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/// Argument and return value labels are passed through TLS variables
22/// __dfsan_arg_tls and __dfsan_retval_tls.
23///
24/// Each byte of application memory is backed by a shadow memory byte. The
25/// shadow byte can represent up to 8 labels. On Linux/x86_64, memory is then
26/// laid out as follows:
27///
28/// +--------------------+ 0x800000000000 (top of memory)
29/// | application 3 |
30/// +--------------------+ 0x700000000000
31/// | invalid |
32/// +--------------------+ 0x610000000000
33/// | origin 1 |
34/// +--------------------+ 0x600000000000
35/// | application 2 |
36/// +--------------------+ 0x510000000000
37/// | shadow 1 |
38/// +--------------------+ 0x500000000000
39/// | invalid |
40/// +--------------------+ 0x400000000000
41/// | origin 3 |
42/// +--------------------+ 0x300000000000
43/// | shadow 3 |
44/// +--------------------+ 0x200000000000
45/// | origin 2 |
46/// +--------------------+ 0x110000000000
47/// | invalid |
48/// +--------------------+ 0x100000000000
49/// | shadow 2 |
50/// +--------------------+ 0x010000000000
51/// | application 1 |
52/// +--------------------+ 0x000000000000
53///
54/// MEM_TO_SHADOW(mem) = mem ^ 0x500000000000
55/// SHADOW_TO_ORIGIN(shadow) = shadow + 0x100000000000
56///
57/// For more information, please refer to the design document:
58/// http://clang.llvm.org/docs/DataFlowSanitizerDesign.html
59//
60//===----------------------------------------------------------------------===//
61
62#include "llvm/Transforms/Instrumentation/DataFlowSanitizer.h"
63#include "llvm/ADT/DenseMap.h"
64#include "llvm/ADT/DenseSet.h"
65#include "llvm/ADT/DepthFirstIterator.h"
66#include "llvm/ADT/SmallPtrSet.h"
67#include "llvm/ADT/SmallVector.h"
68#include "llvm/ADT/StringRef.h"
69#include "llvm/ADT/StringSet.h"
70#include "llvm/ADT/iterator.h"
71#include "llvm/Analysis/DomTreeUpdater.h"
72#include "llvm/Analysis/GlobalsModRef.h"
73#include "llvm/Analysis/TargetLibraryInfo.h"
74#include "llvm/Analysis/ValueTracking.h"
75#include "llvm/IR/Argument.h"
76#include "llvm/IR/AttributeMask.h"
77#include "llvm/IR/Attributes.h"
78#include "llvm/IR/BasicBlock.h"
79#include "llvm/IR/Constant.h"
80#include "llvm/IR/Constants.h"
81#include "llvm/IR/DataLayout.h"
82#include "llvm/IR/DerivedTypes.h"
83#include "llvm/IR/Dominators.h"
84#include "llvm/IR/Function.h"
85#include "llvm/IR/GlobalAlias.h"
86#include "llvm/IR/GlobalValue.h"
87#include "llvm/IR/GlobalVariable.h"
88#include "llvm/IR/IRBuilder.h"
89#include "llvm/IR/InstVisitor.h"
90#include "llvm/IR/InstrTypes.h"
91#include "llvm/IR/Instruction.h"
92#include "llvm/IR/Instructions.h"
93#include "llvm/IR/IntrinsicInst.h"
94#include "llvm/IR/MDBuilder.h"
95#include "llvm/IR/Module.h"
96#include "llvm/IR/PassManager.h"
97#include "llvm/IR/Type.h"
98#include "llvm/IR/User.h"
99#include "llvm/IR/Value.h"
100#include "llvm/Support/Alignment.h"
101#include "llvm/Support/Casting.h"
102#include "llvm/Support/CommandLine.h"
103#include "llvm/Support/ErrorHandling.h"
104#include "llvm/Support/SpecialCaseList.h"
105#include "llvm/Support/VirtualFileSystem.h"
106#include "llvm/TargetParser/Triple.h"
107#include "llvm/Transforms/Utils/BasicBlockUtils.h"
108#include "llvm/Transforms/Utils/Instrumentation.h"
109#include "llvm/Transforms/Utils/Local.h"
110#include <algorithm>
111#include <cassert>
112#include <cstddef>
113#include <cstdint>
114#include <memory>
115#include <set>
116#include <string>
117#include <utility>
118#include <vector>
119
120using namespace llvm;
121
122// This must be consistent with ShadowWidthBits.
123static const Align ShadowTLSAlignment = Align(2);
124
125static const Align MinOriginAlignment = Align(4);
126
127// The size of TLS variables. These constants must be kept in sync with the ones
128// in dfsan.cpp.
129static const unsigned ArgTLSSize = 800;
130static const unsigned RetvalTLSSize = 800;
131
132// The -dfsan-preserve-alignment flag controls whether this pass assumes that
133// alignment requirements provided by the input IR are correct. For example,
134// if the input IR contains a load with alignment 8, this flag will cause
135// the shadow load to have alignment 16. This flag is disabled by default as
136// we have unfortunately encountered too much code (including Clang itself;
137// see PR14291) which performs misaligned access.
138static cl::opt<bool> ClPreserveAlignment(
139 "dfsan-preserve-alignment",
140 cl::desc("respect alignment requirements provided by input IR"), cl::Hidden,
141 cl::init(Val: false));
142
143// The ABI list files control how shadow parameters are passed. The pass treats
144// every function labelled "uninstrumented" in the ABI list file as conforming
145// to the "native" (i.e. unsanitized) ABI. Unless the ABI list contains
146// additional annotations for those functions, a call to one of those functions
147// will produce a warning message, as the labelling behaviour of the function is
148// unknown. The other supported annotations for uninstrumented functions are
149// "functional" and "discard", which are described below under
150// DataFlowSanitizer::WrapperKind.
151// Functions will often be labelled with both "uninstrumented" and one of
152// "functional" or "discard". This will leave the function unchanged by this
153// pass, and create a wrapper function that will call the original.
154//
155// Instrumented functions can also be annotated as "force_zero_labels", which
156// will make all shadow and return values set zero labels.
157// Functions should never be labelled with both "force_zero_labels" and
158// "uninstrumented" or any of the unistrumented wrapper kinds.
159static cl::list<std::string> ClABIListFiles(
160 "dfsan-abilist",
161 cl::desc("File listing native ABI functions and how the pass treats them"),
162 cl::Hidden);
163
164// Controls whether the pass includes or ignores the labels of pointers in load
165// instructions.
166static cl::opt<bool> ClCombinePointerLabelsOnLoad(
167 "dfsan-combine-pointer-labels-on-load",
168 cl::desc("Combine the label of the pointer with the label of the data when "
169 "loading from memory."),
170 cl::Hidden, cl::init(Val: true));
171
172// Controls whether the pass includes or ignores the labels of pointers in
173// stores instructions.
174static cl::opt<bool> ClCombinePointerLabelsOnStore(
175 "dfsan-combine-pointer-labels-on-store",
176 cl::desc("Combine the label of the pointer with the label of the data when "
177 "storing in memory."),
178 cl::Hidden, cl::init(Val: false));
179
180// Controls whether the pass propagates labels of offsets in GEP instructions.
181static cl::opt<bool> ClCombineOffsetLabelsOnGEP(
182 "dfsan-combine-offset-labels-on-gep",
183 cl::desc(
184 "Combine the label of the offset with the label of the pointer when "
185 "doing pointer arithmetic."),
186 cl::Hidden, cl::init(Val: true));
187
188static cl::list<std::string> ClCombineTaintLookupTables(
189 "dfsan-combine-taint-lookup-table",
190 cl::desc(
191 "When dfsan-combine-offset-labels-on-gep and/or "
192 "dfsan-combine-pointer-labels-on-load are false, this flag can "
193 "be used to re-enable combining offset and/or pointer taint when "
194 "loading specific constant global variables (i.e. lookup tables)."),
195 cl::Hidden);
196
197static cl::opt<bool> ClDebugNonzeroLabels(
198 "dfsan-debug-nonzero-labels",
199 cl::desc("Insert calls to __dfsan_nonzero_label on observing a parameter, "
200 "load or return with a nonzero label"),
201 cl::Hidden);
202
203// Experimental feature that inserts callbacks for certain data events.
204// Currently callbacks are only inserted for loads, stores, memory transfers
205// (i.e. memcpy and memmove), and comparisons.
206//
207// If this flag is set to true, the user must provide definitions for the
208// following callback functions:
209// void __dfsan_load_callback(dfsan_label Label, void* addr);
210// void __dfsan_store_callback(dfsan_label Label, void* addr);
211// void __dfsan_mem_transfer_callback(dfsan_label *Start, size_t Len);
212// void __dfsan_cmp_callback(dfsan_label CombinedLabel);
213static cl::opt<bool> ClEventCallbacks(
214 "dfsan-event-callbacks",
215 cl::desc("Insert calls to __dfsan_*_callback functions on data events."),
216 cl::Hidden, cl::init(Val: false));
217
218// Experimental feature that inserts callbacks for conditionals, including:
219// conditional branch, switch, select.
220// This must be true for dfsan_set_conditional_callback() to have effect.
221static cl::opt<bool> ClConditionalCallbacks(
222 "dfsan-conditional-callbacks",
223 cl::desc("Insert calls to callback functions on conditionals."), cl::Hidden,
224 cl::init(Val: false));
225
226// Experimental feature that inserts callbacks for data reaching a function,
227// either via function arguments and loads.
228// This must be true for dfsan_set_reaches_function_callback() to have effect.
229static cl::opt<bool> ClReachesFunctionCallbacks(
230 "dfsan-reaches-function-callbacks",
231 cl::desc("Insert calls to callback functions on data reaching a function."),
232 cl::Hidden, cl::init(Val: false));
233
234// Controls whether the pass tracks the control flow of select instructions.
235static cl::opt<bool> ClTrackSelectControlFlow(
236 "dfsan-track-select-control-flow",
237 cl::desc("Propagate labels from condition values of select instructions "
238 "to results."),
239 cl::Hidden, cl::init(Val: true));
240
241// TODO: This default value follows MSan. DFSan may use a different value.
242static cl::opt<int> ClInstrumentWithCallThreshold(
243 "dfsan-instrument-with-call-threshold",
244 cl::desc("If the function being instrumented requires more than "
245 "this number of origin stores, use callbacks instead of "
246 "inline checks (-1 means never use callbacks)."),
247 cl::Hidden, cl::init(Val: 3500));
248
249// Controls how to track origins.
250// * 0: do not track origins.
251// * 1: track origins at memory store operations.
252// * 2: track origins at memory load and store operations.
253// TODO: track callsites.
254static cl::opt<int> ClTrackOrigins("dfsan-track-origins",
255 cl::desc("Track origins of labels"),
256 cl::Hidden, cl::init(Val: 0));
257
258static cl::opt<bool> ClIgnorePersonalityRoutine(
259 "dfsan-ignore-personality-routine",
260 cl::desc("If a personality routine is marked uninstrumented from the ABI "
261 "list, do not create a wrapper for it."),
262 cl::Hidden, cl::init(Val: false));
263
264static StringRef getGlobalTypeString(const GlobalValue &G) {
265 // Types of GlobalVariables are always pointer types.
266 Type *GType = G.getValueType();
267 // For now we support excluding struct types only.
268 if (StructType *SGType = dyn_cast<StructType>(Val: GType)) {
269 if (!SGType->isLiteral())
270 return SGType->getName();
271 }
272 return "<unknown type>";
273}
274
275namespace {
276
277// Memory map parameters used in application-to-shadow address calculation.
278// Offset = (Addr & ~AndMask) ^ XorMask
279// Shadow = ShadowBase + Offset
280// Origin = (OriginBase + Offset) & ~3ULL
281struct MemoryMapParams {
282 uint64_t AndMask;
283 uint64_t XorMask;
284 uint64_t ShadowBase;
285 uint64_t OriginBase;
286};
287
288} // end anonymous namespace
289
290// NOLINTBEGIN(readability-identifier-naming)
291// aarch64 Linux
292const MemoryMapParams Linux_AArch64_MemoryMapParams = {
293 .AndMask: 0, // AndMask (not used)
294 .XorMask: 0x0B00000000000, // XorMask
295 .ShadowBase: 0, // ShadowBase (not used)
296 .OriginBase: 0x0200000000000, // OriginBase
297};
298
299// x86_64 Linux
300const MemoryMapParams Linux_X86_64_MemoryMapParams = {
301 .AndMask: 0, // AndMask (not used)
302 .XorMask: 0x500000000000, // XorMask
303 .ShadowBase: 0, // ShadowBase (not used)
304 .OriginBase: 0x100000000000, // OriginBase
305};
306// NOLINTEND(readability-identifier-naming)
307
308// loongarch64 Linux
309const MemoryMapParams Linux_LoongArch64_MemoryMapParams = {
310 .AndMask: 0, // AndMask (not used)
311 .XorMask: 0x500000000000, // XorMask
312 .ShadowBase: 0, // ShadowBase (not used)
313 .OriginBase: 0x100000000000, // OriginBase
314};
315
316namespace {
317
318class DFSanABIList {
319 std::unique_ptr<SpecialCaseList> SCL;
320
321public:
322 DFSanABIList() = default;
323
324 void set(std::unique_ptr<SpecialCaseList> List) { SCL = std::move(List); }
325
326 /// Returns whether either this function or its source file are listed in the
327 /// given category.
328 bool isIn(const Function &F, StringRef Category) const {
329 return isIn(M: *F.getParent(), Category) ||
330 SCL->inSection(Section: "dataflow", Prefix: "fun", Query: F.getName(), Category);
331 }
332
333 /// Returns whether this global alias is listed in the given category.
334 ///
335 /// If GA aliases a function, the alias's name is matched as a function name
336 /// would be. Similarly, aliases of globals are matched like globals.
337 bool isIn(const GlobalAlias &GA, StringRef Category) const {
338 if (isIn(M: *GA.getParent(), Category))
339 return true;
340
341 if (isa<FunctionType>(Val: GA.getValueType()))
342 return SCL->inSection(Section: "dataflow", Prefix: "fun", Query: GA.getName(), Category);
343
344 return SCL->inSection(Section: "dataflow", Prefix: "global", Query: GA.getName(), Category) ||
345 SCL->inSection(Section: "dataflow", Prefix: "type", Query: getGlobalTypeString(G: GA),
346 Category);
347 }
348
349 /// Returns whether this module is listed in the given category.
350 bool isIn(const Module &M, StringRef Category) const {
351 return SCL->inSection(Section: "dataflow", Prefix: "src", Query: M.getModuleIdentifier(), Category);
352 }
353};
354
355/// TransformedFunction is used to express the result of transforming one
356/// function type into another. This struct is immutable. It holds metadata
357/// useful for updating calls of the old function to the new type.
358struct TransformedFunction {
359 TransformedFunction(FunctionType *OriginalType, FunctionType *TransformedType,
360 const std::vector<unsigned> &ArgumentIndexMapping)
361 : OriginalType(OriginalType), TransformedType(TransformedType),
362 ArgumentIndexMapping(ArgumentIndexMapping) {}
363
364 // Disallow copies.
365 TransformedFunction(const TransformedFunction &) = delete;
366 TransformedFunction &operator=(const TransformedFunction &) = delete;
367
368 // Allow moves.
369 TransformedFunction(TransformedFunction &&) = default;
370 TransformedFunction &operator=(TransformedFunction &&) = default;
371
372 /// Type of the function before the transformation.
373 FunctionType *OriginalType;
374
375 /// Type of the function after the transformation.
376 FunctionType *TransformedType;
377
378 /// Transforming a function may change the position of arguments. This
379 /// member records the mapping from each argument's old position to its new
380 /// position. Argument positions are zero-indexed. If the transformation
381 /// from F to F' made the first argument of F into the third argument of F',
382 /// then ArgumentIndexMapping[0] will equal 2.
383 std::vector<unsigned> ArgumentIndexMapping;
384};
385
386/// Given function attributes from a call site for the original function,
387/// return function attributes appropriate for a call to the transformed
388/// function.
389AttributeList
390transformFunctionAttributes(const TransformedFunction &TransformedFunction,
391 LLVMContext &Ctx, AttributeList CallSiteAttrs) {
392
393 // Construct a vector of AttributeSet for each function argument.
394 std::vector<llvm::AttributeSet> ArgumentAttributes(
395 TransformedFunction.TransformedType->getNumParams());
396
397 // Copy attributes from the parameter of the original function to the
398 // transformed version. 'ArgumentIndexMapping' holds the mapping from
399 // old argument position to new.
400 for (unsigned I = 0, IE = TransformedFunction.ArgumentIndexMapping.size();
401 I < IE; ++I) {
402 unsigned TransformedIndex = TransformedFunction.ArgumentIndexMapping[I];
403 ArgumentAttributes[TransformedIndex] = CallSiteAttrs.getParamAttrs(ArgNo: I);
404 }
405
406 // Copy annotations on varargs arguments.
407 for (unsigned I = TransformedFunction.OriginalType->getNumParams(),
408 IE = CallSiteAttrs.getNumAttrSets();
409 I < IE; ++I) {
410 ArgumentAttributes.push_back(x: CallSiteAttrs.getParamAttrs(ArgNo: I));
411 }
412
413 return AttributeList::get(C&: Ctx, FnAttrs: CallSiteAttrs.getFnAttrs(),
414 RetAttrs: CallSiteAttrs.getRetAttrs(),
415 ArgAttrs: llvm::ArrayRef(ArgumentAttributes));
416}
417
418class DataFlowSanitizer {
419 friend struct DFSanFunction;
420 friend class DFSanVisitor;
421
422 enum { ShadowWidthBits = 8, ShadowWidthBytes = ShadowWidthBits / 8 };
423
424 enum { OriginWidthBits = 32, OriginWidthBytes = OriginWidthBits / 8 };
425
426 /// How should calls to uninstrumented functions be handled?
427 enum WrapperKind {
428 /// This function is present in an uninstrumented form but we don't know
429 /// how it should be handled. Print a warning and call the function anyway.
430 /// Don't label the return value.
431 WK_Warning,
432
433 /// This function does not write to (user-accessible) memory, and its return
434 /// value is unlabelled.
435 WK_Discard,
436
437 /// This function does not write to (user-accessible) memory, and the label
438 /// of its return value is the union of the label of its arguments.
439 WK_Functional,
440
441 /// Instead of calling the function, a custom wrapper __dfsw_F is called,
442 /// where F is the name of the function. This function may wrap the
443 /// original function or provide its own implementation. WK_Custom uses an
444 /// extra pointer argument to return the shadow. This allows the wrapped
445 /// form of the function type to be expressed in C.
446 WK_Custom
447 };
448
449 Module *Mod;
450 LLVMContext *Ctx;
451 Type *Int8Ptr;
452 IntegerType *OriginTy;
453 PointerType *OriginPtrTy;
454 ConstantInt *ZeroOrigin;
455 /// The shadow type for all primitive types and vector types.
456 IntegerType *PrimitiveShadowTy;
457 PointerType *PrimitiveShadowPtrTy;
458 IntegerType *IntptrTy;
459 ConstantInt *ZeroPrimitiveShadow;
460 Constant *ArgTLS;
461 ArrayType *ArgOriginTLSTy;
462 Constant *ArgOriginTLS;
463 Constant *RetvalTLS;
464 Constant *RetvalOriginTLS;
465 FunctionType *DFSanUnionLoadFnTy;
466 FunctionType *DFSanLoadLabelAndOriginFnTy;
467 FunctionType *DFSanUnimplementedFnTy;
468 FunctionType *DFSanWrapperExternWeakNullFnTy;
469 FunctionType *DFSanSetLabelFnTy;
470 FunctionType *DFSanNonzeroLabelFnTy;
471 FunctionType *DFSanVarargWrapperFnTy;
472 FunctionType *DFSanConditionalCallbackFnTy;
473 FunctionType *DFSanConditionalCallbackOriginFnTy;
474 FunctionType *DFSanReachesFunctionCallbackFnTy;
475 FunctionType *DFSanReachesFunctionCallbackOriginFnTy;
476 FunctionType *DFSanCmpCallbackFnTy;
477 FunctionType *DFSanLoadStoreCallbackFnTy;
478 FunctionType *DFSanMemTransferCallbackFnTy;
479 FunctionType *DFSanChainOriginFnTy;
480 FunctionType *DFSanChainOriginIfTaintedFnTy;
481 FunctionType *DFSanMemOriginTransferFnTy;
482 FunctionType *DFSanMemShadowOriginTransferFnTy;
483 FunctionType *DFSanMemShadowOriginConditionalExchangeFnTy;
484 FunctionType *DFSanMaybeStoreOriginFnTy;
485 FunctionCallee DFSanUnionLoadFn;
486 FunctionCallee DFSanLoadLabelAndOriginFn;
487 FunctionCallee DFSanUnimplementedFn;
488 FunctionCallee DFSanWrapperExternWeakNullFn;
489 FunctionCallee DFSanSetLabelFn;
490 FunctionCallee DFSanNonzeroLabelFn;
491 FunctionCallee DFSanVarargWrapperFn;
492 FunctionCallee DFSanLoadCallbackFn;
493 FunctionCallee DFSanStoreCallbackFn;
494 FunctionCallee DFSanMemTransferCallbackFn;
495 FunctionCallee DFSanConditionalCallbackFn;
496 FunctionCallee DFSanConditionalCallbackOriginFn;
497 FunctionCallee DFSanReachesFunctionCallbackFn;
498 FunctionCallee DFSanReachesFunctionCallbackOriginFn;
499 FunctionCallee DFSanCmpCallbackFn;
500 FunctionCallee DFSanChainOriginFn;
501 FunctionCallee DFSanChainOriginIfTaintedFn;
502 FunctionCallee DFSanMemOriginTransferFn;
503 FunctionCallee DFSanMemShadowOriginTransferFn;
504 FunctionCallee DFSanMemShadowOriginConditionalExchangeFn;
505 FunctionCallee DFSanMaybeStoreOriginFn;
506 SmallPtrSet<Value *, 16> DFSanRuntimeFunctions;
507 MDNode *ColdCallWeights;
508 MDNode *OriginStoreWeights;
509 DFSanABIList ABIList;
510 DenseMap<Value *, Function *> UnwrappedFnMap;
511 AttributeMask ReadOnlyNoneAttrs;
512 StringSet<> CombineTaintLookupTableNames;
513
514 /// Memory map parameters used in calculation mapping application addresses
515 /// to shadow addresses and origin addresses.
516 const MemoryMapParams *MapParams;
517
518 Value *getShadowOffset(Value *Addr, IRBuilder<> &IRB);
519 Value *getShadowAddress(Value *Addr, BasicBlock::iterator Pos);
520 Value *getShadowAddress(Value *Addr, BasicBlock::iterator Pos,
521 Value *ShadowOffset);
522 std::pair<Value *, Value *> getShadowOriginAddress(Value *Addr,
523 Align InstAlignment,
524 BasicBlock::iterator Pos);
525 bool isInstrumented(const Function *F);
526 bool isInstrumented(const GlobalAlias *GA);
527 bool isForceZeroLabels(const Function *F);
528 TransformedFunction getCustomFunctionType(FunctionType *T);
529 WrapperKind getWrapperKind(Function *F);
530 void addGlobalNameSuffix(GlobalValue *GV);
531 void buildExternWeakCheckIfNeeded(IRBuilder<> &IRB, Function *F);
532 Function *buildWrapperFunction(Function *F, StringRef NewFName,
533 GlobalValue::LinkageTypes NewFLink,
534 FunctionType *NewFT);
535 void initializeCallbackFunctions(Module &M);
536 void initializeRuntimeFunctions(Module &M);
537 bool initializeModule(Module &M);
538
539 /// Advances \p OriginAddr to point to the next 32-bit origin and then loads
540 /// from it. Returns the origin's loaded value.
541 Value *loadNextOrigin(BasicBlock::iterator Pos, Align OriginAlign,
542 Value **OriginAddr);
543
544 /// Returns whether the given load byte size is amenable to inlined
545 /// optimization patterns.
546 bool hasLoadSizeForFastPath(uint64_t Size);
547
548 /// Returns whether the pass tracks origins. Supports only TLS ABI mode.
549 bool shouldTrackOrigins();
550
551 /// Returns a zero constant with the shadow type of OrigTy.
552 ///
553 /// getZeroShadow({T1,T2,...}) = {getZeroShadow(T1),getZeroShadow(T2,...}
554 /// getZeroShadow([n x T]) = [n x getZeroShadow(T)]
555 /// getZeroShadow(other type) = i16(0)
556 Constant *getZeroShadow(Type *OrigTy);
557 /// Returns a zero constant with the shadow type of V's type.
558 Constant *getZeroShadow(Value *V);
559
560 /// Checks if V is a zero shadow.
561 bool isZeroShadow(Value *V);
562
563 /// Returns the shadow type of OrigTy.
564 ///
565 /// getShadowTy({T1,T2,...}) = {getShadowTy(T1),getShadowTy(T2),...}
566 /// getShadowTy([n x T]) = [n x getShadowTy(T)]
567 /// getShadowTy(other type) = i16
568 Type *getShadowTy(Type *OrigTy);
569 /// Returns the shadow type of V's type.
570 Type *getShadowTy(Value *V);
571
572 const uint64_t NumOfElementsInArgOrgTLS = ArgTLSSize / OriginWidthBytes;
573
574public:
575 DataFlowSanitizer(const std::vector<std::string> &ABIListFiles);
576
577 bool runImpl(Module &M,
578 llvm::function_ref<TargetLibraryInfo &(Function &)> GetTLI);
579};
580
581struct DFSanFunction {
582 DataFlowSanitizer &DFS;
583 Function *F;
584 DominatorTree DT;
585 bool IsNativeABI;
586 bool IsForceZeroLabels;
587 TargetLibraryInfo &TLI;
588 AllocaInst *LabelReturnAlloca = nullptr;
589 AllocaInst *OriginReturnAlloca = nullptr;
590 DenseMap<Value *, Value *> ValShadowMap;
591 DenseMap<Value *, Value *> ValOriginMap;
592 DenseMap<AllocaInst *, AllocaInst *> AllocaShadowMap;
593 DenseMap<AllocaInst *, AllocaInst *> AllocaOriginMap;
594
595 struct PHIFixupElement {
596 PHINode *Phi;
597 PHINode *ShadowPhi;
598 PHINode *OriginPhi;
599 };
600 std::vector<PHIFixupElement> PHIFixups;
601
602 DenseSet<Instruction *> SkipInsts;
603 std::vector<Value *> NonZeroChecks;
604
605 struct CachedShadow {
606 BasicBlock *Block; // The block where Shadow is defined.
607 Value *Shadow;
608 };
609 /// Maps a value to its latest shadow value in terms of domination tree.
610 DenseMap<std::pair<Value *, Value *>, CachedShadow> CachedShadows;
611 /// Maps a value to its latest collapsed shadow value it was converted to in
612 /// terms of domination tree. When ClDebugNonzeroLabels is on, this cache is
613 /// used at a post process where CFG blocks are split. So it does not cache
614 /// BasicBlock like CachedShadows, but uses domination between values.
615 DenseMap<Value *, Value *> CachedCollapsedShadows;
616 DenseMap<Value *, std::set<Value *>> ShadowElements;
617
618 DFSanFunction(DataFlowSanitizer &DFS, Function *F, bool IsNativeABI,
619 bool IsForceZeroLabels, TargetLibraryInfo &TLI)
620 : DFS(DFS), F(F), IsNativeABI(IsNativeABI),
621 IsForceZeroLabels(IsForceZeroLabels), TLI(TLI) {
622 DT.recalculate(Func&: *F);
623 }
624
625 /// Computes the shadow address for a given function argument.
626 ///
627 /// Shadow = ArgTLS+ArgOffset.
628 Value *getArgTLS(Type *T, unsigned ArgOffset, IRBuilder<> &IRB);
629
630 /// Computes the shadow address for a return value.
631 Value *getRetvalTLS(Type *T, IRBuilder<> &IRB);
632
633 /// Computes the origin address for a given function argument.
634 ///
635 /// Origin = ArgOriginTLS[ArgNo].
636 Value *getArgOriginTLS(unsigned ArgNo, IRBuilder<> &IRB);
637
638 /// Computes the origin address for a return value.
639 Value *getRetvalOriginTLS();
640
641 Value *getOrigin(Value *V);
642 void setOrigin(Instruction *I, Value *Origin);
643 /// Generates IR to compute the origin of the last operand with a taint label.
644 Value *combineOperandOrigins(Instruction *Inst);
645 /// Before the instruction Pos, generates IR to compute the last origin with a
646 /// taint label. Labels and origins are from vectors Shadows and Origins
647 /// correspondingly. The generated IR is like
648 /// Sn-1 != Zero ? On-1: ... S2 != Zero ? O2: S1 != Zero ? O1: O0
649 /// When Zero is nullptr, it uses ZeroPrimitiveShadow. Otherwise it can be
650 /// zeros with other bitwidths.
651 Value *combineOrigins(const std::vector<Value *> &Shadows,
652 const std::vector<Value *> &Origins,
653 BasicBlock::iterator Pos, ConstantInt *Zero = nullptr);
654
655 Value *getShadow(Value *V);
656 void setShadow(Instruction *I, Value *Shadow);
657 /// Generates IR to compute the union of the two given shadows, inserting it
658 /// before Pos. The combined value is with primitive type.
659 Value *combineShadows(Value *V1, Value *V2, BasicBlock::iterator Pos);
660 /// Combines the shadow values of V1 and V2, then converts the combined value
661 /// with primitive type into a shadow value with the original type T.
662 Value *combineShadowsThenConvert(Type *T, Value *V1, Value *V2,
663 BasicBlock::iterator Pos);
664 Value *combineOperandShadows(Instruction *Inst);
665
666 /// Generates IR to load shadow and origin corresponding to bytes [\p
667 /// Addr, \p Addr + \p Size), where addr has alignment \p
668 /// InstAlignment, and take the union of each of those shadows. The returned
669 /// shadow always has primitive type.
670 ///
671 /// When tracking loads is enabled, the returned origin is a chain at the
672 /// current stack if the returned shadow is tainted.
673 std::pair<Value *, Value *> loadShadowOrigin(Value *Addr, uint64_t Size,
674 Align InstAlignment,
675 BasicBlock::iterator Pos);
676
677 void storePrimitiveShadowOrigin(Value *Addr, uint64_t Size,
678 Align InstAlignment, Value *PrimitiveShadow,
679 Value *Origin, BasicBlock::iterator Pos);
680 /// Applies PrimitiveShadow to all primitive subtypes of T, returning
681 /// the expanded shadow value.
682 ///
683 /// EFP({T1,T2, ...}, PS) = {EFP(T1,PS),EFP(T2,PS),...}
684 /// EFP([n x T], PS) = [n x EFP(T,PS)]
685 /// EFP(other types, PS) = PS
686 Value *expandFromPrimitiveShadow(Type *T, Value *PrimitiveShadow,
687 BasicBlock::iterator Pos);
688 /// Collapses Shadow into a single primitive shadow value, unioning all
689 /// primitive shadow values in the process. Returns the final primitive
690 /// shadow value.
691 ///
692 /// CTP({V1,V2, ...}) = UNION(CFP(V1,PS),CFP(V2,PS),...)
693 /// CTP([V1,V2,...]) = UNION(CFP(V1,PS),CFP(V2,PS),...)
694 /// CTP(other types, PS) = PS
695 Value *collapseToPrimitiveShadow(Value *Shadow, BasicBlock::iterator Pos);
696
697 void storeZeroPrimitiveShadow(Value *Addr, uint64_t Size, Align ShadowAlign,
698 BasicBlock::iterator Pos);
699
700 Align getShadowAlign(Align InstAlignment);
701
702 // If ClConditionalCallbacks is enabled, insert a callback after a given
703 // branch instruction using the given conditional expression.
704 void addConditionalCallbacksIfEnabled(Instruction &I, Value *Condition);
705
706 // If ClReachesFunctionCallbacks is enabled, insert a callback for each
707 // argument and load instruction.
708 void addReachesFunctionCallbacksIfEnabled(IRBuilder<> &IRB, Instruction &I,
709 Value *Data);
710
711 bool isLookupTableConstant(Value *P);
712
713private:
714 /// Collapses the shadow with aggregate type into a single primitive shadow
715 /// value.
716 template <class AggregateType>
717 Value *collapseAggregateShadow(AggregateType *AT, Value *Shadow,
718 IRBuilder<> &IRB);
719
720 Value *collapseToPrimitiveShadow(Value *Shadow, IRBuilder<> &IRB);
721
722 /// Returns the shadow value of an argument A.
723 Value *getShadowForTLSArgument(Argument *A);
724
725 /// The fast path of loading shadows.
726 std::pair<Value *, Value *>
727 loadShadowFast(Value *ShadowAddr, Value *OriginAddr, uint64_t Size,
728 Align ShadowAlign, Align OriginAlign, Value *FirstOrigin,
729 BasicBlock::iterator Pos);
730
731 Align getOriginAlign(Align InstAlignment);
732
733 /// Because 4 contiguous bytes share one 4-byte origin, the most accurate load
734 /// is __dfsan_load_label_and_origin. This function returns the union of all
735 /// labels and the origin of the first taint label. However this is an
736 /// additional call with many instructions. To ensure common cases are fast,
737 /// checks if it is possible to load labels and origins without using the
738 /// callback function.
739 ///
740 /// When enabling tracking load instructions, we always use
741 /// __dfsan_load_label_and_origin to reduce code size.
742 bool useCallbackLoadLabelAndOrigin(uint64_t Size, Align InstAlignment);
743
744 /// Returns a chain at the current stack with previous origin V.
745 Value *updateOrigin(Value *V, IRBuilder<> &IRB);
746
747 /// Returns a chain at the current stack with previous origin V if Shadow is
748 /// tainted.
749 Value *updateOriginIfTainted(Value *Shadow, Value *Origin, IRBuilder<> &IRB);
750
751 /// Creates an Intptr = Origin | Origin << 32 if Intptr's size is 64. Returns
752 /// Origin otherwise.
753 Value *originToIntptr(IRBuilder<> &IRB, Value *Origin);
754
755 /// Stores Origin into the address range [StoreOriginAddr, StoreOriginAddr +
756 /// Size).
757 void paintOrigin(IRBuilder<> &IRB, Value *Origin, Value *StoreOriginAddr,
758 uint64_t StoreOriginSize, Align Alignment);
759
760 /// Stores Origin in terms of its Shadow value.
761 /// * Do not write origins for zero shadows because we do not trace origins
762 /// for untainted sinks.
763 /// * Use __dfsan_maybe_store_origin if there are too many origin store
764 /// instrumentations.
765 void storeOrigin(BasicBlock::iterator Pos, Value *Addr, uint64_t Size,
766 Value *Shadow, Value *Origin, Value *StoreOriginAddr,
767 Align InstAlignment);
768
769 /// Convert a scalar value to an i1 by comparing with 0.
770 Value *convertToBool(Value *V, IRBuilder<> &IRB, const Twine &Name = "");
771
772 bool shouldInstrumentWithCall();
773
774 /// Generates IR to load shadow and origin corresponding to bytes [\p
775 /// Addr, \p Addr + \p Size), where addr has alignment \p
776 /// InstAlignment, and take the union of each of those shadows. The returned
777 /// shadow always has primitive type.
778 std::pair<Value *, Value *>
779 loadShadowOriginSansLoadTracking(Value *Addr, uint64_t Size,
780 Align InstAlignment,
781 BasicBlock::iterator Pos);
782 int NumOriginStores = 0;
783};
784
785class DFSanVisitor : public InstVisitor<DFSanVisitor> {
786public:
787 DFSanFunction &DFSF;
788
789 DFSanVisitor(DFSanFunction &DFSF) : DFSF(DFSF) {}
790
791 const DataLayout &getDataLayout() const {
792 return DFSF.F->getDataLayout();
793 }
794
795 // Combines shadow values and origins for all of I's operands.
796 void visitInstOperands(Instruction &I);
797
798 void visitUnaryOperator(UnaryOperator &UO);
799 void visitBinaryOperator(BinaryOperator &BO);
800 void visitBitCastInst(BitCastInst &BCI);
801 void visitCastInst(CastInst &CI);
802 void visitCmpInst(CmpInst &CI);
803 void visitLandingPadInst(LandingPadInst &LPI);
804 void visitGetElementPtrInst(GetElementPtrInst &GEPI);
805 void visitLoadInst(LoadInst &LI);
806 void visitStoreInst(StoreInst &SI);
807 void visitAtomicRMWInst(AtomicRMWInst &I);
808 void visitAtomicCmpXchgInst(AtomicCmpXchgInst &I);
809 void visitReturnInst(ReturnInst &RI);
810 void visitLibAtomicLoad(CallBase &CB);
811 void visitLibAtomicStore(CallBase &CB);
812 void visitLibAtomicExchange(CallBase &CB);
813 void visitLibAtomicCompareExchange(CallBase &CB);
814 void visitCallBase(CallBase &CB);
815 void visitPHINode(PHINode &PN);
816 void visitExtractElementInst(ExtractElementInst &I);
817 void visitInsertElementInst(InsertElementInst &I);
818 void visitShuffleVectorInst(ShuffleVectorInst &I);
819 void visitExtractValueInst(ExtractValueInst &I);
820 void visitInsertValueInst(InsertValueInst &I);
821 void visitAllocaInst(AllocaInst &I);
822 void visitSelectInst(SelectInst &I);
823 void visitMemSetInst(MemSetInst &I);
824 void visitMemTransferInst(MemTransferInst &I);
825 void visitBranchInst(BranchInst &BR);
826 void visitSwitchInst(SwitchInst &SW);
827
828private:
829 void visitCASOrRMW(Align InstAlignment, Instruction &I);
830
831 // Returns false when this is an invoke of a custom function.
832 bool visitWrappedCallBase(Function &F, CallBase &CB);
833
834 // Combines origins for all of I's operands.
835 void visitInstOperandOrigins(Instruction &I);
836
837 void addShadowArguments(Function &F, CallBase &CB, std::vector<Value *> &Args,
838 IRBuilder<> &IRB);
839
840 void addOriginArguments(Function &F, CallBase &CB, std::vector<Value *> &Args,
841 IRBuilder<> &IRB);
842
843 Value *makeAddAcquireOrderingTable(IRBuilder<> &IRB);
844 Value *makeAddReleaseOrderingTable(IRBuilder<> &IRB);
845};
846
847bool LibAtomicFunction(const Function &F) {
848 // This is a bit of a hack because TargetLibraryInfo is a function pass.
849 // The DFSan pass would need to be refactored to be function pass oriented
850 // (like MSan is) in order to fit together nicely with TargetLibraryInfo.
851 // We need this check to prevent them from being instrumented, or wrapped.
852 // Match on name and number of arguments.
853 if (!F.hasName() || F.isVarArg())
854 return false;
855 switch (F.arg_size()) {
856 case 4:
857 return F.getName() == "__atomic_load" || F.getName() == "__atomic_store";
858 case 5:
859 return F.getName() == "__atomic_exchange";
860 case 6:
861 return F.getName() == "__atomic_compare_exchange";
862 default:
863 return false;
864 }
865}
866
867} // end anonymous namespace
868
869DataFlowSanitizer::DataFlowSanitizer(
870 const std::vector<std::string> &ABIListFiles) {
871 std::vector<std::string> AllABIListFiles(std::move(ABIListFiles));
872 llvm::append_range(C&: AllABIListFiles, R&: ClABIListFiles);
873 // FIXME: should we propagate vfs::FileSystem to this constructor?
874 ABIList.set(
875 SpecialCaseList::createOrDie(Paths: AllABIListFiles, FS&: *vfs::getRealFileSystem()));
876
877 CombineTaintLookupTableNames.insert_range(R&: ClCombineTaintLookupTables);
878}
879
880TransformedFunction DataFlowSanitizer::getCustomFunctionType(FunctionType *T) {
881 SmallVector<Type *, 4> ArgTypes;
882
883 // Some parameters of the custom function being constructed are
884 // parameters of T. Record the mapping from parameters of T to
885 // parameters of the custom function, so that parameter attributes
886 // at call sites can be updated.
887 std::vector<unsigned> ArgumentIndexMapping;
888 for (unsigned I = 0, E = T->getNumParams(); I != E; ++I) {
889 Type *ParamType = T->getParamType(i: I);
890 ArgumentIndexMapping.push_back(x: ArgTypes.size());
891 ArgTypes.push_back(Elt: ParamType);
892 }
893 for (unsigned I = 0, E = T->getNumParams(); I != E; ++I)
894 ArgTypes.push_back(Elt: PrimitiveShadowTy);
895 if (T->isVarArg())
896 ArgTypes.push_back(Elt: PrimitiveShadowPtrTy);
897 Type *RetType = T->getReturnType();
898 if (!RetType->isVoidTy())
899 ArgTypes.push_back(Elt: PrimitiveShadowPtrTy);
900
901 if (shouldTrackOrigins()) {
902 for (unsigned I = 0, E = T->getNumParams(); I != E; ++I)
903 ArgTypes.push_back(Elt: OriginTy);
904 if (T->isVarArg())
905 ArgTypes.push_back(Elt: OriginPtrTy);
906 if (!RetType->isVoidTy())
907 ArgTypes.push_back(Elt: OriginPtrTy);
908 }
909
910 return TransformedFunction(
911 T, FunctionType::get(Result: T->getReturnType(), Params: ArgTypes, isVarArg: T->isVarArg()),
912 ArgumentIndexMapping);
913}
914
915bool DataFlowSanitizer::isZeroShadow(Value *V) {
916 Type *T = V->getType();
917 if (!isa<ArrayType>(Val: T) && !isa<StructType>(Val: T)) {
918 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Val: V))
919 return CI->isZero();
920 return false;
921 }
922
923 return isa<ConstantAggregateZero>(Val: V);
924}
925
926bool DataFlowSanitizer::hasLoadSizeForFastPath(uint64_t Size) {
927 uint64_t ShadowSize = Size * ShadowWidthBytes;
928 return ShadowSize % 8 == 0 || ShadowSize == 4;
929}
930
931bool DataFlowSanitizer::shouldTrackOrigins() {
932 static const bool ShouldTrackOrigins = ClTrackOrigins;
933 return ShouldTrackOrigins;
934}
935
936Constant *DataFlowSanitizer::getZeroShadow(Type *OrigTy) {
937 if (!isa<ArrayType>(Val: OrigTy) && !isa<StructType>(Val: OrigTy))
938 return ZeroPrimitiveShadow;
939 Type *ShadowTy = getShadowTy(OrigTy);
940 return ConstantAggregateZero::get(Ty: ShadowTy);
941}
942
943Constant *DataFlowSanitizer::getZeroShadow(Value *V) {
944 return getZeroShadow(OrigTy: V->getType());
945}
946
947static Value *expandFromPrimitiveShadowRecursive(
948 Value *Shadow, SmallVector<unsigned, 4> &Indices, Type *SubShadowTy,
949 Value *PrimitiveShadow, IRBuilder<> &IRB) {
950 if (!isa<ArrayType>(Val: SubShadowTy) && !isa<StructType>(Val: SubShadowTy))
951 return IRB.CreateInsertValue(Agg: Shadow, Val: PrimitiveShadow, Idxs: Indices);
952
953 if (ArrayType *AT = dyn_cast<ArrayType>(Val: SubShadowTy)) {
954 for (unsigned Idx = 0; Idx < AT->getNumElements(); Idx++) {
955 Indices.push_back(Elt: Idx);
956 Shadow = expandFromPrimitiveShadowRecursive(
957 Shadow, Indices, SubShadowTy: AT->getElementType(), PrimitiveShadow, IRB);
958 Indices.pop_back();
959 }
960 return Shadow;
961 }
962
963 if (StructType *ST = dyn_cast<StructType>(Val: SubShadowTy)) {
964 for (unsigned Idx = 0; Idx < ST->getNumElements(); Idx++) {
965 Indices.push_back(Elt: Idx);
966 Shadow = expandFromPrimitiveShadowRecursive(
967 Shadow, Indices, SubShadowTy: ST->getElementType(N: Idx), PrimitiveShadow, IRB);
968 Indices.pop_back();
969 }
970 return Shadow;
971 }
972 llvm_unreachable("Unexpected shadow type");
973}
974
975bool DFSanFunction::shouldInstrumentWithCall() {
976 return ClInstrumentWithCallThreshold >= 0 &&
977 NumOriginStores >= ClInstrumentWithCallThreshold;
978}
979
980Value *DFSanFunction::expandFromPrimitiveShadow(Type *T, Value *PrimitiveShadow,
981 BasicBlock::iterator Pos) {
982 Type *ShadowTy = DFS.getShadowTy(OrigTy: T);
983
984 if (!isa<ArrayType>(Val: ShadowTy) && !isa<StructType>(Val: ShadowTy))
985 return PrimitiveShadow;
986
987 if (DFS.isZeroShadow(V: PrimitiveShadow))
988 return DFS.getZeroShadow(OrigTy: ShadowTy);
989
990 IRBuilder<> IRB(Pos->getParent(), Pos);
991 SmallVector<unsigned, 4> Indices;
992 Value *Shadow = UndefValue::get(T: ShadowTy);
993 Shadow = expandFromPrimitiveShadowRecursive(Shadow, Indices, SubShadowTy: ShadowTy,
994 PrimitiveShadow, IRB);
995
996 // Caches the primitive shadow value that built the shadow value.
997 CachedCollapsedShadows[Shadow] = PrimitiveShadow;
998 return Shadow;
999}
1000
1001template <class AggregateType>
1002Value *DFSanFunction::collapseAggregateShadow(AggregateType *AT, Value *Shadow,
1003 IRBuilder<> &IRB) {
1004 if (!AT->getNumElements())
1005 return DFS.ZeroPrimitiveShadow;
1006
1007 Value *FirstItem = IRB.CreateExtractValue(Agg: Shadow, Idxs: 0);
1008 Value *Aggregator = collapseToPrimitiveShadow(Shadow: FirstItem, IRB);
1009
1010 for (unsigned Idx = 1; Idx < AT->getNumElements(); Idx++) {
1011 Value *ShadowItem = IRB.CreateExtractValue(Agg: Shadow, Idxs: Idx);
1012 Value *ShadowInner = collapseToPrimitiveShadow(Shadow: ShadowItem, IRB);
1013 Aggregator = IRB.CreateOr(LHS: Aggregator, RHS: ShadowInner);
1014 }
1015 return Aggregator;
1016}
1017
1018Value *DFSanFunction::collapseToPrimitiveShadow(Value *Shadow,
1019 IRBuilder<> &IRB) {
1020 Type *ShadowTy = Shadow->getType();
1021 if (!isa<ArrayType>(Val: ShadowTy) && !isa<StructType>(Val: ShadowTy))
1022 return Shadow;
1023 if (ArrayType *AT = dyn_cast<ArrayType>(Val: ShadowTy))
1024 return collapseAggregateShadow<>(AT, Shadow, IRB);
1025 if (StructType *ST = dyn_cast<StructType>(Val: ShadowTy))
1026 return collapseAggregateShadow<>(AT: ST, Shadow, IRB);
1027 llvm_unreachable("Unexpected shadow type");
1028}
1029
1030Value *DFSanFunction::collapseToPrimitiveShadow(Value *Shadow,
1031 BasicBlock::iterator Pos) {
1032 Type *ShadowTy = Shadow->getType();
1033 if (!isa<ArrayType>(Val: ShadowTy) && !isa<StructType>(Val: ShadowTy))
1034 return Shadow;
1035
1036 // Checks if the cached collapsed shadow value dominates Pos.
1037 Value *&CS = CachedCollapsedShadows[Shadow];
1038 if (CS && DT.dominates(Def: CS, User: Pos))
1039 return CS;
1040
1041 IRBuilder<> IRB(Pos->getParent(), Pos);
1042 Value *PrimitiveShadow = collapseToPrimitiveShadow(Shadow, IRB);
1043 // Caches the converted primitive shadow value.
1044 CS = PrimitiveShadow;
1045 return PrimitiveShadow;
1046}
1047
1048void DFSanFunction::addConditionalCallbacksIfEnabled(Instruction &I,
1049 Value *Condition) {
1050 if (!ClConditionalCallbacks) {
1051 return;
1052 }
1053 IRBuilder<> IRB(&I);
1054 Value *CondShadow = getShadow(V: Condition);
1055 CallInst *CI;
1056 if (DFS.shouldTrackOrigins()) {
1057 Value *CondOrigin = getOrigin(V: Condition);
1058 CI = IRB.CreateCall(Callee: DFS.DFSanConditionalCallbackOriginFn,
1059 Args: {CondShadow, CondOrigin});
1060 } else {
1061 CI = IRB.CreateCall(Callee: DFS.DFSanConditionalCallbackFn, Args: {CondShadow});
1062 }
1063 CI->addParamAttr(ArgNo: 0, Kind: Attribute::ZExt);
1064}
1065
1066void DFSanFunction::addReachesFunctionCallbacksIfEnabled(IRBuilder<> &IRB,
1067 Instruction &I,
1068 Value *Data) {
1069 if (!ClReachesFunctionCallbacks) {
1070 return;
1071 }
1072 const DebugLoc &dbgloc = I.getDebugLoc();
1073 Value *DataShadow = collapseToPrimitiveShadow(Shadow: getShadow(V: Data), IRB);
1074 ConstantInt *CILine;
1075 llvm::Value *FilePathPtr;
1076
1077 if (dbgloc.get() == nullptr) {
1078 CILine = llvm::ConstantInt::get(Context&: I.getContext(), V: llvm::APInt(32, 0));
1079 FilePathPtr = IRB.CreateGlobalString(
1080 Str: I.getFunction()->getParent()->getSourceFileName());
1081 } else {
1082 CILine = llvm::ConstantInt::get(Context&: I.getContext(),
1083 V: llvm::APInt(32, dbgloc.getLine()));
1084 FilePathPtr = IRB.CreateGlobalString(Str: dbgloc->getFilename());
1085 }
1086
1087 llvm::Value *FunctionNamePtr =
1088 IRB.CreateGlobalString(Str: I.getFunction()->getName());
1089
1090 CallInst *CB;
1091 std::vector<Value *> args;
1092
1093 if (DFS.shouldTrackOrigins()) {
1094 Value *DataOrigin = getOrigin(V: Data);
1095 args = { DataShadow, DataOrigin, FilePathPtr, CILine, FunctionNamePtr };
1096 CB = IRB.CreateCall(Callee: DFS.DFSanReachesFunctionCallbackOriginFn, Args: args);
1097 } else {
1098 args = { DataShadow, FilePathPtr, CILine, FunctionNamePtr };
1099 CB = IRB.CreateCall(Callee: DFS.DFSanReachesFunctionCallbackFn, Args: args);
1100 }
1101 CB->addParamAttr(ArgNo: 0, Kind: Attribute::ZExt);
1102 CB->setDebugLoc(dbgloc);
1103}
1104
1105Type *DataFlowSanitizer::getShadowTy(Type *OrigTy) {
1106 if (!OrigTy->isSized())
1107 return PrimitiveShadowTy;
1108 if (isa<IntegerType>(Val: OrigTy))
1109 return PrimitiveShadowTy;
1110 if (isa<VectorType>(Val: OrigTy))
1111 return PrimitiveShadowTy;
1112 if (ArrayType *AT = dyn_cast<ArrayType>(Val: OrigTy))
1113 return ArrayType::get(ElementType: getShadowTy(OrigTy: AT->getElementType()),
1114 NumElements: AT->getNumElements());
1115 if (StructType *ST = dyn_cast<StructType>(Val: OrigTy)) {
1116 SmallVector<Type *, 4> Elements;
1117 for (unsigned I = 0, N = ST->getNumElements(); I < N; ++I)
1118 Elements.push_back(Elt: getShadowTy(OrigTy: ST->getElementType(N: I)));
1119 return StructType::get(Context&: *Ctx, Elements);
1120 }
1121 return PrimitiveShadowTy;
1122}
1123
1124Type *DataFlowSanitizer::getShadowTy(Value *V) {
1125 return getShadowTy(OrigTy: V->getType());
1126}
1127
1128bool DataFlowSanitizer::initializeModule(Module &M) {
1129 Triple TargetTriple(M.getTargetTriple());
1130 const DataLayout &DL = M.getDataLayout();
1131
1132 if (TargetTriple.getOS() != Triple::Linux)
1133 report_fatal_error(reason: "unsupported operating system");
1134 switch (TargetTriple.getArch()) {
1135 case Triple::aarch64:
1136 MapParams = &Linux_AArch64_MemoryMapParams;
1137 break;
1138 case Triple::x86_64:
1139 MapParams = &Linux_X86_64_MemoryMapParams;
1140 break;
1141 case Triple::loongarch64:
1142 MapParams = &Linux_LoongArch64_MemoryMapParams;
1143 break;
1144 default:
1145 report_fatal_error(reason: "unsupported architecture");
1146 }
1147
1148 Mod = &M;
1149 Ctx = &M.getContext();
1150 Int8Ptr = PointerType::getUnqual(C&: *Ctx);
1151 OriginTy = IntegerType::get(C&: *Ctx, NumBits: OriginWidthBits);
1152 OriginPtrTy = PointerType::getUnqual(C&: *Ctx);
1153 PrimitiveShadowTy = IntegerType::get(C&: *Ctx, NumBits: ShadowWidthBits);
1154 PrimitiveShadowPtrTy = PointerType::getUnqual(C&: *Ctx);
1155 IntptrTy = DL.getIntPtrType(C&: *Ctx);
1156 ZeroPrimitiveShadow = ConstantInt::getSigned(Ty: PrimitiveShadowTy, V: 0);
1157 ZeroOrigin = ConstantInt::getSigned(Ty: OriginTy, V: 0);
1158
1159 Type *DFSanUnionLoadArgs[2] = {PrimitiveShadowPtrTy, IntptrTy};
1160 DFSanUnionLoadFnTy = FunctionType::get(Result: PrimitiveShadowTy, Params: DFSanUnionLoadArgs,
1161 /*isVarArg=*/false);
1162 Type *DFSanLoadLabelAndOriginArgs[2] = {Int8Ptr, IntptrTy};
1163 DFSanLoadLabelAndOriginFnTy =
1164 FunctionType::get(Result: IntegerType::get(C&: *Ctx, NumBits: 64), Params: DFSanLoadLabelAndOriginArgs,
1165 /*isVarArg=*/false);
1166 DFSanUnimplementedFnTy = FunctionType::get(
1167 Result: Type::getVoidTy(C&: *Ctx), Params: PointerType::getUnqual(C&: *Ctx), /*isVarArg=*/false);
1168 Type *DFSanWrapperExternWeakNullArgs[2] = {Int8Ptr, Int8Ptr};
1169 DFSanWrapperExternWeakNullFnTy =
1170 FunctionType::get(Result: Type::getVoidTy(C&: *Ctx), Params: DFSanWrapperExternWeakNullArgs,
1171 /*isVarArg=*/false);
1172 Type *DFSanSetLabelArgs[4] = {PrimitiveShadowTy, OriginTy,
1173 PointerType::getUnqual(C&: *Ctx), IntptrTy};
1174 DFSanSetLabelFnTy = FunctionType::get(Result: Type::getVoidTy(C&: *Ctx),
1175 Params: DFSanSetLabelArgs, /*isVarArg=*/false);
1176 DFSanNonzeroLabelFnTy = FunctionType::get(Result: Type::getVoidTy(C&: *Ctx), Params: {},
1177 /*isVarArg=*/false);
1178 DFSanVarargWrapperFnTy = FunctionType::get(
1179 Result: Type::getVoidTy(C&: *Ctx), Params: PointerType::getUnqual(C&: *Ctx), /*isVarArg=*/false);
1180 DFSanConditionalCallbackFnTy =
1181 FunctionType::get(Result: Type::getVoidTy(C&: *Ctx), Params: PrimitiveShadowTy,
1182 /*isVarArg=*/false);
1183 Type *DFSanConditionalCallbackOriginArgs[2] = {PrimitiveShadowTy, OriginTy};
1184 DFSanConditionalCallbackOriginFnTy = FunctionType::get(
1185 Result: Type::getVoidTy(C&: *Ctx), Params: DFSanConditionalCallbackOriginArgs,
1186 /*isVarArg=*/false);
1187 Type *DFSanReachesFunctionCallbackArgs[4] = {PrimitiveShadowTy, Int8Ptr,
1188 OriginTy, Int8Ptr};
1189 DFSanReachesFunctionCallbackFnTy =
1190 FunctionType::get(Result: Type::getVoidTy(C&: *Ctx), Params: DFSanReachesFunctionCallbackArgs,
1191 /*isVarArg=*/false);
1192 Type *DFSanReachesFunctionCallbackOriginArgs[5] = {
1193 PrimitiveShadowTy, OriginTy, Int8Ptr, OriginTy, Int8Ptr};
1194 DFSanReachesFunctionCallbackOriginFnTy = FunctionType::get(
1195 Result: Type::getVoidTy(C&: *Ctx), Params: DFSanReachesFunctionCallbackOriginArgs,
1196 /*isVarArg=*/false);
1197 DFSanCmpCallbackFnTy =
1198 FunctionType::get(Result: Type::getVoidTy(C&: *Ctx), Params: PrimitiveShadowTy,
1199 /*isVarArg=*/false);
1200 DFSanChainOriginFnTy =
1201 FunctionType::get(Result: OriginTy, Params: OriginTy, /*isVarArg=*/false);
1202 Type *DFSanChainOriginIfTaintedArgs[2] = {PrimitiveShadowTy, OriginTy};
1203 DFSanChainOriginIfTaintedFnTy = FunctionType::get(
1204 Result: OriginTy, Params: DFSanChainOriginIfTaintedArgs, /*isVarArg=*/false);
1205 Type *DFSanMaybeStoreOriginArgs[4] = {IntegerType::get(C&: *Ctx, NumBits: ShadowWidthBits),
1206 Int8Ptr, IntptrTy, OriginTy};
1207 DFSanMaybeStoreOriginFnTy = FunctionType::get(
1208 Result: Type::getVoidTy(C&: *Ctx), Params: DFSanMaybeStoreOriginArgs, /*isVarArg=*/false);
1209 Type *DFSanMemOriginTransferArgs[3] = {Int8Ptr, Int8Ptr, IntptrTy};
1210 DFSanMemOriginTransferFnTy = FunctionType::get(
1211 Result: Type::getVoidTy(C&: *Ctx), Params: DFSanMemOriginTransferArgs, /*isVarArg=*/false);
1212 Type *DFSanMemShadowOriginTransferArgs[3] = {Int8Ptr, Int8Ptr, IntptrTy};
1213 DFSanMemShadowOriginTransferFnTy =
1214 FunctionType::get(Result: Type::getVoidTy(C&: *Ctx), Params: DFSanMemShadowOriginTransferArgs,
1215 /*isVarArg=*/false);
1216 Type *DFSanMemShadowOriginConditionalExchangeArgs[5] = {
1217 IntegerType::get(C&: *Ctx, NumBits: 8), Int8Ptr, Int8Ptr, Int8Ptr, IntptrTy};
1218 DFSanMemShadowOriginConditionalExchangeFnTy = FunctionType::get(
1219 Result: Type::getVoidTy(C&: *Ctx), Params: DFSanMemShadowOriginConditionalExchangeArgs,
1220 /*isVarArg=*/false);
1221 Type *DFSanLoadStoreCallbackArgs[2] = {PrimitiveShadowTy, Int8Ptr};
1222 DFSanLoadStoreCallbackFnTy =
1223 FunctionType::get(Result: Type::getVoidTy(C&: *Ctx), Params: DFSanLoadStoreCallbackArgs,
1224 /*isVarArg=*/false);
1225 Type *DFSanMemTransferCallbackArgs[2] = {PrimitiveShadowPtrTy, IntptrTy};
1226 DFSanMemTransferCallbackFnTy =
1227 FunctionType::get(Result: Type::getVoidTy(C&: *Ctx), Params: DFSanMemTransferCallbackArgs,
1228 /*isVarArg=*/false);
1229
1230 ColdCallWeights = MDBuilder(*Ctx).createUnlikelyBranchWeights();
1231 OriginStoreWeights = MDBuilder(*Ctx).createUnlikelyBranchWeights();
1232 return true;
1233}
1234
1235bool DataFlowSanitizer::isInstrumented(const Function *F) {
1236 return !ABIList.isIn(F: *F, Category: "uninstrumented");
1237}
1238
1239bool DataFlowSanitizer::isInstrumented(const GlobalAlias *GA) {
1240 return !ABIList.isIn(GA: *GA, Category: "uninstrumented");
1241}
1242
1243bool DataFlowSanitizer::isForceZeroLabels(const Function *F) {
1244 return ABIList.isIn(F: *F, Category: "force_zero_labels");
1245}
1246
1247DataFlowSanitizer::WrapperKind DataFlowSanitizer::getWrapperKind(Function *F) {
1248 if (ABIList.isIn(F: *F, Category: "functional"))
1249 return WK_Functional;
1250 if (ABIList.isIn(F: *F, Category: "discard"))
1251 return WK_Discard;
1252 if (ABIList.isIn(F: *F, Category: "custom"))
1253 return WK_Custom;
1254
1255 return WK_Warning;
1256}
1257
1258void DataFlowSanitizer::addGlobalNameSuffix(GlobalValue *GV) {
1259 std::string GVName = std::string(GV->getName()), Suffix = ".dfsan";
1260 GV->setName(GVName + Suffix);
1261
1262 // Try to change the name of the function in module inline asm. We only do
1263 // this for specific asm directives, currently only ".symver", to try to avoid
1264 // corrupting asm which happens to contain the symbol name as a substring.
1265 // Note that the substitution for .symver assumes that the versioned symbol
1266 // also has an instrumented name.
1267 std::string Asm = GV->getParent()->getModuleInlineAsm();
1268 std::string SearchStr = ".symver " + GVName + ",";
1269 size_t Pos = Asm.find(str: SearchStr);
1270 if (Pos != std::string::npos) {
1271 Asm.replace(pos: Pos, n: SearchStr.size(), str: ".symver " + GVName + Suffix + ",");
1272 Pos = Asm.find(c: '@');
1273
1274 if (Pos == std::string::npos)
1275 report_fatal_error(reason: Twine("unsupported .symver: ", Asm));
1276
1277 Asm.replace(pos: Pos, n: 1, str: Suffix + "@");
1278 GV->getParent()->setModuleInlineAsm(Asm);
1279 }
1280}
1281
1282void DataFlowSanitizer::buildExternWeakCheckIfNeeded(IRBuilder<> &IRB,
1283 Function *F) {
1284 // If the function we are wrapping was ExternWeak, it may be null.
1285 // The original code before calling this wrapper may have checked for null,
1286 // but replacing with a known-to-not-be-null wrapper can break this check.
1287 // When replacing uses of the extern weak function with the wrapper we try
1288 // to avoid replacing uses in conditionals, but this is not perfect.
1289 // In the case where we fail, and accidentally optimize out a null check
1290 // for a extern weak function, add a check here to help identify the issue.
1291 if (GlobalValue::isExternalWeakLinkage(Linkage: F->getLinkage())) {
1292 std::vector<Value *> Args;
1293 Args.push_back(x: F);
1294 Args.push_back(x: IRB.CreateGlobalString(Str: F->getName()));
1295 IRB.CreateCall(Callee: DFSanWrapperExternWeakNullFn, Args);
1296 }
1297}
1298
1299Function *
1300DataFlowSanitizer::buildWrapperFunction(Function *F, StringRef NewFName,
1301 GlobalValue::LinkageTypes NewFLink,
1302 FunctionType *NewFT) {
1303 FunctionType *FT = F->getFunctionType();
1304 Function *NewF = Function::Create(Ty: NewFT, Linkage: NewFLink, AddrSpace: F->getAddressSpace(),
1305 N: NewFName, M: F->getParent());
1306 NewF->copyAttributesFrom(Src: F);
1307 NewF->removeRetAttrs(Attrs: AttributeFuncs::typeIncompatible(
1308 Ty: NewFT->getReturnType(), AS: NewF->getAttributes().getRetAttrs()));
1309
1310 BasicBlock *BB = BasicBlock::Create(Context&: *Ctx, Name: "entry", Parent: NewF);
1311 if (F->isVarArg()) {
1312 NewF->removeFnAttr(Kind: "split-stack");
1313 CallInst::Create(Func: DFSanVarargWrapperFn,
1314 Args: IRBuilder<>(BB).CreateGlobalString(Str: F->getName()), NameStr: "", InsertBefore: BB);
1315 new UnreachableInst(*Ctx, BB);
1316 } else {
1317 auto ArgIt = pointer_iterator<Argument *>(NewF->arg_begin());
1318 std::vector<Value *> Args(ArgIt, ArgIt + FT->getNumParams());
1319
1320 CallInst *CI = CallInst::Create(Func: F, Args, NameStr: "", InsertBefore: BB);
1321 if (FT->getReturnType()->isVoidTy())
1322 ReturnInst::Create(C&: *Ctx, InsertAtEnd: BB);
1323 else
1324 ReturnInst::Create(C&: *Ctx, retVal: CI, InsertBefore: BB);
1325 }
1326
1327 return NewF;
1328}
1329
1330// Initialize DataFlowSanitizer runtime functions and declare them in the module
1331void DataFlowSanitizer::initializeRuntimeFunctions(Module &M) {
1332 LLVMContext &C = M.getContext();
1333 {
1334 AttributeList AL;
1335 AL = AL.addFnAttribute(C, Kind: Attribute::NoUnwind);
1336 AL = AL.addFnAttribute(
1337 C, Attr: Attribute::getWithMemoryEffects(Context&: C, ME: MemoryEffects::readOnly()));
1338 AL = AL.addRetAttribute(C, Kind: Attribute::ZExt);
1339 DFSanUnionLoadFn =
1340 Mod->getOrInsertFunction(Name: "__dfsan_union_load", T: DFSanUnionLoadFnTy, AttributeList: AL);
1341 }
1342 {
1343 AttributeList AL;
1344 AL = AL.addFnAttribute(C, Kind: Attribute::NoUnwind);
1345 AL = AL.addFnAttribute(
1346 C, Attr: Attribute::getWithMemoryEffects(Context&: C, ME: MemoryEffects::readOnly()));
1347 AL = AL.addRetAttribute(C, Kind: Attribute::ZExt);
1348 DFSanLoadLabelAndOriginFn = Mod->getOrInsertFunction(
1349 Name: "__dfsan_load_label_and_origin", T: DFSanLoadLabelAndOriginFnTy, AttributeList: AL);
1350 }
1351 DFSanUnimplementedFn =
1352 Mod->getOrInsertFunction(Name: "__dfsan_unimplemented", T: DFSanUnimplementedFnTy);
1353 DFSanWrapperExternWeakNullFn = Mod->getOrInsertFunction(
1354 Name: "__dfsan_wrapper_extern_weak_null", T: DFSanWrapperExternWeakNullFnTy);
1355 {
1356 AttributeList AL;
1357 AL = AL.addParamAttribute(C&: M.getContext(), ArgNo: 0, Kind: Attribute::ZExt);
1358 AL = AL.addParamAttribute(C&: M.getContext(), ArgNo: 1, Kind: Attribute::ZExt);
1359 DFSanSetLabelFn =
1360 Mod->getOrInsertFunction(Name: "__dfsan_set_label", T: DFSanSetLabelFnTy, AttributeList: AL);
1361 }
1362 DFSanNonzeroLabelFn =
1363 Mod->getOrInsertFunction(Name: "__dfsan_nonzero_label", T: DFSanNonzeroLabelFnTy);
1364 DFSanVarargWrapperFn = Mod->getOrInsertFunction(Name: "__dfsan_vararg_wrapper",
1365 T: DFSanVarargWrapperFnTy);
1366 {
1367 AttributeList AL;
1368 AL = AL.addParamAttribute(C&: M.getContext(), ArgNo: 0, Kind: Attribute::ZExt);
1369 AL = AL.addRetAttribute(C&: M.getContext(), Kind: Attribute::ZExt);
1370 DFSanChainOriginFn = Mod->getOrInsertFunction(Name: "__dfsan_chain_origin",
1371 T: DFSanChainOriginFnTy, AttributeList: AL);
1372 }
1373 {
1374 AttributeList AL;
1375 AL = AL.addParamAttribute(C&: M.getContext(), ArgNo: 0, Kind: Attribute::ZExt);
1376 AL = AL.addParamAttribute(C&: M.getContext(), ArgNo: 1, Kind: Attribute::ZExt);
1377 AL = AL.addRetAttribute(C&: M.getContext(), Kind: Attribute::ZExt);
1378 DFSanChainOriginIfTaintedFn = Mod->getOrInsertFunction(
1379 Name: "__dfsan_chain_origin_if_tainted", T: DFSanChainOriginIfTaintedFnTy, AttributeList: AL);
1380 }
1381 DFSanMemOriginTransferFn = Mod->getOrInsertFunction(
1382 Name: "__dfsan_mem_origin_transfer", T: DFSanMemOriginTransferFnTy);
1383
1384 DFSanMemShadowOriginTransferFn = Mod->getOrInsertFunction(
1385 Name: "__dfsan_mem_shadow_origin_transfer", T: DFSanMemShadowOriginTransferFnTy);
1386
1387 DFSanMemShadowOriginConditionalExchangeFn =
1388 Mod->getOrInsertFunction(Name: "__dfsan_mem_shadow_origin_conditional_exchange",
1389 T: DFSanMemShadowOriginConditionalExchangeFnTy);
1390
1391 {
1392 AttributeList AL;
1393 AL = AL.addParamAttribute(C&: M.getContext(), ArgNo: 0, Kind: Attribute::ZExt);
1394 AL = AL.addParamAttribute(C&: M.getContext(), ArgNo: 3, Kind: Attribute::ZExt);
1395 DFSanMaybeStoreOriginFn = Mod->getOrInsertFunction(
1396 Name: "__dfsan_maybe_store_origin", T: DFSanMaybeStoreOriginFnTy, AttributeList: AL);
1397 }
1398
1399 DFSanRuntimeFunctions.insert(
1400 Ptr: DFSanUnionLoadFn.getCallee()->stripPointerCasts());
1401 DFSanRuntimeFunctions.insert(
1402 Ptr: DFSanLoadLabelAndOriginFn.getCallee()->stripPointerCasts());
1403 DFSanRuntimeFunctions.insert(
1404 Ptr: DFSanUnimplementedFn.getCallee()->stripPointerCasts());
1405 DFSanRuntimeFunctions.insert(
1406 Ptr: DFSanWrapperExternWeakNullFn.getCallee()->stripPointerCasts());
1407 DFSanRuntimeFunctions.insert(
1408 Ptr: DFSanSetLabelFn.getCallee()->stripPointerCasts());
1409 DFSanRuntimeFunctions.insert(
1410 Ptr: DFSanNonzeroLabelFn.getCallee()->stripPointerCasts());
1411 DFSanRuntimeFunctions.insert(
1412 Ptr: DFSanVarargWrapperFn.getCallee()->stripPointerCasts());
1413 DFSanRuntimeFunctions.insert(
1414 Ptr: DFSanLoadCallbackFn.getCallee()->stripPointerCasts());
1415 DFSanRuntimeFunctions.insert(
1416 Ptr: DFSanStoreCallbackFn.getCallee()->stripPointerCasts());
1417 DFSanRuntimeFunctions.insert(
1418 Ptr: DFSanMemTransferCallbackFn.getCallee()->stripPointerCasts());
1419 DFSanRuntimeFunctions.insert(
1420 Ptr: DFSanConditionalCallbackFn.getCallee()->stripPointerCasts());
1421 DFSanRuntimeFunctions.insert(
1422 Ptr: DFSanConditionalCallbackOriginFn.getCallee()->stripPointerCasts());
1423 DFSanRuntimeFunctions.insert(
1424 Ptr: DFSanReachesFunctionCallbackFn.getCallee()->stripPointerCasts());
1425 DFSanRuntimeFunctions.insert(
1426 Ptr: DFSanReachesFunctionCallbackOriginFn.getCallee()->stripPointerCasts());
1427 DFSanRuntimeFunctions.insert(
1428 Ptr: DFSanCmpCallbackFn.getCallee()->stripPointerCasts());
1429 DFSanRuntimeFunctions.insert(
1430 Ptr: DFSanChainOriginFn.getCallee()->stripPointerCasts());
1431 DFSanRuntimeFunctions.insert(
1432 Ptr: DFSanChainOriginIfTaintedFn.getCallee()->stripPointerCasts());
1433 DFSanRuntimeFunctions.insert(
1434 Ptr: DFSanMemOriginTransferFn.getCallee()->stripPointerCasts());
1435 DFSanRuntimeFunctions.insert(
1436 Ptr: DFSanMemShadowOriginTransferFn.getCallee()->stripPointerCasts());
1437 DFSanRuntimeFunctions.insert(
1438 Ptr: DFSanMemShadowOriginConditionalExchangeFn.getCallee()
1439 ->stripPointerCasts());
1440 DFSanRuntimeFunctions.insert(
1441 Ptr: DFSanMaybeStoreOriginFn.getCallee()->stripPointerCasts());
1442}
1443
1444// Initializes event callback functions and declare them in the module
1445void DataFlowSanitizer::initializeCallbackFunctions(Module &M) {
1446 {
1447 AttributeList AL;
1448 AL = AL.addParamAttribute(C&: M.getContext(), ArgNo: 0, Kind: Attribute::ZExt);
1449 DFSanLoadCallbackFn = Mod->getOrInsertFunction(
1450 Name: "__dfsan_load_callback", T: DFSanLoadStoreCallbackFnTy, AttributeList: AL);
1451 }
1452 {
1453 AttributeList AL;
1454 AL = AL.addParamAttribute(C&: M.getContext(), ArgNo: 0, Kind: Attribute::ZExt);
1455 DFSanStoreCallbackFn = Mod->getOrInsertFunction(
1456 Name: "__dfsan_store_callback", T: DFSanLoadStoreCallbackFnTy, AttributeList: AL);
1457 }
1458 DFSanMemTransferCallbackFn = Mod->getOrInsertFunction(
1459 Name: "__dfsan_mem_transfer_callback", T: DFSanMemTransferCallbackFnTy);
1460 {
1461 AttributeList AL;
1462 AL = AL.addParamAttribute(C&: M.getContext(), ArgNo: 0, Kind: Attribute::ZExt);
1463 DFSanCmpCallbackFn = Mod->getOrInsertFunction(Name: "__dfsan_cmp_callback",
1464 T: DFSanCmpCallbackFnTy, AttributeList: AL);
1465 }
1466 {
1467 AttributeList AL;
1468 AL = AL.addParamAttribute(C&: M.getContext(), ArgNo: 0, Kind: Attribute::ZExt);
1469 DFSanConditionalCallbackFn = Mod->getOrInsertFunction(
1470 Name: "__dfsan_conditional_callback", T: DFSanConditionalCallbackFnTy, AttributeList: AL);
1471 }
1472 {
1473 AttributeList AL;
1474 AL = AL.addParamAttribute(C&: M.getContext(), ArgNo: 0, Kind: Attribute::ZExt);
1475 DFSanConditionalCallbackOriginFn =
1476 Mod->getOrInsertFunction(Name: "__dfsan_conditional_callback_origin",
1477 T: DFSanConditionalCallbackOriginFnTy, AttributeList: AL);
1478 }
1479 {
1480 AttributeList AL;
1481 AL = AL.addParamAttribute(C&: M.getContext(), ArgNo: 0, Kind: Attribute::ZExt);
1482 DFSanReachesFunctionCallbackFn =
1483 Mod->getOrInsertFunction(Name: "__dfsan_reaches_function_callback",
1484 T: DFSanReachesFunctionCallbackFnTy, AttributeList: AL);
1485 }
1486 {
1487 AttributeList AL;
1488 AL = AL.addParamAttribute(C&: M.getContext(), ArgNo: 0, Kind: Attribute::ZExt);
1489 DFSanReachesFunctionCallbackOriginFn =
1490 Mod->getOrInsertFunction(Name: "__dfsan_reaches_function_callback_origin",
1491 T: DFSanReachesFunctionCallbackOriginFnTy, AttributeList: AL);
1492 }
1493}
1494
1495bool DataFlowSanitizer::runImpl(
1496 Module &M, llvm::function_ref<TargetLibraryInfo &(Function &)> GetTLI) {
1497 initializeModule(M);
1498
1499 if (ABIList.isIn(M, Category: "skip"))
1500 return false;
1501
1502 const unsigned InitialGlobalSize = M.global_size();
1503 const unsigned InitialModuleSize = M.size();
1504
1505 bool Changed = false;
1506
1507 auto GetOrInsertGlobal = [this, &Changed](StringRef Name,
1508 Type *Ty) -> Constant * {
1509 GlobalVariable *G = Mod->getOrInsertGlobal(Name, Ty);
1510 Changed |= G->getThreadLocalMode() != GlobalVariable::InitialExecTLSModel;
1511 G->setThreadLocalMode(GlobalVariable::InitialExecTLSModel);
1512 return G;
1513 };
1514
1515 // These globals must be kept in sync with the ones in dfsan.cpp.
1516 ArgTLS =
1517 GetOrInsertGlobal("__dfsan_arg_tls",
1518 ArrayType::get(ElementType: Type::getInt64Ty(C&: *Ctx), NumElements: ArgTLSSize / 8));
1519 RetvalTLS = GetOrInsertGlobal(
1520 "__dfsan_retval_tls",
1521 ArrayType::get(ElementType: Type::getInt64Ty(C&: *Ctx), NumElements: RetvalTLSSize / 8));
1522 ArgOriginTLSTy = ArrayType::get(ElementType: OriginTy, NumElements: NumOfElementsInArgOrgTLS);
1523 ArgOriginTLS = GetOrInsertGlobal("__dfsan_arg_origin_tls", ArgOriginTLSTy);
1524 RetvalOriginTLS = GetOrInsertGlobal("__dfsan_retval_origin_tls", OriginTy);
1525
1526 (void)Mod->getOrInsertGlobal(Name: "__dfsan_track_origins", Ty: OriginTy, CreateGlobalCallback: [&] {
1527 Changed = true;
1528 return new GlobalVariable(
1529 M, OriginTy, true, GlobalValue::WeakODRLinkage,
1530 ConstantInt::getSigned(Ty: OriginTy,
1531 V: shouldTrackOrigins() ? ClTrackOrigins : 0),
1532 "__dfsan_track_origins");
1533 });
1534
1535 initializeCallbackFunctions(M);
1536 initializeRuntimeFunctions(M);
1537
1538 std::vector<Function *> FnsToInstrument;
1539 SmallPtrSet<Function *, 2> FnsWithNativeABI;
1540 SmallPtrSet<Function *, 2> FnsWithForceZeroLabel;
1541 SmallPtrSet<Constant *, 1> PersonalityFns;
1542 for (Function &F : M)
1543 if (!F.isIntrinsic() && !DFSanRuntimeFunctions.contains(Ptr: &F) &&
1544 !LibAtomicFunction(F) &&
1545 !F.hasFnAttribute(Kind: Attribute::DisableSanitizerInstrumentation)) {
1546 FnsToInstrument.push_back(x: &F);
1547 if (F.hasPersonalityFn())
1548 PersonalityFns.insert(Ptr: F.getPersonalityFn()->stripPointerCasts());
1549 }
1550
1551 if (ClIgnorePersonalityRoutine) {
1552 for (auto *C : PersonalityFns) {
1553 assert(isa<Function>(C) && "Personality routine is not a function!");
1554 Function *F = cast<Function>(Val: C);
1555 if (!isInstrumented(F))
1556 llvm::erase(C&: FnsToInstrument, V: F);
1557 }
1558 }
1559
1560 // Give function aliases prefixes when necessary, and build wrappers where the
1561 // instrumentedness is inconsistent.
1562 for (GlobalAlias &GA : llvm::make_early_inc_range(Range: M.aliases())) {
1563 // Don't stop on weak. We assume people aren't playing games with the
1564 // instrumentedness of overridden weak aliases.
1565 auto *F = dyn_cast<Function>(Val: GA.getAliaseeObject());
1566 if (!F)
1567 continue;
1568
1569 bool GAInst = isInstrumented(GA: &GA), FInst = isInstrumented(F);
1570 if (GAInst && FInst) {
1571 addGlobalNameSuffix(GV: &GA);
1572 } else if (GAInst != FInst) {
1573 // Non-instrumented alias of an instrumented function, or vice versa.
1574 // Replace the alias with a native-ABI wrapper of the aliasee. The pass
1575 // below will take care of instrumenting it.
1576 Function *NewF =
1577 buildWrapperFunction(F, NewFName: "", NewFLink: GA.getLinkage(), NewFT: F->getFunctionType());
1578 GA.replaceAllUsesWith(V: NewF);
1579 NewF->takeName(V: &GA);
1580 GA.eraseFromParent();
1581 FnsToInstrument.push_back(x: NewF);
1582 }
1583 }
1584
1585 // TODO: This could be more precise.
1586 ReadOnlyNoneAttrs.addAttribute(Val: Attribute::Memory);
1587
1588 // First, change the ABI of every function in the module. ABI-listed
1589 // functions keep their original ABI and get a wrapper function.
1590 for (std::vector<Function *>::iterator FI = FnsToInstrument.begin(),
1591 FE = FnsToInstrument.end();
1592 FI != FE; ++FI) {
1593 Function &F = **FI;
1594 FunctionType *FT = F.getFunctionType();
1595
1596 bool IsZeroArgsVoidRet = (FT->getNumParams() == 0 && !FT->isVarArg() &&
1597 FT->getReturnType()->isVoidTy());
1598
1599 if (isInstrumented(F: &F)) {
1600 if (isForceZeroLabels(F: &F))
1601 FnsWithForceZeroLabel.insert(Ptr: &F);
1602
1603 // Instrumented functions get a '.dfsan' suffix. This allows us to more
1604 // easily identify cases of mismatching ABIs. This naming scheme is
1605 // mangling-compatible (see Itanium ABI), using a vendor-specific suffix.
1606 addGlobalNameSuffix(GV: &F);
1607 } else if (!IsZeroArgsVoidRet || getWrapperKind(F: &F) == WK_Custom) {
1608 // Build a wrapper function for F. The wrapper simply calls F, and is
1609 // added to FnsToInstrument so that any instrumentation according to its
1610 // WrapperKind is done in the second pass below.
1611
1612 // If the function being wrapped has local linkage, then preserve the
1613 // function's linkage in the wrapper function.
1614 GlobalValue::LinkageTypes WrapperLinkage =
1615 F.hasLocalLinkage() ? F.getLinkage()
1616 : GlobalValue::LinkOnceODRLinkage;
1617
1618 Function *NewF = buildWrapperFunction(
1619 F: &F,
1620 NewFName: (shouldTrackOrigins() ? std::string("dfso$") : std::string("dfsw$")) +
1621 std::string(F.getName()),
1622 NewFLink: WrapperLinkage, NewFT: FT);
1623 NewF->removeFnAttrs(Attrs: ReadOnlyNoneAttrs);
1624
1625 // Extern weak functions can sometimes be null at execution time.
1626 // Code will sometimes check if an extern weak function is null.
1627 // This could look something like:
1628 // declare extern_weak i8 @my_func(i8)
1629 // br i1 icmp ne (i8 (i8)* @my_func, i8 (i8)* null), label %use_my_func,
1630 // label %avoid_my_func
1631 // The @"dfsw$my_func" wrapper is never null, so if we replace this use
1632 // in the comparison, the icmp will simplify to false and we have
1633 // accidentally optimized away a null check that is necessary.
1634 // This can lead to a crash when the null extern_weak my_func is called.
1635 //
1636 // To prevent (the most common pattern of) this problem,
1637 // do not replace uses in comparisons with the wrapper.
1638 // We definitely want to replace uses in call instructions.
1639 // Other uses (e.g. store the function address somewhere) might be
1640 // called or compared or both - this case may not be handled correctly.
1641 // We will default to replacing with wrapper in cases we are unsure.
1642 auto IsNotCmpUse = [](Use &U) -> bool {
1643 User *Usr = U.getUser();
1644 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Val: Usr)) {
1645 // This is the most common case for icmp ne null
1646 if (CE->getOpcode() == Instruction::ICmp) {
1647 return false;
1648 }
1649 }
1650 if (Instruction *I = dyn_cast<Instruction>(Val: Usr)) {
1651 if (I->getOpcode() == Instruction::ICmp) {
1652 return false;
1653 }
1654 }
1655 return true;
1656 };
1657 F.replaceUsesWithIf(New: NewF, ShouldReplace: IsNotCmpUse);
1658
1659 UnwrappedFnMap[NewF] = &F;
1660 *FI = NewF;
1661
1662 if (!F.isDeclaration()) {
1663 // This function is probably defining an interposition of an
1664 // uninstrumented function and hence needs to keep the original ABI.
1665 // But any functions it may call need to use the instrumented ABI, so
1666 // we instrument it in a mode which preserves the original ABI.
1667 FnsWithNativeABI.insert(Ptr: &F);
1668
1669 // This code needs to rebuild the iterators, as they may be invalidated
1670 // by the push_back, taking care that the new range does not include
1671 // any functions added by this code.
1672 size_t N = FI - FnsToInstrument.begin(),
1673 Count = FE - FnsToInstrument.begin();
1674 FnsToInstrument.push_back(x: &F);
1675 FI = FnsToInstrument.begin() + N;
1676 FE = FnsToInstrument.begin() + Count;
1677 }
1678 // Hopefully, nobody will try to indirectly call a vararg
1679 // function... yet.
1680 } else if (FT->isVarArg()) {
1681 UnwrappedFnMap[&F] = &F;
1682 *FI = nullptr;
1683 }
1684 }
1685
1686 for (Function *F : FnsToInstrument) {
1687 if (!F || F->isDeclaration())
1688 continue;
1689
1690 removeUnreachableBlocks(F&: *F);
1691
1692 DFSanFunction DFSF(*this, F, FnsWithNativeABI.count(Ptr: F),
1693 FnsWithForceZeroLabel.count(Ptr: F), GetTLI(*F));
1694
1695 if (ClReachesFunctionCallbacks) {
1696 // Add callback for arguments reaching this function.
1697 for (auto &FArg : F->args()) {
1698 Instruction *Next = &F->getEntryBlock().front();
1699 Value *FArgShadow = DFSF.getShadow(V: &FArg);
1700 if (isZeroShadow(V: FArgShadow))
1701 continue;
1702 if (Instruction *FArgShadowInst = dyn_cast<Instruction>(Val: FArgShadow)) {
1703 Next = FArgShadowInst->getNextNode();
1704 }
1705 if (shouldTrackOrigins()) {
1706 if (Instruction *Origin =
1707 dyn_cast<Instruction>(Val: DFSF.getOrigin(V: &FArg))) {
1708 // Ensure IRB insertion point is after loads for shadow and origin.
1709 Instruction *OriginNext = Origin->getNextNode();
1710 if (Next->comesBefore(Other: OriginNext)) {
1711 Next = OriginNext;
1712 }
1713 }
1714 }
1715 IRBuilder<> IRB(Next);
1716 DFSF.addReachesFunctionCallbacksIfEnabled(IRB, I&: *Next, Data: &FArg);
1717 }
1718 }
1719
1720 // DFSanVisitor may create new basic blocks, which confuses df_iterator.
1721 // Build a copy of the list before iterating over it.
1722 SmallVector<BasicBlock *, 4> BBList(depth_first(G: &F->getEntryBlock()));
1723
1724 for (BasicBlock *BB : BBList) {
1725 Instruction *Inst = &BB->front();
1726 while (true) {
1727 // DFSanVisitor may split the current basic block, changing the current
1728 // instruction's next pointer and moving the next instruction to the
1729 // tail block from which we should continue.
1730 Instruction *Next = Inst->getNextNode();
1731 // DFSanVisitor may delete Inst, so keep track of whether it was a
1732 // terminator.
1733 bool IsTerminator = Inst->isTerminator();
1734 if (!DFSF.SkipInsts.count(V: Inst))
1735 DFSanVisitor(DFSF).visit(I: Inst);
1736 if (IsTerminator)
1737 break;
1738 Inst = Next;
1739 }
1740 }
1741
1742 // We will not necessarily be able to compute the shadow for every phi node
1743 // until we have visited every block. Therefore, the code that handles phi
1744 // nodes adds them to the PHIFixups list so that they can be properly
1745 // handled here.
1746 for (DFSanFunction::PHIFixupElement &P : DFSF.PHIFixups) {
1747 for (unsigned Val = 0, N = P.Phi->getNumIncomingValues(); Val != N;
1748 ++Val) {
1749 P.ShadowPhi->setIncomingValue(
1750 i: Val, V: DFSF.getShadow(V: P.Phi->getIncomingValue(i: Val)));
1751 if (P.OriginPhi)
1752 P.OriginPhi->setIncomingValue(
1753 i: Val, V: DFSF.getOrigin(V: P.Phi->getIncomingValue(i: Val)));
1754 }
1755 }
1756
1757 // -dfsan-debug-nonzero-labels will split the CFG in all kinds of crazy
1758 // places (i.e. instructions in basic blocks we haven't even begun visiting
1759 // yet). To make our life easier, do this work in a pass after the main
1760 // instrumentation.
1761 if (ClDebugNonzeroLabels) {
1762 for (Value *V : DFSF.NonZeroChecks) {
1763 BasicBlock::iterator Pos;
1764 if (Instruction *I = dyn_cast<Instruction>(Val: V))
1765 Pos = std::next(x: I->getIterator());
1766 else
1767 Pos = DFSF.F->getEntryBlock().begin();
1768 while (isa<PHINode>(Val: Pos) || isa<AllocaInst>(Val: Pos))
1769 Pos = std::next(x: Pos->getIterator());
1770 IRBuilder<> IRB(Pos->getParent(), Pos);
1771 Value *PrimitiveShadow = DFSF.collapseToPrimitiveShadow(Shadow: V, Pos);
1772 Value *Ne =
1773 IRB.CreateICmpNE(LHS: PrimitiveShadow, RHS: DFSF.DFS.ZeroPrimitiveShadow);
1774 BranchInst *BI = cast<BranchInst>(Val: SplitBlockAndInsertIfThen(
1775 Cond: Ne, SplitBefore: Pos, /*Unreachable=*/false, BranchWeights: ColdCallWeights));
1776 IRBuilder<> ThenIRB(BI);
1777 ThenIRB.CreateCall(Callee: DFSF.DFS.DFSanNonzeroLabelFn, Args: {});
1778 }
1779 }
1780 }
1781
1782 return Changed || !FnsToInstrument.empty() ||
1783 M.global_size() != InitialGlobalSize || M.size() != InitialModuleSize;
1784}
1785
1786Value *DFSanFunction::getArgTLS(Type *T, unsigned ArgOffset, IRBuilder<> &IRB) {
1787 Value *Base = IRB.CreatePointerCast(V: DFS.ArgTLS, DestTy: DFS.IntptrTy);
1788 if (ArgOffset)
1789 Base = IRB.CreateAdd(LHS: Base, RHS: ConstantInt::get(Ty: DFS.IntptrTy, V: ArgOffset));
1790 return IRB.CreateIntToPtr(V: Base, DestTy: PointerType::get(C&: *DFS.Ctx, AddressSpace: 0), Name: "_dfsarg");
1791}
1792
1793Value *DFSanFunction::getRetvalTLS(Type *T, IRBuilder<> &IRB) {
1794 return IRB.CreatePointerCast(V: DFS.RetvalTLS, DestTy: PointerType::get(C&: *DFS.Ctx, AddressSpace: 0),
1795 Name: "_dfsret");
1796}
1797
1798Value *DFSanFunction::getRetvalOriginTLS() { return DFS.RetvalOriginTLS; }
1799
1800Value *DFSanFunction::getArgOriginTLS(unsigned ArgNo, IRBuilder<> &IRB) {
1801 return IRB.CreateConstInBoundsGEP2_64(Ty: DFS.ArgOriginTLSTy, Ptr: DFS.ArgOriginTLS, Idx0: 0,
1802 Idx1: ArgNo, Name: "_dfsarg_o");
1803}
1804
1805Value *DFSanFunction::getOrigin(Value *V) {
1806 assert(DFS.shouldTrackOrigins());
1807 if (!isa<Argument>(Val: V) && !isa<Instruction>(Val: V))
1808 return DFS.ZeroOrigin;
1809 Value *&Origin = ValOriginMap[V];
1810 if (!Origin) {
1811 if (Argument *A = dyn_cast<Argument>(Val: V)) {
1812 if (IsNativeABI)
1813 return DFS.ZeroOrigin;
1814 if (A->getArgNo() < DFS.NumOfElementsInArgOrgTLS) {
1815 Instruction *ArgOriginTLSPos = &*F->getEntryBlock().begin();
1816 IRBuilder<> IRB(ArgOriginTLSPos);
1817 Value *ArgOriginPtr = getArgOriginTLS(ArgNo: A->getArgNo(), IRB);
1818 Origin = IRB.CreateLoad(Ty: DFS.OriginTy, Ptr: ArgOriginPtr);
1819 } else {
1820 // Overflow
1821 Origin = DFS.ZeroOrigin;
1822 }
1823 } else {
1824 Origin = DFS.ZeroOrigin;
1825 }
1826 }
1827 return Origin;
1828}
1829
1830void DFSanFunction::setOrigin(Instruction *I, Value *Origin) {
1831 if (!DFS.shouldTrackOrigins())
1832 return;
1833 assert(!ValOriginMap.count(I));
1834 assert(Origin->getType() == DFS.OriginTy);
1835 ValOriginMap[I] = Origin;
1836}
1837
1838Value *DFSanFunction::getShadowForTLSArgument(Argument *A) {
1839 unsigned ArgOffset = 0;
1840 const DataLayout &DL = F->getDataLayout();
1841 for (auto &FArg : F->args()) {
1842 if (!FArg.getType()->isSized()) {
1843 if (A == &FArg)
1844 break;
1845 continue;
1846 }
1847
1848 unsigned Size = DL.getTypeAllocSize(Ty: DFS.getShadowTy(V: &FArg));
1849 if (A != &FArg) {
1850 ArgOffset += alignTo(Size, A: ShadowTLSAlignment);
1851 if (ArgOffset > ArgTLSSize)
1852 break; // ArgTLS overflows, uses a zero shadow.
1853 continue;
1854 }
1855
1856 if (ArgOffset + Size > ArgTLSSize)
1857 break; // ArgTLS overflows, uses a zero shadow.
1858
1859 Instruction *ArgTLSPos = &*F->getEntryBlock().begin();
1860 IRBuilder<> IRB(ArgTLSPos);
1861 Value *ArgShadowPtr = getArgTLS(T: FArg.getType(), ArgOffset, IRB);
1862 return IRB.CreateAlignedLoad(Ty: DFS.getShadowTy(V: &FArg), Ptr: ArgShadowPtr,
1863 Align: ShadowTLSAlignment);
1864 }
1865
1866 return DFS.getZeroShadow(V: A);
1867}
1868
1869Value *DFSanFunction::getShadow(Value *V) {
1870 if (!isa<Argument>(Val: V) && !isa<Instruction>(Val: V))
1871 return DFS.getZeroShadow(V);
1872 if (IsForceZeroLabels)
1873 return DFS.getZeroShadow(V);
1874 Value *&Shadow = ValShadowMap[V];
1875 if (!Shadow) {
1876 if (Argument *A = dyn_cast<Argument>(Val: V)) {
1877 if (IsNativeABI)
1878 return DFS.getZeroShadow(V);
1879 Shadow = getShadowForTLSArgument(A);
1880 NonZeroChecks.push_back(x: Shadow);
1881 } else {
1882 Shadow = DFS.getZeroShadow(V);
1883 }
1884 }
1885 return Shadow;
1886}
1887
1888void DFSanFunction::setShadow(Instruction *I, Value *Shadow) {
1889 assert(!ValShadowMap.count(I));
1890 ValShadowMap[I] = Shadow;
1891}
1892
1893/// Compute the integer shadow offset that corresponds to a given
1894/// application address.
1895///
1896/// Offset = (Addr & ~AndMask) ^ XorMask
1897Value *DataFlowSanitizer::getShadowOffset(Value *Addr, IRBuilder<> &IRB) {
1898 assert(Addr != RetvalTLS && "Reinstrumenting?");
1899 Value *OffsetLong = IRB.CreatePointerCast(V: Addr, DestTy: IntptrTy);
1900
1901 uint64_t AndMask = MapParams->AndMask;
1902 if (AndMask)
1903 OffsetLong =
1904 IRB.CreateAnd(LHS: OffsetLong, RHS: ConstantInt::get(Ty: IntptrTy, V: ~AndMask));
1905
1906 uint64_t XorMask = MapParams->XorMask;
1907 if (XorMask)
1908 OffsetLong = IRB.CreateXor(LHS: OffsetLong, RHS: ConstantInt::get(Ty: IntptrTy, V: XorMask));
1909 return OffsetLong;
1910}
1911
1912std::pair<Value *, Value *>
1913DataFlowSanitizer::getShadowOriginAddress(Value *Addr, Align InstAlignment,
1914 BasicBlock::iterator Pos) {
1915 // Returns ((Addr & shadow_mask) + origin_base - shadow_base) & ~4UL
1916 IRBuilder<> IRB(Pos->getParent(), Pos);
1917 Value *ShadowOffset = getShadowOffset(Addr, IRB);
1918 Value *ShadowLong = ShadowOffset;
1919 uint64_t ShadowBase = MapParams->ShadowBase;
1920 if (ShadowBase != 0) {
1921 ShadowLong =
1922 IRB.CreateAdd(LHS: ShadowLong, RHS: ConstantInt::get(Ty: IntptrTy, V: ShadowBase));
1923 }
1924 Value *ShadowPtr = IRB.CreateIntToPtr(V: ShadowLong, DestTy: PointerType::get(C&: *Ctx, AddressSpace: 0));
1925 Value *OriginPtr = nullptr;
1926 if (shouldTrackOrigins()) {
1927 Value *OriginLong = ShadowOffset;
1928 uint64_t OriginBase = MapParams->OriginBase;
1929 if (OriginBase != 0)
1930 OriginLong =
1931 IRB.CreateAdd(LHS: OriginLong, RHS: ConstantInt::get(Ty: IntptrTy, V: OriginBase));
1932 const Align Alignment = llvm::assumeAligned(Value: InstAlignment.value());
1933 // When alignment is >= 4, Addr must be aligned to 4, otherwise it is UB.
1934 // So Mask is unnecessary.
1935 if (Alignment < MinOriginAlignment) {
1936 uint64_t Mask = MinOriginAlignment.value() - 1;
1937 OriginLong = IRB.CreateAnd(LHS: OriginLong, RHS: ConstantInt::get(Ty: IntptrTy, V: ~Mask));
1938 }
1939 OriginPtr = IRB.CreateIntToPtr(V: OriginLong, DestTy: OriginPtrTy);
1940 }
1941 return std::make_pair(x&: ShadowPtr, y&: OriginPtr);
1942}
1943
1944Value *DataFlowSanitizer::getShadowAddress(Value *Addr,
1945 BasicBlock::iterator Pos,
1946 Value *ShadowOffset) {
1947 IRBuilder<> IRB(Pos->getParent(), Pos);
1948 return IRB.CreateIntToPtr(V: ShadowOffset, DestTy: PrimitiveShadowPtrTy);
1949}
1950
1951Value *DataFlowSanitizer::getShadowAddress(Value *Addr,
1952 BasicBlock::iterator Pos) {
1953 IRBuilder<> IRB(Pos->getParent(), Pos);
1954 Value *ShadowOffset = getShadowOffset(Addr, IRB);
1955 return getShadowAddress(Addr, Pos, ShadowOffset);
1956}
1957
1958Value *DFSanFunction::combineShadowsThenConvert(Type *T, Value *V1, Value *V2,
1959 BasicBlock::iterator Pos) {
1960 Value *PrimitiveValue = combineShadows(V1, V2, Pos);
1961 return expandFromPrimitiveShadow(T, PrimitiveShadow: PrimitiveValue, Pos);
1962}
1963
1964// Generates IR to compute the union of the two given shadows, inserting it
1965// before Pos. The combined value is with primitive type.
1966Value *DFSanFunction::combineShadows(Value *V1, Value *V2,
1967 BasicBlock::iterator Pos) {
1968 if (DFS.isZeroShadow(V: V1))
1969 return collapseToPrimitiveShadow(Shadow: V2, Pos);
1970 if (DFS.isZeroShadow(V: V2))
1971 return collapseToPrimitiveShadow(Shadow: V1, Pos);
1972 if (V1 == V2)
1973 return collapseToPrimitiveShadow(Shadow: V1, Pos);
1974
1975 auto V1Elems = ShadowElements.find(Val: V1);
1976 auto V2Elems = ShadowElements.find(Val: V2);
1977 if (V1Elems != ShadowElements.end() && V2Elems != ShadowElements.end()) {
1978 if (llvm::includes(Range1&: V1Elems->second, Range2&: V2Elems->second)) {
1979 return collapseToPrimitiveShadow(Shadow: V1, Pos);
1980 }
1981 if (llvm::includes(Range1&: V2Elems->second, Range2&: V1Elems->second)) {
1982 return collapseToPrimitiveShadow(Shadow: V2, Pos);
1983 }
1984 } else if (V1Elems != ShadowElements.end()) {
1985 if (V1Elems->second.count(x: V2))
1986 return collapseToPrimitiveShadow(Shadow: V1, Pos);
1987 } else if (V2Elems != ShadowElements.end()) {
1988 if (V2Elems->second.count(x: V1))
1989 return collapseToPrimitiveShadow(Shadow: V2, Pos);
1990 }
1991
1992 auto Key = std::make_pair(x&: V1, y&: V2);
1993 if (V1 > V2)
1994 std::swap(a&: Key.first, b&: Key.second);
1995 CachedShadow &CCS = CachedShadows[Key];
1996 if (CCS.Block && DT.dominates(A: CCS.Block, B: Pos->getParent()))
1997 return CCS.Shadow;
1998
1999 // Converts inputs shadows to shadows with primitive types.
2000 Value *PV1 = collapseToPrimitiveShadow(Shadow: V1, Pos);
2001 Value *PV2 = collapseToPrimitiveShadow(Shadow: V2, Pos);
2002
2003 IRBuilder<> IRB(Pos->getParent(), Pos);
2004 CCS.Block = Pos->getParent();
2005 CCS.Shadow = IRB.CreateOr(LHS: PV1, RHS: PV2);
2006
2007 std::set<Value *> UnionElems;
2008 if (V1Elems != ShadowElements.end()) {
2009 UnionElems = V1Elems->second;
2010 } else {
2011 UnionElems.insert(x: V1);
2012 }
2013 if (V2Elems != ShadowElements.end()) {
2014 UnionElems.insert(first: V2Elems->second.begin(), last: V2Elems->second.end());
2015 } else {
2016 UnionElems.insert(x: V2);
2017 }
2018 ShadowElements[CCS.Shadow] = std::move(UnionElems);
2019
2020 return CCS.Shadow;
2021}
2022
2023// A convenience function which folds the shadows of each of the operands
2024// of the provided instruction Inst, inserting the IR before Inst. Returns
2025// the computed union Value.
2026Value *DFSanFunction::combineOperandShadows(Instruction *Inst) {
2027 if (Inst->getNumOperands() == 0)
2028 return DFS.getZeroShadow(V: Inst);
2029
2030 Value *Shadow = getShadow(V: Inst->getOperand(i: 0));
2031 for (unsigned I = 1, N = Inst->getNumOperands(); I < N; ++I)
2032 Shadow = combineShadows(V1: Shadow, V2: getShadow(V: Inst->getOperand(i: I)),
2033 Pos: Inst->getIterator());
2034
2035 return expandFromPrimitiveShadow(T: Inst->getType(), PrimitiveShadow: Shadow,
2036 Pos: Inst->getIterator());
2037}
2038
2039void DFSanVisitor::visitInstOperands(Instruction &I) {
2040 Value *CombinedShadow = DFSF.combineOperandShadows(Inst: &I);
2041 DFSF.setShadow(I: &I, Shadow: CombinedShadow);
2042 visitInstOperandOrigins(I);
2043}
2044
2045Value *DFSanFunction::combineOrigins(const std::vector<Value *> &Shadows,
2046 const std::vector<Value *> &Origins,
2047 BasicBlock::iterator Pos,
2048 ConstantInt *Zero) {
2049 assert(Shadows.size() == Origins.size());
2050 size_t Size = Origins.size();
2051 if (Size == 0)
2052 return DFS.ZeroOrigin;
2053 Value *Origin = nullptr;
2054 if (!Zero)
2055 Zero = DFS.ZeroPrimitiveShadow;
2056 for (size_t I = 0; I != Size; ++I) {
2057 Value *OpOrigin = Origins[I];
2058 Constant *ConstOpOrigin = dyn_cast<Constant>(Val: OpOrigin);
2059 if (ConstOpOrigin && ConstOpOrigin->isNullValue())
2060 continue;
2061 if (!Origin) {
2062 Origin = OpOrigin;
2063 continue;
2064 }
2065 Value *OpShadow = Shadows[I];
2066 Value *PrimitiveShadow = collapseToPrimitiveShadow(Shadow: OpShadow, Pos);
2067 IRBuilder<> IRB(Pos->getParent(), Pos);
2068 Value *Cond = IRB.CreateICmpNE(LHS: PrimitiveShadow, RHS: Zero);
2069 Origin = IRB.CreateSelect(C: Cond, True: OpOrigin, False: Origin);
2070 }
2071 return Origin ? Origin : DFS.ZeroOrigin;
2072}
2073
2074Value *DFSanFunction::combineOperandOrigins(Instruction *Inst) {
2075 size_t Size = Inst->getNumOperands();
2076 std::vector<Value *> Shadows(Size);
2077 std::vector<Value *> Origins(Size);
2078 for (unsigned I = 0; I != Size; ++I) {
2079 Shadows[I] = getShadow(V: Inst->getOperand(i: I));
2080 Origins[I] = getOrigin(V: Inst->getOperand(i: I));
2081 }
2082 return combineOrigins(Shadows, Origins, Pos: Inst->getIterator());
2083}
2084
2085void DFSanVisitor::visitInstOperandOrigins(Instruction &I) {
2086 if (!DFSF.DFS.shouldTrackOrigins())
2087 return;
2088 Value *CombinedOrigin = DFSF.combineOperandOrigins(Inst: &I);
2089 DFSF.setOrigin(I: &I, Origin: CombinedOrigin);
2090}
2091
2092Align DFSanFunction::getShadowAlign(Align InstAlignment) {
2093 const Align Alignment = ClPreserveAlignment ? InstAlignment : Align(1);
2094 return Align(Alignment.value() * DFS.ShadowWidthBytes);
2095}
2096
2097Align DFSanFunction::getOriginAlign(Align InstAlignment) {
2098 const Align Alignment = llvm::assumeAligned(Value: InstAlignment.value());
2099 return Align(std::max(a: MinOriginAlignment, b: Alignment));
2100}
2101
2102bool DFSanFunction::isLookupTableConstant(Value *P) {
2103 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Val: P->stripPointerCasts()))
2104 if (GV->isConstant() && GV->hasName())
2105 return DFS.CombineTaintLookupTableNames.count(Key: GV->getName());
2106
2107 return false;
2108}
2109
2110bool DFSanFunction::useCallbackLoadLabelAndOrigin(uint64_t Size,
2111 Align InstAlignment) {
2112 // When enabling tracking load instructions, we always use
2113 // __dfsan_load_label_and_origin to reduce code size.
2114 if (ClTrackOrigins == 2)
2115 return true;
2116
2117 assert(Size != 0);
2118 // * if Size == 1, it is sufficient to load its origin aligned at 4.
2119 // * if Size == 2, we assume most cases Addr % 2 == 0, so it is sufficient to
2120 // load its origin aligned at 4. If not, although origins may be lost, it
2121 // should not happen very often.
2122 // * if align >= 4, Addr must be aligned to 4, otherwise it is UB. When
2123 // Size % 4 == 0, it is more efficient to load origins without callbacks.
2124 // * Otherwise we use __dfsan_load_label_and_origin.
2125 // This should ensure that common cases run efficiently.
2126 if (Size <= 2)
2127 return false;
2128
2129 const Align Alignment = llvm::assumeAligned(Value: InstAlignment.value());
2130 return Alignment < MinOriginAlignment || !DFS.hasLoadSizeForFastPath(Size);
2131}
2132
2133Value *DataFlowSanitizer::loadNextOrigin(BasicBlock::iterator Pos,
2134 Align OriginAlign,
2135 Value **OriginAddr) {
2136 IRBuilder<> IRB(Pos->getParent(), Pos);
2137 *OriginAddr =
2138 IRB.CreateGEP(Ty: OriginTy, Ptr: *OriginAddr, IdxList: ConstantInt::get(Ty: IntptrTy, V: 1));
2139 return IRB.CreateAlignedLoad(Ty: OriginTy, Ptr: *OriginAddr, Align: OriginAlign);
2140}
2141
2142std::pair<Value *, Value *> DFSanFunction::loadShadowFast(
2143 Value *ShadowAddr, Value *OriginAddr, uint64_t Size, Align ShadowAlign,
2144 Align OriginAlign, Value *FirstOrigin, BasicBlock::iterator Pos) {
2145 const bool ShouldTrackOrigins = DFS.shouldTrackOrigins();
2146 const uint64_t ShadowSize = Size * DFS.ShadowWidthBytes;
2147
2148 assert(Size >= 4 && "Not large enough load size for fast path!");
2149
2150 // Used for origin tracking.
2151 std::vector<Value *> Shadows;
2152 std::vector<Value *> Origins;
2153
2154 // Load instructions in LLVM can have arbitrary byte sizes (e.g., 3, 12, 20)
2155 // but this function is only used in a subset of cases that make it possible
2156 // to optimize the instrumentation.
2157 //
2158 // Specifically, when the shadow size in bytes (i.e., loaded bytes x shadow
2159 // per byte) is either:
2160 // - a multiple of 8 (common)
2161 // - equal to 4 (only for load32)
2162 //
2163 // For the second case, we can fit the wide shadow in a 32-bit integer. In all
2164 // other cases, we use a 64-bit integer to hold the wide shadow.
2165 Type *WideShadowTy =
2166 ShadowSize == 4 ? Type::getInt32Ty(C&: *DFS.Ctx) : Type::getInt64Ty(C&: *DFS.Ctx);
2167
2168 IRBuilder<> IRB(Pos->getParent(), Pos);
2169 Value *CombinedWideShadow =
2170 IRB.CreateAlignedLoad(Ty: WideShadowTy, Ptr: ShadowAddr, Align: ShadowAlign);
2171
2172 unsigned WideShadowBitWidth = WideShadowTy->getIntegerBitWidth();
2173 const uint64_t BytesPerWideShadow = WideShadowBitWidth / DFS.ShadowWidthBits;
2174
2175 auto AppendWideShadowAndOrigin = [&](Value *WideShadow, Value *Origin) {
2176 if (BytesPerWideShadow > 4) {
2177 assert(BytesPerWideShadow == 8);
2178 // The wide shadow relates to two origin pointers: one for the first four
2179 // application bytes, and one for the latest four. We use a left shift to
2180 // get just the shadow bytes that correspond to the first origin pointer,
2181 // and then the entire shadow for the second origin pointer (which will be
2182 // chosen by combineOrigins() iff the least-significant half of the wide
2183 // shadow was empty but the other half was not).
2184 Value *WideShadowLo = IRB.CreateShl(
2185 LHS: WideShadow, RHS: ConstantInt::get(Ty: WideShadowTy, V: WideShadowBitWidth / 2));
2186 Shadows.push_back(x: WideShadow);
2187 Origins.push_back(x: DFS.loadNextOrigin(Pos, OriginAlign, OriginAddr: &OriginAddr));
2188
2189 Shadows.push_back(x: WideShadowLo);
2190 Origins.push_back(x: Origin);
2191 } else {
2192 Shadows.push_back(x: WideShadow);
2193 Origins.push_back(x: Origin);
2194 }
2195 };
2196
2197 if (ShouldTrackOrigins)
2198 AppendWideShadowAndOrigin(CombinedWideShadow, FirstOrigin);
2199
2200 // First OR all the WideShadows (i.e., 64bit or 32bit shadow chunks) linearly;
2201 // then OR individual shadows within the combined WideShadow by binary ORing.
2202 // This is fewer instructions than ORing shadows individually, since it
2203 // needs logN shift/or instructions (N being the bytes of the combined wide
2204 // shadow).
2205 for (uint64_t ByteOfs = BytesPerWideShadow; ByteOfs < Size;
2206 ByteOfs += BytesPerWideShadow) {
2207 ShadowAddr = IRB.CreateGEP(Ty: WideShadowTy, Ptr: ShadowAddr,
2208 IdxList: ConstantInt::get(Ty: DFS.IntptrTy, V: 1));
2209 Value *NextWideShadow =
2210 IRB.CreateAlignedLoad(Ty: WideShadowTy, Ptr: ShadowAddr, Align: ShadowAlign);
2211 CombinedWideShadow = IRB.CreateOr(LHS: CombinedWideShadow, RHS: NextWideShadow);
2212 if (ShouldTrackOrigins) {
2213 Value *NextOrigin = DFS.loadNextOrigin(Pos, OriginAlign, OriginAddr: &OriginAddr);
2214 AppendWideShadowAndOrigin(NextWideShadow, NextOrigin);
2215 }
2216 }
2217 for (unsigned Width = WideShadowBitWidth / 2; Width >= DFS.ShadowWidthBits;
2218 Width >>= 1) {
2219 Value *ShrShadow = IRB.CreateLShr(LHS: CombinedWideShadow, RHS: Width);
2220 CombinedWideShadow = IRB.CreateOr(LHS: CombinedWideShadow, RHS: ShrShadow);
2221 }
2222 return {IRB.CreateTrunc(V: CombinedWideShadow, DestTy: DFS.PrimitiveShadowTy),
2223 ShouldTrackOrigins
2224 ? combineOrigins(Shadows, Origins, Pos,
2225 Zero: ConstantInt::getSigned(Ty: IRB.getInt64Ty(), V: 0))
2226 : DFS.ZeroOrigin};
2227}
2228
2229std::pair<Value *, Value *> DFSanFunction::loadShadowOriginSansLoadTracking(
2230 Value *Addr, uint64_t Size, Align InstAlignment, BasicBlock::iterator Pos) {
2231 const bool ShouldTrackOrigins = DFS.shouldTrackOrigins();
2232
2233 // Non-escaped loads.
2234 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val: Addr)) {
2235 const auto SI = AllocaShadowMap.find(Val: AI);
2236 if (SI != AllocaShadowMap.end()) {
2237 IRBuilder<> IRB(Pos->getParent(), Pos);
2238 Value *ShadowLI = IRB.CreateLoad(Ty: DFS.PrimitiveShadowTy, Ptr: SI->second);
2239 const auto OI = AllocaOriginMap.find(Val: AI);
2240 assert(!ShouldTrackOrigins || OI != AllocaOriginMap.end());
2241 return {ShadowLI, ShouldTrackOrigins
2242 ? IRB.CreateLoad(Ty: DFS.OriginTy, Ptr: OI->second)
2243 : nullptr};
2244 }
2245 }
2246
2247 // Load from constant addresses.
2248 SmallVector<const Value *, 2> Objs;
2249 getUnderlyingObjects(V: Addr, Objects&: Objs);
2250 bool AllConstants = true;
2251 for (const Value *Obj : Objs) {
2252 if (isa<Function>(Val: Obj) || isa<BlockAddress>(Val: Obj))
2253 continue;
2254 if (isa<GlobalVariable>(Val: Obj) && cast<GlobalVariable>(Val: Obj)->isConstant())
2255 continue;
2256
2257 AllConstants = false;
2258 break;
2259 }
2260 if (AllConstants)
2261 return {DFS.ZeroPrimitiveShadow,
2262 ShouldTrackOrigins ? DFS.ZeroOrigin : nullptr};
2263
2264 if (Size == 0)
2265 return {DFS.ZeroPrimitiveShadow,
2266 ShouldTrackOrigins ? DFS.ZeroOrigin : nullptr};
2267
2268 // Use callback to load if this is not an optimizable case for origin
2269 // tracking.
2270 if (ShouldTrackOrigins &&
2271 useCallbackLoadLabelAndOrigin(Size, InstAlignment)) {
2272 IRBuilder<> IRB(Pos->getParent(), Pos);
2273 CallInst *Call =
2274 IRB.CreateCall(Callee: DFS.DFSanLoadLabelAndOriginFn,
2275 Args: {Addr, ConstantInt::get(Ty: DFS.IntptrTy, V: Size)});
2276 Call->addRetAttr(Kind: Attribute::ZExt);
2277 return {IRB.CreateTrunc(V: IRB.CreateLShr(LHS: Call, RHS: DFS.OriginWidthBits),
2278 DestTy: DFS.PrimitiveShadowTy),
2279 IRB.CreateTrunc(V: Call, DestTy: DFS.OriginTy)};
2280 }
2281
2282 // Other cases that support loading shadows or origins in a fast way.
2283 Value *ShadowAddr, *OriginAddr;
2284 std::tie(args&: ShadowAddr, args&: OriginAddr) =
2285 DFS.getShadowOriginAddress(Addr, InstAlignment, Pos);
2286
2287 const Align ShadowAlign = getShadowAlign(InstAlignment);
2288 const Align OriginAlign = getOriginAlign(InstAlignment);
2289 Value *Origin = nullptr;
2290 if (ShouldTrackOrigins) {
2291 IRBuilder<> IRB(Pos->getParent(), Pos);
2292 Origin = IRB.CreateAlignedLoad(Ty: DFS.OriginTy, Ptr: OriginAddr, Align: OriginAlign);
2293 }
2294
2295 // When the byte size is small enough, we can load the shadow directly with
2296 // just a few instructions.
2297 switch (Size) {
2298 case 1: {
2299 LoadInst *LI = new LoadInst(DFS.PrimitiveShadowTy, ShadowAddr, "", Pos);
2300 LI->setAlignment(ShadowAlign);
2301 return {LI, Origin};
2302 }
2303 case 2: {
2304 IRBuilder<> IRB(Pos->getParent(), Pos);
2305 Value *ShadowAddr1 = IRB.CreateGEP(Ty: DFS.PrimitiveShadowTy, Ptr: ShadowAddr,
2306 IdxList: ConstantInt::get(Ty: DFS.IntptrTy, V: 1));
2307 Value *Load =
2308 IRB.CreateAlignedLoad(Ty: DFS.PrimitiveShadowTy, Ptr: ShadowAddr, Align: ShadowAlign);
2309 Value *Load1 =
2310 IRB.CreateAlignedLoad(Ty: DFS.PrimitiveShadowTy, Ptr: ShadowAddr1, Align: ShadowAlign);
2311 return {combineShadows(V1: Load, V2: Load1, Pos), Origin};
2312 }
2313 }
2314 bool HasSizeForFastPath = DFS.hasLoadSizeForFastPath(Size);
2315
2316 if (HasSizeForFastPath)
2317 return loadShadowFast(ShadowAddr, OriginAddr, Size, ShadowAlign,
2318 OriginAlign, FirstOrigin: Origin, Pos);
2319
2320 IRBuilder<> IRB(Pos->getParent(), Pos);
2321 CallInst *FallbackCall = IRB.CreateCall(
2322 Callee: DFS.DFSanUnionLoadFn, Args: {ShadowAddr, ConstantInt::get(Ty: DFS.IntptrTy, V: Size)});
2323 FallbackCall->addRetAttr(Kind: Attribute::ZExt);
2324 return {FallbackCall, Origin};
2325}
2326
2327std::pair<Value *, Value *>
2328DFSanFunction::loadShadowOrigin(Value *Addr, uint64_t Size, Align InstAlignment,
2329 BasicBlock::iterator Pos) {
2330 Value *PrimitiveShadow, *Origin;
2331 std::tie(args&: PrimitiveShadow, args&: Origin) =
2332 loadShadowOriginSansLoadTracking(Addr, Size, InstAlignment, Pos);
2333 if (DFS.shouldTrackOrigins()) {
2334 if (ClTrackOrigins == 2) {
2335 IRBuilder<> IRB(Pos->getParent(), Pos);
2336 auto *ConstantShadow = dyn_cast<Constant>(Val: PrimitiveShadow);
2337 if (!ConstantShadow || !ConstantShadow->isZeroValue())
2338 Origin = updateOriginIfTainted(Shadow: PrimitiveShadow, Origin, IRB);
2339 }
2340 }
2341 return {PrimitiveShadow, Origin};
2342}
2343
2344static AtomicOrdering addAcquireOrdering(AtomicOrdering AO) {
2345 switch (AO) {
2346 case AtomicOrdering::NotAtomic:
2347 return AtomicOrdering::NotAtomic;
2348 case AtomicOrdering::Unordered:
2349 case AtomicOrdering::Monotonic:
2350 case AtomicOrdering::Acquire:
2351 return AtomicOrdering::Acquire;
2352 case AtomicOrdering::Release:
2353 case AtomicOrdering::AcquireRelease:
2354 return AtomicOrdering::AcquireRelease;
2355 case AtomicOrdering::SequentiallyConsistent:
2356 return AtomicOrdering::SequentiallyConsistent;
2357 }
2358 llvm_unreachable("Unknown ordering");
2359}
2360
2361Value *StripPointerGEPsAndCasts(Value *V) {
2362 if (!V->getType()->isPointerTy())
2363 return V;
2364
2365 // DFSan pass should be running on valid IR, but we'll
2366 // keep a seen set to ensure there are no issues.
2367 SmallPtrSet<const Value *, 4> Visited;
2368 Visited.insert(Ptr: V);
2369 do {
2370 if (auto *GEP = dyn_cast<GEPOperator>(Val: V)) {
2371 V = GEP->getPointerOperand();
2372 } else if (Operator::getOpcode(V) == Instruction::BitCast) {
2373 V = cast<Operator>(Val: V)->getOperand(i: 0);
2374 if (!V->getType()->isPointerTy())
2375 return V;
2376 } else if (isa<GlobalAlias>(Val: V)) {
2377 V = cast<GlobalAlias>(Val: V)->getAliasee();
2378 }
2379 } while (Visited.insert(Ptr: V).second);
2380
2381 return V;
2382}
2383
2384void DFSanVisitor::visitLoadInst(LoadInst &LI) {
2385 auto &DL = LI.getDataLayout();
2386 uint64_t Size = DL.getTypeStoreSize(Ty: LI.getType());
2387 if (Size == 0) {
2388 DFSF.setShadow(I: &LI, Shadow: DFSF.DFS.getZeroShadow(V: &LI));
2389 DFSF.setOrigin(I: &LI, Origin: DFSF.DFS.ZeroOrigin);
2390 return;
2391 }
2392
2393 // When an application load is atomic, increase atomic ordering between
2394 // atomic application loads and stores to ensure happen-before order; load
2395 // shadow data after application data; store zero shadow data before
2396 // application data. This ensure shadow loads return either labels of the
2397 // initial application data or zeros.
2398 if (LI.isAtomic())
2399 LI.setOrdering(addAcquireOrdering(AO: LI.getOrdering()));
2400
2401 BasicBlock::iterator AfterLi = std::next(x: LI.getIterator());
2402 BasicBlock::iterator Pos = LI.getIterator();
2403 if (LI.isAtomic())
2404 Pos = std::next(x: Pos);
2405
2406 std::vector<Value *> Shadows;
2407 std::vector<Value *> Origins;
2408 Value *PrimitiveShadow, *Origin;
2409 std::tie(args&: PrimitiveShadow, args&: Origin) =
2410 DFSF.loadShadowOrigin(Addr: LI.getPointerOperand(), Size, InstAlignment: LI.getAlign(), Pos);
2411 const bool ShouldTrackOrigins = DFSF.DFS.shouldTrackOrigins();
2412 if (ShouldTrackOrigins) {
2413 Shadows.push_back(x: PrimitiveShadow);
2414 Origins.push_back(x: Origin);
2415 }
2416 if (ClCombinePointerLabelsOnLoad ||
2417 DFSF.isLookupTableConstant(
2418 P: StripPointerGEPsAndCasts(V: LI.getPointerOperand()))) {
2419 Value *PtrShadow = DFSF.getShadow(V: LI.getPointerOperand());
2420 PrimitiveShadow = DFSF.combineShadows(V1: PrimitiveShadow, V2: PtrShadow, Pos);
2421 if (ShouldTrackOrigins) {
2422 Shadows.push_back(x: PtrShadow);
2423 Origins.push_back(x: DFSF.getOrigin(V: LI.getPointerOperand()));
2424 }
2425 }
2426 if (!DFSF.DFS.isZeroShadow(V: PrimitiveShadow))
2427 DFSF.NonZeroChecks.push_back(x: PrimitiveShadow);
2428
2429 Value *Shadow =
2430 DFSF.expandFromPrimitiveShadow(T: LI.getType(), PrimitiveShadow, Pos);
2431 DFSF.setShadow(I: &LI, Shadow);
2432
2433 if (ShouldTrackOrigins) {
2434 DFSF.setOrigin(I: &LI, Origin: DFSF.combineOrigins(Shadows, Origins, Pos));
2435 }
2436
2437 if (ClEventCallbacks) {
2438 IRBuilder<> IRB(Pos->getParent(), Pos);
2439 Value *Addr = LI.getPointerOperand();
2440 CallInst *CI =
2441 IRB.CreateCall(Callee: DFSF.DFS.DFSanLoadCallbackFn, Args: {PrimitiveShadow, Addr});
2442 CI->addParamAttr(ArgNo: 0, Kind: Attribute::ZExt);
2443 }
2444
2445 IRBuilder<> IRB(AfterLi->getParent(), AfterLi);
2446 DFSF.addReachesFunctionCallbacksIfEnabled(IRB, I&: LI, Data: &LI);
2447}
2448
2449Value *DFSanFunction::updateOriginIfTainted(Value *Shadow, Value *Origin,
2450 IRBuilder<> &IRB) {
2451 assert(DFS.shouldTrackOrigins());
2452 return IRB.CreateCall(Callee: DFS.DFSanChainOriginIfTaintedFn, Args: {Shadow, Origin});
2453}
2454
2455Value *DFSanFunction::updateOrigin(Value *V, IRBuilder<> &IRB) {
2456 if (!DFS.shouldTrackOrigins())
2457 return V;
2458 return IRB.CreateCall(Callee: DFS.DFSanChainOriginFn, Args: V);
2459}
2460
2461Value *DFSanFunction::originToIntptr(IRBuilder<> &IRB, Value *Origin) {
2462 const unsigned OriginSize = DataFlowSanitizer::OriginWidthBytes;
2463 const DataLayout &DL = F->getDataLayout();
2464 unsigned IntptrSize = DL.getTypeStoreSize(Ty: DFS.IntptrTy);
2465 if (IntptrSize == OriginSize)
2466 return Origin;
2467 assert(IntptrSize == OriginSize * 2);
2468 Origin = IRB.CreateIntCast(V: Origin, DestTy: DFS.IntptrTy, /* isSigned */ false);
2469 return IRB.CreateOr(LHS: Origin, RHS: IRB.CreateShl(LHS: Origin, RHS: OriginSize * 8));
2470}
2471
2472void DFSanFunction::paintOrigin(IRBuilder<> &IRB, Value *Origin,
2473 Value *StoreOriginAddr,
2474 uint64_t StoreOriginSize, Align Alignment) {
2475 const unsigned OriginSize = DataFlowSanitizer::OriginWidthBytes;
2476 const DataLayout &DL = F->getDataLayout();
2477 const Align IntptrAlignment = DL.getABITypeAlign(Ty: DFS.IntptrTy);
2478 unsigned IntptrSize = DL.getTypeStoreSize(Ty: DFS.IntptrTy);
2479 assert(IntptrAlignment >= MinOriginAlignment);
2480 assert(IntptrSize >= OriginSize);
2481
2482 unsigned Ofs = 0;
2483 Align CurrentAlignment = Alignment;
2484 if (Alignment >= IntptrAlignment && IntptrSize > OriginSize) {
2485 Value *IntptrOrigin = originToIntptr(IRB, Origin);
2486 Value *IntptrStoreOriginPtr =
2487 IRB.CreatePointerCast(V: StoreOriginAddr, DestTy: PointerType::get(C&: *DFS.Ctx, AddressSpace: 0));
2488 for (unsigned I = 0; I < StoreOriginSize / IntptrSize; ++I) {
2489 Value *Ptr =
2490 I ? IRB.CreateConstGEP1_32(Ty: DFS.IntptrTy, Ptr: IntptrStoreOriginPtr, Idx0: I)
2491 : IntptrStoreOriginPtr;
2492 IRB.CreateAlignedStore(Val: IntptrOrigin, Ptr, Align: CurrentAlignment);
2493 Ofs += IntptrSize / OriginSize;
2494 CurrentAlignment = IntptrAlignment;
2495 }
2496 }
2497
2498 for (unsigned I = Ofs; I < (StoreOriginSize + OriginSize - 1) / OriginSize;
2499 ++I) {
2500 Value *GEP = I ? IRB.CreateConstGEP1_32(Ty: DFS.OriginTy, Ptr: StoreOriginAddr, Idx0: I)
2501 : StoreOriginAddr;
2502 IRB.CreateAlignedStore(Val: Origin, Ptr: GEP, Align: CurrentAlignment);
2503 CurrentAlignment = MinOriginAlignment;
2504 }
2505}
2506
2507Value *DFSanFunction::convertToBool(Value *V, IRBuilder<> &IRB,
2508 const Twine &Name) {
2509 Type *VTy = V->getType();
2510 assert(VTy->isIntegerTy());
2511 if (VTy->getIntegerBitWidth() == 1)
2512 // Just converting a bool to a bool, so do nothing.
2513 return V;
2514 return IRB.CreateICmpNE(LHS: V, RHS: ConstantInt::get(Ty: VTy, V: 0), Name);
2515}
2516
2517void DFSanFunction::storeOrigin(BasicBlock::iterator Pos, Value *Addr,
2518 uint64_t Size, Value *Shadow, Value *Origin,
2519 Value *StoreOriginAddr, Align InstAlignment) {
2520 // Do not write origins for zero shadows because we do not trace origins for
2521 // untainted sinks.
2522 const Align OriginAlignment = getOriginAlign(InstAlignment);
2523 Value *CollapsedShadow = collapseToPrimitiveShadow(Shadow, Pos);
2524 IRBuilder<> IRB(Pos->getParent(), Pos);
2525 if (auto *ConstantShadow = dyn_cast<Constant>(Val: CollapsedShadow)) {
2526 if (!ConstantShadow->isZeroValue())
2527 paintOrigin(IRB, Origin: updateOrigin(V: Origin, IRB), StoreOriginAddr, StoreOriginSize: Size,
2528 Alignment: OriginAlignment);
2529 return;
2530 }
2531
2532 if (shouldInstrumentWithCall()) {
2533 IRB.CreateCall(
2534 Callee: DFS.DFSanMaybeStoreOriginFn,
2535 Args: {CollapsedShadow, Addr, ConstantInt::get(Ty: DFS.IntptrTy, V: Size), Origin});
2536 } else {
2537 Value *Cmp = convertToBool(V: CollapsedShadow, IRB, Name: "_dfscmp");
2538 DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Lazy);
2539 Instruction *CheckTerm = SplitBlockAndInsertIfThen(
2540 Cond: Cmp, SplitBefore: &*IRB.GetInsertPoint(), Unreachable: false, BranchWeights: DFS.OriginStoreWeights, DTU: &DTU);
2541 IRBuilder<> IRBNew(CheckTerm);
2542 paintOrigin(IRB&: IRBNew, Origin: updateOrigin(V: Origin, IRB&: IRBNew), StoreOriginAddr, StoreOriginSize: Size,
2543 Alignment: OriginAlignment);
2544 ++NumOriginStores;
2545 }
2546}
2547
2548void DFSanFunction::storeZeroPrimitiveShadow(Value *Addr, uint64_t Size,
2549 Align ShadowAlign,
2550 BasicBlock::iterator Pos) {
2551 IRBuilder<> IRB(Pos->getParent(), Pos);
2552 IntegerType *ShadowTy =
2553 IntegerType::get(C&: *DFS.Ctx, NumBits: Size * DFS.ShadowWidthBits);
2554 Value *ExtZeroShadow = ConstantInt::get(Ty: ShadowTy, V: 0);
2555 Value *ShadowAddr = DFS.getShadowAddress(Addr, Pos);
2556 IRB.CreateAlignedStore(Val: ExtZeroShadow, Ptr: ShadowAddr, Align: ShadowAlign);
2557 // Do not write origins for 0 shadows because we do not trace origins for
2558 // untainted sinks.
2559}
2560
2561void DFSanFunction::storePrimitiveShadowOrigin(Value *Addr, uint64_t Size,
2562 Align InstAlignment,
2563 Value *PrimitiveShadow,
2564 Value *Origin,
2565 BasicBlock::iterator Pos) {
2566 const bool ShouldTrackOrigins = DFS.shouldTrackOrigins() && Origin;
2567
2568 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val: Addr)) {
2569 const auto SI = AllocaShadowMap.find(Val: AI);
2570 if (SI != AllocaShadowMap.end()) {
2571 IRBuilder<> IRB(Pos->getParent(), Pos);
2572 IRB.CreateStore(Val: PrimitiveShadow, Ptr: SI->second);
2573
2574 // Do not write origins for 0 shadows because we do not trace origins for
2575 // untainted sinks.
2576 if (ShouldTrackOrigins && !DFS.isZeroShadow(V: PrimitiveShadow)) {
2577 const auto OI = AllocaOriginMap.find(Val: AI);
2578 assert(OI != AllocaOriginMap.end() && Origin);
2579 IRB.CreateStore(Val: Origin, Ptr: OI->second);
2580 }
2581 return;
2582 }
2583 }
2584
2585 const Align ShadowAlign = getShadowAlign(InstAlignment);
2586 if (DFS.isZeroShadow(V: PrimitiveShadow)) {
2587 storeZeroPrimitiveShadow(Addr, Size, ShadowAlign, Pos);
2588 return;
2589 }
2590
2591 IRBuilder<> IRB(Pos->getParent(), Pos);
2592 Value *ShadowAddr, *OriginAddr;
2593 std::tie(args&: ShadowAddr, args&: OriginAddr) =
2594 DFS.getShadowOriginAddress(Addr, InstAlignment, Pos);
2595
2596 const unsigned ShadowVecSize = 8;
2597 assert(ShadowVecSize * DFS.ShadowWidthBits <= 128 &&
2598 "Shadow vector is too large!");
2599
2600 uint64_t Offset = 0;
2601 uint64_t LeftSize = Size;
2602 if (LeftSize >= ShadowVecSize) {
2603 auto *ShadowVecTy =
2604 FixedVectorType::get(ElementType: DFS.PrimitiveShadowTy, NumElts: ShadowVecSize);
2605 Value *ShadowVec = PoisonValue::get(T: ShadowVecTy);
2606 for (unsigned I = 0; I != ShadowVecSize; ++I) {
2607 ShadowVec = IRB.CreateInsertElement(
2608 Vec: ShadowVec, NewElt: PrimitiveShadow,
2609 Idx: ConstantInt::get(Ty: Type::getInt32Ty(C&: *DFS.Ctx), V: I));
2610 }
2611 do {
2612 Value *CurShadowVecAddr =
2613 IRB.CreateConstGEP1_32(Ty: ShadowVecTy, Ptr: ShadowAddr, Idx0: Offset);
2614 IRB.CreateAlignedStore(Val: ShadowVec, Ptr: CurShadowVecAddr, Align: ShadowAlign);
2615 LeftSize -= ShadowVecSize;
2616 ++Offset;
2617 } while (LeftSize >= ShadowVecSize);
2618 Offset *= ShadowVecSize;
2619 }
2620 while (LeftSize > 0) {
2621 Value *CurShadowAddr =
2622 IRB.CreateConstGEP1_32(Ty: DFS.PrimitiveShadowTy, Ptr: ShadowAddr, Idx0: Offset);
2623 IRB.CreateAlignedStore(Val: PrimitiveShadow, Ptr: CurShadowAddr, Align: ShadowAlign);
2624 --LeftSize;
2625 ++Offset;
2626 }
2627
2628 if (ShouldTrackOrigins) {
2629 storeOrigin(Pos, Addr, Size, Shadow: PrimitiveShadow, Origin, StoreOriginAddr: OriginAddr,
2630 InstAlignment);
2631 }
2632}
2633
2634static AtomicOrdering addReleaseOrdering(AtomicOrdering AO) {
2635 switch (AO) {
2636 case AtomicOrdering::NotAtomic:
2637 return AtomicOrdering::NotAtomic;
2638 case AtomicOrdering::Unordered:
2639 case AtomicOrdering::Monotonic:
2640 case AtomicOrdering::Release:
2641 return AtomicOrdering::Release;
2642 case AtomicOrdering::Acquire:
2643 case AtomicOrdering::AcquireRelease:
2644 return AtomicOrdering::AcquireRelease;
2645 case AtomicOrdering::SequentiallyConsistent:
2646 return AtomicOrdering::SequentiallyConsistent;
2647 }
2648 llvm_unreachable("Unknown ordering");
2649}
2650
2651void DFSanVisitor::visitStoreInst(StoreInst &SI) {
2652 auto &DL = SI.getDataLayout();
2653 Value *Val = SI.getValueOperand();
2654 uint64_t Size = DL.getTypeStoreSize(Ty: Val->getType());
2655 if (Size == 0)
2656 return;
2657
2658 // When an application store is atomic, increase atomic ordering between
2659 // atomic application loads and stores to ensure happen-before order; load
2660 // shadow data after application data; store zero shadow data before
2661 // application data. This ensure shadow loads return either labels of the
2662 // initial application data or zeros.
2663 if (SI.isAtomic())
2664 SI.setOrdering(addReleaseOrdering(AO: SI.getOrdering()));
2665
2666 const bool ShouldTrackOrigins =
2667 DFSF.DFS.shouldTrackOrigins() && !SI.isAtomic();
2668 std::vector<Value *> Shadows;
2669 std::vector<Value *> Origins;
2670
2671 Value *Shadow =
2672 SI.isAtomic() ? DFSF.DFS.getZeroShadow(V: Val) : DFSF.getShadow(V: Val);
2673
2674 if (ShouldTrackOrigins) {
2675 Shadows.push_back(x: Shadow);
2676 Origins.push_back(x: DFSF.getOrigin(V: Val));
2677 }
2678
2679 Value *PrimitiveShadow;
2680 if (ClCombinePointerLabelsOnStore) {
2681 Value *PtrShadow = DFSF.getShadow(V: SI.getPointerOperand());
2682 if (ShouldTrackOrigins) {
2683 Shadows.push_back(x: PtrShadow);
2684 Origins.push_back(x: DFSF.getOrigin(V: SI.getPointerOperand()));
2685 }
2686 PrimitiveShadow = DFSF.combineShadows(V1: Shadow, V2: PtrShadow, Pos: SI.getIterator());
2687 } else {
2688 PrimitiveShadow = DFSF.collapseToPrimitiveShadow(Shadow, Pos: SI.getIterator());
2689 }
2690 Value *Origin = nullptr;
2691 if (ShouldTrackOrigins)
2692 Origin = DFSF.combineOrigins(Shadows, Origins, Pos: SI.getIterator());
2693 DFSF.storePrimitiveShadowOrigin(Addr: SI.getPointerOperand(), Size, InstAlignment: SI.getAlign(),
2694 PrimitiveShadow, Origin, Pos: SI.getIterator());
2695 if (ClEventCallbacks) {
2696 IRBuilder<> IRB(&SI);
2697 Value *Addr = SI.getPointerOperand();
2698 CallInst *CI =
2699 IRB.CreateCall(Callee: DFSF.DFS.DFSanStoreCallbackFn, Args: {PrimitiveShadow, Addr});
2700 CI->addParamAttr(ArgNo: 0, Kind: Attribute::ZExt);
2701 }
2702}
2703
2704void DFSanVisitor::visitCASOrRMW(Align InstAlignment, Instruction &I) {
2705 assert(isa<AtomicRMWInst>(I) || isa<AtomicCmpXchgInst>(I));
2706
2707 Value *Val = I.getOperand(i: 1);
2708 const auto &DL = I.getDataLayout();
2709 uint64_t Size = DL.getTypeStoreSize(Ty: Val->getType());
2710 if (Size == 0)
2711 return;
2712
2713 // Conservatively set data at stored addresses and return with zero shadow to
2714 // prevent shadow data races.
2715 IRBuilder<> IRB(&I);
2716 Value *Addr = I.getOperand(i: 0);
2717 const Align ShadowAlign = DFSF.getShadowAlign(InstAlignment);
2718 DFSF.storeZeroPrimitiveShadow(Addr, Size, ShadowAlign, Pos: I.getIterator());
2719 DFSF.setShadow(I: &I, Shadow: DFSF.DFS.getZeroShadow(V: &I));
2720 DFSF.setOrigin(I: &I, Origin: DFSF.DFS.ZeroOrigin);
2721}
2722
2723void DFSanVisitor::visitAtomicRMWInst(AtomicRMWInst &I) {
2724 visitCASOrRMW(InstAlignment: I.getAlign(), I);
2725 // TODO: The ordering change follows MSan. It is possible not to change
2726 // ordering because we always set and use 0 shadows.
2727 I.setOrdering(addReleaseOrdering(AO: I.getOrdering()));
2728}
2729
2730void DFSanVisitor::visitAtomicCmpXchgInst(AtomicCmpXchgInst &I) {
2731 visitCASOrRMW(InstAlignment: I.getAlign(), I);
2732 // TODO: The ordering change follows MSan. It is possible not to change
2733 // ordering because we always set and use 0 shadows.
2734 I.setSuccessOrdering(addReleaseOrdering(AO: I.getSuccessOrdering()));
2735}
2736
2737void DFSanVisitor::visitUnaryOperator(UnaryOperator &UO) {
2738 visitInstOperands(I&: UO);
2739}
2740
2741void DFSanVisitor::visitBinaryOperator(BinaryOperator &BO) {
2742 visitInstOperands(I&: BO);
2743}
2744
2745void DFSanVisitor::visitBitCastInst(BitCastInst &BCI) {
2746 // Special case: if this is the bitcast (there is exactly 1 allowed) between
2747 // a musttail call and a ret, don't instrument. New instructions are not
2748 // allowed after a musttail call.
2749 if (auto *CI = dyn_cast<CallInst>(Val: BCI.getOperand(i_nocapture: 0)))
2750 if (CI->isMustTailCall())
2751 return;
2752 visitInstOperands(I&: BCI);
2753}
2754
2755void DFSanVisitor::visitCastInst(CastInst &CI) { visitInstOperands(I&: CI); }
2756
2757void DFSanVisitor::visitCmpInst(CmpInst &CI) {
2758 visitInstOperands(I&: CI);
2759 if (ClEventCallbacks) {
2760 IRBuilder<> IRB(&CI);
2761 Value *CombinedShadow = DFSF.getShadow(V: &CI);
2762 CallInst *CallI =
2763 IRB.CreateCall(Callee: DFSF.DFS.DFSanCmpCallbackFn, Args: CombinedShadow);
2764 CallI->addParamAttr(ArgNo: 0, Kind: Attribute::ZExt);
2765 }
2766}
2767
2768void DFSanVisitor::visitLandingPadInst(LandingPadInst &LPI) {
2769 // We do not need to track data through LandingPadInst.
2770 //
2771 // For the C++ exceptions, if a value is thrown, this value will be stored
2772 // in a memory location provided by __cxa_allocate_exception(...) (on the
2773 // throw side) or __cxa_begin_catch(...) (on the catch side).
2774 // This memory will have a shadow, so with the loads and stores we will be
2775 // able to propagate labels on data thrown through exceptions, without any
2776 // special handling of the LandingPadInst.
2777 //
2778 // The second element in the pair result of the LandingPadInst is a
2779 // register value, but it is for a type ID and should never be tainted.
2780 DFSF.setShadow(I: &LPI, Shadow: DFSF.DFS.getZeroShadow(V: &LPI));
2781 DFSF.setOrigin(I: &LPI, Origin: DFSF.DFS.ZeroOrigin);
2782}
2783
2784void DFSanVisitor::visitGetElementPtrInst(GetElementPtrInst &GEPI) {
2785 if (ClCombineOffsetLabelsOnGEP ||
2786 DFSF.isLookupTableConstant(
2787 P: StripPointerGEPsAndCasts(V: GEPI.getPointerOperand()))) {
2788 visitInstOperands(I&: GEPI);
2789 return;
2790 }
2791
2792 // Only propagate shadow/origin of base pointer value but ignore those of
2793 // offset operands.
2794 Value *BasePointer = GEPI.getPointerOperand();
2795 DFSF.setShadow(I: &GEPI, Shadow: DFSF.getShadow(V: BasePointer));
2796 if (DFSF.DFS.shouldTrackOrigins())
2797 DFSF.setOrigin(I: &GEPI, Origin: DFSF.getOrigin(V: BasePointer));
2798}
2799
2800void DFSanVisitor::visitExtractElementInst(ExtractElementInst &I) {
2801 visitInstOperands(I);
2802}
2803
2804void DFSanVisitor::visitInsertElementInst(InsertElementInst &I) {
2805 visitInstOperands(I);
2806}
2807
2808void DFSanVisitor::visitShuffleVectorInst(ShuffleVectorInst &I) {
2809 visitInstOperands(I);
2810}
2811
2812void DFSanVisitor::visitExtractValueInst(ExtractValueInst &I) {
2813 IRBuilder<> IRB(&I);
2814 Value *Agg = I.getAggregateOperand();
2815 Value *AggShadow = DFSF.getShadow(V: Agg);
2816 Value *ResShadow = IRB.CreateExtractValue(Agg: AggShadow, Idxs: I.getIndices());
2817 DFSF.setShadow(I: &I, Shadow: ResShadow);
2818 visitInstOperandOrigins(I);
2819}
2820
2821void DFSanVisitor::visitInsertValueInst(InsertValueInst &I) {
2822 IRBuilder<> IRB(&I);
2823 Value *AggShadow = DFSF.getShadow(V: I.getAggregateOperand());
2824 Value *InsShadow = DFSF.getShadow(V: I.getInsertedValueOperand());
2825 Value *Res = IRB.CreateInsertValue(Agg: AggShadow, Val: InsShadow, Idxs: I.getIndices());
2826 DFSF.setShadow(I: &I, Shadow: Res);
2827 visitInstOperandOrigins(I);
2828}
2829
2830void DFSanVisitor::visitAllocaInst(AllocaInst &I) {
2831 bool AllLoadsStores = true;
2832 for (User *U : I.users()) {
2833 if (isa<LoadInst>(Val: U))
2834 continue;
2835
2836 if (StoreInst *SI = dyn_cast<StoreInst>(Val: U)) {
2837 if (SI->getPointerOperand() == &I)
2838 continue;
2839 }
2840
2841 AllLoadsStores = false;
2842 break;
2843 }
2844 if (AllLoadsStores) {
2845 IRBuilder<> IRB(&I);
2846 DFSF.AllocaShadowMap[&I] = IRB.CreateAlloca(Ty: DFSF.DFS.PrimitiveShadowTy);
2847 if (DFSF.DFS.shouldTrackOrigins()) {
2848 DFSF.AllocaOriginMap[&I] =
2849 IRB.CreateAlloca(Ty: DFSF.DFS.OriginTy, ArraySize: nullptr, Name: "_dfsa");
2850 }
2851 }
2852 DFSF.setShadow(I: &I, Shadow: DFSF.DFS.ZeroPrimitiveShadow);
2853 DFSF.setOrigin(I: &I, Origin: DFSF.DFS.ZeroOrigin);
2854}
2855
2856void DFSanVisitor::visitSelectInst(SelectInst &I) {
2857 Value *CondShadow = DFSF.getShadow(V: I.getCondition());
2858 Value *TrueShadow = DFSF.getShadow(V: I.getTrueValue());
2859 Value *FalseShadow = DFSF.getShadow(V: I.getFalseValue());
2860 Value *ShadowSel = nullptr;
2861 const bool ShouldTrackOrigins = DFSF.DFS.shouldTrackOrigins();
2862 std::vector<Value *> Shadows;
2863 std::vector<Value *> Origins;
2864 Value *TrueOrigin =
2865 ShouldTrackOrigins ? DFSF.getOrigin(V: I.getTrueValue()) : nullptr;
2866 Value *FalseOrigin =
2867 ShouldTrackOrigins ? DFSF.getOrigin(V: I.getFalseValue()) : nullptr;
2868
2869 DFSF.addConditionalCallbacksIfEnabled(I, Condition: I.getCondition());
2870
2871 if (isa<VectorType>(Val: I.getCondition()->getType())) {
2872 ShadowSel = DFSF.combineShadowsThenConvert(T: I.getType(), V1: TrueShadow,
2873 V2: FalseShadow, Pos: I.getIterator());
2874 if (ShouldTrackOrigins) {
2875 Shadows.push_back(x: TrueShadow);
2876 Shadows.push_back(x: FalseShadow);
2877 Origins.push_back(x: TrueOrigin);
2878 Origins.push_back(x: FalseOrigin);
2879 }
2880 } else {
2881 if (TrueShadow == FalseShadow) {
2882 ShadowSel = TrueShadow;
2883 if (ShouldTrackOrigins) {
2884 Shadows.push_back(x: TrueShadow);
2885 Origins.push_back(x: TrueOrigin);
2886 }
2887 } else {
2888 ShadowSel = SelectInst::Create(C: I.getCondition(), S1: TrueShadow, S2: FalseShadow,
2889 NameStr: "", InsertBefore: I.getIterator());
2890 if (ShouldTrackOrigins) {
2891 Shadows.push_back(x: ShadowSel);
2892 Origins.push_back(x: SelectInst::Create(C: I.getCondition(), S1: TrueOrigin,
2893 S2: FalseOrigin, NameStr: "", InsertBefore: I.getIterator()));
2894 }
2895 }
2896 }
2897 DFSF.setShadow(I: &I, Shadow: ClTrackSelectControlFlow ? DFSF.combineShadowsThenConvert(
2898 T: I.getType(), V1: CondShadow,
2899 V2: ShadowSel, Pos: I.getIterator())
2900 : ShadowSel);
2901 if (ShouldTrackOrigins) {
2902 if (ClTrackSelectControlFlow) {
2903 Shadows.push_back(x: CondShadow);
2904 Origins.push_back(x: DFSF.getOrigin(V: I.getCondition()));
2905 }
2906 DFSF.setOrigin(I: &I, Origin: DFSF.combineOrigins(Shadows, Origins, Pos: I.getIterator()));
2907 }
2908}
2909
2910void DFSanVisitor::visitMemSetInst(MemSetInst &I) {
2911 IRBuilder<> IRB(&I);
2912 Value *ValShadow = DFSF.getShadow(V: I.getValue());
2913 Value *ValOrigin = DFSF.DFS.shouldTrackOrigins()
2914 ? DFSF.getOrigin(V: I.getValue())
2915 : DFSF.DFS.ZeroOrigin;
2916 IRB.CreateCall(Callee: DFSF.DFS.DFSanSetLabelFn,
2917 Args: {ValShadow, ValOrigin, I.getDest(),
2918 IRB.CreateZExtOrTrunc(V: I.getLength(), DestTy: DFSF.DFS.IntptrTy)});
2919}
2920
2921void DFSanVisitor::visitMemTransferInst(MemTransferInst &I) {
2922 IRBuilder<> IRB(&I);
2923
2924 // CopyOrMoveOrigin transfers origins by refering to their shadows. So we
2925 // need to move origins before moving shadows.
2926 if (DFSF.DFS.shouldTrackOrigins()) {
2927 IRB.CreateCall(
2928 Callee: DFSF.DFS.DFSanMemOriginTransferFn,
2929 Args: {I.getArgOperand(i: 0), I.getArgOperand(i: 1),
2930 IRB.CreateIntCast(V: I.getArgOperand(i: 2), DestTy: DFSF.DFS.IntptrTy, isSigned: false)});
2931 }
2932
2933 Value *DestShadow = DFSF.DFS.getShadowAddress(Addr: I.getDest(), Pos: I.getIterator());
2934 Value *SrcShadow = DFSF.DFS.getShadowAddress(Addr: I.getSource(), Pos: I.getIterator());
2935 Value *LenShadow =
2936 IRB.CreateMul(LHS: I.getLength(), RHS: ConstantInt::get(Ty: I.getLength()->getType(),
2937 V: DFSF.DFS.ShadowWidthBytes));
2938 auto *MTI = cast<MemTransferInst>(
2939 Val: IRB.CreateCall(FTy: I.getFunctionType(), Callee: I.getCalledOperand(),
2940 Args: {DestShadow, SrcShadow, LenShadow, I.getVolatileCst()}));
2941 MTI->setDestAlignment(DFSF.getShadowAlign(InstAlignment: I.getDestAlign().valueOrOne()));
2942 MTI->setSourceAlignment(DFSF.getShadowAlign(InstAlignment: I.getSourceAlign().valueOrOne()));
2943 if (ClEventCallbacks) {
2944 IRB.CreateCall(
2945 Callee: DFSF.DFS.DFSanMemTransferCallbackFn,
2946 Args: {DestShadow, IRB.CreateZExtOrTrunc(V: I.getLength(), DestTy: DFSF.DFS.IntptrTy)});
2947 }
2948}
2949
2950void DFSanVisitor::visitBranchInst(BranchInst &BR) {
2951 if (!BR.isConditional())
2952 return;
2953
2954 DFSF.addConditionalCallbacksIfEnabled(I&: BR, Condition: BR.getCondition());
2955}
2956
2957void DFSanVisitor::visitSwitchInst(SwitchInst &SW) {
2958 DFSF.addConditionalCallbacksIfEnabled(I&: SW, Condition: SW.getCondition());
2959}
2960
2961static bool isAMustTailRetVal(Value *RetVal) {
2962 // Tail call may have a bitcast between return.
2963 if (auto *I = dyn_cast<BitCastInst>(Val: RetVal)) {
2964 RetVal = I->getOperand(i_nocapture: 0);
2965 }
2966 if (auto *I = dyn_cast<CallInst>(Val: RetVal)) {
2967 return I->isMustTailCall();
2968 }
2969 return false;
2970}
2971
2972void DFSanVisitor::visitReturnInst(ReturnInst &RI) {
2973 if (!DFSF.IsNativeABI && RI.getReturnValue()) {
2974 // Don't emit the instrumentation for musttail call returns.
2975 if (isAMustTailRetVal(RetVal: RI.getReturnValue()))
2976 return;
2977
2978 Value *S = DFSF.getShadow(V: RI.getReturnValue());
2979 IRBuilder<> IRB(&RI);
2980 Type *RT = DFSF.F->getFunctionType()->getReturnType();
2981 unsigned Size = getDataLayout().getTypeAllocSize(Ty: DFSF.DFS.getShadowTy(OrigTy: RT));
2982 if (Size <= RetvalTLSSize) {
2983 // If the size overflows, stores nothing. At callsite, oversized return
2984 // shadows are set to zero.
2985 IRB.CreateAlignedStore(Val: S, Ptr: DFSF.getRetvalTLS(T: RT, IRB), Align: ShadowTLSAlignment);
2986 }
2987 if (DFSF.DFS.shouldTrackOrigins()) {
2988 Value *O = DFSF.getOrigin(V: RI.getReturnValue());
2989 IRB.CreateStore(Val: O, Ptr: DFSF.getRetvalOriginTLS());
2990 }
2991 }
2992}
2993
2994void DFSanVisitor::addShadowArguments(Function &F, CallBase &CB,
2995 std::vector<Value *> &Args,
2996 IRBuilder<> &IRB) {
2997 FunctionType *FT = F.getFunctionType();
2998
2999 auto *I = CB.arg_begin();
3000
3001 // Adds non-variable argument shadows.
3002 for (unsigned N = FT->getNumParams(); N != 0; ++I, --N)
3003 Args.push_back(
3004 x: DFSF.collapseToPrimitiveShadow(Shadow: DFSF.getShadow(V: *I), Pos: CB.getIterator()));
3005
3006 // Adds variable argument shadows.
3007 if (FT->isVarArg()) {
3008 auto *LabelVATy = ArrayType::get(ElementType: DFSF.DFS.PrimitiveShadowTy,
3009 NumElements: CB.arg_size() - FT->getNumParams());
3010 auto *LabelVAAlloca =
3011 new AllocaInst(LabelVATy, getDataLayout().getAllocaAddrSpace(),
3012 "labelva", DFSF.F->getEntryBlock().begin());
3013
3014 for (unsigned N = 0; I != CB.arg_end(); ++I, ++N) {
3015 auto *LabelVAPtr = IRB.CreateStructGEP(Ty: LabelVATy, Ptr: LabelVAAlloca, Idx: N);
3016 IRB.CreateStore(
3017 Val: DFSF.collapseToPrimitiveShadow(Shadow: DFSF.getShadow(V: *I), Pos: CB.getIterator()),
3018 Ptr: LabelVAPtr);
3019 }
3020
3021 Args.push_back(x: IRB.CreateStructGEP(Ty: LabelVATy, Ptr: LabelVAAlloca, Idx: 0));
3022 }
3023
3024 // Adds the return value shadow.
3025 if (!FT->getReturnType()->isVoidTy()) {
3026 if (!DFSF.LabelReturnAlloca) {
3027 DFSF.LabelReturnAlloca = new AllocaInst(
3028 DFSF.DFS.PrimitiveShadowTy, getDataLayout().getAllocaAddrSpace(),
3029 "labelreturn", DFSF.F->getEntryBlock().begin());
3030 }
3031 Args.push_back(x: DFSF.LabelReturnAlloca);
3032 }
3033}
3034
3035void DFSanVisitor::addOriginArguments(Function &F, CallBase &CB,
3036 std::vector<Value *> &Args,
3037 IRBuilder<> &IRB) {
3038 FunctionType *FT = F.getFunctionType();
3039
3040 auto *I = CB.arg_begin();
3041
3042 // Add non-variable argument origins.
3043 for (unsigned N = FT->getNumParams(); N != 0; ++I, --N)
3044 Args.push_back(x: DFSF.getOrigin(V: *I));
3045
3046 // Add variable argument origins.
3047 if (FT->isVarArg()) {
3048 auto *OriginVATy =
3049 ArrayType::get(ElementType: DFSF.DFS.OriginTy, NumElements: CB.arg_size() - FT->getNumParams());
3050 auto *OriginVAAlloca =
3051 new AllocaInst(OriginVATy, getDataLayout().getAllocaAddrSpace(),
3052 "originva", DFSF.F->getEntryBlock().begin());
3053
3054 for (unsigned N = 0; I != CB.arg_end(); ++I, ++N) {
3055 auto *OriginVAPtr = IRB.CreateStructGEP(Ty: OriginVATy, Ptr: OriginVAAlloca, Idx: N);
3056 IRB.CreateStore(Val: DFSF.getOrigin(V: *I), Ptr: OriginVAPtr);
3057 }
3058
3059 Args.push_back(x: IRB.CreateStructGEP(Ty: OriginVATy, Ptr: OriginVAAlloca, Idx: 0));
3060 }
3061
3062 // Add the return value origin.
3063 if (!FT->getReturnType()->isVoidTy()) {
3064 if (!DFSF.OriginReturnAlloca) {
3065 DFSF.OriginReturnAlloca = new AllocaInst(
3066 DFSF.DFS.OriginTy, getDataLayout().getAllocaAddrSpace(),
3067 "originreturn", DFSF.F->getEntryBlock().begin());
3068 }
3069 Args.push_back(x: DFSF.OriginReturnAlloca);
3070 }
3071}
3072
3073bool DFSanVisitor::visitWrappedCallBase(Function &F, CallBase &CB) {
3074 IRBuilder<> IRB(&CB);
3075 switch (DFSF.DFS.getWrapperKind(F: &F)) {
3076 case DataFlowSanitizer::WK_Warning:
3077 CB.setCalledFunction(&F);
3078 IRB.CreateCall(Callee: DFSF.DFS.DFSanUnimplementedFn,
3079 Args: IRB.CreateGlobalString(Str: F.getName()));
3080 DFSF.DFS.buildExternWeakCheckIfNeeded(IRB, F: &F);
3081 DFSF.setShadow(I: &CB, Shadow: DFSF.DFS.getZeroShadow(V: &CB));
3082 DFSF.setOrigin(I: &CB, Origin: DFSF.DFS.ZeroOrigin);
3083 return true;
3084 case DataFlowSanitizer::WK_Discard:
3085 CB.setCalledFunction(&F);
3086 DFSF.DFS.buildExternWeakCheckIfNeeded(IRB, F: &F);
3087 DFSF.setShadow(I: &CB, Shadow: DFSF.DFS.getZeroShadow(V: &CB));
3088 DFSF.setOrigin(I: &CB, Origin: DFSF.DFS.ZeroOrigin);
3089 return true;
3090 case DataFlowSanitizer::WK_Functional:
3091 CB.setCalledFunction(&F);
3092 DFSF.DFS.buildExternWeakCheckIfNeeded(IRB, F: &F);
3093 visitInstOperands(I&: CB);
3094 return true;
3095 case DataFlowSanitizer::WK_Custom:
3096 // Don't try to handle invokes of custom functions, it's too complicated.
3097 // Instead, invoke the dfsw$ wrapper, which will in turn call the __dfsw_
3098 // wrapper.
3099 CallInst *CI = dyn_cast<CallInst>(Val: &CB);
3100 if (!CI)
3101 return false;
3102
3103 const bool ShouldTrackOrigins = DFSF.DFS.shouldTrackOrigins();
3104 FunctionType *FT = F.getFunctionType();
3105 TransformedFunction CustomFn = DFSF.DFS.getCustomFunctionType(T: FT);
3106 std::string CustomFName = ShouldTrackOrigins ? "__dfso_" : "__dfsw_";
3107 CustomFName += F.getName();
3108 FunctionCallee CustomF = DFSF.DFS.Mod->getOrInsertFunction(
3109 Name: CustomFName, T: CustomFn.TransformedType);
3110 if (Function *CustomFn = dyn_cast<Function>(Val: CustomF.getCallee())) {
3111 CustomFn->copyAttributesFrom(Src: &F);
3112
3113 // Custom functions returning non-void will write to the return label.
3114 if (!FT->getReturnType()->isVoidTy()) {
3115 CustomFn->removeFnAttrs(Attrs: DFSF.DFS.ReadOnlyNoneAttrs);
3116 }
3117 }
3118
3119 std::vector<Value *> Args;
3120
3121 // Adds non-variable arguments.
3122 auto *I = CB.arg_begin();
3123 for (unsigned N = FT->getNumParams(); N != 0; ++I, --N) {
3124 Args.push_back(x: *I);
3125 }
3126
3127 // Adds shadow arguments.
3128 const unsigned ShadowArgStart = Args.size();
3129 addShadowArguments(F, CB, Args, IRB);
3130
3131 // Adds origin arguments.
3132 const unsigned OriginArgStart = Args.size();
3133 if (ShouldTrackOrigins)
3134 addOriginArguments(F, CB, Args, IRB);
3135
3136 // Adds variable arguments.
3137 append_range(C&: Args, R: drop_begin(RangeOrContainer: CB.args(), N: FT->getNumParams()));
3138
3139 CallInst *CustomCI = IRB.CreateCall(Callee: CustomF, Args);
3140 CustomCI->setCallingConv(CI->getCallingConv());
3141 CustomCI->setAttributes(transformFunctionAttributes(
3142 TransformedFunction: CustomFn, Ctx&: CI->getContext(), CallSiteAttrs: CI->getAttributes()));
3143
3144 // Update the parameter attributes of the custom call instruction to
3145 // zero extend the shadow parameters. This is required for targets
3146 // which consider PrimitiveShadowTy an illegal type.
3147 for (unsigned N = 0; N < FT->getNumParams(); N++) {
3148 const unsigned ArgNo = ShadowArgStart + N;
3149 if (CustomCI->getArgOperand(i: ArgNo)->getType() ==
3150 DFSF.DFS.PrimitiveShadowTy)
3151 CustomCI->addParamAttr(ArgNo, Kind: Attribute::ZExt);
3152 if (ShouldTrackOrigins) {
3153 const unsigned OriginArgNo = OriginArgStart + N;
3154 if (CustomCI->getArgOperand(i: OriginArgNo)->getType() ==
3155 DFSF.DFS.OriginTy)
3156 CustomCI->addParamAttr(ArgNo: OriginArgNo, Kind: Attribute::ZExt);
3157 }
3158 }
3159
3160 // Loads the return value shadow and origin.
3161 if (!FT->getReturnType()->isVoidTy()) {
3162 LoadInst *LabelLoad =
3163 IRB.CreateLoad(Ty: DFSF.DFS.PrimitiveShadowTy, Ptr: DFSF.LabelReturnAlloca);
3164 DFSF.setShadow(I: CustomCI,
3165 Shadow: DFSF.expandFromPrimitiveShadow(
3166 T: FT->getReturnType(), PrimitiveShadow: LabelLoad, Pos: CB.getIterator()));
3167 if (ShouldTrackOrigins) {
3168 LoadInst *OriginLoad =
3169 IRB.CreateLoad(Ty: DFSF.DFS.OriginTy, Ptr: DFSF.OriginReturnAlloca);
3170 DFSF.setOrigin(I: CustomCI, Origin: OriginLoad);
3171 }
3172 }
3173
3174 CI->replaceAllUsesWith(V: CustomCI);
3175 CI->eraseFromParent();
3176 return true;
3177 }
3178 return false;
3179}
3180
3181Value *DFSanVisitor::makeAddAcquireOrderingTable(IRBuilder<> &IRB) {
3182 constexpr int NumOrderings = (int)AtomicOrderingCABI::seq_cst + 1;
3183 uint32_t OrderingTable[NumOrderings] = {};
3184
3185 OrderingTable[(int)AtomicOrderingCABI::relaxed] =
3186 OrderingTable[(int)AtomicOrderingCABI::acquire] =
3187 OrderingTable[(int)AtomicOrderingCABI::consume] =
3188 (int)AtomicOrderingCABI::acquire;
3189 OrderingTable[(int)AtomicOrderingCABI::release] =
3190 OrderingTable[(int)AtomicOrderingCABI::acq_rel] =
3191 (int)AtomicOrderingCABI::acq_rel;
3192 OrderingTable[(int)AtomicOrderingCABI::seq_cst] =
3193 (int)AtomicOrderingCABI::seq_cst;
3194
3195 return ConstantDataVector::get(Context&: IRB.getContext(), Elts: OrderingTable);
3196}
3197
3198void DFSanVisitor::visitLibAtomicLoad(CallBase &CB) {
3199 // Since we use getNextNode here, we can't have CB terminate the BB.
3200 assert(isa<CallInst>(CB));
3201
3202 IRBuilder<> IRB(&CB);
3203 Value *Size = CB.getArgOperand(i: 0);
3204 Value *SrcPtr = CB.getArgOperand(i: 1);
3205 Value *DstPtr = CB.getArgOperand(i: 2);
3206 Value *Ordering = CB.getArgOperand(i: 3);
3207 // Convert the call to have at least Acquire ordering to make sure
3208 // the shadow operations aren't reordered before it.
3209 Value *NewOrdering =
3210 IRB.CreateExtractElement(Vec: makeAddAcquireOrderingTable(IRB), Idx: Ordering);
3211 CB.setArgOperand(i: 3, v: NewOrdering);
3212
3213 IRBuilder<> NextIRB(CB.getNextNode());
3214 NextIRB.SetCurrentDebugLocation(CB.getDebugLoc());
3215
3216 // TODO: Support ClCombinePointerLabelsOnLoad
3217 // TODO: Support ClEventCallbacks
3218
3219 NextIRB.CreateCall(
3220 Callee: DFSF.DFS.DFSanMemShadowOriginTransferFn,
3221 Args: {DstPtr, SrcPtr, NextIRB.CreateIntCast(V: Size, DestTy: DFSF.DFS.IntptrTy, isSigned: false)});
3222}
3223
3224Value *DFSanVisitor::makeAddReleaseOrderingTable(IRBuilder<> &IRB) {
3225 constexpr int NumOrderings = (int)AtomicOrderingCABI::seq_cst + 1;
3226 uint32_t OrderingTable[NumOrderings] = {};
3227
3228 OrderingTable[(int)AtomicOrderingCABI::relaxed] =
3229 OrderingTable[(int)AtomicOrderingCABI::release] =
3230 (int)AtomicOrderingCABI::release;
3231 OrderingTable[(int)AtomicOrderingCABI::consume] =
3232 OrderingTable[(int)AtomicOrderingCABI::acquire] =
3233 OrderingTable[(int)AtomicOrderingCABI::acq_rel] =
3234 (int)AtomicOrderingCABI::acq_rel;
3235 OrderingTable[(int)AtomicOrderingCABI::seq_cst] =
3236 (int)AtomicOrderingCABI::seq_cst;
3237
3238 return ConstantDataVector::get(Context&: IRB.getContext(), Elts: OrderingTable);
3239}
3240
3241void DFSanVisitor::visitLibAtomicStore(CallBase &CB) {
3242 IRBuilder<> IRB(&CB);
3243 Value *Size = CB.getArgOperand(i: 0);
3244 Value *SrcPtr = CB.getArgOperand(i: 1);
3245 Value *DstPtr = CB.getArgOperand(i: 2);
3246 Value *Ordering = CB.getArgOperand(i: 3);
3247 // Convert the call to have at least Release ordering to make sure
3248 // the shadow operations aren't reordered after it.
3249 Value *NewOrdering =
3250 IRB.CreateExtractElement(Vec: makeAddReleaseOrderingTable(IRB), Idx: Ordering);
3251 CB.setArgOperand(i: 3, v: NewOrdering);
3252
3253 // TODO: Support ClCombinePointerLabelsOnStore
3254 // TODO: Support ClEventCallbacks
3255
3256 IRB.CreateCall(
3257 Callee: DFSF.DFS.DFSanMemShadowOriginTransferFn,
3258 Args: {DstPtr, SrcPtr, IRB.CreateIntCast(V: Size, DestTy: DFSF.DFS.IntptrTy, isSigned: false)});
3259}
3260
3261void DFSanVisitor::visitLibAtomicExchange(CallBase &CB) {
3262 // void __atomic_exchange(size_t size, void *ptr, void *val, void *ret, int
3263 // ordering)
3264 IRBuilder<> IRB(&CB);
3265 Value *Size = CB.getArgOperand(i: 0);
3266 Value *TargetPtr = CB.getArgOperand(i: 1);
3267 Value *SrcPtr = CB.getArgOperand(i: 2);
3268 Value *DstPtr = CB.getArgOperand(i: 3);
3269
3270 // This operation is not atomic for the shadow and origin memory.
3271 // This could result in DFSan false positives or false negatives.
3272 // For now we will assume these operations are rare, and
3273 // the additional complexity to address this is not warrented.
3274
3275 // Current Target to Dest
3276 IRB.CreateCall(
3277 Callee: DFSF.DFS.DFSanMemShadowOriginTransferFn,
3278 Args: {DstPtr, TargetPtr, IRB.CreateIntCast(V: Size, DestTy: DFSF.DFS.IntptrTy, isSigned: false)});
3279
3280 // Current Src to Target (overriding)
3281 IRB.CreateCall(
3282 Callee: DFSF.DFS.DFSanMemShadowOriginTransferFn,
3283 Args: {TargetPtr, SrcPtr, IRB.CreateIntCast(V: Size, DestTy: DFSF.DFS.IntptrTy, isSigned: false)});
3284}
3285
3286void DFSanVisitor::visitLibAtomicCompareExchange(CallBase &CB) {
3287 // bool __atomic_compare_exchange(size_t size, void *ptr, void *expected, void
3288 // *desired, int success_order, int failure_order)
3289 Value *Size = CB.getArgOperand(i: 0);
3290 Value *TargetPtr = CB.getArgOperand(i: 1);
3291 Value *ExpectedPtr = CB.getArgOperand(i: 2);
3292 Value *DesiredPtr = CB.getArgOperand(i: 3);
3293
3294 // This operation is not atomic for the shadow and origin memory.
3295 // This could result in DFSan false positives or false negatives.
3296 // For now we will assume these operations are rare, and
3297 // the additional complexity to address this is not warrented.
3298
3299 IRBuilder<> NextIRB(CB.getNextNode());
3300 NextIRB.SetCurrentDebugLocation(CB.getDebugLoc());
3301
3302 DFSF.setShadow(I: &CB, Shadow: DFSF.DFS.getZeroShadow(V: &CB));
3303
3304 // If original call returned true, copy Desired to Target.
3305 // If original call returned false, copy Target to Expected.
3306 NextIRB.CreateCall(Callee: DFSF.DFS.DFSanMemShadowOriginConditionalExchangeFn,
3307 Args: {NextIRB.CreateIntCast(V: &CB, DestTy: NextIRB.getInt8Ty(), isSigned: false),
3308 TargetPtr, ExpectedPtr, DesiredPtr,
3309 NextIRB.CreateIntCast(V: Size, DestTy: DFSF.DFS.IntptrTy, isSigned: false)});
3310}
3311
3312void DFSanVisitor::visitCallBase(CallBase &CB) {
3313 Function *F = CB.getCalledFunction();
3314 if ((F && F->isIntrinsic()) || CB.isInlineAsm()) {
3315 visitInstOperands(I&: CB);
3316 return;
3317 }
3318
3319 // Calls to this function are synthesized in wrappers, and we shouldn't
3320 // instrument them.
3321 if (F == DFSF.DFS.DFSanVarargWrapperFn.getCallee()->stripPointerCasts())
3322 return;
3323
3324 LibFunc LF;
3325 if (DFSF.TLI.getLibFunc(CB, F&: LF)) {
3326 // libatomic.a functions need to have special handling because there isn't
3327 // a good way to intercept them or compile the library with
3328 // instrumentation.
3329 switch (LF) {
3330 case LibFunc_atomic_load:
3331 if (!isa<CallInst>(Val: CB)) {
3332 llvm::errs() << "DFSAN -- cannot instrument invoke of libatomic load. "
3333 "Ignoring!\n";
3334 break;
3335 }
3336 visitLibAtomicLoad(CB);
3337 return;
3338 case LibFunc_atomic_store:
3339 visitLibAtomicStore(CB);
3340 return;
3341 default:
3342 break;
3343 }
3344 }
3345
3346 // TODO: These are not supported by TLI? They are not in the enum.
3347 if (F && F->hasName() && !F->isVarArg()) {
3348 if (F->getName() == "__atomic_exchange") {
3349 visitLibAtomicExchange(CB);
3350 return;
3351 }
3352 if (F->getName() == "__atomic_compare_exchange") {
3353 visitLibAtomicCompareExchange(CB);
3354 return;
3355 }
3356 }
3357
3358 DenseMap<Value *, Function *>::iterator UnwrappedFnIt =
3359 DFSF.DFS.UnwrappedFnMap.find(Val: CB.getCalledOperand());
3360 if (UnwrappedFnIt != DFSF.DFS.UnwrappedFnMap.end())
3361 if (visitWrappedCallBase(F&: *UnwrappedFnIt->second, CB))
3362 return;
3363
3364 IRBuilder<> IRB(&CB);
3365
3366 const bool ShouldTrackOrigins = DFSF.DFS.shouldTrackOrigins();
3367 FunctionType *FT = CB.getFunctionType();
3368 const DataLayout &DL = getDataLayout();
3369
3370 // Stores argument shadows.
3371 unsigned ArgOffset = 0;
3372 for (unsigned I = 0, N = FT->getNumParams(); I != N; ++I) {
3373 if (ShouldTrackOrigins) {
3374 // Ignore overflowed origins
3375 Value *ArgShadow = DFSF.getShadow(V: CB.getArgOperand(i: I));
3376 if (I < DFSF.DFS.NumOfElementsInArgOrgTLS &&
3377 !DFSF.DFS.isZeroShadow(V: ArgShadow))
3378 IRB.CreateStore(Val: DFSF.getOrigin(V: CB.getArgOperand(i: I)),
3379 Ptr: DFSF.getArgOriginTLS(ArgNo: I, IRB));
3380 }
3381
3382 unsigned Size =
3383 DL.getTypeAllocSize(Ty: DFSF.DFS.getShadowTy(OrigTy: FT->getParamType(i: I)));
3384 // Stop storing if arguments' size overflows. Inside a function, arguments
3385 // after overflow have zero shadow values.
3386 if (ArgOffset + Size > ArgTLSSize)
3387 break;
3388 IRB.CreateAlignedStore(Val: DFSF.getShadow(V: CB.getArgOperand(i: I)),
3389 Ptr: DFSF.getArgTLS(T: FT->getParamType(i: I), ArgOffset, IRB),
3390 Align: ShadowTLSAlignment);
3391 ArgOffset += alignTo(Size, A: ShadowTLSAlignment);
3392 }
3393
3394 Instruction *Next = nullptr;
3395 if (!CB.getType()->isVoidTy()) {
3396 if (InvokeInst *II = dyn_cast<InvokeInst>(Val: &CB)) {
3397 if (II->getNormalDest()->getSinglePredecessor()) {
3398 Next = &II->getNormalDest()->front();
3399 } else {
3400 BasicBlock *NewBB =
3401 SplitEdge(From: II->getParent(), To: II->getNormalDest(), DT: &DFSF.DT);
3402 Next = &NewBB->front();
3403 }
3404 } else {
3405 assert(CB.getIterator() != CB.getParent()->end());
3406 Next = CB.getNextNode();
3407 }
3408
3409 // Don't emit the epilogue for musttail call returns.
3410 if (isa<CallInst>(Val: CB) && cast<CallInst>(Val&: CB).isMustTailCall())
3411 return;
3412
3413 // Loads the return value shadow.
3414 IRBuilder<> NextIRB(Next);
3415 unsigned Size = DL.getTypeAllocSize(Ty: DFSF.DFS.getShadowTy(V: &CB));
3416 if (Size > RetvalTLSSize) {
3417 // Set overflowed return shadow to be zero.
3418 DFSF.setShadow(I: &CB, Shadow: DFSF.DFS.getZeroShadow(V: &CB));
3419 } else {
3420 LoadInst *LI = NextIRB.CreateAlignedLoad(
3421 Ty: DFSF.DFS.getShadowTy(V: &CB), Ptr: DFSF.getRetvalTLS(T: CB.getType(), IRB&: NextIRB),
3422 Align: ShadowTLSAlignment, Name: "_dfsret");
3423 DFSF.SkipInsts.insert(V: LI);
3424 DFSF.setShadow(I: &CB, Shadow: LI);
3425 DFSF.NonZeroChecks.push_back(x: LI);
3426 }
3427
3428 if (ShouldTrackOrigins) {
3429 LoadInst *LI = NextIRB.CreateLoad(Ty: DFSF.DFS.OriginTy,
3430 Ptr: DFSF.getRetvalOriginTLS(), Name: "_dfsret_o");
3431 DFSF.SkipInsts.insert(V: LI);
3432 DFSF.setOrigin(I: &CB, Origin: LI);
3433 }
3434
3435 DFSF.addReachesFunctionCallbacksIfEnabled(IRB&: NextIRB, I&: CB, Data: &CB);
3436 }
3437}
3438
3439void DFSanVisitor::visitPHINode(PHINode &PN) {
3440 Type *ShadowTy = DFSF.DFS.getShadowTy(V: &PN);
3441 PHINode *ShadowPN = PHINode::Create(Ty: ShadowTy, NumReservedValues: PN.getNumIncomingValues(), NameStr: "",
3442 InsertBefore: PN.getIterator());
3443
3444 // Give the shadow phi node valid predecessors to fool SplitEdge into working.
3445 Value *PoisonShadow = PoisonValue::get(T: ShadowTy);
3446 for (BasicBlock *BB : PN.blocks())
3447 ShadowPN->addIncoming(V: PoisonShadow, BB);
3448
3449 DFSF.setShadow(I: &PN, Shadow: ShadowPN);
3450
3451 PHINode *OriginPN = nullptr;
3452 if (DFSF.DFS.shouldTrackOrigins()) {
3453 OriginPN = PHINode::Create(Ty: DFSF.DFS.OriginTy, NumReservedValues: PN.getNumIncomingValues(), NameStr: "",
3454 InsertBefore: PN.getIterator());
3455 Value *PoisonOrigin = PoisonValue::get(T: DFSF.DFS.OriginTy);
3456 for (BasicBlock *BB : PN.blocks())
3457 OriginPN->addIncoming(V: PoisonOrigin, BB);
3458 DFSF.setOrigin(I: &PN, Origin: OriginPN);
3459 }
3460
3461 DFSF.PHIFixups.push_back(x: {.Phi: &PN, .ShadowPhi: ShadowPN, .OriginPhi: OriginPN});
3462}
3463
3464PreservedAnalyses DataFlowSanitizerPass::run(Module &M,
3465 ModuleAnalysisManager &AM) {
3466 // Return early if nosanitize_dataflow module flag is present for the module.
3467 if (checkIfAlreadyInstrumented(M, Flag: "nosanitize_dataflow"))
3468 return PreservedAnalyses::all();
3469 auto GetTLI = [&](Function &F) -> TargetLibraryInfo & {
3470 auto &FAM =
3471 AM.getResult<FunctionAnalysisManagerModuleProxy>(IR&: M).getManager();
3472 return FAM.getResult<TargetLibraryAnalysis>(IR&: F);
3473 };
3474 if (!DataFlowSanitizer(ABIListFiles).runImpl(M, GetTLI))
3475 return PreservedAnalyses::all();
3476
3477 PreservedAnalyses PA = PreservedAnalyses::none();
3478 // GlobalsAA is considered stateless and does not get invalidated unless
3479 // explicitly invalidated; PreservedAnalyses::none() is not enough. Sanitizers
3480 // make changes that require GlobalsAA to be invalidated.
3481 PA.abandon<GlobalsAA>();
3482 return PA;
3483}
3484