1 //===- AddressSanitizer.cpp - memory error detector -----------------------===//
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
7 //===----------------------------------------------------------------------===//
9 // This file is a part of AddressSanitizer, an address basic correctness
11 // Details of the algorithm:
12 // https://github.com/google/sanitizers/wiki/AddressSanitizerAlgorithm
14 // FIXME: This sanitizer does not yet handle scalable vectors
16 //===----------------------------------------------------------------------===//
18 #include "llvm/Transforms/Instrumentation/AddressSanitizer.h"
19 #include "llvm/ADT/ArrayRef.h"
20 #include "llvm/ADT/DenseMap.h"
21 #include "llvm/ADT/DepthFirstIterator.h"
22 #include "llvm/ADT/SmallPtrSet.h"
23 #include "llvm/ADT/SmallVector.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/ADT/StringExtras.h"
26 #include "llvm/ADT/StringRef.h"
27 #include "llvm/ADT/Triple.h"
28 #include "llvm/ADT/Twine.h"
29 #include "llvm/Analysis/MemoryBuiltins.h"
30 #include "llvm/Analysis/StackSafetyAnalysis.h"
31 #include "llvm/Analysis/TargetLibraryInfo.h"
32 #include "llvm/Analysis/ValueTracking.h"
33 #include "llvm/BinaryFormat/MachO.h"
34 #include "llvm/Demangle/Demangle.h"
35 #include "llvm/IR/Argument.h"
36 #include "llvm/IR/Attributes.h"
37 #include "llvm/IR/BasicBlock.h"
38 #include "llvm/IR/Comdat.h"
39 #include "llvm/IR/Constant.h"
40 #include "llvm/IR/Constants.h"
41 #include "llvm/IR/DIBuilder.h"
42 #include "llvm/IR/DataLayout.h"
43 #include "llvm/IR/DebugInfoMetadata.h"
44 #include "llvm/IR/DebugLoc.h"
45 #include "llvm/IR/DerivedTypes.h"
46 #include "llvm/IR/Function.h"
47 #include "llvm/IR/GlobalAlias.h"
48 #include "llvm/IR/GlobalValue.h"
49 #include "llvm/IR/GlobalVariable.h"
50 #include "llvm/IR/IRBuilder.h"
51 #include "llvm/IR/InlineAsm.h"
52 #include "llvm/IR/InstVisitor.h"
53 #include "llvm/IR/InstrTypes.h"
54 #include "llvm/IR/Instruction.h"
55 #include "llvm/IR/Instructions.h"
56 #include "llvm/IR/IntrinsicInst.h"
57 #include "llvm/IR/Intrinsics.h"
58 #include "llvm/IR/LLVMContext.h"
59 #include "llvm/IR/MDBuilder.h"
60 #include "llvm/IR/Metadata.h"
61 #include "llvm/IR/Module.h"
62 #include "llvm/IR/Type.h"
63 #include "llvm/IR/Use.h"
64 #include "llvm/IR/Value.h"
65 #include "llvm/MC/MCSectionMachO.h"
66 #include "llvm/Support/Casting.h"
67 #include "llvm/Support/CommandLine.h"
68 #include "llvm/Support/Debug.h"
69 #include "llvm/Support/ErrorHandling.h"
70 #include "llvm/Support/MathExtras.h"
71 #include "llvm/Support/raw_ostream.h"
72 #include "llvm/Transforms/Instrumentation.h"
73 #include "llvm/Transforms/Instrumentation/AddressSanitizerCommon.h"
74 #include "llvm/Transforms/Instrumentation/AddressSanitizerOptions.h"
75 #include "llvm/Transforms/Utils/ASanStackFrameLayout.h"
76 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
77 #include "llvm/Transforms/Utils/Local.h"
78 #include "llvm/Transforms/Utils/ModuleUtils.h"
79 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
92 #define DEBUG_TYPE "asan"
94 static const uint64_t kDefaultShadowScale = 3;
95 static const uint64_t kDefaultShadowOffset32 = 1ULL << 29;
96 static const uint64_t kDefaultShadowOffset64 = 1ULL << 44;
97 static const uint64_t kDynamicShadowSentinel =
98 std::numeric_limits<uint64_t>::max();
99 static const uint64_t kSmallX86_64ShadowOffsetBase = 0x7FFFFFFF; // < 2G.
100 static const uint64_t kSmallX86_64ShadowOffsetAlignMask = ~0xFFFULL;
101 static const uint64_t kLinuxKasan_ShadowOffset64 = 0xdffffc0000000000;
102 static const uint64_t kPPC64_ShadowOffset64 = 1ULL << 44;
103 static const uint64_t kSystemZ_ShadowOffset64 = 1ULL << 52;
104 static const uint64_t kMIPS_ShadowOffsetN32 = 1ULL << 29;
105 static const uint64_t kMIPS32_ShadowOffset32 = 0x0aaa0000;
106 static const uint64_t kMIPS64_ShadowOffset64 = 1ULL << 37;
107 static const uint64_t kAArch64_ShadowOffset64 = 1ULL << 36;
108 static const uint64_t kRISCV64_ShadowOffset64 = 0xd55550000;
109 static const uint64_t kFreeBSD_ShadowOffset32 = 1ULL << 30;
110 static const uint64_t kFreeBSD_ShadowOffset64 = 1ULL << 46;
111 static const uint64_t kFreeBSDAArch64_ShadowOffset64 = 1ULL << 47;
112 static const uint64_t kFreeBSDKasan_ShadowOffset64 = 0xdffff7c000000000;
113 static const uint64_t kNetBSD_ShadowOffset32 = 1ULL << 30;
114 static const uint64_t kNetBSD_ShadowOffset64 = 1ULL << 46;
115 static const uint64_t kNetBSDKasan_ShadowOffset64 = 0xdfff900000000000;
116 static const uint64_t kPS_ShadowOffset64 = 1ULL << 40;
117 static const uint64_t kWindowsShadowOffset32 = 3ULL << 28;
118 static const uint64_t kEmscriptenShadowOffset = 0;
120 // The shadow memory space is dynamically allocated.
121 static const uint64_t kWindowsShadowOffset64 = kDynamicShadowSentinel;
123 static const size_t kMinStackMallocSize = 1 << 6; // 64B
124 static const size_t kMaxStackMallocSize = 1 << 16; // 64K
125 static const uintptr_t kCurrentStackFrameMagic = 0x41B58AB3;
126 static const uintptr_t kRetiredStackFrameMagic = 0x45E0360E;
128 const char kAsanModuleCtorName[] = "asan.module_ctor";
129 const char kAsanModuleDtorName[] = "asan.module_dtor";
130 static const uint64_t kAsanCtorAndDtorPriority = 1;
131 // On Emscripten, the system needs more than one priorities for constructors.
132 static const uint64_t kAsanEmscriptenCtorAndDtorPriority = 50;
133 const char kAsanReportErrorTemplate[] = "__asan_report_";
134 const char kAsanRegisterGlobalsName[] = "__asan_register_globals";
135 const char kAsanUnregisterGlobalsName[] = "__asan_unregister_globals";
136 const char kAsanRegisterImageGlobalsName[] = "__asan_register_image_globals";
137 const char kAsanUnregisterImageGlobalsName[] =
138 "__asan_unregister_image_globals";
139 const char kAsanRegisterElfGlobalsName[] = "__asan_register_elf_globals";
140 const char kAsanUnregisterElfGlobalsName[] = "__asan_unregister_elf_globals";
141 const char kAsanPoisonGlobalsName[] = "__asan_before_dynamic_init";
142 const char kAsanUnpoisonGlobalsName[] = "__asan_after_dynamic_init";
143 const char kAsanInitName[] = "__asan_init";
144 const char kAsanVersionCheckNamePrefix[] = "__asan_version_mismatch_check_v";
145 const char kAsanPtrCmp[] = "__sanitizer_ptr_cmp";
146 const char kAsanPtrSub[] = "__sanitizer_ptr_sub";
147 const char kAsanHandleNoReturnName[] = "__asan_handle_no_return";
148 static const int kMaxAsanStackMallocSizeClass = 10;
149 const char kAsanStackMallocNameTemplate[] = "__asan_stack_malloc_";
150 const char kAsanStackMallocAlwaysNameTemplate[] =
151 "__asan_stack_malloc_always_";
152 const char kAsanStackFreeNameTemplate[] = "__asan_stack_free_";
153 const char kAsanGenPrefix[] = "___asan_gen_";
154 const char kODRGenPrefix[] = "__odr_asan_gen_";
155 const char kSanCovGenPrefix[] = "__sancov_gen_";
156 const char kAsanSetShadowPrefix[] = "__asan_set_shadow_";
157 const char kAsanPoisonStackMemoryName[] = "__asan_poison_stack_memory";
158 const char kAsanUnpoisonStackMemoryName[] = "__asan_unpoison_stack_memory";
160 // ASan version script has __asan_* wildcard. Triple underscore prevents a
161 // linker (gold) warning about attempting to export a local symbol.
162 const char kAsanGlobalsRegisteredFlagName[] = "___asan_globals_registered";
164 const char kAsanOptionDetectUseAfterReturn[] =
165 "__asan_option_detect_stack_use_after_return";
167 const char kAsanShadowMemoryDynamicAddress[] =
168 "__asan_shadow_memory_dynamic_address";
170 const char kAsanAllocaPoison[] = "__asan_alloca_poison";
171 const char kAsanAllocasUnpoison[] = "__asan_allocas_unpoison";
173 const char kAMDGPUAddressSharedName[] = "llvm.amdgcn.is.shared";
174 const char kAMDGPUAddressPrivateName[] = "llvm.amdgcn.is.private";
176 // Accesses sizes are powers of two: 1, 2, 4, 8, 16.
177 static const size_t kNumberOfAccessSizes = 5;
179 static const uint64_t kAllocaRzSize = 32;
181 // ASanAccessInfo implementation constants.
182 constexpr size_t kCompileKernelShift = 0;
183 constexpr size_t kCompileKernelMask = 0x1;
184 constexpr size_t kAccessSizeIndexShift = 1;
185 constexpr size_t kAccessSizeIndexMask = 0xf;
186 constexpr size_t kIsWriteShift = 5;
187 constexpr size_t kIsWriteMask = 0x1;
189 // Command-line flags.
191 static cl::opt<bool> ClEnableKasan(
192 "asan-kernel", cl::desc("Enable KernelAddressSanitizer instrumentation"),
193 cl::Hidden, cl::init(false));
195 static cl::opt<bool> ClRecover(
197 cl::desc("Enable recovery mode (continue-after-error)."),
198 cl::Hidden, cl::init(false));
200 static cl::opt<bool> ClInsertVersionCheck(
201 "asan-guard-against-version-mismatch",
202 cl::desc("Guard against compiler/runtime version mismatch."),
203 cl::Hidden, cl::init(true));
205 // This flag may need to be replaced with -f[no-]asan-reads.
206 static cl::opt<bool> ClInstrumentReads("asan-instrument-reads",
207 cl::desc("instrument read instructions"),
208 cl::Hidden, cl::init(true));
210 static cl::opt<bool> ClInstrumentWrites(
211 "asan-instrument-writes", cl::desc("instrument write instructions"),
212 cl::Hidden, cl::init(true));
215 ClUseStackSafety("asan-use-stack-safety", cl::Hidden, cl::init(false),
216 cl::Hidden, cl::desc("Use Stack Safety analysis results"),
219 static cl::opt<bool> ClInstrumentAtomics(
220 "asan-instrument-atomics",
221 cl::desc("instrument atomic instructions (rmw, cmpxchg)"), cl::Hidden,
225 ClInstrumentByval("asan-instrument-byval",
226 cl::desc("instrument byval call arguments"), cl::Hidden,
229 static cl::opt<bool> ClAlwaysSlowPath(
230 "asan-always-slow-path",
231 cl::desc("use instrumentation with slow path for all accesses"), cl::Hidden,
234 static cl::opt<bool> ClForceDynamicShadow(
235 "asan-force-dynamic-shadow",
236 cl::desc("Load shadow address into a local variable for each function"),
237 cl::Hidden, cl::init(false));
240 ClWithIfunc("asan-with-ifunc",
241 cl::desc("Access dynamic shadow through an ifunc global on "
242 "platforms that support this"),
243 cl::Hidden, cl::init(true));
245 static cl::opt<bool> ClWithIfuncSuppressRemat(
246 "asan-with-ifunc-suppress-remat",
247 cl::desc("Suppress rematerialization of dynamic shadow address by passing "
248 "it through inline asm in prologue."),
249 cl::Hidden, cl::init(true));
251 // This flag limits the number of instructions to be instrumented
252 // in any given BB. Normally, this should be set to unlimited (INT_MAX),
253 // but due to http://llvm.org/bugs/show_bug.cgi?id=12652 we temporary
255 static cl::opt<int> ClMaxInsnsToInstrumentPerBB(
256 "asan-max-ins-per-bb", cl::init(10000),
257 cl::desc("maximal number of instructions to instrument in any given BB"),
260 // This flag may need to be replaced with -f[no]asan-stack.
261 static cl::opt<bool> ClStack("asan-stack", cl::desc("Handle stack memory"),
262 cl::Hidden, cl::init(true));
263 static cl::opt<uint32_t> ClMaxInlinePoisoningSize(
264 "asan-max-inline-poisoning-size",
266 "Inline shadow poisoning for blocks up to the given size in bytes."),
267 cl::Hidden, cl::init(64));
269 static cl::opt<AsanDetectStackUseAfterReturnMode> ClUseAfterReturn(
270 "asan-use-after-return",
271 cl::desc("Sets the mode of detection for stack-use-after-return."),
273 clEnumValN(AsanDetectStackUseAfterReturnMode::Never, "never",
274 "Never detect stack use after return."),
276 AsanDetectStackUseAfterReturnMode::Runtime, "runtime",
277 "Detect stack use after return if "
278 "binary flag 'ASAN_OPTIONS=detect_stack_use_after_return' is set."),
279 clEnumValN(AsanDetectStackUseAfterReturnMode::Always, "always",
280 "Always detect stack use after return.")),
281 cl::Hidden, cl::init(AsanDetectStackUseAfterReturnMode::Runtime));
283 static cl::opt<bool> ClRedzoneByvalArgs("asan-redzone-byval-args",
284 cl::desc("Create redzones for byval "
285 "arguments (extra copy "
286 "required)"), cl::Hidden,
289 static cl::opt<bool> ClUseAfterScope("asan-use-after-scope",
290 cl::desc("Check stack-use-after-scope"),
291 cl::Hidden, cl::init(false));
293 // This flag may need to be replaced with -f[no]asan-globals.
294 static cl::opt<bool> ClGlobals("asan-globals",
295 cl::desc("Handle global objects"), cl::Hidden,
298 static cl::opt<bool> ClInitializers("asan-initialization-order",
299 cl::desc("Handle C++ initializer order"),
300 cl::Hidden, cl::init(true));
302 static cl::opt<bool> ClInvalidPointerPairs(
303 "asan-detect-invalid-pointer-pair",
304 cl::desc("Instrument <, <=, >, >=, - with pointer operands"), cl::Hidden,
307 static cl::opt<bool> ClInvalidPointerCmp(
308 "asan-detect-invalid-pointer-cmp",
309 cl::desc("Instrument <, <=, >, >= with pointer operands"), cl::Hidden,
312 static cl::opt<bool> ClInvalidPointerSub(
313 "asan-detect-invalid-pointer-sub",
314 cl::desc("Instrument - operations with pointer operands"), cl::Hidden,
317 static cl::opt<unsigned> ClRealignStack(
318 "asan-realign-stack",
319 cl::desc("Realign stack to the value of this flag (power of two)"),
320 cl::Hidden, cl::init(32));
322 static cl::opt<int> ClInstrumentationWithCallsThreshold(
323 "asan-instrumentation-with-call-threshold",
325 "If the function being instrumented contains more than "
326 "this number of memory accesses, use callbacks instead of "
327 "inline checks (-1 means never use callbacks)."),
328 cl::Hidden, cl::init(7000));
330 static cl::opt<std::string> ClMemoryAccessCallbackPrefix(
331 "asan-memory-access-callback-prefix",
332 cl::desc("Prefix for memory access callbacks"), cl::Hidden,
333 cl::init("__asan_"));
335 static cl::opt<bool> ClKasanMemIntrinCallbackPrefix(
336 "asan-kernel-mem-intrinsic-prefix",
337 cl::desc("Use prefix for memory intrinsics in KASAN mode"), cl::Hidden,
341 ClInstrumentDynamicAllocas("asan-instrument-dynamic-allocas",
342 cl::desc("instrument dynamic allocas"),
343 cl::Hidden, cl::init(true));
345 static cl::opt<bool> ClSkipPromotableAllocas(
346 "asan-skip-promotable-allocas",
347 cl::desc("Do not instrument promotable allocas"), cl::Hidden,
350 // These flags allow to change the shadow mapping.
351 // The shadow mapping looks like
352 // Shadow = (Mem >> scale) + offset
354 static cl::opt<int> ClMappingScale("asan-mapping-scale",
355 cl::desc("scale of asan shadow mapping"),
356 cl::Hidden, cl::init(0));
358 static cl::opt<uint64_t>
359 ClMappingOffset("asan-mapping-offset",
360 cl::desc("offset of asan shadow mapping [EXPERIMENTAL]"),
361 cl::Hidden, cl::init(0));
363 // Optimization flags. Not user visible, used mostly for testing
364 // and benchmarking the tool.
366 static cl::opt<bool> ClOpt("asan-opt", cl::desc("Optimize instrumentation"),
367 cl::Hidden, cl::init(true));
369 static cl::opt<bool> ClOptimizeCallbacks("asan-optimize-callbacks",
370 cl::desc("Optimize callbacks"),
371 cl::Hidden, cl::init(false));
373 static cl::opt<bool> ClOptSameTemp(
374 "asan-opt-same-temp", cl::desc("Instrument the same temp just once"),
375 cl::Hidden, cl::init(true));
377 static cl::opt<bool> ClOptGlobals("asan-opt-globals",
378 cl::desc("Don't instrument scalar globals"),
379 cl::Hidden, cl::init(true));
381 static cl::opt<bool> ClOptStack(
382 "asan-opt-stack", cl::desc("Don't instrument scalar stack variables"),
383 cl::Hidden, cl::init(false));
385 static cl::opt<bool> ClDynamicAllocaStack(
386 "asan-stack-dynamic-alloca",
387 cl::desc("Use dynamic alloca to represent stack variables"), cl::Hidden,
390 static cl::opt<uint32_t> ClForceExperiment(
391 "asan-force-experiment",
392 cl::desc("Force optimization experiment (for testing)"), cl::Hidden,
396 ClUsePrivateAlias("asan-use-private-alias",
397 cl::desc("Use private aliases for global variables"),
398 cl::Hidden, cl::init(false));
401 ClUseOdrIndicator("asan-use-odr-indicator",
402 cl::desc("Use odr indicators to improve ODR reporting"),
403 cl::Hidden, cl::init(false));
406 ClUseGlobalsGC("asan-globals-live-support",
407 cl::desc("Use linker features to support dead "
408 "code stripping of globals"),
409 cl::Hidden, cl::init(true));
411 // This is on by default even though there is a bug in gold:
412 // https://sourceware.org/bugzilla/show_bug.cgi?id=19002
414 ClWithComdat("asan-with-comdat",
415 cl::desc("Place ASan constructors in comdat sections"),
416 cl::Hidden, cl::init(true));
418 static cl::opt<AsanDtorKind> ClOverrideDestructorKind(
419 "asan-destructor-kind",
420 cl::desc("Sets the ASan destructor kind. The default is to use the value "
421 "provided to the pass constructor"),
422 cl::values(clEnumValN(AsanDtorKind::None, "none", "No destructors"),
423 clEnumValN(AsanDtorKind::Global, "global",
424 "Use global destructors")),
425 cl::init(AsanDtorKind::Invalid), cl::Hidden);
429 static cl::opt<int> ClDebug("asan-debug", cl::desc("debug"), cl::Hidden,
432 static cl::opt<int> ClDebugStack("asan-debug-stack", cl::desc("debug stack"),
433 cl::Hidden, cl::init(0));
435 static cl::opt<std::string> ClDebugFunc("asan-debug-func", cl::Hidden,
436 cl::desc("Debug func"));
438 static cl::opt<int> ClDebugMin("asan-debug-min", cl::desc("Debug min inst"),
439 cl::Hidden, cl::init(-1));
441 static cl::opt<int> ClDebugMax("asan-debug-max", cl::desc("Debug max inst"),
442 cl::Hidden, cl::init(-1));
444 STATISTIC(NumInstrumentedReads, "Number of instrumented reads");
445 STATISTIC(NumInstrumentedWrites, "Number of instrumented writes");
446 STATISTIC(NumOptimizedAccessesToGlobalVar,
447 "Number of optimized accesses to global vars");
448 STATISTIC(NumOptimizedAccessesToStackVar,
449 "Number of optimized accesses to stack vars");
453 /// This struct defines the shadow mapping using the rule:
454 /// shadow = (mem >> Scale) ADD-or-OR Offset.
455 /// If InGlobal is true, then
456 /// extern char __asan_shadow[];
457 /// shadow = (mem >> Scale) + &__asan_shadow
458 struct ShadowMapping {
465 } // end anonymous namespace
467 static ShadowMapping getShadowMapping(const Triple &TargetTriple, int LongSize,
469 bool IsAndroid = TargetTriple.isAndroid();
470 bool IsIOS = TargetTriple.isiOS() || TargetTriple.isWatchOS() ||
471 TargetTriple.isDriverKit();
472 bool IsMacOS = TargetTriple.isMacOSX();
473 bool IsFreeBSD = TargetTriple.isOSFreeBSD();
474 bool IsNetBSD = TargetTriple.isOSNetBSD();
475 bool IsPS = TargetTriple.isPS();
476 bool IsLinux = TargetTriple.isOSLinux();
477 bool IsPPC64 = TargetTriple.getArch() == Triple::ppc64 ||
478 TargetTriple.getArch() == Triple::ppc64le;
479 bool IsSystemZ = TargetTriple.getArch() == Triple::systemz;
480 bool IsX86_64 = TargetTriple.getArch() == Triple::x86_64;
481 bool IsMIPSN32ABI = TargetTriple.getEnvironment() == Triple::GNUABIN32;
482 bool IsMIPS32 = TargetTriple.isMIPS32();
483 bool IsMIPS64 = TargetTriple.isMIPS64();
484 bool IsArmOrThumb = TargetTriple.isARM() || TargetTriple.isThumb();
485 bool IsAArch64 = TargetTriple.getArch() == Triple::aarch64;
486 bool IsRISCV64 = TargetTriple.getArch() == Triple::riscv64;
487 bool IsWindows = TargetTriple.isOSWindows();
488 bool IsFuchsia = TargetTriple.isOSFuchsia();
489 bool IsEmscripten = TargetTriple.isOSEmscripten();
490 bool IsAMDGPU = TargetTriple.isAMDGPU();
492 ShadowMapping Mapping;
494 Mapping.Scale = kDefaultShadowScale;
495 if (ClMappingScale.getNumOccurrences() > 0) {
496 Mapping.Scale = ClMappingScale;
499 if (LongSize == 32) {
501 Mapping.Offset = kDynamicShadowSentinel;
502 else if (IsMIPSN32ABI)
503 Mapping.Offset = kMIPS_ShadowOffsetN32;
505 Mapping.Offset = kMIPS32_ShadowOffset32;
507 Mapping.Offset = kFreeBSD_ShadowOffset32;
509 Mapping.Offset = kNetBSD_ShadowOffset32;
511 Mapping.Offset = kDynamicShadowSentinel;
513 Mapping.Offset = kWindowsShadowOffset32;
514 else if (IsEmscripten)
515 Mapping.Offset = kEmscriptenShadowOffset;
517 Mapping.Offset = kDefaultShadowOffset32;
518 } else { // LongSize == 64
519 // Fuchsia is always PIE, which means that the beginning of the address
520 // space is always available.
524 Mapping.Offset = kPPC64_ShadowOffset64;
526 Mapping.Offset = kSystemZ_ShadowOffset64;
527 else if (IsFreeBSD && IsAArch64)
528 Mapping.Offset = kFreeBSDAArch64_ShadowOffset64;
529 else if (IsFreeBSD && !IsMIPS64) {
531 Mapping.Offset = kFreeBSDKasan_ShadowOffset64;
533 Mapping.Offset = kFreeBSD_ShadowOffset64;
534 } else if (IsNetBSD) {
536 Mapping.Offset = kNetBSDKasan_ShadowOffset64;
538 Mapping.Offset = kNetBSD_ShadowOffset64;
540 Mapping.Offset = kPS_ShadowOffset64;
541 else if (IsLinux && IsX86_64) {
543 Mapping.Offset = kLinuxKasan_ShadowOffset64;
545 Mapping.Offset = (kSmallX86_64ShadowOffsetBase &
546 (kSmallX86_64ShadowOffsetAlignMask << Mapping.Scale));
547 } else if (IsWindows && IsX86_64) {
548 Mapping.Offset = kWindowsShadowOffset64;
550 Mapping.Offset = kMIPS64_ShadowOffset64;
552 Mapping.Offset = kDynamicShadowSentinel;
553 else if (IsMacOS && IsAArch64)
554 Mapping.Offset = kDynamicShadowSentinel;
556 Mapping.Offset = kAArch64_ShadowOffset64;
558 Mapping.Offset = kRISCV64_ShadowOffset64;
560 Mapping.Offset = (kSmallX86_64ShadowOffsetBase &
561 (kSmallX86_64ShadowOffsetAlignMask << Mapping.Scale));
563 Mapping.Offset = kDefaultShadowOffset64;
566 if (ClForceDynamicShadow) {
567 Mapping.Offset = kDynamicShadowSentinel;
570 if (ClMappingOffset.getNumOccurrences() > 0) {
571 Mapping.Offset = ClMappingOffset;
574 // OR-ing shadow offset if more efficient (at least on x86) if the offset
575 // is a power of two, but on ppc64 we have to use add since the shadow
576 // offset is not necessary 1/8-th of the address space. On SystemZ,
577 // we could OR the constant in a single instruction, but it's more
578 // efficient to load it once and use indexed addressing.
579 Mapping.OrShadowOffset = !IsAArch64 && !IsPPC64 && !IsSystemZ && !IsPS &&
581 !(Mapping.Offset & (Mapping.Offset - 1)) &&
582 Mapping.Offset != kDynamicShadowSentinel;
583 bool IsAndroidWithIfuncSupport =
584 IsAndroid && !TargetTriple.isAndroidVersionLT(21);
585 Mapping.InGlobal = ClWithIfunc && IsAndroidWithIfuncSupport && IsArmOrThumb;
591 void getAddressSanitizerParams(const Triple &TargetTriple, int LongSize,
592 bool IsKasan, uint64_t *ShadowBase,
593 int *MappingScale, bool *OrShadowOffset) {
594 auto Mapping = getShadowMapping(TargetTriple, LongSize, IsKasan);
595 *ShadowBase = Mapping.Offset;
596 *MappingScale = Mapping.Scale;
597 *OrShadowOffset = Mapping.OrShadowOffset;
600 ASanAccessInfo::ASanAccessInfo(int32_t Packed)
602 AccessSizeIndex((Packed >> kAccessSizeIndexShift) & kAccessSizeIndexMask),
603 IsWrite((Packed >> kIsWriteShift) & kIsWriteMask),
604 CompileKernel((Packed >> kCompileKernelShift) & kCompileKernelMask) {}
606 ASanAccessInfo::ASanAccessInfo(bool IsWrite, bool CompileKernel,
607 uint8_t AccessSizeIndex)
608 : Packed((IsWrite << kIsWriteShift) +
609 (CompileKernel << kCompileKernelShift) +
610 (AccessSizeIndex << kAccessSizeIndexShift)),
611 AccessSizeIndex(AccessSizeIndex), IsWrite(IsWrite),
612 CompileKernel(CompileKernel) {}
616 static uint64_t getRedzoneSizeForScale(int MappingScale) {
617 // Redzone used for stack and globals is at least 32 bytes.
618 // For scales 6 and 7, the redzone has to be 64 and 128 bytes respectively.
619 return std::max(32U, 1U << MappingScale);
622 static uint64_t GetCtorAndDtorPriority(Triple &TargetTriple) {
623 if (TargetTriple.isOSEmscripten()) {
624 return kAsanEmscriptenCtorAndDtorPriority;
626 return kAsanCtorAndDtorPriority;
632 /// AddressSanitizer: instrument the code in module to find memory bugs.
633 struct AddressSanitizer {
634 AddressSanitizer(Module &M, const StackSafetyGlobalInfo *SSGI,
635 bool CompileKernel = false, bool Recover = false,
636 bool UseAfterScope = false,
637 AsanDetectStackUseAfterReturnMode UseAfterReturn =
638 AsanDetectStackUseAfterReturnMode::Runtime)
639 : CompileKernel(ClEnableKasan.getNumOccurrences() > 0 ? ClEnableKasan
641 Recover(ClRecover.getNumOccurrences() > 0 ? ClRecover : Recover),
642 UseAfterScope(UseAfterScope || ClUseAfterScope),
643 UseAfterReturn(ClUseAfterReturn.getNumOccurrences() ? ClUseAfterReturn
646 C = &(M.getContext());
647 LongSize = M.getDataLayout().getPointerSizeInBits();
648 IntptrTy = Type::getIntNTy(*C, LongSize);
649 Int8PtrTy = Type::getInt8PtrTy(*C);
650 Int32Ty = Type::getInt32Ty(*C);
651 TargetTriple = Triple(M.getTargetTriple());
653 Mapping = getShadowMapping(TargetTriple, LongSize, this->CompileKernel);
655 assert(this->UseAfterReturn != AsanDetectStackUseAfterReturnMode::Invalid);
658 uint64_t getAllocaSizeInBytes(const AllocaInst &AI) const {
659 uint64_t ArraySize = 1;
660 if (AI.isArrayAllocation()) {
661 const ConstantInt *CI = dyn_cast<ConstantInt>(AI.getArraySize());
662 assert(CI && "non-constant array size");
663 ArraySize = CI->getZExtValue();
665 Type *Ty = AI.getAllocatedType();
666 uint64_t SizeInBytes =
667 AI.getModule()->getDataLayout().getTypeAllocSize(Ty);
668 return SizeInBytes * ArraySize;
671 /// Check if we want (and can) handle this alloca.
672 bool isInterestingAlloca(const AllocaInst &AI);
674 bool ignoreAccess(Instruction *Inst, Value *Ptr);
675 void getInterestingMemoryOperands(
676 Instruction *I, SmallVectorImpl<InterestingMemoryOperand> &Interesting);
678 void instrumentMop(ObjectSizeOffsetVisitor &ObjSizeVis,
679 InterestingMemoryOperand &O, bool UseCalls,
680 const DataLayout &DL);
681 void instrumentPointerComparisonOrSubtraction(Instruction *I);
682 void instrumentAddress(Instruction *OrigIns, Instruction *InsertBefore,
683 Value *Addr, uint32_t TypeSize, bool IsWrite,
684 Value *SizeArgument, bool UseCalls, uint32_t Exp);
685 Instruction *instrumentAMDGPUAddress(Instruction *OrigIns,
686 Instruction *InsertBefore, Value *Addr,
687 uint32_t TypeSize, bool IsWrite,
688 Value *SizeArgument);
689 void instrumentUnusualSizeOrAlignment(Instruction *I,
690 Instruction *InsertBefore, Value *Addr,
691 uint32_t TypeSize, bool IsWrite,
692 Value *SizeArgument, bool UseCalls,
694 Value *createSlowPathCmp(IRBuilder<> &IRB, Value *AddrLong,
695 Value *ShadowValue, uint32_t TypeSize);
696 Instruction *generateCrashCode(Instruction *InsertBefore, Value *Addr,
697 bool IsWrite, size_t AccessSizeIndex,
698 Value *SizeArgument, uint32_t Exp);
699 void instrumentMemIntrinsic(MemIntrinsic *MI);
700 Value *memToShadow(Value *Shadow, IRBuilder<> &IRB);
701 bool suppressInstrumentationSiteForDebug(int &Instrumented);
702 bool instrumentFunction(Function &F, const TargetLibraryInfo *TLI);
703 bool maybeInsertAsanInitAtFunctionEntry(Function &F);
704 bool maybeInsertDynamicShadowAtFunctionEntry(Function &F);
705 void markEscapedLocalAllocas(Function &F);
708 friend struct FunctionStackPoisoner;
710 void initializeCallbacks(Module &M);
712 bool LooksLikeCodeInBug11395(Instruction *I);
713 bool GlobalIsLinkerInitialized(GlobalVariable *G);
714 bool isSafeAccess(ObjectSizeOffsetVisitor &ObjSizeVis, Value *Addr,
715 uint64_t TypeSize) const;
717 /// Helper to cleanup per-function state.
718 struct FunctionStateRAII {
719 AddressSanitizer *Pass;
721 FunctionStateRAII(AddressSanitizer *Pass) : Pass(Pass) {
722 assert(Pass->ProcessedAllocas.empty() &&
723 "last pass forgot to clear cache");
724 assert(!Pass->LocalDynamicShadow);
727 ~FunctionStateRAII() {
728 Pass->LocalDynamicShadow = nullptr;
729 Pass->ProcessedAllocas.clear();
739 AsanDetectStackUseAfterReturnMode UseAfterReturn;
743 ShadowMapping Mapping;
744 FunctionCallee AsanHandleNoReturnFunc;
745 FunctionCallee AsanPtrCmpFunction, AsanPtrSubFunction;
746 Constant *AsanShadowGlobal;
748 // These arrays is indexed by AccessIsWrite, Experiment and log2(AccessSize).
749 FunctionCallee AsanErrorCallback[2][2][kNumberOfAccessSizes];
750 FunctionCallee AsanMemoryAccessCallback[2][2][kNumberOfAccessSizes];
752 // These arrays is indexed by AccessIsWrite and Experiment.
753 FunctionCallee AsanErrorCallbackSized[2][2];
754 FunctionCallee AsanMemoryAccessCallbackSized[2][2];
756 FunctionCallee AsanMemmove, AsanMemcpy, AsanMemset;
757 Value *LocalDynamicShadow = nullptr;
758 const StackSafetyGlobalInfo *SSGI;
759 DenseMap<const AllocaInst *, bool> ProcessedAllocas;
761 FunctionCallee AMDGPUAddressShared;
762 FunctionCallee AMDGPUAddressPrivate;
765 class ModuleAddressSanitizer {
767 ModuleAddressSanitizer(Module &M, bool CompileKernel = false,
768 bool Recover = false, bool UseGlobalsGC = true,
769 bool UseOdrIndicator = false,
770 AsanDtorKind DestructorKind = AsanDtorKind::Global)
771 : CompileKernel(ClEnableKasan.getNumOccurrences() > 0 ? ClEnableKasan
773 Recover(ClRecover.getNumOccurrences() > 0 ? ClRecover : Recover),
774 UseGlobalsGC(UseGlobalsGC && ClUseGlobalsGC && !this->CompileKernel),
775 // Enable aliases as they should have no downside with ODR indicators.
776 UsePrivateAlias(UseOdrIndicator || ClUsePrivateAlias),
777 UseOdrIndicator(UseOdrIndicator || ClUseOdrIndicator),
778 // Not a typo: ClWithComdat is almost completely pointless without
779 // ClUseGlobalsGC (because then it only works on modules without
780 // globals, which are rare); it is a prerequisite for ClUseGlobalsGC;
781 // and both suffer from gold PR19002 for which UseGlobalsGC constructor
782 // argument is designed as workaround. Therefore, disable both
783 // ClWithComdat and ClUseGlobalsGC unless the frontend says it's ok to
785 UseCtorComdat(UseGlobalsGC && ClWithComdat && !this->CompileKernel),
786 DestructorKind(DestructorKind) {
787 C = &(M.getContext());
788 int LongSize = M.getDataLayout().getPointerSizeInBits();
789 IntptrTy = Type::getIntNTy(*C, LongSize);
790 TargetTriple = Triple(M.getTargetTriple());
791 Mapping = getShadowMapping(TargetTriple, LongSize, this->CompileKernel);
793 if (ClOverrideDestructorKind != AsanDtorKind::Invalid)
794 this->DestructorKind = ClOverrideDestructorKind;
795 assert(this->DestructorKind != AsanDtorKind::Invalid);
798 bool instrumentModule(Module &);
801 void initializeCallbacks(Module &M);
803 bool InstrumentGlobals(IRBuilder<> &IRB, Module &M, bool *CtorComdat);
804 void InstrumentGlobalsCOFF(IRBuilder<> &IRB, Module &M,
805 ArrayRef<GlobalVariable *> ExtendedGlobals,
806 ArrayRef<Constant *> MetadataInitializers);
807 void InstrumentGlobalsELF(IRBuilder<> &IRB, Module &M,
808 ArrayRef<GlobalVariable *> ExtendedGlobals,
809 ArrayRef<Constant *> MetadataInitializers,
810 const std::string &UniqueModuleId);
811 void InstrumentGlobalsMachO(IRBuilder<> &IRB, Module &M,
812 ArrayRef<GlobalVariable *> ExtendedGlobals,
813 ArrayRef<Constant *> MetadataInitializers);
815 InstrumentGlobalsWithMetadataArray(IRBuilder<> &IRB, Module &M,
816 ArrayRef<GlobalVariable *> ExtendedGlobals,
817 ArrayRef<Constant *> MetadataInitializers);
819 GlobalVariable *CreateMetadataGlobal(Module &M, Constant *Initializer,
820 StringRef OriginalName);
821 void SetComdatForGlobalMetadata(GlobalVariable *G, GlobalVariable *Metadata,
822 StringRef InternalSuffix);
823 Instruction *CreateAsanModuleDtor(Module &M);
825 const GlobalVariable *getExcludedAliasedGlobal(const GlobalAlias &GA) const;
826 bool shouldInstrumentGlobal(GlobalVariable *G) const;
827 bool ShouldUseMachOGlobalsSection() const;
828 StringRef getGlobalMetadataSection() const;
829 void poisonOneInitializer(Function &GlobalInit, GlobalValue *ModuleName);
830 void createInitializerPoisonCalls(Module &M, GlobalValue *ModuleName);
831 uint64_t getMinRedzoneSizeForGlobal() const {
832 return getRedzoneSizeForScale(Mapping.Scale);
834 uint64_t getRedzoneSizeForGlobal(uint64_t SizeInBytes) const;
835 int GetAsanVersion(const Module &M) const;
840 bool UsePrivateAlias;
841 bool UseOdrIndicator;
843 AsanDtorKind DestructorKind;
847 ShadowMapping Mapping;
848 FunctionCallee AsanPoisonGlobals;
849 FunctionCallee AsanUnpoisonGlobals;
850 FunctionCallee AsanRegisterGlobals;
851 FunctionCallee AsanUnregisterGlobals;
852 FunctionCallee AsanRegisterImageGlobals;
853 FunctionCallee AsanUnregisterImageGlobals;
854 FunctionCallee AsanRegisterElfGlobals;
855 FunctionCallee AsanUnregisterElfGlobals;
857 Function *AsanCtorFunction = nullptr;
858 Function *AsanDtorFunction = nullptr;
861 // Stack poisoning does not play well with exception handling.
862 // When an exception is thrown, we essentially bypass the code
863 // that unpoisones the stack. This is why the run-time library has
864 // to intercept __cxa_throw (as well as longjmp, etc) and unpoison the entire
865 // stack in the interceptor. This however does not work inside the
866 // actual function which catches the exception. Most likely because the
867 // compiler hoists the load of the shadow value somewhere too high.
868 // This causes asan to report a non-existing bug on 453.povray.
869 // It sounds like an LLVM bug.
870 struct FunctionStackPoisoner : public InstVisitor<FunctionStackPoisoner> {
872 AddressSanitizer &ASan;
877 ShadowMapping Mapping;
879 SmallVector<AllocaInst *, 16> AllocaVec;
880 SmallVector<AllocaInst *, 16> StaticAllocasToMoveUp;
881 SmallVector<Instruction *, 8> RetVec;
883 FunctionCallee AsanStackMallocFunc[kMaxAsanStackMallocSizeClass + 1],
884 AsanStackFreeFunc[kMaxAsanStackMallocSizeClass + 1];
885 FunctionCallee AsanSetShadowFunc[0x100] = {};
886 FunctionCallee AsanPoisonStackMemoryFunc, AsanUnpoisonStackMemoryFunc;
887 FunctionCallee AsanAllocaPoisonFunc, AsanAllocasUnpoisonFunc;
889 // Stores a place and arguments of poisoning/unpoisoning call for alloca.
890 struct AllocaPoisonCall {
891 IntrinsicInst *InsBefore;
896 SmallVector<AllocaPoisonCall, 8> DynamicAllocaPoisonCallVec;
897 SmallVector<AllocaPoisonCall, 8> StaticAllocaPoisonCallVec;
898 bool HasUntracedLifetimeIntrinsic = false;
900 SmallVector<AllocaInst *, 1> DynamicAllocaVec;
901 SmallVector<IntrinsicInst *, 1> StackRestoreVec;
902 AllocaInst *DynamicAllocaLayout = nullptr;
903 IntrinsicInst *LocalEscapeCall = nullptr;
905 bool HasInlineAsm = false;
906 bool HasReturnsTwiceCall = false;
909 FunctionStackPoisoner(Function &F, AddressSanitizer &ASan)
910 : F(F), ASan(ASan), DIB(*F.getParent(), /*AllowUnresolved*/ false),
911 C(ASan.C), IntptrTy(ASan.IntptrTy),
912 IntptrPtrTy(PointerType::get(IntptrTy, 0)), Mapping(ASan.Mapping),
913 PoisonStack(ClStack &&
914 !Triple(F.getParent()->getTargetTriple()).isAMDGPU()) {}
916 bool runOnFunction() {
920 if (ClRedzoneByvalArgs)
921 copyArgsPassedByValToAllocas();
923 // Collect alloca, ret, lifetime instructions etc.
924 for (BasicBlock *BB : depth_first(&F.getEntryBlock())) visit(*BB);
926 if (AllocaVec.empty() && DynamicAllocaVec.empty()) return false;
928 initializeCallbacks(*F.getParent());
930 if (HasUntracedLifetimeIntrinsic) {
931 // If there are lifetime intrinsics which couldn't be traced back to an
932 // alloca, we may not know exactly when a variable enters scope, and
933 // therefore should "fail safe" by not poisoning them.
934 StaticAllocaPoisonCallVec.clear();
935 DynamicAllocaPoisonCallVec.clear();
938 processDynamicAllocas();
939 processStaticAllocas();
942 LLVM_DEBUG(dbgs() << F);
947 // Arguments marked with the "byval" attribute are implicitly copied without
948 // using an alloca instruction. To produce redzones for those arguments, we
949 // copy them a second time into memory allocated with an alloca instruction.
950 void copyArgsPassedByValToAllocas();
952 // Finds all Alloca instructions and puts
953 // poisoned red zones around all of them.
954 // Then unpoison everything back before the function returns.
955 void processStaticAllocas();
956 void processDynamicAllocas();
958 void createDynamicAllocasInitStorage();
960 // ----------------------- Visitors.
961 /// Collect all Ret instructions, or the musttail call instruction if it
962 /// precedes the return instruction.
963 void visitReturnInst(ReturnInst &RI) {
964 if (CallInst *CI = RI.getParent()->getTerminatingMustTailCall())
965 RetVec.push_back(CI);
967 RetVec.push_back(&RI);
970 /// Collect all Resume instructions.
971 void visitResumeInst(ResumeInst &RI) { RetVec.push_back(&RI); }
973 /// Collect all CatchReturnInst instructions.
974 void visitCleanupReturnInst(CleanupReturnInst &CRI) { RetVec.push_back(&CRI); }
976 void unpoisonDynamicAllocasBeforeInst(Instruction *InstBefore,
978 IRBuilder<> IRB(InstBefore);
979 Value *DynamicAreaPtr = IRB.CreatePtrToInt(SavedStack, IntptrTy);
980 // When we insert _asan_allocas_unpoison before @llvm.stackrestore, we
981 // need to adjust extracted SP to compute the address of the most recent
982 // alloca. We have a special @llvm.get.dynamic.area.offset intrinsic for
984 if (!isa<ReturnInst>(InstBefore)) {
985 Function *DynamicAreaOffsetFunc = Intrinsic::getDeclaration(
986 InstBefore->getModule(), Intrinsic::get_dynamic_area_offset,
989 Value *DynamicAreaOffset = IRB.CreateCall(DynamicAreaOffsetFunc, {});
991 DynamicAreaPtr = IRB.CreateAdd(IRB.CreatePtrToInt(SavedStack, IntptrTy),
996 AsanAllocasUnpoisonFunc,
997 {IRB.CreateLoad(IntptrTy, DynamicAllocaLayout), DynamicAreaPtr});
1000 // Unpoison dynamic allocas redzones.
1001 void unpoisonDynamicAllocas() {
1002 for (Instruction *Ret : RetVec)
1003 unpoisonDynamicAllocasBeforeInst(Ret, DynamicAllocaLayout);
1005 for (Instruction *StackRestoreInst : StackRestoreVec)
1006 unpoisonDynamicAllocasBeforeInst(StackRestoreInst,
1007 StackRestoreInst->getOperand(0));
1010 // Deploy and poison redzones around dynamic alloca call. To do this, we
1011 // should replace this call with another one with changed parameters and
1012 // replace all its uses with new address, so
1013 // addr = alloca type, old_size, align
1015 // new_size = (old_size + additional_size) * sizeof(type)
1016 // tmp = alloca i8, new_size, max(align, 32)
1017 // addr = tmp + 32 (first 32 bytes are for the left redzone).
1018 // Additional_size is added to make new memory allocation contain not only
1019 // requested memory, but also left, partial and right redzones.
1020 void handleDynamicAllocaCall(AllocaInst *AI);
1022 /// Collect Alloca instructions we want (and can) handle.
1023 void visitAllocaInst(AllocaInst &AI) {
1024 if (!ASan.isInterestingAlloca(AI)) {
1025 if (AI.isStaticAlloca()) {
1026 // Skip over allocas that are present *before* the first instrumented
1027 // alloca, we don't want to move those around.
1028 if (AllocaVec.empty())
1031 StaticAllocasToMoveUp.push_back(&AI);
1036 if (!AI.isStaticAlloca())
1037 DynamicAllocaVec.push_back(&AI);
1039 AllocaVec.push_back(&AI);
1042 /// Collect lifetime intrinsic calls to check for use-after-scope
1044 void visitIntrinsicInst(IntrinsicInst &II) {
1045 Intrinsic::ID ID = II.getIntrinsicID();
1046 if (ID == Intrinsic::stackrestore) StackRestoreVec.push_back(&II);
1047 if (ID == Intrinsic::localescape) LocalEscapeCall = &II;
1048 if (!ASan.UseAfterScope)
1050 if (!II.isLifetimeStartOrEnd())
1052 // Found lifetime intrinsic, add ASan instrumentation if necessary.
1053 auto *Size = cast<ConstantInt>(II.getArgOperand(0));
1054 // If size argument is undefined, don't do anything.
1055 if (Size->isMinusOne()) return;
1056 // Check that size doesn't saturate uint64_t and can
1057 // be stored in IntptrTy.
1058 const uint64_t SizeValue = Size->getValue().getLimitedValue();
1059 if (SizeValue == ~0ULL ||
1060 !ConstantInt::isValueValidForType(IntptrTy, SizeValue))
1062 // Find alloca instruction that corresponds to llvm.lifetime argument.
1063 // Currently we can only handle lifetime markers pointing to the
1064 // beginning of the alloca.
1065 AllocaInst *AI = findAllocaForValue(II.getArgOperand(1), true);
1067 HasUntracedLifetimeIntrinsic = true;
1070 // We're interested only in allocas we can handle.
1071 if (!ASan.isInterestingAlloca(*AI))
1073 bool DoPoison = (ID == Intrinsic::lifetime_end);
1074 AllocaPoisonCall APC = {&II, AI, SizeValue, DoPoison};
1075 if (AI->isStaticAlloca())
1076 StaticAllocaPoisonCallVec.push_back(APC);
1077 else if (ClInstrumentDynamicAllocas)
1078 DynamicAllocaPoisonCallVec.push_back(APC);
1081 void visitCallBase(CallBase &CB) {
1082 if (CallInst *CI = dyn_cast<CallInst>(&CB)) {
1083 HasInlineAsm |= CI->isInlineAsm() && &CB != ASan.LocalDynamicShadow;
1084 HasReturnsTwiceCall |= CI->canReturnTwice();
1088 // ---------------------- Helpers.
1089 void initializeCallbacks(Module &M);
1091 // Copies bytes from ShadowBytes into shadow memory for indexes where
1092 // ShadowMask is not zero. If ShadowMask[i] is zero, we assume that
1093 // ShadowBytes[i] is constantly zero and doesn't need to be overwritten.
1094 void copyToShadow(ArrayRef<uint8_t> ShadowMask, ArrayRef<uint8_t> ShadowBytes,
1095 IRBuilder<> &IRB, Value *ShadowBase);
1096 void copyToShadow(ArrayRef<uint8_t> ShadowMask, ArrayRef<uint8_t> ShadowBytes,
1097 size_t Begin, size_t End, IRBuilder<> &IRB,
1099 void copyToShadowInline(ArrayRef<uint8_t> ShadowMask,
1100 ArrayRef<uint8_t> ShadowBytes, size_t Begin,
1101 size_t End, IRBuilder<> &IRB, Value *ShadowBase);
1103 void poisonAlloca(Value *V, uint64_t Size, IRBuilder<> &IRB, bool DoPoison);
1105 Value *createAllocaForLayout(IRBuilder<> &IRB, const ASanStackFrameLayout &L,
1107 PHINode *createPHI(IRBuilder<> &IRB, Value *Cond, Value *ValueIfTrue,
1108 Instruction *ThenTerm, Value *ValueIfFalse);
1111 } // end anonymous namespace
1113 void ModuleAddressSanitizerPass::printPipeline(
1114 raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) {
1115 static_cast<PassInfoMixin<ModuleAddressSanitizerPass> *>(this)->printPipeline(
1116 OS, MapClassName2PassName);
1118 if (Options.CompileKernel)
1123 ModuleAddressSanitizerPass::ModuleAddressSanitizerPass(
1124 const AddressSanitizerOptions &Options, bool UseGlobalGC,
1125 bool UseOdrIndicator, AsanDtorKind DestructorKind)
1126 : Options(Options), UseGlobalGC(UseGlobalGC),
1127 UseOdrIndicator(UseOdrIndicator), DestructorKind(DestructorKind) {}
1129 PreservedAnalyses ModuleAddressSanitizerPass::run(Module &M,
1130 ModuleAnalysisManager &MAM) {
1131 ModuleAddressSanitizer ModuleSanitizer(M, Options.CompileKernel,
1132 Options.Recover, UseGlobalGC,
1133 UseOdrIndicator, DestructorKind);
1134 bool Modified = false;
1135 auto &FAM = MAM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager();
1136 const StackSafetyGlobalInfo *const SSGI =
1137 ClUseStackSafety ? &MAM.getResult<StackSafetyGlobalAnalysis>(M) : nullptr;
1138 for (Function &F : M) {
1139 AddressSanitizer FunctionSanitizer(M, SSGI, Options.CompileKernel,
1140 Options.Recover, Options.UseAfterScope,
1141 Options.UseAfterReturn);
1142 const TargetLibraryInfo &TLI = FAM.getResult<TargetLibraryAnalysis>(F);
1143 Modified |= FunctionSanitizer.instrumentFunction(F, &TLI);
1145 Modified |= ModuleSanitizer.instrumentModule(M);
1146 return Modified ? PreservedAnalyses::none() : PreservedAnalyses::all();
1149 static size_t TypeSizeToSizeIndex(uint32_t TypeSize) {
1150 size_t Res = countTrailingZeros(TypeSize / 8);
1151 assert(Res < kNumberOfAccessSizes);
1155 /// Check if \p G has been created by a trusted compiler pass.
1156 static bool GlobalWasGeneratedByCompiler(GlobalVariable *G) {
1157 // Do not instrument @llvm.global_ctors, @llvm.used, etc.
1158 if (G->getName().startswith("llvm.") ||
1159 // Do not instrument gcov counter arrays.
1160 G->getName().startswith("__llvm_gcov_ctr") ||
1161 // Do not instrument rtti proxy symbols for function sanitizer.
1162 G->getName().startswith("__llvm_rtti_proxy"))
1165 // Do not instrument asan globals.
1166 if (G->getName().startswith(kAsanGenPrefix) ||
1167 G->getName().startswith(kSanCovGenPrefix) ||
1168 G->getName().startswith(kODRGenPrefix))
1174 static bool isUnsupportedAMDGPUAddrspace(Value *Addr) {
1175 Type *PtrTy = cast<PointerType>(Addr->getType()->getScalarType());
1176 unsigned int AddrSpace = PtrTy->getPointerAddressSpace();
1177 if (AddrSpace == 3 || AddrSpace == 5)
1182 Value *AddressSanitizer::memToShadow(Value *Shadow, IRBuilder<> &IRB) {
1184 Shadow = IRB.CreateLShr(Shadow, Mapping.Scale);
1185 if (Mapping.Offset == 0) return Shadow;
1186 // (Shadow >> scale) | offset
1188 if (LocalDynamicShadow)
1189 ShadowBase = LocalDynamicShadow;
1191 ShadowBase = ConstantInt::get(IntptrTy, Mapping.Offset);
1192 if (Mapping.OrShadowOffset)
1193 return IRB.CreateOr(Shadow, ShadowBase);
1195 return IRB.CreateAdd(Shadow, ShadowBase);
1198 // Instrument memset/memmove/memcpy
1199 void AddressSanitizer::instrumentMemIntrinsic(MemIntrinsic *MI) {
1200 IRBuilder<> IRB(MI);
1201 if (isa<MemTransferInst>(MI)) {
1203 isa<MemMoveInst>(MI) ? AsanMemmove : AsanMemcpy,
1204 {IRB.CreatePointerCast(MI->getOperand(0), IRB.getInt8PtrTy()),
1205 IRB.CreatePointerCast(MI->getOperand(1), IRB.getInt8PtrTy()),
1206 IRB.CreateIntCast(MI->getOperand(2), IntptrTy, false)});
1207 } else if (isa<MemSetInst>(MI)) {
1210 {IRB.CreatePointerCast(MI->getOperand(0), IRB.getInt8PtrTy()),
1211 IRB.CreateIntCast(MI->getOperand(1), IRB.getInt32Ty(), false),
1212 IRB.CreateIntCast(MI->getOperand(2), IntptrTy, false)});
1214 MI->eraseFromParent();
1217 /// Check if we want (and can) handle this alloca.
1218 bool AddressSanitizer::isInterestingAlloca(const AllocaInst &AI) {
1219 auto PreviouslySeenAllocaInfo = ProcessedAllocas.find(&AI);
1221 if (PreviouslySeenAllocaInfo != ProcessedAllocas.end())
1222 return PreviouslySeenAllocaInfo->getSecond();
1224 bool IsInteresting =
1225 (AI.getAllocatedType()->isSized() &&
1226 // alloca() may be called with 0 size, ignore it.
1227 ((!AI.isStaticAlloca()) || getAllocaSizeInBytes(AI) > 0) &&
1228 // We are only interested in allocas not promotable to registers.
1229 // Promotable allocas are common under -O0.
1230 (!ClSkipPromotableAllocas || !isAllocaPromotable(&AI)) &&
1231 // inalloca allocas are not treated as static, and we don't want
1232 // dynamic alloca instrumentation for them as well.
1233 !AI.isUsedWithInAlloca() &&
1234 // swifterror allocas are register promoted by ISel
1235 !AI.isSwiftError());
1237 ProcessedAllocas[&AI] = IsInteresting;
1238 return IsInteresting;
1241 bool AddressSanitizer::ignoreAccess(Instruction *Inst, Value *Ptr) {
1242 // Instrument acesses from different address spaces only for AMDGPU.
1243 Type *PtrTy = cast<PointerType>(Ptr->getType()->getScalarType());
1244 if (PtrTy->getPointerAddressSpace() != 0 &&
1245 !(TargetTriple.isAMDGPU() && !isUnsupportedAMDGPUAddrspace(Ptr)))
1248 // Ignore swifterror addresses.
1249 // swifterror memory addresses are mem2reg promoted by instruction
1250 // selection. As such they cannot have regular uses like an instrumentation
1251 // function and it makes no sense to track them as memory.
1252 if (Ptr->isSwiftError())
1255 // Treat memory accesses to promotable allocas as non-interesting since they
1256 // will not cause memory violations. This greatly speeds up the instrumented
1257 // executable at -O0.
1258 if (auto AI = dyn_cast_or_null<AllocaInst>(Ptr))
1259 if (ClSkipPromotableAllocas && !isInterestingAlloca(*AI))
1262 if (SSGI != nullptr && SSGI->stackAccessIsSafe(*Inst) &&
1263 findAllocaForValue(Ptr))
1269 void AddressSanitizer::getInterestingMemoryOperands(
1270 Instruction *I, SmallVectorImpl<InterestingMemoryOperand> &Interesting) {
1271 // Do not instrument the load fetching the dynamic shadow address.
1272 if (LocalDynamicShadow == I)
1275 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1276 if (!ClInstrumentReads || ignoreAccess(I, LI->getPointerOperand()))
1278 Interesting.emplace_back(I, LI->getPointerOperandIndex(), false,
1279 LI->getType(), LI->getAlign());
1280 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
1281 if (!ClInstrumentWrites || ignoreAccess(I, SI->getPointerOperand()))
1283 Interesting.emplace_back(I, SI->getPointerOperandIndex(), true,
1284 SI->getValueOperand()->getType(), SI->getAlign());
1285 } else if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(I)) {
1286 if (!ClInstrumentAtomics || ignoreAccess(I, RMW->getPointerOperand()))
1288 Interesting.emplace_back(I, RMW->getPointerOperandIndex(), true,
1289 RMW->getValOperand()->getType(), None);
1290 } else if (AtomicCmpXchgInst *XCHG = dyn_cast<AtomicCmpXchgInst>(I)) {
1291 if (!ClInstrumentAtomics || ignoreAccess(I, XCHG->getPointerOperand()))
1293 Interesting.emplace_back(I, XCHG->getPointerOperandIndex(), true,
1294 XCHG->getCompareOperand()->getType(), None);
1295 } else if (auto CI = dyn_cast<CallInst>(I)) {
1296 if (CI->getIntrinsicID() == Intrinsic::masked_load ||
1297 CI->getIntrinsicID() == Intrinsic::masked_store) {
1298 bool IsWrite = CI->getIntrinsicID() == Intrinsic::masked_store;
1299 // Masked store has an initial operand for the value.
1300 unsigned OpOffset = IsWrite ? 1 : 0;
1301 if (IsWrite ? !ClInstrumentWrites : !ClInstrumentReads)
1304 auto BasePtr = CI->getOperand(OpOffset);
1305 if (ignoreAccess(I, BasePtr))
1307 Type *Ty = IsWrite ? CI->getArgOperand(0)->getType() : CI->getType();
1308 MaybeAlign Alignment = Align(1);
1309 // Otherwise no alignment guarantees. We probably got Undef.
1310 if (auto *Op = dyn_cast<ConstantInt>(CI->getOperand(1 + OpOffset)))
1311 Alignment = Op->getMaybeAlignValue();
1312 Value *Mask = CI->getOperand(2 + OpOffset);
1313 Interesting.emplace_back(I, OpOffset, IsWrite, Ty, Alignment, Mask);
1315 for (unsigned ArgNo = 0; ArgNo < CI->arg_size(); ArgNo++) {
1316 if (!ClInstrumentByval || !CI->isByValArgument(ArgNo) ||
1317 ignoreAccess(I, CI->getArgOperand(ArgNo)))
1319 Type *Ty = CI->getParamByValType(ArgNo);
1320 Interesting.emplace_back(I, ArgNo, false, Ty, Align(1));
1326 static bool isPointerOperand(Value *V) {
1327 return V->getType()->isPointerTy() || isa<PtrToIntInst>(V);
1330 // This is a rough heuristic; it may cause both false positives and
1331 // false negatives. The proper implementation requires cooperation with
1333 static bool isInterestingPointerComparison(Instruction *I) {
1334 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(I)) {
1335 if (!Cmp->isRelational())
1340 return isPointerOperand(I->getOperand(0)) &&
1341 isPointerOperand(I->getOperand(1));
1344 // This is a rough heuristic; it may cause both false positives and
1345 // false negatives. The proper implementation requires cooperation with
1347 static bool isInterestingPointerSubtraction(Instruction *I) {
1348 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
1349 if (BO->getOpcode() != Instruction::Sub)
1354 return isPointerOperand(I->getOperand(0)) &&
1355 isPointerOperand(I->getOperand(1));
1358 bool AddressSanitizer::GlobalIsLinkerInitialized(GlobalVariable *G) {
1359 // If a global variable does not have dynamic initialization we don't
1360 // have to instrument it. However, if a global does not have initializer
1361 // at all, we assume it has dynamic initializer (in other TU).
1362 if (!G->hasInitializer())
1365 if (G->hasSanitizerMetadata() && G->getSanitizerMetadata().IsDynInit)
1371 void AddressSanitizer::instrumentPointerComparisonOrSubtraction(
1374 FunctionCallee F = isa<ICmpInst>(I) ? AsanPtrCmpFunction : AsanPtrSubFunction;
1375 Value *Param[2] = {I->getOperand(0), I->getOperand(1)};
1376 for (Value *&i : Param) {
1377 if (i->getType()->isPointerTy())
1378 i = IRB.CreatePointerCast(i, IntptrTy);
1380 IRB.CreateCall(F, Param);
1383 static void doInstrumentAddress(AddressSanitizer *Pass, Instruction *I,
1384 Instruction *InsertBefore, Value *Addr,
1385 MaybeAlign Alignment, unsigned Granularity,
1386 uint32_t TypeSize, bool IsWrite,
1387 Value *SizeArgument, bool UseCalls,
1389 // Instrument a 1-, 2-, 4-, 8-, or 16- byte access with one check
1390 // if the data is properly aligned.
1391 if ((TypeSize == 8 || TypeSize == 16 || TypeSize == 32 || TypeSize == 64 ||
1393 (!Alignment || *Alignment >= Granularity || *Alignment >= TypeSize / 8))
1394 return Pass->instrumentAddress(I, InsertBefore, Addr, TypeSize, IsWrite,
1395 nullptr, UseCalls, Exp);
1396 Pass->instrumentUnusualSizeOrAlignment(I, InsertBefore, Addr, TypeSize,
1397 IsWrite, nullptr, UseCalls, Exp);
1400 static void instrumentMaskedLoadOrStore(AddressSanitizer *Pass,
1401 const DataLayout &DL, Type *IntptrTy,
1402 Value *Mask, Instruction *I,
1403 Value *Addr, MaybeAlign Alignment,
1404 unsigned Granularity, Type *OpType,
1405 bool IsWrite, Value *SizeArgument,
1406 bool UseCalls, uint32_t Exp) {
1407 auto *VTy = cast<FixedVectorType>(OpType);
1408 uint64_t ElemTypeSize = DL.getTypeStoreSizeInBits(VTy->getScalarType());
1409 unsigned Num = VTy->getNumElements();
1410 auto Zero = ConstantInt::get(IntptrTy, 0);
1411 for (unsigned Idx = 0; Idx < Num; ++Idx) {
1412 Value *InstrumentedAddress = nullptr;
1413 Instruction *InsertBefore = I;
1414 if (auto *Vector = dyn_cast<ConstantVector>(Mask)) {
1415 // dyn_cast as we might get UndefValue
1416 if (auto *Masked = dyn_cast<ConstantInt>(Vector->getOperand(Idx))) {
1417 if (Masked->isZero())
1418 // Mask is constant false, so no instrumentation needed.
1420 // If we have a true or undef value, fall through to doInstrumentAddress
1421 // with InsertBefore == I
1425 Value *MaskElem = IRB.CreateExtractElement(Mask, Idx);
1426 Instruction *ThenTerm = SplitBlockAndInsertIfThen(MaskElem, I, false);
1427 InsertBefore = ThenTerm;
1430 IRBuilder<> IRB(InsertBefore);
1431 InstrumentedAddress =
1432 IRB.CreateGEP(VTy, Addr, {Zero, ConstantInt::get(IntptrTy, Idx)});
1433 doInstrumentAddress(Pass, I, InsertBefore, InstrumentedAddress, Alignment,
1434 Granularity, ElemTypeSize, IsWrite, SizeArgument,
1439 void AddressSanitizer::instrumentMop(ObjectSizeOffsetVisitor &ObjSizeVis,
1440 InterestingMemoryOperand &O, bool UseCalls,
1441 const DataLayout &DL) {
1442 Value *Addr = O.getPtr();
1444 // Optimization experiments.
1445 // The experiments can be used to evaluate potential optimizations that remove
1446 // instrumentation (assess false negatives). Instead of completely removing
1447 // some instrumentation, you set Exp to a non-zero value (mask of optimization
1448 // experiments that want to remove instrumentation of this instruction).
1449 // If Exp is non-zero, this pass will emit special calls into runtime
1450 // (e.g. __asan_report_exp_load1 instead of __asan_report_load1). These calls
1451 // make runtime terminate the program in a special way (with a different
1452 // exit status). Then you run the new compiler on a buggy corpus, collect
1453 // the special terminations (ideally, you don't see them at all -- no false
1454 // negatives) and make the decision on the optimization.
1455 uint32_t Exp = ClForceExperiment;
1457 if (ClOpt && ClOptGlobals) {
1458 // If initialization order checking is disabled, a simple access to a
1459 // dynamically initialized global is always valid.
1460 GlobalVariable *G = dyn_cast<GlobalVariable>(getUnderlyingObject(Addr));
1461 if (G && (!ClInitializers || GlobalIsLinkerInitialized(G)) &&
1462 isSafeAccess(ObjSizeVis, Addr, O.TypeSize)) {
1463 NumOptimizedAccessesToGlobalVar++;
1468 if (ClOpt && ClOptStack) {
1469 // A direct inbounds access to a stack variable is always valid.
1470 if (isa<AllocaInst>(getUnderlyingObject(Addr)) &&
1471 isSafeAccess(ObjSizeVis, Addr, O.TypeSize)) {
1472 NumOptimizedAccessesToStackVar++;
1478 NumInstrumentedWrites++;
1480 NumInstrumentedReads++;
1482 unsigned Granularity = 1 << Mapping.Scale;
1484 instrumentMaskedLoadOrStore(this, DL, IntptrTy, O.MaybeMask, O.getInsn(),
1485 Addr, O.Alignment, Granularity, O.OpType,
1486 O.IsWrite, nullptr, UseCalls, Exp);
1488 doInstrumentAddress(this, O.getInsn(), O.getInsn(), Addr, O.Alignment,
1489 Granularity, O.TypeSize, O.IsWrite, nullptr, UseCalls,
1494 Instruction *AddressSanitizer::generateCrashCode(Instruction *InsertBefore,
1495 Value *Addr, bool IsWrite,
1496 size_t AccessSizeIndex,
1497 Value *SizeArgument,
1499 IRBuilder<> IRB(InsertBefore);
1500 Value *ExpVal = Exp == 0 ? nullptr : ConstantInt::get(IRB.getInt32Ty(), Exp);
1501 CallInst *Call = nullptr;
1504 Call = IRB.CreateCall(AsanErrorCallbackSized[IsWrite][0],
1505 {Addr, SizeArgument});
1507 Call = IRB.CreateCall(AsanErrorCallbackSized[IsWrite][1],
1508 {Addr, SizeArgument, ExpVal});
1512 IRB.CreateCall(AsanErrorCallback[IsWrite][0][AccessSizeIndex], Addr);
1514 Call = IRB.CreateCall(AsanErrorCallback[IsWrite][1][AccessSizeIndex],
1518 Call->setCannotMerge();
1522 Value *AddressSanitizer::createSlowPathCmp(IRBuilder<> &IRB, Value *AddrLong,
1524 uint32_t TypeSize) {
1525 size_t Granularity = static_cast<size_t>(1) << Mapping.Scale;
1526 // Addr & (Granularity - 1)
1527 Value *LastAccessedByte =
1528 IRB.CreateAnd(AddrLong, ConstantInt::get(IntptrTy, Granularity - 1));
1529 // (Addr & (Granularity - 1)) + size - 1
1530 if (TypeSize / 8 > 1)
1531 LastAccessedByte = IRB.CreateAdd(
1532 LastAccessedByte, ConstantInt::get(IntptrTy, TypeSize / 8 - 1));
1533 // (uint8_t) ((Addr & (Granularity-1)) + size - 1)
1535 IRB.CreateIntCast(LastAccessedByte, ShadowValue->getType(), false);
1536 // ((uint8_t) ((Addr & (Granularity-1)) + size - 1)) >= ShadowValue
1537 return IRB.CreateICmpSGE(LastAccessedByte, ShadowValue);
1540 Instruction *AddressSanitizer::instrumentAMDGPUAddress(
1541 Instruction *OrigIns, Instruction *InsertBefore, Value *Addr,
1542 uint32_t TypeSize, bool IsWrite, Value *SizeArgument) {
1543 // Do not instrument unsupported addrspaces.
1544 if (isUnsupportedAMDGPUAddrspace(Addr))
1546 Type *PtrTy = cast<PointerType>(Addr->getType()->getScalarType());
1547 // Follow host instrumentation for global and constant addresses.
1548 if (PtrTy->getPointerAddressSpace() != 0)
1549 return InsertBefore;
1550 // Instrument generic addresses in supported addressspaces.
1551 IRBuilder<> IRB(InsertBefore);
1552 Value *AddrLong = IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy());
1553 Value *IsShared = IRB.CreateCall(AMDGPUAddressShared, {AddrLong});
1554 Value *IsPrivate = IRB.CreateCall(AMDGPUAddressPrivate, {AddrLong});
1555 Value *IsSharedOrPrivate = IRB.CreateOr(IsShared, IsPrivate);
1556 Value *Cmp = IRB.CreateICmpNE(IRB.getTrue(), IsSharedOrPrivate);
1557 Value *AddrSpaceZeroLanding =
1558 SplitBlockAndInsertIfThen(Cmp, InsertBefore, false);
1559 InsertBefore = cast<Instruction>(AddrSpaceZeroLanding);
1560 return InsertBefore;
1563 void AddressSanitizer::instrumentAddress(Instruction *OrigIns,
1564 Instruction *InsertBefore, Value *Addr,
1565 uint32_t TypeSize, bool IsWrite,
1566 Value *SizeArgument, bool UseCalls,
1568 if (TargetTriple.isAMDGPU()) {
1569 InsertBefore = instrumentAMDGPUAddress(OrigIns, InsertBefore, Addr,
1570 TypeSize, IsWrite, SizeArgument);
1575 IRBuilder<> IRB(InsertBefore);
1576 size_t AccessSizeIndex = TypeSizeToSizeIndex(TypeSize);
1577 const ASanAccessInfo AccessInfo(IsWrite, CompileKernel, AccessSizeIndex);
1579 if (UseCalls && ClOptimizeCallbacks) {
1580 const ASanAccessInfo AccessInfo(IsWrite, CompileKernel, AccessSizeIndex);
1581 Module *M = IRB.GetInsertBlock()->getParent()->getParent();
1583 Intrinsic::getDeclaration(M, Intrinsic::asan_check_memaccess),
1584 {IRB.CreatePointerCast(Addr, Int8PtrTy),
1585 ConstantInt::get(Int32Ty, AccessInfo.Packed)});
1589 Value *AddrLong = IRB.CreatePointerCast(Addr, IntptrTy);
1592 IRB.CreateCall(AsanMemoryAccessCallback[IsWrite][0][AccessSizeIndex],
1595 IRB.CreateCall(AsanMemoryAccessCallback[IsWrite][1][AccessSizeIndex],
1596 {AddrLong, ConstantInt::get(IRB.getInt32Ty(), Exp)});
1601 IntegerType::get(*C, std::max(8U, TypeSize >> Mapping.Scale));
1602 Type *ShadowPtrTy = PointerType::get(ShadowTy, 0);
1603 Value *ShadowPtr = memToShadow(AddrLong, IRB);
1604 Value *CmpVal = Constant::getNullValue(ShadowTy);
1605 Value *ShadowValue =
1606 IRB.CreateLoad(ShadowTy, IRB.CreateIntToPtr(ShadowPtr, ShadowPtrTy));
1608 Value *Cmp = IRB.CreateICmpNE(ShadowValue, CmpVal);
1609 size_t Granularity = 1ULL << Mapping.Scale;
1610 Instruction *CrashTerm = nullptr;
1612 if (ClAlwaysSlowPath || (TypeSize < 8 * Granularity)) {
1613 // We use branch weights for the slow path check, to indicate that the slow
1614 // path is rarely taken. This seems to be the case for SPEC benchmarks.
1615 Instruction *CheckTerm = SplitBlockAndInsertIfThen(
1616 Cmp, InsertBefore, false, MDBuilder(*C).createBranchWeights(1, 100000));
1617 assert(cast<BranchInst>(CheckTerm)->isUnconditional());
1618 BasicBlock *NextBB = CheckTerm->getSuccessor(0);
1619 IRB.SetInsertPoint(CheckTerm);
1620 Value *Cmp2 = createSlowPathCmp(IRB, AddrLong, ShadowValue, TypeSize);
1622 CrashTerm = SplitBlockAndInsertIfThen(Cmp2, CheckTerm, false);
1624 BasicBlock *CrashBlock =
1625 BasicBlock::Create(*C, "", NextBB->getParent(), NextBB);
1626 CrashTerm = new UnreachableInst(*C, CrashBlock);
1627 BranchInst *NewTerm = BranchInst::Create(CrashBlock, NextBB, Cmp2);
1628 ReplaceInstWithInst(CheckTerm, NewTerm);
1631 CrashTerm = SplitBlockAndInsertIfThen(Cmp, InsertBefore, !Recover);
1634 Instruction *Crash = generateCrashCode(CrashTerm, AddrLong, IsWrite,
1635 AccessSizeIndex, SizeArgument, Exp);
1636 Crash->setDebugLoc(OrigIns->getDebugLoc());
1639 // Instrument unusual size or unusual alignment.
1640 // We can not do it with a single check, so we do 1-byte check for the first
1641 // and the last bytes. We call __asan_report_*_n(addr, real_size) to be able
1642 // to report the actual access size.
1643 void AddressSanitizer::instrumentUnusualSizeOrAlignment(
1644 Instruction *I, Instruction *InsertBefore, Value *Addr, uint32_t TypeSize,
1645 bool IsWrite, Value *SizeArgument, bool UseCalls, uint32_t Exp) {
1646 IRBuilder<> IRB(InsertBefore);
1647 Value *Size = ConstantInt::get(IntptrTy, TypeSize / 8);
1648 Value *AddrLong = IRB.CreatePointerCast(Addr, IntptrTy);
1651 IRB.CreateCall(AsanMemoryAccessCallbackSized[IsWrite][0],
1654 IRB.CreateCall(AsanMemoryAccessCallbackSized[IsWrite][1],
1655 {AddrLong, Size, ConstantInt::get(IRB.getInt32Ty(), Exp)});
1657 Value *LastByte = IRB.CreateIntToPtr(
1658 IRB.CreateAdd(AddrLong, ConstantInt::get(IntptrTy, TypeSize / 8 - 1)),
1660 instrumentAddress(I, InsertBefore, Addr, 8, IsWrite, Size, false, Exp);
1661 instrumentAddress(I, InsertBefore, LastByte, 8, IsWrite, Size, false, Exp);
1665 void ModuleAddressSanitizer::poisonOneInitializer(Function &GlobalInit,
1666 GlobalValue *ModuleName) {
1667 // Set up the arguments to our poison/unpoison functions.
1668 IRBuilder<> IRB(&GlobalInit.front(),
1669 GlobalInit.front().getFirstInsertionPt());
1671 // Add a call to poison all external globals before the given function starts.
1672 Value *ModuleNameAddr = ConstantExpr::getPointerCast(ModuleName, IntptrTy);
1673 IRB.CreateCall(AsanPoisonGlobals, ModuleNameAddr);
1675 // Add calls to unpoison all globals before each return instruction.
1676 for (auto &BB : GlobalInit.getBasicBlockList())
1677 if (ReturnInst *RI = dyn_cast<ReturnInst>(BB.getTerminator()))
1678 CallInst::Create(AsanUnpoisonGlobals, "", RI);
1681 void ModuleAddressSanitizer::createInitializerPoisonCalls(
1682 Module &M, GlobalValue *ModuleName) {
1683 GlobalVariable *GV = M.getGlobalVariable("llvm.global_ctors");
1687 ConstantArray *CA = dyn_cast<ConstantArray>(GV->getInitializer());
1691 for (Use &OP : CA->operands()) {
1692 if (isa<ConstantAggregateZero>(OP)) continue;
1693 ConstantStruct *CS = cast<ConstantStruct>(OP);
1695 // Must have a function or null ptr.
1696 if (Function *F = dyn_cast<Function>(CS->getOperand(1))) {
1697 if (F->getName() == kAsanModuleCtorName) continue;
1698 auto *Priority = cast<ConstantInt>(CS->getOperand(0));
1699 // Don't instrument CTORs that will run before asan.module_ctor.
1700 if (Priority->getLimitedValue() <= GetCtorAndDtorPriority(TargetTriple))
1702 poisonOneInitializer(*F, ModuleName);
1707 const GlobalVariable *
1708 ModuleAddressSanitizer::getExcludedAliasedGlobal(const GlobalAlias &GA) const {
1709 // In case this function should be expanded to include rules that do not just
1710 // apply when CompileKernel is true, either guard all existing rules with an
1711 // 'if (CompileKernel) { ... }' or be absolutely sure that all these rules
1712 // should also apply to user space.
1713 assert(CompileKernel && "Only expecting to be called when compiling kernel");
1715 const Constant *C = GA.getAliasee();
1717 // When compiling the kernel, globals that are aliased by symbols prefixed
1718 // by "__" are special and cannot be padded with a redzone.
1719 if (GA.getName().startswith("__"))
1720 return dyn_cast<GlobalVariable>(C->stripPointerCastsAndAliases());
1725 bool ModuleAddressSanitizer::shouldInstrumentGlobal(GlobalVariable *G) const {
1726 Type *Ty = G->getValueType();
1727 LLVM_DEBUG(dbgs() << "GLOBAL: " << *G << "\n");
1729 if (G->hasSanitizerMetadata() && G->getSanitizerMetadata().NoAddress)
1731 if (!Ty->isSized()) return false;
1732 if (!G->hasInitializer()) return false;
1733 // Globals in address space 1 and 4 are supported for AMDGPU.
1734 if (G->getAddressSpace() &&
1735 !(TargetTriple.isAMDGPU() && !isUnsupportedAMDGPUAddrspace(G)))
1737 if (GlobalWasGeneratedByCompiler(G)) return false; // Our own globals.
1738 // Two problems with thread-locals:
1739 // - The address of the main thread's copy can't be computed at link-time.
1740 // - Need to poison all copies, not just the main thread's one.
1741 if (G->isThreadLocal()) return false;
1742 // For now, just ignore this Global if the alignment is large.
1743 if (G->getAlignment() > getMinRedzoneSizeForGlobal()) return false;
1745 // For non-COFF targets, only instrument globals known to be defined by this
1747 // FIXME: We can instrument comdat globals on ELF if we are using the
1748 // GC-friendly metadata scheme.
1749 if (!TargetTriple.isOSBinFormatCOFF()) {
1750 if (!G->hasExactDefinition() || G->hasComdat())
1753 // On COFF, don't instrument non-ODR linkages.
1754 if (G->isInterposable())
1758 // If a comdat is present, it must have a selection kind that implies ODR
1759 // semantics: no duplicates, any, or exact match.
1760 if (Comdat *C = G->getComdat()) {
1761 switch (C->getSelectionKind()) {
1763 case Comdat::ExactMatch:
1764 case Comdat::NoDeduplicate:
1766 case Comdat::Largest:
1767 case Comdat::SameSize:
1772 if (G->hasSection()) {
1773 // The kernel uses explicit sections for mostly special global variables
1774 // that we should not instrument. E.g. the kernel may rely on their layout
1775 // without redzones, or remove them at link time ("discard.*"), etc.
1779 StringRef Section = G->getSection();
1781 // Globals from llvm.metadata aren't emitted, do not instrument them.
1782 if (Section == "llvm.metadata") return false;
1783 // Do not instrument globals from special LLVM sections.
1784 if (Section.contains("__llvm") || Section.contains("__LLVM"))
1787 // Do not instrument function pointers to initialization and termination
1788 // routines: dynamic linker will not properly handle redzones.
1789 if (Section.startswith(".preinit_array") ||
1790 Section.startswith(".init_array") ||
1791 Section.startswith(".fini_array")) {
1795 // Do not instrument user-defined sections (with names resembling
1796 // valid C identifiers)
1797 if (TargetTriple.isOSBinFormatELF()) {
1798 if (llvm::all_of(Section,
1799 [](char c) { return llvm::isAlnum(c) || c == '_'; }))
1803 // On COFF, if the section name contains '$', it is highly likely that the
1804 // user is using section sorting to create an array of globals similar to
1805 // the way initialization callbacks are registered in .init_array and
1806 // .CRT$XCU. The ATL also registers things in .ATL$__[azm]. Adding redzones
1807 // to such globals is counterproductive, because the intent is that they
1808 // will form an array, and out-of-bounds accesses are expected.
1809 // See https://github.com/google/sanitizers/issues/305
1810 // and http://msdn.microsoft.com/en-US/en-en/library/bb918180(v=vs.120).aspx
1811 if (TargetTriple.isOSBinFormatCOFF() && Section.contains('$')) {
1812 LLVM_DEBUG(dbgs() << "Ignoring global in sorted section (contains '$'): "
1817 if (TargetTriple.isOSBinFormatMachO()) {
1818 StringRef ParsedSegment, ParsedSection;
1819 unsigned TAA = 0, StubSize = 0;
1821 cantFail(MCSectionMachO::ParseSectionSpecifier(
1822 Section, ParsedSegment, ParsedSection, TAA, TAAParsed, StubSize));
1824 // Ignore the globals from the __OBJC section. The ObjC runtime assumes
1825 // those conform to /usr/lib/objc/runtime.h, so we can't add redzones to
1827 if (ParsedSegment == "__OBJC" ||
1828 (ParsedSegment == "__DATA" && ParsedSection.startswith("__objc_"))) {
1829 LLVM_DEBUG(dbgs() << "Ignoring ObjC runtime global: " << *G << "\n");
1832 // See https://github.com/google/sanitizers/issues/32
1833 // Constant CFString instances are compiled in the following way:
1834 // -- the string buffer is emitted into
1835 // __TEXT,__cstring,cstring_literals
1836 // -- the constant NSConstantString structure referencing that buffer
1837 // is placed into __DATA,__cfstring
1838 // Therefore there's no point in placing redzones into __DATA,__cfstring.
1839 // Moreover, it causes the linker to crash on OS X 10.7
1840 if (ParsedSegment == "__DATA" && ParsedSection == "__cfstring") {
1841 LLVM_DEBUG(dbgs() << "Ignoring CFString: " << *G << "\n");
1844 // The linker merges the contents of cstring_literals and removes the
1846 if (ParsedSegment == "__TEXT" && (TAA & MachO::S_CSTRING_LITERALS)) {
1847 LLVM_DEBUG(dbgs() << "Ignoring a cstring literal: " << *G << "\n");
1853 if (CompileKernel) {
1854 // Globals that prefixed by "__" are special and cannot be padded with a
1856 if (G->getName().startswith("__"))
1863 // On Mach-O platforms, we emit global metadata in a separate section of the
1864 // binary in order to allow the linker to properly dead strip. This is only
1865 // supported on recent versions of ld64.
1866 bool ModuleAddressSanitizer::ShouldUseMachOGlobalsSection() const {
1867 if (!TargetTriple.isOSBinFormatMachO())
1870 if (TargetTriple.isMacOSX() && !TargetTriple.isMacOSXVersionLT(10, 11))
1872 if (TargetTriple.isiOS() /* or tvOS */ && !TargetTriple.isOSVersionLT(9))
1874 if (TargetTriple.isWatchOS() && !TargetTriple.isOSVersionLT(2))
1876 if (TargetTriple.isDriverKit())
1882 StringRef ModuleAddressSanitizer::getGlobalMetadataSection() const {
1883 switch (TargetTriple.getObjectFormat()) {
1884 case Triple::COFF: return ".ASAN$GL";
1885 case Triple::ELF: return "asan_globals";
1886 case Triple::MachO: return "__DATA,__asan_globals,regular";
1891 case Triple::DXContainer:
1893 "ModuleAddressSanitizer not implemented for object file format");
1894 case Triple::UnknownObjectFormat:
1897 llvm_unreachable("unsupported object format");
1900 void ModuleAddressSanitizer::initializeCallbacks(Module &M) {
1901 IRBuilder<> IRB(*C);
1903 // Declare our poisoning and unpoisoning functions.
1905 M.getOrInsertFunction(kAsanPoisonGlobalsName, IRB.getVoidTy(), IntptrTy);
1906 AsanUnpoisonGlobals =
1907 M.getOrInsertFunction(kAsanUnpoisonGlobalsName, IRB.getVoidTy());
1909 // Declare functions that register/unregister globals.
1910 AsanRegisterGlobals = M.getOrInsertFunction(
1911 kAsanRegisterGlobalsName, IRB.getVoidTy(), IntptrTy, IntptrTy);
1912 AsanUnregisterGlobals = M.getOrInsertFunction(
1913 kAsanUnregisterGlobalsName, IRB.getVoidTy(), IntptrTy, IntptrTy);
1915 // Declare the functions that find globals in a shared object and then invoke
1916 // the (un)register function on them.
1917 AsanRegisterImageGlobals = M.getOrInsertFunction(
1918 kAsanRegisterImageGlobalsName, IRB.getVoidTy(), IntptrTy);
1919 AsanUnregisterImageGlobals = M.getOrInsertFunction(
1920 kAsanUnregisterImageGlobalsName, IRB.getVoidTy(), IntptrTy);
1922 AsanRegisterElfGlobals =
1923 M.getOrInsertFunction(kAsanRegisterElfGlobalsName, IRB.getVoidTy(),
1924 IntptrTy, IntptrTy, IntptrTy);
1925 AsanUnregisterElfGlobals =
1926 M.getOrInsertFunction(kAsanUnregisterElfGlobalsName, IRB.getVoidTy(),
1927 IntptrTy, IntptrTy, IntptrTy);
1930 // Put the metadata and the instrumented global in the same group. This ensures
1931 // that the metadata is discarded if the instrumented global is discarded.
1932 void ModuleAddressSanitizer::SetComdatForGlobalMetadata(
1933 GlobalVariable *G, GlobalVariable *Metadata, StringRef InternalSuffix) {
1934 Module &M = *G->getParent();
1935 Comdat *C = G->getComdat();
1937 if (!G->hasName()) {
1938 // If G is unnamed, it must be internal. Give it an artificial name
1939 // so we can put it in a comdat.
1940 assert(G->hasLocalLinkage());
1941 G->setName(Twine(kAsanGenPrefix) + "_anon_global");
1944 if (!InternalSuffix.empty() && G->hasLocalLinkage()) {
1945 std::string Name = std::string(G->getName());
1946 Name += InternalSuffix;
1947 C = M.getOrInsertComdat(Name);
1949 C = M.getOrInsertComdat(G->getName());
1952 // Make this IMAGE_COMDAT_SELECT_NODUPLICATES on COFF. Also upgrade private
1953 // linkage to internal linkage so that a symbol table entry is emitted. This
1954 // is necessary in order to create the comdat group.
1955 if (TargetTriple.isOSBinFormatCOFF()) {
1956 C->setSelectionKind(Comdat::NoDeduplicate);
1957 if (G->hasPrivateLinkage())
1958 G->setLinkage(GlobalValue::InternalLinkage);
1963 assert(G->hasComdat());
1964 Metadata->setComdat(G->getComdat());
1967 // Create a separate metadata global and put it in the appropriate ASan
1968 // global registration section.
1970 ModuleAddressSanitizer::CreateMetadataGlobal(Module &M, Constant *Initializer,
1971 StringRef OriginalName) {
1972 auto Linkage = TargetTriple.isOSBinFormatMachO()
1973 ? GlobalVariable::InternalLinkage
1974 : GlobalVariable::PrivateLinkage;
1975 GlobalVariable *Metadata = new GlobalVariable(
1976 M, Initializer->getType(), false, Linkage, Initializer,
1977 Twine("__asan_global_") + GlobalValue::dropLLVMManglingEscape(OriginalName));
1978 Metadata->setSection(getGlobalMetadataSection());
1982 Instruction *ModuleAddressSanitizer::CreateAsanModuleDtor(Module &M) {
1983 AsanDtorFunction = Function::createWithDefaultAttr(
1984 FunctionType::get(Type::getVoidTy(*C), false),
1985 GlobalValue::InternalLinkage, 0, kAsanModuleDtorName, &M);
1986 AsanDtorFunction->addFnAttr(Attribute::NoUnwind);
1987 // Ensure Dtor cannot be discarded, even if in a comdat.
1988 appendToUsed(M, {AsanDtorFunction});
1989 BasicBlock *AsanDtorBB = BasicBlock::Create(*C, "", AsanDtorFunction);
1991 return ReturnInst::Create(*C, AsanDtorBB);
1994 void ModuleAddressSanitizer::InstrumentGlobalsCOFF(
1995 IRBuilder<> &IRB, Module &M, ArrayRef<GlobalVariable *> ExtendedGlobals,
1996 ArrayRef<Constant *> MetadataInitializers) {
1997 assert(ExtendedGlobals.size() == MetadataInitializers.size());
1998 auto &DL = M.getDataLayout();
2000 SmallVector<GlobalValue *, 16> MetadataGlobals(ExtendedGlobals.size());
2001 for (size_t i = 0; i < ExtendedGlobals.size(); i++) {
2002 Constant *Initializer = MetadataInitializers[i];
2003 GlobalVariable *G = ExtendedGlobals[i];
2004 GlobalVariable *Metadata =
2005 CreateMetadataGlobal(M, Initializer, G->getName());
2006 MDNode *MD = MDNode::get(M.getContext(), ValueAsMetadata::get(G));
2007 Metadata->setMetadata(LLVMContext::MD_associated, MD);
2008 MetadataGlobals[i] = Metadata;
2010 // The MSVC linker always inserts padding when linking incrementally. We
2011 // cope with that by aligning each struct to its size, which must be a power
2013 unsigned SizeOfGlobalStruct = DL.getTypeAllocSize(Initializer->getType());
2014 assert(isPowerOf2_32(SizeOfGlobalStruct) &&
2015 "global metadata will not be padded appropriately");
2016 Metadata->setAlignment(assumeAligned(SizeOfGlobalStruct));
2018 SetComdatForGlobalMetadata(G, Metadata, "");
2021 // Update llvm.compiler.used, adding the new metadata globals. This is
2022 // needed so that during LTO these variables stay alive.
2023 if (!MetadataGlobals.empty())
2024 appendToCompilerUsed(M, MetadataGlobals);
2027 void ModuleAddressSanitizer::InstrumentGlobalsELF(
2028 IRBuilder<> &IRB, Module &M, ArrayRef<GlobalVariable *> ExtendedGlobals,
2029 ArrayRef<Constant *> MetadataInitializers,
2030 const std::string &UniqueModuleId) {
2031 assert(ExtendedGlobals.size() == MetadataInitializers.size());
2033 // Putting globals in a comdat changes the semantic and potentially cause
2034 // false negative odr violations at link time. If odr indicators are used, we
2035 // keep the comdat sections, as link time odr violations will be dectected on
2036 // the odr indicator symbols.
2037 bool UseComdatForGlobalsGC = UseOdrIndicator;
2039 SmallVector<GlobalValue *, 16> MetadataGlobals(ExtendedGlobals.size());
2040 for (size_t i = 0; i < ExtendedGlobals.size(); i++) {
2041 GlobalVariable *G = ExtendedGlobals[i];
2042 GlobalVariable *Metadata =
2043 CreateMetadataGlobal(M, MetadataInitializers[i], G->getName());
2044 MDNode *MD = MDNode::get(M.getContext(), ValueAsMetadata::get(G));
2045 Metadata->setMetadata(LLVMContext::MD_associated, MD);
2046 MetadataGlobals[i] = Metadata;
2048 if (UseComdatForGlobalsGC)
2049 SetComdatForGlobalMetadata(G, Metadata, UniqueModuleId);
2052 // Update llvm.compiler.used, adding the new metadata globals. This is
2053 // needed so that during LTO these variables stay alive.
2054 if (!MetadataGlobals.empty())
2055 appendToCompilerUsed(M, MetadataGlobals);
2057 // RegisteredFlag serves two purposes. First, we can pass it to dladdr()
2058 // to look up the loaded image that contains it. Second, we can store in it
2059 // whether registration has already occurred, to prevent duplicate
2062 // Common linkage ensures that there is only one global per shared library.
2063 GlobalVariable *RegisteredFlag = new GlobalVariable(
2064 M, IntptrTy, false, GlobalVariable::CommonLinkage,
2065 ConstantInt::get(IntptrTy, 0), kAsanGlobalsRegisteredFlagName);
2066 RegisteredFlag->setVisibility(GlobalVariable::HiddenVisibility);
2068 // Create start and stop symbols.
2069 GlobalVariable *StartELFMetadata = new GlobalVariable(
2070 M, IntptrTy, false, GlobalVariable::ExternalWeakLinkage, nullptr,
2071 "__start_" + getGlobalMetadataSection());
2072 StartELFMetadata->setVisibility(GlobalVariable::HiddenVisibility);
2073 GlobalVariable *StopELFMetadata = new GlobalVariable(
2074 M, IntptrTy, false, GlobalVariable::ExternalWeakLinkage, nullptr,
2075 "__stop_" + getGlobalMetadataSection());
2076 StopELFMetadata->setVisibility(GlobalVariable::HiddenVisibility);
2078 // Create a call to register the globals with the runtime.
2079 IRB.CreateCall(AsanRegisterElfGlobals,
2080 {IRB.CreatePointerCast(RegisteredFlag, IntptrTy),
2081 IRB.CreatePointerCast(StartELFMetadata, IntptrTy),
2082 IRB.CreatePointerCast(StopELFMetadata, IntptrTy)});
2084 // We also need to unregister globals at the end, e.g., when a shared library
2086 if (DestructorKind != AsanDtorKind::None) {
2087 IRBuilder<> IrbDtor(CreateAsanModuleDtor(M));
2088 IrbDtor.CreateCall(AsanUnregisterElfGlobals,
2089 {IRB.CreatePointerCast(RegisteredFlag, IntptrTy),
2090 IRB.CreatePointerCast(StartELFMetadata, IntptrTy),
2091 IRB.CreatePointerCast(StopELFMetadata, IntptrTy)});
2095 void ModuleAddressSanitizer::InstrumentGlobalsMachO(
2096 IRBuilder<> &IRB, Module &M, ArrayRef<GlobalVariable *> ExtendedGlobals,
2097 ArrayRef<Constant *> MetadataInitializers) {
2098 assert(ExtendedGlobals.size() == MetadataInitializers.size());
2100 // On recent Mach-O platforms, use a structure which binds the liveness of
2101 // the global variable to the metadata struct. Keep the list of "Liveness" GV
2102 // created to be added to llvm.compiler.used
2103 StructType *LivenessTy = StructType::get(IntptrTy, IntptrTy);
2104 SmallVector<GlobalValue *, 16> LivenessGlobals(ExtendedGlobals.size());
2106 for (size_t i = 0; i < ExtendedGlobals.size(); i++) {
2107 Constant *Initializer = MetadataInitializers[i];
2108 GlobalVariable *G = ExtendedGlobals[i];
2109 GlobalVariable *Metadata =
2110 CreateMetadataGlobal(M, Initializer, G->getName());
2112 // On recent Mach-O platforms, we emit the global metadata in a way that
2113 // allows the linker to properly strip dead globals.
2114 auto LivenessBinder =
2115 ConstantStruct::get(LivenessTy, Initializer->getAggregateElement(0u),
2116 ConstantExpr::getPointerCast(Metadata, IntptrTy));
2117 GlobalVariable *Liveness = new GlobalVariable(
2118 M, LivenessTy, false, GlobalVariable::InternalLinkage, LivenessBinder,
2119 Twine("__asan_binder_") + G->getName());
2120 Liveness->setSection("__DATA,__asan_liveness,regular,live_support");
2121 LivenessGlobals[i] = Liveness;
2124 // Update llvm.compiler.used, adding the new liveness globals. This is
2125 // needed so that during LTO these variables stay alive. The alternative
2126 // would be to have the linker handling the LTO symbols, but libLTO
2127 // current API does not expose access to the section for each symbol.
2128 if (!LivenessGlobals.empty())
2129 appendToCompilerUsed(M, LivenessGlobals);
2131 // RegisteredFlag serves two purposes. First, we can pass it to dladdr()
2132 // to look up the loaded image that contains it. Second, we can store in it
2133 // whether registration has already occurred, to prevent duplicate
2136 // common linkage ensures that there is only one global per shared library.
2137 GlobalVariable *RegisteredFlag = new GlobalVariable(
2138 M, IntptrTy, false, GlobalVariable::CommonLinkage,
2139 ConstantInt::get(IntptrTy, 0), kAsanGlobalsRegisteredFlagName);
2140 RegisteredFlag->setVisibility(GlobalVariable::HiddenVisibility);
2142 IRB.CreateCall(AsanRegisterImageGlobals,
2143 {IRB.CreatePointerCast(RegisteredFlag, IntptrTy)});
2145 // We also need to unregister globals at the end, e.g., when a shared library
2147 if (DestructorKind != AsanDtorKind::None) {
2148 IRBuilder<> IrbDtor(CreateAsanModuleDtor(M));
2149 IrbDtor.CreateCall(AsanUnregisterImageGlobals,
2150 {IRB.CreatePointerCast(RegisteredFlag, IntptrTy)});
2154 void ModuleAddressSanitizer::InstrumentGlobalsWithMetadataArray(
2155 IRBuilder<> &IRB, Module &M, ArrayRef<GlobalVariable *> ExtendedGlobals,
2156 ArrayRef<Constant *> MetadataInitializers) {
2157 assert(ExtendedGlobals.size() == MetadataInitializers.size());
2158 unsigned N = ExtendedGlobals.size();
2161 // On platforms that don't have a custom metadata section, we emit an array
2162 // of global metadata structures.
2163 ArrayType *ArrayOfGlobalStructTy =
2164 ArrayType::get(MetadataInitializers[0]->getType(), N);
2165 auto AllGlobals = new GlobalVariable(
2166 M, ArrayOfGlobalStructTy, false, GlobalVariable::InternalLinkage,
2167 ConstantArray::get(ArrayOfGlobalStructTy, MetadataInitializers), "");
2168 if (Mapping.Scale > 3)
2169 AllGlobals->setAlignment(Align(1ULL << Mapping.Scale));
2171 IRB.CreateCall(AsanRegisterGlobals,
2172 {IRB.CreatePointerCast(AllGlobals, IntptrTy),
2173 ConstantInt::get(IntptrTy, N)});
2175 // We also need to unregister globals at the end, e.g., when a shared library
2177 if (DestructorKind != AsanDtorKind::None) {
2178 IRBuilder<> IrbDtor(CreateAsanModuleDtor(M));
2179 IrbDtor.CreateCall(AsanUnregisterGlobals,
2180 {IRB.CreatePointerCast(AllGlobals, IntptrTy),
2181 ConstantInt::get(IntptrTy, N)});
2185 // This function replaces all global variables with new variables that have
2186 // trailing redzones. It also creates a function that poisons
2187 // redzones and inserts this function into llvm.global_ctors.
2188 // Sets *CtorComdat to true if the global registration code emitted into the
2189 // asan constructor is comdat-compatible.
2190 bool ModuleAddressSanitizer::InstrumentGlobals(IRBuilder<> &IRB, Module &M,
2192 *CtorComdat = false;
2194 // Build set of globals that are aliased by some GA, where
2195 // getExcludedAliasedGlobal(GA) returns the relevant GlobalVariable.
2196 SmallPtrSet<const GlobalVariable *, 16> AliasedGlobalExclusions;
2197 if (CompileKernel) {
2198 for (auto &GA : M.aliases()) {
2199 if (const GlobalVariable *GV = getExcludedAliasedGlobal(GA))
2200 AliasedGlobalExclusions.insert(GV);
2204 SmallVector<GlobalVariable *, 16> GlobalsToChange;
2205 for (auto &G : M.globals()) {
2206 if (!AliasedGlobalExclusions.count(&G) && shouldInstrumentGlobal(&G))
2207 GlobalsToChange.push_back(&G);
2210 size_t n = GlobalsToChange.size();
2216 auto &DL = M.getDataLayout();
2218 // A global is described by a structure
2221 // size_t size_with_redzone;
2222 // const char *name;
2223 // const char *module_name;
2224 // size_t has_dynamic_init;
2225 // size_t padding_for_windows_msvc_incremental_link;
2226 // size_t odr_indicator;
2227 // We initialize an array of such structures and pass it to a run-time call.
2228 StructType *GlobalStructTy =
2229 StructType::get(IntptrTy, IntptrTy, IntptrTy, IntptrTy, IntptrTy,
2230 IntptrTy, IntptrTy, IntptrTy);
2231 SmallVector<GlobalVariable *, 16> NewGlobals(n);
2232 SmallVector<Constant *, 16> Initializers(n);
2234 bool HasDynamicallyInitializedGlobals = false;
2236 // We shouldn't merge same module names, as this string serves as unique
2237 // module ID in runtime.
2238 GlobalVariable *ModuleName = createPrivateGlobalForString(
2239 M, M.getModuleIdentifier(), /*AllowMerging*/ false, kAsanGenPrefix);
2241 for (size_t i = 0; i < n; i++) {
2242 GlobalVariable *G = GlobalsToChange[i];
2244 GlobalValue::SanitizerMetadata MD;
2245 if (G->hasSanitizerMetadata())
2246 MD = G->getSanitizerMetadata();
2248 // TODO: Symbol names in the descriptor can be demangled by the runtime
2249 // library. This could save ~0.4% of VM size for a private large binary.
2250 std::string NameForGlobal = llvm::demangle(G->getName().str());
2251 GlobalVariable *Name =
2252 createPrivateGlobalForString(M, NameForGlobal,
2253 /*AllowMerging*/ true, kAsanGenPrefix);
2255 Type *Ty = G->getValueType();
2256 const uint64_t SizeInBytes = DL.getTypeAllocSize(Ty);
2257 const uint64_t RightRedzoneSize = getRedzoneSizeForGlobal(SizeInBytes);
2258 Type *RightRedZoneTy = ArrayType::get(IRB.getInt8Ty(), RightRedzoneSize);
2260 StructType *NewTy = StructType::get(Ty, RightRedZoneTy);
2261 Constant *NewInitializer = ConstantStruct::get(
2262 NewTy, G->getInitializer(), Constant::getNullValue(RightRedZoneTy));
2264 // Create a new global variable with enough space for a redzone.
2265 GlobalValue::LinkageTypes Linkage = G->getLinkage();
2266 if (G->isConstant() && Linkage == GlobalValue::PrivateLinkage)
2267 Linkage = GlobalValue::InternalLinkage;
2268 GlobalVariable *NewGlobal = new GlobalVariable(
2269 M, NewTy, G->isConstant(), Linkage, NewInitializer, "", G,
2270 G->getThreadLocalMode(), G->getAddressSpace());
2271 NewGlobal->copyAttributesFrom(G);
2272 NewGlobal->setComdat(G->getComdat());
2273 NewGlobal->setAlignment(MaybeAlign(getMinRedzoneSizeForGlobal()));
2274 // Don't fold globals with redzones. ODR violation detector and redzone
2275 // poisoning implicitly creates a dependence on the global's address, so it
2276 // is no longer valid for it to be marked unnamed_addr.
2277 NewGlobal->setUnnamedAddr(GlobalValue::UnnamedAddr::None);
2279 // Move null-terminated C strings to "__asan_cstring" section on Darwin.
2280 if (TargetTriple.isOSBinFormatMachO() && !G->hasSection() &&
2282 auto Seq = dyn_cast<ConstantDataSequential>(G->getInitializer());
2283 if (Seq && Seq->isCString())
2284 NewGlobal->setSection("__TEXT,__asan_cstring,regular");
2287 // Transfer the debug info and type metadata. The payload starts at offset
2288 // zero so we can copy the metadata over as is.
2289 NewGlobal->copyMetadata(G, 0);
2292 Indices2[0] = IRB.getInt32(0);
2293 Indices2[1] = IRB.getInt32(0);
2295 G->replaceAllUsesWith(
2296 ConstantExpr::getGetElementPtr(NewTy, NewGlobal, Indices2, true));
2297 NewGlobal->takeName(G);
2298 G->eraseFromParent();
2299 NewGlobals[i] = NewGlobal;
2301 Constant *ODRIndicator = ConstantExpr::getNullValue(IRB.getInt8PtrTy());
2302 GlobalValue *InstrumentedGlobal = NewGlobal;
2304 bool CanUsePrivateAliases =
2305 TargetTriple.isOSBinFormatELF() || TargetTriple.isOSBinFormatMachO() ||
2306 TargetTriple.isOSBinFormatWasm();
2307 if (CanUsePrivateAliases && UsePrivateAlias) {
2308 // Create local alias for NewGlobal to avoid crash on ODR between
2309 // instrumented and non-instrumented libraries.
2310 InstrumentedGlobal =
2311 GlobalAlias::create(GlobalValue::PrivateLinkage, "", NewGlobal);
2314 // ODR should not happen for local linkage.
2315 if (NewGlobal->hasLocalLinkage()) {
2316 ODRIndicator = ConstantExpr::getIntToPtr(ConstantInt::get(IntptrTy, -1),
2317 IRB.getInt8PtrTy());
2318 } else if (UseOdrIndicator) {
2319 // With local aliases, we need to provide another externally visible
2320 // symbol __odr_asan_XXX to detect ODR violation.
2321 auto *ODRIndicatorSym =
2322 new GlobalVariable(M, IRB.getInt8Ty(), false, Linkage,
2323 Constant::getNullValue(IRB.getInt8Ty()),
2324 kODRGenPrefix + NameForGlobal, nullptr,
2325 NewGlobal->getThreadLocalMode());
2327 // Set meaningful attributes for indicator symbol.
2328 ODRIndicatorSym->setVisibility(NewGlobal->getVisibility());
2329 ODRIndicatorSym->setDLLStorageClass(NewGlobal->getDLLStorageClass());
2330 ODRIndicatorSym->setAlignment(Align(1));
2331 ODRIndicator = ODRIndicatorSym;
2334 Constant *Initializer = ConstantStruct::get(
2336 ConstantExpr::getPointerCast(InstrumentedGlobal, IntptrTy),
2337 ConstantInt::get(IntptrTy, SizeInBytes),
2338 ConstantInt::get(IntptrTy, SizeInBytes + RightRedzoneSize),
2339 ConstantExpr::getPointerCast(Name, IntptrTy),
2340 ConstantExpr::getPointerCast(ModuleName, IntptrTy),
2341 ConstantInt::get(IntptrTy, MD.IsDynInit),
2342 Constant::getNullValue(IntptrTy),
2343 ConstantExpr::getPointerCast(ODRIndicator, IntptrTy));
2345 if (ClInitializers && MD.IsDynInit)
2346 HasDynamicallyInitializedGlobals = true;
2348 LLVM_DEBUG(dbgs() << "NEW GLOBAL: " << *NewGlobal << "\n");
2350 Initializers[i] = Initializer;
2353 // Add instrumented globals to llvm.compiler.used list to avoid LTO from
2354 // ConstantMerge'ing them.
2355 SmallVector<GlobalValue *, 16> GlobalsToAddToUsedList;
2356 for (size_t i = 0; i < n; i++) {
2357 GlobalVariable *G = NewGlobals[i];
2358 if (G->getName().empty()) continue;
2359 GlobalsToAddToUsedList.push_back(G);
2361 appendToCompilerUsed(M, ArrayRef<GlobalValue *>(GlobalsToAddToUsedList));
2363 std::string ELFUniqueModuleId =
2364 (UseGlobalsGC && TargetTriple.isOSBinFormatELF()) ? getUniqueModuleId(&M)
2367 if (!ELFUniqueModuleId.empty()) {
2368 InstrumentGlobalsELF(IRB, M, NewGlobals, Initializers, ELFUniqueModuleId);
2370 } else if (UseGlobalsGC && TargetTriple.isOSBinFormatCOFF()) {
2371 InstrumentGlobalsCOFF(IRB, M, NewGlobals, Initializers);
2372 } else if (UseGlobalsGC && ShouldUseMachOGlobalsSection()) {
2373 InstrumentGlobalsMachO(IRB, M, NewGlobals, Initializers);
2375 InstrumentGlobalsWithMetadataArray(IRB, M, NewGlobals, Initializers);
2378 // Create calls for poisoning before initializers run and unpoisoning after.
2379 if (HasDynamicallyInitializedGlobals)
2380 createInitializerPoisonCalls(M, ModuleName);
2382 LLVM_DEBUG(dbgs() << M);
2387 ModuleAddressSanitizer::getRedzoneSizeForGlobal(uint64_t SizeInBytes) const {
2388 constexpr uint64_t kMaxRZ = 1 << 18;
2389 const uint64_t MinRZ = getMinRedzoneSizeForGlobal();
2392 if (SizeInBytes <= MinRZ / 2) {
2393 // Reduce redzone size for small size objects, e.g. int, char[1]. MinRZ is
2394 // at least 32 bytes, optimize when SizeInBytes is less than or equal to
2396 RZ = MinRZ - SizeInBytes;
2398 // Calculate RZ, where MinRZ <= RZ <= MaxRZ, and RZ ~ 1/4 * SizeInBytes.
2399 RZ = std::max(MinRZ, std::min(kMaxRZ, (SizeInBytes / MinRZ / 4) * MinRZ));
2401 // Round up to multiple of MinRZ.
2402 if (SizeInBytes % MinRZ)
2403 RZ += MinRZ - (SizeInBytes % MinRZ);
2406 assert((RZ + SizeInBytes) % MinRZ == 0);
2411 int ModuleAddressSanitizer::GetAsanVersion(const Module &M) const {
2412 int LongSize = M.getDataLayout().getPointerSizeInBits();
2413 bool isAndroid = Triple(M.getTargetTriple()).isAndroid();
2415 // 32-bit Android is one version ahead because of the switch to dynamic
2417 Version += (LongSize == 32 && isAndroid);
2421 bool ModuleAddressSanitizer::instrumentModule(Module &M) {
2422 initializeCallbacks(M);
2424 // Create a module constructor. A destructor is created lazily because not all
2425 // platforms, and not all modules need it.
2426 if (CompileKernel) {
2427 // The kernel always builds with its own runtime, and therefore does not
2428 // need the init and version check calls.
2429 AsanCtorFunction = createSanitizerCtor(M, kAsanModuleCtorName);
2431 std::string AsanVersion = std::to_string(GetAsanVersion(M));
2432 std::string VersionCheckName =
2433 ClInsertVersionCheck ? (kAsanVersionCheckNamePrefix + AsanVersion) : "";
2434 std::tie(AsanCtorFunction, std::ignore) =
2435 createSanitizerCtorAndInitFunctions(M, kAsanModuleCtorName,
2436 kAsanInitName, /*InitArgTypes=*/{},
2437 /*InitArgs=*/{}, VersionCheckName);
2440 bool CtorComdat = true;
2442 IRBuilder<> IRB(AsanCtorFunction->getEntryBlock().getTerminator());
2443 InstrumentGlobals(IRB, M, &CtorComdat);
2446 const uint64_t Priority = GetCtorAndDtorPriority(TargetTriple);
2448 // Put the constructor and destructor in comdat if both
2449 // (1) global instrumentation is not TU-specific
2450 // (2) target is ELF.
2451 if (UseCtorComdat && TargetTriple.isOSBinFormatELF() && CtorComdat) {
2452 AsanCtorFunction->setComdat(M.getOrInsertComdat(kAsanModuleCtorName));
2453 appendToGlobalCtors(M, AsanCtorFunction, Priority, AsanCtorFunction);
2454 if (AsanDtorFunction) {
2455 AsanDtorFunction->setComdat(M.getOrInsertComdat(kAsanModuleDtorName));
2456 appendToGlobalDtors(M, AsanDtorFunction, Priority, AsanDtorFunction);
2459 appendToGlobalCtors(M, AsanCtorFunction, Priority);
2460 if (AsanDtorFunction)
2461 appendToGlobalDtors(M, AsanDtorFunction, Priority);
2467 void AddressSanitizer::initializeCallbacks(Module &M) {
2468 IRBuilder<> IRB(*C);
2469 // Create __asan_report* callbacks.
2470 // IsWrite, TypeSize and Exp are encoded in the function name.
2471 for (int Exp = 0; Exp < 2; Exp++) {
2472 for (size_t AccessIsWrite = 0; AccessIsWrite <= 1; AccessIsWrite++) {
2473 const std::string TypeStr = AccessIsWrite ? "store" : "load";
2474 const std::string ExpStr = Exp ? "exp_" : "";
2475 const std::string EndingStr = Recover ? "_noabort" : "";
2477 SmallVector<Type *, 3> Args2 = {IntptrTy, IntptrTy};
2478 SmallVector<Type *, 2> Args1{1, IntptrTy};
2480 Type *ExpType = Type::getInt32Ty(*C);
2481 Args2.push_back(ExpType);
2482 Args1.push_back(ExpType);
2484 AsanErrorCallbackSized[AccessIsWrite][Exp] = M.getOrInsertFunction(
2485 kAsanReportErrorTemplate + ExpStr + TypeStr + "_n" + EndingStr,
2486 FunctionType::get(IRB.getVoidTy(), Args2, false));
2488 AsanMemoryAccessCallbackSized[AccessIsWrite][Exp] = M.getOrInsertFunction(
2489 ClMemoryAccessCallbackPrefix + ExpStr + TypeStr + "N" + EndingStr,
2490 FunctionType::get(IRB.getVoidTy(), Args2, false));
2492 for (size_t AccessSizeIndex = 0; AccessSizeIndex < kNumberOfAccessSizes;
2493 AccessSizeIndex++) {
2494 const std::string Suffix = TypeStr + itostr(1ULL << AccessSizeIndex);
2495 AsanErrorCallback[AccessIsWrite][Exp][AccessSizeIndex] =
2496 M.getOrInsertFunction(
2497 kAsanReportErrorTemplate + ExpStr + Suffix + EndingStr,
2498 FunctionType::get(IRB.getVoidTy(), Args1, false));
2500 AsanMemoryAccessCallback[AccessIsWrite][Exp][AccessSizeIndex] =
2501 M.getOrInsertFunction(
2502 ClMemoryAccessCallbackPrefix + ExpStr + Suffix + EndingStr,
2503 FunctionType::get(IRB.getVoidTy(), Args1, false));
2508 const std::string MemIntrinCallbackPrefix =
2509 (CompileKernel && !ClKasanMemIntrinCallbackPrefix)
2511 : ClMemoryAccessCallbackPrefix;
2512 AsanMemmove = M.getOrInsertFunction(MemIntrinCallbackPrefix + "memmove",
2513 IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
2514 IRB.getInt8PtrTy(), IntptrTy);
2515 AsanMemcpy = M.getOrInsertFunction(MemIntrinCallbackPrefix + "memcpy",
2516 IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
2517 IRB.getInt8PtrTy(), IntptrTy);
2518 AsanMemset = M.getOrInsertFunction(MemIntrinCallbackPrefix + "memset",
2519 IRB.getInt8PtrTy(), IRB.getInt8PtrTy(),
2520 IRB.getInt32Ty(), IntptrTy);
2522 AsanHandleNoReturnFunc =
2523 M.getOrInsertFunction(kAsanHandleNoReturnName, IRB.getVoidTy());
2525 AsanPtrCmpFunction =
2526 M.getOrInsertFunction(kAsanPtrCmp, IRB.getVoidTy(), IntptrTy, IntptrTy);
2527 AsanPtrSubFunction =
2528 M.getOrInsertFunction(kAsanPtrSub, IRB.getVoidTy(), IntptrTy, IntptrTy);
2529 if (Mapping.InGlobal)
2530 AsanShadowGlobal = M.getOrInsertGlobal("__asan_shadow",
2531 ArrayType::get(IRB.getInt8Ty(), 0));
2533 AMDGPUAddressShared = M.getOrInsertFunction(
2534 kAMDGPUAddressSharedName, IRB.getInt1Ty(), IRB.getInt8PtrTy());
2535 AMDGPUAddressPrivate = M.getOrInsertFunction(
2536 kAMDGPUAddressPrivateName, IRB.getInt1Ty(), IRB.getInt8PtrTy());
2539 bool AddressSanitizer::maybeInsertAsanInitAtFunctionEntry(Function &F) {
2540 // For each NSObject descendant having a +load method, this method is invoked
2541 // by the ObjC runtime before any of the static constructors is called.
2542 // Therefore we need to instrument such methods with a call to __asan_init
2543 // at the beginning in order to initialize our runtime before any access to
2544 // the shadow memory.
2545 // We cannot just ignore these methods, because they may call other
2546 // instrumented functions.
2547 if (F.getName().find(" load]") != std::string::npos) {
2548 FunctionCallee AsanInitFunction =
2549 declareSanitizerInitFunction(*F.getParent(), kAsanInitName, {});
2550 IRBuilder<> IRB(&F.front(), F.front().begin());
2551 IRB.CreateCall(AsanInitFunction, {});
2557 bool AddressSanitizer::maybeInsertDynamicShadowAtFunctionEntry(Function &F) {
2558 // Generate code only when dynamic addressing is needed.
2559 if (Mapping.Offset != kDynamicShadowSentinel)
2562 IRBuilder<> IRB(&F.front().front());
2563 if (Mapping.InGlobal) {
2564 if (ClWithIfuncSuppressRemat) {
2565 // An empty inline asm with input reg == output reg.
2566 // An opaque pointer-to-int cast, basically.
2567 InlineAsm *Asm = InlineAsm::get(
2568 FunctionType::get(IntptrTy, {AsanShadowGlobal->getType()}, false),
2569 StringRef(""), StringRef("=r,0"),
2570 /*hasSideEffects=*/false);
2571 LocalDynamicShadow =
2572 IRB.CreateCall(Asm, {AsanShadowGlobal}, ".asan.shadow");
2574 LocalDynamicShadow =
2575 IRB.CreatePointerCast(AsanShadowGlobal, IntptrTy, ".asan.shadow");
2578 Value *GlobalDynamicAddress = F.getParent()->getOrInsertGlobal(
2579 kAsanShadowMemoryDynamicAddress, IntptrTy);
2580 LocalDynamicShadow = IRB.CreateLoad(IntptrTy, GlobalDynamicAddress);
2585 void AddressSanitizer::markEscapedLocalAllocas(Function &F) {
2586 // Find the one possible call to llvm.localescape and pre-mark allocas passed
2587 // to it as uninteresting. This assumes we haven't started processing allocas
2588 // yet. This check is done up front because iterating the use list in
2589 // isInterestingAlloca would be algorithmically slower.
2590 assert(ProcessedAllocas.empty() && "must process localescape before allocas");
2592 // Try to get the declaration of llvm.localescape. If it's not in the module,
2593 // we can exit early.
2594 if (!F.getParent()->getFunction("llvm.localescape")) return;
2596 // Look for a call to llvm.localescape call in the entry block. It can't be in
2598 for (Instruction &I : F.getEntryBlock()) {
2599 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&I);
2600 if (II && II->getIntrinsicID() == Intrinsic::localescape) {
2601 // We found a call. Mark all the allocas passed in as uninteresting.
2602 for (Value *Arg : II->args()) {
2603 AllocaInst *AI = dyn_cast<AllocaInst>(Arg->stripPointerCasts());
2604 assert(AI && AI->isStaticAlloca() &&
2605 "non-static alloca arg to localescape");
2606 ProcessedAllocas[AI] = false;
2613 bool AddressSanitizer::suppressInstrumentationSiteForDebug(int &Instrumented) {
2614 bool ShouldInstrument =
2615 ClDebugMin < 0 || ClDebugMax < 0 ||
2616 (Instrumented >= ClDebugMin && Instrumented <= ClDebugMax);
2618 return !ShouldInstrument;
2621 bool AddressSanitizer::instrumentFunction(Function &F,
2622 const TargetLibraryInfo *TLI) {
2625 if (F.getLinkage() == GlobalValue::AvailableExternallyLinkage) return false;
2626 if (!ClDebugFunc.empty() && ClDebugFunc == F.getName()) return false;
2627 if (F.getName().startswith("__asan_")) return false;
2629 bool FunctionModified = false;
2631 // If needed, insert __asan_init before checking for SanitizeAddress attr.
2632 // This function needs to be called even if the function body is not
2634 if (maybeInsertAsanInitAtFunctionEntry(F))
2635 FunctionModified = true;
2637 // Leave if the function doesn't need instrumentation.
2638 if (!F.hasFnAttribute(Attribute::SanitizeAddress)) return FunctionModified;
2640 if (F.hasFnAttribute(Attribute::DisableSanitizerInstrumentation))
2641 return FunctionModified;
2643 LLVM_DEBUG(dbgs() << "ASAN instrumenting:\n" << F << "\n");
2645 initializeCallbacks(*F.getParent());
2647 FunctionStateRAII CleanupObj(this);
2649 FunctionModified |= maybeInsertDynamicShadowAtFunctionEntry(F);
2651 // We can't instrument allocas used with llvm.localescape. Only static allocas
2652 // can be passed to that intrinsic.
2653 markEscapedLocalAllocas(F);
2655 // We want to instrument every address only once per basic block (unless there
2656 // are calls between uses).
2657 SmallPtrSet<Value *, 16> TempsToInstrument;
2658 SmallVector<InterestingMemoryOperand, 16> OperandsToInstrument;
2659 SmallVector<MemIntrinsic *, 16> IntrinToInstrument;
2660 SmallVector<Instruction *, 8> NoReturnCalls;
2661 SmallVector<BasicBlock *, 16> AllBlocks;
2662 SmallVector<Instruction *, 16> PointerComparisonsOrSubtracts;
2664 // Fill the set of memory operations to instrument.
2665 for (auto &BB : F) {
2666 AllBlocks.push_back(&BB);
2667 TempsToInstrument.clear();
2668 int NumInsnsPerBB = 0;
2669 for (auto &Inst : BB) {
2670 if (LooksLikeCodeInBug11395(&Inst)) return false;
2671 // Skip instructions inserted by another instrumentation.
2672 if (Inst.hasMetadata(LLVMContext::MD_nosanitize))
2674 SmallVector<InterestingMemoryOperand, 1> InterestingOperands;
2675 getInterestingMemoryOperands(&Inst, InterestingOperands);
2677 if (!InterestingOperands.empty()) {
2678 for (auto &Operand : InterestingOperands) {
2679 if (ClOpt && ClOptSameTemp) {
2680 Value *Ptr = Operand.getPtr();
2681 // If we have a mask, skip instrumentation if we've already
2682 // instrumented the full object. But don't add to TempsToInstrument
2683 // because we might get another load/store with a different mask.
2684 if (Operand.MaybeMask) {
2685 if (TempsToInstrument.count(Ptr))
2686 continue; // We've seen this (whole) temp in the current BB.
2688 if (!TempsToInstrument.insert(Ptr).second)
2689 continue; // We've seen this temp in the current BB.
2692 OperandsToInstrument.push_back(Operand);
2695 } else if (((ClInvalidPointerPairs || ClInvalidPointerCmp) &&
2696 isInterestingPointerComparison(&Inst)) ||
2697 ((ClInvalidPointerPairs || ClInvalidPointerSub) &&
2698 isInterestingPointerSubtraction(&Inst))) {
2699 PointerComparisonsOrSubtracts.push_back(&Inst);
2700 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(&Inst)) {
2702 IntrinToInstrument.push_back(MI);
2705 if (auto *CB = dyn_cast<CallBase>(&Inst)) {
2706 // A call inside BB.
2707 TempsToInstrument.clear();
2708 if (CB->doesNotReturn())
2709 NoReturnCalls.push_back(CB);
2711 if (CallInst *CI = dyn_cast<CallInst>(&Inst))
2712 maybeMarkSanitizerLibraryCallNoBuiltin(CI, TLI);
2714 if (NumInsnsPerBB >= ClMaxInsnsToInstrumentPerBB) break;
2718 bool UseCalls = (ClInstrumentationWithCallsThreshold >= 0 &&
2719 OperandsToInstrument.size() + IntrinToInstrument.size() >
2720 (unsigned)ClInstrumentationWithCallsThreshold);
2721 const DataLayout &DL = F.getParent()->getDataLayout();
2722 ObjectSizeOpts ObjSizeOpts;
2723 ObjSizeOpts.RoundToAlign = true;
2724 ObjectSizeOffsetVisitor ObjSizeVis(DL, TLI, F.getContext(), ObjSizeOpts);
2727 int NumInstrumented = 0;
2728 for (auto &Operand : OperandsToInstrument) {
2729 if (!suppressInstrumentationSiteForDebug(NumInstrumented))
2730 instrumentMop(ObjSizeVis, Operand, UseCalls,
2731 F.getParent()->getDataLayout());
2732 FunctionModified = true;
2734 for (auto Inst : IntrinToInstrument) {
2735 if (!suppressInstrumentationSiteForDebug(NumInstrumented))
2736 instrumentMemIntrinsic(Inst);
2737 FunctionModified = true;
2740 FunctionStackPoisoner FSP(F, *this);
2741 bool ChangedStack = FSP.runOnFunction();
2743 // We must unpoison the stack before NoReturn calls (throw, _exit, etc).
2744 // See e.g. https://github.com/google/sanitizers/issues/37
2745 for (auto CI : NoReturnCalls) {
2746 IRBuilder<> IRB(CI);
2747 IRB.CreateCall(AsanHandleNoReturnFunc, {});
2750 for (auto Inst : PointerComparisonsOrSubtracts) {
2751 instrumentPointerComparisonOrSubtraction(Inst);
2752 FunctionModified = true;
2755 if (ChangedStack || !NoReturnCalls.empty())
2756 FunctionModified = true;
2758 LLVM_DEBUG(dbgs() << "ASAN done instrumenting: " << FunctionModified << " "
2761 return FunctionModified;
2764 // Workaround for bug 11395: we don't want to instrument stack in functions
2765 // with large assembly blobs (32-bit only), otherwise reg alloc may crash.
2766 // FIXME: remove once the bug 11395 is fixed.
2767 bool AddressSanitizer::LooksLikeCodeInBug11395(Instruction *I) {
2768 if (LongSize != 32) return false;
2769 CallInst *CI = dyn_cast<CallInst>(I);
2770 if (!CI || !CI->isInlineAsm()) return false;
2771 if (CI->arg_size() <= 5)
2773 // We have inline assembly with quite a few arguments.
2777 void FunctionStackPoisoner::initializeCallbacks(Module &M) {
2778 IRBuilder<> IRB(*C);
2779 if (ASan.UseAfterReturn == AsanDetectStackUseAfterReturnMode::Always ||
2780 ASan.UseAfterReturn == AsanDetectStackUseAfterReturnMode::Runtime) {
2781 const char *MallocNameTemplate =
2782 ASan.UseAfterReturn == AsanDetectStackUseAfterReturnMode::Always
2783 ? kAsanStackMallocAlwaysNameTemplate
2784 : kAsanStackMallocNameTemplate;
2785 for (int Index = 0; Index <= kMaxAsanStackMallocSizeClass; Index++) {
2786 std::string Suffix = itostr(Index);
2787 AsanStackMallocFunc[Index] = M.getOrInsertFunction(
2788 MallocNameTemplate + Suffix, IntptrTy, IntptrTy);
2789 AsanStackFreeFunc[Index] =
2790 M.getOrInsertFunction(kAsanStackFreeNameTemplate + Suffix,
2791 IRB.getVoidTy(), IntptrTy, IntptrTy);
2794 if (ASan.UseAfterScope) {
2795 AsanPoisonStackMemoryFunc = M.getOrInsertFunction(
2796 kAsanPoisonStackMemoryName, IRB.getVoidTy(), IntptrTy, IntptrTy);
2797 AsanUnpoisonStackMemoryFunc = M.getOrInsertFunction(
2798 kAsanUnpoisonStackMemoryName, IRB.getVoidTy(), IntptrTy, IntptrTy);
2801 for (size_t Val : {0x00, 0xf1, 0xf2, 0xf3, 0xf5, 0xf8}) {
2802 std::ostringstream Name;
2803 Name << kAsanSetShadowPrefix;
2804 Name << std::setw(2) << std::setfill('0') << std::hex << Val;
2805 AsanSetShadowFunc[Val] =
2806 M.getOrInsertFunction(Name.str(), IRB.getVoidTy(), IntptrTy, IntptrTy);
2809 AsanAllocaPoisonFunc = M.getOrInsertFunction(
2810 kAsanAllocaPoison, IRB.getVoidTy(), IntptrTy, IntptrTy);
2811 AsanAllocasUnpoisonFunc = M.getOrInsertFunction(
2812 kAsanAllocasUnpoison, IRB.getVoidTy(), IntptrTy, IntptrTy);
2815 void FunctionStackPoisoner::copyToShadowInline(ArrayRef<uint8_t> ShadowMask,
2816 ArrayRef<uint8_t> ShadowBytes,
2817 size_t Begin, size_t End,
2819 Value *ShadowBase) {
2823 const size_t LargestStoreSizeInBytes =
2824 std::min<size_t>(sizeof(uint64_t), ASan.LongSize / 8);
2826 const bool IsLittleEndian = F.getParent()->getDataLayout().isLittleEndian();
2828 // Poison given range in shadow using larges store size with out leading and
2829 // trailing zeros in ShadowMask. Zeros never change, so they need neither
2830 // poisoning nor up-poisoning. Still we don't mind if some of them get into a
2831 // middle of a store.
2832 for (size_t i = Begin; i < End;) {
2833 if (!ShadowMask[i]) {
2834 assert(!ShadowBytes[i]);
2839 size_t StoreSizeInBytes = LargestStoreSizeInBytes;
2840 // Fit store size into the range.
2841 while (StoreSizeInBytes > End - i)
2842 StoreSizeInBytes /= 2;
2844 // Minimize store size by trimming trailing zeros.
2845 for (size_t j = StoreSizeInBytes - 1; j && !ShadowMask[i + j]; --j) {
2846 while (j <= StoreSizeInBytes / 2)
2847 StoreSizeInBytes /= 2;
2851 for (size_t j = 0; j < StoreSizeInBytes; j++) {
2853 Val |= (uint64_t)ShadowBytes[i + j] << (8 * j);
2855 Val = (Val << 8) | ShadowBytes[i + j];
2858 Value *Ptr = IRB.CreateAdd(ShadowBase, ConstantInt::get(IntptrTy, i));
2859 Value *Poison = IRB.getIntN(StoreSizeInBytes * 8, Val);
2860 IRB.CreateAlignedStore(
2861 Poison, IRB.CreateIntToPtr(Ptr, Poison->getType()->getPointerTo()),
2864 i += StoreSizeInBytes;
2868 void FunctionStackPoisoner::copyToShadow(ArrayRef<uint8_t> ShadowMask,
2869 ArrayRef<uint8_t> ShadowBytes,
2870 IRBuilder<> &IRB, Value *ShadowBase) {
2871 copyToShadow(ShadowMask, ShadowBytes, 0, ShadowMask.size(), IRB, ShadowBase);
2874 void FunctionStackPoisoner::copyToShadow(ArrayRef<uint8_t> ShadowMask,
2875 ArrayRef<uint8_t> ShadowBytes,
2876 size_t Begin, size_t End,
2877 IRBuilder<> &IRB, Value *ShadowBase) {
2878 assert(ShadowMask.size() == ShadowBytes.size());
2879 size_t Done = Begin;
2880 for (size_t i = Begin, j = Begin + 1; i < End; i = j++) {
2881 if (!ShadowMask[i]) {
2882 assert(!ShadowBytes[i]);
2885 uint8_t Val = ShadowBytes[i];
2886 if (!AsanSetShadowFunc[Val])
2889 // Skip same values.
2890 for (; j < End && ShadowMask[j] && Val == ShadowBytes[j]; ++j) {
2893 if (j - i >= ClMaxInlinePoisoningSize) {
2894 copyToShadowInline(ShadowMask, ShadowBytes, Done, i, IRB, ShadowBase);
2895 IRB.CreateCall(AsanSetShadowFunc[Val],
2896 {IRB.CreateAdd(ShadowBase, ConstantInt::get(IntptrTy, i)),
2897 ConstantInt::get(IntptrTy, j - i)});
2902 copyToShadowInline(ShadowMask, ShadowBytes, Done, End, IRB, ShadowBase);
2905 // Fake stack allocator (asan_fake_stack.h) has 11 size classes
2906 // for every power of 2 from kMinStackMallocSize to kMaxAsanStackMallocSizeClass
2907 static int StackMallocSizeClass(uint64_t LocalStackSize) {
2908 assert(LocalStackSize <= kMaxStackMallocSize);
2909 uint64_t MaxSize = kMinStackMallocSize;
2910 for (int i = 0;; i++, MaxSize *= 2)
2911 if (LocalStackSize <= MaxSize) return i;
2912 llvm_unreachable("impossible LocalStackSize");
2915 void FunctionStackPoisoner::copyArgsPassedByValToAllocas() {
2916 Instruction *CopyInsertPoint = &F.front().front();
2917 if (CopyInsertPoint == ASan.LocalDynamicShadow) {
2918 // Insert after the dynamic shadow location is determined
2919 CopyInsertPoint = CopyInsertPoint->getNextNode();
2920 assert(CopyInsertPoint);
2922 IRBuilder<> IRB(CopyInsertPoint);
2923 const DataLayout &DL = F.getParent()->getDataLayout();
2924 for (Argument &Arg : F.args()) {
2925 if (Arg.hasByValAttr()) {
2926 Type *Ty = Arg.getParamByValType();
2927 const Align Alignment =
2928 DL.getValueOrABITypeAlignment(Arg.getParamAlign(), Ty);
2930 AllocaInst *AI = IRB.CreateAlloca(
2932 (Arg.hasName() ? Arg.getName() : "Arg" + Twine(Arg.getArgNo())) +
2934 AI->setAlignment(Alignment);
2935 Arg.replaceAllUsesWith(AI);
2937 uint64_t AllocSize = DL.getTypeAllocSize(Ty);
2938 IRB.CreateMemCpy(AI, Alignment, &Arg, Alignment, AllocSize);
2943 PHINode *FunctionStackPoisoner::createPHI(IRBuilder<> &IRB, Value *Cond,
2945 Instruction *ThenTerm,
2946 Value *ValueIfFalse) {
2947 PHINode *PHI = IRB.CreatePHI(IntptrTy, 2);
2948 BasicBlock *CondBlock = cast<Instruction>(Cond)->getParent();
2949 PHI->addIncoming(ValueIfFalse, CondBlock);
2950 BasicBlock *ThenBlock = ThenTerm->getParent();
2951 PHI->addIncoming(ValueIfTrue, ThenBlock);
2955 Value *FunctionStackPoisoner::createAllocaForLayout(
2956 IRBuilder<> &IRB, const ASanStackFrameLayout &L, bool Dynamic) {
2959 Alloca = IRB.CreateAlloca(IRB.getInt8Ty(),
2960 ConstantInt::get(IRB.getInt64Ty(), L.FrameSize),
2963 Alloca = IRB.CreateAlloca(ArrayType::get(IRB.getInt8Ty(), L.FrameSize),
2964 nullptr, "MyAlloca");
2965 assert(Alloca->isStaticAlloca());
2967 assert((ClRealignStack & (ClRealignStack - 1)) == 0);
2968 uint64_t FrameAlignment = std::max(L.FrameAlignment, uint64_t(ClRealignStack));
2969 Alloca->setAlignment(Align(FrameAlignment));
2970 return IRB.CreatePointerCast(Alloca, IntptrTy);
2973 void FunctionStackPoisoner::createDynamicAllocasInitStorage() {
2974 BasicBlock &FirstBB = *F.begin();
2975 IRBuilder<> IRB(dyn_cast<Instruction>(FirstBB.begin()));
2976 DynamicAllocaLayout = IRB.CreateAlloca(IntptrTy, nullptr);
2977 IRB.CreateStore(Constant::getNullValue(IntptrTy), DynamicAllocaLayout);
2978 DynamicAllocaLayout->setAlignment(Align(32));
2981 void FunctionStackPoisoner::processDynamicAllocas() {
2982 if (!ClInstrumentDynamicAllocas || DynamicAllocaVec.empty()) {
2983 assert(DynamicAllocaPoisonCallVec.empty());
2987 // Insert poison calls for lifetime intrinsics for dynamic allocas.
2988 for (const auto &APC : DynamicAllocaPoisonCallVec) {
2989 assert(APC.InsBefore);
2991 assert(ASan.isInterestingAlloca(*APC.AI));
2992 assert(!APC.AI->isStaticAlloca());
2994 IRBuilder<> IRB(APC.InsBefore);
2995 poisonAlloca(APC.AI, APC.Size, IRB, APC.DoPoison);
2996 // Dynamic allocas will be unpoisoned unconditionally below in
2997 // unpoisonDynamicAllocas.
2998 // Flag that we need unpoison static allocas.
3001 // Handle dynamic allocas.
3002 createDynamicAllocasInitStorage();
3003 for (auto &AI : DynamicAllocaVec)
3004 handleDynamicAllocaCall(AI);
3005 unpoisonDynamicAllocas();
3008 /// Collect instructions in the entry block after \p InsBefore which initialize
3009 /// permanent storage for a function argument. These instructions must remain in
3010 /// the entry block so that uninitialized values do not appear in backtraces. An
3011 /// added benefit is that this conserves spill slots. This does not move stores
3012 /// before instrumented / "interesting" allocas.
3013 static void findStoresToUninstrumentedArgAllocas(
3014 AddressSanitizer &ASan, Instruction &InsBefore,
3015 SmallVectorImpl<Instruction *> &InitInsts) {
3016 Instruction *Start = InsBefore.getNextNonDebugInstruction();
3017 for (Instruction *It = Start; It; It = It->getNextNonDebugInstruction()) {
3018 // Argument initialization looks like:
3019 // 1) store <Argument>, <Alloca> OR
3020 // 2) <CastArgument> = cast <Argument> to ...
3021 // store <CastArgument> to <Alloca>
3022 // Do not consider any other kind of instruction.
3024 // Note: This covers all known cases, but may not be exhaustive. An
3025 // alternative to pattern-matching stores is to DFS over all Argument uses:
3026 // this might be more general, but is probably much more complicated.
3027 if (isa<AllocaInst>(It) || isa<CastInst>(It))
3029 if (auto *Store = dyn_cast<StoreInst>(It)) {
3030 // The store destination must be an alloca that isn't interesting for
3031 // ASan to instrument. These are moved up before InsBefore, and they're
3032 // not interesting because allocas for arguments can be mem2reg'd.
3033 auto *Alloca = dyn_cast<AllocaInst>(Store->getPointerOperand());
3034 if (!Alloca || ASan.isInterestingAlloca(*Alloca))
3037 Value *Val = Store->getValueOperand();
3038 bool IsDirectArgInit = isa<Argument>(Val);
3039 bool IsArgInitViaCast =
3040 isa<CastInst>(Val) &&
3041 isa<Argument>(cast<CastInst>(Val)->getOperand(0)) &&
3042 // Check that the cast appears directly before the store. Otherwise
3043 // moving the cast before InsBefore may break the IR.
3044 Val == It->getPrevNonDebugInstruction();
3045 bool IsArgInit = IsDirectArgInit || IsArgInitViaCast;
3049 if (IsArgInitViaCast)
3050 InitInsts.push_back(cast<Instruction>(Val));
3051 InitInsts.push_back(Store);
3055 // Do not reorder past unknown instructions: argument initialization should
3056 // only involve casts and stores.
3061 void FunctionStackPoisoner::processStaticAllocas() {
3062 if (AllocaVec.empty()) {
3063 assert(StaticAllocaPoisonCallVec.empty());
3067 int StackMallocIdx = -1;
3068 DebugLoc EntryDebugLocation;
3069 if (auto SP = F.getSubprogram())
3070 EntryDebugLocation =
3071 DILocation::get(SP->getContext(), SP->getScopeLine(), 0, SP);
3073 Instruction *InsBefore = AllocaVec[0];
3074 IRBuilder<> IRB(InsBefore);
3076 // Make sure non-instrumented allocas stay in the entry block. Otherwise,
3077 // debug info is broken, because only entry-block allocas are treated as
3078 // regular stack slots.
3079 auto InsBeforeB = InsBefore->getParent();
3080 assert(InsBeforeB == &F.getEntryBlock());
3081 for (auto *AI : StaticAllocasToMoveUp)
3082 if (AI->getParent() == InsBeforeB)
3083 AI->moveBefore(InsBefore);
3085 // Move stores of arguments into entry-block allocas as well. This prevents
3086 // extra stack slots from being generated (to house the argument values until
3087 // they can be stored into the allocas). This also prevents uninitialized
3088 // values from being shown in backtraces.
3089 SmallVector<Instruction *, 8> ArgInitInsts;
3090 findStoresToUninstrumentedArgAllocas(ASan, *InsBefore, ArgInitInsts);
3091 for (Instruction *ArgInitInst : ArgInitInsts)
3092 ArgInitInst->moveBefore(InsBefore);
3094 // If we have a call to llvm.localescape, keep it in the entry block.
3095 if (LocalEscapeCall) LocalEscapeCall->moveBefore(InsBefore);
3097 SmallVector<ASanStackVariableDescription, 16> SVD;
3098 SVD.reserve(AllocaVec.size());
3099 for (AllocaInst *AI : AllocaVec) {
3100 ASanStackVariableDescription D = {AI->getName().data(),
3101 ASan.getAllocaSizeInBytes(*AI),
3103 AI->getAlign().value(),
3110 // Minimal header size (left redzone) is 4 pointers,
3111 // i.e. 32 bytes on 64-bit platforms and 16 bytes in 32-bit platforms.
3112 uint64_t Granularity = 1ULL << Mapping.Scale;
3113 uint64_t MinHeaderSize = std::max((uint64_t)ASan.LongSize / 2, Granularity);
3114 const ASanStackFrameLayout &L =
3115 ComputeASanStackFrameLayout(SVD, Granularity, MinHeaderSize);
3117 // Build AllocaToSVDMap for ASanStackVariableDescription lookup.
3118 DenseMap<const AllocaInst *, ASanStackVariableDescription *> AllocaToSVDMap;
3119 for (auto &Desc : SVD)
3120 AllocaToSVDMap[Desc.AI] = &Desc;
3122 // Update SVD with information from lifetime intrinsics.
3123 for (const auto &APC : StaticAllocaPoisonCallVec) {
3124 assert(APC.InsBefore);
3126 assert(ASan.isInterestingAlloca(*APC.AI));
3127 assert(APC.AI->isStaticAlloca());
3129 ASanStackVariableDescription &Desc = *AllocaToSVDMap[APC.AI];
3130 Desc.LifetimeSize = Desc.Size;
3131 if (const DILocation *FnLoc = EntryDebugLocation.get()) {
3132 if (const DILocation *LifetimeLoc = APC.InsBefore->getDebugLoc().get()) {
3133 if (LifetimeLoc->getFile() == FnLoc->getFile())
3134 if (unsigned Line = LifetimeLoc->getLine())
3135 Desc.Line = std::min(Desc.Line ? Desc.Line : Line, Line);
3140 auto DescriptionString = ComputeASanStackFrameDescription(SVD);
3141 LLVM_DEBUG(dbgs() << DescriptionString << " --- " << L.FrameSize << "\n");
3142 uint64_t LocalStackSize = L.FrameSize;
3143 bool DoStackMalloc =
3144 ASan.UseAfterReturn != AsanDetectStackUseAfterReturnMode::Never &&
3145 !ASan.CompileKernel && LocalStackSize <= kMaxStackMallocSize;
3146 bool DoDynamicAlloca = ClDynamicAllocaStack;
3147 // Don't do dynamic alloca or stack malloc if:
3148 // 1) There is inline asm: too often it makes assumptions on which registers
3150 // 2) There is a returns_twice call (typically setjmp), which is
3151 // optimization-hostile, and doesn't play well with introduced indirect
3152 // register-relative calculation of local variable addresses.
3153 DoDynamicAlloca &= !HasInlineAsm && !HasReturnsTwiceCall;
3154 DoStackMalloc &= !HasInlineAsm && !HasReturnsTwiceCall;
3156 Value *StaticAlloca =
3157 DoDynamicAlloca ? nullptr : createAllocaForLayout(IRB, L, false);
3160 Value *LocalStackBase;
3161 Value *LocalStackBaseAlloca;
3162 uint8_t DIExprFlags = DIExpression::ApplyOffset;
3164 if (DoStackMalloc) {
3165 LocalStackBaseAlloca =
3166 IRB.CreateAlloca(IntptrTy, nullptr, "asan_local_stack_base");
3167 if (ASan.UseAfterReturn == AsanDetectStackUseAfterReturnMode::Runtime) {
3168 // void *FakeStack = __asan_option_detect_stack_use_after_return
3169 // ? __asan_stack_malloc_N(LocalStackSize)
3171 // void *LocalStackBase = (FakeStack) ? FakeStack :
3172 // alloca(LocalStackSize);
3173 Constant *OptionDetectUseAfterReturn = F.getParent()->getOrInsertGlobal(
3174 kAsanOptionDetectUseAfterReturn, IRB.getInt32Ty());
3175 Value *UseAfterReturnIsEnabled = IRB.CreateICmpNE(
3176 IRB.CreateLoad(IRB.getInt32Ty(), OptionDetectUseAfterReturn),
3177 Constant::getNullValue(IRB.getInt32Ty()));
3179 SplitBlockAndInsertIfThen(UseAfterReturnIsEnabled, InsBefore, false);
3180 IRBuilder<> IRBIf(Term);
3181 StackMallocIdx = StackMallocSizeClass(LocalStackSize);
3182 assert(StackMallocIdx <= kMaxAsanStackMallocSizeClass);
3183 Value *FakeStackValue =
3184 IRBIf.CreateCall(AsanStackMallocFunc[StackMallocIdx],
3185 ConstantInt::get(IntptrTy, LocalStackSize));
3186 IRB.SetInsertPoint(InsBefore);
3187 FakeStack = createPHI(IRB, UseAfterReturnIsEnabled, FakeStackValue, Term,
3188 ConstantInt::get(IntptrTy, 0));
3190 // assert(ASan.UseAfterReturn == AsanDetectStackUseAfterReturnMode:Always)
3191 // void *FakeStack = __asan_stack_malloc_N(LocalStackSize);
3192 // void *LocalStackBase = (FakeStack) ? FakeStack :
3193 // alloca(LocalStackSize);
3194 StackMallocIdx = StackMallocSizeClass(LocalStackSize);
3195 FakeStack = IRB.CreateCall(AsanStackMallocFunc[StackMallocIdx],
3196 ConstantInt::get(IntptrTy, LocalStackSize));
3198 Value *NoFakeStack =
3199 IRB.CreateICmpEQ(FakeStack, Constant::getNullValue(IntptrTy));
3201 SplitBlockAndInsertIfThen(NoFakeStack, InsBefore, false);
3202 IRBuilder<> IRBIf(Term);
3203 Value *AllocaValue =
3204 DoDynamicAlloca ? createAllocaForLayout(IRBIf, L, true) : StaticAlloca;
3206 IRB.SetInsertPoint(InsBefore);
3207 LocalStackBase = createPHI(IRB, NoFakeStack, AllocaValue, Term, FakeStack);
3208 IRB.CreateStore(LocalStackBase, LocalStackBaseAlloca);
3209 DIExprFlags |= DIExpression::DerefBefore;
3211 // void *FakeStack = nullptr;
3212 // void *LocalStackBase = alloca(LocalStackSize);
3213 FakeStack = ConstantInt::get(IntptrTy, 0);
3215 DoDynamicAlloca ? createAllocaForLayout(IRB, L, true) : StaticAlloca;
3216 LocalStackBaseAlloca = LocalStackBase;
3219 // It shouldn't matter whether we pass an `alloca` or a `ptrtoint` as the
3220 // dbg.declare address opereand, but passing a `ptrtoint` seems to confuse
3221 // later passes and can result in dropped variable coverage in debug info.
3222 Value *LocalStackBaseAllocaPtr =
3223 isa<PtrToIntInst>(LocalStackBaseAlloca)
3224 ? cast<PtrToIntInst>(LocalStackBaseAlloca)->getPointerOperand()
3225 : LocalStackBaseAlloca;
3226 assert(isa<AllocaInst>(LocalStackBaseAllocaPtr) &&
3227 "Variable descriptions relative to ASan stack base will be dropped");
3229 // Replace Alloca instructions with base+offset.
3230 for (const auto &Desc : SVD) {
3231 AllocaInst *AI = Desc.AI;
3232 replaceDbgDeclare(AI, LocalStackBaseAllocaPtr, DIB, DIExprFlags,
3234 Value *NewAllocaPtr = IRB.CreateIntToPtr(
3235 IRB.CreateAdd(LocalStackBase, ConstantInt::get(IntptrTy, Desc.Offset)),
3237 AI->replaceAllUsesWith(NewAllocaPtr);
3240 // The left-most redzone has enough space for at least 4 pointers.
3241 // Write the Magic value to redzone[0].
3242 Value *BasePlus0 = IRB.CreateIntToPtr(LocalStackBase, IntptrPtrTy);
3243 IRB.CreateStore(ConstantInt::get(IntptrTy, kCurrentStackFrameMagic),
3245 // Write the frame description constant to redzone[1].
3246 Value *BasePlus1 = IRB.CreateIntToPtr(
3247 IRB.CreateAdd(LocalStackBase,
3248 ConstantInt::get(IntptrTy, ASan.LongSize / 8)),
3250 GlobalVariable *StackDescriptionGlobal =
3251 createPrivateGlobalForString(*F.getParent(), DescriptionString,
3252 /*AllowMerging*/ true, kAsanGenPrefix);
3253 Value *Description = IRB.CreatePointerCast(StackDescriptionGlobal, IntptrTy);
3254 IRB.CreateStore(Description, BasePlus1);
3255 // Write the PC to redzone[2].
3256 Value *BasePlus2 = IRB.CreateIntToPtr(
3257 IRB.CreateAdd(LocalStackBase,
3258 ConstantInt::get(IntptrTy, 2 * ASan.LongSize / 8)),
3260 IRB.CreateStore(IRB.CreatePointerCast(&F, IntptrTy), BasePlus2);
3262 const auto &ShadowAfterScope = GetShadowBytesAfterScope(SVD, L);
3264 // Poison the stack red zones at the entry.
3265 Value *ShadowBase = ASan.memToShadow(LocalStackBase, IRB);
3266 // As mask we must use most poisoned case: red zones and after scope.
3267 // As bytes we can use either the same or just red zones only.
3268 copyToShadow(ShadowAfterScope, ShadowAfterScope, IRB, ShadowBase);
3270 if (!StaticAllocaPoisonCallVec.empty()) {
3271 const auto &ShadowInScope = GetShadowBytes(SVD, L);
3273 // Poison static allocas near lifetime intrinsics.
3274 for (const auto &APC : StaticAllocaPoisonCallVec) {
3275 const ASanStackVariableDescription &Desc = *AllocaToSVDMap[APC.AI];
3276 assert(Desc.Offset % L.Granularity == 0);
3277 size_t Begin = Desc.Offset / L.Granularity;
3278 size_t End = Begin + (APC.Size + L.Granularity - 1) / L.Granularity;
3280 IRBuilder<> IRB(APC.InsBefore);
3281 copyToShadow(ShadowAfterScope,
3282 APC.DoPoison ? ShadowAfterScope : ShadowInScope, Begin, End,
3287 SmallVector<uint8_t, 64> ShadowClean(ShadowAfterScope.size(), 0);
3288 SmallVector<uint8_t, 64> ShadowAfterReturn;
3290 // (Un)poison the stack before all ret instructions.
3291 for (Instruction *Ret : RetVec) {
3292 IRBuilder<> IRBRet(Ret);
3293 // Mark the current frame as retired.
3294 IRBRet.CreateStore(ConstantInt::get(IntptrTy, kRetiredStackFrameMagic),
3296 if (DoStackMalloc) {
3297 assert(StackMallocIdx >= 0);
3298 // if FakeStack != 0 // LocalStackBase == FakeStack
3299 // // In use-after-return mode, poison the whole stack frame.
3300 // if StackMallocIdx <= 4
3301 // // For small sizes inline the whole thing:
3302 // memset(ShadowBase, kAsanStackAfterReturnMagic, ShadowSize);
3303 // **SavedFlagPtr(FakeStack) = 0
3305 // __asan_stack_free_N(FakeStack, LocalStackSize)
3307 // <This is not a fake stack; unpoison the redzones>
3309 IRBRet.CreateICmpNE(FakeStack, Constant::getNullValue(IntptrTy));
3310 Instruction *ThenTerm, *ElseTerm;
3311 SplitBlockAndInsertIfThenElse(Cmp, Ret, &ThenTerm, &ElseTerm);
3313 IRBuilder<> IRBPoison(ThenTerm);
3314 if (StackMallocIdx <= 4) {
3315 int ClassSize = kMinStackMallocSize << StackMallocIdx;
3316 ShadowAfterReturn.resize(ClassSize / L.Granularity,
3317 kAsanStackUseAfterReturnMagic);
3318 copyToShadow(ShadowAfterReturn, ShadowAfterReturn, IRBPoison,
3320 Value *SavedFlagPtrPtr = IRBPoison.CreateAdd(
3322 ConstantInt::get(IntptrTy, ClassSize - ASan.LongSize / 8));
3323 Value *SavedFlagPtr = IRBPoison.CreateLoad(
3324 IntptrTy, IRBPoison.CreateIntToPtr(SavedFlagPtrPtr, IntptrPtrTy));
3325 IRBPoison.CreateStore(
3326 Constant::getNullValue(IRBPoison.getInt8Ty()),
3327 IRBPoison.CreateIntToPtr(SavedFlagPtr, IRBPoison.getInt8PtrTy()));
3329 // For larger frames call __asan_stack_free_*.
3330 IRBPoison.CreateCall(
3331 AsanStackFreeFunc[StackMallocIdx],
3332 {FakeStack, ConstantInt::get(IntptrTy, LocalStackSize)});
3335 IRBuilder<> IRBElse(ElseTerm);
3336 copyToShadow(ShadowAfterScope, ShadowClean, IRBElse, ShadowBase);
3338 copyToShadow(ShadowAfterScope, ShadowClean, IRBRet, ShadowBase);
3342 // We are done. Remove the old unused alloca instructions.
3343 for (auto AI : AllocaVec) AI->eraseFromParent();
3346 void FunctionStackPoisoner::poisonAlloca(Value *V, uint64_t Size,
3347 IRBuilder<> &IRB, bool DoPoison) {
3348 // For now just insert the call to ASan runtime.
3349 Value *AddrArg = IRB.CreatePointerCast(V, IntptrTy);
3350 Value *SizeArg = ConstantInt::get(IntptrTy, Size);
3352 DoPoison ? AsanPoisonStackMemoryFunc : AsanUnpoisonStackMemoryFunc,
3353 {AddrArg, SizeArg});
3356 // Handling llvm.lifetime intrinsics for a given %alloca:
3357 // (1) collect all llvm.lifetime.xxx(%size, %value) describing the alloca.
3358 // (2) if %size is constant, poison memory for llvm.lifetime.end (to detect
3359 // invalid accesses) and unpoison it for llvm.lifetime.start (the memory
3360 // could be poisoned by previous llvm.lifetime.end instruction, as the
3361 // variable may go in and out of scope several times, e.g. in loops).
3362 // (3) if we poisoned at least one %alloca in a function,
3363 // unpoison the whole stack frame at function exit.
3364 void FunctionStackPoisoner::handleDynamicAllocaCall(AllocaInst *AI) {
3365 IRBuilder<> IRB(AI);
3367 const Align Alignment = std::max(Align(kAllocaRzSize), AI->getAlign());
3368 const uint64_t AllocaRedzoneMask = kAllocaRzSize - 1;
3370 Value *Zero = Constant::getNullValue(IntptrTy);
3371 Value *AllocaRzSize = ConstantInt::get(IntptrTy, kAllocaRzSize);
3372 Value *AllocaRzMask = ConstantInt::get(IntptrTy, AllocaRedzoneMask);
3374 // Since we need to extend alloca with additional memory to locate
3375 // redzones, and OldSize is number of allocated blocks with
3376 // ElementSize size, get allocated memory size in bytes by
3377 // OldSize * ElementSize.
3378 const unsigned ElementSize =
3379 F.getParent()->getDataLayout().getTypeAllocSize(AI->getAllocatedType());
3381 IRB.CreateMul(IRB.CreateIntCast(AI->getArraySize(), IntptrTy, false),
3382 ConstantInt::get(IntptrTy, ElementSize));
3384 // PartialSize = OldSize % 32
3385 Value *PartialSize = IRB.CreateAnd(OldSize, AllocaRzMask);
3387 // Misalign = kAllocaRzSize - PartialSize;
3388 Value *Misalign = IRB.CreateSub(AllocaRzSize, PartialSize);
3390 // PartialPadding = Misalign != kAllocaRzSize ? Misalign : 0;
3391 Value *Cond = IRB.CreateICmpNE(Misalign, AllocaRzSize);
3392 Value *PartialPadding = IRB.CreateSelect(Cond, Misalign, Zero);
3394 // AdditionalChunkSize = Alignment + PartialPadding + kAllocaRzSize
3395 // Alignment is added to locate left redzone, PartialPadding for possible
3396 // partial redzone and kAllocaRzSize for right redzone respectively.
3397 Value *AdditionalChunkSize = IRB.CreateAdd(
3398 ConstantInt::get(IntptrTy, Alignment.value() + kAllocaRzSize),
3401 Value *NewSize = IRB.CreateAdd(OldSize, AdditionalChunkSize);
3403 // Insert new alloca with new NewSize and Alignment params.
3404 AllocaInst *NewAlloca = IRB.CreateAlloca(IRB.getInt8Ty(), NewSize);
3405 NewAlloca->setAlignment(Alignment);
3407 // NewAddress = Address + Alignment
3409 IRB.CreateAdd(IRB.CreatePtrToInt(NewAlloca, IntptrTy),
3410 ConstantInt::get(IntptrTy, Alignment.value()));
3412 // Insert __asan_alloca_poison call for new created alloca.
3413 IRB.CreateCall(AsanAllocaPoisonFunc, {NewAddress, OldSize});
3415 // Store the last alloca's address to DynamicAllocaLayout. We'll need this
3416 // for unpoisoning stuff.
3417 IRB.CreateStore(IRB.CreatePtrToInt(NewAlloca, IntptrTy), DynamicAllocaLayout);
3419 Value *NewAddressPtr = IRB.CreateIntToPtr(NewAddress, AI->getType());
3421 // Replace all uses of AddessReturnedByAlloca with NewAddressPtr.
3422 AI->replaceAllUsesWith(NewAddressPtr);
3424 // We are done. Erase old alloca from parent.
3425 AI->eraseFromParent();
3428 // isSafeAccess returns true if Addr is always inbounds with respect to its
3429 // base object. For example, it is a field access or an array access with
3430 // constant inbounds index.
3431 bool AddressSanitizer::isSafeAccess(ObjectSizeOffsetVisitor &ObjSizeVis,
3432 Value *Addr, uint64_t TypeSize) const {
3433 SizeOffsetType SizeOffset = ObjSizeVis.compute(Addr);
3434 if (!ObjSizeVis.bothKnown(SizeOffset)) return false;
3435 uint64_t Size = SizeOffset.first.getZExtValue();
3436 int64_t Offset = SizeOffset.second.getSExtValue();
3437 // Three checks are required to ensure safety:
3438 // . Offset >= 0 (since the offset is given from the base ptr)
3439 // . Size >= Offset (unsigned)
3440 // . Size - Offset >= NeededSize (unsigned)
3441 return Offset >= 0 && Size >= uint64_t(Offset) &&
3442 Size - uint64_t(Offset) >= TypeSize / 8;