList.isIn(*F, "custom"))
1120 return WK_Custom;
1121
1122 return WK_Warning;
1123}
1124
1125void DataFlowSanitizer::addGlobalNameSuffix(GlobalValue *GV) {
1126 std::string GVName = std::string(GV->getName()), Suffix = ".dfsan";
1127 GV->setName(GVName + Suffix);
1128
1129 // Try to change the name of the function in module inline asm. We only do
1130 // this for specific asm directives, currently only ".symver", to try to avoid
1131 // corrupting asm which happens to contain the symbol name as a substring.
1132 // Note that the substitution for .symver assumes that the versioned symbol
1133 // also has an instrumented name.
1134 std::string Asm = GV->getParent()->getModuleInlineAsm();
1135 std::string SearchStr = ".symver " + GVName + ",";
1136 size_t Pos = Asm.find(SearchStr);
1137 if (Pos != std::string::npos) {
1138 Asm.replace(Pos, SearchStr.size(), ".symver " + GVName + Suffix + ",");
1139 Pos = Asm.find("@");
1140
1141 if (Pos == std::string::npos)
1142 report_fatal_error("unsupported .symver: " + Asm);
1143
1144 Asm.replace(Pos, 1, Suffix + "@");
1145 GV->getParent()->setModuleInlineAsm(Asm);
1146 }
1147}
1148
1149Function *
1150DataFlowSanitizer::buildWrapperFunction(Function *F, StringRef NewFName,
1151 GlobalValue::LinkageTypes NewFLink,
1152 FunctionType *NewFT) {
1153 FunctionType *FT = F->getFunctionType();
1154 Function *NewF = Function::Create(NewFT, NewFLink, F->getAddressSpace(),
1155 NewFName, F->getParent());
1156 NewF->copyAttributesFrom(F);
1157 NewF->removeAttributes(
1158 AttributeList::ReturnIndex,
1159 AttributeFuncs::typeIncompatible(NewFT->getReturnType()));
1160
1161 BasicBlock *BB = BasicBlock::Create(*Ctx, "entry", NewF);
1162 if (F->isVarArg()) {
1163 NewF->removeAttributes(AttributeList::FunctionIndex,
1164 AttrBuilder().addAttribute("split-stack"));
1165 CallInst::Create(DFSanVarargWrapperFn,
1166 IRBuilder<>(BB).CreateGlobalStringPtr(F->getName()), "",
1167 BB);
1168 new UnreachableInst(*Ctx, BB);
1169 } else {
1170 auto ArgIt = pointer_iterator<Argument *>(NewF->arg_begin());
1171 std::vector<Value *> Args(ArgIt, ArgIt + FT->getNumParams());
1172
1173 CallInst *CI = CallInst::Create(F, Args, "", BB);
1174 if (FT->getReturnType()->isVoidTy())
1175 ReturnInst::Create(*Ctx, BB);
1176 else
1177 ReturnInst::Create(*Ctx, CI, BB);
1178 }
1179
1180 return NewF;
1181}
1182
1183Constant *DataFlowSanitizer::getOrBuildTrampolineFunction(FunctionType *FT,
1184 StringRef FName) {
1185 FunctionType *FTT = getTrampolineFunctionType(FT);
1186 FunctionCallee C = Mod->getOrInsertFunction(FName, FTT);
1187 Function *F = dyn_cast<Function>(C.getCallee());
1188 if (F && F->isDeclaration()) {
1189 F->setLinkage(GlobalValue::LinkOnceODRLinkage);
1190 BasicBlock *BB = BasicBlock::Create(*Ctx, "entry", F);
1191 std::vector<Value *> Args;
1192 Function::arg_iterator AI = F->arg_begin() + 1;
1193 for (unsigned N = FT->getNumParams(); N != 0; ++AI, --N)
1194 Args.push_back(&*AI);
1195 CallInst *CI = CallInst::Create(FT, &*F->arg_begin(), Args, "", BB);
1196 Type *RetType = FT->getReturnType();
1197 ReturnInst *RI = RetType->isVoidTy() ? ReturnInst::Create(*Ctx, BB)
1198 : ReturnInst::Create(*Ctx, CI, BB);
1199
1200 // F is called by a wrapped custom function with primitive shadows. So
1201 // its arguments and return value need conversion.
1202 DFSanFunction DFSF(*this, F, /*IsNativeABI=*/true);
1203 Function::arg_iterator ValAI = F->arg_begin(), ShadowAI = AI;
1204 ++ValAI;
1205 for (unsigned N = FT->getNumParams(); N != 0; ++ValAI, ++ShadowAI, --N) {
1206 Value *Shadow =
1207 DFSF.expandFromPrimitiveShadow(ValAI->getType(), &*ShadowAI, CI);
1208 DFSF.ValShadowMap[&*ValAI] = Shadow;
1209 }
1210 Function::arg_iterator RetShadowAI = ShadowAI;
1211 const bool ShouldTrackOrigins = shouldTrackOrigins();
1212 if (ShouldTrackOrigins) {
1213 ValAI = F->arg_begin();
1214 ++ValAI;
1215 Function::arg_iterator OriginAI = ShadowAI;
1216 if (!RetType->isVoidTy())
1217 ++OriginAI;
1218 for (unsigned N = FT->getNumParams(); N != 0; ++ValAI, ++OriginAI, --N) {
1219 DFSF.ValOriginMap[&*ValAI] = &*OriginAI;
1220 }
1221 }
1222 DFSanVisitor(DFSF).visitCallInst(*CI);
1223 if (!RetType->isVoidTy()) {
1224 Value *PrimitiveShadow = DFSF.collapseToPrimitiveShadow(
1225 DFSF.getShadow(RI->getReturnValue()), RI);
1226 new StoreInst(PrimitiveShadow, &*RetShadowAI, RI);
1227 if (ShouldTrackOrigins) {
1228 Value *Origin = DFSF.getOrigin(RI->getReturnValue());
1229 new StoreInst(Origin, &*std::prev(F->arg_end()), RI);
1230 }
1231 }
1232 }
1233
1234 return cast<Constant>(C.getCallee());
1235}
1236
1237// Initialize DataFlowSanitizer runtime functions and declare them in the module
1238void DataFlowSanitizer::initializeRuntimeFunctions(Module &M) {
1239 {
1240 AttributeList AL;
1241 AL = AL.addAttribute(M.getContext(), AttributeList::FunctionIndex,
1242 Attribute::NoUnwind);
1243 AL = AL.addAttribute(M.getContext(), AttributeList::FunctionIndex,
1244 Attribute::ReadOnly);
1245 AL = AL.addAttribute(M.getContext(), AttributeList::ReturnIndex,
1246 Attribute::ZExt);
1247 DFSanUnionLoadFn =
1248 Mod->getOrInsertFunction("__dfsan_union_load", DFSanUnionLoadFnTy, AL);
1249 }
1250 {
1251 AttributeList AL;
1252 AL = AL.addAttribute(M.getContext(), AttributeList::FunctionIndex,
1253 Attribute::NoUnwind);
1254 AL = AL.addAttribute(M.getContext(), AttributeList::FunctionIndex,
1255 Attribute::ReadOnly);
1256 AL = AL.addAttribute(M.getContext(), AttributeList::ReturnIndex,
1257 Attribute::ZExt);
1258 DFSanLoadLabelAndOriginFn = Mod->getOrInsertFunction(
1259 "__dfsan_load_label_and_origin", DFSanLoadLabelAndOriginFnTy, AL);
1260 }
1261 DFSanUnimplementedFn =
1262 Mod->getOrInsertFunction("__dfsan_unimplemented", DFSanUnimplementedFnTy);
1263 {
1264 AttributeList AL;
1265 AL = AL.addParamAttribute(M.getContext(), 0, Attribute::ZExt);
1266 AL = AL.addParamAttribute(M.getContext(), 1, Attribute::ZExt);
1267 DFSanSetLabelFn =
1268 Mod->getOrInsertFunction("__dfsan_set_label", DFSanSetLabelFnTy, AL);
1269 }
1270 DFSanNonzeroLabelFn =
1271 Mod->getOrInsertFunction("__dfsan_nonzero_label", DFSanNonzeroLabelFnTy);
1272 DFSanVarargWrapperFn = Mod->getOrInsertFunction("__dfsan_vararg_wrapper",
1273 DFSanVarargWrapperFnTy);
1274 {
1275 AttributeList AL;
1276 AL = AL.addParamAttribute(M.getContext(), 0, Attribute::ZExt);
1277 AL = AL.addAttribute(M.getContext(), AttributeList::ReturnIndex,
1278 Attribute::ZExt);
1279 DFSanChainOriginFn = Mod->getOrInsertFunction("__dfsan_chain_origin",
1280 DFSanChainOriginFnTy, AL);
1281 }
1282 {
1283 AttributeList AL;
1284 AL = AL.addParamAttribute(M.getContext(), 0, Attribute::ZExt);
1285 AL = AL.addParamAttribute(M.getContext(), 1, Attribute::ZExt);
1286 AL = AL.addAttribute(M.getContext(), AttributeList::ReturnIndex,
1287 Attribute::ZExt);
1288 DFSanChainOriginIfTaintedFn = Mod->getOrInsertFunction(
1289 "__dfsan_chain_origin_if_tainted", DFSanChainOriginIfTaintedFnTy, AL);
1290 }
1291 DFSanMemOriginTransferFn = Mod->getOrInsertFunction(
1292 "__dfsan_mem_origin_transfer", DFSanMemOriginTransferFnTy);
1293
1294 {
1295 AttributeList AL;
1296 AL = AL.addParamAttribute(M.getContext(), 0, Attribute::ZExt);
1297 AL = AL.addParamAttribute(M.getContext(), 3, Attribute::ZExt);
1298 DFSanMaybeStoreOriginFn = Mod->getOrInsertFunction(
1299 "__dfsan_maybe_store_origin", DFSanMaybeStoreOriginFnTy, AL);
1300 }
1301
1302 DFSanRuntimeFunctions.insert(
1303 DFSanUnionLoadFn.getCallee()->stripPointerCasts());
1304 DFSanRuntimeFunctions.insert(
1305 DFSanLoadLabelAndOriginFn.getCallee()->stripPointerCasts());
1306 DFSanRuntimeFunctions.insert(
1307 DFSanUnimplementedFn.getCallee()->stripPointerCasts());
1308 DFSanRuntimeFunctions.insert(
1309 DFSanSetLabelFn.getCallee()->stripPointerCasts());
1310 DFSanRuntimeFunctions.insert(
1311 DFSanNonzeroLabelFn.getCallee()->stripPointerCasts());
1312 DFSanRuntimeFunctions.insert(
1313 DFSanVarargWrapperFn.getCallee()->stripPointerCasts());
1314 DFSanRuntimeFunctions.insert(
1315 DFSanLoadCallbackFn.getCallee()->stripPointerCasts());
1316 DFSanRuntimeFunctions.insert(
1317 DFSanStoreCallbackFn.getCallee()->stripPointerCasts());
1318 DFSanRuntimeFunctions.insert(
1319 DFSanMemTransferCallbackFn.getCallee()->stripPointerCasts());
1320 DFSanRuntimeFunctions.insert(
1321 DFSanCmpCallbackFn.getCallee()->stripPointerCasts());
1322 DFSanRuntimeFunctions.insert(
1323 DFSanChainOriginFn.getCallee()->stripPointerCasts());
1324 DFSanRuntimeFunctions.insert(
1325 DFSanChainOriginIfTaintedFn.getCallee()->stripPointerCasts());
1326 DFSanRuntimeFunctions.insert(
1327 DFSanMemOriginTransferFn.getCallee()->stripPointerCasts());
1328 DFSanRuntimeFunctions.insert(
1329 DFSanMaybeStoreOriginFn.getCallee()->stripPointerCasts());
1330}
1331
1332// Initializes event callback functions and declare them in the module
1333void DataFlowSanitizer::initializeCallbackFunctions(Module &M) {
1334 DFSanLoadCallbackFn = Mod->getOrInsertFunction("__dfsan_load_callback",
1335 DFSanLoadStoreCallbackFnTy);
1336 DFSanStoreCallbackFn = Mod->getOrInsertFunction("__dfsan_store_callback",
1337 DFSanLoadStoreCallbackFnTy);
1338 DFSanMemTransferCallbackFn = Mod->getOrInsertFunction(
1339 "__dfsan_mem_transfer_callback", DFSanMemTransferCallbackFnTy);
1340 DFSanCmpCallbackFn =
1341 Mod->getOrInsertFunction("__dfsan_cmp_callback", DFSanCmpCallbackFnTy);
1342}
1343
1344void DataFlowSanitizer::injectMetadataGlobals(Module &M) {
1345 // These variables can be used:
1346 // - by the runtime (to discover what the shadow width was, during
1347 // compilation)
1348 // - in testing (to avoid hardcoding the shadow width and type but instead
1349 // extract them by pattern matching)
1350 Type *IntTy = Type::getInt32Ty(*Ctx);
1351 (void)Mod->getOrInsertGlobal("__dfsan_shadow_width_bits", IntTy, [&] {
1352 return new GlobalVariable(
1353 M, IntTy, /*isConstant=*/true, GlobalValue::WeakODRLinkage,
1354 ConstantInt::get(IntTy, ShadowWidthBits), "__dfsan_shadow_width_bits");
1355 });
1356 (void)Mod->getOrInsertGlobal("__dfsan_shadow_width_bytes", IntTy, [&] {
1357 return new GlobalVariable(M, IntTy, /*isConstant=*/true,
1358 GlobalValue::WeakODRLinkage,
1359 ConstantInt::get(IntTy, ShadowWidthBytes),
1360 "__dfsan_shadow_width_bytes");
1361 });
1362}
1363
1364bool DataFlowSanitizer::runImpl(Module &M) {
1365 initializeModule(M);
1366
1367 if (ABIList.isIn(M, "skip"))
1368 return false;
1369
1370 const unsigned InitialGlobalSize = M.global_size();
1371 const unsigned InitialModuleSize = M.size();
1372
1373 bool Changed = false;
1374
1375 auto GetOrInsertGlobal = [this, &Changed](StringRef Name,
1376 Type *Ty) -> Constant * {
1377 Constant *C = Mod->getOrInsertGlobal(Name, Ty);
1378 if (GlobalVariable *G = dyn_cast<GlobalVariable>(C)) {
1379 Changed |= G->getThreadLocalMode() != GlobalVariable::InitialExecTLSModel;
1380 G->setThreadLocalMode(GlobalVariable::InitialExecTLSModel);
1381 }
1382 return C;
1383 };
1384
1385 // These globals must be kept in sync with the ones in dfsan.cpp.
1386 ArgTLS =
1387 GetOrInsertGlobal("__dfsan_arg_tls",
1388 ArrayType::get(Type::getInt64Ty(*Ctx), ArgTLSSize / 8));
1389 RetvalTLS = GetOrInsertGlobal(
1390 "__dfsan_retval_tls",
1391 ArrayType::get(Type::getInt64Ty(*Ctx), RetvalTLSSize / 8));
1392 ArgOriginTLSTy = ArrayType::get(OriginTy, NumOfElementsInArgOrgTLS);
1393 ArgOriginTLS = GetOrInsertGlobal("__dfsan_arg_origin_tls", ArgOriginTLSTy);
1394 RetvalOriginTLS = GetOrInsertGlobal("__dfsan_retval_origin_tls", OriginTy);
1395
1396 (void)Mod->getOrInsertGlobal("__dfsan_track_origins", OriginTy, [&] {
1397 Changed = true;
1398 return new GlobalVariable(
1399 M, OriginTy, true, GlobalValue::WeakODRLinkage,
1400 ConstantInt::getSigned(OriginTy,
1401 shouldTrackOrigins() ? ClTrackOrigins : 0),
1402 "__dfsan_track_origins");
1403 });
1404
1405 injectMetadataGlobals(M);
1406
1407 initializeCallbackFunctions(M);
1408 initializeRuntimeFunctions(M);
1409
1410 std::vector<Function *> FnsToInstrument;
1411 SmallPtrSet<Function *, 2> FnsWithNativeABI;
1412 for (Function &F : M)
1413 if (!F.isIntrinsic() && !DFSanRuntimeFunctions.contains(&F))
1414 FnsToInstrument.push_back(&F);
1415
1416 // Give function aliases prefixes when necessary, and build wrappers where the
1417 // instrumentedness is inconsistent.
1418 for (Module::alias_iterator AI = M.alias_begin(), AE = M.alias_end();
1419 AI != AE;) {
1420 GlobalAlias *GA = &*AI;
1421 ++AI;
1422 // Don't stop on weak. We assume people aren't playing games with the
1423 // instrumentedness of overridden weak aliases.
1424 auto *F = dyn_cast<Function>(GA->getBaseObject());
1425 if (!F)
1426 continue;
1427
1428 bool GAInst = isInstrumented(GA), FInst = isInstrumented(F);
1429 if (GAInst && FInst) {
1430 addGlobalNameSuffix(GA);
1431 } else if (GAInst != FInst) {
1432 // Non-instrumented alias of an instrumented function, or vice versa.
1433 // Replace the alias with a native-ABI wrapper of the aliasee. The pass
1434 // below will take care of instrumenting it.
1435 Function *NewF =
1436 buildWrapperFunction(F, "", GA->getLinkage(), F->getFunctionType());
1437 GA->replaceAllUsesWith(ConstantExpr::getBitCast(NewF, GA->getType()));
1438 NewF->takeName(GA);
1439 GA->eraseFromParent();
1440 FnsToInstrument.push_back(NewF);
1441 }
1442 }
1443
1444 ReadOnlyNoneAttrs.addAttribute(Attribute::ReadOnly)
1445 .addAttribute(Attribute::ReadNone);
1446
1447 // First, change the ABI of every function in the module. ABI-listed
1448 // functions keep their original ABI and get a wrapper function.
1449 for (std::vector<Function *>::iterator FI = FnsToInstrument.begin(),
1450 FE = FnsToInstrument.end();
1451 FI != FE; ++FI) {
1452 Function &F = **FI;
1453 FunctionType *FT = F.getFunctionType();
1454
1455 bool IsZeroArgsVoidRet = (FT->getNumParams() == 0 && !FT->isVarArg() &&
1456 FT->getReturnType()->isVoidTy());
1457
1458 if (isInstrumented(&F)) {
1459 // Instrumented functions get a '.dfsan' suffix. This allows us to more
1460 // easily identify cases of mismatching ABIs. This naming scheme is
1461 // mangling-compatible (see Itanium ABI), using a vendor-specific suffix.
1462 if (getInstrumentedABI() == IA_Args && !IsZeroArgsVoidRet) {
1463 FunctionType *NewFT = getArgsFunctionType(FT);
1464 Function *NewF = Function::Create(NewFT, F.getLinkage(),
1465 F.getAddressSpace(), "", &M);
1466 NewF->copyAttributesFrom(&F);
1467 NewF->removeAttributes(
1468 AttributeList::ReturnIndex,
1469 AttributeFuncs::typeIncompatible(NewFT->getReturnType()));
1470 for (Function::arg_iterator FArg = F.arg_begin(),
1471 NewFArg = NewF->arg_begin(),
1472 FArgEnd = F.arg_end();
1473 FArg != FArgEnd; ++FArg, ++NewFArg) {
1474 FArg->replaceAllUsesWith(&*NewFArg);
1475 }
1476 NewF->getBasicBlockList().splice(NewF->begin(), F.getBasicBlockList());
1477
1478 for (Function::user_iterator UI = F.user_begin(), UE = F.user_end();
1479 UI != UE;) {
1480 BlockAddress *BA = dyn_cast<BlockAddress>(*UI);
1481 ++UI;
1482 if (BA) {
1483 BA->replaceAllUsesWith(
1484 BlockAddress::get(NewF, BA->getBasicBlock()));
1485 delete BA;
1486 }
1487 }
1488 F.replaceAllUsesWith(
1489 ConstantExpr::getBitCast(NewF, PointerType::getUnqual(FT)));
1490 NewF->takeName(&F);
1491 F.eraseFromParent();
1492 *FI = NewF;
1493 addGlobalNameSuffix(NewF);
1494 } else {
1495 addGlobalNameSuffix(&F);
1496 }
1497 } else if (!IsZeroArgsVoidRet || getWrapperKind(&F) == WK_Custom) {
1498 // Build a wrapper function for F. The wrapper simply calls F, and is
1499 // added to FnsToInstrument so that any instrumentation according to its
1500 // WrapperKind is done in the second pass below.
1501 FunctionType *NewFT =
1502 getInstrumentedABI() == IA_Args ? getArgsFunctionType(FT) : FT;
1503
1504 // If the function being wrapped has local linkage, then preserve the
1505 // function's linkage in the wrapper function.
1506 GlobalValue::LinkageTypes WrapperLinkage =
1507 F.hasLocalLinkage() ? F.getLinkage()
1508 : GlobalValue::LinkOnceODRLinkage;
1509
1510 Function *NewF = buildWrapperFunction(
1511 &F,
1512 (shouldTrackOrigins() ? std::string("dfso$") : std::string("dfsw$")) +
1513 std::string(F.getName()),
1514 WrapperLinkage, NewFT);
1515 if (getInstrumentedABI() == IA_TLS)
1516 NewF->removeAttributes(AttributeList::FunctionIndex, ReadOnlyNoneAttrs);
1517
1518 Value *WrappedFnCst =
1519 ConstantExpr::getBitCast(NewF, PointerType::getUnqual(FT));
1520 F.replaceAllUsesWith(WrappedFnCst);
1521
1522 UnwrappedFnMap[WrappedFnCst] = &F;
1523 *FI = NewF;
1524
1525 if (!F.isDeclaration()) {
1526 // This function is probably defining an interposition of an
1527 // uninstrumented function and hence needs to keep the original ABI.
1528 // But any functions it may call need to use the instrumented ABI, so
1529 // we instrument it in a mode which preserves the original ABI.
1530 FnsWithNativeABI.insert(&F);
1531
1532 // This code needs to rebuild the iterators, as they may be invalidated
1533 // by the push_back, taking care that the new range does not include
1534 // any functions added by this code.
1535 size_t N = FI - FnsToInstrument.begin(),
1536 Count = FE - FnsToInstrument.begin();
1537 FnsToInstrument.push_back(&F);
1538 FI = FnsToInstrument.begin() + N;
1539 FE = FnsToInstrument.begin() + Count;
1540 }
1541 // Hopefully, nobody will try to indirectly call a vararg
1542 // function... yet.
1543 } else if (FT->isVarArg()) {
1544 UnwrappedFnMap[&F] = &F;
1545 *FI = nullptr;
1546 }
1547 }
1548
1549 for (Function *F : FnsToInstrument) {
1550 if (!F || F->isDeclaration())
1551 continue;
1552
1553 removeUnreachableBlocks(*F);
1554
1555 DFSanFunction DFSF(*this, F, FnsWithNativeABI.count(F));
1556
1557 // DFSanVisitor may create new basic blocks, which confuses df_iterator.
1558 // Build a copy of the list before iterating over it.
1559 SmallVector<BasicBlock *, 4> BBList(depth_first(&F->getEntryBlock()));
1560
1561 for (BasicBlock *BB : BBList) {
1562 Instruction *Inst = &BB->front();
1563 while (true) {
1564 // DFSanVisitor may split the current basic block, changing the current
1565 // instruction's next pointer and moving the next instruction to the
1566 // tail block from which we should continue.
1567 Instruction *Next = Inst->getNextNode();
1568 // DFSanVisitor may delete Inst, so keep track of whether it was a
1569 // terminator.
1570 bool IsTerminator = Inst->isTerminator();
1571 if (!DFSF.SkipInsts.count(Inst))
1572 DFSanVisitor(DFSF).visit(Inst);
1573 if (IsTerminator)
1574 break;
1575 Inst = Next;
1576 }
1577 }
1578
1579 // We will not necessarily be able to compute the shadow for every phi node
1580 // until we have visited every block. Therefore, the code that handles phi
1581 // nodes adds them to the PHIFixups list so that they can be properly
1582 // handled here.
1583 for (DFSanFunction::PHIFixupElement &P : DFSF.PHIFixups) {
1584 for (unsigned Val = 0, N = P.Phi->getNumIncomingValues(); Val != N;
1585 ++Val) {
1586 P.ShadowPhi->setIncomingValue(
1587 Val, DFSF.getShadow(P.Phi->getIncomingValue(Val)));
1588 if (P.OriginPhi)
1589 P.OriginPhi->setIncomingValue(
1590 Val, DFSF.getOrigin(P.Phi->getIncomingValue(Val)));
1591 }
1592 }
1593
1594 // -dfsan-debug-nonzero-labels will split the CFG in all kinds of crazy
1595 // places (i.e. instructions in basic blocks we haven't even begun visiting
1596 // yet). To make our life easier, do this work in a pass after the main
1597 // instrumentation.
1598 if (ClDebugNonzeroLabels) {
1599 for (Value *V : DFSF.NonZeroChecks) {
1600 Instruction *Pos;
1601 if (Instruction *I = dyn_cast<Instruction>(V))
1602 Pos = I->getNextNode();
1603 else
1604 Pos = &DFSF.F->getEntryBlock().front();
1605 while (isa<PHINode>(Pos) || isa<AllocaInst>(Pos))
1606 Pos = Pos->getNextNode();
1607 IRBuilder<> IRB(Pos);
1608 Value *PrimitiveShadow = DFSF.collapseToPrimitiveShadow(V, Pos);
1609 Value *Ne =
1610 IRB.CreateICmpNE(PrimitiveShadow, DFSF.DFS.ZeroPrimitiveShadow);
1611 BranchInst *BI = cast<BranchInst>(SplitBlockAndInsertIfThen(
1612 Ne, Pos, /*Unreachable=*/false, ColdCallWeights));
1613 IRBuilder<> ThenIRB(BI);
1614 ThenIRB.CreateCall(DFSF.DFS.DFSanNonzeroLabelFn, {});
1615 }
1616 }
1617 }
1618
1619 return Changed || !FnsToInstrument.empty() ||
1620 M.global_size() != InitialGlobalSize || M.size() != InitialModuleSize;
1621}
1622
1623Value *DFSanFunction::getArgTLS(Type *T, unsigned ArgOffset, IRBuilder<> &IRB) {
1624 Value *Base = IRB.CreatePointerCast(DFS.ArgTLS, DFS.IntptrTy);
1625 if (ArgOffset)
1626 Base = IRB.CreateAdd(Base, ConstantInt::get(DFS.IntptrTy, ArgOffset));
1627 return IRB.CreateIntToPtr(Base, PointerType::get(DFS.getShadowTy(T), 0),
1628 "_dfsarg");
1629}
1630
1631Value *DFSanFunction::getRetvalTLS(Type *T, IRBuilder<> &IRB) {
1632 return IRB.CreatePointerCast(
1633 DFS.RetvalTLS, PointerType::get(DFS.getShadowTy(T), 0), "_dfsret");
1634}
1635
1636Value *DFSanFunction::getRetvalOriginTLS() { return DFS.RetvalOriginTLS; }
1637
1638Value *DFSanFunction::getArgOriginTLS(unsigned ArgNo, IRBuilder<> &IRB) {
1639 return IRB.CreateConstGEP2_64(DFS.ArgOriginTLSTy, DFS.ArgOriginTLS, 0, ArgNo,
1640 "_dfsarg_o");
1641}
1642
1643Value *DFSanFunction::getOrigin(Value *V) {
1644 assert(DFS.shouldTrackOrigins())((void)0);
1645 if (!isa<Argument>(V) && !isa<Instruction>(V))
1646 return DFS.ZeroOrigin;
1647 Value *&Origin = ValOriginMap[V];
1648 if (!Origin) {
1649 if (Argument *A = dyn_cast<Argument>(V)) {
1650 if (IsNativeABI)
1651 return DFS.ZeroOrigin;
1652 switch (IA) {
1653 case DataFlowSanitizer::IA_TLS: {
1654 if (A->getArgNo() < DFS.NumOfElementsInArgOrgTLS) {
1655 Instruction *ArgOriginTLSPos = &*F->getEntryBlock().begin();
1656 IRBuilder<> IRB(ArgOriginTLSPos);
1657 Value *ArgOriginPtr = getArgOriginTLS(A->getArgNo(), IRB);
1658 Origin = IRB.CreateLoad(DFS.OriginTy, ArgOriginPtr);
1659 } else {
1660 // Overflow
1661 Origin = DFS.ZeroOrigin;
1662 }
1663 break;
1664 }
1665 case DataFlowSanitizer::IA_Args: {
1666 Origin = DFS.ZeroOrigin;
1667 break;
1668 }
1669 }
1670 } else {
1671 Origin = DFS.ZeroOrigin;
1672 }
1673 }
1674 return Origin;
1675}
1676
1677void DFSanFunction::setOrigin(Instruction *I, Value *Origin) {
1678 if (!DFS.shouldTrackOrigins())
1679 return;
1680 assert(!ValOriginMap.count(I))((void)0);
1681 assert(Origin->getType() == DFS.OriginTy)((void)0);
1682 ValOriginMap[I] = Origin;
1683}
1684
1685Value *DFSanFunction::getShadowForTLSArgument(Argument *A) {
1686 unsigned ArgOffset = 0;
1687 const DataLayout &DL = F->getParent()->getDataLayout();
1688 for (auto &FArg : F->args()) {
1689 if (!FArg.getType()->isSized()) {
1690 if (A == &FArg)
1691 break;
1692 continue;
1693 }
1694
1695 unsigned Size = DL.getTypeAllocSize(DFS.getShadowTy(&FArg));
1696 if (A != &FArg) {
1697 ArgOffset += alignTo(Size, ShadowTLSAlignment);
1698 if (ArgOffset > ArgTLSSize)
1699 break; // ArgTLS overflows, uses a zero shadow.
1700 continue;
1701 }
1702
1703 if (ArgOffset + Size > ArgTLSSize)
1704 break; // ArgTLS overflows, uses a zero shadow.
1705
1706 Instruction *ArgTLSPos = &*F->getEntryBlock().begin();
1707 IRBuilder<> IRB(ArgTLSPos);
1708 Value *ArgShadowPtr = getArgTLS(FArg.getType(), ArgOffset, IRB);
1709 return IRB.CreateAlignedLoad(DFS.getShadowTy(&FArg), ArgShadowPtr,
1710 ShadowTLSAlignment);
1711 }
1712
1713 return DFS.getZeroShadow(A);
1714}
1715
1716Value *DFSanFunction::getShadow(Value *V) {
1717 if (!isa<Argument>(V) && !isa<Instruction>(V))
1718 return DFS.getZeroShadow(V);
1719 Value *&Shadow = ValShadowMap[V];
1720 if (!Shadow) {
1721 if (Argument *A = dyn_cast<Argument>(V)) {
1722 if (IsNativeABI)
1723 return DFS.getZeroShadow(V);
1724 switch (IA) {
1725 case DataFlowSanitizer::IA_TLS: {
1726 Shadow = getShadowForTLSArgument(A);
1727 break;
1728 }
1729 case DataFlowSanitizer::IA_Args: {
1730 unsigned ArgIdx = A->getArgNo() + F->arg_size() / 2;
1731 Function::arg_iterator Arg = F->arg_begin();
1732 std::advance(Arg, ArgIdx);
1733 Shadow = &*Arg;
1734 assert(Shadow->getType() == DFS.PrimitiveShadowTy)((void)0);
1735 break;
1736 }
1737 }
1738 NonZeroChecks.push_back(Shadow);
1739 } else {
1740 Shadow = DFS.getZeroShadow(V);
1741 }
1742 }
1743 return Shadow;
1744}
1745
1746void DFSanFunction::setShadow(Instruction *I, Value *Shadow) {
1747 assert(!ValShadowMap.count(I))((void)0);
1748 assert(DFS.shouldTrackFieldsAndIndices() ||((void)0)
1749 Shadow->getType() == DFS.PrimitiveShadowTy)((void)0);
1750 ValShadowMap[I] = Shadow;
1751}
1752
1753/// Compute the integer shadow offset that corresponds to a given
1754/// application address.
1755///
1756/// Offset = (Addr & ~AndMask) ^ XorMask
1757Value *DataFlowSanitizer::getShadowOffset(Value *Addr, IRBuilder<> &IRB) {
1758 assert(Addr != RetvalTLS && "Reinstrumenting?")((void)0);
1759 Value *OffsetLong = IRB.CreatePointerCast(Addr, IntptrTy);
1760
1761 uint64_t AndMask = MapParams->AndMask;
1762 if (AndMask)
1763 OffsetLong =
1764 IRB.CreateAnd(OffsetLong, ConstantInt::get(IntptrTy, ~AndMask));
1765
1766 uint64_t XorMask = MapParams->XorMask;
1767 if (XorMask)
1768 OffsetLong = IRB.CreateXor(OffsetLong, ConstantInt::get(IntptrTy, XorMask));
1769 return OffsetLong;
1770}
1771
1772std::pair<Value *, Value *>
1773DataFlowSanitizer::getShadowOriginAddress(Value *Addr, Align InstAlignment,
1774 Instruction *Pos) {
1775 // Returns ((Addr & shadow_mask) + origin_base - shadow_base) & ~4UL
1776 IRBuilder<> IRB(Pos);
1777 Value *ShadowOffset = getShadowOffset(Addr, IRB);
1778 Value *ShadowLong = ShadowOffset;
1779 uint64_t ShadowBase = MapParams->ShadowBase;
1780 if (ShadowBase != 0) {
1781 ShadowLong =
1782 IRB.CreateAdd(ShadowLong, ConstantInt::get(IntptrTy, ShadowBase));
1783 }
1784 IntegerType *ShadowTy = IntegerType::get(*Ctx, ShadowWidthBits);
1785 Value *ShadowPtr =
1786 IRB.CreateIntToPtr(ShadowLong, PointerType::get(ShadowTy, 0));
1787 Value *OriginPtr = nullptr;
1788 if (shouldTrackOrigins()) {
1789 Value *OriginLong = ShadowOffset;
1790 uint64_t OriginBase = MapParams->OriginBase;
1791 if (OriginBase != 0)
1792 OriginLong =
1793 IRB.CreateAdd(OriginLong, ConstantInt::get(IntptrTy, OriginBase));
1794 const Align Alignment = llvm::assumeAligned(InstAlignment.value());
1795 // When alignment is >= 4, Addr must be aligned to 4, otherwise it is UB.
1796 // So Mask is unnecessary.
1797 if (Alignment < MinOriginAlignment) {
1798 uint64_t Mask = MinOriginAlignment.value() - 1;
1799 OriginLong = IRB.CreateAnd(OriginLong, ConstantInt::get(IntptrTy, ~Mask));
1800 }
1801 OriginPtr = IRB.CreateIntToPtr(OriginLong, OriginPtrTy);
1802 }
1803 return std::make_pair(ShadowPtr, OriginPtr);
1804}
1805
1806Value *DataFlowSanitizer::getShadowAddress(Value *Addr, Instruction *Pos,
1807 Value *ShadowOffset) {
1808 IRBuilder<> IRB(Pos);
1809 return IRB.CreateIntToPtr(ShadowOffset, PrimitiveShadowPtrTy);
1810}
1811
1812Value *DataFlowSanitizer::getShadowAddress(Value *Addr, Instruction *Pos) {
1813 IRBuilder<> IRB(Pos);
1814 Value *ShadowOffset = getShadowOffset(Addr, IRB);
1815 return getShadowAddress(Addr, Pos, ShadowOffset);
1816}
1817
1818Value *DFSanFunction::combineShadowsThenConvert(Type *T, Value *V1, Value *V2,
1819 Instruction *Pos) {
1820 Value *PrimitiveValue = combineShadows(V1, V2, Pos);
1821 return expandFromPrimitiveShadow(T, PrimitiveValue, Pos);
1822}
1823
1824// Generates IR to compute the union of the two given shadows, inserting it
1825// before Pos. The combined value is with primitive type.
1826Value *DFSanFunction::combineShadows(Value *V1, Value *V2, Instruction *Pos) {
1827 if (DFS.isZeroShadow(V1))
1828 return collapseToPrimitiveShadow(V2, Pos);
1829 if (DFS.isZeroShadow(V2))
1830 return collapseToPrimitiveShadow(V1, Pos);
1831 if (V1 == V2)
1832 return collapseToPrimitiveShadow(V1, Pos);
1833
1834 auto V1Elems = ShadowElements.find(V1);
1835 auto V2Elems = ShadowElements.find(V2);
1836 if (V1Elems != ShadowElements.end() && V2Elems != ShadowElements.end()) {
1837 if (std::includes(V1Elems->second.begin(), V1Elems->second.end(),
1838 V2Elems->second.begin(), V2Elems->second.end())) {
1839 return collapseToPrimitiveShadow(V1, Pos);
1840 }
1841 if (std::includes(V2Elems->second.begin(), V2Elems->second.end(),
1842 V1Elems->second.begin(), V1Elems->second.end())) {
1843 return collapseToPrimitiveShadow(V2, Pos);
1844 }
1845 } else if (V1Elems != ShadowElements.end()) {
1846 if (V1Elems->second.count(V2))
1847 return collapseToPrimitiveShadow(V1, Pos);
1848 } else if (V2Elems != ShadowElements.end()) {
1849 if (V2Elems->second.count(V1))
1850 return collapseToPrimitiveShadow(V2, Pos);
1851 }
1852
1853 auto Key = std::make_pair(V1, V2);
1854 if (V1 > V2)
1855 std::swap(Key.first, Key.second);
1856 CachedShadow &CCS = CachedShadows[Key];
1857 if (CCS.Block && DT.dominates(CCS.Block, Pos->getParent()))
1858 return CCS.Shadow;
1859
1860 // Converts inputs shadows to shadows with primitive types.
1861 Value *PV1 = collapseToPrimitiveShadow(V1, Pos);
1862 Value *PV2 = collapseToPrimitiveShadow(V2, Pos);
1863
1864 IRBuilder<> IRB(Pos);
1865 CCS.Block = Pos->getParent();
1866 CCS.Shadow = IRB.CreateOr(PV1, PV2);
1867
1868 std::set<Value *> UnionElems;
1869 if (V1Elems != ShadowElements.end()) {
1870 UnionElems = V1Elems->second;
1871 } else {
1872 UnionElems.insert(V1);
1873 }
1874 if (V2Elems != ShadowElements.end()) {
1875 UnionElems.insert(V2Elems->second.begin(), V2Elems->second.end());
1876 } else {
1877 UnionElems.insert(V2);
1878 }
1879 ShadowElements[CCS.Shadow] = std::move(UnionElems);
1880
1881 return CCS.Shadow;
1882}
1883
1884// A convenience function which folds the shadows of each of the operands
1885// of the provided instruction Inst, inserting the IR before Inst. Returns
1886// the computed union Value.
1887Value *DFSanFunction::combineOperandShadows(Instruction *Inst) {
1888 if (Inst->getNumOperands() == 0)
1889 return DFS.getZeroShadow(Inst);
1890
1891 Value *Shadow = getShadow(Inst->getOperand(0));
1892 for (unsigned I = 1, N = Inst->getNumOperands(); I < N; ++I)
1893 Shadow = combineShadows(Shadow, getShadow(Inst->getOperand(I)), Inst);
1894
1895 return expandFromPrimitiveShadow(Inst->getType(), Shadow, Inst);
1896}
1897
1898void DFSanVisitor::visitInstOperands(Instruction &I) {
1899 Value *CombinedShadow = DFSF.combineOperandShadows(&I);
1900 DFSF.setShadow(&I, CombinedShadow);
1901 visitInstOperandOrigins(I);
1902}
1903
1904Value *DFSanFunction::combineOrigins(const std::vector<Value *> &Shadows,
1905 const std::vector<Value *> &Origins,
1906 Instruction *Pos, ConstantInt *Zero) {
1907 assert(Shadows.size() == Origins.size())((void)0);
1908 size_t Size = Origins.size();
1909 if (Size == 0)
1910 return DFS.ZeroOrigin;
1911 Value *Origin = nullptr;
1912 if (!Zero)
1913 Zero = DFS.ZeroPrimitiveShadow;
1914 for (size_t I = 0; I != Size; ++I) {
1915 Value *OpOrigin = Origins[I];
1916 Constant *ConstOpOrigin = dyn_cast<Constant>(OpOrigin);
1917 if (ConstOpOrigin && ConstOpOrigin->isNullValue())
1918 continue;
1919 if (!Origin) {
1920 Origin = OpOrigin;
1921 continue;
1922 }
1923 Value *OpShadow = Shadows[I];
1924 Value *PrimitiveShadow = collapseToPrimitiveShadow(OpShadow, Pos);
1925 IRBuilder<> IRB(Pos);
1926 Value *Cond = IRB.CreateICmpNE(PrimitiveShadow, Zero);
1927 Origin = IRB.CreateSelect(Cond, OpOrigin, Origin);
1928 }
1929 return Origin ? Origin : DFS.ZeroOrigin;
1930}
1931
1932Value *DFSanFunction::combineOperandOrigins(Instruction *Inst) {
1933 size_t Size = Inst->getNumOperands();
1934 std::vector<Value *> Shadows(Size);
1935 std::vector<Value *> Origins(Size);
1936 for (unsigned I = 0; I != Size; ++I) {
1937 Shadows[I] = getShadow(Inst->getOperand(I));
1938 Origins[I] = getOrigin(Inst->getOperand(I));
1939 }
1940 return combineOrigins(Shadows, Origins, Inst);
1941}
1942
1943void DFSanVisitor::visitInstOperandOrigins(Instruction &I) {
1944 if (!DFSF.DFS.shouldTrackOrigins())
1945 return;
1946 Value *CombinedOrigin = DFSF.combineOperandOrigins(&I);
1947 DFSF.setOrigin(&I, CombinedOrigin);
1948}
1949
1950Align DFSanFunction::getShadowAlign(Align InstAlignment) {
1951 const Align Alignment = ClPreserveAlignment ? InstAlignment : Align(1);
5
Assuming the condition is true
6
'?' condition is true
7
The value 255 is assigned to 'Alignment.ShiftValue'
1952 return Align(Alignment.value() * DFS.ShadowWidthBytes);
8
Calling 'Align::value'
1953}
1954
1955Align DFSanFunction::getOriginAlign(Align InstAlignment) {
1956 const Align Alignment = llvm::assumeAligned(InstAlignment.value());
1957 return Align(std::max(MinOriginAlignment, Alignment));
1958}
1959
1960bool DFSanFunction::useCallbackLoadLabelAndOrigin(uint64_t Size,
1961 Align InstAlignment) {
1962 // When enabling tracking load instructions, we always use
1963 // __dfsan_load_label_and_origin to reduce code size.
1964 if (ClTrackOrigins == 2)
1965 return true;
1966
1967 assert(Size != 0)((void)0);
1968 // * if Size == 1, it is sufficient to load its origin aligned at 4.
1969 // * if Size == 2, we assume most cases Addr % 2 == 0, so it is sufficient to
1970 // load its origin aligned at 4. If not, although origins may be lost, it
1971 // should not happen very often.
1972 // * if align >= 4, Addr must be aligned to 4, otherwise it is UB. When
1973 // Size % 4 == 0, it is more efficient to load origins without callbacks.
1974 // * Otherwise we use __dfsan_load_label_and_origin.
1975 // This should ensure that common cases run efficiently.
1976 if (Size <= 2)
1977 return false;
1978
1979 const Align Alignment = llvm::assumeAligned(InstAlignment.value());
1980 return Alignment < MinOriginAlignment || !DFS.hasLoadSizeForFastPath(Size);
1981}
1982
1983Value *DataFlowSanitizer::loadNextOrigin(Instruction *Pos, Align OriginAlign,
1984 Value **OriginAddr) {
1985 IRBuilder<> IRB(Pos);
1986 *OriginAddr =
1987 IRB.CreateGEP(OriginTy, *OriginAddr, ConstantInt::get(IntptrTy, 1));
1988 return IRB.CreateAlignedLoad(OriginTy, *OriginAddr, OriginAlign);
1989}
1990
1991std::pair<Value *, Value *> DFSanFunction::loadShadowFast(
1992 Value *ShadowAddr, Value *OriginAddr, uint64_t Size, Align ShadowAlign,
1993 Align OriginAlign, Value *FirstOrigin, Instruction *Pos) {
1994 const bool ShouldTrackOrigins = DFS.shouldTrackOrigins();
1995 const uint64_t ShadowSize = Size * DFS.ShadowWidthBytes;
1996
1997 assert(Size >= 4 && "Not large enough load size for fast path!")((void)0);
1998
1999 // Used for origin tracking.
2000 std::vector<Value *> Shadows;
2001 std::vector<Value *> Origins;
2002
2003 // Load instructions in LLVM can have arbitrary byte sizes (e.g., 3, 12, 20)
2004 // but this function is only used in a subset of cases that make it possible
2005 // to optimize the instrumentation.
2006 //
2007 // Specifically, when the shadow size in bytes (i.e., loaded bytes x shadow
2008 // per byte) is either:
2009 // - a multiple of 8 (common)
2010 // - equal to 4 (only for load32)
2011 //
2012 // For the second case, we can fit the wide shadow in a 32-bit integer. In all
2013 // other cases, we use a 64-bit integer to hold the wide shadow.
2014 Type *WideShadowTy =
2015 ShadowSize == 4 ? Type::getInt32Ty(*DFS.Ctx) : Type::getInt64Ty(*DFS.Ctx);
2016
2017 IRBuilder<> IRB(Pos);
2018 Value *WideAddr = IRB.CreateBitCast(ShadowAddr, WideShadowTy->getPointerTo());
2019 Value *CombinedWideShadow =
2020 IRB.CreateAlignedLoad(WideShadowTy, WideAddr, ShadowAlign);
2021
2022 unsigned WideShadowBitWidth = WideShadowTy->getIntegerBitWidth();
2023 const uint64_t BytesPerWideShadow = WideShadowBitWidth / DFS.ShadowWidthBits;
2024
2025 auto AppendWideShadowAndOrigin = [&](Value *WideShadow, Value *Origin) {
2026 if (BytesPerWideShadow > 4) {
2027 assert(BytesPerWideShadow == 8)((void)0);
2028 // The wide shadow relates to two origin pointers: one for the first four
2029 // application bytes, and one for the latest four. We use a left shift to
2030 // get just the shadow bytes that correspond to the first origin pointer,
2031 // and then the entire shadow for the second origin pointer (which will be
2032 // chosen by combineOrigins() iff the least-significant half of the wide
2033 // shadow was empty but the other half was not).
2034 Value *WideShadowLo = IRB.CreateShl(
2035 WideShadow, ConstantInt::get(WideShadowTy, WideShadowBitWidth / 2));
2036 Shadows.push_back(WideShadow);
2037 Origins.push_back(DFS.loadNextOrigin(Pos, OriginAlign, &OriginAddr));
2038
2039 Shadows.push_back(WideShadowLo);
2040 Origins.push_back(Origin);
2041 } else {
2042 Shadows.push_back(WideShadow);
2043 Origins.push_back(Origin);
2044 }
2045 };
2046
2047 if (ShouldTrackOrigins)
2048 AppendWideShadowAndOrigin(CombinedWideShadow, FirstOrigin);
2049
2050 // First OR all the WideShadows (i.e., 64bit or 32bit shadow chunks) linearly;
2051 // then OR individual shadows within the combined WideShadow by binary ORing.
2052 // This is fewer instructions than ORing shadows individually, since it
2053 // needs logN shift/or instructions (N being the bytes of the combined wide
2054 // shadow).
2055 for (uint64_t ByteOfs = BytesPerWideShadow; ByteOfs < Size;
2056 ByteOfs += BytesPerWideShadow) {
2057 WideAddr = IRB.CreateGEP(WideShadowTy, WideAddr,
2058 ConstantInt::get(DFS.IntptrTy, 1));
2059 Value *NextWideShadow =
2060 IRB.CreateAlignedLoad(WideShadowTy, WideAddr, ShadowAlign);
2061 CombinedWideShadow = IRB.CreateOr(CombinedWideShadow, NextWideShadow);
2062 if (ShouldTrackOrigins) {
2063 Value *NextOrigin = DFS.loadNextOrigin(Pos, OriginAlign, &OriginAddr);
2064 AppendWideShadowAndOrigin(NextWideShadow, NextOrigin);
2065 }
2066 }
2067 for (unsigned Width = WideShadowBitWidth / 2; Width >= DFS.ShadowWidthBits;
2068 Width >>= 1) {
2069 Value *ShrShadow = IRB.CreateLShr(CombinedWideShadow, Width);
2070 CombinedWideShadow = IRB.CreateOr(CombinedWideShadow, ShrShadow);
2071 }
2072 return {IRB.CreateTrunc(CombinedWideShadow, DFS.PrimitiveShadowTy),
2073 ShouldTrackOrigins
2074 ? combineOrigins(Shadows, Origins, Pos,
2075 ConstantInt::getSigned(IRB.getInt64Ty(), 0))
2076 : DFS.ZeroOrigin};
2077}
2078
2079std::pair<Value *, Value *> DFSanFunction::loadShadowOriginSansLoadTracking(
2080 Value *Addr, uint64_t Size, Align InstAlignment, Instruction *Pos) {
2081 const bool ShouldTrackOrigins = DFS.shouldTrackOrigins();
2082
2083 // Non-escaped loads.
2084 if (AllocaInst *AI = dyn_cast<AllocaInst>(Addr)) {
2085 const auto SI = AllocaShadowMap.find(AI);
2086 if (SI != AllocaShadowMap.end()) {
2087 IRBuilder<> IRB(Pos);
2088 Value *ShadowLI = IRB.CreateLoad(DFS.PrimitiveShadowTy, SI->second);
2089 const auto OI = AllocaOriginMap.find(AI);
2090 assert(!ShouldTrackOrigins || OI != AllocaOriginMap.end())((void)0);
2091 return {ShadowLI, ShouldTrackOrigins
2092 ? IRB.CreateLoad(DFS.OriginTy, OI->second)
2093 : nullptr};
2094 }
2095 }
2096
2097 // Load from constant addresses.
2098 SmallVector<const Value *, 2> Objs;
2099 getUnderlyingObjects(Addr, Objs);
2100 bool AllConstants = true;
2101 for (const Value *Obj : Objs) {
2102 if (isa<Function>(Obj) || isa<BlockAddress>(Obj))
2103 continue;
2104 if (isa<GlobalVariable>(Obj) && cast<GlobalVariable>(Obj)->isConstant())
2105 continue;
2106
2107 AllConstants = false;
2108 break;
2109 }
2110 if (AllConstants)
2111 return {DFS.ZeroPrimitiveShadow,
2112 ShouldTrackOrigins ? DFS.ZeroOrigin : nullptr};
2113
2114 if (Size == 0)
2115 return {DFS.ZeroPrimitiveShadow,
2116 ShouldTrackOrigins ? DFS.ZeroOrigin : nullptr};
2117
2118 // Use callback to load if this is not an optimizable case for origin
2119 // tracking.
2120 if (ShouldTrackOrigins &&
2121 useCallbackLoadLabelAndOrigin(Size, InstAlignment)) {
2122 IRBuilder<> IRB(Pos);
2123 CallInst *Call =
2124 IRB.CreateCall(DFS.DFSanLoadLabelAndOriginFn,
2125 {IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()),
2126 ConstantInt::get(DFS.IntptrTy, Size)});
2127 Call->addAttribute(AttributeList::ReturnIndex, Attribute::ZExt);
2128 return {IRB.CreateTrunc(IRB.CreateLShr(Call, DFS.OriginWidthBits),
2129 DFS.PrimitiveShadowTy),
2130 IRB.CreateTrunc(Call, DFS.OriginTy)};
2131 }
2132
2133 // Other cases that support loading shadows or origins in a fast way.
2134 Value *ShadowAddr, *OriginAddr;
2135 std::tie(ShadowAddr, OriginAddr) =
2136 DFS.getShadowOriginAddress(Addr, InstAlignment, Pos);
2137
2138 const Align ShadowAlign = getShadowAlign(InstAlignment);
2139 const Align OriginAlign = getOriginAlign(InstAlignment);
2140 Value *Origin = nullptr;
2141 if (ShouldTrackOrigins) {
2142 IRBuilder<> IRB(Pos);
2143 Origin = IRB.CreateAlignedLoad(DFS.OriginTy, OriginAddr, OriginAlign);
2144 }
2145
2146 // When the byte size is small enough, we can load the shadow directly with
2147 // just a few instructions.
2148 switch (Size) {
2149 case 1: {
2150 LoadInst *LI = new LoadInst(DFS.PrimitiveShadowTy, ShadowAddr, "", Pos);
2151 LI->setAlignment(ShadowAlign);
2152 return {LI, Origin};
2153 }
2154 case 2: {
2155 IRBuilder<> IRB(Pos);
2156 Value *ShadowAddr1 = IRB.CreateGEP(DFS.PrimitiveShadowTy, ShadowAddr,
2157 ConstantInt::get(DFS.IntptrTy, 1));
2158 Value *Load =
2159 IRB.CreateAlignedLoad(DFS.PrimitiveShadowTy, ShadowAddr, ShadowAlign);
2160 Value *Load1 =
2161 IRB.CreateAlignedLoad(DFS.PrimitiveShadowTy, ShadowAddr1, ShadowAlign);
2162 return {combineShadows(Load, Load1, Pos), Origin};
2163 }
2164 }
2165 bool HasSizeForFastPath = DFS.hasLoadSizeForFastPath(Size);
2166
2167 if (HasSizeForFastPath)
2168 return loadShadowFast(ShadowAddr, OriginAddr, Size, ShadowAlign,
2169 OriginAlign, Origin, Pos);
2170
2171 IRBuilder<> IRB(Pos);
2172 CallInst *FallbackCall = IRB.CreateCall(
2173 DFS.DFSanUnionLoadFn, {ShadowAddr, ConstantInt::get(DFS.IntptrTy, Size)});
2174 FallbackCall->addAttribute(AttributeList::ReturnIndex, Attribute::ZExt);
2175 return {FallbackCall, Origin};
2176}
2177
2178std::pair<Value *, Value *> DFSanFunction::loadShadowOrigin(Value *Addr,
2179 uint64_t Size,
2180 Align InstAlignment,
2181 Instruction *Pos) {
2182 Value *PrimitiveShadow, *Origin;
2183 std::tie(PrimitiveShadow, Origin) =
2184 loadShadowOriginSansLoadTracking(Addr, Size, InstAlignment, Pos);
2185 if (DFS.shouldTrackOrigins()) {
2186 if (ClTrackOrigins == 2) {
2187 IRBuilder<> IRB(Pos);
2188 auto *ConstantShadow = dyn_cast<Constant>(PrimitiveShadow);
2189 if (!ConstantShadow || !ConstantShadow->isZeroValue())
2190 Origin = updateOriginIfTainted(PrimitiveShadow, Origin, IRB);
2191 }
2192 }
2193 return {PrimitiveShadow, Origin};
2194}
2195
2196static AtomicOrdering addAcquireOrdering(AtomicOrdering AO) {
2197 switch (AO) {
2198 case AtomicOrdering::NotAtomic:
2199 return AtomicOrdering::NotAtomic;
2200 case AtomicOrdering::Unordered:
2201 case AtomicOrdering::Monotonic:
2202 case AtomicOrdering::Acquire:
2203 return AtomicOrdering::Acquire;
2204 case AtomicOrdering::Release:
2205 case AtomicOrdering::AcquireRelease:
2206 return AtomicOrdering::AcquireRelease;
2207 case AtomicOrdering::SequentiallyConsistent:
2208 return AtomicOrdering::SequentiallyConsistent;
2209 }
2210 llvm_unreachable("Unknown ordering")__builtin_unreachable();
2211}
2212
2213void DFSanVisitor::visitLoadInst(LoadInst &LI) {
2214 auto &DL = LI.getModule()->getDataLayout();
2215 uint64_t Size = DL.getTypeStoreSize(LI.getType());
2216 if (Size == 0) {
2217 DFSF.setShadow(&LI, DFSF.DFS.getZeroShadow(&LI));
2218 DFSF.setOrigin(&LI, DFSF.DFS.ZeroOrigin);
2219 return;
2220 }
2221
2222 // When an application load is atomic, increase atomic ordering between
2223 // atomic application loads and stores to ensure happen-before order; load
2224 // shadow data after application data; store zero shadow data before
2225 // application data. This ensure shadow loads return either labels of the
2226 // initial application data or zeros.
2227 if (LI.isAtomic())
2228 LI.setOrdering(addAcquireOrdering(LI.getOrdering()));
2229
2230 Instruction *Pos = LI.isAtomic() ? LI.getNextNode() : &LI;
2231 std::vector<Value *> Shadows;
2232 std::vector<Value *> Origins;
2233 Value *PrimitiveShadow, *Origin;
2234 std::tie(PrimitiveShadow, Origin) =
2235 DFSF.loadShadowOrigin(LI.getPointerOperand(), Size, LI.getAlign(), Pos);
2236 const bool ShouldTrackOrigins = DFSF.DFS.shouldTrackOrigins();
2237 if (ShouldTrackOrigins) {
2238 Shadows.push_back(PrimitiveShadow);
2239 Origins.push_back(Origin);
2240 }
2241 if (ClCombinePointerLabelsOnLoad) {
2242 Value *PtrShadow = DFSF.getShadow(LI.getPointerOperand());
2243 PrimitiveShadow = DFSF.combineShadows(PrimitiveShadow, PtrShadow, Pos);
2244 if (ShouldTrackOrigins) {
2245 Shadows.push_back(PtrShadow);
2246 Origins.push_back(DFSF.getOrigin(LI.getPointerOperand()));
2247 }
2248 }
2249 if (!DFSF.DFS.isZeroShadow(PrimitiveShadow))
2250 DFSF.NonZeroChecks.push_back(PrimitiveShadow);
2251
2252 Value *Shadow =
2253 DFSF.expandFromPrimitiveShadow(LI.getType(), PrimitiveShadow, Pos);
2254 DFSF.setShadow(&LI, Shadow);
2255
2256 if (ShouldTrackOrigins) {
2257 DFSF.setOrigin(&LI, DFSF.combineOrigins(Shadows, Origins, Pos));
2258 }
2259
2260 if (ClEventCallbacks) {
2261 IRBuilder<> IRB(Pos);
2262 Value *Addr8 = IRB.CreateBitCast(LI.getPointerOperand(), DFSF.DFS.Int8Ptr);
2263 IRB.CreateCall(DFSF.DFS.DFSanLoadCallbackFn, {PrimitiveShadow, Addr8});
2264 }
2265}
2266
2267Value *DFSanFunction::updateOriginIfTainted(Value *Shadow, Value *Origin,
2268 IRBuilder<> &IRB) {
2269 assert(DFS.shouldTrackOrigins())((void)0);
2270 return IRB.CreateCall(DFS.DFSanChainOriginIfTaintedFn, {Shadow, Origin});
2271}
2272
2273Value *DFSanFunction::updateOrigin(Value *V, IRBuilder<> &IRB) {
2274 if (!DFS.shouldTrackOrigins())
2275 return V;
2276 return IRB.CreateCall(DFS.DFSanChainOriginFn, V);
2277}
2278
2279Value *DFSanFunction::originToIntptr(IRBuilder<> &IRB, Value *Origin) {
2280 const unsigned OriginSize = DataFlowSanitizer::OriginWidthBytes;
2281 const DataLayout &DL = F->getParent()->getDataLayout();
2282 unsigned IntptrSize = DL.getTypeStoreSize(DFS.IntptrTy);
2283 if (IntptrSize == OriginSize)
2284 return Origin;
2285 assert(IntptrSize == OriginSize * 2)((void)0);
2286 Origin = IRB.CreateIntCast(Origin, DFS.IntptrTy, /* isSigned */ false);
2287 return IRB.CreateOr(Origin, IRB.CreateShl(Origin, OriginSize * 8));
2288}
2289
2290void DFSanFunction::paintOrigin(IRBuilder<> &IRB, Value *Origin,
2291 Value *StoreOriginAddr,
2292 uint64_t StoreOriginSize, Align Alignment) {
2293 const unsigned OriginSize = DataFlowSanitizer::OriginWidthBytes;
2294 const DataLayout &DL = F->getParent()->getDataLayout();
2295 const Align IntptrAlignment = DL.getABITypeAlign(DFS.IntptrTy);
2296 unsigned IntptrSize = DL.getTypeStoreSize(DFS.IntptrTy);
2297 assert(IntptrAlignment >= MinOriginAlignment)((void)0);
2298 assert(IntptrSize >= OriginSize)((void)0);
2299
2300 unsigned Ofs = 0;
2301 Align CurrentAlignment = Alignment;
2302 if (Alignment >= IntptrAlignment && IntptrSize > OriginSize) {
2303 Value *IntptrOrigin = originToIntptr(IRB, Origin);
2304 Value *IntptrStoreOriginPtr = IRB.CreatePointerCast(
2305 StoreOriginAddr, PointerType::get(DFS.IntptrTy, 0));
2306 for (unsigned I = 0; I < StoreOriginSize / IntptrSize; ++I) {
2307 Value *Ptr =
2308 I ? IRB.CreateConstGEP1_32(DFS.IntptrTy, IntptrStoreOriginPtr, I)
2309 : IntptrStoreOriginPtr;
2310 IRB.CreateAlignedStore(IntptrOrigin, Ptr, CurrentAlignment);
2311 Ofs += IntptrSize / OriginSize;
2312 CurrentAlignment = IntptrAlignment;
2313 }
2314 }
2315
2316 for (unsigned I = Ofs; I < (StoreOriginSize + OriginSize - 1) / OriginSize;
2317 ++I) {
2318 Value *GEP = I ? IRB.CreateConstGEP1_32(DFS.OriginTy, StoreOriginAddr, I)
2319 : StoreOriginAddr;
2320 IRB.CreateAlignedStore(Origin, GEP, CurrentAlignment);
2321 CurrentAlignment = MinOriginAlignment;
2322 }
2323}
2324
2325Value *DFSanFunction::convertToBool(Value *V, IRBuilder<> &IRB,
2326 const Twine &Name) {
2327 Type *VTy = V->getType();
2328 assert(VTy->isIntegerTy())((void)0);
2329 if (VTy->getIntegerBitWidth() == 1)
2330 // Just converting a bool to a bool, so do nothing.
2331 return V;
2332 return IRB.CreateICmpNE(V, ConstantInt::get(VTy, 0), Name);
2333}
2334
2335void DFSanFunction::storeOrigin(Instruction *Pos, Value *Addr, uint64_t Size,
2336 Value *Shadow, Value *Origin,
2337 Value *StoreOriginAddr, Align InstAlignment) {
2338 // Do not write origins for zero shadows because we do not trace origins for
2339 // untainted sinks.
2340 const Align OriginAlignment = getOriginAlign(InstAlignment);
2341 Value *CollapsedShadow = collapseToPrimitiveShadow(Shadow, Pos);
2342 IRBuilder<> IRB(Pos);
2343 if (auto *ConstantShadow = dyn_cast<Constant>(CollapsedShadow)) {
2344 if (!ConstantShadow->isZeroValue())
2345 paintOrigin(IRB, updateOrigin(Origin, IRB), StoreOriginAddr, Size,
2346 OriginAlignment);
2347 return;
2348 }
2349
2350 if (shouldInstrumentWithCall()) {
2351 IRB.CreateCall(DFS.DFSanMaybeStoreOriginFn,
2352 {CollapsedShadow,
2353 IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()),
2354 ConstantInt::get(DFS.IntptrTy, Size), Origin});
2355 } else {
2356 Value *Cmp = convertToBool(CollapsedShadow, IRB, "_dfscmp");
2357 Instruction *CheckTerm = SplitBlockAndInsertIfThen(
2358 Cmp, &*IRB.GetInsertPoint(), false, DFS.OriginStoreWeights, &DT);
2359 IRBuilder<> IRBNew(CheckTerm);
2360 paintOrigin(IRBNew, updateOrigin(Origin, IRBNew), StoreOriginAddr, Size,
2361 OriginAlignment);
2362 ++NumOriginStores;
2363 }
2364}
2365
2366void DFSanFunction::storeZeroPrimitiveShadow(Value *Addr, uint64_t Size,
2367 Align ShadowAlign,
2368 Instruction *Pos) {
2369 IRBuilder<> IRB(Pos);
2370 IntegerType *ShadowTy =
2371 IntegerType::get(*DFS.Ctx, Size * DFS.ShadowWidthBits);
2372 Value *ExtZeroShadow = ConstantInt::get(ShadowTy, 0);
2373 Value *ShadowAddr = DFS.getShadowAddress(Addr, Pos);
2374 Value *ExtShadowAddr =
2375 IRB.CreateBitCast(ShadowAddr, PointerType::getUnqual(ShadowTy));
2376 IRB.CreateAlignedStore(ExtZeroShadow, ExtShadowAddr, ShadowAlign);
2377 // Do not write origins for 0 shadows because we do not trace origins for
2378 // untainted sinks.
2379}
2380
2381void DFSanFunction::storePrimitiveShadowOrigin(Value *Addr, uint64_t Size,
2382 Align InstAlignment,
2383 Value *PrimitiveShadow,
2384 Value *Origin,
2385 Instruction *Pos) {
2386 const bool ShouldTrackOrigins = DFS.shouldTrackOrigins() && Origin;
2387
2388 if (AllocaInst *AI = dyn_cast<AllocaInst>(Addr)) {
2389 const auto SI = AllocaShadowMap.find(AI);
2390 if (SI != AllocaShadowMap.end()) {
2391 IRBuilder<> IRB(Pos);
2392 IRB.CreateStore(PrimitiveShadow, SI->second);
2393
2394 // Do not write origins for 0 shadows because we do not trace origins for
2395 // untainted sinks.
2396 if (ShouldTrackOrigins && !DFS.isZeroShadow(PrimitiveShadow)) {
2397 const auto OI = AllocaOriginMap.find(AI);
2398 assert(OI != AllocaOriginMap.end() && Origin)((void)0);
2399 IRB.CreateStore(Origin, OI->second);
2400 }
2401 return;
2402 }
2403 }
2404
2405 const Align ShadowAlign = getShadowAlign(InstAlignment);
2406 if (DFS.isZeroShadow(PrimitiveShadow)) {
2407 storeZeroPrimitiveShadow(Addr, Size, ShadowAlign, Pos);
2408 return;
2409 }
2410
2411 IRBuilder<> IRB(Pos);
2412 Value *ShadowAddr, *OriginAddr;
2413 std::tie(ShadowAddr, OriginAddr) =
2414 DFS.getShadowOriginAddress(Addr, InstAlignment, Pos);
2415
2416 const unsigned ShadowVecSize = 8;
2417 assert(ShadowVecSize * DFS.ShadowWidthBits <= 128 &&((void)0)
2418 "Shadow vector is too large!")((void)0);
2419
2420 uint64_t Offset = 0;
2421 uint64_t LeftSize = Size;
2422 if (LeftSize >= ShadowVecSize) {
2423 auto *ShadowVecTy =
2424 FixedVectorType::get(DFS.PrimitiveShadowTy, ShadowVecSize);
2425 Value *ShadowVec = UndefValue::get(ShadowVecTy);
2426 for (unsigned I = 0; I != ShadowVecSize; ++I) {
2427 ShadowVec = IRB.CreateInsertElement(
2428 ShadowVec, PrimitiveShadow,
2429 ConstantInt::get(Type::getInt32Ty(*DFS.Ctx), I));
2430 }
2431 Value *ShadowVecAddr =
2432 IRB.CreateBitCast(ShadowAddr, PointerType::getUnqual(ShadowVecTy));
2433 do {
2434 Value *CurShadowVecAddr =
2435 IRB.CreateConstGEP1_32(ShadowVecTy, ShadowVecAddr, Offset);
2436 IRB.CreateAlignedStore(ShadowVec, CurShadowVecAddr, ShadowAlign);
2437 LeftSize -= ShadowVecSize;
2438 ++Offset;
2439 } while (LeftSize >= ShadowVecSize);
2440 Offset *= ShadowVecSize;
2441 }
2442 while (LeftSize > 0) {
2443 Value *CurShadowAddr =
2444 IRB.CreateConstGEP1_32(DFS.PrimitiveShadowTy, ShadowAddr, Offset);
2445 IRB.CreateAlignedStore(PrimitiveShadow, CurShadowAddr, ShadowAlign);
2446 --LeftSize;
2447 ++Offset;
2448 }
2449
2450 if (ShouldTrackOrigins) {
2451 storeOrigin(Pos, Addr, Size, PrimitiveShadow, Origin, OriginAddr,
2452 InstAlignment);
2453 }
2454}
2455
2456static AtomicOrdering addReleaseOrdering(AtomicOrdering AO) {
2457 switch (AO) {
2458 case AtomicOrdering::NotAtomic:
2459 return AtomicOrdering::NotAtomic;
2460 case AtomicOrdering::Unordered:
2461 case AtomicOrdering::Monotonic:
2462 case AtomicOrdering::Release:
2463 return AtomicOrdering::Release;
2464 case AtomicOrdering::Acquire:
2465 case AtomicOrdering::AcquireRelease:
2466 return AtomicOrdering::AcquireRelease;
2467 case AtomicOrdering::SequentiallyConsistent:
2468 return AtomicOrdering::SequentiallyConsistent;
2469 }
2470 llvm_unreachable("Unknown ordering")__builtin_unreachable();
2471}
2472
2473void DFSanVisitor::visitStoreInst(StoreInst &SI) {
2474 auto &DL = SI.getModule()->getDataLayout();
2475 Value *Val = SI.getValueOperand();
2476 uint64_t Size = DL.getTypeStoreSize(Val->getType());
2477 if (Size == 0)
2478 return;
2479
2480 // When an application store is atomic, increase atomic ordering between
2481 // atomic application loads and stores to ensure happen-before order; load
2482 // shadow data after application data; store zero shadow data before
2483 // application data. This ensure shadow loads return either labels of the
2484 // initial application data or zeros.
2485 if (SI.isAtomic())
2486 SI.setOrdering(addReleaseOrdering(SI.getOrdering()));
2487
2488 const bool ShouldTrackOrigins =
2489 DFSF.DFS.shouldTrackOrigins() && !SI.isAtomic();
2490 std::vector<Value *> Shadows;
2491 std::vector<Value *> Origins;
2492
2493 Value *Shadow =
2494 SI.isAtomic() ? DFSF.DFS.getZeroShadow(Val) : DFSF.getShadow(Val);
2495
2496 if (ShouldTrackOrigins) {
2497 Shadows.push_back(Shadow);
2498 Origins.push_back(DFSF.getOrigin(Val));
2499 }
2500
2501 Value *PrimitiveShadow;
2502 if (ClCombinePointerLabelsOnStore) {
2503 Value *PtrShadow = DFSF.getShadow(SI.getPointerOperand());
2504 if (ShouldTrackOrigins) {
2505 Shadows.push_back(PtrShadow);
2506 Origins.push_back(DFSF.getOrigin(SI.getPointerOperand()));
2507 }
2508 PrimitiveShadow = DFSF.combineShadows(Shadow, PtrShadow, &SI);
2509 } else {
2510 PrimitiveShadow = DFSF.collapseToPrimitiveShadow(Shadow, &SI);
2511 }
2512 Value *Origin = nullptr;
2513 if (ShouldTrackOrigins)
2514 Origin = DFSF.combineOrigins(Shadows, Origins, &SI);
2515 DFSF.storePrimitiveShadowOrigin(SI.getPointerOperand(), Size, SI.getAlign(),
2516 PrimitiveShadow, Origin, &SI);
2517 if (ClEventCallbacks) {
2518 IRBuilder<> IRB(&SI);
2519 Value *Addr8 = IRB.CreateBitCast(SI.getPointerOperand(), DFSF.DFS.Int8Ptr);
2520 IRB.CreateCall(DFSF.DFS.DFSanStoreCallbackFn, {PrimitiveShadow, Addr8});
2521 }
2522}
2523
2524void DFSanVisitor::visitCASOrRMW(Align InstAlignment, Instruction &I) {
2525 assert(isa<AtomicRMWInst>(I) || isa<AtomicCmpXchgInst>(I))((void)0);
2526
2527 Value *Val = I.getOperand(1);
2528 const auto &DL = I.getModule()->getDataLayout();
2529 uint64_t Size = DL.getTypeStoreSize(Val->getType());
2530 if (Size == 0)
2
Assuming 'Size' is not equal to 0
3
Taking false branch
2531 return;
2532
2533 // Conservatively set data at stored addresses and return with zero shadow to
2534 // prevent shadow data races.
2535 IRBuilder<> IRB(&I);
2536 Value *Addr = I.getOperand(0);
2537 const Align ShadowAlign = DFSF.getShadowAlign(InstAlignment);
4
Calling 'DFSanFunction::getShadowAlign'
2538 DFSF.storeZeroPrimitiveShadow(Addr, Size, ShadowAlign, &I);
2539 DFSF.setShadow(&I, DFSF.DFS.getZeroShadow(&I));
2540 DFSF.setOrigin(&I, DFSF.DFS.ZeroOrigin);
2541}
2542
2543void DFSanVisitor::visitAtomicRMWInst(AtomicRMWInst &I) {
2544 visitCASOrRMW(I.getAlign(), I);
2545 // TODO: The ordering change follows MSan. It is possible not to change
2546 // ordering because we always set and use 0 shadows.
2547 I.setOrdering(addReleaseOrdering(I.getOrdering()));
2548}
2549
2550void DFSanVisitor::visitAtomicCmpXchgInst(AtomicCmpXchgInst &I) {
2551 visitCASOrRMW(I.getAlign(), I);
1
Calling 'DFSanVisitor::visitCASOrRMW'
2552 // TODO: The ordering change follows MSan. It is possible not to change
2553 // ordering because we always set and use 0 shadows.
2554 I.setSuccessOrdering(addReleaseOrdering(I.getSuccessOrdering()));
2555}
2556
2557void DFSanVisitor::visitUnaryOperator(UnaryOperator &UO) {
2558 visitInstOperands(UO);
2559}
2560
2561void DFSanVisitor::visitBinaryOperator(BinaryOperator &BO) {
2562 visitInstOperands(BO);
2563}
2564
2565void DFSanVisitor::visitBitCastInst(BitCastInst &BCI) {
2566 if (DFSF.DFS.getInstrumentedABI() == DataFlowSanitizer::IA_TLS) {
2567 // Special case: if this is the bitcast (there is exactly 1 allowed) between
2568 // a musttail call and a ret, don't instrument. New instructions are not
2569 // allowed after a musttail call.
2570 if (auto *CI = dyn_cast<CallInst>(BCI.getOperand(0)))
2571 if (CI->isMustTailCall())
2572 return;
2573 }
2574 // TODO: handle musttail call returns for IA_Args.
2575 visitInstOperands(BCI);
2576}
2577
2578void DFSanVisitor::visitCastInst(CastInst &CI) { visitInstOperands(CI); }
2579
2580void DFSanVisitor::visitCmpInst(CmpInst &CI) {
2581 visitInstOperands(CI);
2582 if (ClEventCallbacks) {
2583 IRBuilder<> IRB(&CI);
2584 Value *CombinedShadow = DFSF.getShadow(&CI);
2585 IRB.CreateCall(DFSF.DFS.DFSanCmpCallbackFn, CombinedShadow);
2586 }
2587}
2588
2589void DFSanVisitor::visitLandingPadInst(LandingPadInst &LPI) {
2590 // We do not need to track data through LandingPadInst.
2591 //
2592 // For the C++ exceptions, if a value is thrown, this value will be stored
2593 // in a memory location provided by __cxa_allocate_exception(...) (on the
2594 // throw side) or __cxa_begin_catch(...) (on the catch side).
2595 // This memory will have a shadow, so with the loads and stores we will be
2596 // able to propagate labels on data thrown through exceptions, without any
2597 // special handling of the LandingPadInst.
2598 //
2599 // The second element in the pair result of the LandingPadInst is a
2600 // register value, but it is for a type ID and should never be tainted.
2601 DFSF.setShadow(&LPI, DFSF.DFS.getZeroShadow(&LPI));
2602 DFSF.setOrigin(&LPI, DFSF.DFS.ZeroOrigin);
2603}
2604
2605void DFSanVisitor::visitGetElementPtrInst(GetElementPtrInst &GEPI) {
2606 if (ClCombineOffsetLabelsOnGEP) {
2607 visitInstOperands(GEPI);
2608 return;
2609 }
2610
2611 // Only propagate shadow/origin of base pointer value but ignore those of
2612 // offset operands.
2613 Value *BasePointer = GEPI.getPointerOperand();
2614 DFSF.setShadow(&GEPI, DFSF.getShadow(BasePointer));
2615 if (DFSF.DFS.shouldTrackOrigins())
2616 DFSF.setOrigin(&GEPI, DFSF.getOrigin(BasePointer));
2617}
2618
2619void DFSanVisitor::visitExtractElementInst(ExtractElementInst &I) {
2620 visitInstOperands(I);
2621}
2622
2623void DFSanVisitor::visitInsertElementInst(InsertElementInst &I) {
2624 visitInstOperands(I);
2625}
2626
2627void DFSanVisitor::visitShuffleVectorInst(ShuffleVectorInst &I) {
2628 visitInstOperands(I);
2629}
2630
2631void DFSanVisitor::visitExtractValueInst(ExtractValueInst &I) {
2632 if (!DFSF.DFS.shouldTrackFieldsAndIndices()) {
2633 visitInstOperands(I);
2634 return;
2635 }
2636
2637 IRBuilder<> IRB(&I);
2638 Value *Agg = I.getAggregateOperand();
2639 Value *AggShadow = DFSF.getShadow(Agg);
2640 Value *ResShadow = IRB.CreateExtractValue(AggShadow, I.getIndices());
2641 DFSF.setShadow(&I, ResShadow);
2642 visitInstOperandOrigins(I);
2643}
2644
2645void DFSanVisitor::visitInsertValueInst(InsertValueInst &I) {
2646 if (!DFSF.DFS.shouldTrackFieldsAndIndices()) {
2647 visitInstOperands(I);
2648 return;
2649 }
2650
2651 IRBuilder<> IRB(&I);
2652 Value *AggShadow = DFSF.getShadow(I.getAggregateOperand());
2653 Value *InsShadow = DFSF.getShadow(I.getInsertedValueOperand());
2654 Value *Res = IRB.CreateInsertValue(AggShadow, InsShadow, I.getIndices());
2655 DFSF.setShadow(&I, Res);
2656 visitInstOperandOrigins(I);
2657}
2658
2659void DFSanVisitor::visitAllocaInst(AllocaInst &I) {
2660 bool AllLoadsStores = true;
2661 for (User *U : I.users()) {
2662 if (isa<LoadInst>(U))
2663 continue;
2664
2665 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
2666 if (SI->getPointerOperand() == &I)
2667 continue;
2668 }
2669
2670 AllLoadsStores = false;
2671 break;
2672 }
2673 if (AllLoadsStores) {
2674 IRBuilder<> IRB(&I);
2675 DFSF.AllocaShadowMap[&I] = IRB.CreateAlloca(DFSF.DFS.PrimitiveShadowTy);
2676 if (DFSF.DFS.shouldTrackOrigins()) {
2677 DFSF.AllocaOriginMap[&I] =
2678 IRB.CreateAlloca(DFSF.DFS.OriginTy, nullptr, "_dfsa");
2679 }
2680 }
2681 DFSF.setShadow(&I, DFSF.DFS.ZeroPrimitiveShadow);
2682 DFSF.setOrigin(&I, DFSF.DFS.ZeroOrigin);
2683}
2684
2685void DFSanVisitor::visitSelectInst(SelectInst &I) {
2686 Value *CondShadow = DFSF.getShadow(I.getCondition());
2687 Value *TrueShadow = DFSF.getShadow(I.getTrueValue());
2688 Value *FalseShadow = DFSF.getShadow(I.getFalseValue());
2689 Value *ShadowSel = nullptr;
2690 const bool ShouldTrackOrigins = DFSF.DFS.shouldTrackOrigins();
2691 std::vector<Value *> Shadows;
2692 std::vector<Value *> Origins;
2693 Value *TrueOrigin =
2694 ShouldTrackOrigins ? DFSF.getOrigin(I.getTrueValue()) : nullptr;
2695 Value *FalseOrigin =
2696 ShouldTrackOrigins ? DFSF.getOrigin(I.getFalseValue()) : nullptr;
2697
2698 if (isa<VectorType>(I.getCondition()->getType())) {
2699 ShadowSel = DFSF.combineShadowsThenConvert(I.getType(), TrueShadow,
2700 FalseShadow, &I);
2701 if (ShouldTrackOrigins) {
2702 Shadows.push_back(TrueShadow);
2703 Shadows.push_back(FalseShadow);
2704 Origins.push_back(TrueOrigin);
2705 Origins.push_back(FalseOrigin);
2706 }
2707 } else {
2708 if (TrueShadow == FalseShadow) {
2709 ShadowSel = TrueShadow;
2710 if (ShouldTrackOrigins) {
2711 Shadows.push_back(TrueShadow);
2712 Origins.push_back(TrueOrigin);
2713 }
2714 } else {
2715 ShadowSel =
2716 SelectInst::Create(I.getCondition(), TrueShadow, FalseShadow, "", &I);
2717 if (ShouldTrackOrigins) {
2718 Shadows.push_back(ShadowSel);
2719 Origins.push_back(SelectInst::Create(I.getCondition(), TrueOrigin,
2720 FalseOrigin, "", &I));
2721 }
2722 }
2723 }
2724 DFSF.setShadow(&I, ClTrackSelectControlFlow
2725 ? DFSF.combineShadowsThenConvert(
2726 I.getType(), CondShadow, ShadowSel, &I)
2727 : ShadowSel);
2728 if (ShouldTrackOrigins) {
2729 if (ClTrackSelectControlFlow) {
2730 Shadows.push_back(CondShadow);
2731 Origins.push_back(DFSF.getOrigin(I.getCondition()));
2732 }
2733 DFSF.setOrigin(&I, DFSF.combineOrigins(Shadows, Origins, &I));
2734 }
2735}
2736
2737void DFSanVisitor::visitMemSetInst(MemSetInst &I) {
2738 IRBuilder<> IRB(&I);
2739 Value *ValShadow = DFSF.getShadow(I.getValue());
2740 Value *ValOrigin = DFSF.DFS.shouldTrackOrigins()
2741 ? DFSF.getOrigin(I.getValue())
2742 : DFSF.DFS.ZeroOrigin;
2743 IRB.CreateCall(
2744 DFSF.DFS.DFSanSetLabelFn,
2745 {ValShadow, ValOrigin,
2746 IRB.CreateBitCast(I.getDest(), Type::getInt8PtrTy(*DFSF.DFS.Ctx)),
2747 IRB.CreateZExtOrTrunc(I.getLength(), DFSF.DFS.IntptrTy)});
2748}
2749
2750void DFSanVisitor::visitMemTransferInst(MemTransferInst &I) {
2751 IRBuilder<> IRB(&I);
2752
2753 // CopyOrMoveOrigin transfers origins by refering to their shadows. So we
2754 // need to move origins before moving shadows.
2755 if (DFSF.DFS.shouldTrackOrigins()) {
2756 IRB.CreateCall(
2757 DFSF.DFS.DFSanMemOriginTransferFn,
2758 {IRB.CreatePointerCast(I.getArgOperand(0), IRB.getInt8PtrTy()),
2759 IRB.CreatePointerCast(I.getArgOperand(1), IRB.getInt8PtrTy()),
2760 IRB.CreateIntCast(I.getArgOperand(2), DFSF.DFS.IntptrTy, false)});
2761 }
2762
2763 Value *RawDestShadow = DFSF.DFS.getShadowAddress(I.getDest(), &I);
2764 Value *SrcShadow = DFSF.DFS.getShadowAddress(I.getSource(), &I);
2765 Value *LenShadow =
2766 IRB.CreateMul(I.getLength(), ConstantInt::get(I.getLength()->getType(),
2767 DFSF.DFS.ShadowWidthBytes));
2768 Type *Int8Ptr = Type::getInt8PtrTy(*DFSF.DFS.Ctx);
2769 Value *DestShadow = IRB.CreateBitCast(RawDestShadow, Int8Ptr);
2770 SrcShadow = IRB.CreateBitCast(SrcShadow, Int8Ptr);
2771 auto *MTI = cast<MemTransferInst>(
2772 IRB.CreateCall(I.getFunctionType(), I.getCalledOperand(),
2773 {DestShadow, SrcShadow, LenShadow, I.getVolatileCst()}));
2774 if (ClPreserveAlignment) {
2775 MTI->setDestAlignment(I.getDestAlign() * DFSF.DFS.ShadowWidthBytes);
2776 MTI->setSourceAlignment(I.getSourceAlign() * DFSF.DFS.ShadowWidthBytes);
2777 } else {
2778 MTI->setDestAlignment(Align(DFSF.DFS.ShadowWidthBytes));
2779 MTI->setSourceAlignment(Align(DFSF.DFS.ShadowWidthBytes));
2780 }
2781 if (ClEventCallbacks) {
2782 IRB.CreateCall(DFSF.DFS.DFSanMemTransferCallbackFn,
2783 {RawDestShadow,
2784 IRB.CreateZExtOrTrunc(I.getLength(), DFSF.DFS.IntptrTy)});
2785 }
2786}
2787
2788static bool isAMustTailRetVal(Value *RetVal) {
2789 // Tail call may have a bitcast between return.
2790 if (auto *I = dyn_cast<BitCastInst>(RetVal)) {
2791 RetVal = I->getOperand(0);
2792 }
2793 if (auto *I = dyn_cast<CallInst>(RetVal)) {
2794 return I->isMustTailCall();
2795 }
2796 return false;
2797}
2798
2799void DFSanVisitor::visitReturnInst(ReturnInst &RI) {
2800 if (!DFSF.IsNativeABI && RI.getReturnValue()) {
2801 switch (DFSF.IA) {
2802 case DataFlowSanitizer::IA_TLS: {
2803 // Don't emit the instrumentation for musttail call returns.
2804 if (isAMustTailRetVal(RI.getReturnValue()))
2805 return;
2806
2807 Value *S = DFSF.getShadow(RI.getReturnValue());
2808 IRBuilder<> IRB(&RI);
2809 Type *RT = DFSF.F->getFunctionType()->getReturnType();
2810 unsigned Size =
2811 getDataLayout().getTypeAllocSize(DFSF.DFS.getShadowTy(RT));
2812 if (Size <= RetvalTLSSize) {
2813 // If the size overflows, stores nothing. At callsite, oversized return
2814 // shadows are set to zero.
2815 IRB.CreateAlignedStore(S, DFSF.getRetvalTLS(RT, IRB),
2816 ShadowTLSAlignment);
2817 }
2818 if (DFSF.DFS.shouldTrackOrigins()) {
2819 Value *O = DFSF.getOrigin(RI.getReturnValue());
2820 IRB.CreateStore(O, DFSF.getRetvalOriginTLS());
2821 }
2822 break;
2823 }
2824 case DataFlowSanitizer::IA_Args: {
2825 // TODO: handle musttail call returns for IA_Args.
2826
2827 IRBuilder<> IRB(&RI);
2828 Type *RT = DFSF.F->getFunctionType()->getReturnType();
2829 Value *InsVal =
2830 IRB.CreateInsertValue(UndefValue::get(RT), RI.getReturnValue(), 0);
2831 Value *InsShadow =
2832 IRB.CreateInsertValue(InsVal, DFSF.getShadow(RI.getReturnValue()), 1);
2833 RI.setOperand(0, InsShadow);
2834 break;
2835 }
2836 }
2837 }
2838}
2839
2840void DFSanVisitor::addShadowArguments(Function &F, CallBase &CB,
2841 std::vector<Value *> &Args,
2842 IRBuilder<> &IRB) {
2843 FunctionType *FT = F.getFunctionType();
2844
2845 auto *I = CB.arg_begin();
2846
2847 // Adds non-variable argument shadows.
2848 for (unsigned N = FT->getNumParams(); N != 0; ++I, --N)
2849 Args.push_back(DFSF.collapseToPrimitiveShadow(DFSF.getShadow(*I), &CB));
2850
2851 // Adds variable argument shadows.
2852 if (FT->isVarArg()) {
2853 auto *LabelVATy = ArrayType::get(DFSF.DFS.PrimitiveShadowTy,
2854 CB.arg_size() - FT->getNumParams());
2855 auto *LabelVAAlloca =
2856 new AllocaInst(LabelVATy, getDataLayout().getAllocaAddrSpace(),
2857 "labelva", &DFSF.F->getEntryBlock().front());
2858
2859 for (unsigned N = 0; I != CB.arg_end(); ++I, ++N) {
2860 auto *LabelVAPtr = IRB.CreateStructGEP(LabelVATy, LabelVAAlloca, N);
2861 IRB.CreateStore(DFSF.collapseToPrimitiveShadow(DFSF.getShadow(*I), &CB),
2862 LabelVAPtr);
2863 }
2864
2865 Args.push_back(IRB.CreateStructGEP(LabelVATy, LabelVAAlloca, 0));
2866 }
2867
2868 // Adds the return value shadow.
2869 if (!FT->getReturnType()->isVoidTy()) {
2870 if (!DFSF.LabelReturnAlloca) {
2871 DFSF.LabelReturnAlloca = new AllocaInst(
2872 DFSF.DFS.PrimitiveShadowTy, getDataLayout().getAllocaAddrSpace(),
2873 "labelreturn", &DFSF.F->getEntryBlock().front());
2874 }
2875 Args.push_back(DFSF.LabelReturnAlloca);
2876 }
2877}
2878
2879void DFSanVisitor::addOriginArguments(Function &F, CallBase &CB,
2880 std::vector<Value *> &Args,
2881 IRBuilder<> &IRB) {
2882 FunctionType *FT = F.getFunctionType();
2883
2884 auto *I = CB.arg_begin();
2885
2886 // Add non-variable argument origins.
2887 for (unsigned N = FT->getNumParams(); N != 0; ++I, --N)
2888 Args.push_back(DFSF.getOrigin(*I));
2889
2890 // Add variable argument origins.
2891 if (FT->isVarArg()) {
2892 auto *OriginVATy =
2893 ArrayType::get(DFSF.DFS.OriginTy, CB.arg_size() - FT->getNumParams());
2894 auto *OriginVAAlloca =
2895 new AllocaInst(OriginVATy, getDataLayout().getAllocaAddrSpace(),
2896 "originva", &DFSF.F->getEntryBlock().front());
2897
2898 for (unsigned N = 0; I != CB.arg_end(); ++I, ++N) {
2899 auto *OriginVAPtr = IRB.CreateStructGEP(OriginVATy, OriginVAAlloca, N);
2900 IRB.CreateStore(DFSF.getOrigin(*I), OriginVAPtr);
2901 }
2902
2903 Args.push_back(IRB.CreateStructGEP(OriginVATy, OriginVAAlloca, 0));
2904 }
2905
2906 // Add the return value origin.
2907 if (!FT->getReturnType()->isVoidTy()) {
2908 if (!DFSF.OriginReturnAlloca) {
2909 DFSF.OriginReturnAlloca = new AllocaInst(
2910 DFSF.DFS.OriginTy, getDataLayout().getAllocaAddrSpace(),
2911 "originreturn", &DFSF.F->getEntryBlock().front());
2912 }
2913 Args.push_back(DFSF.OriginReturnAlloca);
2914 }
2915}
2916
2917bool DFSanVisitor::visitWrappedCallBase(Function &F, CallBase &CB) {
2918 IRBuilder<> IRB(&CB);
2919 switch (DFSF.DFS.getWrapperKind(&F)) {
2920 case DataFlowSanitizer::WK_Warning:
2921 CB.setCalledFunction(&F);
2922 IRB.CreateCall(DFSF.DFS.DFSanUnimplementedFn,
2923 IRB.CreateGlobalStringPtr(F.getName()));
2924 DFSF.setShadow(&CB, DFSF.DFS.getZeroShadow(&CB));
2925 DFSF.setOrigin(&CB, DFSF.DFS.ZeroOrigin);
2926 return true;
2927 case DataFlowSanitizer::WK_Discard:
2928 CB.setCalledFunction(&F);
2929 DFSF.setShadow(&CB, DFSF.DFS.getZeroShadow(&CB));
2930 DFSF.setOrigin(&CB, DFSF.DFS.ZeroOrigin);
2931 return true;
2932 case DataFlowSanitizer::WK_Functional:
2933 CB.setCalledFunction(&F);
2934 visitInstOperands(CB);
2935 return true;
2936 case DataFlowSanitizer::WK_Custom:
2937 // Don't try to handle invokes of custom functions, it's too complicated.
2938 // Instead, invoke the dfsw$ wrapper, which will in turn call the __dfsw_
2939 // wrapper.
2940 CallInst *CI = dyn_cast<CallInst>(&CB);
2941 if (!CI)
2942 return false;
2943
2944 const bool ShouldTrackOrigins = DFSF.DFS.shouldTrackOrigins();
2945 FunctionType *FT = F.getFunctionType();
2946 TransformedFunction CustomFn = DFSF.DFS.getCustomFunctionType(FT);
2947 std::string CustomFName = ShouldTrackOrigins ? "__dfso_" : "__dfsw_";
2948 CustomFName += F.getName();
2949 FunctionCallee CustomF = DFSF.DFS.Mod->getOrInsertFunction(
2950 CustomFName, CustomFn.TransformedType);
2951 if (Function *CustomFn = dyn_cast<Function>(CustomF.getCallee())) {
2952 CustomFn->copyAttributesFrom(&F);
2953
2954 // Custom functions returning non-void will write to the return label.
2955 if (!FT->getReturnType()->isVoidTy()) {
2956 CustomFn->removeAttributes(AttributeList::FunctionIndex,
2957 DFSF.DFS.ReadOnlyNoneAttrs);
2958 }
2959 }
2960
2961 std::vector<Value *> Args;
2962
2963 // Adds non-variable arguments.
2964 auto *I = CB.arg_begin();
2965 for (unsigned N = FT->getNumParams(); N != 0; ++I, --N) {
2966 Type *T = (*I)->getType();
2967 FunctionType *ParamFT;
2968 if (isa<PointerType>(T) &&
2969 (ParamFT = dyn_cast<FunctionType>(T->getPointerElementType()))) {
2970 std::string TName = "dfst";
2971 TName += utostr(FT->getNumParams() - N);
2972 TName += "$";
2973 TName += F.getName();
2974 Constant *Trampoline =
2975 DFSF.DFS.getOrBuildTrampolineFunction(ParamFT, TName);
2976 Args.push_back(Trampoline);
2977 Args.push_back(
2978 IRB.CreateBitCast(*I, Type::getInt8PtrTy(*DFSF.DFS.Ctx)));
2979 } else {
2980 Args.push_back(*I);
2981 }
2982 }
2983
2984 // Adds shadow arguments.
2985 const unsigned ShadowArgStart = Args.size();
2986 addShadowArguments(F, CB, Args, IRB);
2987
2988 // Adds origin arguments.
2989 const unsigned OriginArgStart = Args.size();
2990 if (ShouldTrackOrigins)
2991 addOriginArguments(F, CB, Args, IRB);
2992
2993 // Adds variable arguments.
2994 append_range(Args, drop_begin(CB.args(), FT->getNumParams()));
2995
2996 CallInst *CustomCI = IRB.CreateCall(CustomF, Args);
2997 CustomCI->setCallingConv(CI->getCallingConv());
2998 CustomCI->setAttributes(transformFunctionAttributes(
2999 CustomFn, CI->getContext(), CI->getAttributes()));
3000
3001 // Update the parameter attributes of the custom call instruction to
3002 // zero extend the shadow parameters. This is required for targets
3003 // which consider PrimitiveShadowTy an illegal type.
3004 for (unsigned N = 0; N < FT->getNumParams(); N++) {
3005 const unsigned ArgNo = ShadowArgStart + N;
3006 if (CustomCI->getArgOperand(ArgNo)->getType() ==
3007 DFSF.DFS.PrimitiveShadowTy)
3008 CustomCI->addParamAttr(ArgNo, Attribute::ZExt);
3009 if (ShouldTrackOrigins) {
3010 const unsigned OriginArgNo = OriginArgStart + N;
3011 if (CustomCI->getArgOperand(OriginArgNo)->getType() ==
3012 DFSF.DFS.OriginTy)
3013 CustomCI->addParamAttr(OriginArgNo, Attribute::ZExt);
3014 }
3015 }
3016
3017 // Loads the return value shadow and origin.
3018 if (!FT->getReturnType()->isVoidTy()) {
3019 LoadInst *LabelLoad =
3020 IRB.CreateLoad(DFSF.DFS.PrimitiveShadowTy, DFSF.LabelReturnAlloca);
3021 DFSF.setShadow(CustomCI, DFSF.expandFromPrimitiveShadow(
3022 FT->getReturnType(), LabelLoad, &CB));
3023 if (ShouldTrackOrigins) {
3024 LoadInst *OriginLoad =
3025 IRB.CreateLoad(DFSF.DFS.OriginTy, DFSF.OriginReturnAlloca);
3026 DFSF.setOrigin(CustomCI, OriginLoad);
3027 }
3028 }
3029
3030 CI->replaceAllUsesWith(CustomCI);
3031 CI->eraseFromParent();
3032 return true;
3033 }
3034 return false;
3035}
3036
3037void DFSanVisitor::visitCallBase(CallBase &CB) {
3038 Function *F = CB.getCalledFunction();
3039 if ((F && F->isIntrinsic()) || CB.isInlineAsm()) {
3040 visitInstOperands(CB);
3041 return;
3042 }
3043
3044 // Calls to this function are synthesized in wrappers, and we shouldn't
3045 // instrument them.
3046 if (F == DFSF.DFS.DFSanVarargWrapperFn.getCallee()->stripPointerCasts())
3047 return;
3048
3049 DenseMap<Value *, Function *>::iterator UnwrappedFnIt =
3050 DFSF.DFS.UnwrappedFnMap.find(CB.getCalledOperand());
3051 if (UnwrappedFnIt != DFSF.DFS.UnwrappedFnMap.end())
3052 if (visitWrappedCallBase(*UnwrappedFnIt->second, CB))
3053 return;
3054
3055 IRBuilder<> IRB(&CB);
3056
3057 const bool ShouldTrackOrigins = DFSF.DFS.shouldTrackOrigins();
3058 FunctionType *FT = CB.getFunctionType();
3059 if (DFSF.DFS.getInstrumentedABI() == DataFlowSanitizer::IA_TLS) {
3060 // Stores argument shadows.
3061 unsigned ArgOffset = 0;
3062 const DataLayout &DL = getDataLayout();
3063 for (unsigned I = 0, N = FT->getNumParams(); I != N; ++I) {
3064 if (ShouldTrackOrigins) {
3065 // Ignore overflowed origins
3066 Value *ArgShadow = DFSF.getShadow(CB.getArgOperand(I));
3067 if (I < DFSF.DFS.NumOfElementsInArgOrgTLS &&
3068 !DFSF.DFS.isZeroShadow(ArgShadow))
3069 IRB.CreateStore(DFSF.getOrigin(CB.getArgOperand(I)),
3070 DFSF.getArgOriginTLS(I, IRB));
3071 }
3072
3073 unsigned Size =
3074 DL.getTypeAllocSize(DFSF.DFS.getShadowTy(FT->getParamType(I)));
3075 // Stop storing if arguments' size overflows. Inside a function, arguments
3076 // after overflow have zero shadow values.
3077 if (ArgOffset + Size > ArgTLSSize)
3078 break;
3079 IRB.CreateAlignedStore(
3080 DFSF.getShadow(CB.getArgOperand(I)),
3081 DFSF.getArgTLS(FT->getParamType(I), ArgOffset, IRB),
3082 ShadowTLSAlignment);
3083 ArgOffset += alignTo(Size, ShadowTLSAlignment);
3084 }
3085 }
3086
3087 Instruction *Next = nullptr;
3088 if (!CB.getType()->isVoidTy()) {
3089 if (InvokeInst *II = dyn_cast<InvokeInst>(&CB)) {
3090 if (II->getNormalDest()->getSinglePredecessor()) {
3091 Next = &II->getNormalDest()->front();
3092 } else {
3093 BasicBlock *NewBB =
3094 SplitEdge(II->getParent(), II->getNormalDest(), &DFSF.DT);
3095 Next = &NewBB->front();
3096 }
3097 } else {
3098 assert(CB.getIterator() != CB.getParent()->end())((void)0);
3099 Next = CB.getNextNode();
3100 }
3101
3102 if (DFSF.DFS.getInstrumentedABI() == DataFlowSanitizer::IA_TLS) {
3103 // Don't emit the epilogue for musttail call returns.
3104 if (isa<CallInst>(CB) && cast<CallInst>(CB).isMustTailCall())
3105 return;
3106
3107 // Loads the return value shadow.
3108 IRBuilder<> NextIRB(Next);
3109 const DataLayout &DL = getDataLayout();
3110 unsigned Size = DL.getTypeAllocSize(DFSF.DFS.getShadowTy(&CB));
3111 if (Size > RetvalTLSSize) {
3112 // Set overflowed return shadow to be zero.
3113 DFSF.setShadow(&CB, DFSF.DFS.getZeroShadow(&CB));
3114 } else {
3115 LoadInst *LI = NextIRB.CreateAlignedLoad(
3116 DFSF.DFS.getShadowTy(&CB), DFSF.getRetvalTLS(CB.getType(), NextIRB),
3117 ShadowTLSAlignment, "_dfsret");
3118 DFSF.SkipInsts.insert(LI);
3119 DFSF.setShadow(&CB, LI);
3120 DFSF.NonZeroChecks.push_back(LI);
3121 }
3122
3123 if (ShouldTrackOrigins) {
3124 LoadInst *LI = NextIRB.CreateLoad(
3125 DFSF.DFS.OriginTy, DFSF.getRetvalOriginTLS(), "_dfsret_o");
3126 DFSF.SkipInsts.insert(LI);
3127 DFSF.setOrigin(&CB, LI);
3128 }
3129 }
3130 }
3131
3132 // Do all instrumentation for IA_Args down here to defer tampering with the
3133 // CFG in a way that SplitEdge may be able to detect.
3134 if (DFSF.DFS.getInstrumentedABI() == DataFlowSanitizer::IA_Args) {
3135 // TODO: handle musttail call returns for IA_Args.
3136
3137 FunctionType *NewFT = DFSF.DFS.getArgsFunctionType(FT);
3138 Value *Func =
3139 IRB.CreateBitCast(CB.getCalledOperand(), PointerType::getUnqual(NewFT));
3140
3141 const unsigned NumParams = FT->getNumParams();
3142
3143 // Copy original arguments.
3144 auto *ArgIt = CB.arg_begin(), *ArgEnd = CB.arg_end();
3145 std::vector<Value *> Args(NumParams);
3146 std::copy_n(ArgIt, NumParams, Args.begin());
3147
3148 // Add shadow arguments by transforming original arguments.
3149 std::generate_n(std::back_inserter(Args), NumParams,
3150 [&]() { return DFSF.getShadow(*ArgIt++); });
3151
3152 if (FT->isVarArg()) {
3153 unsigned VarArgSize = CB.arg_size() - NumParams;
3154 ArrayType *VarArgArrayTy =
3155 ArrayType::get(DFSF.DFS.PrimitiveShadowTy, VarArgSize);
3156 AllocaInst *VarArgShadow =
3157 new AllocaInst(VarArgArrayTy, getDataLayout().getAllocaAddrSpace(),
3158 "", &DFSF.F->getEntryBlock().front());
3159 Args.push_back(IRB.CreateConstGEP2_32(VarArgArrayTy, VarArgShadow, 0, 0));
3160
3161 // Copy remaining var args.
3162 unsigned GepIndex = 0;
3163 std::for_each(ArgIt, ArgEnd, [&](Value *Arg) {
3164 IRB.CreateStore(
3165 DFSF.getShadow(Arg),
3166 IRB.CreateConstGEP2_32(VarArgArrayTy, VarArgShadow, 0, GepIndex++));
3167 Args.push_back(Arg);
3168 });
3169 }
3170
3171 CallBase *NewCB;
3172 if (InvokeInst *II = dyn_cast<InvokeInst>(&CB)) {
3173 NewCB = IRB.CreateInvoke(NewFT, Func, II->getNormalDest(),
3174 II->getUnwindDest(), Args);
3175 } else {
3176 NewCB = IRB.CreateCall(NewFT, Func, Args);
3177 }
3178 NewCB->setCallingConv(CB.getCallingConv());
3179 NewCB->setAttributes(CB.getAttributes().removeAttributes(
3180 *DFSF.DFS.Ctx, AttributeList::ReturnIndex,
3181 AttributeFuncs::typeIncompatible(NewCB->getType())));
3182
3183 if (Next) {
3184 ExtractValueInst *ExVal = ExtractValueInst::Create(NewCB, 0, "", Next);
3185 DFSF.SkipInsts.insert(ExVal);
3186 ExtractValueInst *ExShadow = ExtractValueInst::Create(NewCB, 1, "", Next);
3187 DFSF.SkipInsts.insert(ExShadow);
3188 DFSF.setShadow(ExVal, ExShadow);
3189 DFSF.NonZeroChecks.push_back(ExShadow);
3190
3191 CB.replaceAllUsesWith(ExVal);
3192 }
3193
3194 CB.eraseFromParent();
3195 }
3196}
3197
3198void DFSanVisitor::visitPHINode(PHINode &PN) {
3199 Type *ShadowTy = DFSF.DFS.getShadowTy(&PN);
3200 PHINode *ShadowPN =
3201 PHINode::Create(ShadowTy, PN.getNumIncomingValues(), "", &PN);
3202
3203 // Give the shadow phi node valid predecessors to fool SplitEdge into working.
3204 Value *UndefShadow = UndefValue::get(ShadowTy);
3205 for (BasicBlock *BB : PN.blocks())
3206 ShadowPN->addIncoming(UndefShadow, BB);
3207
3208 DFSF.setShadow(&PN, ShadowPN);
3209
3210 PHINode *OriginPN = nullptr;
3211 if (DFSF.DFS.shouldTrackOrigins()) {
3212 OriginPN =
3213 PHINode::Create(DFSF.DFS.OriginTy, PN.getNumIncomingValues(), "", &PN);
3214 Value *UndefOrigin = UndefValue::get(DFSF.DFS.OriginTy);
3215 for (BasicBlock *BB : PN.blocks())
3216 OriginPN->addIncoming(UndefOrigin, BB);
3217 DFSF.setOrigin(&PN, OriginPN);
3218 }
3219
3220 DFSF.PHIFixups.push_back({&PN, ShadowPN, OriginPN});
3221}
3222
3223namespace {
3224class DataFlowSanitizerLegacyPass : public ModulePass {
3225private:
3226 std::vector<std::string> ABIListFiles;
3227
3228public:
3229 static char ID;
3230
3231 DataFlowSanitizerLegacyPass(
3232 const std::vector<std::string> &ABIListFiles = std::vector<std::string>())
3233 : ModulePass(ID), ABIListFiles(ABIListFiles) {}
3234
3235 bool runOnModule(Module &M) override {
3236 return DataFlowSanitizer(ABIListFiles).runImpl(M);
3237 }
3238};
3239} // namespace
3240
3241char DataFlowSanitizerLegacyPass::ID;
3242
3243INITIALIZE_PASS(DataFlowSanitizerLegacyPass, "dfsan",static void *initializeDataFlowSanitizerLegacyPassPassOnce(PassRegistry
&Registry) { PassInfo *PI = new PassInfo( "DataFlowSanitizer: dynamic data flow analysis."
, "dfsan", &DataFlowSanitizerLegacyPass::ID, PassInfo::NormalCtor_t
(callDefaultCtor<DataFlowSanitizerLegacyPass>), false, false
); Registry.registerPass(*PI, true); return PI; } static llvm
::once_flag InitializeDataFlowSanitizerLegacyPassPassFlag; void
llvm::initializeDataFlowSanitizerLegacyPassPass(PassRegistry
&Registry) { llvm::call_once(InitializeDataFlowSanitizerLegacyPassPassFlag
, initializeDataFlowSanitizerLegacyPassPassOnce, std::ref(Registry
)); }
3244 "DataFlowSanitizer: dynamic data flow analysis.", false, false)static void *initializeDataFlowSanitizerLegacyPassPassOnce(PassRegistry
&Registry) { PassInfo *PI = new PassInfo( "DataFlowSanitizer: dynamic data flow analysis."
, "dfsan", &DataFlowSanitizerLegacyPass::ID, PassInfo::NormalCtor_t
(callDefaultCtor<DataFlowSanitizerLegacyPass>), false, false
); Registry.registerPass(*PI, true); return PI; } static llvm
::once_flag InitializeDataFlowSanitizerLegacyPassPassFlag; void
llvm::initializeDataFlowSanitizerLegacyPassPass(PassRegistry
&Registry) { llvm::call_once(InitializeDataFlowSanitizerLegacyPassPassFlag
, initializeDataFlowSanitizerLegacyPassPassOnce, std::ref(Registry
)); }
3245
3246ModulePass *llvm::createDataFlowSanitizerLegacyPassPass(
3247 const std::vector<std::string> &ABIListFiles) {
3248 return new DataFlowSanitizerLegacyPass(ABIListFiles);
3249}
3250
3251PreservedAnalyses DataFlowSanitizerPass::run(Module &M,
3252 ModuleAnalysisManager &AM) {
3253 if (DataFlowSanitizer(ABIListFiles).runImpl(M)) {
3254 return PreservedAnalyses::none();
3255 }
3256 return PreservedAnalyses::all();
3257}

/usr/src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Support/Alignment.h

1//===-- llvm/Support/Alignment.h - Useful alignment functions ---*- C++ -*-===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file contains types to represent alignments.
10// They are instrumented to guarantee some invariants are preserved and prevent
11// invalid manipulations.
12//
13// - Align represents an alignment in bytes, it is always set and always a valid
14// power of two, its minimum value is 1 which means no alignment requirements.
15//
16// - MaybeAlign is an optional type, it may be undefined or set. When it's set
17// you can get the underlying Align type by using the getValue() method.
18//
19//===----------------------------------------------------------------------===//
20
21#ifndef LLVM_SUPPORT_ALIGNMENT_H_
22#define LLVM_SUPPORT_ALIGNMENT_H_
23
24#include "llvm/ADT/Optional.h"
25#include "llvm/Support/MathExtras.h"
26#include <cassert>
27#ifndef NDEBUG1
28#include <string>
29#endif // NDEBUG
30
31namespace llvm {
32
33#define ALIGN_CHECK_ISPOSITIVE(decl) \
34 assert(decl > 0 && (#decl " should be defined"))((void)0)
35
36/// This struct is a compact representation of a valid (non-zero power of two)
37/// alignment.
38/// It is suitable for use as static global constants.
39struct Align {
40private:
41 uint8_t ShiftValue = 0; /// The log2 of the required alignment.
42 /// ShiftValue is less than 64 by construction.
43
44 friend struct MaybeAlign;
45 friend unsigned Log2(Align);
46 friend bool operator==(Align Lhs, Align Rhs);
47 friend bool operator!=(Align Lhs, Align Rhs);
48 friend bool operator<=(Align Lhs, Align Rhs);
49 friend bool operator>=(Align Lhs, Align Rhs);
50 friend bool operator<(Align Lhs, Align Rhs);
51 friend bool operator>(Align Lhs, Align Rhs);
52 friend unsigned encode(struct MaybeAlign A);
53 friend struct MaybeAlign decodeMaybeAlign(unsigned Value);
54
55 /// A trivial type to allow construction of constexpr Align.
56 /// This is currently needed to workaround a bug in GCC 5.3 which prevents
57 /// definition of constexpr assign operators.
58 /// https://stackoverflow.com/questions/46756288/explicitly-defaulted-function-cannot-be-declared-as-constexpr-because-the-implic
59 /// FIXME: Remove this, make all assign operators constexpr and introduce user
60 /// defined literals when we don't have to support GCC 5.3 anymore.
61 /// https://llvm.org/docs/GettingStarted.html#getting-a-modern-host-c-toolchain
62 struct LogValue {
63 uint8_t Log;
64 };
65
66public:
67 /// Default is byte-aligned.
68 constexpr Align() = default;
69 /// Do not perform checks in case of copy/move construct/assign, because the
70 /// checks have been performed when building `Other`.
71 constexpr Align(const Align &Other) = default;
72 constexpr Align(Align &&Other) = default;
73 Align &operator=(const Align &Other) = default;
74 Align &operator=(Align &&Other) = default;
75
76 explicit Align(uint64_t Value) {
77 assert(Value > 0 && "Value must not be 0")((void)0);
78 assert(llvm::isPowerOf2_64(Value) && "Alignment is not a power of 2")((void)0);
79 ShiftValue = Log2_64(Value);
80 assert(ShiftValue < 64 && "Broken invariant")((void)0);
81 }
82
83 /// This is a hole in the type system and should not be abused.
84 /// Needed to interact with C for instance.
85 uint64_t value() const { return uint64_t(1) << ShiftValue; }
9
The result of the left shift is undefined due to shifting by '255', which is greater or equal to the width of type 'uint64_t'
86
87 /// Allow constructions of constexpr Align.
88 template <size_t kValue> constexpr static LogValue Constant() {
89 return LogValue{static_cast<uint8_t>(CTLog2<kValue>())};
90 }
91
92 /// Allow constructions of constexpr Align from types.
93 /// Compile time equivalent to Align(alignof(T)).
94 template <typename T> constexpr static LogValue Of() {
95 return Constant<std::alignment_of<T>::value>();
96 }
97
98 /// Constexpr constructor from LogValue type.
99 constexpr Align(LogValue CA) : ShiftValue(CA.Log) {}
100};
101
102/// Treats the value 0 as a 1, so Align is always at least 1.
103inline Align assumeAligned(uint64_t Value) {
104 return Value ? Align(Value) : Align();
105}
106
107/// This struct is a compact representation of a valid (power of two) or
108/// undefined (0) alignment.
109struct MaybeAlign : public llvm::Optional<Align> {
110private:
111 using UP = llvm::Optional<Align>;
112
113public:
114 /// Default is undefined.
115 MaybeAlign() = default;
116 /// Do not perform checks in case of copy/move construct/assign, because the
117 /// checks have been performed when building `Other`.
118 MaybeAlign(const MaybeAlign &Other) = default;
119 MaybeAlign &operator=(const MaybeAlign &Other) = default;
120 MaybeAlign(MaybeAlign &&Other) = default;
121 MaybeAlign &operator=(MaybeAlign &&Other) = default;
122
123 /// Use llvm::Optional<Align> constructor.
124 using UP::UP;
125
126 explicit MaybeAlign(uint64_t Value) {
127 assert((Value == 0 || llvm::isPowerOf2_64(Value)) &&((void)0)
128 "Alignment is neither 0 nor a power of 2")((void)0);
129 if (Value)
130 emplace(Value);
131 }
132
133 /// For convenience, returns a valid alignment or 1 if undefined.
134 Align valueOrOne() const { return hasValue() ? getValue() : Align(); }
135};
136
137/// Checks that SizeInBytes is a multiple of the alignment.
138inline bool isAligned(Align Lhs, uint64_t SizeInBytes) {
139 return SizeInBytes % Lhs.value() == 0;
140}
141
142/// Checks that Addr is a multiple of the alignment.
143inline bool isAddrAligned(Align Lhs, const void *Addr) {
144 return isAligned(Lhs, reinterpret_cast<uintptr_t>(Addr));
145}
146
147/// Returns a multiple of A needed to store `Size` bytes.
148inline uint64_t alignTo(uint64_t Size, Align A) {
149 const uint64_t Value = A.value();
150 // The following line is equivalent to `(Size + Value - 1) / Value * Value`.
151
152 // The division followed by a multiplication can be thought of as a right
153 // shift followed by a left shift which zeros out the extra bits produced in
154 // the bump; `~(Value - 1)` is a mask where all those bits being zeroed out
155 // are just zero.
156
157 // Most compilers can generate this code but the pattern may be missed when
158 // multiple functions gets inlined.
159 return (Size + Value - 1) & ~(Value - 1U);
160}
161
162/// If non-zero \p Skew is specified, the return value will be a minimal integer
163/// that is greater than or equal to \p Size and equal to \p A * N + \p Skew for
164/// some integer N. If \p Skew is larger than \p A, its value is adjusted to '\p
165/// Skew mod \p A'.
166///
167/// Examples:
168/// \code
169/// alignTo(5, Align(8), 7) = 7
170/// alignTo(17, Align(8), 1) = 17
171/// alignTo(~0LL, Align(8), 3) = 3
172/// \endcode
173inline uint64_t alignTo(uint64_t Size, Align A, uint64_t Skew) {
174 const uint64_t Value = A.value();
175 Skew %= Value;
176 return ((Size + Value - 1 - Skew) & ~(Value - 1U)) + Skew;
177}
178
179/// Returns a multiple of A needed to store `Size` bytes.
180/// Returns `Size` if current alignment is undefined.
181inline uint64_t alignTo(uint64_t Size, MaybeAlign A) {
182 return A ? alignTo(Size, A.getValue()) : Size;
183}
184
185/// Aligns `Addr` to `Alignment` bytes, rounding up.
186inline uintptr_t alignAddr(const void *Addr, Align Alignment) {
187 uintptr_t ArithAddr = reinterpret_cast<uintptr_t>(Addr);
188 assert(static_cast<uintptr_t>(ArithAddr + Alignment.value() - 1) >=((void)0)
189 ArithAddr &&((void)0)
190 "Overflow")((void)0);
191 return alignTo(ArithAddr, Alignment);
192}
193
194/// Returns the offset to the next integer (mod 2**64) that is greater than
195/// or equal to \p Value and is a multiple of \p Align.
196inline uint64_t offsetToAlignment(uint64_t Value, Align Alignment) {
197 return alignTo(Value, Alignment) - Value;
198}
199
200/// Returns the necessary adjustment for aligning `Addr` to `Alignment`
201/// bytes, rounding up.
202inline uint64_t offsetToAlignedAddr(const void *Addr, Align Alignment) {
203 return offsetToAlignment(reinterpret_cast<uintptr_t>(Addr), Alignment);
204}
205
206/// Returns the log2 of the alignment.
207inline unsigned Log2(Align A) { return A.ShiftValue; }
208
209/// Returns the alignment that satisfies both alignments.
210/// Same semantic as MinAlign.
211inline Align commonAlignment(Align A, Align B) { return std::min(A, B); }
212
213/// Returns the alignment that satisfies both alignments.
214/// Same semantic as MinAlign.
215inline Align commonAlignment(Align A, uint64_t Offset) {
216 return Align(MinAlign(A.value(), Offset));
217}
218
219/// Returns the alignment that satisfies both alignments.
220/// Same semantic as MinAlign.
221inline MaybeAlign commonAlignment(MaybeAlign A, MaybeAlign B) {
222 return A && B ? commonAlignment(*A, *B) : A ? A : B;
223}
224
225/// Returns the alignment that satisfies both alignments.
226/// Same semantic as MinAlign.
227inline MaybeAlign commonAlignment(MaybeAlign A, uint64_t Offset) {
228 return MaybeAlign(MinAlign((*A).value(), Offset));
229}
230
231/// Returns a representation of the alignment that encodes undefined as 0.
232inline unsigned encode(MaybeAlign A) { return A ? A->ShiftValue + 1 : 0; }
233
234/// Dual operation of the encode function above.
235inline MaybeAlign decodeMaybeAlign(unsigned Value) {
236 if (Value == 0)
237 return MaybeAlign();
238 Align Out;
239 Out.ShiftValue = Value - 1;
240 return Out;
241}
242
243/// Returns a representation of the alignment, the encoded value is positive by
244/// definition.
245inline unsigned encode(Align A) { return encode(MaybeAlign(A)); }
246
247/// Comparisons between Align and scalars. Rhs must be positive.
248inline bool operator==(Align Lhs, uint64_t Rhs) {
249 ALIGN_CHECK_ISPOSITIVE(Rhs);
250 return Lhs.value() == Rhs;
251}
252inline bool operator!=(Align Lhs, uint64_t Rhs) {
253 ALIGN_CHECK_ISPOSITIVE(Rhs);
254 return Lhs.value() != Rhs;
255}
256inline bool operator<=(Align Lhs, uint64_t Rhs) {
257 ALIGN_CHECK_ISPOSITIVE(Rhs);
258 return Lhs.value() <= Rhs;
259}
260inline bool operator>=(Align Lhs, uint64_t Rhs) {
261 ALIGN_CHECK_ISPOSITIVE(Rhs);
262 return Lhs.value() >= Rhs;
263}
264inline bool operator<(Align Lhs, uint64_t Rhs) {
265 ALIGN_CHECK_ISPOSITIVE(Rhs);
266 return Lhs.value() < Rhs;
267}
268inline bool operator>(Align Lhs, uint64_t Rhs) {
269 ALIGN_CHECK_ISPOSITIVE(Rhs);
270 return Lhs.value() > Rhs;
271}
272
273/// Comparisons between MaybeAlign and scalars.
274inline bool operator==(MaybeAlign Lhs, uint64_t Rhs) {
275 return Lhs ? (*Lhs).value() == Rhs : Rhs == 0;
276}
277inline bool operator!=(MaybeAlign Lhs, uint64_t Rhs) {
278 return Lhs ? (*Lhs).value() != Rhs : Rhs != 0;
279}
280
281/// Comparisons operators between Align.
282inline bool operator==(Align Lhs, Align Rhs) {
283 return Lhs.ShiftValue == Rhs.ShiftValue;
284}
285inline bool operator!=(Align Lhs, Align Rhs) {
286 return Lhs.ShiftValue != Rhs.ShiftValue;
287}
288inline bool operator<=(Align Lhs, Align Rhs) {
289 return Lhs.ShiftValue <= Rhs.ShiftValue;
290}
291inline bool operator>=(Align Lhs, Align Rhs) {
292 return Lhs.ShiftValue >= Rhs.ShiftValue;
293}
294inline bool operator<(Align Lhs, Align Rhs) {
295 return Lhs.ShiftValue < Rhs.ShiftValue;
296}
297inline bool operator>(Align Lhs, Align Rhs) {
298 return Lhs.ShiftValue > Rhs.ShiftValue;
299}
300
301// Don't allow relational comparisons with MaybeAlign.
302bool operator<=(Align Lhs, MaybeAlign Rhs) = delete;
303bool operator>=(Align Lhs, MaybeAlign Rhs) = delete;
304bool operator<(Align Lhs, MaybeAlign Rhs) = delete;
305bool operator>(Align Lhs, MaybeAlign Rhs) = delete;
306
307bool operator<=(MaybeAlign Lhs, Align Rhs) = delete;
308bool operator>=(MaybeAlign Lhs, Align Rhs) = delete;
309bool operator<(MaybeAlign Lhs, Align Rhs) = delete;
310bool operator>(MaybeAlign Lhs, Align Rhs) = delete;
311
312bool operator<=(MaybeAlign Lhs, MaybeAlign Rhs) = delete;
313bool operator>=(MaybeAlign Lhs, MaybeAlign Rhs) = delete;
314bool operator<(MaybeAlign Lhs, MaybeAlign Rhs) = delete;
315bool operator>(MaybeAlign Lhs, MaybeAlign Rhs) = delete;
316
317inline Align operator*(Align Lhs, uint64_t Rhs) {
318 assert(Rhs > 0 && "Rhs must be positive")((void)0);
319 return Align(Lhs.value() * Rhs);
320}
321
322inline MaybeAlign operator*(MaybeAlign Lhs, uint64_t Rhs) {
323 assert(Rhs > 0 && "Rhs must be positive")((void)0);
324 return Lhs ? Lhs.getValue() * Rhs : MaybeAlign();
325}
326
327inline Align operator/(Align Lhs, uint64_t Divisor) {
328 assert(llvm::isPowerOf2_64(Divisor) &&((void)0)
329 "Divisor must be positive and a power of 2")((void)0);
330 assert(Lhs != 1 && "Can't halve byte alignment")((void)0);
331 return Align(Lhs.value() / Divisor);
332}
333
334inline MaybeAlign operator/(MaybeAlign Lhs, uint64_t Divisor) {
335 assert(llvm::isPowerOf2_64(Divisor) &&((void)0)
336 "Divisor must be positive and a power of 2")((void)0);
337 return Lhs ? Lhs.getValue() / Divisor : MaybeAlign();
338}
339
340inline Align max(MaybeAlign Lhs, Align Rhs) {
341 return Lhs && *Lhs > Rhs ? *Lhs : Rhs;
342}
343
344inline Align max(Align Lhs, MaybeAlign Rhs) {
345 return Rhs && *Rhs > Lhs ? *Rhs : Lhs;
346}
347
348#ifndef NDEBUG1
349// For usage in LLVM_DEBUG macros.
350inline std::string DebugStr(const Align &A) {
351 return std::to_string(A.value());
352}
353// For usage in LLVM_DEBUG macros.
354inline std::string DebugStr(const MaybeAlign &MA) {
355 if (MA)
356 return std::to_string(MA->value());
357 return "None";
358}
359#endif // NDEBUG
360
361#undef ALIGN_CHECK_ISPOSITIVE
362
363} // namespace llvm
364
365#endif // LLVM_SUPPORT_ALIGNMENT_H_